text stringlengths 2.85k 2.55M | label class label 11 classes |
|---|---|
ALMOST MAXIMAL GROWTH OF THE HILBERT FUNCTION
arXiv:1403.1409v1 [math.AG] 6 Mar 2014
LUCA CHIANTINI AND JUAN MIGLIORE∗
A BSTRACT. Let A = S/J be a standard artinian graded algebra over the polynomial ring S. A
theorem of Macaulay dictates the possible growth of the Hilbert function of A from any degree
to the next, and if this growth is the maximal possible then strong consequences have been given
by Gotzmann. It can be phrased in terms of the base locus of the linear system defined by the
relevant component(s) of J. If J is the artinian reduction of the ideal of a finite set of points in
projective space then this maximal growth for A was shown by Bigatti, Geramita and the second
author to imply strong geometric consequences for the points. We now suppose that the growth
of the Hilbert function is one less than maximal. This again has (not as) strong consequences
for the base locus defined by the relevant component. And when J is the artinian reduction of
the ideal of a finite set of points in projective space, we prove that almost maximal growth again
forces geometric consequences.
C ONTENTS
1. Introduction
2. Preliminaries
3. Forcing a base locus
4. Growth type (. . . , k, k − 1): when there is a non-empty base locus
5. Growth type (. . . , k, k − 1): the basepoint free case
6. Extending Gotzmann’s theorem
7. Further questions
References
1
3
5
16
25
32
33
34
1. I NTRODUCTION
Let S = K[x1 . . . , xr ], where K is an
L algebraically closed field of any characteristic. Let
J ⊂ S be a homogeneous ideal, so J = t∈Z [J]t . The Hilbert function of S/J is the numerical
function defined by hS/J (t) = dimK [S/J]t . For any positive integer n, a theorem of Macaulay
gives an upper bound for hS/J (n + 1) in terms of hS/J (n) and n. When hS/J (n + 1) achieves
2010 Mathematics Subject Classification. Primary 13D40; Secondary 14M05 .
Key words and phrases. Hilbert function, maximal growth, Macaulay’s theorem, binomial expansion, base
locus.
∗
Part of the work for this paper was done while this author was sponsored by the National Security Agency
under Grant Number H98230-12-1-0204, and by the Simons Foundation under grant #208579. He also thanks the
Italian CNR-GNSAGA.
2
L. CHIANTINI AND J. MIGLIORE
this bound, we say that S/J has maximal growth in degree n. If hS/J (n+1) fails by 1 to achieve
the bound, we say that S/J has almost maximal growth in degree n.
In this paper we give results assuming almost maximal growth for S/J in degree n to degree
n + 1. We consider two settings. First, the setting of arbitrary homogeneous ideals. Secondly,
we will restrict ourselves to the artinian reductions of ideals of sets of points: if Z ⊂ Pr is
a finite set of points, with homogeneous ideal IZ , and L is a general linear form, the ideal
Z ,Li
⊂ S is the artinian reduction of IZ by L.
J = hIhLi
For fixed n, the projectivization of the vector space [J]n is a linear system of hypersurfaces
in Pr−1 . As such, it may be basepoint-free or it may have a non-empty base locus. The starting
observation of our research is that if S/J has maximal growth, then a result of Gotzmann
[18] forces the existence of a non-empty base locus, and it gives the dimension and degree of
this locus (as a scheme) through the Hilbert polynomial. For the artinian reduction of a set of
points Z, in [5] Bigatti, Geramita and the second author apply Gotzmann’s result, giving careful
information about the decomposition of Z into the subset lying on the base locus and the subset
off the base locus (this is also a generalization of an old result of Davis for points in P2 , see
[14]). These results are recalled in section 2.
In the case of almost maximal growth, Gotzmann’s result no longer applies, and one focus
of this paper is the fact that more possibilities arise than in the case of maximal growth. We
give results that explain what kinds of base loci can occur, depending on the kind of almost
maximal growth involved. (This latter is a technicality arising from Macaulay’s theorem and
will be made precise in sections 2 and 3.) In almost all cases, a base locus is forced, but now it
can have one of two possible dimensions, and different degrees; see for instance theorem 3.1,
example 3.3 and theorem 3.4. These results are given in section 3.
In the only case where a base locus is not forced, it can still happen that a base locus exists.
This is the topic of section 4. In this case, turning to the artinian reductions of ideals of sets of
points, we give results about a decomposition of Z, analogous to the one mentioned above.
The last possibility is that there is no base locus. This is dealt with in section 5. In this
situation we find our most surprising result. Despite the absence of a base locus, it still turns out
that many of the points are forced to lie on a plane. This is Theorem 5.1. It is very reminiscent
of a result of Maroscia [24], which says essentially that if the value of the Hilbert function of
S/J takes two or more consecutive values less than r then many of the points must lie on a
linear space of specified dimension.
There are several motivations for studying the geometry of sets of points, when the artinian
reduction achieves almost maximal growth. We should mention that a similar analysis, in the
case of sets of points with uniform position properties, allowed results for the theory of curves:
bounds on the genus, existence of special linear series, postulation of nodes of general plane
projections (see [9], [10], [8]). Recent applications to the study of symmetric tensors (for which
the uniform position properties may not hold) can be found in [1], [7], [2].
At the end of the paper, we make a series of remarks on related problems and further extensions.
ALMOST MAXIMAL GROWTH OF THE HILBERT FUNCTION
3
2. P RELIMINARIES
Let S = K[x1 , . . . , xr ], where K is an algebraically
closed field of arbitrary characteristic.
L
Let J ⊂ S be a homogeneous ideal, so J = t∈Z [J]t .
Definition 2.1. The Hilbert function of S/J is the numerical function defined by hS/J (t) =
dim[S/J]t for t ≥ 0.
We begin by recalling results of Macaulay and Green that we will need in this paper.
Definition 2.2. Let k and i be positive integers. The i-binomial expansion of k is
kj
ki−1
ki
,
+ ... +
+
k(i) =
j
i−1
i
where ki > ki−1 > ... > kj ≥ j ≥ 1. We remark that such an expansion always exists and it is
unique (see, e.g., [6] Lemma 4.2.6).
Notation 2.3. Following [4], we define, for any integers a and b,
b
kj + b
ki−1 + b
ki + b
,
+ ... +
k(i) a =
+
j+a
i−1+a
i+a
where we set mq = 0 whenever m < q or q < 0. Furthermore, we will set
−1
1
k hii = k(i) 1 and khii = k(i) 0 .
Theorem 2.4. Let L ∈ [S]1 be a general linear form and let A = S/J be a standard graded
′
algebra. Denote by hn the degree n entry of the Hilbert function of A and by hn the degree n
entry of the Hilbert function of A/LA. Then:
(i) (Macaulay) hn+1 ≤ hhni
n .
′
(ii) (Green) hn ≤ (hn )hni .
(iii) (Gotzmann Persistence Theorem) Assume that hA (t) has maximal growth from degree n
to degree n + 1 and that J has no minimal generator in degree ≥ n + 2. Then
d
hA (n + d) = (hA (n))(n)
d
for all d ≥ 0. In particular, the Hilbert function equals the Hilbert polynomial in all
degrees ≥ n.
Proof. (i) See [6], Theorem 4.2.10.
(ii) See [19], Theorem 1.
(iii) See [18].
Definition 2.5. If hA (n + 1) = hA (n)hni , we say that hA has maximal growth in degree n. We
also sometimes say that A has maximal growth from degree n to degree n + 1. If hA (n + 1) =
hA (n)hni − 1, we say that hA (or simply A) has almost maximal growth in degree n.
One can see from the result of Gotzmann that there are different kinds of maximal growth of
the Hilbert function, depending on the degree of the corresponding Hilbert polynomial.
4
L. CHIANTINI AND J. MIGLIORE
Definition 2.6. Fix positive integers n and k. Suppose that the n-binomial expansion of k is
kn−1
kn
+ ....
+
k=
n−1
n
Then the MG-dimension of k in degree n, denoted MG(k, n), is kn − n.
Indeed, if hA (n) = knn + (lower degree terms), hA has maximal growth from degree n to
degree n + 1 and J has no generators in degree ≥ n + 2, then the Hilbert polynomial of A has
degree kn − n, by Gotzmann’s theorem. This means that the base locus of the linear system is
a scheme of dimension kn − n in Pr−1 .
Notice that we reduce to the case where J has no minimal generators in degree ≥ i if we
substitute J with its truncation J≤i , i.e. the ideal generated by the generators of J which have
degree ≤ i. The ideal J≤i defines the base locus of [Ji ], although it is not necessarily saturated.
Given any value hA (n), the above discussion shows that MG(k, n) is the dimension for the
base locus of [J]n , in case hA (n + 1) is the maximum allowed by Macaulay’s theorem.
Let R = K[x0 , . . . , xr ] and let Z ⊂ Pn be a reduced, zero-dimensional scheme. We will
denote by hZ the Hilbert function of R/IZ . Let L ∈ [R]1 be a general linear form. The ideal
Z ,Li
J = hIhLi
⊂ R/hLi ∼
= S is the artinian reduction of R/IZ , and its Hilbert function is the
h-vector of Z, or of R/IZ . It is possible for R/IZ to fail to have maximal growth from degree
n to degree n + 1, but nevertheless S/J does have maximal growth there (see [5] Example 1.3
(a)).
The main idea of [5] was that if Z ⊂ Pr is a reduced, zero-dimensional scheme with artinian
reduction J, such that S/J has maximal growth in degree n (even if it is not the case that R/IZ
has maximal growth there), this still has strong implications for the geometry of Z. The proofs
in [5] heavily used the Gotzmann Persistence Theorem.
Our interest in this paper focuses on the situation where the growth of the first difference of
the Hilbert function is almost maximal as defined in Definition 2.5. As mentioned above, there
are different kinds of maximal growth, so correspondingly there are different cases of almost
maximal growth to be analyzed. Although this paper should be viewed as a generalization of
parts of [5], it should also be noted that [5] was motivated by the following theorem of E.D.
Davis, which we state more in the language of this paper. (He actually proved a bit more,
including the fact that he did not assume that Z is reduced).
Theorem 2.7 (see [14] Theorem 4.1 and Corollary 4.2). Let Z ⊂ P2 be a reduced, finite set of
points with homogeneous ideal IZ . Assume that ∆hZ (n) = ∆hZ (n + 1) = k. Then [IZ ]n and
[IZ ]n+1 have a GCD, say F , of degree k. Furthermore, F defines a reduced curve and
(a)
(b)
(c)
(d)
(IZ : F ) is the saturated ideal of the subset, Z2 , of Z not lying on the curve defined by F .
(IZ , F ) is the saturated ideal of the subset of Z1 lying on F .
∆hZ (t) = ∆hZ2 (t − k) + hZ1 (t) for all t.
∆hZ2 (t) = 0 for t ≥ n − k.
Remark 2.8. A scheme X with ideal sheaf IX is t-regular if H i (IX (t − i) = 0 for i > 0.
As in [5], we will need the following results from [20].
ALMOST MAXIMAL GROWTH OF THE HILBERT FUNCTION
5
(a) ([20] Theorem 1.1) If X ⊂ Pn is a reduced irreducible non-degenerate curve of degree d
then IX is (d + 2 − n)-regular.
(b) ([20] Remark (1), page 497) Let X ⊂ Pn be a reduced but not necessarily irreducible
curve. Suppose X has irreducible components Xi of degree di , and that Xi spans a
Pni ⊂ Pn . Set
di + 2 − ni , if di ≥ 2;
mi =
1
if di = 1 (i.e., if Xi is a line)
P
Then X is ( mi )-regular.
3. F ORCING
A BASE LOCUS
In this section we investigate when almost maximal growth of the Hilbert function of a standard graded algebra S/J from degree n to degree n + 1 forces the existence of a base locus
in [J]n . We do not yet assume that J is the artinian reduction of the ideal of a set of points in
projective space.
The next definition merely gives a notation for the dimension of the base locus of the component of an ideal in a degree n under the assumption that the quotient has maximal growth
(rather than almost maximal growth) from degree n to degree n + 1.
Let J ⊂ S = k[x1 , . . . , xr ] be a homogeneous ideal and let A = S/J. Suppose that hA has
maximal growth from degree n to degree n + 1. Then the base locus of [J]n and the base locus
of [J]n+1 , viewed in Pr−1 , coincide and this locus is a scheme of dimension MG(hA (n), n).
If J is the artinian reduction of the homogeneous ideal of a zero-dimensional scheme Z by a
general linear form, then [IZ ]n has base locus of dimension MG(hA (n), n) + 1, and the same
holds for [IZ ]n+1 . This is the idea behind [5]. We now extend this idea to our setting.
The first step is the case where the MG-dimension of S/J is 1. We assume that the growth of
the Hilbert function of S/J is one less than maximal, and we will show that the dimension of the
base locus is either the same as that occurring for maximal growth, or one less. In Example 3.3
we show that both possibilities for the dimension of the base locus can occur, and we indicate
some possibilities for the degree of the base locus (recalling that for maximal growth the degree
and dimension are both forced).
Theorem 3.1. Let J ⊂ S be a homogeneous ideal, and let A = S/J. Assume that
n+1
+ (lower terms)
(1)
hA (n) =
n
is the n-binomial expansion of hA (n). Equivalently, assume that n + 1 ≤ hA (n) < n+2
.
2
hni
Assume further that hA (n + 1) = hA (n) − 1, i.e. that the growth is one less than maximal.
Then the base locus of [J]n in Pr−1 exists and it is a scheme of dimension either 0 or 1. In other
words, the dimension of the base locus is either MG(hA (n), n) − 1 or MG(hA (n), n).
Proof. For any homogeneous ideal I ⊂ S and linear form L we have the exact sequence
6
L. CHIANTINI AND J. MIGLIORE
0
I:L
(−1)
I
(2)
S/I(−1)
×L
S/I
S/(I, L)
0
S/(I : L)(−1)
0
0
We shall use this sequence repeatedly, especially the second short exact sequence. Notice
that S/(I : L)(−1) is the image of S/I(−1) in S/I under multiplication by L.
We first note that under our hypotheses, the base locus cannot have dimension > 1. Indeed,
the Hilbert polynomial of S/J≤n has degree at most 1. We just have to show that the base locus
is not empty.
Let ℓ be a general linear form. For convenience let us denote by
k = hA (n)
p = dim hS/(J,ℓ) (n)
s = dim hS/(J,ℓ) (n + 1)
Thus we have the table
degree i 0
1
dim[S/J]i 1
r
dim[S/(J : ℓ)]i−1
1
dim[S/(J, ℓ)]i 1 r − 1
...
n
n+1
...
k
k hni − 1
. . . k − p k hni − 1 − s
...
p
s
There exist q and m giving the following binomial expansions:
n−q+2
n
n+1
n−q−m+1
n−q
+···+
+
k =
+···+
+
n−q+1
n−1
n
n−q−m+1
n−q
|
{z
} |
{z
}
q
terms
m
terms
n−q+2
n
n+1
+m
+···+
+
=
n−q+1
n−1
n
and
n−q+3
n+1
n+2
+ m.
+···+
+
k =
n−q+2
n
n+1
By Green’s theorem, and then by Macaulay’s theorem we have
hni
p ≤ q ≤ n and
s ≤ p ≤ q.
Notice that
m ≤ n − q, and in particular if n = q then m = 0.
We have from the condition (1) that n < hA (n) = k < n+2
= n+2
.
2
n
From the exact sequence (2) we also know that
k hni − 1 − s ≤ k
ALMOST MAXIMAL GROWTH OF THE HILBERT FUNCTION
7
so, by the properties of binomials,
s ≥
k hni − k− 1
n−q+2
n+2
−1
−
=
n−q+1
n+1
= q − 1.
All together we have
q − 1 ≤ s ≤ p ≤ q ≤ n.
This means that the possibilities for (p, s) are (q − 1, q − 1), (q, q − 1) or (q, q). The first and
third of these represent maximal growth for S/(J, ℓ) since q ≤ n, so the base locus for [(J, ℓ)]n
is zero-dimensional, by Theorem 2.4 (iii). Thus (since ℓ was general) the base locus for [J]n
has dimension 1 = MG(hA (n), n).
Assume now p = q, s = q − 1. This means
dim[S/(J : ℓ)]n−1 = k − q, and dim[S/(J : ℓ)]n = k hni − q,
i.e. we have the table
degree i 0
1
dim[S/J]i 1
r
dim[S/(J : ℓ)]i−1
1
dim[S/(J, ℓ)]i 1 r − 1
...
n
n+1
...
k
k hni − 1
. . . k − q k hni − q
...
q
q−1
(Notice that the passage from k to k hni − 1 occurs from degree n to degree n + 1, while the
growth from k − q to k hni − q is from degree n − 1 to degree n.) We have
n+1
n−q+2
n
+m−q
+
+···+
k − q = hA (n) − q =
n
n −1
n − q + 1
n−q+2
n
+n+1+m−q
+···+
=
n−q+1
n−1
and similarly
k
hni
−q =
n−q+3
n+1
+ n + 2 + m − q.
+···+
n−q+2
n
We have the following cases.
◮ Case 1. If n = q (hence m = 0) then
k−q =
and
2
n
+1
+···+
1
n−1
3
n+1
+2
+···+
k −q =
2
n
which exceeds maximal growth from degree n − 1 to degree n, so this case cannot occur.
hni
8
L. CHIANTINI AND J. MIGLIORE
k−q
=
=
n+1
n
◮ Case 2. If n > q and m < n − q then we get
−q =
n
n−q+2
n−q+1
+ ···+
+
n−1
n−q+1
n−q
n−q+1
n−q
, so that
+m
n
n−q+2
n−q+1
n−q−2
n−q−1
+ ···+
+
+
+
+ . . .
n−1
n−q+1
n−q
n−q−2
n−q−1
{z
}
|
m times
and similarly
k
hni
−q
=
=
n+1
n−q+3
n−q+2
+ ··· +
+
+m
n
n−q+2
n−q+1
n−q−1
n+1
n−q+3
n−q+2
n−q
+
+ . . .
+ ···+
+
+
n−q
n−q−1
n
n−q+2
n−q+1
{z
}
|
m times
which is clearly maximal growth. In this case we have that the base locus of [J : ℓ]n−1
and of [J : ℓ]n has dimension 1, so the base locus of [J]n has dimension at least 1 =
MG(hA (n), n) since [J]n ⊂ [J : ℓ]n .
◮ Case 3. If n > q and m = n − q then we have as above
n−q+1
n−q+2
n
+m
+
+···+
k−q =
n−q
n−q+1
n−1
and
n
−
q
+
2
n
−
q
+
3
n
+
1
+m
+
+···+
k hni − q =
n−q+1
n−q+2
n
but now m > n−q−1
+ · · · + 11 , so we cannot conclude as in the previous case.
n−q−1
Instead, we consider subcases as follows.
• 3A. If m = n − q = 1, let us first summarize our assumptions at this point.
1
3
n+1
+
+···+
k =
2 1
n
2
4
n
+
2
+
+···+
k hni =
2
3
n+1
p = q = n−1
s = q− 1 =
n−2
n+1
2
3
n
+1=
+
+···+
k−q =
n−1
2 1
n−1
1
3
n
+
1
+
+···+
k hni − q =
1
2
n
ALMOST MAXIMAL GROWTH OF THE HILBERT FUNCTION
9
Notice that the growth from k − q in degree n − 1 to k hni − q in degree n in this case
is two less than maximal growth. A key observation is that we have
k = k hni − q;
that is,
dim[S/J]n = dim[S/(J : ℓ)]n .
This means that the multiplication ×ℓ : [S/J]n → [S/J]n+1 is injective. Let J1 =
(J : ℓ). Let ℓ2 be a general linear form. Since J does not depend on ℓ, multiplication
by ℓ2 is also injective. Then from the commutative diagram
(3)
0 → [S/J1 ]n−1
↓ ×ℓ2
0 → [S/J1 ]n
0
↓
×ℓ
−→ [S/J]n → [S/(J, ℓ)]n → 0
↓ ×ℓ2
↓ ×ℓ2
×ℓ
−→ [S/J]n+1 → [S/(J, ℓ)]n+1 → 0
and by the Snake Lemma, we see that ×ℓ2 : [S/J1 ]n−1 → [S/J1 ]n is also injective.
We obtain
(4)
dim[S/J1 ]n−1 = dim[S/(J1 : ℓ2 )]n−1 .
Making a similar analysis but using J1 in place of J, we have, after a short calculation,
degree i 0
1
dim[S/J1 ]i 1
r
dim[S/(J1 : ℓ2 )]i−1
1
dim[S/(J1 , ℓ2 )]i 1 r − 1
...
...
...
...
n − 1
n+1
n−1
n+1
−
n−1
p1
p1
n+1
n
n
+ · · · + 32 +
n+1
n−1
1
1
n−1
for some p1 . Combining Green’s theorem (in degree n − 1) and Macaulay’s theorem
on the bottom line, we obtain
n − 1 ≤ p1 ≤ n + 1.
On the other hand, from the second line and again invoking Macaulay’s theorem
(remembering the shift), we have
n
n+1
− p1 ≥
n−2
n−1
n
(since n−2
is the smallest value in degree n − 2 that could grow to n+1
in degree
n−1
n − 1), or
p1 ≤ n.
Thus p1 is either equal to n − 1 or n.
10
L. CHIANTINI AND J. MIGLIORE
* If p1 = n then the second line of the last table gives
n
,
dim[S/(J1 : ℓ2 )]n−2 =
n−2
so this represents maximal growth for S/(J1 : ℓ2 ) from degree n − 2 to degree
n − 1. From the corresponding Hilbert polynomial we see that [(J1 : ℓ2 )]n−2
and [(J1 : ℓ2 )]n−1 are the degree n − 2 and n − 1 components of the saturated
ideal of a linear space Λ of dimension 2. We have
[IΛ ]n−1 = [(J1 : ℓ2 )]n−1 = [J1 ]n−1 = [(J : ℓ)]n−1 .
On the other hand, we have observed that dim[S/(J : ℓ)]n is two less than the
maximum possible. Thus (J : ℓ) picks up two minimal generators in degree
n, so the base locus of [J : ℓ]n is at least zero-dimensional. Since we have also
seen in this case that [J]n = [J : ℓ]n , we see that the base locus of [J]n is at
least zero dimensional, so we have our desired result.
* If p1 = n−1, then S/(J1 , ℓ2 ) has maximal growth from degree n−1 to degree
n. This means that ([J1 , ℓ2 )]n−1 and [(J1 , ℓ2 )]n have a base locus consisting of
a zero-dimensional scheme. Thus [J1 ]n−1 and [J1 ]n have a base locus consisting of a 1-dimensional scheme, since ℓ2 is general. In particular, [(J : ℓ)]n
has a 1-dimensional base locus. Since [J]n ⊂ [(J : ℓ)]n , we have the desired
result.
• 3B. If m = n − q > 1 then as before we summarize our current assumptions.
1
n−q
n−q+2
n+1
+···+
+
+···+
k =
1
n − q + 1 n − q
n
2
n
−
q
+
1
n
−
q
+
3
n
+
2
+···+
+
+···+
k hni =
2
n−q+1
n−q+2
n+1
p = q= s +1
n−q
n
+···+
k−q =
n − q − 1
n − 1
1
n
−
q
n
−
q
+
2
n
+
1
+···+
+
+···+
k hni − q =
1
n−q
n−q+1
n
Notice that again, we have
k = k hni − q, i.e. dim[S/J]n = dim[S/(J : ℓ)]n ,
×ℓ
so [S/J]n −→ [S/J]n+1 is injective. We again let J1 = (J : ℓ) and ℓ2 a general
×ℓ2
linear form. Again using the Snake Lemma, we obtain that [S/J1 ]n−1 −→
[S/J1 ]n
is injective, so
dim[S/J1 ]n−1 = dim[S/(J1 : ℓ2 )]n−1
ALMOST MAXIMAL GROWTH OF THE HILBERT FUNCTION
11
Let p1 = dim[S/(J1 , ℓ2 )]n−1 and p2 = dim[S/(J1 , ℓ2 )]n . We compute
p2 = dim[S/J
1 ]n − dim S/(Ji1 : ℓh2 )]n−1
h
n+1
n−q+2
n−q
=
+
+
·
·
·
+
+···+
n−q+1
hn
i n−q
n
n−q
− n−1
+ · · · + n−q−1
= q.
i
1
1
We thus have the table
n−1
degree i . . .
dim[S/J1 ]i . . .
dim[S/(J1 : ℓ2 )]i−1 . . .
dim[S/(J1 , ℓ2 )]i . . .
n
n−1
n
n−1
+···+
+···+
p1
n−q
n−q−1
n−q
n−q−1
− p1
h
n
i
n−q+2
+
·
·
·
+
n
n−q+1
h
i
n−q
+ n−q + · · · + 11
n
n−q
+
·
·
·
+
n−1
n−q−1
q
n+1
From Green’s theorem, the first and the third lines of the table show that p1 ≤ q + 1.
From one of the assumptions in Case 3 we have q < n, so from Macaulay’s theorem
applied to S/(J1 , ℓ2 ) we obtain p1 ≥ q. Combining, we have
q ≤ p1 ≤ q + 1.
We have the following possibilities.
* If p1 = q then the bottom line of the above table represents maximal growth
from degree n−1 to degree n. This means that the base loci of [(J1 , ℓ2 )]n−1 and
[(J1 , ℓ2 )]n have dimension 0. Hence [J1 ]n−1 and [J1 ]n have a 1-dimensional
base locus. In particular, [(J : ℓ)]n has a 1-dimensional base locus. Since
[J]n ⊂ [(J : ℓ)]n , we have the desired result.
* Assume that p1 = q + 1 and n − q − 1 > 1. We note that since p1 = q + 1,
and since n − q − 1 > 1, we have
n−q
n−1
+ n − q − 1.
+···+
dim[S/(J1 : ℓ2 )]n−2 =
n−q−1
n−2
n−q−1
n−q
n−1
,
+
+···+
=
n−q−2
n−q−1
n−2
so we have maximal growth for S/(J1 : ℓ2 ) from degree n − 2 to degree n − 1,
giving a 1-dimensional base locus. The argument is then the same as above.
* Assume finally that p1 = q + 1 and n − q − 1 = 1. After a short computation,
one sees that
n
.
dim[S/(J1 : ℓ2 )]n−2 =
n−2
So now we have the table
12
L. CHIANTINI AND J. MIGLIORE
degree i
dim[S/J1 ]i
dim[S/(J1 : ℓ2 )]i−1
dim[S/(J1 , ℓ2 )]i
...
...
...
...
n − 1
+ · · · +
n
n
n−1
n−2
2
1
q+1
n
n+1
+
· · · + 43 +
n
n
+ · · · + 21
n−1
2
q
The exact sequence (2) splits into two short exact sequences as indicated (but
use I = J1 and L = ℓ2 ). The fact that
dim[S/J1 ]n−1 = dim[S/(J1 : ℓ2 )]n−1
means that [ J1J:ℓ1 2 ]n−1 = 0, so from the long exact sequence (2) we see that
×ℓ
2
×ℓ2 : [S/J1 ]n−1 −→
[S/J1 ]n is an injection. Thus for a general linear form ℓ3 ,
×ℓ3
[S/J1 ]n−1 −→ [S/J1 ]n is also injective. Setting J2 = (J1 : ℓ2 ), we again ob×ℓ3
tain by the Snake Lemma (as we did in (3)) an injection [S/(J1 : ℓ2 )]n−2 −→
[S/(J1 : ℓ2 )]n−1 . We now have the following table.
degree i
dim[S/J2 ]i
dim[S/(J2 : ℓ3 )]i−1
dim[S/(J2 , ℓ3 )]i
...
...
...
...
n − 2
n
n−2
n
n−2
− p3
p3
n − 1
+ · · · +
n
n
n−1
n−2
2
1
n−1
where the bottom right dimension is the result of a simple computation. Now
combining Green’s theorem and Macaulay’s theorem, we obtain n−1 ≤ p3 ≤
n − 1, i.e. p3 = n − 1. But then
n−1
n
.
− (n − 1) =
dim[S/(J2 : ℓ3 )]n−3 =
n−3
n−2
This means that S/(J2 : ℓ3 ) has maximal growth from degree n − 3 to n − 2,
so its base locus is 2-dimensional. In fact, by Gotzmann’s theorem, we can
compute the Hilbert polynomial and obtain that the base locus is a plane Λ,
and [J2 : ℓ3 ]n−2 = [IΛ ]n−2 . From the above table we then obtain that [J2 ]n−2 =
[IΛ ]n−2 . Since the growth of the Hilbert function of J2 from degree n − 2 to
degree n − 1 is one less than maximal, the base locus of [J2 ]n−1 is (precisely)
1-dimensional. Indeed, it must be the degree n−1 component of a plane curve
of degree n − 1. From the previous table we have (recalling J2 = J1 : ℓ2 ) that
[J2 ]n−1 = [J1 ]n−1 . The growth of S/J1 from degree n − 1 to degree n is again
one less than maximal, so [J1 ]n has at least a 0-dimensional base locus. In the
same way we obtain that [J1 ]n = [J]n , so we obtain the desired result.
This completes the proof.
Remark 3.2. When
the base locus has dimension 1, its degree is at most the number d of terms
a+1
of the form a in the n-binomial expansion of hA (n).
Examples exist where this bound is achieved. This can immediately be seen by choosing an
ideal J1 for which S/J1 has Hilbert function hA (n) in degree n and has maximal growth to
ALMOST MAXIMAL GROWTH OF THE HILBERT FUNCTION
13
degree n + 1. Then form the ideal J by adding one minimal generator to J1 (say a general form)
in degree n + 1.
The next example shows that all degrees ≤ d can occur at least for some values of hA (n).
Example 3.3. In this example we show that all the possibilities left open by Theorem 3.1 can
occur. The computation were performed with the aid of the computer Algebra Package CoCoA
([13]).
Let S = k[x, y, z], n = 6 and
1
2
3
5
6
7
.
+
+
+
+
+
hA (6) = 21 =
1
2
3
4
5
6
Then maximal growth would correspond to hA (7) = 24, so the assumption of Theorem 3.1 is
that hA (7) = 23. Notice that a value of 21 in degree 6 allows a base locus of at most a curve of
degree 3. Let
I = hx6 , x5 y, x5 z, x4 y 2 , x4 yz, x4 z 2 i
J1 = hI, x3 y 3 , x3 y 2 z 2 i
J2 = hI, x2 y 4 i
J3 = hI, xy 5i
J4 = hF, Gi : IP
where P is a general point in P2 and hF, Gi is a general complete intersection of type (4, 6) in
IP . Notice that I and J1 are lex-segment ideals, and J1 has a minimal generator in degree 7
while none of J2 , J3 or J4 do. (J4 has generators in degrees 4, 6 and 8.) One can check that
S/J1 , S/J2 , S/J3 and S/J4 all satisfy the assumption hA (6) = 21 and hA (7) = 23. The Hilbert
polynomials are as follows:
HP (S/J1)
HP (S/J2)
HP (S/J3)
HP (S/J4)
=
=
=
=
3t + 2
2t + 9
t + 16
23 for
for t ≥ 7;
for t ≥ 6;
for t ≥ 7;
t ≥ 7;
Hence the base locus of [J1 ]6 and [J1 ]7 is a curve of degree 3 (defined by x3 ), the base locus of
[J2 ]6 and [J2 ]7 is a curve of degree 2 (defined by x2 ), the base locus of [J3 ]6 and [J3 ]7 is a curve
of degree 1 (defined by x) and the base locus of [J4 ]6 and [J4 ]7 is a zero-dimensional scheme
of degree 24 defined by the complete intersection hF, Gi (since the last generator reduces the
degree by 1, thanks to a standard liaison computation).
When the value of the Hilbert function hA (n) is bigger than or equal to n+2
, we have a
2
similar behavior.
Theorem
3.4. Let J ⊂ S be a homogeneous ideal, and let A = S/J. Assume that hA (n) ≥
n+2
. Assume further that hA (n + 1) = hA (n)hni − 1, i.e. that the growth is one less than
2
maximal. Then the base locus of [J]n in Pr−1 is a scheme of dimension either MG(hA (n), n)−1
or MG(hA (n), n).
Proof. We will use induction on r, the number of variables in S. If r = 3 then [A]≤n = [S]≤n
by the assumption that hA (n) ≥ n+2
(in fact equality must hold), so the base locus is all of
2
14
L. CHIANTINI AND J. MIGLIORE
P2 . Notice that in this case [J]n+1 has a base locus consisting of a curve of degree n + 1, by the
assumption of almost maximal growth.
Let us write the n-binomial expansion of hA (n):
ki
kn−1
kn
+···+
+
(5)
hA (n) =
i
n−1
n
where kn > kn−1 > · · · > ki ≥ i ≥ 1. The condition that hA (n) ≥ n+2
means that
n
kn ≥ n + 2. Hence hA (n)hni ≥ n + 1.
We have
ki + 1
kn−1 + 1
kn + 1
hni
,
+···+
+
hA (n) =
i+1
n
n+1
ki − 1
kn−1 − 1
kn − 1
+···+
+
(6)
hA (n)hni =
i
n−1
n
and
(7)
hni
[hA (n)hni ]
=
ki
kn−1
kn
+···+
+
i+1
n
n+1
where some of the binomials may be zero in the last two equalities. Indeed, all the terms are
non-zero if and only if ki > i.
Let h be a general linear form in S. Mimicking a proof in [6], page 172, we have the following
sequence of inequalities and equalities:
(8)
[hA (n)hni ]hni ≥
≥
≥
=
=
hA/hA (n)hni
hA/hA (n + 1)
hA (n + 1) − hA (n)
hA (n)hni − 1 − hA (n)
[hA (n)hni ]hni − 1.
(Green’s theorem)
(Macaulay’s theorem)
(hypothesis)
The third line comes from the exact sequence
(9)
×h
A(−1) −→ A → A/hA → 0
and the last line is a straightforward computation from the definitions. This means that one of
the inequalities must be strict (differing by 1) and the others must be equalities.
Suppose that the first line of (8) is a strict inequality and the rest are equalities, i.e. we have
(10)
[hA (n)hni ]hni − 1 =
=
=
=
hA/hA (n)hni
hA/hA (n + 1)
hA (n + 1) − hA (n)
hA (n)hni − 1 − hA (n)
There may or may not exist an integer m such that mhni = [hA (n)hni ]hni − 1. If it does not,
then the first line of
(8) must be an equality. If it does exist then the n-binomial expansion of m
kn −1
begins with n , i.e. MG(hA/hA (n), n) = kn − 1 − n ≥ 1. The second equality in (10) gives
ALMOST MAXIMAL GROWTH OF THE HILBERT FUNCTION
that A/hA has maximal growth from degree n to degree n + 1, so in fact
h
15
(J,h)
(h)
i
n
has base locus
of dimension kn − n − 1 ≥ 1, and [J]n has base locus of dimension kn − n = MG(hA (n), n).
Now suppose that the second line of (8) is a strict inequality and the rest are equalities, i.e.
we have
[hA (n)hni ]hni − 1 =
=
=
=
(11)
hA/hA (n)hni − 1
hA/hA (n + 1)
hA (n + 1) − hA (n)
hA (n)hni − hA (n) − 1.
In particular, we have (a) hA (n)hni = hA/hA (n) and (b) the growth of hA/hA is one less than
maximal from degree n to degree n + 1. From (a) and the fact that kn ≥ n + 2, by (6) we obtain
MG(hA/hA
h (n),
i n) = (kn − 1) − n ≥ 1. If (kn − 1) − n = 1, Theorem 3.1 shows that the base
locus of
(J,h)
(h)
n
is a scheme of dimension either 0 or 1, so [J]n defines a scheme of dimension
either 1 or 2, hi.e. either
i MG(hA (n), n) − 1 or MG(hA (n), n). If instead (kn − 1) − n ≥ 2 then
(J,h)
defines a scheme of dimension either (kn − 1) − n or (kn − 1) − n − 1.
by induction (h)
n
Then [J]n defines a scheme of dimension either kn − n or kn − n − 1, i.e. either MG(hA (n), n)
or MG(hA (n), n) − 1.
Finally, suppose that the third line of (8) is a strict inequality and the rest are equalities, i.e.
we have
(12)
[hA (n)hni ]hni − 1 =
=
=
=
hA/hA (n)hni − 1
hA/hA (n + 1) − 1
hA (n + 1) − hA (n)
hA (n)hni − hA (n) − 1.
We again have hA (n)hni = hA/hA (n) but this time the second equation of (12) shows that the
growth of hA/hA is maximal from degree n to degree n+1. We hagainihave MG(hA/hA (n), n) =
kn −n−1. This time maximal growth implies the base locus of (J,h)
has dimension kn −n−1,
(h)
so the base locus of [J]n has dimension kn − n = MG(hA (n), n).
Corollary 3.5. Let Z ⊂ Pr be a finite set of points, let L be a general linear form, and let
Z ,L)
⊂ S = R/(L), where R = k[x0 , . . . , xr ]. Let A = S/J. For some integer n, assume
J = (I(L)
that A satisfies the condition of one of the two previous theorems. Then the base locus of [IZ ]n
is a scheme of dimension either MG(hA (n), n) or MG(hA (n), n) + 1.
Remark 3.6. In the setting of Corollary 3.5, it is clear that we expect many of the points to lie
on the base locus. However, we do not have a good estimate for the number of such points.
Nevertheless, we carefully analyze this question in the next section for the one remaining type
of almost maximal growth (see Proposition 4.13). Furthermore, in section 5, even when there
is no higher dimensional base locus we show the surprising result that there is a “hidden” linear
variety containing many of the points.
16
L. CHIANTINI AND J. MIGLIORE
4. G ROWTH
TYPE
(. . . , k, k − 1):
WHEN THERE IS A NON - EMPTY BASE LOCUS
The first case not covered in the last section is the following. Let 2 ≤ k ≤ n. Then an ideal
J ⊂ S = k[x1 , . . . , xr ] with hS/J (n) = k must satisfy hS/J (n + 1) ≤ k by Macaulay’s theorem.
Thus “almost maximal growth” means hS/J (n + 1) = k − 1 ≥ 1. In this case we will see that
[J]n may or may not have a base locus, and that the setting of arbitrary homogeneous ideals
becomes almost trivial, but the setting of artinian reductions of ideals of sets of points is very
interesting.
Thus we now let Z ⊂ Pr be a finite set of reduced points. Let L be a general linear form in
Z ,L
R = k[x0 , . . . , xr ], and let J := IhLi)
⊂ S := R/hLi ∼
= k[x1 , . . . , xr ]. Throughout this section
we make the following assumption.
(13)
Z is a reduced finite set of points whose h-vector has values k and k − 1 in
degrees n and n + 1 respectively, where n ≥ k ≥ 2, and is zero thereafter.
The assumption that the h-vector is zero starting in degree n + 2 is without loss of generality.
Indeed, we can always remove points one at a time so that the Hilbert function is continually
truncated at the top ([26], lemma 5 and proposition 7.iii). This means that the h-vector is
unchanged in all except the last non-zero value, where it drops by one. We repeat this until we
have the desired 0 in degree n + 2.
We recall the relevant result from [5] for maximal growth. (We have suitably revised the
statement to fit in with our situation.)
Theorem 4.1 ([5] Theorem 3.6). Let Z ⊂ Pr be a reduced set of points. Assume that for some
integer n, ∆hZ (n) = ∆hZ (n + 1) = k, where n ≥ k ≥ 1. Then
(a) h[IZ ]≤n i is the saturated ideal of a curve, V , of degree k (not necessarily unmixed).
Furthermore V is reduced.
Let C be the unmixed one-dimensional part of V . Let Z1 be the subset of Z lying on C and let
Z2 be the “residual” set.
(b) h[IZ1 ]≤n i = IC .
(c)
∆hC (t), for t ≤ n + 1;
∆hZ1 (t) =
∆hZ (t), for t ≥ n.
In particular, ∆hZ1 (t) = k for all k ≤ t ≤ n + 1.
The starting point of the proof of the above result was that the maximal growth condition
forces the base locus of the degree n (and degree n+1) component of J to be a zero-dimensional
scheme of degree k (eventually proven to be reduced), which then lifts to a curve of degree k
containing many of the points of Z.
Unfortunately, the main results of [5] used Gotzmann’s theorem, which holds only in the
case of maximal growth. As a result, there are more possibilities in the setting of (13). Thus,
our analysis will be based on a consideration of the possibilities for the base locus of [J]n and
[J]n+1 .
ALMOST MAXIMAL GROWTH OF THE HILBERT FUNCTION
17
If Z ⊂ P2 then the analysis was already known to Davis [14]. Thus we will assume
(14)
The h-vector of R/IZ has the form (1, h1 , . . . , k, k − 1, 0), where the k occurs in
degree n ≥ k ≥ 2, and 3 ≤ h1 ≤ r.
Lemma 4.2. Under the assumptions of (14), let W be the base locus of [J]n . Then either
dim W = 0 with deg W ≤ k or W = ∅. In the latter case, the linear system |[J]n | is basepoint
free.
Proof. If W were one-dimensional or more (as a subscheme of the Pr−1 defined by L) then the
elements of [J]n all lie in [IW ]n . This means that the Hilbert function of S/J is greater than or
equal to that of S/IW in degree n, which is not possible since hS/IW (n) ≥ n + 1. This also
shows deg W ≤ k. Hence if the base locus of [J]n is not empty, it is zero-dimensional. The last
assertion is obvious.
We will see that both possibilities allowed by Lemma 4.2 actually can occur. In this section
we deal with the first of the two, i.e. the case dim W = 0. We first show that the base locus can
have any degree ≤ k.
Proposition 4.3. Fix any integer d with 1 ≤ d ≤ k ≤ n. Then there exists a set of points
Z ⊂ Pr satisfying (14), for which [J]n has base locus consisting of precisely d points. That is,
[IZ ]n has a base locus whose 1-dimensional component is a reduced curve of degree d.
Proof. We proceed in three steps. First, fix a plane Γ ⊂ Pr and choose a smooth plane curve C
of degree d on Γ. Choose a general set, Z1 , of
d
+ (n − d + 2)d + (d − 1)
1 + · · · + d + d + · · · + d + (d − 1) =
2
points on C. The h-vector of Z1 is (1, . . . , d, d, . . . , d, d − 1) where the last d occurs in degree
n. (The first d entries increase monotonically by 1, but if d = 1 then there is no increase.)
For the second step, if d = k then we do nothing and proceed to the third step. If, instead, we
have d < k, then we want to add points on Γ so that ∆hR/IZ (n) = k, ∆hR/IZ (n + 1) = k − 1,
and the base locus in degree n and n + 1 remains a curve of degree d. We first add one more
general point to Z1 on C, making the value of the h-vector in degree n + 1 now d. Let Z2′ be a
general complete intersection in Γ of type
k + n − 2d + 1
k + n − 2d + 1
,
.
2
2
Observe that the h-vector of Z2′ in degree n − d is k − d, and the value in degree n − d + 1
is k − d − 1. Since Z2′ is in uniform position, we can remove any k−d−1
points from Z2′ to
2
obtain a set of points Z2 whose h-vector agrees with that of Z2′ up to degree n − d + 1 and
is zero thereafter. The base locus of the ideal of Z2 in degree n − d is zero-dimensional. Let
Z = Z1 ∪ Z2 . We claim that for t ≤ n + 1 we have
∆hR/IZ (t) = ∆hR/IZ2 (t − d) + hR/IC (t).
It is enough to think of these algebras as quotients of k[x, y, z]. Then this follows from the
work of Davis, but can also be seen directly with a simple computation using Bezout’s theorem,
18
L. CHIANTINI AND J. MIGLIORE
since for t ≤ n + 1 any plane curve of degree t containing Z1 must contain C as a component,
so dim[IZ ]t = dim[IZ2 ]t−d . With this formula we obtain that the h-vector of Z has value k in
degree n and k − 1 in degree n + 1 as desired. Also, the base locus of IZ in degrees n and n + 1
has only C as 1-dimensional base locus.
For the third step, observe that Z as constructed so far has h1 = 2. We now merely add up to
r − 2 general points to obtain 3 ≤ h1 ≤ r, and we are finished.
We would like to point out, though, that in Proposition 4.3 there are at least
k
+ k(n − k + 3) − 1
1 + 2 + 3 + · · · + k + k + · · · + k + (k − 1) =
2
points on a plane, which also contains the 1-dimensional component. It is possible to construct
examples where the curve (and hence many of the points of Z) does not lie on a plane. We give
an example.
Example 4.4. Consider the case of finite sets of points in P3 and suppose n = 8 and k = 7. We
show that for any 1 ≤ d ≤ 7 there is a set of points Z ⊂ P3 such that
• ∆hR/IZ (8) = 7 and ∆hR/IZ (9) = 6;
• the base locus of [IZ ]8 is a non-degenerate, smooth curve, C, of degree d.
Furthermore, we show that there is a set of points whose h-vector ends (. . . , 7, 6) and whose
base locus for the component in degree 8 is zero-dimensional (in fact it consists only of points
of Z). This is in stark contrast to the h-vector (. . . , 7, 7), where the base locus must be a curve
of degree 7.
For fixed d, 1 ≤ d ≤ 7, let C be a smooth, irreducible arithmetically Cohen-Macaulay curve
of degree d whose h-vector is as short as possible, and let Z1 be a general set of points on C
with h-vector as follows:
d
1
2
3
4
h-vector of Z1
(1, 1, 1, 1, 1, 1, 1, 1, 1)
(1, 2, 2, 2, 2, 2, 2, 2, 2)
(1, 3, 3, 3, 3, 3, 3, 3, 3)
(1, 3, 4, 4, 4, 4, 4, 4, 4)
d
5
6
7
h-vector of Z1
(1, 3, 5, 5, 5, 5, 5, 5, 5)
(1, 3, 6, 6, 6, 6, 6, 6, 6, 6)
(1, 3, 6, 7, 7, 7, 7, 7, 7, 7)
(Note that the length of the h-vector in the last two cases is one more than the lengths in the
previous cases.)
In the case d = 7 we simply remove any point P from Z1 and we are finished: Z = Z1 −{P }
and the h-vector is (1, 3, 6, 7, 7, 7, 7, 7, 7, 6).
In the case d = 6 we add 81 general points of P3 to Z1 , arriving at the h-vector
(1, 3, 6, 10, 15, 21, 28, 36, 7, 6).
Now assume 1 ≤ d ≤ 5. Let F1 , F2 be a regular sequence in IC of type (7, 9 − d), hence
also a regular sequence of type (7, 9 − d) in IZ1 . Let L be a general linear form. The ideal
hL, F1 , F2 i is the saturated ideal of a complete intersection, Z2 , of type (7, 9 − d) in the plane
defined by L, hence of 63 − 7d reduced points. Thanks to [[22] Lemma 4.8 and Remark 4.10]
ALMOST MAXIMAL GROWTH OF THE HILBERT FUNCTION
19
or [[25] Proposition 5.4.5], the ideal L · IZ1 + hF1 , F2 i is the saturated ideal of the union, Z, of
Z1 and Z2 . Its h-vector is
deg 0 1 2 3 4 5 6 7
8
9
Z1
1 ∗ ∗ d d d d
d
d
Z2 1 2 3 4 ∗ ∗ ∗ ∗ 7 − d 6 − d . . .
1 3 ∗ ∗ ∗ ∗ ∗ ∗
7
6
...
By [17], a subset of Z1 ∪ Z2 can be found by removing a suitable set of points of Z2 , so that the
h-vector of the resulting set of points is (1, 3, . . . , 7, 6) (with the 7 in degree 8). We thus have
k = 7, n = 8 so the numerical assumptions are satisfied.
Now, in degrees 7 and 8, Z2 does not have a one-dimensional component in the base locus,
hence one can check from the form of the ideal given above (and confirm on CoCoA) that the
one-dimensional base locus of Z in degree 8 and degree 9 has degree d.
To arrange for the base locus to be zero-dimensional, we replace the set of points on a curve
with any set of points with regularity less than 7 (say), and let hF1 , F2 i be a complete intersection
of type (7, 9) and apply the same construction.
We now ask what consequences are imposed by the assumption that the base locus of [J]n
(or [J]n+1 ) is zero-dimensional of degree d. We first recall the following result.
Proposition 4.5 ([5] Corollary 3.7). Let Z ⊂ Pr be a reduced finite set of points. Assume that
for some n ≥ d, ∆hR/IZ (n) = d and that the saturated ideal h[IZ ]≤n isat defines a curve of
degree d. Then all the conclusions of Theorem 4.1 continue to hold.
Corollary 4.6. Assume that [J]n+1 has a base locus consisting of a zero-dimensional scheme
in Pr−1 of degree k − 1, together with the other assumptions highlighted so far in this section.
Then all the conclusions of Theorem 4.1 continue to hold. In particular, [IZ ]n+1 has a reduced
1-dimensional component of degree k − 1.
Proposition 4.7. Suppose that IZ has a minimal generator of degree n + 1, together with our
assumption (14). Then all the conclusions of Theorem 4.1 continue to hold. In particular, [IZ ]n
has a reduced 1-dimensional component of degree k.
Proof. If IZ has such minimal generator then so does the artinian reduction J. Removing this
generator from J gives a new artinian ideal J ′ that agrees with J through degree n, and whose
Hilbert function has the value k degree n. This means that [J ′ ]n , and hence also [J]n , has a
base locus consisting of a zero-dimensional scheme of degree k. It follows that [IZ ]n has a base
locus with a one-dimensional component of degree k. Thus Proposition 4.5 applies and we are
done.
Corollary 4.8. In the setting of Proposition 4.7, at least
k
+ k(n − k + 3) − 1
2
points of Z lie on a curve of degree k. In the setting of Corollary 4.6, at least
k−1
+ (k − 1)(n − k + 4)
2
20
L. CHIANTINI AND J. MIGLIORE
points lie on a curve of degree k − 1.
Proof. We will use several times in this paper the observation that if Z1 is a set of points lying
on a curve C of some given degree d then the sum of the entries of the h-vector (i.e. the number
of points) of Z1 is minimized when C is a plane curve, since then Theorem 4.1 (once we know
that it applies) gives that the h-vector is of the form (1, 2, 3, . . . , (d − 1), d, d, . . . , d, ǫ (where
ǫ ≤ d) so this allows the h-vector to grow as slowly as possible. By Theorem 4.1, since the
degree of the curve is k and the curve is reduced, it follows that
|Z1 | ≥ 1 + 2 + · · · + (k − 1) + k + k + · · · + k + (k − 1)
where the two (k − 1)’s occur in degrees k − 2 and n + 1. The result is then an easy calculation.
The same idea gives the second result.
This leads to the next reduction.
(15)
From now on we assume that J, and hence IZ , has no minimal generator
of degree n + 1.
We believe that something like Proposition 4.7 should continue to hold.
Conjecture 4.9. Assume that [J]n has a base locus consisting of a zero-dimensional scheme
in Pr−1 of degree d, together with the other assumptions highlighted so far in this section. As
noted above, we have d ≤ k. Then
(a) h[IZ ]≤n+1 i is the saturated ideal of a reduced curve, V , of degree d (not necessarily
unmixed).
Let C be the unmixed one-dimensional part of V . Let Z1 be the subset of Z lying on C and let
Z2 be the “residual” set.
(b) h[IZ1 ]≤n+1 i = IC .
(c) ∆hZ1 (t) = ∆hC (t), for t ≤ n + 1. In particular, ∆hZ1 (t) = d for all d ≤ t ≤ n + 1.
Question 4.10. We will see in section 5 that in the situation where [J]n is basepoint free, many
points of Z have to lie on a plane. If the above conjecture is true, does it also follow that many
points of Z2 have to lie on a plane? We do not have much information about the Hilbert function
of Z2 , even conjecturally.
Lemma 4.11. Every
H 1 IZ (d − 1) = 0.
0-dimensional subscheme Z
of length d in
Pr
satisfies
Proof. Consider the h-vector, h, of Z. Since the h-vector is the Hilbert function of R/(IZ , L)
for a general linear form L, it follows that if h(i) = 0 for some i then h(i + j) = 0 for all
j > 0. But then h is, at worst, of the form 1 1 1 ...1 0 ... so that h(d) = 0, and the conclusion
follows.
By regularity, it follows that every 0-dimensional subscheme Z of length d has a homogeneous ideal generated in degree ≤ d.
ALMOST MAXIMAL GROWTH OF THE HILBERT FUNCTION
21
Lemma 4.12. Given an unmixed reduced curve C of degree d (even reducible) in Pr , then
H 1 IC (d − 1) = 0.
Proof. Well known. See for instance Remark 2.8.
Here is the main new fact.
Proposition 4.13. Assume that the h-vector of R/IZ has the form (1, h1 , . . . , k, k − 1, 0), where
the k occurs in degree n ≥ k ≥ 2. Assume that J, and hence IZ , has no minimal generator
of degree n + 1. Assume that [J]n has a base locus consisting of a zero-dimensional scheme
in Pr−1 of degree d′ . Call C ′ the union of all positive dimensional parts of the saturation of
h[IZ ]≤n i, so deg C ′ = d′ . Let C = (C ′ )red and assume that deg C = d.
Then C has dimension 1 and degree d ≤ k − 1. Furthermore,
d
+ 5d points of Z;
• if k − 1 − d ≤ d + 1 then C contains at least 2 ·
2
• If d + 1 ≤ k − 1 − d then C contains at least d2 + d(n − k + 4) points of Z
Proof. The fact that C ′ is a (possibly non reduced) curve of degree d′ contained in the base
locus of [IZ ]n follows at once from the fact that L is a general hyperplane.
Call Z1 the subset of Z lying on C ′ . Notice that indeed Z1 lies on C. Define Z − C = Z2 ,
where Z − C is defined by the homogeneous ideal IZ : IC . Since Z is reduced, we have
Z = Z1 ∪ Z2 .
Since IZ has no minimal generators in degree n + 1, we have for any i ≤ n + 1 that
[IZ ]i ⊆ [IC∪Z2 ]i = [IC∪Z ]i ⊆ [IZ ]i ,
so all of these vector spaces are equal. In particular, IZ coincides up to degree n + 1 with
IC∪Z = IC∪Z2 .
Let L1 ∼
= Pr−1 be a general hyperplane and call W the intersection of L1 and C. Note first
that the h-vector of Z ′ = Z ∪ W coincides with the h-vector of Z for i = 0, . . . , n + 1. Indeed
Z ′ sits in the base locus of h[IZ ]≤n+1 i. In particular the h-vector of Z ′ is k − 1 in degree n + 1.
We claim that the h-vector of Z ′ is 0 in degree n + 3, i.e. that Z ′ is separated by forms of
degree n + 2. To see this, note that the sequence
×L
0 → H 0 IZ (n + 1) −→1 H 0 IZ (n + 2) → H 0 OL1 (n + 2) → 0
is exact (identifying L1 with Pr−1 ), because the h-vector of Z is 0 in degree n + 2, hence
H 1 IZ (n + 1) = 0. With this in mind, we can prove that if Y is any subscheme of Z ′ whose
length is the length of Z ′ minus 1, then there exists a form f of degree n + 2 passing through Y
and not through Z ′ .
◮ If Z ′ − Y is supported in Z, then one takes a form f ′ of degree n + 1 passing through Z ∩ Y
(it exists because Z is separated in degree n + 1) and takes f = f ′ L1 .
◮ If Z ′ − Y is supported in L1 , then one takes one form of degree n + 2 in L1 = Pr−1 which
passes through Y ∩ W and not through W (it exists, by Lemma 4.11) and lifts it to a form f
containing Z, via the surjection
H 0 IZ (n + 2) → H 0 OL1 (n + 2) → 0.
22
L. CHIANTINI AND J. MIGLIORE
It follows that the h-vector of Z ′ is k, k − 1, d, 0 in degrees n, n + 1, n + 2, n + 3 respectively.
This proves in particular that d ≤ k − 1.
If d = k − 1, we apply Theorem 4.1 to Z ′ and prove that C has degree d = k − 1. Moreover
C contains a subscheme of Z ′ of length at least (k − 1)(n + 2 − (k − 1)). Since the length of
Z ′ − Z is d = k − 1, the conclusion follows.
Assume then that d < k − 1. By construction the base locus of [IZ ′ ]n+1 contains C. We do
not know if the base locus of [IZ ′ ]n+2 contains C, so we have to consider both possibilities.
Case 1: The base locus of [IZ ′ ]n+2 does not contain C. Call q1 the dimension of the vector
space
[IZ ′ ]n+2 /([IZ ′ ]n+2 ) ∩ [IC ]n+2 ).
We would like to add to Z ′ a set of q1 “general” points of C. From a geometric point of view,
the significance of the number q1 is the following.
This number q1 is the number of points on C that we need to add to
Z ∪ (L1 ∩ C) (where L1 is a general plane) in order to get a set Z1′ such that all
the elements of H 0 IZ1′ (n + 2) contain C.
We note here that the numbers qi introduced below have an analogous geometric meaning: qi is
′
the number of points on C that we need to add to Zi−1
∪ (Li ∩ C) in order to get a set Zi′ such
that all elements of H 0 IZi′ (n + 1 + i) contain C.
The first problem is what this means if C is not irreducible. This will be resolved by the
following procedure. Let {F1 , . . . , Fm } be a basis for ([IZ ′ ]n+2 ∩ [IC ]n+2 ). We have that
dim[IZ ′ ]n+2 = m + q1 , so we can complete this to a basis for [IZ ′ ]n+2 as follows. It is possible that [IZ ′ ]n+2 has a one-dimensional base locus, C1 , whose support is a union of some (but
not all) components of C. Choose a component of C that is not a component of C1 , and let P1
be a general point of this component and let A1 = Z ′ ∪{P1 }. Clearly dim[IA1 ]n+2 = m+q1 −1.
Applying the same procedure but with A1 instead of Z ′ , we obtain a point P2 on some component of C and form A2 = Z ′ ∪ {P1 , P2 } with dim[IA2 ]n+2 = m + q1 − 2. Continuing
in this way, we obtain a set of points {P1 , . . . , Pq1 } so that if Z1′ = Z ′ ∪ {P1 , . . . , Pq1 } then
dim[IZ1′ ]n+2 = dim[IZ ′ ]n+2 − q1 . Since dim[IZ1′ ]n+1 = dim[IZ ′ ]n+1 (the new points all lie in the
base locus), we see that the length of the resulting scheme Z1′ is length(Z) + d + q1, its h-vector
is equal to the h-vector of Z up to degree n + 1, and the h-vector has value d + q1 in degree
n + 2.
It follows that Z1′ is separated in degree n + 2 and it has length d + q1 +length(Z). Moreover
d + q1 ≤ k − 1. If d + q1 = k − 1, the h-vector of Z1′ ends with k − 1, k − 1 in degrees
n + 1, n + 2. As above, we can apply Theorem 4.1 and find that C has degree d = k − 1, and
contains a subset of Z1′ of length at least (k − 1)(n + 2 − (k − 1)). Since the length of Z1′ − Z
is d + q1 = k − 1, the conclusion follows. As above, we can apply Theorem 4.1. We obtain
that IZ1′ has no minimal generator in degree n + 2, and that the base locus of the components
in degrees n + 1 and n + 2 is a curve of degree k − 1. Since we also assumed that there is no
minimal generator in degree n + 1, we get that this curve is C, and that hence d = k − 1 and
q1 = 0 (we get a little bit more: we get that the curve is C ′ , i.e. C ′ is reduced). The estimate for
the number of points of Z that lie on C will be carried out at the end of the proof.
ALMOST MAXIMAL GROWTH OF THE HILBERT FUNCTION
23
Assume d + q1 < k − 1. Fix another general hyperplane L2 and repeat the procedure, i.e. add
to Z1′ the set L2 ∩ C and, if the dimension q2 of
[IZ1′ ∪(L2 ∩C) ]n+3 /([IZ1′ ∪(L2 ∩C) ]n+3 ∩ [IC ]n+3 )
is non-zero, also add a set of q2 “general” points on C as above. Call Z2′ the resulting set. As
above, we have a surjection
H 0 IZ1′ (n + 3) → H 0 OL2 (n + 3) → 0
which guarantees that the h-vector of Z2′ ends with k − 1, d + q1 , d + q2 , 0 in degrees n + 1, n +
2, n + 3, n + 4 respectively.
We continue in the same way obtaining, after i steps, a 0-dimensional scheme Zi′ whose hvector ends with k − 1, d + q1 , . . . , d + qi , 0 in degrees n + 1, n + 2, . . . , n + i + 1, n + i + 2
respectively.
We now claim that the qi are strictly decreasing until they reach the value 0. If this were not
the case, then we would have two consecutive degrees in which the first difference of the Hilbert
function of our set of points Zi′ would have the same value d + qi . This forces the base locus of
[IZi′ ]n+1+i to contain a curve of degree d + qi > d. But the artinian reduction of R/IZi′ agrees
with that of R/IZ in degrees n and n + 1, and this base locus has only degree d. The base locus
cannot grow as the degree increases, so this is impossible.
Thus the process stops after m steps, where m is the first value of i such that qi = 0, since
then qi = 0 for all i > m as well, and so we have maximal growth.
Claim: m ≤ min{k − 1 − d, d + 1}.
The fact that the number of steps is at most k − 1 − d is clear: since d + q1 < k − 1, we get
q1 ≤ k − d − 2 so there are at most k − d − 2 non-zero qi , plus qm = 0. We will show that in
fact qd+1 = qd+2 = 0. To see this, note first that C ∪ Zi′ = C ∪ Z2 for any i, since the added
points all lie on C, and recall that Z2 is separated in degree n + 1. It follows that the sequence
0 → H 0 IC∪Z2 (d + n + 1) → H 0 IC (d + n + 1) → H 0 OZ2 → 0
is exact. Indeed, the only issue is the surjectivity of the last map. But if P is a point of Z2 then
there is a form F of degree n + 1 vanishing at the remaining points of Z2 but not at P , and since
IC is generated in degree ≤ d there is a form G of degree d vanishing on C but not at P . The
product F G demonstrates the desired surjectivity.
As a consequence, we have
H 1 IC∪Z2 (d + n + 1) ⊆ H 1 IC (d + n + 1) = 0.
Hence the sequence
0 → H 0 IC∪Z2 (d + n + 1) → H 0 IZ2 (d + n + 1) → H 0 OC (d + n + 1) → 0
is exact. It also means that for any i the map
H 0 IZi′ (d + n + 1) → H 0ID|C (d + n + 1)
surjects, where D is the set on C given by C ∩ Zi′ . Of course H 0 ID|C (d + n + 1) may be zero,
and certainly is zero for large i.
24
L. CHIANTINI AND J. MIGLIORE
Now we prove that qd+1 is 0. As noted above, this number qd+1 is the number of points on
C that we need to add to Zd′ ∪ (Ld+1 ∩ C) (where Ld+1 is a general plane) in order to get a
′
′
(n + d + 2) all contain C. So, if we can prove that
set Zd+1
such that the elements of H 0 IZd+1
′
Zd ∪ (Ld+1 ∩ C) has this property (i.e. surfaces of degree n + d + 2 through Zd′ ∪ (Ld+1 ∩ C)
all contain C), then we must have qd+1 = 0.
To see this, suppose to the contrary that there exists a surface F of degree n + d + 2 passing
through Zd′ ∪ (Ld+1 ∩ C) and not containing C. Then F cuts a divisor E in OC (n + d + 2)
which contains D = C ∩ Zd′ and Ld+1 ∩ C. Then E ′ = E − (Ld+1 ∩ C) is a non-trivial divisor
in OC (n + d + 1) containing D. Thus, by the surjection above, E ′ can be lifted to a surface F ′
of degree d + n + 1 which contains D and Z2 , i.e. F ′ contains Zd′ and does not contain C. This
is impossible by the construction of Zd′ .
′
Thus qd+1 = 0. In the same way we have qd+2 = 0, i.e. Zd+2
has an h-vector which ends in
degree d + n + 3 and has value d in degrees d + n + 2 and d + n + 3.
We now give the estimate for the number of points of Z lying on C in Case 1. We know that
after m steps the one-dimensional part of the base locus is already C, without adding points,
′
and that thus in degree n + 1 + m, the decomposition of Zm
according to Theorem 4.1 holds.
′
The total number of points of Zm that lie on C is minimized in the case that C is a plane
curve, so that the h-vector grows as slowly as possible to reach the value d. Thus the number of
′
points of Zm
that lie on C is at least
(1 + 2 + · · · + (d − 1) + d + d + · · · + d
′
where the final d occurs in degree n + 1 + m. Thus Zm
contains at least
d
d
+ d(n + m − d + 3).
+ d[(n + 1 + m) − (d − 2)] =
2
2
′
To obtain Zm
we added to Z a total of md + (q1 + q2 + · · ·+ qm ) points (where qm = 0). Bearing
in mind that the qi are strictly decreasing as long as they are positive, and that q1 ≤ k − 2 − d,
we have the following lower bound for the cardinality of the original set, Z1 , of points lying on
C.
d
d
+d(n−d+3)−[q1 +· · ·+qm ].
+d(n+m−d+3)−[md + (q1 + · · · + qm )] =
|Z1 | ≥
2
2
By the above claim, we have m ≤ min{k − 1 − d, d + 1}, where m is the smallest index such
that qm = 0. This gives us two possibilities.
ALMOST MAXIMAL GROWTH OF THE HILBERT FUNCTION
25
Assume first that m ≤ k − 1 − d ≤ d + 1, i.e. k ≤ 2d + 2. Notice that this implies
d − (k − 2 − d) ≥ 0, and recall also that k ≤ n and that q1 ≤ k − 2 − d. Then
d
+ d(n − d + 3) − [1 + 2 + · · · + (k − 2 − d)]
|Z1| ≥
2
d
+ d(k − d + 3) − [1 + 2 + · · · + (k − 2 − d)]
≥
2
d
+ [(d − 1) + (d − 2) + · · · + (d − (k − 2 − d))] + 5d
≥
2
d
+ [(d − 1) + (d − 2) + · · · + 1 + 0] + 5d
≥
2
d
d
+ 5d
+
=
2
2
d
+ 5d.
= 2·
2
Now assume that m ≤ d + 1 ≤ k − 1 − d, i.e. k ≥ 2d + 2. We have that qd+1 = 0, so we
maximize q1 + · · · + qm by (k − 2 − d) + · · · + (k − 2d − 1) since the qi are strictly decreasing
and there are at most d non-zero ones.
d
+ d(n − d + 3) − [(k − 2 − d) + · · · + (k − 2d − 1)]
|Z1 | ≥
2
d
d
+ d(n − d + 3) − [2k − 3d + −3]
=
2
2
= d2 + d(n − k + 4)
Case 2: The base locus of [IZ ′ ]n+2 contains C. In this case q1 and q2 are both zero, so the
computation is as before but easier. Clearly in this case there are more points than in Case 1, so
the claimed bound still holds.
5. G ROWTH
TYPE
(. . . , k, k − 1):
THE BASEPOINT FREE CASE
We continue to assume that dim[S/J]n = k ≥ 2 and dim[S/J]n+1 = k − 1. The other
possibility allowed by Lemma 4.2 is that the linear systems |[J]n | and |[J]n+1 | are basepoint
free, i.e. that h[J]≤n i is artinian (hence also h[J]≤n+1 i is artinian):
(16)
From now on we assume that the degree n component of J has no base locus.
Hence the same is true also in degree n + 1.
26
L. CHIANTINI AND J. MIGLIORE
Our goal in this section is to prove the following theorem.
Theorem 5.1. Assume that [J]n and [J]n+1 are base point free, together with the other assumptions highlighted so far in this section. Then there is a distinguished plane Λ such that if Z1 is
the subset of Z lying on Λ and Z2 is the subset of Z lying off Λ then
∆hR/IZ1 (t) = ∆hR/IZ (t)
for t ≥ n
and
∆hR/IZ2 (t) = 0
for t ≥ n − 1.
This will take some preparation, and we devote the rest of this section to its proof.
We know that S/J is artinian, with Hilbert function having values k and k − 1 in degrees
n and n + 1 respectively, and then zero. Clearly the dimension of the socle of S/J in degree
n + 1 is k − 1. Suppose that S/J has a non-zero socle element f in degree n. Then S/(J, f ) has
Hilbert function with value k − 1 in both degrees n and n + 1. Thus the components of (J, f ) in
degrees n and n + 1 have a non-empty base locus. But in degree n + 1, J and (J, f ) coincide.
Thus we have a contradiction to (16).
(17)
From now on we assume that S/J has no non-zero socle in degree n.
It follows that the canonical module of S/J has no minimal generators in the second degree.
Consequently, the same is true for the canonical module of R/IZ :
(18)
From now on we assume that the canonical module of R/IZ has no minimal
generator in its second degree.
With the assumptions (15), (16), (17), (18) in place, we now prove the geometric consequences as described in the theorem.
Let ℓ be an arbitrary (not general) linear form. We have a multiplication
×ℓ : [S/J]n → [S/J]n+1 ,
where these components have dimensions k and k − 1 respectively. Choosing bases for [S/J]n
and [S/J]n+1 and finding the matrices for ×x1 , ×x2 , . . . , ×xr with respect to these bases, we
can represent the multiplication by a linear form ℓ = a1 x1 + a2 x2 + · · · + ar xr as a (k − 1) × k
matrix, A, whose entries are linear forms, mi,j , in the dual variables a1 , . . . , ar . Let us write
m1,1 . . . m1,k
..
..
A=
.
.
mk−1,1 . . . mk−1,k
The vanishing locus of the maximal minors of this matrix corresponds to the set of linear forms
ℓ for which ×ℓ fails to have rank k − 1. This locus has codimension at least 2 in Pr−1 (the dual
projective space).
Remark 5.2. As an aside, suppose that r = 3. Notice that the expected vanishing
locus for the
k
maximal minors of a (k − 1) × k matrix of linear forms is a finite set of 2 points in the dual
projective plane, corresponding to k2 linear forms. If it happens that the degree n component
of J has a base locus consisting of k distinct points, no three on a line, then the linear forms
that fail to give a surjectivity are exactly the k2 lines passing through two of the points.
ALMOST MAXIMAL GROWTH OF THE HILBERT FUNCTION
27
We first note that if ℓ is general then ×ℓ is surjective. Indeed, we have the exact sequence
×ℓ
[S/J]n −→ [S/J]n+1 → [S/(J, ℓ)]n+1 → 0
and Green’s theorem (cf. Theorem 2.4) together with the assumption k ≤ n gives that the last
vector space is zero.
Now we want to describe the possible ranks for arbitrary ℓ. There are essentially four possibilities for the rank of ×ℓ, a priori:
Case 1: rk (×ℓ) = k − 1 (i.e. is surjective) for all ℓ.
This is impossible. As noted above, the vanishing locus of the maximal minors of A has
codimension ≤ 2, so there is at least a finite number of points (depending also on r) in the dual
projective space where this rank is < k − 1.
For the remaining cases, we assume that ×ℓ fails to have maximal rank, and we consider the
exact sequence from (2)
(19)
×ℓ
0 → [S/(J : ℓ)]i−1 −→ [S/J]i → [S/(J, ℓ)]i → 0
for different values of i. We are interested in knowing what are the consequences for ℓ if it fails
to give a multiplication of maximal rank. Note that rk (×ℓ) = dim[S/(J : ℓ)]i−1 .
Case 2: For i = n + 1, rk (×ℓ) = k − 2, so dim[(S/(J : ℓ)]n = k − 2.
We represent the relevant data from (2) in the following table. Since (2) is a short exact
sequence of graded modules, we obtain a collection of short exact sequences of vector spaces,
coming from the homogeneous components of the various degrees, i. Each column in the table
below represents the dimensions of the corresponding vector spaces for the degree given in the
top row. By exactness, the second and third entries must add up to the first entry.
degree i 0
1
dim[S/J]i 1
r
dim[S/(J : ℓ)]i−1
1
dim[S/(J, ℓ)]i 1 r − 1
...
n
n+1
...
k
k−1
...
p
k−2
... k − p
1
where p = dim[S/(J : ℓ)]n−1 ≤ k. For simple Hilbert function reasons, we cannot have
0 ≤ p ≤ k − 3 (looking at the middle line) or p = k (looking at the bottom line). If p = k − 1
then S/(J, ℓ) has maximal growth from degree n to degree n + 1, and hence the base locus of
the component of (J, ℓ) in degree n (and n + 1) is non-empty, meaning that the same is true of
J, violating (16). (Notice that we do not need ℓ to be general in order to reach this conclusion.)
Finally, if p = k − 2 then S/(J : ℓ) has maximal growth from degree n − 1 to degree n, so
[J : ℓ]n has a base locus. But (J : ℓ) ⊃ J, so this means that [J]n has a base locus, again
violating (16).
Having established the idea in Case 2, we combine most of the remaining cases as Case 3.
Case 3: 1 ≤ s = rk (×ℓ) ≤ k − 3.
Then we have the following table:
28
L. CHIANTINI AND J. MIGLIORE
1
degree i 0
dim[S/J]i 1
r
dim[S/(J : ℓ)]i−1
1
dim[S/(J, ℓ)]i 1 r − 1
...
n
n+1
...
k
k−1
...
p
s
... k − p k − 1 −s
Now for Hilbert function reasons p cannot be equal to 0, 1, . . . , s − 1 (looking at the second
row). Looking at the last row, k − p cannot be equal to 0, 1, . . . , k − 2 − s; that is, we cannot
have p ≥ s + 2. This means that we only have to consider p = s and p = s + 1. For this we
argue exactly as in Case 2.
Case 4: rk (×ℓ) = 0.
Since the other cases have been ruled out by our assumptions, this case must hold. This means
that ×ℓ is the zero map on S/J from degree n to degree n + 1, i.e. that [J]n+1 = [(J, ℓ)]n+1 .
Thus [(IZ , L)]n+1 = [(IZ , L, ℓ̄)]n+1 for any lifting ℓ̄ of ℓ to R (note that L and ℓ̄ are independent).
We now have the table
degree i 0
1
dim[S/J]i 1
r
dim[S/(J : ℓ)]i−1
1
dim[S/(J, ℓ)]i 1 r − 1
...
n
n+1
...
k
k−1
...
p
0
... k − p k − 1
It follows immediately that p = 0, i.e. that (×ℓ) is also the zero map from degree n − 1 to
degree n, and hence that [J]n = [(J, ℓ)]n . In particular, [J]n contains ℓ · [S]n−1 and [J]n+1
contains ℓ · [S]n . Thus [(IZ , L)]n = [(IZ , L, ℓ̄)]n for any lifting ℓ̄ of ℓ to R.
We summarize these facts:
(a) There exists a linear form ℓ ∈ S = R/(L) such that ×ℓ : [S/J]n−1 → [S/J]n
and ×ℓ : [S/J]n → [S/J]n+1 are both the zero map.
(20)
(b) For any lifting ℓ̄ of ℓ to R, we have [J]i = [(J, ℓ)]i , i.e. [(IZ , L)]i =
[(IZ , L, ℓ̄)]i for i = n and n + 1.
By duality,
(21)
The linear form ℓ also annihilates the first and second components of the
canonical module of S/J.
We now return to the matrix, A = (mi,j ), that defines the multiplication from degree n to
degree n+1. Since ×ℓ is surjective for general ℓ, A drops rank in codimension 1 or codimension
2. On the other hand, we have seen that when ×ℓ is not surjective, it is the zero map. We obtain:
Lemma 5.3. The ideal hm1,1 , . . . , mk−1,k i generated by the entries of A defines a codimension 2
linear variety in the dual projective space Pr−1 . Thus up to change of variables we may assume
that the mi,j involve only two (dual) variables. In particular, there exist r − 2 independent
ALMOST MAXIMAL GROWTH OF THE HILBERT FUNCTION
29
linear forms in [S]1 that annihilate [S/J]n−1 and [S/J]n . When r = 3, ℓ is unique up to scalar
multiplication.
Proof. We know that the set of ℓ for which ×ℓ is the zero map is a set of codimension ≤ 2,
and it is defined by the ideal hm1,1 , . . . , mk−1,k i. Clearly this defines a linear variety. Suppose
that this variety has codimension 1. Then there is a linear form in the dual variables a1 , . . . , ar
such that each mi,j is a scalar multiple of this linear form. Changing basis if necessary, we
can suppose that this linear form is a1 . This means that for any linear form ℓ involving only
x2 , . . . , xr , we have ×ℓ : [S/J]n → [S/J]n+1 is the zero map, and the same is true replacing n
by n − 1 as noted above. This means that
hx2 , . . . , xr i · [S]n−1 ⊂ [J]n .
But this forces dim[S/J]n and dim[S/J]n+1 both to be ≤ 1, contradicting our assumption that
k > 1. Thus we have shown the first assertion. The second, third and fourth statements are then
immediate.
Lemma 5.4. Let M be a finitely generated graded R-module of depth ≥ 1 and let L be a
general linear form. Let N = M/LM. Without loss of generality say that the initial degree of
M is 0. Assume that there exist linear forms ℓ1 , . . . , ℓr−2 ∈ S = R/(L) that annihilate N0 and
N1 . For each i let ℓ̄i be any lifting of ℓi to R. Then for each i, the image of ×ℓ̄i : M0 → M1
is contained inside the image of ×L : M0 → M1 , and similarly for ×ℓ̄i : M1 → M2 . For a
general lifting ℓ̄, these images actually coincide.
Proof. The first assertion follows from the following commutative diagram (j = 0, 1):
0
↓
0
↓
×ℓ̄
i
Mj−1 −→
↓ ×L
Mj
↓
Nj
↓
0
×ℓ̄
i
−→
×0
−→
Mj
↓ ×L
Mj+1
↓
Nj+1
↓
0
The second assertion follows from a diagram chase, since a general ℓ̄ will be a general element
of the pencil spanned by L and ℓ̄, so it will be a non-zerodivisor for M, since L is.
Corollary 5.5. In the setting of Lemma 5.4, suppose that L and L′ are general linear forms, and
let ℓ1 , . . . , ℓr−2 and ℓ′1 , . . . , ℓ′r−2 be corresponding linear forms in R/L and R/L′ respectively
that annihilate the components Ni and Ni′ = [M/L′ M]i , respectively, where i = 0, 1. Let ℓ̄i and
ℓ̄′i be general liftings. If M has no minimal generator in degree 1, and dim M1 > 2 · dim M0
then L, L′ , ℓ̄1 , . . . , ℓ̄r−2 , ℓ̄′1 , . . . , ℓ̄′r−2 define a point in Pr .
Proof. We have seen that ℓ1 , . . . , ℓr−2 are independent in [S]1 , so L, ℓ̄1 , . . . , ℓ̄r−2 already define
a line in Pr , as do L′ , ℓ̄′1 , . . . , ℓ̄′r−2. Since L′ is general with respect to L and all the ℓi , the linear
30
L. CHIANTINI AND J. MIGLIORE
forms define at most a point. If the statement is not true, then together all these linear forms
span [R]1 . Let us compute the dimension of the image of the multiplication
φ : [R]1 ⊗ M0 → M1 .
By Lemma 5.4, the image of ×L : M0 → M1 contains the image of ×ℓ̄i : M0 → M1 for each i,
and similarly the image of ×L′ : M0 → M1 contains the image of ×ℓ̄′i : M0 → M1 for each i.
Hence if m1 , . . . , ms form a basis of M0 , then Lm1 , . . . , Lms , L′ m1 , . . . , L′ ms are a basis for
the image of φ, so the image has dimension ≤ 2 · dim M0 . Since M has no minimal generator
in degree 1, we have a contradiction. The result follows immediately.
Now let M be the canonical module of R/IZ , suitable shifted so that it begins in degree 0.
Note that N is not the canonical module of S/J, but it is a twist of this module. We have
dim M0 = k − 1 and dim M1 = 2k − 1. We know that M has no minimal generator in degree
1 by (18). Thanks to (21), all the assumptions of Corollary 5.5 are satisfied.
Now let us interpret this geometrically. The general linear form L defines a general hyperplane HL and contains no point of Z. The linear forms 0 6= ℓi ∈ S (1 ≤ i ≤ r − 2)
together define a line λℓ in HL ⊂ Pr . Similarly we have a general hyperplane HL′ and a line
λℓ′ ⊂ HL′ ⊂ Pr . The meaning of Corollary 5.5 is that λℓ and λℓ′ meet in a point. Hence they
span a plane Λ.
Recall the “Linear Lemma” of [11]:
Lemma 5.6. Any set of m-planes such that any two of them meet in an (m − 1)-plane, either is
contained in some fixed Pm+1 or has an (m − 1)-plane for base locus.
In our case, taking m = 1, we have that the lines λℓ as L ranges over R1 either lie in a 2-plane
or have a point as base locus. But in the latter case, choosing a hyperplane that avoids this point
leads to a contradiction. Thus all the lines obtained in this way lie in the plane Λ. Notice that
Λ does not depend on the original choice of L, so we may assume that L is general even with
respect to Λ.
Let H be a general element of [IΛ ]1 , and by abuse of notation we denote by H also the
hyperplane in Pr defined by this linear form. Let
IZ1 = (IZ + IΛ )sat
sat
IZ +IΛ
IZ1 |Λ =
⊂ R/IΛ
IΛ sat
IZ +(H)
IZ1 |H =
⊂ R/hHi
(H)
IZ 2 = IZ : H = IZ : IΛ .
So Z1 is the subset of Z lying on Λ, and Z2 is the subset lying off Λ. Let L be a general linear
I +hLi
form. Let J2 = Z2hLi . Let h be the restriction of H to R/hLi. We have the commutative
ALMOST MAXIMAL GROWTH OF THE HILBERT FUNCTION
31
diagram
h
0 →
IZ :H
IZ
h
0 →
i
IZ :H
IZ
n−1
i
J:h
0 →
J
n
n
0
↓
×H
→ [R/IZ ]n−1 −→
→
→
↓ ×L
[R/IZ ]n
↓
[S/J]n
↓
0
0
↓
[R/IZ ]n
→
↓ ×L
−→ [R/IZ ]n+1 →
×H
×0
−→
↓
[S/J]n+1
↓
0
[R/(IZ , H)]n
→ 0
[R/(IZ , H)]n+1
→ 0
→ [R/(IZ , H, h)]n+1
where the indicated zero map is multiplication by h. We will also shortly consider this diagram
in degree one less. We obtain (again) [J : h]n = [S]n and [R/(IZ , H, h)]n+1 = [S/J]n+1 .
Now, the image of the middle map ×H in this diagram is [R/IZ2 ]n , and the commutativity
of this diagram means that reduction of [R/IZ2 ]n modulo L is zero. As indicated, everything
continues to work in degree one less. Thus we obtain
∆hR/IZ2 (n − 1) = ∆hR/IZ2 (n) = 0.
(22)
Now consider the diagram
0
↓
0 → [R/IZ2 ]n−1
↓ ×L
0 → [R/IZ2 ]n
↓
[S/J2 ]n
↓
0
0
↓
×H
−→ [R/IZ ]n →
[R/(IZ , H)]n
→ 0
↓ ×L
↓ ×L
×H
−→ [R/IZ ]n+1 → [R/(IZ , H)]n+1 → 0
↓
↓
[S/J]n+1
[R/(IZ , H, L]n+1
↓
↓
0
0
We obtain:
(23)
the induced map [S/J2 ]n → [S/J]n+1 is the zero map.
Applying the Snake Lemma and looking at the rightmost column, we thus get
×L
0 → [S/J2 ]n → [R/(IZ , H)]n −→ [R/(IZ , H)]n+1 → [S/J]n+1 → 0
from which we conclude
(24)
∆hR/IZ (n + 1) = ∆hR/(IZ ,H) (n + 1) + ∆hR/IZ2 (n).
The same holds in degree one less thanks to our observations above. Combining with (22) we
have
(25)
∆hR/IZ (n) = ∆hR/(IZ ,H) (n) and ∆hR/IZ (n + 1) = ∆hR/(IZ ,H) (n + 1).
32
L. CHIANTINI AND J. MIGLIORE
Now consider the exact sequences
×H
0 → IZ2 (−1) −→ IZ →
IZ + hHi
→ 0 and
hHi
×H
0 → IZ2 (−1) −→ IZ → IZ1 |H → 0.
+hHi
Equation (22) implies h1 (IZ2 (n − 2)) = 0. Thus IZhHi
is saturated in degrees ≥ n − 1. Since
the Hilbert function of Z1 is the same whether viewed as a subscheme of Pr or as a subscheme
of H or as a subscheme of Λ, we obtain
(26)
∆hR/IZ (t) = ∆hR/IZ1 (t)
for all t ≥ n.
This concludes the proof of Theorem 5.1.
Corollary 5.7. In the setting of Theorem 5.1, at least
k+1
+ (k + 1)(n − k + 2) − 3
2
points of Z must lie on a plane.
Proof. The approach is almost identical to the proof of Corollary 4.8, but now there is a slight
twist. Since [J]n is basepoint free, the h-vector cannot be
(1, 2, . . . , k − 1, k, k, . . . , k, k − 1)
since in this case the component in degree n is still a curve. Thus the smallest possible h-vector
is
(1, 2, . . . , k − 1, k, k + 1, k + 1, . . . , k + 1, k, k − 1)
and this gives the desired bound.
6. E XTENDING G OTZMANN ’ S
THEOREM
We have noted that in [5], use was made of Gotzmann’s Persistence Theorem, which described the behavior of the Hilbert function of Z1 , the subset of Z lying on the base locus (under
the assumption of maximal growth), assuming that no additional generators were present in its
ideal. Although it is not the focus of this paper, it is still of interest to know what the behavior of
the Hilbert function can be when we have almost maximal growth and no additional generators.
To illustrate that something can be said, we consider the situation of Theorem 5.1.
Note that even if Z ⊂ Pr , its artinian reduction agrees in degrees ≥ n with the artinian
reduction of Z1 , the subset of Z lying on the plane Λ. Thus without loss of generality we
can assume that our almost maximal growth arises in the setting of an algebra S/J where
S = k[x, y].
Proposition 6.1. Let S = k[x, y], and let J be a homogeneous ideal in S such that for some
integer n the following hold:
(a) The linear system defined by [J]n has no base locus;
(b) hS/J (n + 1) = hS/J (n) − 1;
(c) J has no minimal generators in degree > n.
ALMOST MAXIMAL GROWTH OF THE HILBERT FUNCTION
33
Then for all j ≥ n we have hS/J (j + 1) = max{hS/J (j) − 1, 0}.
Proof. The minimal free resolution of J has the form
0 → F2 → F1 → J → 0
where all summands S(−j) of F1 satisfy j ≤ n. Condition (a) guarantees that J contains a
regular sequence of two forms of degree n.
Now suppose that the assertion is not true. Then there exists some k ≥ n + 1 for which
hS/J (k) = hS/J (k − 1) − 1 and hS/J (k + 1) ≤ hS/J (k) − 2.
Let c be a complete intersection of type (n, n) in I. It links J to a homogeneous artinian ideal
J ′ ⊂ S. We have
hS/J ′ (2n − k − 3) > hS/J ′ (2n − k − 2) = hS/J ′ (2n − k − 1).
Since the latter two represent maximal growth (or possibly 0), a result of [12] (see also [15]
Theorem 3.4) gives that S/J ′ has a non-zero socle element in degree 2n − k − 3. Hence the
minimal free resolution of S/J ′ has a free summand S(−2n + k + 1) in the last free module. By
linkage, this means that J has a minimal generator in degree k+1 ≥ n+2, contradicting (c).
7. F URTHER
QUESTIONS
Finally, we present some open problems which we leave for future study.
1. When the base locus is of dimension ≥ 2, can we obtain a good bound on the number of
points of Z lying on the base locus?
2. Are there other results extending Gotzmann’s theorem besides the one given in section
6?
3. What happens if we do not assume that Z is reduced? Can we still obtain similar results?
In this case, Z1 becomes the subscheme of Z lying on the base locus, which may or may
not be reduced.
4. Can we obtain similar results for Z of higher dimension?
5. What happens when Z is a reduced set of points in uniform position? In [5] several nice
consequences were obtained. Perhaps similar results can be found here. In particular,
must the base locus be irreducible and must all the points of Z lie on it?
6. We believe that Theorem 3.1 and Theorem 3.4 can be extended to “maximal growth
minus 2,” and indeed to “maximal growth minus ℓ” for ℓ ≤ r − 2. Specifically, there
should be ℓ + 1 possibilities for the dimension of the base locus in general.
7. In [9] and [10], when the set Z ⊂ P3 is the general hyperplane section of an irreducible
curve in P4 , a result similar to Theorem 5.1 is obtained, with a different approach. More
generally, when there is a large monodromy group acting on Z, the existence of a surface
of degree ℓ containing Z is proven when “maximal growth minus ℓ” holds, for some ℓ.
We wonder if the approach introduced above can extend the results of [9] and [10] to any
sets of points in (very?) uniform position.
8. As we mentioned in the Introduction, we have in mind some applications of our results
on almost maximal growth to the study of symmetric tensors. What other applications
can we obtain from these results?
34
L. CHIANTINI AND J. MIGLIORE
R EFERENCES
[1] E. Ballico and A. Bernardi, A partial stratification of secant varieties of Veronese varieties via curvilinear
subschemes, Sarajevo J. Math. 8 (2012), 33–52.
[2] E. Ballico and L. Chiantini, A criterion for detecting the identifiability of symmetric tensors of size three, Diff.
Geom. Applic. 30 (2012), 233–237.
[3] A. M. Bigatti, Upper Bounds for the Betti Numbers of a Given Hilbert Function, Comm. Algebra 21 (1993),
2317–2334.
[4] A.M. Bigatti and A.V. Geramita: Level Algebras, Lex Segments and Minimal Hilbert Functions, Comm. in
Algebra 31 (2003), 1427-1451.
[5] A.M. Bigatti, A.V. Geramita and J. Migliore, Geometric consequences of extremal behavior in a theorem of
Macaulay, Trans. Amer. Math. Soc. 346 (1994), 203–235.
[6] W. Bruns and J. Herzog: Cohen-Macaulay rings, Cambridge studies in advanced mathematics, No. 39, Revised edition (1998), Cambridge, U.K.
[7] J. Buczyński, A. Ginensky and J. M. Landsberg, Determinantal equations for secant varieties and the
Eisenbud-Koh-Stillman conjecture, J. London Math. Soc. 88 (2013), 1–24.
[8] L. Chiantini, N. Chiarli and S. Greco, Halphen conditions and postulation of nodes, Adv. Geom. 5 (2005),
237–264.
[9] L. Chiantini, C. Ciliberto and V. di Gennaro, The genus of projective curves in P4 verifying certain flag
conditions, Manuscr. Math. 88 (1995), 119–134.
[10] L. Chiantini and C. Ciliberto, Towards a Halphen theory of linear series on curves, Trans. Amer. Math. Soc.
351 (1999), 2197–2212.
[11] L. Chiantini and M. Coppens, Grassmannians of secant varieties, Forum Math. 13 (2001), 615–628.
[12] Y. Cho and A. Iarrobino, Hilbert functions of level algebras, J. Algebra 241 (2001), 745–758.
[13] CoCoATeam, CoCoA: a system for doing Computations in Commutative Algebra. Available at
http://cocoa.dima.unige.it
[14] E.D. Davis, Complete Intersections of Codimension 2 in Pr : The Bezout-Jacobi-Segre Theorem Revisited,
Rend. Sem. Mat. Univers. Politecn. Torino, Vol. 43, 2 (1985), 333–353.
[15] A. Geramita, T. Harima, J. Migliore and Y. Shin, “The Hilbert function of a level algebra,” Memoirs AMS
vol. 186, no. 872 (2007).
[16] A.V. Geramita, M. Kreuzer and L. Robbiano, Cayley-Bacharach schemes and their canonical modules, Trans.
Amer. Math. Soc. 339 (1993), 163–189.
[17] A.V. Geramita, P. Maroscia and L. Roberts, The Hilbert function of a reduced K-algebra, J. London Math.
Soc. (2) 28 (1983), no. 3, 443–452.
[18] G. Gotzmann, Eine Bedingung für die Flachheit und das Hilbertpolynom eines graduierten Ringes, Math. Z.
158 (1978), 61–70.
[19] M. Green, Restrictions of linear series to hyperplanes, and some results of Macaulay and Gotzmann, Algebraic curves and projective geometry (1988), 76-86, Trento; Lecture Notes in Math. 1389 (1989), Springer,
Berlin.
[20] L. Gruson, R. Lazarsfeld and C. Peskine, On a theorem of Castelnuovo and the equations defining space
curves, Invent. Math. 72 (1983), 491–506.
[21] H. A. Hulett, Maximum Betti Numbers of Homogeneous Ideals with a Given Hilbert Function, Comm. Algebra. 21 (1993), 2335–2350.
[22] J.O. Kleppe, J. Migliore, R.M. Miró-Roig, U. Nagel and C. Peterson, “Gorenstein liaison, complete intersection liaison invariants and unobstructedness,” Mem. Amer. Math. Soc. 154 (2001), no. 732.
[23] F.S. Macaulay, Some properties of enumeration in the theory of modular systems, Proc. London Math. Soc.
26 (1927), 531–555.
[24] P. Maroscia, Some problems and results on finite sets of points in Pn , Lecture Notes in Math., vol. 997,
Algebraic Geometry – Open Problems (Proceedings, Ravello 1982), Springer, 1983.
ALMOST MAXIMAL GROWTH OF THE HILBERT FUNCTION
35
[25] J. Migliore, “Introduction to Liaison Theory and Deficiency Modules,” Progress in Math., vol. 165,
Birkhäuser Boston, Massachusetts, 1998.
[26] F. Orecchia and I. Ramella, The conductor of one-dimensional local rings with reduced associated graded
ring, Ric. Mat. XLIV (1995), 231–236.
[27] K. Pardue, Deformation Classes of Graded Modules and Maximal Betti Numbers, Illinois J. Math. 40 (1996),
564–585.
(L. Chiantini) D IPARTIMENTO DI I NGEGNERIA DELL’I NFORMAZIONE E S CIENZE M ATEMATICHE , U NIVER S IENA , V IA ROMA 56 (S.N ICCOLO ’), 53100 S IENA , I TALY
E-mail address: luca.chiantini@unisi.it
URL: http://www.diism.unisi.it/˜chiantini
SITA ‘ DI
(J. Migliore) D EPARTMENT OF M ATHEMATICS , U NIVERSITY
E-mail address: migliore.1@nd.edu
URL: http://www.nd.edu/˜jmiglior/
OF
N OTRE DAME , N OTRE DAME , IN 46556
| 0math.AC |
Exploration of Pattern-Matching Techniques for
Lossy Compression on Cosmology Simulation
Data Sets
Dingwen Tao1 , Sheng Di2 , Zizhong Chen1 , and Franck Cappello2,3
arXiv:1707.08205v2 [cs.IT] 6 Aug 2017
1
3
University of California, Riverside, CA, USA
{dtao001, chen}@cs.ucr.edu
2
Argonne National Laboratory, IL, USA
{sdi1, cappello}@anl.gov
University of Illinois at Urbana-Champaign, IL, USA
Abstract. Because of the vast volume of data being produced by today’s scientific simulations, lossy compression allowing user-controlled
information loss can significantly reduce the data size and the I/O burden. However, for large-scale cosmology simulation, such as the Hardware/Hybrid Accelerated Cosmology Code (HACC), where memory overhead constraints restrict compression to only one snapshot at a time, the
lossy compression ratio is extremely limited because of the fairly low spatial coherence and high irregularity of the data. In this work, we propose
a pattern-matching (similarity searching) technique to optimize the prediction accuracy and compression ratio of SZ lossy compressor on the
HACC data sets. We evaluate our proposed method with different configurations and compare it with state-of-the-art lossy compressors. Experiments show that our proposed optimization approach can improve
the prediction accuracy and reduce the compressed size of quantization
codes compared with SZ. We present several lessons useful for future
research involving pattern-matching techniques for lossy compression.
1
Introduction
Because of ever-increasing parallel execution scale, today’s scientific simulations
are producing volumes of data too large to be accommodated in storage systems.
The limitation comes from the limited storage capacity and I/O bandwidth of
parallel file systems in production facilities. Cosmology simulations such as the
Hardware/Hybrid Accelerated Cosmology Code (HACC) [12] are typical examples of parallel executions facing this issue. HACC solves an N-body problem
involving domain decomposition, a medium-/long-range force solver based on
a particle-mesh method, and a short-range force solver based on a particleparticle/particle-mesh algorithm. According to cosmology researchers, the number of particles to simulate can be up to 3.5 trillion in today’s simulations (and
even more in the future), which leads to 60 PB of data to store; yet a system
such as the Mira supercomputer has only 26 PB of file system storage. Currently,
HACC users rely on decimation in time, storing only a fraction of the simulation
snapshots, to reduce the pressure on the storage system. A reduction factor of
80% to 90% is commonly used. At exascale, temporal decimation will not be
enough to address the limitations of the storage system: snapshots will be so
large (each in the range of 5 PB) that the time to store each snapshot (83 minutes on a storage system offering a sustained bandwidth of 1 TB/s) will become
a serious problem. HACC is not a special case. As indicated by [11], nearly 2.5
PB of data were produced by the Community Earth System Model for the Coupled Model Intercomparison Project (CMIP) 5, which further introduced 170
TB of postprocessing data submitted to the Earth System Grid [3]. Estimates of
the raw data requirements for the CMIP6 project exceed 10 PB [2]. At exascale,
storing each full snapshot in this case would also take too long, however, so that
on-line/in situ compression of each snapshot is needed.
In this paper, we explore pattern-matching techniques for lossy compression,
focusing on individual snapshots of the scientific data sets produced by cosmology simulations. Because of the constraints of memory consumption, we cannot
leverage the smoothness of a particle’s trajectory (such as smoothness along the
time dimension) to reduce the data size; hence, we must perform compression
on individual snapshots. Unlike the mesh data produced by conventional simulations, such as fluid dynamics, the data of particles in cosmology simulations,
such as coordinate and velocity data, are stored in separate 1D arrays. In the
HACC application, the indices of each 1D array are kept consistent for the same
cosmology particle. Specifically, the HACC simulation data contains six 1D arrays: three coordinate fields (xx, yy, zz) and three velocity fields (vx, vy, vz).
Because of the lack of correlation between adjacent particles in the HACC data
set, state-of-the-art lossy compressors, such as FPZIP [16], ZFP [15] and SZ [9,
20], reach relatively low compression ratios/factors (2 to 5 with the error bound
set to 10−4 ).
The rest of the paper is organized as follows. In Section 3, we formulate
the data compression problem based on cosmology simulation data sets and the
assessment of several state-of-the-art lossy compressors on the HACC data sets.
In Section 4, we discuss the well-known dictionary-based lossless compression
algorithm LZ77 and propose our pattern-matching-based optimization method
for SZ lossy compression for low spatial coherence and highly irregular data, such
as the velocity variables in the HACC data sets. In Section 5, we evaluate the
compression ratios of our proposed optimization method and compare it with
one variant of the SZ lossy compressor. We discuss related work in Section 2 and
provide conclusions in Section 6.
2
Related Work
Data compression has been extensively studied for decades and can be split into
two categories: lossless compression and lossy compression. The main limiation of
the lossless compressors (such as GZIP [9]) is their fairly low compression ratio on
scientific data sets composed of floating-point values, as confirmed by [10,20,22].
Data Prediction
Quantization
Variable-length
Encoding
Unpredictable
Data Compression
Dictionary-based
Encoding
SZ Compression Framework
Fig. 1. Overview of SZ lossy compression algorithm.
Recently, many lossy compressors have been designed and implemented for
scientific data. Most of them are designed for mesh data sets, which are expected
to have strong coherence among the nearby data in the data set, but the quality
of their compression declines on cosmology simulation data sets. For example,
SZ [10, 22] has five main steps including (1) data prediction for each point by
its preceding neighbors in the multidimensional space, (2) error-controlled linear
quantization, (3) customized Huffman coding [13] (i.e., variable-length encoding) to shrink the data size significantly, (4) unpredictable data compression,
and (5) customized LZ77 coding (i.e., dictionary-based encoding). The compression framework of SZ is shown in Figure 1. ZFP [16] splits the whole data set into
many small blocks with an edge size of 4 along each dimension and compresses
the data in each block separately by a series of carefully designed steps (including
alignment of exponent, orthogonal transform, fixed-point integer conversion, and
binary representation analysis with bit-plane encoding). FPZIP [17] adopts predictive coding and ignores insignificant bit planes in the mantissa based on the
analysis of IEEE 754 binary representation [7]. SSEM [21] splits data into a highfrequency part and low-frequency part by wavelet transform [8] and then uses
vector quantization and GZIP. ISABELA [15] sorts the data and then performs
the data compression by B-spline interpolation; but it has to store an extra index
array to record the original location for each point, and it suffers significantly
from low compression ratio. Compression of particle simulation data sets has also
been studied for years, but most of the methods proposed are based on smooth
temporal trajectory of the same particles, which requires loading/keeping multiple snapshots during the compression/simulation [1, 6, 14, 18, 23]. Thus, they
are not suitable for extremely large-scale simulation in which only one snapshot
is allowed to be loaded into the memory. Omeltchenko et al. [19] proposed a
lossy compression method (called CPC2000 in this paper) that does not rely on
temporal coherence and relies on only a single snapshot. Its main steps involve
reorganizing all particles in the space onto a zigzag-similar space-filling curve [5],
sorting the particles based on the R-indices by a radix-similar sorting method in
each block, and compressing the difference of the adjacent indices by adaptive
variable-length coding.
3
Problem Formulation
Scientific data compression algorithms can be classified into two categories: lossless compression and lossy compression. The main limitation of lossless compressors is their limited data reduction capability, that is, up to 2:1 in general [20] and
even lower on cosmology simulation simulation data sets. In this work, therefore,
we focus on lossy compression methods for cosmology simulations.
Cosmology simulations generate multiple snapshots. Because of considerations of memory consumption, we focus on single-snapshot compression without
using temporal coherence in this work. Such simulations contain many variables
each representing one data field of particles. In the HACC simulation data considered in this study, the variables are stored in separate 1D arrays. Specifically,
each snapshot of HACC simulation contains six single-precision floating-point
variables: xx, yy, zz, vx, vy, and vz. The first three indicate coordinate information, and the other three indicate velocity along the three dimensions. The
six variables are stored in separate floating-point arrays. Unlike regular multidimensional mesh data, the particle elements in each array are allowed to be
reordered in the reconstructed data set, whereas the locations or indices of the
elements with regard to the same particle must be consistent across arrays.
The main objective of our work is to optimize the single-snapshot lossy compression ratio for cosmology simulation data sets, provided that the compression
errors are controlled within a user-specified bound for each data point. Compression ratio is the ratio of the original data size to the compressed data size. Table
1 shows the compression ratios of several state-of-the-art lossy compressors on
the HACC data sets under the value-range-based relative error bound 10−4 , denoted by ebrel = 10−4 . The version of the SZ lossy compressor we focus on in this
work is “SZ-LV”, which is based on the last-value prediction model. Note that
for CPC2000, ZFP, and SZ, we use the absolute error bounds computed based on
ebrel = 10−4 and the value range of each variable; for FPZIP, we set the number
of retained bits to 21 as approximate ebrel = 10−4 for all the variables. The SZ
lossy compressor has higher compression ratios on the coordinate variables (i.e.,
xx, yy, zz) than on the velocity variables (i.e., vx, vy, vz). Therefore, in this work
we focus on optimizing the prediction accuracy and compression ratios based on
SZ lossy compression for the velocity variables in the HACC data.
Table 1. Compression ratios of different variables with different compressors on HACC
data sets under value-range-based relative error bound 10−4 .
Compressor
CPC2000
FPZIP
ZFP
SZ
xx
7.1
5.8
2.3
8.2
yy
7.1
5.7
2.3
8.3
zz
7.1
4.4
2.2
5.9
vx
2.3
2.2
2.3
4.0
vy
2.3
2.2
2.3
4.0
vz
2.3
2.2
2.3
4.0
4
Pattern-Matching Techniques for Lossy Compression
In this section, we first discuss the well-known dictionary-based lossless compression algorithm Lempel-Ziv 77 (LZ77). It can encode a sequence of symbols
and compress the input source by using the information of recently frequent consecutive symbols. Inspired by LZ77’s classic idea, we then propose our patternmatching-based lossy compression method, called SZ-PM. Because of different
input sources, we propose many tailored designs for dealing with lossy compression and floating-point scientific data.
4.1
LZ77: string matching based lossless compression
while look-ahead buffer is not empty do
go backwards in search buffer to find longest match of the look-ahead
buffer;
if match found then
output (offset, length, next symbol in look-ahead buffer);
shift sliding window by length+1;
else
output (0, 0, first symbol in look-ahead buffer);
shift sliding window by 1;
end
end
Algorithm 1: Pseudo code of the LZ77 algorithm
The Lempel-Ziv 77 (LZ77) lossless compression algorithm is the first LempelZiv compression algorithm. Unlike scientific data compression, LZ77 is designed
for encoding a sequence of symbols byte by byte based on a dictionary constructed from a portion of the recently encoded sequence. Specifically, LZ77
encodes the input sequence through a sliding window composed of two buffers,
a search buffer and a look-ahead buffer, as shown in Figure 2. The search buffer
contains the most recently compressed symbols, while the look-ahead buffer contains multiple uncompressed symbols. The algorithm searches the longest prefix
of the look-ahead buffer that is also contained in the search buffer. The details of
LZ77 are shown in Algorithm 1. The LZ77 algorithm searches all the consecutive
symbols in the search buffer to identify whether these symbols match the consecutive symbols in the look-ahead buffer. The offset in the algorithm represents
the distance of the longest match’s first symbol (in the search buffer) from the
look-ahead buffer, and length represents the length of the longest match. Therefore, the general idea of LZ77 is to save storage by using the information from
the recent symbol sequences based on a string-matching approach. It inspires us
to design a similar matching technique for lossy scientific data compression.
4.2
SZ-PM: pattern-matching-based lossy compression
We propose a pattern-matching-based lossy compression method called SZ-PM.
The idea of pattern matching is similar to the string matching idea used in LZ77.
Dictionary
Input sequence
Encoded
sequence
Search buffer
Look ahead
buffer
Not encoded
sequence
Sliding window
Fig. 2. Overview of LZ77 lossless compression algorithm.
It is also designed to use the information of recent floating-point sequences with
similar pattern in order to improve the prediction accuracy and compression
ratio of SZ lossy compression for irregular data. Unlike the lossless compression
algorithm for symbols (one byte per symbol), however, the lossy compression
for scientific data is designed mainly for single/double floating-point data (4/8
bytes per value) and can tolerate compression errors within user-controlled error bounds. Therefore, we can design many tailored features for the patternmatching method.
Let us first define necessary notations. Similar to LZ77, our algorithm also
maintains two buffers in the sliding window during the compression: a search
buffer and a look-ahead buffer. Let the search buffer size be m and the lookahead buffer size be n. Here the buffer size represents the number of data
points in the buffer. Let the m compressed data points in the search buffer be
{s1 , s2 , ..., sm } and the n uncompressed data points in the look-ahead buffer be
{l1 , l2 , ..., ln }. Let the m−n+1 sequences with length of n in the search buffer to
be X1 , X2 , ..., Xm−n+1 , where X1 = {s1 , ..., sn }, X2 = {s2 , ..., sn+1 }, ..., Xm−n+1 =
{sm−n+1 , ..., sm }. Let the one sequence with length of n in the search buffer be
Y = {l1 , l2 , ..., ln }.
We now describe our tailored designs of pattern matching for lossy compression and scientific data. For compression, (1) we fix the length of matching
sequences to be the size of look-ahead buffer (i.e., n). In other words, we attempt to identify the most similar sequence in the search buffer for the whole
look-ahead buffer with length n. (2) We sort the n data points in each sequence,
including X1 , X2 , ..., Xm−n+1 from the search buffer and Y from the look-ahead
buffer. (3) For each sorted sequence, we subtract the mean value of the sequence
from each value. In other words, we shift the sequence by its mean value as
X = (x1 − X, x2 − X, ..., xn − X) and Y = (y1 − Y , y2 − Y , ..., yn − Y ), where
n
n
P
P
X = n1
xi and Y = n1
yi . (4) We attempt to match the sequences from
i=1
i=1
the search buffer for the look-ahead buffer, but we relax the “matching” condition. Specifically, the matching condition of LZ77 algorithm is that two symbol sequences are exactly the same; but in our algorithm we define two shifted
floating-data sequences X = (x1 , x2 , ..., xn ) and Y = (y1 , y2 , ..., yn ) as “matched”
n
P
if ( |xi − yi |p )1/p < θ, where θ is a given threshold, X is one shifted sequence
i=1
from the search buffer, and Y is the shifted sequence of the look-ahead buffer.
Note that the search buffer can have multiple matched sequences. (5) We pick
the matched sequence X ∗ with the smallest distance from the multiple matched
sequences as the most similar sequence for Y . We denote the values in X ∗ by
{x∗1 , x∗2 , ..., x∗n }. We name this matching process as “pattern matching” and the
sequence X ∗ as the “pattern matched sequence” for Y . (6) We always shift the
sliding window by length of n after we go over the m − n + 1 sequences in the
search buffer. Unlike LZ77, we also shift the sliding window by length of n, even
if we cannot find a matched sequence under the threshold θ. (7) We use X ∗ as
the prediction sequence for Y , if the pattern matched sequence can be found.
Specifically, we take x∗i − X ∗ as the prediction value for yi − Y of data point i.
We use SZ’s original prediction model proposed in [22] to generate the prediction
values for Y , if no matched sequence exists in the search buffer. Therefore, we
must use an extra bit, denoted by bitpredmd , to represent the prediction method
of each sequence. For example, we use bitpredmd = 0 to indicate that the sequence
is predicted by pattern-matching method and bitpredmd = 1 to indicate that the
sequence is predicted by SZ’s original prediction model. (8) Similar to LZ77, if
the sequence is predicted by the pattern-matching method, we still have to store
the offset; but we do not need to store the length due to the fixed length. We also
have to store the mean value of Y in order to reconstruct the data during the
decompression. (9) We use the linear quantization method and the customized
Huffman coding proposed in [22] to encode the differences between prediction
values and real values for Y and compress the quantization codes based on the
user-set error bound. Because of space limitations, we do not describe them in
detail here.
For decompression, we use the same decompression method proposed in [22]
to construct the differences between prediction values and real values for each
sequence. For example, in decompressing the sequence Y , we denote the difference of data point i in Y by yidif f . We then construct the prediction values of Y
by its corresponding prediction method known from bitpredmd . If bitpredmd indicates Y is predicted by SZ’s original prediction model during the compression,
we construct its prediction values using the same process described in [22]; if
bitpredmd indicates Y is predicted by the pattern-matching approach during the
compression, we use the stored offset and mean value to construct the prediction values. Specifically, we can construct the prediction value of data point i
by yipred = x∗i − X ∗ + Y , where X ∗ is the pattern-matched sequence that has
already been decompressed. After constructing the prediction values for Y , we
can reconstruct the value of data point i by yidecomp = yipred + yidif f .
Algorithm 2 shows the pseudo code of our proposed pattern-matching-based
lossy compression method. Figure 3 shows an example of two pattern-matched
sequences transformed by sorting and shifting. We have several remarks here. (1)
For our matching condition, we treat the two n-length floating-point sequences
as two data points in the n-dimensional space and define them as “matched”
if their distance in Lp norm is smaller than the threshold θ. According to [4],
we set θ to 0.5 of the search buffer size. (2) From our initial study we find that
p > 1 cannot reduce the size of the compressed quantization codes on the HACC
data; hence we set p = 1/2 in our algorithm and the following evaluation. (We
will research the optimal p in the future.) (3) As a result of the sorting process,
the reconstructed data is recorded in one sequence. But as described in Section
3, the particle elements in each 1D array are allowed to be reordered in the
reconstructed data sets. Hence, we do not have to extra storage to record the
initial index information. (4) We use extra memory space to sort and shift the
sequences without any modifications of the original data. The reason for sorting
and shifting is to increase the possibility of matching sequences due to the high
irregularity of the data and the relatively large value range of the floating-point
data.
while look-ahead buffer is not empty do
sequence Y is composed of the n data points of the look-ahead buffer;
search buffer contains m − n + 1 sequences {X1 , X2 , ..., Xm−n+1 };
sort each sequence including X1 , X2 , ..., Xm−n+1 and Y ;
compare sorted Y with {X1 , X2 , ..., Xm−n+1 } and find sequence X ∗
with the smallest distance (in Lp norm) from Y, i.e., dist(X ∗ , Y );
if dist(X ∗ , Y ) < θ then
bitpredmd = 0;
store (offset, mean value Y );
prediction values of Y are calculated by yipred = x∗i − X ∗ + Y ;
else
bitpredmd = 1;
use SZ’s original prediction model to predict values of Y ;
end
calculate differences between real value yi and prediction value yipred ;
encode differences using linear quantization method based on user-set
error bound;
compute and record decompressed value;
shift sliding window by length of n;
end
compress linear quantization codes using Huffman coding;
compress unpredictable data by SZ’s binary representation analysis;
Algorithm 2: Pseudo code of SZ-PM algorithm
5
Empirical Evaluation
In this section, we evaluate our proposed lossy compression method, SZ-PM, on
the velocity variables in the HACC data sets, and we compare it with the SZ
lossy compressor [22]. Note that the SZ lossy compressor we evaluate in this
study is a variant of the original SZ. It first splits the original data into multiple
segments. The segment size is consistent with the look-ahead buffer size. It then
performs a sorting within each segment. After that, it conducts the original SZ
compression on the transformed data. The reason of using this variant version
Sequence X*
1
0.8
0.6
0.4
0.2
0
-0.2
-0.4
-0.6
-0.8
-1
1
2
3
4
5
6
7
Sequence Y
8
9 10 11 12 13 14 15 16
Quantization Interval
(a) SZ
9.0%
8.0%
7.0%
6.0%
5.0%
4.0%
3.0%
2.0%
1.0%
0.0%
1
30
59
88
117
146
175
204
233
262
291
320
349
378
407
436
465
494
Probablity
9.0%
8.0%
7.0%
6.0%
5.0%
4.0%
3.0%
2.0%
1.0%
0.0%
1
30
59
88
117
146
175
204
233
262
291
320
349
378
407
436
465
494
Probability
Fig. 3. Example of two pattern matched sequences after sorting and shifting.
Quantization Interval
(b) SZ-PM
Fig. 4. Distribution produced by linear quantization encoder in (a) SZ and (b) SZ-PM
on the velocity variable vx in the HACC data sets with 511 quantization intervals.
is that we want to evaluate the effects of the pattern-matching method without
impact from the sorting technique and to compare SZ and SZ-PM in a fair level.
As described in [22], the distribution produced by linear quantization encoder can significantly affect the performance of Huffman coding [13]. Generally
speaking, the more concentrated the distribution, the higher the compression
ratio that the Huffman coding can achieve. Figure 4 shows the distributions produced by linear quantization encoder in the SZ and our proposed SZ-PM lossy
compression method on the velocity variable vx in the HACC data sets. Note
that we use a 10−4 value-range-based relative error bound and 511 quantization
intervals. Based on our observation, 511 quantization intervals can cover more
than 99.9% data points during the linear quantization in this case. The figure
illustrates that our proposed SZ-PM can improve the prediction accuracy and
make the distribution of quantization code more concentrated. (We will show
the incremental results in detail later.)
Table 2. Evaluation of our proposed SZ-PM on the velocity variable vx in the HACC
data sets with different sizes of sorting/matching sequence.
Size of
Size of
Ratio of
Size of
Size of
Overall
Compression
Quantization Code bitpredmd PM Sequence
Offset
Mean Value Bit-rate
Ratio
(bits/value)
(bits/value)
(%)
(bits/value) (bits/value) (bits/value)
CPC2000
/
/
/
/
/
13.9
2.30
SZ(8)
7.31
/
/
/
/
7.3
4.38
SZ-PM(8)
5.45
1/8
99.6%
1.25
3.98
10.8
2.96
SZ(16)
6.75
/
/
/
/
6.8
4.74
SZ-PM(16)
6.01
1/16
93.1%
0.58
1.86
8.5
3.76
SZ(32)
6.16
/
/
/
/
6.2
5.19
SZ-PM(32)
6.07
1/32
66.3%
0.04
0.66
6.8
4.71
Table 2 shows the experimental results of our evaluation for SZ-PM on the
HACC data sets. In the experiments, we set the search buffer size to 1024;
hence, we need to use 10 bits (210 = 1024) to represent the offset value for each
sequence that is predicted by the pattern-matching method during compression.
We test SZ-PM with different configurations of three look-ahead buffer sizes:
8, 16, and 32. The size of each category presented in the table is the atomized
size (i.e., bits per value). Note that the original data type of the HACC data
is single floating-point (i.e., 32 bits per value); hence, the compression ratio
can be calculated by 32/overall size. The number in each bracket represents
the segment size/sequence size; for example, SZ(8) means that the segment size
used for sorting in SZ is 8, and SZ-PM(8) means that the length of sequence used
in the pattern matching is 8. The column “Ratio of PM Sequence” means the
ratio of the sequences predicted by the pattern matching during compression.
We make several observattions from Table 2. (1) SZ-PM can improve the
prediction accuracy and reduce the size of the compressed quantization codes.
(2) The shorter the matching sequence is, the more accurately the SZ-PM can
predict. (3) For SZ-PM, the shorter the matching sequence is, the smaller the
compressed quantization codes will be; however, for SZ, on the contrary, the
longer the segment is, the smaller the compressed quantization codes will be. (4)
The longer the matching sequence is, the less the storage overhead that the offset
and mean values will have. (5) The reduced size of the compressed quantization
codes, achieved from the improvement of the prediction accuracy by SZ-PM, is
counteracted by the incremental overhead of storing offset and mean values.
From these observations, we derive some useful lessons for future research
with respect to the pattern-matching techniques in lossy compression as follows.
(1) Our proposed pattern-matching technique can enhance the prediction accuracy and reduce the size of compressed quantization codes, but the improvement
is not enough to cover the extra overhead introduced by storing offset and mean
values. (2) We should further improve the prediction accuracy using a more advanced pattern-matching technique. (3) We should reduce/eliminate the extra
overhead of offset and mean values, especially the mean values of floating-point
data type. For example, we may shift the sequence by the value of the first element in the sequence; consequently, we do not need to store the mean values.
(4) Currently, we consider reordering only one variable in the HACC data sets.
In future research, we need to consider the impact of reordering one variable to
the other variables, since we have to make all the variables consistent.
6
Conclusion
In this work, we explored pattern-matching techniques for lossy compression
based on the SZ compressor. The experiments demonstrate that our proposed optimization method, SZ-PM, can improve the prediction accuracy and reduce the
size of compressed quantization codes on the HACC velocity data, but the compression ratio cannot be improved because of storing extra information. We plan
to explore ways to improve the prediction accuracy with the pattern-matching
technique and to reduce the storage of extra information.
Acknowledgments
This material is based upon work supported by the U.S. Department of Energy, Office of Science, under contract number DE-AC02-06CH11357. The submitted manuscript has been created by UChicago
Argonne, LLC, Operator of Argonne National Laboratory (“Argonne”). Argonne, a U.S. Department
of Energy Office of Science laboratory, is operated under Contract No. DE-AC02-06CH11357. The
U.S. Government retains for itself, and others acting on its behalf, a paid-up nonexclusive, irrevocable worldwide license in said article to reproduce, prepare derivative works, distribute copies
to the public, and perform publicly and display publicly, by or on behalf of the Government. The
Department of Energy will provide public access to these results of federally sponsored research in
accordance with the DOE Public Access Plan. http://energy.gov/downloads/doe-public-access-plan.
References
1. Ahmed, N., Natarajan, T., Rao, K.R.: Discrete cosine transform. IEEE Transactions on Computers 100(1), 90–93 (1974)
2. Baker, A.H., Xu, H., Dennis, J.M., Levy, M.N., Nychka, D., Mickelson, S.A., Edwards, J., Vertenstein, M., Wegener, A.: A methodology for evaluating the impact
of data compression on climate simulation data. In: HPDC’14. pp. 203–214 (2014)
3. Bernholdt, D., Bharathi, S., Brown, D., Chanchio, K., Chen, M., Chervenak, A.,
Cinquini, L., Drach, B., Foster, I., Fox, P., et al.: The Earth System Grid: Supporting the next generation of climate modeling research. Proceedings of the IEEE
93(3), 485–495 (2005)
4. Chan, K.P., Fu, A.W.C.: Efficient time series matching by wavelets. In: Proceedings
of the 15th International Conference on Data Engineering. pp. 126–133. IEEE
(1999)
5. Chanussot, J., Lambert, P.: Total ordering based on space filling curves for multivalued morphology. Computational Imaging and Vision 12, 51–58 (1998)
6. Chen, Z., Son, S.W., Hendrix, W., Agrawal, A., Liao, W., Choudhary, A.N.: NUMARCK: machine learning algorithm for resiliency and checkpointing. In: SC 2014.
pp. 733–744 (2014)
7. Committee, I.S., et al.: 754-2008 IEEE standard for floating-point arithmetic. IEEE
Computer Society Std 2008 (2008)
8. Daubechies, I.: The wavelet transform, time-frequency localization and signal analysis. IEEE transactions on information theory 36(5), 961–1005 (1990)
9. Deutsch, L.P.: GZIP file format specification version 4.3 (1996)
10. Di, S., Cappello, F.: Fast error-bounded lossy HPC data compression with SZ. In:
2016 IEEE International Parallel and Distributed Processing Symposium, IPDPS
2016, Chicago, IL, USA, May 23-27, 2016. pp. 730–739 (2016)
11. Gleckler, P.J., Durack, P.J., Stouffer, R.J., Johnson, G.C., Forest, C.E.: Industrialera global ocean heat uptake doubles in recent decades. Nature Climate Change
(2016)
12. Habib, S., Pope, A., Finkel, H., Frontiere, N., Heitmann, K., Daniel, D., Fasel,
P., Morozov, V., Zagaris, G., Peterka, T., et al.: Hacc: Simulating sky surveys on
state-of-the-art supercomputing architectures. New Astronomy 42, 49–65 (2016)
13. Huffman, D.A., et al.: A method for the construction of minimum-redundancy
codes. Proceedings of the IRE 40(9), 1098–1101 (1952)
14. Kumar, A., Zhu, X., Tu, Y.C., Pandit, S.: Compression in molecular simulation
datasets. In: International Conference on Intelligent Science and Big Data Engineering. pp. 22–29. Springer (2013)
15. Lakshminarasimhan, S., Shah, N., Ethier, S., Ku, S., Chang, C., Klasky, S.,
Latham, R., Ross, R.B., Samatova, N.F.: ISABELA for effective in situ compression of scientific data. Concurrency and Computation: Practice and Experience
25(4), 524–540 (2013)
16. Lindstrom, P.: Fixed-rate compressed floating-point arrays. IEEE transactions on
visualization and computer graphics 20(12), 2674–2683 (2014)
17. Lindstrom, P., Isenburg, M.: Fast and efficient compression of floating-point data.
TVCG 12(5), 1245–1250 (2006)
18. Meyer, T., Ferrer-Costa, C., Pérez, A., Rueda, M., Bidon-Chanal, A., Luque, F.J.,
Laughton, C., Orozco, M.: Essential dynamics: a tool for efficient trajectory compression and management. Journal of Chemical Theory and Computation 2(2),
251–258 (Mar 2006)
19. Omeltchenko, A., Campbell, T.J., Kalia, R.K., Liu, X., Nakano, A., Vashishta,
P.: Scalable i/o of large-scale molecular dynamics simulations: A data-compression
algorithm. Computer Physics Communications 131(1), 78 – 85 (2000)
20. Ratanaworabhan, P., Ke, J., Burtscher, M.: Fast lossless compression of scientific
floating-point data. In: Data Compression Conference, 2006. DCC 2006. Proceedings. pp. 133–142. IEEE (2006)
21. Sasaki, N., Sato, K., Endo, T., Matsuoka, S.: Exploration of lossy compression
for application-level checkpoint/restart. In: Parallel and Distributed Processing
Symposium (IPDPS), 2015 IEEE International. pp. 914–922. IEEE (2015)
22. Tao, D., Di, S., Chen, Z., Cappello, F.: Significantly improving lossy compression
for scientific data sets based on multidimensional prediction and error-controlled
quantization. In: 2017 IEEE International Parallel and Distributed Processing
Symposium, IPDPS 2017, Orlando, Florida, USA, May 29-June 2, 2017. pp. 1129–
1139 (2017)
23. Yang, D.Y., Grama, A., Sarin, V.: Bounded-error compression of particle data
from hierarchical approximate methods. In: Proceedings of the 1999 ACM/IEEE
Conference on Supercomputing. SC ’99, ACM, New York, NY, USA (1999)
| 7cs.IT |
arXiv:1802.07795v1 [quant-ph] 21 Feb 2018
Communication Complexity of One-Shot Remote
State Preparation ∗
Shima Bab Hadiashar
U. Waterloo
†
Ashwin Nayak
U. Waterloo
‡
Renato Renner
ETH Zurich
§
Abstract
Quantum teleportation uses prior shared entanglement and classical communication to send an unknown quantum state from one party to another. Remote state preparation (RSP) is a similar distributed
task in which the sender knows the entire classical description of the state to be sent. (This may also be
viewed as the task of non-oblivious compression of a single sample from an ensemble of quantum states.)
We study the communication complexity of approximate remote state preparation, in which the goal is
to prepare an approximation of the desired quantum state.
Jain [Quant. Inf. & Comp., 2006] showed that the worst-case communication complexity of approximate RSP can be bounded from above in terms of the maximum possible information in an encoding.
He also showed that this quantity is a lower bound for communication complexity of (exact) remote
state preparation. In this work, we tightly characterize the worst-case and average-case communication
complexity of remote state preparation in terms of non-asymptotic information-theoretic quantities.
We also show that the average-case communication complexity of RSP can be much smaller than the
worst-case one. In the process, we show that n bits cannot be communicated with less than n transmitted
bits in LOCC protocols. This strengthens a result due to Nayak and Salzman [J. ACM, 2006] and may
be of independent interest.
1
Introduction
Quantum teleportation [3] is an archetypical protocol in information processing that is impossible in the
absence of quantum resources like shared entanglement. Through quantum teleportation, one party is able
to communicate an arbitrary qubit state to another party using only two classical bits of communication
and a previously shared maximally entangled pair of qubits. The two classical bits of communication and
a maximally entangled pair of qubits are both necessary and sufficient for the task. This is a remarkable
phenomenon, as the entire classical description of the state being communicated is potentially infinite in
length.
In Ref. [33], Lo introduced a similar distributed task in which the sender (called Alice in the literature) knows
a classical description of the quantum state. This task is called remote state preparation (RSP). In particular,
remote state preparation is a task involving two parties, Alice and Bob, who share qubits in an entangled
∗ Much
of the work in this article was reported in S.B.’s Master’s thesis [2].
of Combinatorics and Optimization, and Institute for Quantum Computing, University of Waterloo, 200
University Ave. W., Waterloo, ON, N2L 3G1, Canada. Email: sbabhadi@uwaterloo.ca . Research supported in part by
NSERC Canada.
‡ Department of Combinatorics and Optimization, and Institute for Quantum Computing, University of Waterloo, 200
University Ave. W., Waterloo, ON, N2L 3G1, Canada. Email: ashwin.nayak@uwaterloo.ca . Research supported in part by
NSERC Canada.
§ Institute for Theoretical Physics, ETH Zurich, Wolfgang-Pauli-Str. 27, 8093 Zurich, Switzerland.
Email:
renner@itp.phys.ethz.ch .
† Department
1
state. Alice is given the description of a state, Q(x), chosen from a subset of quantum states {Q(1), . . . , Q(n)},
and their goal is to prepare that quantum state on Bob’s side using only local operations and classical
communication (LOCC). This may also be viewed as the task of compression (which is non-oblivious at the
sender’s end), of a single sample from an ensemble of quantum states with entanglement-assisted classical
communication.
We say an RSP protocol is oblivious to Bob if he can get no more information about the prepared state
than what is contained in a single copy of the state [32]. A relaxed version of RSP is approximate remote state preparation (ARSP) in which we wish to prepare an approximation σx of the specified quantum state Q(x). We define the error of a protocol for approximate remote state preparation in terms of
the fidelity between Q(x)√and σx . We say a protocol has worst-case error at most ǫ, if for every x ∈
− ǫ2 . Similarly, a protocol
{1, . . . , n}, F(Q(x), σx ) ≥ 1P
√ has average-case error at most ǫ with respect to a
n
probability distribution p, if x=1 px F(Q(x), σx ) ≥ 1 − ǫ2 .
Lo [33] gave several examples of ensembles which can be remotely prepared using a one-way communication
protocol with classical communication cost less than that in quantum teleportation. However, he conjectured
that to prepare arbitrary pure n-qubit states remotely, Alice has to necessarily send the same number of
classical bits as in quantum teleportation i.e., 2n classical bits. The task has been studied extensively since
then, largely in the asymptotic setting.
Bennett et al. [4] showed that in the presence of a large amount of shared entanglement, Alice can prepare
general quantum states on Bob’s side with the asymptotic classical communication rate of one bit per qubit.
This amount of classical communication from Alice to Bob is also necessary by causality [33]. They also
showed that unlike for quantum teleportation, there is a trade-off between the communication cost and
the amount of entanglement in remote state preparation. In particular, they proved that at the cost of
using more entanglement, the communication cost of preparing a one-qubit state ranges from one bit in the
high entanglement limit to an infinite number of bits in the case of no previously shared entanglement. In
addition, they suggested that the Lo conjecture is true in a more restricted setting, such as when the protocol
is faithful and oblivious to Bob [4]. (A protocol is said to be faithful if it is exact and deterministic.)
Devetak and Berger [17] found an analytic expression for the trade-off curve between the shared entanglement
and classical communication of teleportation based RSP protocols in the low-entanglement region (less than 1
singlet state per qubit). They conjectured that teleportation based protocols are optimal among all lowentanglement protocols. Later, Leung and Shor [32] proved the Lo conjecture for a special case. They
proved that if a one-way RSP protocol for a generic ensemble of pure states is faithful and oblivious to Bob,
then it necessarily uses at least as much classical communication as in teleportation. (A generic ensemble
is an ensemble of states whose density matrices span the operators in the input Hilbert space.) Hayashi,
Hashimoto and Horibe [21] showed that in order to remotely prepare one qubit in an arbitrary state using
a one-way faithful, but not necessarily oblivious protocol, Alice requires two classical bits of communication
as in teleportation.
Berry and Sanders [7] studied ARSP, the approximation variant of RSP, of an ensemble E of mixed states
(which might be entangled with some other system on Alice’s part) such that their entanglement with
other systems does not change significantly. They showed that approximate remote state preparation with
arbitrary small average-case error ǫ can be done asymptotically using communication per prepared state
arbitrarily close to the Holevo information χ(E ) of the ensemble. (See Section 2.4 for a definition of Holevo
information.) Later Bennett, Hayden, Leung, Shor, and Winter [6] proved that approximate remote state
preparation with small worst-case error ǫ requires an asymptotic rate of one bit of classical communication
per qubit from Alice to Bob. They also showed that this amount of classical communication is sufficient.
Moreover, they derived the exact trade-off curve between shared entangled bits and classical communication
bits for an arbitrary ensemble of candidate states.
Jain [25] studied remote state preparation in the one-shot scenario. He considered the total communication
cost when given access to an arbitrary amount of entanglement. He showed that the communication cost
required for exact remote state preparation is at least T(Q)/2 and ARSP with worst-case error at most ǫ can
2
Protocol Type
Faithful RSP [4]
Faithful RSP [21]
Faithful and
oblivious RSP [32]
ARSP with small
average-case
error [7]
ARSP with small
worst-case error [6]
Exact RSP [25]
ARSP with
worst-case error
ǫ [25]
Conditions
an arbitrary state,
one-way communication,
in asymptotics
one pure qubit in a
general state,
one-way communication
a generic ensemble of
pure states,
one-way communication
an ensemble E of mixed
states preserving their
entanglement,
one-way communication,
in asymptotics
an arbitrary pure state,
two-way communication,
in asymptotics
an arbitrary state,
two-way communication,
in one-shot scenario
an arbitrary state,
one-way communication,
in one-shot scenario
Entanglement
Classical Communication
high entanglement
= 1 classical bit per qubit
= 1 ebit(singlet)
per qubit
= 2 classical bit
= 1 ebit(singlet)
per qubit
= 2 classical bit per qubit
no limit
≈ χ(E ) classical bits per
prepared state
= 1 ebit(singlet)
per qubit
= 1 classical bit per qubit
from Alice to Bob
no limit
≥ T(Q)/2
no limit
≤
(1−
√
8
2 (4T(Q)
1−ǫ2 )
+ 7)
Table 1: A summary of previous works on communication cost of Remote State Preparation
8
√
2 (4T(Q) + 7), where T(Q) denotes the maximum
(1− 1−ǫ2 )
possible information in an encoding Q. (A precise definition can be found in Section 2.4.)
be accomplished with communication at most
These abovementioned results on remote state preparation are summarized in Table 1 .
1.1
Our results
Intuitively, relaxing the remote state preparation problem so that Bob produces some approximation to the
ideal state should lower the communication complexity of the task. This suggests that the bounds provided
by Jain [25] are not tight.
In this work, we characterize the communication complexity of remote state preparation in two different
cases. First, we consider ARSP with average-case error at most ǫ, and bound its communication complexity
by the smooth max-information Bob has about Alice’s input. (See Section 2.4 for a precise definition of this
quantity.) Then we consider ARSP with worst-case error at most ǫ, and give lower and upper bounds for
its communication complexity in terms of smooth max-relative entropy and show that these bounds may be
arbitrarily tighter than that in Ref. [25].
Our main results about the remote state preparation problem are summarized below, using notions introduced in Section
√ 2. Recall that a protocol has worst-case error at most ǫ, if for every x ∈ {1, . . . , n},
F(Q(x), σx ) ≥ P1 − ǫ2 , and a protocol√has average-case error at most ǫ with respect to a probability distribution p, if nx=1 px F(Q(x), σx ) ≥ 1 − ǫ2 . We denote the average-case communication complexity of
ARSP by Q∗p (RSP(S, Q), ǫ), and the worst-case communication complexity of ARSP by Q∗ (RSP(S, Q), ǫ).
3
Theorem 1.1. For any finite set S, and set of quantum states {Q(x) : x ∈ S},
P let p be a probability
distribution over S and ρAB (p) ∈ D(H′ ⊗H) be the bipartite quantum state ρAB (p) = x∈S px |xihx|A ⊗Q(x)B .
Then
1. For any fixed ǫ ∈ (0, 1], we have
Iǫmax (A : B)ρ(p)
≤
Q∗p (RSP(S, Q), ǫ)
≤
ǫ
√
2 2
Imax
(A : B)ρ(p) + f (ǫ) ,
where f (ǫ) ∈ Θ(log log 1ǫ ) is a function of ǫ, and Iǫmax (A : B) denotes the smooth max-information part
B has about part A.
2. For any fixed ǫ ∈ (0, 1] and for any 0 < δ < 1 − ǫ2 , we have
√ 2
2(ǫ +δ)
min max Dmax
(Q(x)kσ) + g1 (ǫ, δ) ≤ Q∗ (RSP(S, Q), ǫ)
σ∈D(H) x∈S
≤
min
√
ǫ
2
1+ǫ
max Dmax
(Q(x)kσ) + g2 (ǫ) ,
σ∈D(H) x∈S
3
where g1 , g2 are functions such that g1 (ǫ, δ) ∈ Θ log ǫ2δ+δ , g2 (ǫ) ∈ Θ(log log 1ǫ ), and Dǫmax (Q(x)kσ)
denotes the smooth max-relative entropy of Q(x) with respect to σ.
It is relatively straightforward to show that the one-shot information expressions appearing in the above
theorem are continuous in ǫ. This indicates the tightness of the characterization. In fact, a bound on the
difference between lower and upper bounds in the above theorem, in terms of the ensemble, may be inferred
from the continuity property.
We remark that the quantity appearing in the second part of the theorem is similar to the notion of information radius. It may be possible to relate the quantity to smooth max-information with respect to
a distribution over S using ideas from Ref. [18, Lemma 3] (which extends Ref. [50, Lemma 14]), and the
connection between max-relative entropy and the sandwiched Rényi relative entropy. Finally, earlier works
have considered remote state preparation of states drawn from infinite sets of states. We discuss how the
bounds in Theorem 1.1 may be applied to that case in Appendix B.
The communication cost of ARSP may decrease dramatically when more error is allowed,and if we consider
average-case error instead of worst-case error. In particular, we show that for every ǫ ∈ [0, √12 ), there exists
a set of n quantum states for which there is a log n gap between the worst-case error and average-case error
remote preparation of that set. In addition, for a special set of quantum states, we derive a gap between
the worst-case error and average-case communication complexity in terms of ǫ. This confirms our intuition
that the more skewed the probability distribution is, the bigger the gap between worst-case and average-case
error variants may be.
In the process of establishing the first gap described above, we strengthen a result due to Nayak and
Salzman [35]; we prove a bound on the communication required by any LOCC protocol for transmitting a
uniformly random n bit string with some probability p. This bound is optimal, and may be of independent
interest.
Theorem 1.2. Let Y be the output of Bob in any two-way LOCC protocol in which Alice receives a uniformly
distributed n-bit input X (that is not known to Bob, and is independent of their joint quantum state). Let mA
be the total number of bits Alice sends to Bob and p := Pr[Y = X] be the probability that Bob obtains the
output X. Then
mA ≥ n + log p .
4
Worst-case protocols for ARSP capture precisely the task of compression in one-way communication complexity. Average-case protocols for ARSP are relevant in the distributional setting in communication complexity,
and in asymptotic information theory. The results in this paper thus supercede those due to Jain, Radhakrishnan, and Sen [29] (and due to Touchette [47] for the same setting). We also show how a characterization
due to Berry and Sanders [7] may be reproduced from ours, via a quantum asymptotic equipartition property
(cf. Theorem 2.6). Thus, we believe the results presented here have wider ramifications.
1.2
Organization
The organization of this paper is as follows. In Section 2, we review some concepts, fix notation, and the
terminology used in the paper. Then we define remote state preparation, and explain an efficient protocol
for this problem introduced in Ref. [29]. In Section 3 and Section 4, we give bounds on average-case error
and worst-case communication complexity of ARSP, respectively. We make some observations, including a
comparison with previously known results in Section 5. We analyze LOCC protocols for communicating a
uniformly random n bit string in Section 6. The paper ends with a summary of our results and an outlook
in Section 7. In the Appendix, we present the proofs of some properties of information-theoretic quantities,
and discuss remote state preparation of states drawn from an infinite set.
Acknowledgments
We are grateful to Matthias Christandl for discussions which led to the research reported in this article.
S.B. thanks Marco Tomamichel for his help with the proof of Theorem 4.2 during her internship at CQT,
Singapore. We also thank the reviewers and the Associate Editor, Mark Wilde, for their comments and
suggestions.
2
Preliminaries
In this section, we review some notions in quantum computing and quantum information theory, such as
LOCC protocols, quantum communication complexity, asymptotic and non-asymptotic quantum information
theory, as well as some mathematical tools like the minimax theorem. We also define remote state preparation
formally and describe a non-trivial protocol for this problem. We refer the reader to the books by Nielsen
and Chuang [37] and Watrous [49] for basic notions and results in quantum information, and largely only
describe the potentially non-standard notation and terminology we use.
2.1
Some basic notions
We denote Hilbert spaces either by capital script letters like H and K, or as Cm where m is the dimension of
the Hilbert space. We concern ourselves only with finite dimensional Hilbert spaces in this article. We denote
the set of all linear operators from H to K by L(H, K). We abbreviate L(H, H) as L(H). We denote the set
of all positive semidefinite operators in H by Pos(H). An operator A is called sub-normal if it is positive
semidefinite and has trace at most 1. (The term “subnormalized” is also often used for such operators.)
We denote the identity operator on a Hilbert space by 1 and the set of all unitary operators on space H
by U(H).
We call a physical quantum system with a finite number of degrees of freedom a register. Every register
is associated with a Hilbert space. We denote registers by capital letters, e.g., X, Y and Z. We use the
5
notation |X| to denote the dimension of the Hilbert space associated with register X. The state of a register X
is modelled as a density operator, i.e., a positive semidefinite operator with trace one, and is called a quantum
state. We denote density operators by lower case Greek letters (e.g., ρ, σ, . . . ), and the set of all density
operators over a Hilbert space H by D(H). We may also denote a state by ρX to indicate its register X.
A bipartite register XY with Hilbert space H ⊗ K is called a classical-quantum
register in the context of
P
an information processing task, if it only assumes states of the form i pi |ei ihei | ⊗ ρi where {|ei i} is the
standard basis of H and p is a probability distribution over the basis.
√ states are
P In that case we say
Pthe
√ that
classical on X. For any ω ∈ Pos(H) with spectral decomposition i λi |ψi ihψi |, we let ω = i λi |ψi ihψi |.
We denote the partial trace over Hilbert space K of a quantum state ρAB ∈ D(H ⊗ K) by either TrK (ρAB )
or TrB (ρAB ). We say that ρAB ∈ D(H ⊗ K) is an extension of ρA ∈ D(H) if TrK (ρAB ) = ρA .
We call completely positive and trace preserving linear maps √
L(H) → L(K) quantum channels. Quantum
measurements are quantum channels with Kraus operators { Ea ⊗ |ai : a ∈ Γ}, where Γ is the set of
outcomes of
P the measurement and Ea is a positive semidefinite operator associated with the outcome a ∈ Γ
such that a∈Γ Ea = 1. We refer to the operators Ea as measurement operators.
The fidelity F(ρ, σ) between two quantum states ρ and σ, is defined as
q
√ √
F(ρ, σ) := Tr
ρσ ρ .
In the literature, fidelity is sometimes defined as the square of the above quantity. Fidelity may be extended
to sub-normal states ρ, σ as follows:
q
p
√ √
F(ρ, σ) := Tr
ρ σ ρ + (1 − Tr(ρ)) (1 − Tr(σ)) .
The fidelity function is monotone under the application of quantum channels, and is jointly concave over the
set of quantum states. Other useful properties of fidelity are stated in the following propositions.
Proposition 2.1. For any quantum state ρ and sub-normal state σ, it holds that
F(ρ, σ)2
≤
Tr(σ) .
Proposition 2.2. Let ρ, σ ∈ D(H) be two quantum states. Then
1 + F(ρ, σ)
=
max {F(ρ, ξ)2 + F(σ, ξ)2 : ξ ∈ D(H)} .
For a proof of the above property, see Ref. [36, Lemma 3.3].
We use the purified distance (see Ref. [45]) as a metric for sub-normal states. This is an extension of the
metrics developed in Refs. [40, 41, 20, 42]. Suppose that ρ and σ are two sub-normal states. Then the
purified distance of ρ and σ is defined as
p
P(ρ, σ) :=
1 − F(ρ, σ)2 .
There are other metrics over sub-normal states, such as the trace distance. However, we choose purified
distance since it turns out to be more convenient to use in non-asymptotic quantum information theory.
Let ρ ∈ D(H) be a quantum state and ǫ ∈ [0, 1). Then, we define
Bǫ (ρ)
:=
{ρ̃ ∈ Pos(H) : P(ρ, ρ̃) ≤ ǫ, Tr ρ̃ ≤ 1}
as the ball of sub-normal states that are within purified distance ǫ of ρ. We say that σ is ǫ-close to ρ,
or equivalently, σ is an ǫ-approximation of ρ, if σ ∈ Bǫ (ρ). The following property of purified distance
states that any state ρ′A that is ǫ-close to ρA may be extended to a state ρ′AB that is ǫ-close to any given
extension ρAB of ρA .
6
Proposition 2.3. Let ρA ∈ D(HA ) be a quantum state in the Hilbert space HA and ρAB ∈ D(HA ⊗HB ) be an
extension of ρA over the Hilbert space HA ⊗HB , i.e. ρA = TrB (ρAB ). Let ρ′A ∈ Bǫ (ρA ) be an ǫ-approximation
of ρA . Then there exists ρ′AB ∈ Bǫ (ρAB ) such that ρ′A = TrB (ρ′AB ).
Proof: Let |vi ∈ D(HA′ ⊗ HB ′ ⊗ HA ⊗ HB ) be a purification of ρAB and therefore also of ρA , and |v ′ i ∈
D(HA′ ⊗ HB ′ ⊗ HA ⊗ HB ) be a purification of ρ′A , such that F(ρA , ρ′A ) = |hv|v ′ i|. Such |vi and |v ′ i exist
by the Uhlmann theorem. Define ρ′AB = TrA′ B ′ (|v ′ ihv ′ |). By definition, we have F(ρA , ρ′A ) = F(ρAB , ρ′AB ).
Therefore ρ′AB ∈ Bǫ (ρAB ).
The above property is in fact an extension of the Uhlmann theorem for purified distance.
2.2
LOCC protocols
The notion of LOCC, short for local operations and classical communication, plays an important role in
quantum information, especially in the study of properties of entanglement (see, e.g., Ref. [5]). This notion
has been described formally in terms of quantum instruments in Ref. [14]. In this article, we only study
two-party LOCC protocols, in which one party receives a classical input, and the other party produces a
quantum output. We describe these protocols informally below.
Suppose we have two parties, Alice and Bob, who communicate with each other using only classical bits,
share parts of a possibly entangled quantum state, and are allowed to perform any local quantum channels
on their registers. We call the registers (or qubits) accessible by only one of the parties private registers
(or qubits). Alice is given a classical input; Bob does not receive any input. Let A be the register which
holds Alice’s input, Y0 := P0 V0 and Z0 := Q0 W0 be Alice’s and Bob’s initial classical-quantum private
registers, respectively. Registers Pi and Qi are classical registers with Alice and Bob, respectively, after
the ith message. These registers hold the message transcript thus far. Initially, P0 , Q0 are both empty.
Registers V0 and W0 are initialized to a quantum state independent of the inputs. Note that the state
in V0 W0 might be entangled across the registers. If there are k messages, Pk+1 and Vk+1 denote Alice’s
final classical and quantum registers, respectively, and Qk+1 Wk+1 denote Bob’s, potentially after a local
operation. Register A remains unchanged throughout the protocol. Bob produces the output, which is a
sub-register B of Qk+1 Wk+1 .
A one-way LOCC protocol is an LOCC protocol in which the communication consists of one message from
Alice to Bob. The three steps of the protocol are:
1) Alice measures her register V0 , obtains the outcome in register P1 (and a residual state in V1 ). The
measurement is controlled by her input in A.
2) Alice sends a copy of her measurement outcome to Bob, in classical register M . Bob sets Q1 = M .
3) Bob measures his register W1 (which is the same as W0 ), controlled by the register Q1 . The outcome
and residual state are stored in classical-quantum registers Q2 W2 , where Q2 includes Q1 . The output
of the protocol is a designated sub-register B of his registers Q2 W2
A two-way LOCC protocol is a protocol with communication in both directions, from Alice to Bob and Bob
to Alice. It has several rounds of communication in which the two parties alternately do a local measurement
and send a message. Either party may start or end the protocol. Suppose in round i, it is Alice’s turn. Then
• First, Alice measures her quantum register in that round, Vi−1 , controlled by her input A and her
classical register Pi−1 . She copies the outcome Mi in a fresh register Ni . The register Pi := Pi−1 Ni .
7
• Alice then sends Mi to Bob using mi classical bits, and Bob includes the received message Mi in his
transcript register: Qi := Qi−1 Mi .
Bob’s actions are similar in a round in which it is his turn (except that he does not have any input), using
registers Qi Wi . At the end of a protocol with k rounds of communication, Bob makes a measurement on
the quantum register Wk controlled by Qk , and he includes the outcome Mk+1 of the measurement in the
register Qk+1 . A pre-designated sub-register B of Qk+1 Wk+1 is the output of the protocol.
2.3
Quantum communication complexity
Quantum communication complexity was introduced by Yao [51], and has been studied extensively since.
Here we describe it in the context of LOCC protocols.
Let X, Y be two finite sets, Z be a set (not necessarily finite), and f ⊆ X × Y × Z be a relation such that
for every (x, y) ∈ X × Y , there exists some z ∈ Z such that (x, y, z) ∈ f . The sets X, Y, Z might be sets of
quantum states. For example, in remote state preparation Z is the set of quantum states over some space.
In an LOCC protocol, Alice and Bob get as their inputs x ∈ X and y ∈ Y , respectively, and their goal is
to output an element z ∈ Z such that (x, y, z) ∈ f . In the protocols we consider, one party may not get
any input, e.g., Y may be empty. Also, in general the output of the protocol is probabilistic. If Wx,y is the
random output that the protocol produces on inputs (x, y), we define the error of the protocol as
δ
:=
max Pr((x, y, Wx,y ) 6∈ f ) .
x∈X,y∈Y
We then say the protocol computes f with error δ.
Definition 2.1. The entanglement-assisted communication complexity of f with error δ is defined as the
minimum number of bits exchanged in an LOCC protocol computing f with error δ.
Now consider a relation f ⊆ X × Y × Z, with Z = D(H), the set of quantum states over H. In this
context we may allow a protocol to produce an approximation to the desired quantum state. Suppose the
output quantum state that an LOCC protocol for f produces on inputs (x, y) is denoted by wxy . Let p be a
probability distribution over X × Y . We say a protocol computes an approximation of f with average-case
error at most ǫ if there are quantum states {zxy : x ∈ X, y ∈ Y, (x, y, zxy ) ∈ f } such that
p
X
pxy F(wxy , zxy ) ≥
1 − ǫ2 .
x∈X,y∈Y
P
The above condition may equivalently
P be written as P(ζ, ω) ≤ ǫ, where ζ := x,y pxy |xyihxy| ⊗ zxy is an
ideal input-output state, and ω := x,y pxy |xyihxy| ⊗ wxy is the actual input-output state of the protocol.
Definition 2.2. The average-case communication complexity of f is defined as the minimum number of
bits exchanged in an LOCC protocol computing an approximation of f with average-case error at most ǫ,
and is denoted by Q∗p (f, ǫ).
Similarly, we say a protocol computes an approximation of f with worst-case error at most ǫ if there are
quantum states {zxy : x ∈ X, y ∈ Y, (x, y, zxy ) ∈ f } such that
max P(wxy , zxy )
x∈X,y∈Y
≤
ǫ .
Definition 2.3. The worst-case communication complexity of f is defined as the minimum number of bits
exchanged in an LOCC protocol computing an approximation of f with worst-case error at most ǫ, and is
denoted by Q∗ (f, ǫ).
Note that “error” here refers to the quality of approximation in the output state. The result of any probabilistic error made by the protocol is included in the output state, and hence this kind of error is reflected
in the quality of approximation.
8
2.4
Quantum information theory
Let X be a register in quantum state ρ ∈ D(H). Then the von Neumann entropy S(ρ) of X is defined as
:=
S(ρ)
−Tr(ρ log ρ) .
Let X and Y be two registers in quantum states ρX ∈ D(H) and σY ∈ D(H), respectively. The relative
entropy denoted by S(ρX kσY ) is defined as
:=
S(ρX kσY )
Tr (ρX log ρX − ρX log σY )
if supp(ρ) ⊆ supp(σ), and as ∞ otherwise. Suppose that ρXY ∈ D(H ⊗ K) is the joint state of registers X
and Y , then the mutual information of X and Y is defined as
:=
I(X : Y )ρ
S(ρX ) + S(ρY ) − S(ρXY ) ,
where ρX = TrY (ρXY ) and ρY = TrX (ρXY ). When the register whose state is ρ is clear from the context,
we may omit it from the subscript of ρ. Similarly, when the state ρ of the registers XY is clear from the
context, we may omit it from the subscript of I(X : Y ).
For ρ, σ ∈ D(H), the observational divergence [28] between ρ and σ is defined as
Tr(M ρ)
Dobs (ρkσ) := sup Tr(M ρ) log
: 0 ≤ M ≤ 1, Tr(M σ) 6= 0 .
Tr(M σ)
P
Let E = (pj , ρj ) : 1 ≤ j ≤ n be an ensemble of quantum states, i.e., 0 ≤ pj ≤ 1 for 1 ≤ j ≤ n, nj=1 pj = 1,
and ρj ∈ D(H) are quantum states over the same space. The Holevo information of E , denoted as χ(E ), is
defined as
n
X
pj S(ρj kρ) ,
χ(E ) :=
j=1
where ρ is the ensemble average, i.e., ρ =
denoted as Dobs (E ), as
Pn
Dobs (E )
j=1
pj ρj . Similarly, we define the divergence information of E ,
:=
n
X
j=1
pj Dobs (ρj kρ) .
Let S be a set, and Q : S → D(H) be a function which “encodes” each x ∈ S asP
a quantum state. Let p be
a probability distribution over S, and ρAB (p) be the bipartite state ρAB (p) := x px |xihx|A ⊗ Q(x)B . We
define the maximum possible information in Q [25], denoted by T(Q), as
T(Q)
:=
max I(A : B)ρ(p) ,
p
where the maximum is taken over all probability distributions p over S.
P
Note that for a classical-quantum state ρAB = nj=1 pj |jihj| ⊗ ρj , the mutual information of A and B is
equal to the Holevo information of the quantum ensemble E = (pj , ρj ) : 1 ≤ j ≤ n , i.e., χ(E ) = I(A : B),
and therefore T(Q) ≥ χ(E ).
Most of the entropic quantities defined above arise naturally in the analysis of information processing tasks
in the asymptotic setting, i.e., when the available resources may be used to jointly complete arbitrarily
long sequences of tasks on independent, identically distributed (iid) inputs. The asymptotic setting is an
idealization that may not be realistic in certain scenarios. More often, we are faced with single instances of
a task which we wish to accomplish with the fewest resources. Recently, researchers have begun to formally
study tasks in the non-iid or one-shot setting, and the entropic notions that arise therein. Several one-shot
9
entropic concepts have been implicit in traditional (iid) information theory and in communication complexity.
For example, Jain, Radhakrishnan, and Sen implicitly studied the concept of smooth max-relative entropy in
Ref. [28]. However, non-asymptotic concepts were formalized only later (see, e.g., Refs. [43, 44, 16]). In this
work, we use one-shot entropic quantities to tightly characterize the communication complexity of remote
state preparation.
Let ρ, σ ∈ D(H) be two quantum states. The max-relative entropy of ρ with respect to σ is defined as
Dmax (ρkσ)
min{λ : ρ ≤ 2λ σ} ,
:=
when supp(ρ) ⊆ supp(σ), and is ∞ otherwise [16]. This notion captures how two states ρ, σ behave relative
to each other under the application of a measurement. For a bipartite quantum state ρAB ∈ D(H′ ⊗ H), the
max-information part B has about part A [9] is defined as
Imax (A : B)ρ
:=
min Dmax (ρAB k ρA ⊗ σB ) .
σ∈D(H)
Note that this quantity is asymmetric with respect to the parts A and B. As for mutual information,
we include the state as a subscript only when it is not clear from the context. The smoothed versions of
these quantities come into play when approximations are allowed in the tasks at hand. Smooth max-relative
entropy is defined as
Dǫmax (ρkσ) :=
min
Dmax (ρ̃kσ) ,
ǫ
ρ̃∈B (ρ)
and smooth max-information is defined as
Iǫmax (A : B)ρ
:=
min Imax (A : B)ρ̃ .
ρ̃∈Bǫ (ρ)
There are several ways to define max-information using max-relative entropy [15]. We choose the above
definition in this work since it can be used to characterize average-case communication complexity of the
remote state preparation problem.
The following are some properties of max-information we use. Both the exact and smooth versions of this
quantity are monotonic under the application of a quantum channel [9].
Proposition 2.4 (Monotonicity under quantum channels). Let Φ : L(H′ ) → L(K) be a quantum channel,
ρAB a bipartite sub-normal state over H′ ⊗ H, σAB ∈ D(H′ ⊗ H) a bipartite quantum state, and ǫ ∈ [0, 1].
Then
Imax (A′ : B)ρ′
Iǫmax (A′
: B)σ′
≤
≤
Imax (A : B)ρ ,
Iǫmax (A
and
: B)σ ,
′
where A′ , B denote two parts of the states ρ′A′B := (Φ ⊗ 1)(ρ) and σA
′B := (Φ ⊗ 1)(σ).
For a classical-quantum state ρAB , the value of smooth max-information is achieved by a classical-quantum
state ρ′AB that is ǫ-close to ρAB . A proof is included in Appendix A.
Proposition 2.5. Let ρAB ∈ D(H′ ⊗ H) be a bipartite quantum state that is classical on A. For any ǫ ≥ 0,
there exists ρ′AB ∈ Bǫ (ρAB ) ∩ D(H′ ⊗ H) classical on A such that
Iǫmax (A : B)ρ
=
Imax (A : B)ρ′ .
Smooth max-information satisfies the Asymptotic Equipartition Property, as proven by Berta, Christandl,
and Renner [9]. Let H denote the binary entropy function H(α) := −α log α − (1 − α) log(1 − α).
10
Theorem 2.6 (Quantum Asymptotic Equipartition property). Let ǫ > 0, n an integer such that n ≥
2(1 − ǫ2 ), and ρAB ∈ D(HAB ). Then
I(A : B)ρ −
3
H(ǫ) − 2ǫ log(|A| |B|)
n
≤
1 ǫ
I
(A : B)ρ⊗n ,
n max
(2.1)
and
1 ǫ
I
(A : B)ρ⊗n
n max
√
where ξ(ǫ) = 8 13 − 4 log ǫ (2 +
1
2
2
ǫ2
ξ(ǫ)
,
I(A : B)ρ + √ − log
n
n
24
≤
(2.2)
log |A|). Therefore,
lim lim
ǫ→0 n→∞
1 ǫ
I
(A : B)ρ⊗n
n max
=
I(A : B)ρ .
For ǫ ∈ [0, 1), the ǫ-hypothesis testing relative entropy [48] of two quantum states ρ, σ ∈ D(H) is defined as
Dǫh (ρkσ)
:=
− log
βǫ (ρkσ)
,
1−ǫ
where
βǫ (ρkσ)
:=
inf {hQ, σi | 0 ≤ Q ≤ 1 and hQ, ρi ≥ 1 − ǫ} .
(2.3)
The infimum in the above definition is always achieved and βǫ (ρkσ) is between 0 and 1. In this definition,
we interpret (Q, 1 − Q) as a measurement for distinguishing ρ from σ, i.e., as a strategy in hypothesis testing.
So βǫ (ρkσ) corresponds to the minimum probability of incorrectly identifying σ when ρ is identified correctly
with probability at least 1 − ǫ. This one-shot entropic quantity has been studied for a long time either
implicitly (see, e.g., Refs. [24, 38]) or explicitly, albeit without giving it a name (see, e.g., Refs. [13, 12]). It
also arises in the context of channel coding [23, 48] and other tasks [22].
The error in hypothesis testing may only increase under the action of a quantum channel. This has been
known for some time; see, e.g., Ref. [10, Eq. (44)] for a proof.
Proposition 2.7 (Data Processing Inequality). Let ρ, σ ∈ D(H) for some Hilbert space H, and Φ : L(H) →
L(K) be a quantum channel. Then
βǫ (ρkσ)
≤
βǫ (Φ(ρ) k Φ(σ)) .
The following two properties have been proved implicitly by Matthews and Wehner [34]. For completeness,
we include their proofs in Appendix A.
Hypothesis testing error satisfies a restricted form of joint convexity in its two arguments.
Proposition 2.8. Let ρAB (p) ∈ D(HA ⊗ HB ) be a state classical on A such that the distribution on A is
given by the probability vector p. Let ρA (p) = TrB (ρAB (p)), and σ ∈ D(HB ) be a quantum state on Hilbert
space HB . Then the function βǫ (ρAB (p) k ρA (p) ⊗ σ) is convex with respect to p.
Hypothesis testing error is concave in its second argument.
Proposition 2.9. For any fixed quantum state ρ ∈ D(H), the function βǫ (ρkσ) is a concave function with
respect to σ.
It turns out that hypothesis testing relative entropy is closely related to smooth max-relative entropy, as
captured by the following theorem.
11
Theorem 2.10 ([19, 46]). Let ρ, σ ∈ D(H) be two quantum states in Hilbert space H. For any ǫ ∈ (0, 1)
and δ ∈ (0, ǫ), the following inequalities hold:
√
2(1−ǫ)
Dmax
(ρkσ) ≤ Dǫh (ρkσ) ,
and
(2.4)
√
1−ǫ
Dmax
(ρkσ)
2.5
≥
Dǫ−δ
h (ρkσ) − log
ǫ(1 − ǫ + δ)
− 3 log 3 .
δ3
(2.5)
The minimax theorem
The minimax theorem is a powerful result that provides conditions under which switching the order of
minimization and maximization in certain optimization problems does not change the optimum.
Theorem 2.11 ([39]). Let n be a positive integer, and A1 , A2 be non-empty, convex and compact subsets
of Rn . Let f : A1 × A2 → R be a continuous function such that
1. ∀a2 ∈ A2 , the set {a1 ∈ A1 : (∀a′1 ∈ A1 ) f (a1 , a2 ) ≥ f (a′1 , a2 )} is convex.
2. ∀a1 ∈ A1 , the set {a2 ∈ A2 : (∀a′2 ∈ A2 ) f (a1 , a2 ) ≤ f (a1 , a′2 )} is convex.
Then
max min f (a1 , a2 )
a1 ∈A1 a2 ∈A2
2.6
=
min max f (a1 , a2 ) .
a2 ∈A2 a1 ∈A1
Remote state preparation
Let S be a finite, non-empty set, and let Q : S → D(H) be a function that maps each element x ∈ S to a
quantum state Q(x) over the Hilbert space H. Recall that remote state preparation, denoted as RSP(S, Q),
is a communication task in which one party, Alice, is given an input x ∈ S, and engages in an LOCC protocol
with another party, Bob, so that Bob is able to prepare Q(x). The function Q is known to both parties. In
the approximate remote state preparation, we allow Bob to prepare an approximation σx ∈ D(H) to Q(x).
We consider two notions of error in approximation: worst case and average case. Let ǫ ∈ [0, 1], and let p be
a probability distribution on S. We say a protocol for RSP(S, Q) makes worst-case error ǫ if P(σx , Q(x)) ≤ ǫ
for each x ∈ S. We say a protocol for RSP(S, Q) makes average-case error ǫ w.r.t. the distribution p over S
if the purified distance between the ideal and actual joint input-output states is at most ǫ. By the definitions
of purified distance and fidelity, this condition is equivalent to
p
X
px F(σx , Q(x)) ≥
1 − ǫ2 .
x∈S
In Sections 3 and 4, we characterize the communication complexity of this problem for the two different
kinds of approximation. We emphasize that Alice and Bob communicate with a noiseless classical channel,
they have access to an arbitrarily large amount of entanglement of their choice, and they have unlimited
computational power.
A straightforward protocol for approximate remote state preparation is as follows. Alice sends her input x
directly to Bob and Bob creates the desired state Q(x). Thus Bob prepares the target state with zero error
(ǫ = 0) using ⌈log(n + 1)⌉ bits of classical communication, where n = |S|.
Jain, Radhakrishnan, and Sen [28, 29] proposed the following, potentially more efficient protocol, which we
call the JRS protocol in the sequel. Let K be a Hilbert space with dim(K) ≥ dim(H) and {σx }x∈S ⊆ D(H)
12
be a set of quantum states such that for all x ∈ S, P(σx , Q(x)) ≤ δ for some δ ∈ [0, 1]. Suppose that for
some λ ∈ [0, ∞) and some σ ∈ D(H), we have
σx
≤
2λ σ
=
2−λ σx + (1 − 2−λ )ξx ,
for all x ∈ S .
(2.6)
This can be rewritten for a fixed x ∈ S as
σ
where ξx ∈ D(H) is a quantum state. Let |vx i ∈ K ⊗ H be a purification of σx in the Hilbert space K ⊗ H,
and |ux i ∈ K ⊗ H be a purification of ξx . Then
p
√
|wx i =
2−λ |0i|vx i +
1 − 2−λ |1i|ux i ,
is a purification of σ. Let |wi be an arbitrary but fixed purification of σ in C2 ⊗ K ⊗ H. By the unitary
equivalence of purifications, there is a unitary operation Ux on the space C2 ⊗K which transforms |wi to |wx i.
We are ready to describe the JRS protocol.
JRS Protocol: Alice and Bob agree on a parameter t, that depends on the quality of approximation they
desire. Initially, Alice and Bob share t copies of the quantum state |wi. The registers corresponding to
Hilbert spaces C2 and K in the ith copy of |wi are called Ci and Ki , respectively, and are held by Alice. The
register corresponding to the Hilbert space H is called Hi and is held by Bob.
1. On getting input x, Alice performs the unitary operation Ux on registers Ci Ki for each i ∈ [t]. This
transforms all copies of |wi to copies of |wx i. Then she measures the register Ci for all i ∈ [t]. If at
least one of the measurement outcomes, say the jth, is equal to zero, she sends the index j to Bob,
using ⌈log(t + 1)⌉ bits. (She may choose to send any such index.) Otherwise, if the outcomes of all t
measurements are equal to one, she sends 0 to Bob.
2. On receiving an integer k, where 0 ≤ k ≤ t, Bob outputs the state in register Hk if k ∈ [t], and outputs
the maximally mixed state over H if k = 0.
The output of this protocol is
the output state is
1
|H|
with probability 1 − 2−λ
t
and σx with the remaining probability. Hence,
1
.
|H|
By choosing the approximation parameter δ small enough and t large enough, Bob produces a state σ̃x with
the desired accuracy. We use this protocol to give upper bounds on the worst-case error and average-case
communication complexity of RSP(S, Q).
σ̃x
3
=
1 − 1 − 2−λ
t
σx + 1 − 2−λ
t
Average-case communication complexity
Let p be a probability distribution over S and Q∗p (RSP(S, Q), ǫ) denote the average-case entanglementassisted communication complexity of approximate remote state preparation (ARSP), with respect to p,
and with (average) error at most ǫ. We characterize this quantity in terms of smooth max-information, a
one-shot analogue of mutual information.
3.1
An upper bound
First, we show that the average-case communication complexity with error ǫ of ARSP is bounded above
essentially by Iδmax (A : B)ρ(p) , where ρ(p) is the ideal joint state of Alice’s input and Bob’s output, and δ ∈
Θ(ǫ). To do so, we use the JRS protocol described in Section 2.6.
13
Theorem 3.1. For any finite set S, function Q : S → D(H), and ǫ ∈ (0, 1], let p beP
a probability distribution
over S and ρAB (p) ∈ D(H′ ⊗ H) be the bipartite classical-quantum state ρAB (p) = x∈S px |xihx|A ⊗ Q(x)B .
Then
8
Q∗p (RSP(S, Q), ǫ) ≤ Iδmax (A : B)ρ(p) + log2 ln 2 + 2 ,
ǫ
√
where δ = ǫ/2 2.
Proof: Fix some ǫ ∈ (0, 1], and let λ be equal to Iδmax (A : B)ρ(p) with δ as in the statement of the
′
δ
′
theorem. By Proposition 2.5,
P there exist quantum
Pstates ρAB ∈ Bx (ρAB ) and σB ∈′ D(H)
P such that ρAB ≤
λ ′
′
x
2 ρA ⊗ σB ,where ρAB = x qx |xihx| ⊗ σB with x qx = 1 and σB ∈ D(H), and ρA = x qx |xihx|. Then
x
σB
2λ σB ,
≤
(3.1)
x
for all x ∈ S with qx 6= 0. For each x ∈ S with qx = 0, we assume, w.l.o.g., that σB
= σB . Inequality (3.1)
is in the form of inequality (2.6) and therefore we may execute the JRS protocol with a suitable choice of
parameter t. Initially, Alice and Bob share t copies of entangled state |wi, where |wi is a purification of σB .
x
Alice gets input x with probability px . They perform the protocol for approximating state σB
from σB . The
final joint state of Alice’s input and Bob’s output is
X
x
ρ̃AB =
,
px |xihx| ⊗ σ̃B
x∈S
where
x
σ̃B
=
Therefore,
F(ρ̃AB , ρ′AB )
=
F
1 − 1 − 2−λ
X
x∈S
t
px |xihx| ⊗
−λ t
≥
1 − (1 − 2
=
1 − (1 − 2−λ )t
≥
−λ t
1 − (1 − 2
)
x
σB
+ 1 − 2−λ
x
σ̃B
,
X
x∈S
F
X
x∈S
t
qx |xihx| ⊗
px |xihx| ⊗
X√
px qx
1
.
dim(H)
x
σB
x
σB
,
!
X
x∈S
qx |xihx| ⊗
x
σB
!
x∈S
p
)
1 − δ2 ,
where the first inequality follows from the joint concavity of fidelity. The last inequality follows from
monotonicity under quantum channels:
X√
px qx = F(ρ′A , ρA ) ≥ F(ρ′AB , ρAB ) .
x∈S
In addition, by Proposition 2.2,
F(ρ̃AB , ρAB )
≥
≥
≥
F(ρ̃AB , ρ′AB )2 + F(ρAB , ρ′AB )2 − 1
2
1 − (1 − 2−λ )t (1 − δ 2 ) + (1 − δ 2 ) − 1
p
1 − ǫ2 ,
√
where the last inequality is derived using inequalities ln(1 − x) ≤ −x and 1 − x ≤ 1 − x2 , which hold for x ∈
√
λ
P
[0, 1), and the parameter values δ = ǫ/2 2 and t = 2 ln ǫ82 . Since F(ρ̃AB , ρAB ) = x∈S px F(σ̃x , Q(x)),
the protocol has average-case error at most ǫ.
14
The communication cost of this protocol is ⌈log(t + 1)⌉. So the communication complexity of approximate
remote state preparation with average-case error ǫ is
Q∗p (RSP(S, Q), ǫ)
≤
⌈log(t + 1)⌉
≤
λ + log2 ln
8
+2 ,
ǫ2
as required.
We have not attempted to optimize the upper bound derived above. It is possible that the parameter δ and
the ǫ-dependent additive term be improved further.
3.2
A lower bound
Next, we show that the average-case communication complexity of any protocol for approximate remote
state preparation is bounded from below by Iǫmax (A : B)ρ(p) . In order to do this, we strengthen a property of
smooth max-information due to Berta, Christandl, and Renner [9, Lemma B.12], in the case of a tripartite
state ρMAB that is classical on M .
Lemma 3.2. Let ǫ ≥ 0 and ρMAB ∈ D(M ⊗ H′ ⊗ H) be any tripartite quantum state over registers M , A
and B such that ρ is classical on M . Then
Iǫmax (A : MB)
Iǫmax (A : B) + log |M | .
≤
Proof: Fix σB ∈ D(H) and ρ̃AB ∈ Bǫ (ρAB ) such that Iǫmax (A : B) = Dmax (ρ̃AB k ρ̃A ⊗ σB ). Let λ denote this
max-relative entropy, i.e., λ is the minimum non-negative real number for which ρ̃AB ≤ 2λ ρ̃A ⊗ σB . Then
1
⊗ ρ̃AB
|M |
2λ
≤
1
⊗ ρ̃A ⊗ σB .
|M |
(3.2)
By Proposition 2.3, there exists some extension ρ′MAB of ρ̃AB such that ρ′MAB ∈ Bǫ (ρMAB ). By construction,
we have TrM (ρ′MAB ) = ρ̃AB . Consider the quantum-to-classical channel Φ : L(M) → L(M) defined by
X
hei |X|ei i|ei ihei |
Φ(X) =
i
for all X ∈ L(M), where {|ei i} is the standard orthonormal basis for the Hilbert space M. The state (Φ ⊗
1)(ρ′MAB ) is classical on M , and is an extension of ρ̃AB . Define ρ̃MAB := (Φ ⊗ 1)(ρ′MAB ). Since ρ′MAB ∈
Bǫ (ρMAB ), by monotonicity of fidelity under quantum channels and because ρMAB is classical on M , we
have ρ̃MAB ∈ Bǫ (ρMAB ). So ρ̃MAB may be written as
X
i
γi |ei ihei | ⊗ σAB
,
ρ̃MAB =
i
i
σAB
where all
are normalized and
tion (3.2), we can conclude that
P
i
γi ≤ 1. We have ρ̃MAB ≤ 1M ⊗ ρ̃AB . Combining this with Equa-
ρ̃MAB
≤
λ
2 |M |
1M
⊗ ρ̃A ⊗ σB
|M |
and consequently,
Dmax ρ̃MAB
1M
⊗ ρ̃A ⊗ σB
|M |
≤
λ + log |M | .
By the definition of smooth max-information, this implies that
Iǫmax (A : MB)
≤
15
λ + log |M | ,
as required.
Remark: The above lemma could alternatively be derived from an analogous inequality for α-Rényi mutual
information [31, Equation (2.25)]). Taking the limit as α → ∞ gives us the inequality for max-information
(i.e., for ǫ = 0). We may extend this to any ǫ ≥ 0 by smoothing arguments similar to those in the above
proof.
Using this lemma, we bound the average-case communication complexity of RSP(S, Q) from below.
Theorem 3.3. For any finite set S, function Q : S → D(H), and probability distribution p over S, let ρ(p)
be the bipartite quantum state
X
ρ(p) =
px |xihx|A ⊗ Q(x)B .
x∈S
For any ǫ ∈ [0, 1], we have
Q∗p (RSP(S, Q), ǫ)
≥
Iǫmax (A : B)ρ(p) .
Proof: In this proof we follow the notation and convention described in Section 2.2. Consider a k-round
LOCC protocol Π for RSP(S, Q) with average-case error ǫ. Suppose Bob sends the first message, and the
joint state in Alice and Bob’s registers (excluding the input register A) after the message is φ. As Bob receives
no input, the joint state φ is known to both parties. Hence, the rest of the protocol can be considered as
a new LOCC protocol, with the same output, in which the initial shared state of parties is φ, and Alice
starts the protocol. The communication cost of this new protocol is less than the communication cost of the
original one. Therefore, it suffices to show the lower bound for protocols in which Alice starts.
Let A be Alice’s input register, and Yi := Pi Vi and Zi := Qi Wi be Alice’s and Bob’s classical-quantum private
registers, respectively, after the ith round of the protocol for i ≥ 0. Initially, A and Z0 are independent, and
so
Imax (A : Z0 )
=
0 .
(3.3)
Consider the ith round of a two-way LOCC protocol. The communication in each round is either from Alice
to Bob (for odd i) or from Bob to Alice (for even i).
Odd round i: In this case, Alice measures her private qubits Vi−1 controlled by Pi−1 and A. She includes
the outcome of her measurement Mi in the register Pi (recall that Pi = Pi−1 Mi ), and sends a copy of Mi to
Bob using mi := ⌈log(|Mi | + 1)⌉ bits of communication. Then Bob includes the received message Mi in Qi
(recall that Qi = Qi−1 Mi ). Thus,
Imax (A : Zi )
≤ Imax (A : Zi−1 ) + log |Mi |
(by Lemma 3.2)
≤ Imax (A : Zi−1 ) + mi .
(3.4)
Even round i: In this case, Bob measures his private qubits Wi−1 controlled by Qi−1 . He includes the
outcome of his measurement Mi in Qi , and sends a copy of Mi to Alice using mi = ⌈log(|Mi | + 1)⌉ bits of
communication. Alice includes the received message in Pi . Thus,
Imax (A : Zi )
≤ Imax (A : Zi−1 ) .
(by Proposition 2.4)
Combining Eqs. (3.4) and (3.5) recursively, we get
Imax (A : Zk )
≤
Imax (A : Z0 ) +
X
1≤i≤k
i odd
16
mi
=
X
1≤i≤k
i odd
mi ,
(3.5)
P
after k rounds of communication. Let m := 1≤i≤k, odd mi . At the end of the protocol, Bob applies a quantum channel on his register Zk to get the output B. By monotonicity of max-information (Proposition 2.4),
we have
Imax (A : B)ρ′ (p) ≤ m ,
P
where ρ′ (p) = x px |xihx| ⊗ σx is the bipartite quantum state of registers AB, and m is the number of bits of
communication from Alice to Bob. In addition, protocol Π guarantees that ρ′ (p) is within purified distance ǫ
of ρ(p). Therefore, we conclude the theorem.
4
Worst-case communication complexity
In this section, we characterize the worst-case communication complexity of remote state preparation, denoted as Q∗ (RSP(S, Q), ǫ), in terms of smooth max-relative entropy.
4.1
An upper bound
We show that for some fixed ǫ ∈ (0, 1], the worst-case communication complexity of the approximate remote
state preparation problem is bounded from above essentially by
min max Dδmax (Q(x)kσ) ,
σ∈D(H) x∈S
where δ ∈ Θ(ǫ). As for the average case, we utilize the JRS protocol presented in Section 2.6.
Theorem 4.1. Let S be a non-empty finite set, Q : S → D(H) be a function from S to the set of density
operators in the Hilbert space H, and ǫ ∈ [0, 1]. Then
Q∗ (RSP(S, Q), ǫ)
where δ =
≤
min max Dδmax (Q(x)kσ) + log2 (1 + ǫ2 ) + log2 ln
σ∈D(H) x∈S
2
+2 ,
ǫ4
√ ǫ
.
1+ǫ2
Proof: Let α := minσ∈D(H) maxx∈S Dδmax (Q(x)kσ) and σ ′ be the quantum state for which the minimum is
achieved, i.e., α = maxx∈S Dδmax (Q(x)kσ ′ ). By definition, for all x ∈ S there exists some σx ∈ Bδ (Q(x)) such
that
σ ′ ≥ 2−α σx .
1
Since P(σx , Q(x)) ≤ δ, we have F(σx , Q(x))2 ≥ 1 − δ 2 . So, by Proposition 2.1, Tr(σx ) ≥ 1 − δ 2 = 1+ǫ
2 for
σx
:=
all x ∈ S. For each x ∈ S, define ρx
Tr(σx ) . Then for all x ∈ S, ρx is a quantum state δ-close to Q(x),
i.e., ρx ∈ Bδ (Q(x)) ∩ D(H) , and
σ′
≥
2−α Tr(σx ) ρx
≥
2−α
ρx .
1 + ǫ2
This inequality is precisely in the form of inequality (2.6). Now we run the JRS protocol to approximate Q(x),
with t = 2α (1 + ǫ2 ) ln ǫ24 . At the end of this protocol, Bob’s output is
σ̃x
where κ = α + log(1 + ǫ2 ).
:=
1 − 1 − 2−κ
t
17
σx + (1 − 2−κ )t
1
,
dim(H)
By joint concavity of fidelity, and because σx is
F(Q(x), σ̃x )
≥
√ ǫ
-close
1+ǫ2
1 − (1 − 2−κ )t F(Q(x), σx )
to Q(x), we have
1 − (1 − 2−κ )t
√
1 + ǫ2
≥
≥
p
1 − ǫ2 .
√
Here we appealed to the inequalities ln(1 − x) ≤ −x and 1 − x ≤ 1 − x2 (for x ∈ [0, 1)), and the definition
of κ and t. Thus, the purified distance of Q(x) and σ̃x is at most ǫ, and the protocol performs remote state
preparation with worst-case error ǫ. The communication cost of this protocol is ⌈log(t + 1)⌉. Hence, we have
Q∗ (RSP(S, Q), ǫ)
≤
⌈log(t + 1)⌉
≤
α + log2 (1 + ǫ2 ) + log2 ln
2
+2 ,
ǫ4
the stated upper bound.
4.2
A lower bound
By definition, any protocol with worst-case error at most ǫ is also a protocol with average-case error at most ǫ.
As a consequence, any lower bound for average-case communication complexity is also a lower bound for
worst-case communication complexity. In particular, by Theorem 3.3, for each probability distribution p,
Iǫmax (A : B)ρ(p) is a lower bound for the worst-case communication complexity of remote state preparation.
Therefore,
max Iǫmax (A : B)ρ(p) ≤ Q∗ (RSP(S, Q), ǫ) ,
(4.1)
p
where the maximum is over all probability distributions p on the set S. In the following theorem, we give a
lower bound for Q∗ (RSP(S, Q), ǫ) in terms of max-relative entropy using Equation (4.1).
Theorem 4.2. Let S be a non-empty finite set, Q : S → D(H) be a function from S to the set of density
operators in Hilbert space H, ǫ ∈ (0, 1], and δ ∈ (0, 1 − ǫ2 ). Then
min max Dγmax (Q(x)kσ) − log
σ∈D(H) x∈S
where γ =
(1 − ǫ2 )(ǫ2 + δ)
− 3 log 3
δ3
≤
Q∗ (RSP(S, Q), ǫ) ,
p
2(ǫ2 + δ).
Proof: By definition of the smooth max-information, Eq. (4.1) implies that
max min
p
σ∈D(H)
Dǫmax (ρAB (p) k ρA (p) ⊗ σ)
≤
Q∗ (RSP(S, Q), ǫ) ,
(4.2)
whereas the upper bound shown in Theorem 4.1 is
min max Dδmax (Q(x)kσ) .
σ∈D(H) x∈S
If the minimax theorem held for the above expression, the theorem would follow. However, smooth maxrelative entropy Dǫmax is neither convex nor concave in its arguments, and the minimax theorem does not
apply directly. Instead, we appeal to Theorem 2.10, and approximate it with hypothesis testing relative
entropy Dǫh , and write it in terms of the hypothesis testing error βǫ . This measure satisfies the hypotheses of
the minimax theorem (cf. Proposition 2.9 and 2.8). We then apply the minimax theorem, and finally return
to Dǫmax to derive the lower bound.
18
By Theorem 2.10, we have
max min
p
σ∈D(H)
Dǫmax (ρAB (p) k ρA (p) ⊗ σ)
Dλh (ρAB (p) k ρA (p) ⊗ σ) − f (ǫ, δ)
− log βλ (ρAB (p) k ρA (p) ⊗ σ)
≥
max min
=
max min
=
+ log(1 − λ) − f (ǫ, δ)
− log min max βλ (ρAB (p) k ρA (p) ⊗ σ)
p
p
σ∈D(H)
σ∈D(H)
p
σ∈D(H)
+ log(1 − λ) − f (ǫ, δ) ,
2
where f (ǫ, δ) = log (1−ǫ δ)(ǫ
3
2
+δ)
+ 3 log 3 and λ = 1 − ǫ2 − δ.
Let A1 be the set of all probability distributions p over S, and A2 be the set of all quantum states σ ∈ D(H).
Viewing σ as an element of the real vector space of Hermitian operators in L(H), A1 and A2 are non-empty,
convex and compact subsets of Rn for some positive integer n. The quantity βλ (ρAB (p) k ρA (p) ⊗ σ) is a
continuous function of its arguments. Moreover, by Proposition 2.8 and Proposition 2.9, it satisfies both
conditions of the minimax theorem, Theorem 2.11. Thus, we conclude that
max min Dǫmax (ρAB (p) k ρA (p) ⊗ σ) ≥ − log max min βλ (ρAB (p) k ρA (p) ⊗ σ)
p
σ∈D(H)
σ∈D(H)
=
p
+ log(1 − λ) − f (ǫ, δ)
min max Dλh (ρAB (p) k ρA (p) ⊗ σ) − f (ǫ, δ)
σ∈D(H)
p
min max Dγmax (ρAB (p) k ρA (p) ⊗ σ) − f (ǫ, δ)
≥
σ∈D(H)
≥
σ∈D(H) x∈S
p
min max Dγmax (Q(x)kσ) − f (ǫ, δ) ,
(4.3)
p
p
where γ = 2(1 − λ) = 2(ǫ2 + δ). In the second inequality above, we use Theorem 2.10 to move between
hypothesis testing relative entropy and max-relative entropy. Combining Eqs. (4.3) and (4.2), we get the
lower bound for the worst-case communication complexity of ARSP.
5
Some observations
In earlier sections, we characterized the communication complexity of the approximate remote state preparation problem (ARSP) for both worst-case error and average-case error. We now discuss the results, especially
in light of previous work.
5.1
A comparison with previous works
In Section 4, we derived bounds on the worst-case communication complexity of ARSP. Jain [25] showed
that the worst-case communication complexity of ARSP of a sequence of quantum states (Q(x) : x ∈ S) is
bounded from above in terms of the “maximum possible information” T(Q) as:
8(4T(Q) + 7)
√
2 ,
1 − 1 − ǫ2
where ǫ is the approximation error. (See Section 2.4 for a definition of T(Q).)
19
(5.1)
We observe that for certain sets of states there is a large separation between the bound established in
Theorem 4.1, and Equation (5.1). Specifically, the upper bound in Theorem 4.1 may be asymptotically
smaller than the bound in Equation (5.1).
The separation follows from a combination of two pieces of work. The first is an information-theoretic result,
the Substate theorem due to Jain, Radhakrishnan, and Sen [30], which relates the smooth max-relative
entropy of two states to their observational divergence. The precise form of the statement below is due to
Jain and Nayak [26].
Theorem 5.1 (Substate theorem [30, 26]). Let H be a Hilbert space, and let ρ, σ ∈ D(H) be quantum states
such that supp(ρ) ⊆ supp(σ). For any ǫ ∈ (0, 1),
Dǫmax (ρkσ)
≤
Dobs (ρkσ)
1
+ log
.
2
ǫ
1 − ǫ2
The second result is due to Jain, Nayak, and Su [27], who constructed an ensemble of quantum states for
which there is a large separation between its Holevo and Divergence information. (See Section 2.4 for a
definition of these two information quantities.)
Theorem 5.2. Let n be a positive integer, and H be a Hilbert space of dimension n. For every positive real
2
number k ≥ 1 such that log2 n > 36kP
, there is a finite set S and an ensemble E = {(λx , ξx ) : x ∈ S} of
quantum states ξx ∈ D(H) with ξ := x∈S λx ξx = n1 , such that Dobs (Q(x)kξ) = Dobs (E ) = k for all x ∈ S
and χ(E ) ∈ Θ(k log log n).
Jain et al. [27] also showed that this is the best separation possible for an ensemble of quantum states with
a completely mixed ensemble average.
Putting these together, we get:
Theorem 5.3. Let δ ∈ (0, 1] and H be Hilbert space with dimension n. Then, for every positive real
number k ≥ 1 such that log2 n > 36k 2 , there is a finite set S and a function Q : S → D(H) such that T(Q) ∈
Ω(k log log n) while
k
1
min max Dδmax (Q(x)kσ) ≤
+ log
.
2
σ∈D(H) x∈S
δ
1 − δ2
Proof: Let S be the set S and E = {(λx , ξx ) : x ∈ S} the ensemble givenP
by Theorem 5.2. Let Q : S → D(H)
be the function such that Q(x) = ξx for all x ∈ S. Suppose that ξ := x∈S λx ξx is the ensemble average.
Then we have
min max Dδmax (Q(x)kσ)
σ∈ D(H) x∈S
max Dδmax (Q(x)kξ)
≤
x∈S
maxx Dobs (Q(x)kξ)
1
+ log
δ2
1 − δ2
1
k
+ log
,
=
δ2
1 − δ2
where the second inequality is derived using the Substate theorem (Theorem 5.1). Moreover, by definition of
the maximum possible information T(Q), we have T(Q) ≥ χ(E ). This gives us the existence of the required
function Q.
≤
Jain [25] also gave a lower bound of T(Q)/2 for exact remote state preparation. The above observation also
implies that allowing remote state preparation with non-zero error in approximating the state may decrease
the communication cost asymptotically. By Theorem 5.3, we get a function Q for which the worst-case
complexity with zero error Q∗ (RSP(S, Q), 0) ∈ Ω(k log log n), while for any ǫ ∈ (0, 1], the complexity with
error ǫ is
1
k
+ log
,
Q∗ (RSP(S, Q), ǫ) ≤
2
δ
1 − δ2
ǫ
where δ := 2√1+ǫ
.
2
20
5.2
Average-case error vs. worst-case error
Requiring bounded worst-case error in approximating states in remote state preparation is more demanding,
and potentially requires more communication, as compared to the average case. Here we quantify how much
more expensive it could be.
For the rest of this subsection, we let n be a positive integer, fix S = {1, 2, . . . , 2n }, H = span {|xi : x ∈ S},
and define Q : S → D(H) by Q(x) = |xihx| for all x ∈ S.
√
Proposition 5.4. For every ǫ ∈ [0, 1/ 2 ), there is a probability distribution pǫ over the set S such that
Q∗pǫ (RSP(S, Q), ǫ) = 0, while Q∗ (RSP(S, Q), ǫ) ≥ n.
Using quantum teleportation, any set of quantum states in space H can be prepared with zero error with
communication cost 2n. Thus, the above separation is maximal, up to the factor of 2.
To prove Proposition 5.4, we first analyze worst-case error protocols.
√
Lemma 5.5. For any ǫ ∈ [0, 1/ 2 ), Q∗ (RSP(S, Q), ǫ) ≥ n.
Proof: Given any ARSP protocol Π for the given set of states Q, we construct an LOCC protocol Π′ for
transmitting n bits:
Protocol Π′
1. Alice, with input x ∈ S, and Bob (with no input) simulate the protocol Π.
2. Let σx be the output of Π, obtained by Bob. Bob measures σx according to the projective measurement (|yihy| : y ∈ S).
The communication complexity of Π′ equals that of Π.
Suppose Alice is given a uniformly random input, and let X be the corresponding random variable. Let Y
be the random variable corresponding to Bob’s output in Π′ . Then, by the monotonicity of fidelity under
quantum channels, the success probability of Π′ is
1 X
F(σx , Q(x))2 ≥ 1 − ǫ2 .
Pr[Y = X] ≥
2n x
By Theorem 1.2, the communication cost of Π′ , and therefore of Π, is at least n+log(1−ǫ2 ). Since ǫ ∈ [0, √12 ),
we have log(1 − ǫ2 ) > −1. So Q∗ (RSP(S, Q), ǫ) ≥ n.
We show that the complexity of the task drops drastically, if average-case error is considered.
√
Lemma 5.6. For every ǫ ∈ [0, 1/ 2 ), There is a probability distribution pǫ over the set S such that
Q∗pǫ (RSP(S, Q), ǫ) = 0.
Proof: Fix some x0 ∈ S. Let pǫ be the probability distribution defined by
(√
1 − ǫ2 x = x0
√
pǫ,x =
1− 1−ǫ2
x 6= x0 .
2n −1
Consider the protocol Π in which Alice does not send any message to Bob, and Bob always prepares the
state Q(x0 ) = |x0 ihx0 |. The final joint state of the input-output registers in the protocol Π is
X
ρ′AB =
pǫ,x |xihx| ⊗ Q(x0 )
x∈S
21
and the communication cost is zero. Denoting by ρAB the ideal input-output state, we have
p
F(ρAB , ρ′AB ) ≥
1 − ǫ2 .
So Q∗p (RSP(S, Q), ǫ) = 0.
Thus we conclude Proposition 5.4. In fact we can construct an ensemble independent of ǫ, which exhibits a
similar disparity between worst and average-case ARSP.
√
Proposition 5.7. There is a probability distribution p over S such that for every ǫ ∈ [0, 1/ 2 ), we have
Q∗p (RSP(S, Q), ǫ)
≤
n
2 o
+2 .
log min 2n , log2 2
ǫ
Proof: Let m := 2n . Define p as the geometrically decreasing probability distribution
(
1
x ∈ {1, . . . , m − 1}
2x
px =
.
1
x=m
2m−1
Now consider the following protocol Π for ARSP. If Alice’s input x belongs to the set {1, . . . , t} with t =
min{⌈log ǫ22 ⌉, m}, then she sends x to Bob. Otherwise, she sends a random number chosen from the
set {1, . . . , t} to Bob. After receiving Alice’s message y, Bob outputs the state Q(y).
In protocol Π, the final state of Alice and Bob is of the form
ρ′AB
:=
m
X
x=1
px |xihx| ⊗ σx ,
where σx = Q(x) for x ≤ t. Consequently
F(ρAB , ρ′AB )
=
m
X
px F(Q(x), σx )
x=1
≥
t
X
x=1
px
≥
p
1 − ǫ2 .
Therefore, the average-case error is at most ǫ, and the communication is ⌈log t⌉. This implies that
Q∗p (RSP(S, Q), ǫ)
as claimed.
≤
n
2 o
log min 2n , log 2
+2 ,
ǫ
This example illustrates that the more sharply skewed the probability distribution over Q, the bigger the
gap between the worst-case and the average-case is. The example in Lemma 5.6 is a limiting case of such a
distribution.
5.3
Connection to the asymptotic case
It is worth mentioning that our bounds for the average-case communication complexity of ARSP in the
one-shot scenario also gives the optimal bounds in the asymptotic scenario established earlier by Berry and
Sanders [7]. This can be derived using the Quantum Asymptotic Equipartition Property of max-information,
i.e., Theorem 2.6. In the asymptotic scenario, Alice is given n independent and identically distributed inputs.
Using the notation from Section 3, the target joint state of Alice’s input and Bob’s output is ρ(p)⊗n , and
the goal is to prepare it approximately on Bob’s side with average error ǫ.
22
Let q∗p (RSP(S, Q), ǫ) denote the asymptotic rate of communication complexity of ARSP with average error ǫ.
This is the limit of the communication complexity of preparing ρ(p)⊗n with average-case error ǫ, divided
by n, as n → ∞. By Theorems 3.1 and 3.3, we have
8
1 δ
1 ǫ
∗
⊗n
⊗n
Imax (A : B)ρ(p) + log2 ln 2 + 2 ,
I
(A : B)ρ(p)
≤ qp (RSP(S, Q), ǫ) ≤
lim
lim
n→∞ n
n→∞ n max
ǫ
where δ =
ǫ
√
.
2 2
So by inequalities (2.1) and (2.2) in Theorem 2.6, we get the following bounds:
I(A : B)ρ(p) − 2ǫ log(|A| |B|)
6
≤
q∗p (RSP(S, Q), ǫ)
≤
I(A : B)ρ(p) .
On LOCC protocols for transmitting bits
In this section, we digress from the main theme of this article; we characterize the communication required
to convey classical bits through LOCC protocols as in Theorem 1.2. We have used this in Section 5 to
highlight a key difference between worst-case and average-case protocols for remote state preparation.
Consider the following communication task T :
Two physically separated parties, Alice and Bob, have unlimited computational power and can
communicate with each other. Alice is given a uniformly random n-bit string X unknown to
Bob, that is independent of their initial state. Alice and Bob communicate with each other so
that Bob learns X with probability at least p ∈ (0, 1].
Consider a classical communication protocol in which Alice sends exactly ⌈n − log p1 ⌉ bits of X, and Bob
chooses uniformly random bits as his guess for the remaining bits. Then the probability that Bob correctly
decodes Alice’s message is at least p. In this section, we show that even if we allow Alice and Bob to use
LOCC protocols, the classical communication complexity of the task T does not decrease. In other words, in
any (potentially two-way) LOCC protocol for this task, Alice sends at least n + log p bits in order to achieve
success probability at least p (Theorem 1.2). Nayak and Salzman [35] showed that in any two-way quantum
communication protocol with shared entanglement for the task T , Alice sends at least 12 (n + log p) qubits to
Bob. We obtain Theorem 1.2 by strengthening their proof.
6.1
Preparation
In LOCC protocols we assume that Alice and Bob each have access to an arbitrarily large but finite supply
of qubits in some fixed basis state, say |0̄i. Without loss of generality, we further assume that during a
protocol, each party performs some unitary operation followed by the measurement of a subset of qubits in
the standard basis. Note that any measurement can be implemented in this manner [37, Sec 2.2.8]. Further,
if the subset of qubits measured is of size k, we may assume that it consists of the leftmost k qubits.
We state some properties of protocols and states from Ref. [35] which are used later in this section. For
completeness we include their proofs here.
Proposition 6.1 ([35]). In any communication protocol with prior entanglement and local quantum channels,
we may assume that the initial shared quantum state is of the form
X
(1A ⊗ Λ)
|riA |riB ,
r∈{0,1}e
for some Λ :=
P
r∈{0,1}e
√
P
λr |rihr| with λr ≥ 0, r∈{0,1}e λr = 1, and for some integer e ≥ 1.
23
Proof: Without loss of generality, assume
and Bob hold eA and eB qubits of the initial state,
P that Alice
√
respectively, where eB ≥ eA . Let |φi = i∈{0,1}eA γi |ai iA |bi iB be a Schmidt decomposition of the initial
shared state.
√
P
We define a new protocol in which Alice and Bob start with the shared state |ψi := r∈{0,1}eB λr |riA |riB ,
where λ0̄s = γs for s ∈ {0, 1}eA and is zero otherwise. The state simplifies to
X √
γi |0̄, iiA |0̄, iiB .
i∈{0,1}eA
Using appropriate local unitary operators, Alice and Bob produce the state |φi (tensored with some fixed
pure state), and then run the original protocol.
Proposition 6.2 ([35]). For any linear transformation T on e qubits and any orthonormal set {|φa i : a ∈
{0, 1}e } over e′ ≥ e qubits,
X
X
T |ai ⊗ |φa i =
|ai ⊗ T̃ |φa i ,
a∈{0,1}e
a∈{0,1}e
P
where T̃ is any transformation on e′ qubits such that for all a′ ∈ {0, 1}e , T̃ |φa′ i = a∈{0,1}e ha′ |T |ai|φa i.
If T is a unitary operation, then we may take T̃ to be a unitary operation on e′ qubits.
Proof: Since the set {|ai : a ∈ {0, 1}e } is an orthonormal basis for the Hilbert space of e qubits, we have
X
XX
T |ai|φa i =
ha′ |T |ai|a′ i|φa i
a
a′
X
|a′ i
a∈{0,1}e
=
a′
=
X
a′
′
X
a
ha′ |T |ai|φa i
|a iT̃ |φa′ i ,
as claimed. The second part of the proposition is straightforward.
We also use this property in the following form in our analysis. The proof is straightforward, and is omitted.
P
Corollary 6.3. For any controlled unitary operation T :=
z∈{0,1}m |zihz| ⊗ Tz on a classical-quantum
register with m bits and e qubits, and collections of orthonormal sets {|ψza i : a ∈ {0, 1}e} over e′ qubits
m
with e′ ≥ e and z ∈ {0, 1} ,
X
X
X
X
|zai ⊗ T̃ (|zi|ψza i) ,
T |zai ⊗ |zi|ψza i =
z∈{0,1}m a∈{0,1}e
z∈{0,1}m a∈{0,1}e
where T̃ := z∈{0,1}m |zihz| ⊗ T̃z , and (T̃z ) is a sequence of unitary transformations on e′ qubits such that
P
for all z ∈ {0, 1}m and a′ ∈ {0, 1}e, T̃z |ψza′ i = a∈{0,1}e ha′ |Tz |ai|ψza i.
P
6.2
One-way LOCC protocols
As a warm-up, we prove the analogue of Theorem 1.2 for one-way LOCC protocols.
Theorem 6.4. Let Y be Bob’s output in any one-way LOCC protocol for task T when Alice receives uniformly
distributed n-bit input X. Let p := Pr[Y = X] be the probability that Bob gets the output X. Then
m
≥
n − log
1
,
p
where m is the number of classical bits Alice sends to Bob in the protocol.
24
P
Proof: Using Proposition 6.1, we assume that the initial shared entangled state is r∈{0,1}e |riΛ|ri for
√
P
P
some Λ :=
λr |rihr| with λr ≥ 0 and
r∈{0,1}e
r∈{0,1}e λr = 1, and some e ≥ 1. As explained in
Section 6.1, first Alice performs a unitary transformation on her part of the initial state depending on her
input X and measures the left-most m qubits in the standard basis. Let Ux be the unitary operation Alice
uses when she is given x as input. After the unitary operation Ux is performed, the joint state is
X
X
(Ux ⊗ 1)(1 ⊗ Λ)
|ri ⊗ |ri = (1 ⊗ Λ)(Ux ⊗ 1)
|ri ⊗ |ri
r∈{0,1}e
r∈{0,1}e
=
(1 ⊗ Λ)(1 ⊗ UxT )
X
=
r∈{0,1}e
X
r∈{0,1}e
|ri ⊗ |ri
(By Proposition 6.2)
|riΛUxT |ri .
Then Alice measures the state as described above and sends Bob the outcome of her measurement. Bob’s
state after this step is
X
ξx =
|zihz| ⊗ ΛUxT (|zihz| ⊗ 1)U x Λ∗ .
z∈{0,1}m
Note that
ξx
=
(1 ⊗ Λ)
X
z∈{0,1}m
|zihz| ⊗ UxT (|zihz| ⊗ 1)U x (1 ⊗ Λ∗ )
≤
(1 ⊗ ΛΛ∗ ) ,
(6.1)
where the identity operator acts on a 2m dimensional space. Finally, Bob performs a projective measurement {Py }y∈{0,1}n on his qubits, and gets as outcome the random variable Y . The success probability p of
the protocol is
X
Pr[X = x] Pr[Y = x|X = x]
Pr[X = Y ] =
x∈{0,1}n
=
X
x∈{0,1}n
≤
=
=
1
Tr (Px ξx )
2n
1 X
Tr Px (1 ⊗ ΛΛ∗ )
n
2 x
(By equation (6.1))
1
Tr(1 ⊗ ΛΛ∗ )
2n
2m
.
2n
We conclude that m ≥ n + log p.
6.3
The extension to two-way LOCC protocols
We now extend the above result to any two-way LOCC protocol. In particular we prove Theorem 1.2, which
we restate here for convenience.
Theorem 6.5. Let Y be Bob’s output in any two-way LOCC protocol for task T when Alice receives uniformly
distributed n-bit input X. Let mA be the total number of bits Alice sends to Bob, and p := Pr[Y = X] be the
probability that Bob produces output X. Then
mA
≥
n − log
25
1
.
p
To prove the theorem, we characterise the joint state of Alice and Bob at the end of a bounded round LOCC
protocol.
Lemma 6.6. Let Π be a bounded round LOCC protocol. Let e be the initial number of qubits with each of
Alice and Bob, q be the total number of bits sent by Alice to Bob, q ′ be the total number of bits sent by Bob
to Alice, and m be the total number of bits exchanged in Π (so m = q + q ′ ). Then Alice and Bob’s joint state
at the end of the protocol (before the measurement for producing the output) can be written as
X
X
|z, rihz, s|A ⊗ Λ|φz,r ihφz,s |B Λ∗ ,
z∈{0,1}m r,s∈{0,1}e−q
where
1. A and B are classical-quantum registers with m-bit classical parts that contain the transcript of the
protocol; register A is with Alice, and B with Bob,
2. Λ is a linear transformation that maps classical-quantum states with m bits and e qubits to classicalquantum states of the same form, depends only on the initial joint state and the unitary transformations
applied by Bob, and satisfies Tr(ΛΛ∗ ) = 2q ; and
3. {|φz,r i} is an orthonormal set of classical-quantum states of the form |φz,r i := |zi|ψz,r i over m-bits
and e qubits, and depends only on the initial joint state and the unitary transformations applied by
Alice.
Proof: Suppose that Π is a t-round LOCC protocol. Let ρi be the joint state of Alice and Bob after i-th
round, and mi be the total number of bits exchanged by Alice and Bob in the first i rounds, of which qi bits
are sent by Alice, for 1 ≤ i ≤ t. Let ρ0 be their initial state.
We prove the lemma by induction on t.
Base Case:
have
Suppose that Π is a zero communication LOCC protocol, i.e., t = 0. By Proposition 6.1, we
X
|rihs| ⊗ Λ|rihs|Λ∗ ,
ρ0 =
r,s∈{0,1}e
√
P
where Λ = r∈{0,1}e λr |rihr| for some λr ≥ 0 and r λr = 1. Since Tr(ΛΛ∗ ) = 1, the state ρ0 satisfies the
claimed properties.
P
Induction Hypothesis: Suppose the lemma holds for any l-round LOCC protocol, for some l ≥ 0.
Inductive Step: Suppose that Π is an (l + 1)-round protocol. By the induction hypothesis, after the first l
rounds of communication we have
X
X
|z, rihz, s| ⊗ Λl |φz,r ihφz,s |Λ∗l ,
ρl =
z∈{0,1}ml r,s∈{0,1}e−ql
where Λl and |φz,r i satisfy the properties stated in the lemma. In particular, suppose |φz,r i := |zi|ψz,r i for
each z, r. We show that at the end of the protocol ρl+1 is in the required form as well. Consider the (l + 1)-th
round of Π.
Case (1): Suppose P
that the communication in the last round is from Alice to Bob. Alice applies a unitary
transformation U := z |zihz|⊗Uz , which acts on the quantum part of her register, controlled by the classical
part of her register. She then measures the k leftmost qubits in the standard basis, appends the outcome to
26
the message transcript in her classical register, and sends the outcome a of her measurement to Bob. The
joint state after applying U is
(U ⊗ 1)(1 ⊗ Λl )
X
r,s∈{0,1}e−ql
z∈{0,1}ml
=
=
(1 ⊗ Λl )
"
(1 ⊗ Λl )
"
X
r,s,z
X
r,s,z
|z, rihz, s| ⊗ |φz,r ihφz,s | (1 ⊗ Λ∗l )(U ∗ ⊗ 1)
#
∗
U |z, rihz, s|U ⊗ |φz,r ihφz,s | (1 ⊗ Λ∗l )
|z, rihz, s| ⊗ Ũ |φz,r ihφz,s |Ũ
∗
#
(1 ⊗ Λ∗l ) ,
P
where Ũ :=
z |zihz| ⊗ Ũz is the unitary operation given by Corollary 6.3. After Alice performs her
measurement and sends the measurement outcome a to Bob, say he stores the message in register M . Denote
by 1M ⊗ Λl Ũ the operator Λl Ũ on the registers originally with Bob, extended to include the register M .
(The order of the operators in tensor product does not represent the order of the registers.) The joint state
then may be expressed as below.
ρl+1
=
X
r ′,s′ ∈{0,1}e−(ql +k)
a∈{0,1}k
z∈{0,1}ml
|za, r′ ihza, s′ | ⊗ (1M ⊗ Λl Ũ )(|zihz| ⊗ |aiha|M ⊗ |ψz,ar′ ihψz,as′ |)(1M ⊗ Ũ ∗ Λ∗l ) ,
where Λl Ũ acts on the classical-quantum register with Bob before the message was sent. We define Λl+1 :=
1M ⊗ Λl , and |φz′ ,r′ i := |zai ⊗ Ũz |ψz,ar′ i, where z ′ := za. Noting that ml+1 = ml + k and ql+1 = ql + k, we
have
X
ρl+1 =
|z ′ , r′ ihz ′ , s′ | ⊗ Λl+1 |φz′ ,r′ ihφz′ ,s′ |Λ∗l+1 .
r ′ ,s′ ∈{0,1}e−ql+1
z ′ ∈{0,1}ml+1
Further note that Tr(Λl+1 Λ∗l+1 ) = 2ql+1 and {|φz′ ,r′ i} is an orthonormal set of the claimed form.
Case (2): Suppose that
P the communication in the last round is from Bob to Alice. Bob applies a unitary
transformation V :=
z |zihz| ⊗ Vz to the quantum part of his register, controlled by the classical part
of his register. Then he measures the k leftmost qubits (say in sub-register L) in the standard basis, and
appends the outcome b to the message transcript, in classical register M . Finally, he sends the outcome b of
the measurement to Alice. Denote by 1M ⊗ (hb|L ⊗ 1)V Λl , the extension of the operator (hb|L ⊗ 1)V Λl to
include the register M . (Here, the order of the operators in tensor product does not represent the order of
the registers on which they act. The same applies to the operator Λl+1 defined below.) The joint state then
is as follows.
X
|zb, rihzb, s| ⊗ (1M ⊗ (hb|L ⊗ 1)V Λl )(|zbihzb| ⊗ |ψz,r ihψz,s |)(1M ⊗ Λ∗l V ∗ (|biL ⊗ 1)) .
ρl+1 =
r,s∈{0,1}e−ql
b∈{0,1}k
z∈{0,1}ml
P
Note that ql+1 = ql , and ml+1 = ml +k. Define Λl+1 := b |bihb|M ⊗(hb|L ⊗1)V Λl and |φz′ ,r′ i = |zbi⊗|ψz,br′ i,
where z ′ := zb. It is straightforward to verify that Tr(Λl+1 Λ∗l+1 ) = 2ql+1 , the set {|φz′ ,r′ i} is of the claimed
form, and
X
|z ′ , r′ ihz ′ , s′ | ⊗ Λl+1 |φz′ ,r′ ihφz′ ,s′ |Λ∗l+1 .
ρl+1 =
r ′ ,s′ ∈{0,1}e−ql+1
z ′ ∈{0,1}ml+1
27
This completes the proof.
We are ready to prove Theorem 1.2, restated in this section as Theorem 6.5.
Proof of Theorem 6.5: By Lemma 6.6, at the end of any two-way LOCC protocol, when Alice has
n
input x ∈ {0, 1} , Bob’s state before performing his final measurement to get Y is
X
Λ|φz,r (x)ihφz,r (x)|Λ∗ ,
ξx =
r∈{0,1}e−mA
z∈{0,1}m
for some linear transformation Λ with Tr(ΛΛ∗ ) = 2mA and orthonormal set {|φz,r (x)i}z,r . The transformation Λ only depends on Bob’s unitary operations and the initial state, and is therefore independent of Alice’s
input x. Note that
ξx ≤ ΛΛ∗ .
(6.2)
After Bob performs his final projective measurement {Py }y∈{0,1}n and gets the output Y , the probability of
correctly recovering an input X chosen uniformly at random is
1 X
p := Pr[Y = X] =
Tr(Px ξx )
2n
n
x∈{0,1}
≤
1 X
Tr(Px ΛΛ∗ )
2n x
1
Tr(ΛΛ∗ )
2n
Therefore, we have mA ≥ n − log p1 , as required.
=
7
=
(Equation (6.2))
2 mA
.
2n
Conclusion
In this article, we studied the communication complexity of remote state preparation in the one-shot scenario.
Our main results can be summarized as follows:
• The communication complexity of remote state preparation with bounded average-case error ǫ can be
characterized tightly in terms of the smooth max-information Bob’s output has about Alice’s input.
• The communication complexity of remote state preparation with bounded worst-case error ǫ can be
characterized in terms of a similar natural expression involving smooth max-relative entropy.
The bounds we derive for the worst-case communication complexity are provably tighter than earlier ones.
We also show out how protocols that guarantee low worst-case error necessarily use more communication
than those that require low error on average. In the process, we strengthen a lower bound on LOCC protocols
for transmitting classical bits.
In this work, we focused on the remote preparation of a possibly mixed quantum state. However, often the
quantum state to be remotely prepared is entangled with other systems (“the environment”). We can consider
the problem of preparing an approximation of the quantum state such that its entanglement with other
systems does not change significantly. This problem has been studied in asymptotic scenario [6, 7]. Berta [8]
implicitly studied this problem in the one-shot scenario by considering the quantum state merging problem,
and showed that the minimal entanglement cost needed for this problem is equal to minus the ǫ-smooth
conditional min-entropy of Alice’s register conditioned on the environment, while classical communication is
allowed for free. Note that the entanglement cost is defined as the difference between the number of bits of
pure entanglement at the beginning and at the end of the process. It would be interesting to characterize the
minimum classical communication of such “faithful” ARSP in terms of non-asymptotic information theoretic
quantities.
28
References
[1] Scott Aaronson. Guest column: NP-complete problems and physical reality. SIGACT News, 36(1):30–
52, March 2005.
[2] Shima Bab Hadiashar. Communication Complexity of Remote State Preparation. Master’s Thesis,
University of Waterloo, Waterloo, Canada, May 2014.
[3] Charles H. Bennett, Gilles Brassard, Claude Crépeau, Richard Jozsa, Asher Peres, and William K.
Wootters. Teleporting an unknown quantum state via dual classical and Einstein-Podolsky-Rosen channels. Physical Review Letters, 70(13):1895–1899, March 1993.
[4] Charles H. Bennett, David P. DiVincenzo, Peter W. Shor, John A. Smolin, Barbara M. Terhal, and
William K. Wootters. Remote state preparation. Physical Review Letters, 87(7):077902, July 2001.
[5] Charles H. Bennett, David P. DiVincenzo, John A. Smolin, and William K. Wootters. Mixed-state
entanglement and quantum error correction. Physical Review A, 54:3824–3851, Nov 1996.
[6] Charles H. Bennett, Patrick Hayden, Debbie W. Leung, Peter W. Shor, and Andreas Winter. Remote
preparation of quantum states. IEEE Transactions on Information Theory, 51(1):56–74, January 2005.
[7] Dominic W. Berry and Barry C. Sanders. Optimal remote state preparation. Physical Review Letters,
90:057901, February 2003.
[8] Mario Berta. Single-Shot Quantum State Merging. Diploma thesis, ETH, Zurich, February 2008.
[9] Mario Berta, Matthias Christandl, and Renato Renner. The Quantum Reverse Shannon theorem based
on one-shot information theory. Communications in Mathematical Physics, 306(3):579–615, September
2011.
[10] Igor Bjelakovic and Rainer Siegmund-Schultze. Quantum Stein’s lemma revisited, inequalities for quantum entropies, and a concavity theorem of Lieb. Technical Report arXiv:quant-ph/0307170v2, arXiv.org,
2003.
[11] Lenore Blum, Mike Shub, and Steve Smale. On a theory of computation and complexity over the
real numbers: NP-completeness, recursive functions and universal machines. Bulletin of the American
Mathematical Society, 21:1–46, 1989.
[12] Fernando G. S. L. Brandão and Nilanjana Datta. One-shot rates for entanglement manipulation under
non-entangling maps. IEEE Transactions on Information Theory, 57(3):1754–1760, March 2011.
[13] Francesco Buscemi and Nilanjana Datta. The quantum capacity of channels with arbitrarily correlated
noise. IEEE Transactions on Information Theory, 56(3):1447–1460, March 2010.
[14] Eric Chitambar, Debbie Leung, Laura Mancinska, Maris Ozols, and Andreas Winter. Everything You
Always Wanted to Know About LOCC (But Were Afraid to Ask). Communications in Mathematical
Physics, 328(1):303–326, 2014.
[15] Nikola Ciganovic, Normand J. Beaudry, and Renato Renner. Smooth max-information as one-shot
generalization for mutual information. Information Theory, IEEE Transactions on, 60(3):1573–1581,
March 2014.
[16] Nilanjana Datta. Min- and max-relative entropies and a new entanglement monotone. IEEE Transactions on Information Theory, 55(6):2816–2826, June 2009.
[17] Igor Devetak and Toby Berger. Low-entanglement remote state preparation. Physical Review Letters,
87(19), October 2001.
[18] Dawei Ding and Mark M. Wilde. Strong converse exponents for the feedback-assisted classical capacity
of entanglement-breaking channels. Technical Report arXiv:1506.02228v4 [quant-ph], arXiv.org, 2017.
29
[19] Frédéric Dupuis, Lea Kraemer, Philippe Faist, Joseph M. Renes, and Renato Renner. Generalized
entropies. In Proceedings of the XVIIth International Congress on Mathematical Physics, Aalborg,
Denmark, 2012, pages 134–153, August 2012.
[20] Alexei Gilchrist, Nathan K. Langford, and Michael A. Nielsen. Distance measures to compare real and
ideal quantum processes. Physical Review A, 71:062310, Jun 2005.
[21] Akihisa Hayashi, Takeji Hashimoto, and Minoru Horibe. Remote state preparation without oblivious
conditions. Physical Review A, 67:052302, May 2003.
[22] Masahito Hayashi.
Role of hypothesis testing in quantum information.
arXiv:1709.07701 [quant-ph], arXiv.org, 2017.
Technical Report
[23] Masato Hayashi and Hiroshi Nagaoka. A general formula for the classical capacity of a general quantum
channel. In Proceedings of the IEEE International Symposium on Information Theory. IEEE, 2002.
[24] Fumio Hiai and Dénes Petz. The proper formula for relative entropy and its asymptotics in quantum
probability. Communications in Mathematical Physics, 143(1):99–114, 1991.
[25] Rahul Jain. Communication complexity of remote state preparation with entanglement. Quantum
Information & Computation, 6(4):461–464, July 2006.
[26] Rahul Jain and Ashwin Nayak. Short proofs of the quantum Substate theorem. IEEE Transactions on
Information Theory, 58(6):3664–3669, June 2012.
[27] Rahul Jain, Ashwin Nayak, and Yi Su. A separation between divergence and Holevo information for
ensembles. Mathematical Structures in Computer Science, 20(5):977–993, 2010.
[28] Rahul Jain, Jaikumar Radhakrishnan, and Pranab Sen. Privacy and interaction in quantum communication complexity and a theorem about the relative entropy of quantum states. In Proceedings of the
43rd Annual IEEE Symposium on Foundations of Computer Science, pages 429–438, 2002.
[29] Rahul Jain, Jaikumar Radhakrishnan, and Pranab Sen. Prior entanglement, message compression and
privacy in quantum communication. In Proceedings of the Twentieth Annual IEEE Conference on
Computational Complexity, 2005, pages 285–296, June 2005.
[30] Rahul Jain, Jaikumar Radhakrishnan, and Pranab Sen. A property of quantum relative entropy with an
application to privacy in quantum communication. Journal of the ACM, 56(6):33:1–33:32, September
2009.
[31] Felix Leditzky, Mark M. Wilde, and Nilanjana Datta. Strong converse theorems using Rényi entropies.
Journal of Mathematical Physics, 57(8):082202, 2016.
[32] Debbie W. Leung and Peter W. Shor. Oblivious remote state preparation. Physical Review Letters,
90(12):127905, March 2003.
[33] Hoi-Kwong Lo. Classical-communication cost in distributed quantum-information processing: A generalization of quantum-communication complexity. Physical Review A, 62(1):012313, June 2000.
[34] William Matthews and Stephanie Wehner. Finite blocklength converse bounds for quantum channels.
IEEE Transactions on Information Theory, 60(11):7317–7329, 2014.
[35] Ashwin Nayak and Julia Salzman. Limits on the ability of quantum states to convey classical messages.
Journal of the ACM, 53(1):184 – 206, January 2006.
[36] Ashwin Nayak and Peter Shor. Bit-commitment-based quantum coin flipping. Physical Review A,
67:012304, January 2003.
[37] Michael A. Nielsen and Isaac L. Chuang. Quantum Computation and Quantum Information. Cambridge
university press, 2010.
30
[38] Tomohiro Ogawa and Hiroshi Nagaoka. Strong converse and Stein’s lemma in quantum hypothesis
testing. IEEE Transactions on Information Theory, 46(7):2428–2433, 2000.
[39] Martin J. Osborne and Ariel Rubinstein. A Course in Game Theory. MIT Press, 1994.
[40] Alexey E. Rastegin. Relative error of state-dependent cloning. Physical Review A, 66(4):042304, 2002.
[41] Alexey E. Rastegin. A lower bound on the relative error of mixed-state cloning and related operations.
Journal of Optics B: Quantum and Semiclassical Optics, 5(6):S647, 2003.
[42] Alexey E. Rastegin. Sine distance for quantum states. Technical Report arxiv:quant-ph/0602112,
arXiv.org, 2006.
[43] Renato Renner. Security of Quantum Key Distribution. PHD thesis, ETH, Zurich, December 2005.
[44] Renato Renner and Stefan Wolf. Smooth Rényi entropy and applications. In IEEE International
Symposium on Information Theory, page 233, 2004.
[45] Marco Tomamichel, Roger Colbeck, and Renato Renner. Duality between smooth min- and maxentropies. IEEE Transactions on Information Theory, 56(9):4674–4681, September 2010.
[46] Marco Tomamichel and Masahito Hayashi. A hierarchy of information quantities for finite block length
analysis of quantum tasks. IEEE Transactions on Information Theory, 59(11):7693–7710, November
2013.
[47] Dave Touchette. Quantum information complexity. In Proceedings of the Forty-Seventh Annual ACM
on Symposium on Theory of Computing, pages 317–326. ACM, 2015.
[48] Ligong Wang and Renato Renner. One-shot classical-quantum capacity and hypothesis testing. Physical
Review Letters, 108:200501, May 2012.
[49] John Watrous. Theory of Quantum Information. 2016. Book draft, September 2016. Available at
https://cs.uwaterloo.ca/~watrous/TQI/ .
[50] Mark M. Wilde, Andreas Winter, and Dong Yang. Strong converse for the classical capacity of
entanglement-breaking and hadamard channels via a sandwiched Réenyi relative entropy. Communications in Mathematical Physics, 331(2):593–622, October 2014.
[51] Andrew Chi-Chih Yao. Quantum circuit complexity. In Proceedings of the 34th Annual Symposium on
Foundations of Computer Science, 1993, pages 352–361, November 1993.
A
Some properties of entropic quantities
In this section, we present the proofs of some properties of information-theoretic quantities stated in Section 2.4. For convenience, we restate the properties here.
Proposition A.1 (Proposition 2.5). Let ρAB ∈ D(H′ ⊗ H) be a bipartite quantum state that is classical
on A. For any ǫ ≥ 0, there exists ρ′AB ∈ Bǫ (ρAB ) ∩ D(H′ ⊗ H) classical on A such that
Iǫmax (A : B)ρ
=
Imax (A : B)ρ′ .
Proof: Let λ = Iǫmax (A : B)ρ , and ρ̃AB ∈ Bǫ (ρAB ) and σB ∈ D(H) be two quantum states for which
ρ̃AB
≤
2λ ρ̃A ⊗ σB .
31
Without loss of generality, we assume that ρ̃AB has trace equal to one, i.e., ρ̃AB ∈ Bǫ (ρAB ) ∩ D(H′ ⊗ H). If
not, we consider the state ωAB := Tr(ρ̃AB
ρ̃AB ) instead of ρ̃AB . Since ρ has trace 1, P(ω, ρ) ≤ P(ρ̃, ρ). Further,
λ
ωAB ≤ 2 ωA ⊗ σB .
Let ΦA : L(H) → L(H) be a quantum-to-classical channel such that:
X
hei |X|ei i|ei ihei |
ΦA (X) =
i
for all X ∈ L(H), where {|ei i} is the standard basis for L(H). Let ρ′AB = (ΦA ⊗ 1B )(ρ̃AB ). By the definition
of ρ′AB and the monotonicity of purified distance ρ′AB ∈ Bǫ (ρAB ) ∩ D(H′ ⊗ H).
By optimality of ρ̃AB , we have
Iǫmax (A : B)ρ
=
Imax (A : B)ρ̃
≤
Imax (A : B)ρ′ ,
and by Proposition 2.4, monotonicity of smooth max-information, we have
Imax (A : B)ρ′
≤
Imax (A : B)ρ̃ .
Iǫmax (A : B)ρ
=
Imax (A : B)ρ′ ,
Therefore, we conclude that
where ρ′AB ∈ Bǫ (ρAB ) ∩ D(H′ ⊗ H) and is classical on A.
Proposition A.2 (Proposition 2.8). Let ρAB (p) ∈ D(HA ⊗ HB ) be a state classical on A such that the
distribution on A is given by the probability vector p. Let ρA (p) = TrB (ρAB (p)), and σ ∈ D(HB ) be a
quantum state on Hilbert space HB . Then the function βǫ (ρAB (p) k ρA (p) ⊗ σ) is convex with respect to p.
Proof: Let p0 and p1 be two arbitrary probability distributions on the standard basis of HA . For λ ∈ [0, 1],
let q = λp0 + (1 − λ)p1 . We show that
βǫ (ρAB (q) k ρA (q) ⊗ σ)
≤
λ βǫ (ρAB (p0 ) k ρA (p0 ) ⊗ σ) + (1 − λ) βǫ (ρAB (p1 ) k ρA (p1 ) ⊗ σ) ,
which proves the claim.
Let Φ : L(HA ) → L(C2 ⊗ HA ) be the quantum channel with Kraus operators Aa,x =
0 (a)
1 (a)
a and x ∈ {0, 1}, where αa0 := λ pq(a)
and αa1 = (1 − λ) pq(a)
. Then we have
ρXAB (q)
=
(Φ ⊗ 1B )(ρAB (q))
=
√ a
αx |xi ⊗ |aiha| for all
λ |0ih0| ⊗ ρAB (p0 ) + (1 − λ) |1ih1| ⊗ ρAB (p1 ) .
Since ρXAB (q) is an extension of ρAB (q), using Proposition 2.7 twice, we get
βǫ (ρAB (q) k ρA (q) ⊗ σ)
βǫ (ρXAB (q) k ρXA (q) ⊗ σ) .
=
(A.1)
ǫ
For each x ∈ {0, 1}, let Qx be the
Pmeasurement operator that achieves β (ρAB (px ) k ρA (px ) ⊗ σ). Consider
the measurement operator Q := x∈{0,1} |xihx| ⊗ Qx . This satisfies
hQ, ρXAB (q)i
=
λ hQ0 , ρAB (p0 )i + (1 − λ) hQ1 , ρAB (p1 )i
≥
1−ǫ ,
by definition of Q0 , Q1 . By Eq. (A.1) and the definition of βǫ , we get
βǫ (ρAB (q) k ρA (q) ⊗ σ)
= βǫ (ρXAB (q) k ρXA (q) ⊗ σ)
≤ hQ, ρXA (q) ⊗ σi
= λ hQ0 , ρA (p0 ) ⊗ σi + (1 − λ) hQ1 , ρA (p1 ) ⊗ σi
= λ βǫ (ρAB (p0 ) k ρA (p0 ) ⊗ σ) + (1 − λ) βǫ (ρAB (p1 ) k ρA (p1 ) ⊗ σ) ,
as we set out to prove.
32
Proposition A.3 (Proposition 2.9). For any fixed quantum state ρ ∈ D(H), the function βǫ (ρkσ) is a
concave function with respect to σ.
Proof: For any choice of σ0 , σ1 ∈ D(H) and λ ∈ [0, 1], let Q be the measurement operator that achieves
hypothesis testing error βǫ (ρ k λσ0 + (1 − λ)σ1 ). Then
βǫ (ρ k λσ0 + (1 − λ)σ1 )
=
=
≥
hQ, λσ0 + (1 − λ)σ1 i
λ hQ, σ0 i + (1 − λ) hQ, σ1 i
λ βǫ (ρkσ0 ) + (1 − λ) βǫ (ρkσ1 ) ,
since hQ, ρi ≥ 1 − ǫ.
B
Preparing states from an infinite set
In this section, we discuss remote state preparation of states drawn from an infinite set. This scenario has
been studied by Lo [33] and in later works on the topic.
In remote state preparation, Alice’s input is supposed to provide a complete description of the state to be
prepared at Bob’s end. In any physically realistic model of computation, the description necessarily has finite
bit-length (see, e.g., Ref. [1]). For instance, if a d-dimensional quantum state is described by specifying Θ(d2 )
complex entries in the corresponding d×d matrix, the complex numbers would have to be specified with finite
precision. This implies that the input set S (following the notation in Section 2.6) is necessarily countable.
This point has not been addressed in previous works.
To meaningfully consider the preparation states drawn from an uncountable set, we may instead consider
approximations drawn from a suitable countable set. For example, instead of the set D(H) of all quantum
states over a d-dimensional space H, we may instead study the countably dense set of states whose matrix
representations only have complex entries with rational real and imaginary parts. Such states have unique
finite-length representations. (Similar approximation is also implicit in the case of RSP of a finite set of
states, when the corresponding matrices involve irrational numbers.)
Another approach, perhaps only of theoretical interest, would be to allow the local operations in an LOCC
protocol to be defined on a suitable generalization of the Real RAM model due to Blum, Shub, and Smale [11].
We do not attempt to define such a model of computation here. For our purposes, it would suffice to assume
a model which enables the implementation of quantum operations such as unitary operations controlled by
the registers holding real numbers in finite time.
We assume that we take one of the abovementioned approaches in the analysis in this section. The underlying
idea, that of approximating states from an infinite set with those from a net , probably applies in other
reasonable approaches as well.
As before, we restrict ourselves to states over a finite dimensional Hilbert space H.
Definition B.1. Let ν ∈ (0, 1] and D ⊆ D(H) be any set of quantum states. A ν-net N in D is a subset
of D such that for any state ρ ∈ D, there is a state σ ∈ N such that P(ρ, σ) < ν.
We argue that every subset of finite-dimensional states admits a finite net.
Proposition B.1. For every ν ∈ (0, 1], and every set D ⊆ D(H) of quantum states, there is a finite ν-net
in D.
33
Proof: Since D(H) is compact, it has a finite cover (Bi ) consisting of open balls of radius ν/2. This is
also a cover for any subset D of quantum states. Let N be a subset of D constructed by taking one point
from Bi ∩ D, whenever this intersection is non-empty. We claim that this is a finite ν-net in D.
Consider a state ρ ∈ D. Since (Bi ) is a cover for the set of all quantum states, ρ ∈ Bj for some j. By
construction, there is a state σ ∈ N from Bj ∩ D. Since ρ, σ both belong to the same ball Bj of radius ν/2,
we have P(ρ, σ) < ν. So N is a ν-net in D.
Suppose S is an infinite set, and Q : S → D(H) is a one-to-one function mapping each element of S to a quantum state. (We view an element x ∈ S as a description, i.e., unique encoding, of the quantum state Q(x).)
Define R := Q(S) as the image of S under Q; this is the set of quantum states under consideration. We fix
an approximation parameter ν > 0 of our choice, and a finite ν-net N in R, and let T := Q−1 (N ) be the set
of inputs corresponding to N . We bound the communication required for remote state preparation of states
from R with that for states from N . We may then appeal to Theorem 1.1 to infer bounds on RSP(S, Q).
Worst-case error. We first consider the simpler case, that of worst-case error ǫ > 0. Any protocol for
RSP(S, Q) with worst-case error ǫ is also a protocol for RSP(T, Q) as T is a subset of S. So we have
Q∗ (RSP(T, Q), ǫ)
Q∗ (RSP(S, Q), ǫ) .
≤
Now suppose Π is a protocol for RSP(T, Q) with communication cost c and worst case error ǫ. We design
a protocol Π′ for RSP(S, Q) as follows. Given an x ∈ S, Alice chooses y ∈ T such that P(Q(x), Q(y)) ≤ ν,
and prepares an approximation of Q(y) on Bob’s side using protocol Π. Suppose Bob’s output is σy . Then
P(Q(x), σy )
≤
P(Q(x), Q(y)) + P(Q(y), σy )
≤
ν +ǫ .
So Π′ is a protocol for RSP(S, Q) with communication cost c, and worst case error ǫ + ν. Therefore,
Q∗ (RSP(S, Q), ǫ + ν)
≤
Q∗ (RSP(T, Q), ǫ) .
Putting the two together, for ν, ǫ such that 0 < ν < ǫ, we get
Q∗ (RSP(T, Q), ǫ)
≤
Q∗ (RSP(S, Q), ǫ)
≤
Q∗ (RSP(T, Q), ǫ − ν) .
Average-case error. Next we consider approximate RSP with average error at most ǫ ∈ (0, 1] with respect
to a probability measure µ on the set of states R. For simplicity, we only consider the case when the open
sets in R generated by the metric P are measurable. Since Q is injective, we may equivalently consider µ as
a probability measure on S.
Let (ρi ) be an enumeration of the states in N , and (Bi ) be open balls of radius ν centred at ρi with respect
to the metric P. Since N is a ν-net in R, we have R ⊆ ∪i Bi . Define the function f : R → N as f (σ) := ρi
for all states σ ∈ (Bi ∩ R) \ (∪j<i Bj ). The function f maps each quantum state ρ ∈ R to a quantum state
in N such that P(ρ, f (ρ)) < ν. Moreover, it is measurable.
The function f induces a probability distribution p on N in the natural way:
pρi
:=
µ(f −1 (ρi ))
for ρi ∈ N . We may view the distribution p as being over the corresponding set T of inputs: for y ∈ T such
that Q(y) = ρi , we define py := pρi .
We relate protocols for RSP(S, Q) with average error ǫ with respect to µ to protocols for RSP(T, Q) with
average error “close” to ǫ with respect to p.
34
Lemma B.2. Suppose Π is a protocol for RSP(S, Q) with communication cost c and average error ǫ with
respect to µ. Then there is a protocol Π′ for RSP(T, Q) with communication cost c and average error at
most ν + ǫ with respect to p.
Proof: For y ∈ T , define Ry := f −1 (Q(y)), the set of states in R that are mapped to Q(y) ∈ N . Define Sy :=
Q−1 (Ry ), the set of inputs corresponding to the states in Ry . Note that (Ry ) is a partition of R and (Sy )
of S. Since f is measurable, Ry is a measurable set. When Ry has non-zero measure, we define a probability
measure µy on Ry as µy (W ) := µ(W )/µ(Ry ) for all measurable sets W ⊆ Ry . We also view µy as a
probability measure on Sy .
We now construct the protocol Π′ for RSP(T, Q) as follows. Given y ∈ T , Alice selects an input x ∈ Sy
randomly with respect to the probability measure µy and runs the protocol Π on this input.
The communication in Π′ is also c. Suppose σx is the output of the protocol Π when the input is x. Then
the average error of the protocol Π′ is
X Z
X Z
P(Q(y), σx ) dµy (x) =
py
P(Q(y), σx ) dµ(x)
y∈T
x∈Sy
y∈T
≤
x∈Sy
X Z
y∈T
+
X Z
y∈T
≤
P(Q(y), Q(x)) dµ(x)
x∈Sy
P(Q(x), σx ) dµ(x)
x∈Sy
ν +ǫ ,
as claimed.
Conversely, we can also derive a protocol for RSP(S, Q) from one for RSP(T, Q).
Lemma B.3. Suppose Π is a protocol for RSP(T, Q) with communication cost c and average error ǫ with
respect to the distribution p. There exists a protocol Π′ for RSP(S, Q) with communication cost c and average
error at most ǫ + ν with respect to µ.
Proof: In the protocol Π′ , given input x ∈ S, Alice runs the protocol Π on input y defined as y :=
Q−1 (f (Q(x))). This is the input corresponding to the state in the ν-net to which f maps Q(x). The
communication cost of Π′ is also c.
Suppose the output of Π′ on input x is σx . Note that f maps all states Q(x) for x ∈ Sy to the same
value Q(y), and therefore the outputs σx for all inputs x ∈ Sy are equal to σy .
The average error of the protocol with respect to µ is
Z
Z
Z
P(f (Q(x)), σx ) dµ(x)
P(Q(x), f (Q(x))) dµ(x) +
P(Q(x), σx ) dµ(x) ≤
x∈S
x∈S
x∈S
X Z
≤ ν+
P(f (Q(x)), σx ) dµ(x)
x∈Sy
y∈T
=
ν+
X Z
y∈T
=
ν+
X
P(Q(y), σy ) dµ(x)
x∈Sy
py P(Q(y), σy )
y∈T
≤
ν+ǫ ,
35
where in the third step we have used the abovementioned property that f is constant on Sy .
For ν, ǫ such that ν < ǫ, the above two lemmata imply that
Q∗p (RSP(T, Q), ǫ + ν)
≤
Q∗µ (RSP(S, Q), ǫ)
36
≤
Q∗p (RSP(T, Q), ǫ − ν) .
| 7cs.IT |
arXiv:1712.05291v1 [cs.GT] 14 Dec 2017
A Bayesian Clearing Mechanism for Combinatorial Auctions
Gianluca Brero
University of Zurich
Sébastien Lahaie
Google Research
brero@ifi.uzh.ch
slahaie@google.com
Abstract
We cast the problem of combinatorial auction design in a Bayesian framework in order to
incorporate prior information into the auction process and minimize the number of rounds to
convergence. We first develop a generative model of agent valuations and market prices such
that clearing prices become maximum a posteriori estimates given observed agent valuations.
This generative model then forms the basis of an auction process which alternates between
refining estimates of agent valuations and computing candidate clearing prices. We provide
an implementation of the auction using assumed density filtering to estimate valuations and
expectation maximization to compute prices. An empirical evaluation over a range of valuation
domains demonstrates that our Bayesian auction mechanism is highly competitive against the
combinatorial clock auction in terms of rounds to convergence, even under the most favorable
choices of price increment for this baseline.
Introduction
Combinatorial auctions address the problem of allocating multiple distinct items among agents
who may view the items as complements or substitutes. In such auctions, agents can place bids
on entire packages of items in order to express complex preferences, leading to higher allocative
efficiency. Nevertheless, bidding in a combinatorial auction places a substantial cognitive burden
on agents, because the process of valuing even a single bundle can be a costly exercise [Kwasnica
et al., 2005; Parkes, 2006]. There is therefore great interest in developing iterative combinatorial
auctions, which help to guide the bidding process using price feedback, and in devising techniques
to limit the number of rounds needed to reach convergence (ideally in the dozens rather than
hundreds) [Petrakis et al., 2012; Bichler et al., 2017].
In this work, we propose to incorporate prior information on agent valuations into the auction
procedure in a principled manner, thereby achieving a low number of rounds in practice. We
cast the problem of combinatorial auction design in a Bayesian framework by developing a joint
generative model of agent valuations and market prices. Our generative model defines a likelihood
function for clearing prices given agent valuations. If these valuations are observed, the maximum
a posteriori (MAP) estimate for prices corresponds to market clearing prices. If they remain latent,
valuations can be marginalized away, weighed by their own likelihood according to observed bids.
This forms the basis for an auction scheme to solve the more general clearing problem where
valuations are unknown.
We consider settings where several indivisible items are up for sale, and agents have superadditive valuation functions over bundles of items (i.e., the items are pure complements). We
provide an auction implementation using item prices consisting of two components. In the knowledge update component, we maintain a Gaussian posterior over agent valuations, which is updated
as new bids are placed using assumed density filtering [Opper and Winther, 1998]. Prior information
can be incorporated into the auction by suitably initializing this component. The knowledge update
1
step presumes that agents follow myopic best-response strategies and bid on utility-maximizing
bundles at each round. Accordingly, we discuss an extension to our auction scheme using multiple
price trajectories that incentivizes this behavior in ex post Nash equilibrium. In the price update
component, we obtain an analytical expression for the clearing price objective, based on the
Gaussian model of valuations that the auction maintains. We establish that the form of the objective
is suitable for optimization using expectation maximization. By alternating the two components,
we obtain an intuitive and tractable auction scheme where agents place bids, knowledge over latent
valuations is updated given bids, and prices are updated given current knowledge of valuations.
For evaluation purposes, we first illustrate our auction on a stylized instance to gain insight
into the auction’s behavior under both unbiased and biased prior information. We then conduct
simulation experiments to compare our auction implementation against a combinatorial clock
auction that updates prices according to excess demand, which is the standard price update scheme
used in practice [Ausubel and Baranov, 2014]. The prior information in our Bayesian auction
is obtained by fitting a Gaussian process prior on a training sample of valuations. The baseline
clock auction is parametrized by a step size, or price increment. We find in our experiments
that our Bayesian auction is competitive against the strongest possible version of the baseline
auction, where the price increment is chosen separately for each instance to lead to the fewest
possible rounds. In particular, the Bayesian auction almost matches the strongest possible version
of baseline auction in terms of number of instances cleared, and uses fewer rounds on average
when it is able to clear.
Preliminaries
We consider a setting with m distinct and indivisible items, held by a single seller. The items are
to be allocated among n agents (i.e., buyers). We will use the notation [n] = {1, . . . , n}, so that
[n] and [m] denote the index sets of agents and items, respectively. There is unit supply of each
item. A bundle is a subset of the set of items. We associate each bundle with its indicator vector,
and denote the set of bundles as X = {0, 1}m . The component-wise inequality x ≤ x0 therefore
means that bundle x is contained in bundle x0 . The empty bundle is denoted by ∅.
Each agent i is single-minded so that its valuation can be encoded via a pair (xi , wi ) where
xi ∈ X is a bundle and wi ∈ R+ is a non-negative value (i.e., willingness to pay) for the bundle.
The agent’s valuation function vi : X → R+ is defined as vi (x) = wi if x ≥ xi , and vi (x) = 0
otherwise. In words, the agent only derives positive value if it acquires all the items in xi (which are
therefore complements), and any further item is superfluous. Our auction and results all extend to
agents with OR valuations, which are concise representations of super-additive valuations [Nisan,
2000].1 This is due to the fact that an agent with an OR valuation will behave and bid in our
auction exactly like a set of single-minded agents, under myopic best-response [Parkes, 1999].
Under super-additive valuations, items are pure complements, and complementarities are a key
motivation for using package bidding. For the sake of simplicity, however, we limit the exposition
to single-minded agents.
An allocation is represented as a vector of bundles y = (y1 , . . . , yn ), listing the bundle that
each agent obtains (possibly ∅). An allocation is feasible if the listed bundles are pairwise disjoint
(i.e., each item is allocated to at most one agent). We denote the set of feasible allocations by F.
The purpose of running a combinatorial auction is to find an efficient allocation of the items to
1 More formally, an OR valuation takes the form v(x) = max{v (x ) + v (x ) : x + x = x, x ∧ x = 0},
1 1
2 2
1
2
1
2
where v1 and v2 are themselves OR valuations or single-minded.
2
2
the agents, meaning an allocation that maximizes the total value
P to the 0agents. More formally,
a feasible allocation y ∈ F is efficient if y ∈ arg maxy0 ∈F i∈[n] vi (yi ). However, an iterative
auction proceeds via a price adjustment process, so prices will be our central object of study, rather
than allocations. The allocation in an iterative auction is adjusted according to agents’ responses to
prices.
Clearing Prices In the context of a combinatorial auction, we encode prices as a non-negative
function θ : X → R+ over the bundles. We assume that prices are normalized and monotone:
θ(∅) = 0, and θ(x) ≤ θ(x0 ) if x ≤ x0 . An iterative auction adjusts prices to balance demand and
supply. To formalize this notion, we need several additional concepts. We assume that agents have
quasi-linear utility, so that the utility to agent i of obtaining bundle x at prices θ is vi (x) − θ(x).
The indirect utility function provides the maximum utility that agent i can achieve, when faced
with prices θ, by choosing among bundles from X :
Vi (θ) = max { vi (x) − θ(x) : x ∈ X } .
(1)
Note that for single-minded agents, the indirect utility reduces to Vi (θ; xi , wi ) = (wi − θ(xi ))+ ,
where the notation (a)+ = max{a, 0} refers to the positive part of the argument. It will sometimes
be useful to make explicit the parametrization of the indirect utility on the agent’s type (xi , wi ), as
we have just done. The demand set of agent i is defined as Di (θ) = {x ∈ X : vi (x) − θ(x) = Vi (θ)}.
Similarly, the indirect revenue function provides the maximum revenue that the seller can achieve,
when faced with prices θ, by selecting among feasible allocations:
X
R(θ) = max
θ(yi ) : y ∈ F .
(2)
i∈[n]
The P
seller’s supply set consists of the feasible allocations that maximize revenue: S(θ) = {y ∈
F : i∈[n] θ(yi ) = R(θ)}.
We say that prices θ are clearing prices if there is a feasible allocation y such that, at prices θ,
the seller’s revenue is maximized, and each agent’s utility is maximized. Formally, we require the
following conditions: y ∈ S(θ) and yi ∈ Di (θ) for all i ∈ [n]. We say that the clearing prices θ
support allocation y.
It is a standard result that the set of allocations supported by any given clearing prices θ
coincides with the set of efficient allocations. (This is a special case of the Fundamental Theorems
of Welfare Economics [Mas-Colell et al., 1995, 16.C–D].) Moreover, Bikhchandani and Ostroy
[2002] have shown that clearing prices exist and coincide with the minimizers of the following
objective function, which corresponds to the linear programming dual of the problem of allocating
the items efficiently:
X
Vi (θ) + R(θ).
(3)
i∈[n]
This is a piece-wise linear, convex function of the price function θ. Importantly, this result is
guaranteed to hold only if the prices are an unrestricted function over the bundles (except for
non-negativity and normalization). In practice, it is common to use certain parametrizations for
2 This is in contrast to the goal of maximizing revenue. In auction design, one typically begins with an efficient auction,
which is then modified (e.g., using reserve prices) to achieve optimal revenue [Myerson, 1981]. We therefore consider the
problem of designing an efficient auction as more fundamental.
3
Initial
prices
Does
demand
meet
supply?
Observe
demand/
supply
yes
Output
allocation
no
Price
update
Knowledge
update
Figure 1: Bayesian iterative auction.
the prices. For instance, taking θ(x) = p · x for some vector p ∈ Rm
+ corresponds to using
linear prices (i.e., item prices). These parametrizations may not achieve the unrestricted minimum
in (3); in particular, linear clearing prices may not exist. We will use unrestricted prices in the
development of our auction, and postpone the question of price parametrization until needed to
achieve a practical implementation.
It is useful to view (3) as a potential function that quantifies how close prices θ are to supporting
an efficient allocation. Indeed, if some prices achieve a value of (3) that differs from the optimum
by an additive error of δ, then the agents (and seller) can be induced to accept an efficient trade
using transfers totaling δ. In the following, we will therefore refer to the function
X
Vi (θ) − R(θ)
(4)
U (θ; v) = exp −
i∈[n]
as the clearing potential for the valuation profile v = (v1 , ..., vn ), which will capture, in a formal
sense, how likely a price function θ is to clearing valuation profile v.
Iterative Auction and Incentives The goal of our paper is to design an iterative auction that
exploits the auctioneer’s prior knowledge over agent valuations in order to speed up the clearing
process. The auction proceeds over rounds. Agents report their demand at the current prices and, if
the market is not cleared, the information provided by the agents is used to update the knowledge
about their valuations. Candidate clearing prices are computed based on the updated knowledge,
and the procedure iterates. A schematic representation of the auction process is presented in
Figure 1. The knowledge update and price update components constitute the core of the auction
that must be implemented.
The correctness of our auction relies on the agents following a strategy of myopic best-response
bidding, meaning that each agent bids on a utility-maximizing bundle at each round. There is
evidence that myopic bidding may be a reasonable assumption in practice. For instance, in the FCC
broadband spectrum auction, jump bids were the exception [Cramton, 1997]. Nonetheless, a robust
auction design should incentivize agents to follow the appropriate strategies. For this purpose, we
can use an extension of our auction that maintains n + 1 price trajectories in order to compute
clearing prices when all agents are present, and when each agent is removed in turn. This allows
one to compute final VCG payments and bring myopic best-response bidding into an ex post Nash
4
equilibrium [Gul and Stacchetti, 2000; Bikhchandani and Ostroy, 2006]. The technique of using
multiple trajectories was previously used by Ausubel [2006] and Mishra and Parkes [2007] among
others. We will provide a more precise treatment of incentives in the formal description of our
auction mechanism.
Generative Model
The purpose of this section is to define a probabilistic relationship between prices and valuations
that will allow us to use the auctioneer’s prior knowledge over valuations to make inferences over
clearing prices. We write w = (w1 , . . . , wn ) and x = (x1 , . . . , xn ) for the vectors of agents’
values and bundles, and denote the probabilistic model as P (x, w, θ). Below, our convention
is that Q refers to distributions—possibly unnormalized—that form the building blocks of the
generative model, whereas P refers to the normalized distribution resulting from the generative
model. We represent the prior
Q knowledge of the auctioneer over agent valuations via the probability
density function Q(w) = i∈[n] Q(wi ).
The structure of our probability model is inspired by the work of Sollich [2002], who provides
a Bayesian interpretation of the support vector machine (SVM) objective. To establish a proper
relationship between prices and valuations, the key is to require that
P (x, w, θ) ∝ U (θ; x, w)Q(w),
(5)
where U (θ; x, w) is the clearing potential introduced in (4), adapted to single-minded valuations.
Under this joint probability model, we have that the posterior probability of prices θ takes the form
P (θ | x, w) ∝ U (θ; x, w).
(6)
Therefore, the maximum a posteriori (MAP) estimate maximizes U (θ; x, w), or equivalently
minimizes (3), and corresponds to clearing prices.
To establish that a probability model of the form (5) is possible—namely, that it can indeed
be normalized—we will derive it as the result of a generative model. This process may be of
independent interest as a means of generating agents together with market prices.
1. Draw prices θ according to Q(θ) ∝ exp [−R(θ)] .
2. For each agent i ∈ [n]:
– Draw wi ∈ R+ from Q(wi ).
– Draw xi ∈ X from
Q(xi | wi , θ) =
1
exp [−Vi (θ; xi , wi )] .
2m
– With probability 1 − ν(wi , θ), restart from step 1, where
X
ν(w, θ) =
Q(x | w, θ).
x∈X
Above, we must ensure that the Q(θ) prior normalizes; this is the case under our assumption that
the domain of θ falls within the positive orthant. The prior distribution Q(wi ) on value wi is left
5
free in the model, so that it may correspond to the auctioneer’s prior in practice. Note that the
bundle likelihood Q(xi | wi , θ) is not normalized; because Vi (θ; xi , wi ) ≥ 0, summing over the set
of bundles leads to the aggregate probability mass ν(wi , θ) ≤ 1. Rather than normalizing by this
quantity, we use the “remaining probability” 1 − ν(wi , θ) of not drawing any bundle to restart the
process. Because of the possible restart, the agent types (bundle-value pairs) and clearing prices
are not independent in the overall generative distribution. In particular, the number of agents n in
the economy affects the distribution of prices.
The following proposition confirms that our model satisfies (5). All proofs are deferred to the
appendix.
Proposition 1. The generative model of agent types and prices takes the form
P (x, w, θ) ∝ U (θ; x, w)Q(w).
(7)
The generative process defines a probability distribution over prices once valuations are observed,
but during the auction the valuations remain latent, and must be inferred based on observed bids
placed across rounds. Under appropriate incentives, the auctioneer can infer valuations assuming
that the agents follow myopic best-response bidding. However, if there are any bidding errors or
corruption in communication, assuming exact best-response can cause singularities in the inference
process (e.g., there may be no valuation consistent with all observed bids). To guard against
this, our mechanism will integrate bids as if they were generated from the following stochastic
model: Let bi ∈ {−1, +1} be an indicator variable to denote whether the agent bids on bundle xi
(bi = +1) or not (bi = −1); the latter is equivalent to bidding on ∅. If the cost of bundle xi is ci ,
then the choice of bid follows the probability distribution
Q(bi = +1 | ci , wi ) ∝ Φ(β(wi − ci )),
(8)
where Φ is the cumulative distribution function of the standard normal, and β > 0 is a scalar
parameter. This is known as the probit variant of approximate best-response, which arises from
random utility models [Train, 2009]. As β → ∞, we obtain exact best-response: the agent bids
on xi if and only if this bundle yields positive utility under bundle cost ci . Using a large (but
finite) β allows the auctioneer to model agents as essentially following a best-response strategy,
but occasionally allowing for bidding errors or inconsistencies.
Auction Description
Our auction proceeds over rounds; we use k to denote the current round, and ` to index the rounds
up to k. At each round, prices are updated, which imputes a cost to each agent’s bundle. Let c`i be
the cost of agent i’s desired bundle xi in round ` according to the current prices. The prices at each
round should not be confused with the latent clearing prices θ, which we are trying to compute as
a MAP estimate of the generative model. Given its value wi and the bundle cost c`i , agent i places
bid b`i ∈ {−1, +1} in round `.
We write c` = (c`1 , . . . , c`n ) to denote the vector of bundle costs in round `, and c(`) =
1
(c , . . . , c` ) to denote the vector of costs up to round `. For brevity we also write c = c(k) to
denote the vector of all costs up to the current round. We use the notation b` , b(`) , and b to denote
the analogous vectors of bids. The bundle costs and agent bids in a round depend on the current
prices, which themselves depend on the bids placed in all earlier rounds. Assuming that the first
round prices are zero, we have the following intuitive posterior over bids and costs.
6
Lemma 1. The posterior distribution over bids and costs placed during the auction, given the
generated prices and agent types, is given by
P (b, c | x, w, θ) = Q(b | c, w)
k
Y
P (c` | b(`−1) , c(`−1) , x),
`=1
where b` and c` are the vectors of agent bids and costs at round `, and b(`) and c(`) are the vectors
of agent bids and costs up to round `.
We see that the posterior over bids and costs does not depend on the underlying clearing prices
θ, conditional on agent types (x, w), because the initial prices and agent valuations fully determine
how the auction proceeds. More specifically, the posterior decomposes into the likelihood of the
observed bids Q(b|c, w) under stochastic model (8), times the likelihood of the observed sequence
of costs. The latter does not involve w, because current round prices are fully determined by the
bids and costs of previous rounds. Our auction is based on the following characterization of the
overall posterior over prices and agent values.
Proposition 2. The posterior distribution of latent variables (w, θ) given observed variables
(b, c, x) takes the form
P (w, θ | b, c, x) ∝ Q(b | c, w)Q(w) · Q(x | w, θ)Q(θ),
where the proportionality constant depends solely on (b, c, x).
The posterior factors into two terms, which motivates our auction procedure. The first term,
Q(b | c, w)Q(w), can be construed as a posterior over agent values given bids and costs, since
Q(w) corresponds to a prior and Q(b | c, w) corresponds to a likelihood. We will maintain an
approximation P̂ (w) to this posterior over agent values and update it as new bids are placed in
response to bundle costs. This is the knowledge update component.
Recalling (6), the second term Q(x | w, θ)Q(θ) ∝ U (θ; x, w) in the posterior corresponds
(up to a constant factor) to the price posterior given knowledge of agent types. This leads to an
approximation to the price posterior when values remain latent:
Z
P (θ | b, c, x) ≈ dw P̂ (w) · Q(x | w, θ)Q(θ).
(9)
Here we have simply integrated the full posterior as given by Proposition 2, and made use of our
approximation to the value posterior. (We have also omitted the normalization constant.) In the
context of an auction, we quote a specific price function to the agents, rather than a distribution
over prices. Therefore, in the price update component, we will compute and quote the MAP
estimate of prices by maximizing (9). Note that if we have exact knowledge of agent values (i.e.,
P̂ is a point mass), computing the MAP estimate is equivalent to minimizing (3) and to computing
clearing prices, as one would expect.
Knowledge Update We observe that the value posterior Q(b | c, w)Q(w) consists of a separate
Qk
factor for each agent i, taking the form Q(wi ) `=1 Q(b`i | c`i , wi ), where k is the current round.
This represents a posterior on agent i’s individual value wi . To obtain an approximation to this
posterior, we use an online scheme known as assumed density filtering, which is a special case of
expectation propagation [Cowell et al., 1996; Minka, 2001; Opper and Winther, 1998]. Under this
7
approach, a Gaussian distribution P̂ (wi ; mi , σi2 ) is used to approximate the posterior; its mean mi
and variance σi2 are updated at each round given the bidding observations. The Gaussian is initially
set to approximate the prior Q(wi ) via moment matching: mi and σi2 are set to the mean and
variance of this prior Q. In each later round ` = 1, . . . , k the posterior is again updated by matching
the moments of Q(b`i | c`i , wi )P̂ (wi ; mi , σi2 ), which is an online update. Using moment matching
as an approximation is justified by the fact that it corresponds to minimizing the Kullback-Leibler
divergence KL(QkP̂ ) under the constraint that P̂ is Gaussian.
Due to the form of the likelihood (8) and the fact that P̂ is Gaussian, the update has a closed-form
solution [see Williams and Rasmussen, 2006, p. 74]:
σi4 β 2 N (zi )
N (zi )
b` σ 2 βN (zi )
, σi2 ← σi2 −
z
+
mi ← mi + p i i 2
i
1 + σi2 β 2 Φ(zi )
Φ(zi )
1 + σi β 2 Φ(zi )
where N and Φ are the probability density and cumulative
p distribution functions of the standard
normal, respectively, and where zi = b`i β(mi − c`i )/ 1 + σi2 β 2 . Recall that β is a positive
parameter characterizing the extent to which the auctioneer assumes that agents make mistakes
in placing best-response bids. Since β is positive, the mean mi is updated in the direction of the
bid b`i . On the other hand, the variance σi2 is strictly decreasing, thus ensuring that the beliefs over
bidder values converge to a point mass in the limit as the rounds progress, and that the auction
converges to a final vector of prices.
Price Update To implement the price update component we need an algorithm to maximize the
approximate posterior (9). This posterior factors into Q(θ) and a term for each agent i, which we
denote as
Z
Li (θ) = dwi P̂i (wi ; mi , σi2 )Q(xi | wi , θ).
Because Qi (xi | wi , θ) has an exponential form, and P̂i is a Gaussian, this integral has a closed
form solution (see appendix). Let qi ∈ {0, 1} be a binary auxiliary variable. We have Li (θ) ∝
P1
qi =0 Li (θ, qi ), where we define
mi − θ(xi )
σi2
Li (θ, 1) = Φ
, and
− σi exp θ(xi ) − mi +
σi
2
θ(xi ) − mi
Li (θ, 0) = Φ
.
σi
Here Φ is again the cumulative distribution function of the standard normal. To summarize, taking
the log of (9), the objective we seek to maximize with respect to θ is
log Q(θ) +
X
log
1
X
Li (θ, qi ).
(10)
qi =0
i∈[n]
Now, because Φ is log-concave, both Li (θ, 0) and Li (θ, 1) are log-concave in θ. Ignoring the first
term for an instant, we see that the objective consists of a sum of mixtures Li (θ) of log-concave
functions for each agent. This kind of objective is well-suited to optimization using the expectationmaximization (EM) algorithm [Dempster et al., 1977]. The qi amount to “latent” variables and the
“marginal” likelihood appears within the objective (10). (However, we do not claim any intuitive
interpretation for the latent qi —they are simply used to fit the objective into the mold of EM.)
8
The remaining term is log Q(θ), which is −R(θ) up to an additive constant. Recalling the
definition of the seller’s indirect utility (2), we see that this term is very complex for unrestricted
θ, because the set of feasible allocations F has a very complicated structure. To address this we
m
will impose a linear structure on prices:
Pmθ(x) = p · x where p ∈ R+ denote item prices. With
this parametrization, we have R(θ) = j=1 pj , because any allocation that allocates all the items
maximizes revenue under linear prices. The log Q(θ) term therefore becomes a linear term in p,
which is straightforward to incorporate within the EM algorithm.
Incentive Compatibility Our auction converges to an efficient allocation and clearing prices
under myopic best-response bidding, but to ensure that agents follow such a strategy, they must
be incentivized to do so. The standard technique used to achieve this in the literature on iterative
auctions is to charge VCG payments upon completion [Gul and Stacchetti, 2000; Bikhchandani
and Ostroy, 2006]. But whereas VCG payments (together with an efficient allocation) induce
truthful bidding in dominant strategies for single-shot auctions, weaker results hold for iterative
auctions.
A strategy profile constitutes an ex post Nash equilibrium if no agent would like to deviate
from its strategy, holding the others’ strategies fixed, even with knowledge of the private valuations
of the other agents. Gul and Stacchetti [2000] prove the following result:
Theorem 1 (Gul and Stacchetti, 2000). Truthful myopic best-response bidding is an ex post Nash
equilibrium in an iterative auction that myopically-implements the VCG outcome.
Above, the VCG outcome refers to an efficient allocation along with VCG payments, and an auction
myopically-implements this outcome if the auction converges to it under myopic best-response
bidding. The reason that truthfulness only holds in ex post Nash equilibrium, rather than dominant
strategies, is that profitable deviations may exist if another agent bids in a manner inconsistent
with any valuation.
Our auction already computes the efficient allocation under these conditions by virtue of
converging to clearing prices. To compute VCG payments, we can simply extend or auction
drawing on the idea of multiple price trajectories: the usual trajectory traced by our auction,
and the trajectories that would result if each agent were removed in turn. This technique was
previously used by Ausubel [2006] and Mishra and Parkes [2007]. In this extended design, at
each round, agents place bids against n + 1 different price vectors. Upon completion, the agents
place last-and-final bids for their allocated bundles, thereby communicating their value for the
allocations; importantly, agents do not need to communicate values for any bundles they did not
win. This information is precisely what is needed to compute VCG payments [see, e.g., Parkes and
Ungar, 2000].
Empirical Evaluation
In this section we evaluate our Bayesian auction design with two different kinds of experiments:
a small experiment to illustrate the behavior of our auction under biased and unbiased prior
information, and a larger-scale experiment to compare our auction against a competitive baseline.
Our simulations are conducted in Matlab. In all our experiments, we assume that agents best
respond to the proposed prices (i.e., they always bid on their most profitable bundle), and that
the auctioneer considers their bids as if they were generated from the response model presented
in (8) with β = 10. However, simulations where real bids follows (8) with β = 4 provide results
9
5
Number of rounds
4
3
2
1
Unbiased
Biased
0
0
5
10
15
20
Standard deviation
Figure 2: Auction rounds in LLG.
similar to the ones presented. For the price update, the objective (10) is maximized using the
“active-set” algorithm in Matlab. To avoid numerical singularities we place a lower bound of 0.01
on the variance of valuation estimates.
LLG Experiments We consider an instance of the Local-Local-Global (LLG) domain [Ausubel
and Milgrom, 2006], which has been considered several times in the combinatorial auctions
literature. There are two items and three single-minded agents. Two of the agents are local,
meaning that they are interested in just one item, respectively the first and second item. The last
agent is global in the sense that it is interested in both items.
The two local agents have a value of 4 for their respective items, and the global one has a value
of 10 for both. The items are efficiently allocated when they are both assigned to the global agent,
and linear prices are expressive enough to clear the market (e.g., we can use a price of 4 for each
item).
We assume that the auctioneer has accurate knowledge of the local agents’ values: Q(w1 ) =
N (w1 ; m1 = 4, σ12 = 0.01) and Q(w2 ) = N (w2 ; m2 = 4, σ22 = 0.01). Note the very low
variance, reflecting certainty. We test how different kinds of prior knowledge over the global
agent’s value affect the number of rounds that the Bayesian auction takes to clear the market. In
the first case we assume unbiased prior knowledge: m3 = 10. In the second case we assume that it
is biased below: m3 = 4. Here, the auctioneer tends to allocate to the local agents instead of the
global one.
Figure 2 plots the number of rounds that our Bayesian auction takes to clear the market against
the standard deviation σ3 of the prior over the global agent’s value. We see that, in the unbiased
scenario, the number of rounds monotonically increases as the standard deviation grows. This
can be easily explained since increasing the standard deviation only adds noise to the exact prior
estimate. In the biased scenario, we have an optimal range of standard deviations between 2 and
5. If the standard deviation is too low, the auction needs many observations to correct the biased
prior. If it is too high, the low confidence leads to many rounds because the auctioneer needs to
refine its estimate of the value regardless of the bias.
10
0.6
0.4
0.2
0
Paths
Regions Arbitrary Scheduling
Figure 3: Cleared instances in CATS.
Number of rounds
0.8
100
Number of rounds
Fraction of cleared instances
SIO
SAOc
Bayes
100
Paths
Regions
Arbitrary
Scheduling
50
0
50
0
SIO
SAOr Bayes
SIO
SAOr Bayes
Figure 4: Auction rounds in CATS.
CATS Experiments For our second set of experiments, we generate instances using the Combinatorial Auction Test Suite (CATS), which offers four generator distributions: paths, regions,
arbitrary, and scheduling [Leyton-Brown et al., 2000]. These are meant to model realistic
domains such as truck routes, real estate, and pollution rights. We generate 1000 instances from
each distribution, each with 12 items and 10 single-minded agents.
The instances are generated as follows. First, 100 input files with 1000 bids each are generated.
Each input file is partitioned into a “training set” and “test set”, each with 500 bids. From the
test set, 10 bids (representing 10 single-minded agents) are sampled uniformly at random. The
training set is used to fit the prior knowledge of our Bayesian auction. Specifically, we fit a linear
regression model of bundle value according to items contained, using a Gaussian process with a
linear covariance function, leading to Gaussian prior knowledge. The fit is performed using the
publicly available GPML Matlab code [Williams and Rasmussen, 2006].
As a baseline we implemented a standard linear-price auction scheme closely related to the
combinatorial clock auction [Ausubel and Baranov, 2014]. The scheme is parametrized by a
positive step√
size τ . At each round `, the price of an item is incremented by its excess demand,
scaled by τ / `. The excess demand of an item is the number of bidded bundles that contain it,
minus the number of units of the item offered by the seller at current prices. This can be viewed as
a subgradient
descent method for computing clearing prices, for which a step size proportional to
√
1/ ` yields the optimal worst-case convergence rate [Bertsekas, 2015, Chap. 3].
Both the Bayesian auction and the baseline subgradient auction use linear prices, but these
may not be expressive enough to support an efficient allocation in instances generated by CATS.
We therefore set a limit of 100 rounds for each auction run, and record reaching this limit as a
failure to clear the market.
On each instance we run a single Bayesian auction, and 100 subgradient auctions with the
step size uniformly spanning the interval from zero to the maximum agent value. This leads to
several baseline results. The standard instance optimized (SIO) results refer to the performance of
the baseline when using the optimal step size for each instance. The standard average clearingoptimized (SAOc) results refer to the performance of the baseline under the fixed step size that
leads to the best clearing performance on average, for each valuation domain. Analogously, the
standard average round-optimized (SAOr) results refer to baseline performance under the step
11
size leading to lowest average number of rounds. For each instance, the step size that leads to
the lowest number of rounds naturally leads to the best clearing rate. But the fixed step sizes that
optimize these two criteria on average may be different. Note that SIO is an extremely competitive
baseline, since it is optimized for each instance; a priori, we hoped to be competitive against it, but
did not expect to beat it. The SAOc and SAOr baselines reflect more realistic performance that
could be achieved in practice.
We first consider clearing performance. The results are reported in Figure 3. We find that the
Bayesian auction is competitive with SIO on all four domains, and that it always outperforms
SAOc. In fact, our auction even outperforms SIO on the arbitrary domain. This means that it
was able to clear some instances that the subgradient auction could not clear within 100 rounds at
any step size. In general, there is good agreement between our Bayesian auction and the baselines
on which instances can be cleared or not according to the 100-round criterion. This indicates that
failure to clear is typically a property of the instance rather than the algorithm.
Figure 4 summarizes the distributions of rounds needed to achieve clearing using box plots. To
enable fair comparisons, for this plot we only considered instances that were cleared by all auction
types: the Bayesian auction, SAOr, and SIO. This yields 770 valid instances for paths, 910 for
regions, 624 for arbitrary and 955 for scheduling. The mean rounds for the Bayesian
auction, SAOr, and SIO are always statistically different at the 0.01 level. We see from the plot
that, in terms of the median number of rounds, the Bayesian auction clearly outperforms SAOr,
but also remarkably outperforms SIO. Furthermore, the distribution of rounds for the Bayesian
auction has a much lower spread than the baselines. It is able to clear almost all instances in less
than 10 rounds.
Conclusion
In this work we developed a Bayesian clearing mechanism for combinatorial auctions that allows
one to incorporate prior information into the auction process in a principled manner. Our auction
mechanism is based on a joint generative model of valuations and prices such that clearing prices
are the MAP estimate given observed valuations. Our empirical evaluation confirmed that our
Bayesian mechanism performs remarkably well against a conventional clock auction scheme, in
terms of rounds to convergence. Our auction’s performance simply relies on reasonable priors for
valuations, rather than careful tuning of price increments.
We believe that the Bayesian perspective on auction design developed in this paper could be
leveraged to improve other aspects beyond rounds to convergence. For instance, the Bayesian
paradigm offers a principled way to select hyperparameters [MacKay, 1992]; in our context, this
could be used to choose the right structure of prices (linear, nonlinear) to clear the market, a priori.
The knowledge update component could also form the basis of more refined activity rules; for
instance, one could reject bids that are highly unlikely, given the valuation posterior based on
previous bids. We intend to pursue these directions in future work.
12
References
Lawrence M. Ausubel. An efficient dynamic auction for heterogeneous commodities. The
American Economic Review, 96(3):602–629, 2006.
Lawrence M. Ausubel and Oleg V. Baranov. A practical guide to the combinatorial clock auction.
Technical report, University of Maryland, 2014.
Lawrence M Ausubel and Paul Milgrom. The lovely but lonely vickrey auction. In Peter Cramton,
Yoav Shoham, and Richard Steinberg, editors, Combinatorial auctions. The MIT Press, 2006.
Dimitri P. Bertsekas. Convex optimization algorithms. Athena Scientific, 2015.
Martin Bichler, Zhen Hao, and Gediminas Adomavicius. Coalition-based pricing in ascending
combinatorial auctions. Information Systems Research, 2017. Forthcoming.
Sushil Bikhchandani and Joseph M. Ostroy. The package assignment model. Journal of Economic
Theory, 107(2):377–406, 2002.
Sushil Bikhchandani and Joseph M Ostroy. Ascending price vickrey auctions. Games and
Economic Behavior, 55(2):215–241, 2006.
Robert G. Cowell, A. Philip Dawid, and Paola Sebastiani. A comparison of sequential learning
methods for incomplete data. Bayesian Statistics, 5:533–542, 1996.
Peter Cramton. The FCC spectrum auctions: An early assessment. Journal of Economics &
Management Strategy, 6(3):431–495, 1997.
Arthur P. Dempster, Nan M. Laird, and Donald B. Rubin. Maximum likelihood from incomplete
data via the EM algorithm. Journal of the Royal Statistical Society, Series B, pages 1–38, 1977.
Faruk Gul and Ennio Stacchetti. The english auction with differentiated commodities. Journal of
Economic Theory, 92(1):66–95, 2000.
Anthony M. Kwasnica, John O. Ledyard, Dave Porter, and Christine DeMartini. A new and
improved design for multiobject iterative auctions. Management Science, 51(3):419–434, 2005.
Kevin Leyton-Brown, Mark Pearson, and Yoav Shoham. Towards a universal test suite for
combinatorial auction algorithms. In Proceedings of the 2nd ACM Conference on Electronic
Commerce, pages 66–76. ACM, 2000.
David J. C. MacKay. Bayesian interpolation. Neural computation, 4(3):415–447, 1992.
Andreu Mas-Colell, Michael D. Whinston, and Jerry R. Green. Microeconomic theory. Oxford
university Press, 1995.
Thomas P. Minka. A family of algorithms for approximate Bayesian inference. PhD thesis,
Massachusetts Institute of Technology, 2001.
Debasis Mishra and David C. Parkes. Ascending price vickrey auctions for general valuations.
Journal of Economic Theory, 132(1):335–366, 2007.
Roger B Myerson. Optimal auction design. Mathematics of operations research, 6(1):58–73,
1981.
13
Noam Nisan. Bidding and allocation in combinatorial auctions. In Proceedings of the 2nd ACM
Conference on Electronic Commerce, pages 1–12. ACM, 2000.
Manfred Opper and Ole Winther. A Bayesian approach to online learning. Online Learning in
Neural Networks, pages 363–378, 1998.
David C. Parkes. iBundle: an efficient ascending price bundle auction. In Proceedings of the 1st
ACM Conference on Electronic Commerce, pages 148–157. ACM, 1999.
David C. Parkes. Iterative combinatorial auctions. In Peter Cramton, Yoav Shoham, and Richard
Steinberg, editors, Combinatorial auctions. The MIT Press, 2006.
David C. Parkes and Lyle H. Ungar. Preventing strategic manipulation in iterative auctions:
Proxy agents and price-adjustment. In Proceedings of the 17th AAAI Conference on Artificial
Intelligence, pages 82–89, 2000.
Ioannis Petrakis, Georg Ziegler, and Martin Bichler. Ascending combinatorial auctions with
allocation constraints: On game theoretical and computational properties of generic pricing
rules. Information Systems Research, 24(3):768–786, 2012.
Peter Sollich. Bayesian methods for support vector machines: Evidence and predictive class
probabilities. Machine learning, 46(1-3):21–52, 2002.
Kenneth E. Train. Discrete choice methods with simulation. Cambridge University Press, 2009.
Christopher K. Williams and Carl E. Rasmussen. Gaussian Processes for Machine Learning. The
MIT Press, 2006.
14
Proof of Proposition 1
Q
Q
Proof. Let ν(w, θ) = i ν(wi , θ) and Q(w) = i Q(wi ).
We denote the probability of a restart as
Z Z
r =1−
dw dθ ν(w, θ)Q(w)Q(θ),
(11)
Q
where dθ is a shorthand for x∈X dθ(x). (Note that any price function θ can be seen as a 2m dimensional vector of positive real numbers.) The probability that (x, w, θ) is drawn after ` restarts
will then be
P (x, w, θ, `) = Q(x | w, θ)Q(w)Q(θ) · r` .
(12)
Thus,
P (x, w, θ)
=
∞
X
Q(x | w, θ)Q(w)Q(θ) · r`
`=0
=
(13)
Q(x | w, θ)Q(w)Q(θ)
.
1−r
Recalling equation (4), we have that
U (θ; x, w) ∝ Q(x | w, θ)Q(θ).
(14)
P (x, w, θ) ∝ U (θ; x, w)Q(w).
(15)
We can conclude that
Proof of Lemma 1
Proof.
P (b, c | x, w, θ)
=
k
Y
P (b` , c` | b(`−1) , c(`−1) , x, w, θ)
`=1
=
k
Y
Q(b` | c` , w)P (c` | b(`−1) , c(`−1) , x)
`=1
= Q(b | c, w)
k
Y
P (c` | b(`−1) , c(`−1) , x)
`=1
The third line follows because bids only depend on bundle costs and values according to (8), and
costs are independent of values and clearing prices given previous round bids and costs, which
fully determine the current round prices.
15
Proof of Proposition 2
Proof. We have that
P (w, θ | b, c, x) ∝ P (b, c | x, w, θ)P (x, w, θ)
∝ Q(b | c, w)P (x, w, θ)
∝ Q(b | c, w)Q(x | w, θ)Q(w)Q(θ).
The second line follows from Lemma 1, and the last line from Proposition 1.
Price Likelihood Derivation
We aim to solve the integral
Z
dwi P̂i (wi ; mi , σi2 )Q(xi | wi , θ)
=
1
2m
Z
dwi N (wi |mi , σi2 ) exp [−(wi − θ(xi ))+ ],
where N (wi |mi , σi2 ) is the probability density function of the normal distribution with mean mi
and variance σi2 . We can rewrite the integral as
1
2m
Z
+∞
dwi N (wi |mi , σi2 ) exp [−(wi − θ(xi ))]
θ(xi )
+
1
2m
Z
θ(xi )
dwi N (wi |mi , σi2 ).
−∞
Now, since
N (wi |mi , σi2 ) exp [−(wi − θ(xi ))]
= N (wi |mi − σi2 , σi2 ) exp [−(mi − σi2 /2 − θ(xi ))],
and given that Φ(−z) = 1 − Φ(z), we can rewrite our integral as
1
mi − θ(xi )
1
θ(xi ) − mi
σi2
Φ
+
Φ
−
σ
.
exp
θ(x
)
−
m
+
i
i
i
2m
σi
2
2m
σi
16
| 2cs.AI |
arXiv:1512.02170v2 [math.RT] 11 Jan 2016
The Littlewood-Richardson rule for wreath
products with symmetric groups and the quiver
of the category F ≀ FIn
Itamar Stein∗
Department of Mathematics
Bar Ilan University
52900 Ramat Gan
Israel
Abstract
We give a new proof for the Littlewood-Richardson rule for the wreath
product F ≀Sn where F is a finite group. Our proof does not use symmetric
functions but more elementary representation theoretic tools. We also
derive a branching rule for inducing the natural embedding of F ≀ Sn to
F ≀ Sn+1 . We then apply the generalized Littlewood-Richardson rule for
computing the ordinary quiver of the category F ≀ FIn where FIn is the
category of all injective functions between subsets of an n-element set.
1
Introduction
Let G be a finite group and let H ≤ G be a subgroup. By Maschke’s theorem,
group algebras (over C) are semisimple so every group representation is a finite
direct sum of irreducible ones. A basic question in the representation theory
of finite groups is the following one: Let V be some H-representation, what
∗ This paper is part of the author’s PHD thesis, being carried out under the supervision of
Prof. Stuart Margolis. The author’s research was supported by Grant No. 2012080 from the
United States-Israel Binational Science Foundation (BSF).
1
is the decomposition of the inducted representation IndG
H V into irreducible Grepresentations? Alternatively, let U be some G-representation, what is the
decomposition of the restriction ResG
H U into irreducible H-representations? By
Frobenius reciprocity, answering one of these questions essentially answers the
other one. Moreover, since both induction and restriction are additive, it is
enough to consider the case where U or V are irreducible representations. Even
considering this reduction, the question, in general, is very difficult. If G = Sn
the answer is known for certain natural choices of H and these solutions are often
called branching rules. The most classical case is where H = Sn−1 viewed as
the subgroup of all permutations that fix n. An important generalization is the
Littlewood-Richardson rule which gives the answer for the case H = Sk × Sn−k .
Let F and G be finite groups such that G acts on the left of a finite set X. We
denote by F ≀X G the wreath product of F and G. The representation theory
of F ≀X G is a well-studied subject (see [3] and [6, Chapter 4]) and the case
G = Sn with the natural action on {1, . . . , n} is of special importance. Finding
generalizations for the branching rules is a natural question. The “classical”
branching rule for inducing from F ≀ Sn to F ≀ Sn+1 was found by Pushkarev
[10]. In this paper we generalize the Littlewood-Richardson rule to the group
F ≀ Sn . After the present paper was already circulating, we became aware of the
paper [5] by Ingram, Jing and Stitzinger, where the same result was obtained
using symmetric functions. However, our approach is different. We use only
elementary representation theoretic tools and base our proof on the explicit
description of the irreducible representations of F ≀ Sn . In Section §5 we use the
generalized Littlewood-Richardson rule to retrieve Pushkarev’s result.
Then we turn to give an application to the representation theory of a natural
family of categories. Denote by FI the category of finite sets and injective
functions. The representation theory of FI is currently under active research
which was initiated in [4]. There is also research on the representation theory
of the wreath product F ≀ FI (see [8, 11, 13]). We will be interested here in the
finite version of this category. We denote by FIn the category of all subsets of
{1, . . . , n} and injective functions. In Section §6 we will give a description of the
ordinary quiver of the algebra of F ≀ FIn . The case where F is the trivial group
was originally done by [2] and a simple proof was later given in [9]. For the
general case, we imitate the method of [9] but where they use usual branching
rule for Sn we will use the generalization for F ≀ Sn .
2
2
2.1
Preliminaries
Wreath product
Throughout this paper F and G will be finite groups such that G acts on the
left of some finite set X. An element f ∈ F X is a function from X to F . Note
that F X is a group, multiplication being defined componentwise. We can also
define a left action of G on F X by
(g ∗ f )(x) = f (g −1 x).
It is easy to verify that this is indeed a left action.
Definition 2.1. The wreath product of F with G denoted F ≀X G is the semidirect product F X ⋊ G. In other words, it is the set F X × G with multiplication
given by
(f, g) · (f ′ , g ′ ) = (f (g ∗ f ′ ), gg ′ ).
If H ≤ G is a subgroup, we can restrict the action on X to H and get the group
F ≀X H which is a subgroup of F ≀X G.
Clearly, Sn acts on {1, . . . , k} (for n ≤ k) by permuting the first n elements and
fixing the other ones. We refer to this action as the standard action of Sn on
{1, . . . , k}. In this case we denote the wreath product with F by F ≀k Sn . The
focus of this paper will be the case where k = n and we will simply denote this
group by F ≀ Sn . There is a very natural way to think of this group. Recall
that we can identify Sn with the group of permutation matrices, that is, a
permutation π ∈ Sn can be identified with an n × n matrix A where
Ai,j
1 π(j) = i
=
.
0 otherwise
Similarly, we can identify F ≀ Sn with a group of matrices, but here the non-zero
entries can be any element of F . In other words, the tuple (f, π) is identified
with the n × n matrix A where
Ai,j
f (i) π(j) = i
.
=
0
otherwise
The multiplication of F ≀ Sn is then identified with matrix multiplication when
3
one assumes
0 · a = a · 0 = 0,
a+0=0+a=a
for every a ∈ F .
The following fact will be of use (see [3, Proposition 2.1.3] for proof).
Lemma 2.2 (Distributivity of the wreath product). Let G1 and G2 be groups
acting on disjoint sets X1 and X2 respectively, so G1 × G2 acts on the disjoint
˙ 2 . Then
union X = X1 ∪X
(F ≀X1 G1 ) × (F ≀X2 G2 ) ∼
= F ≀X (G1 × G2 ).
2.2
Complex group representations
We only consider representations over C in this paper. A G-representation is a
pair (U, ρ) where U is a finite dimensional vector space and ρ : G → End(U )
is a group homomorphism. This is equivalent to an action of G on the vector
space U by linear transformations. We will sometimes omit the homomorphism
and say that U is a G-representation. For u ∈ U we will usually write g · u
or even gu instead of ρ(g)(u). Let U and V be two G-representations. We
∼ V ) if there is a vector space
say that U is isomorphic to V (and write U =
isomorphism T : U → V such that T (g · u) = g · T (u) for every g ∈ G and u ∈ U .
The direct sum U ⊕ V of two G-representations is again a G-representation
according to g · (u + v) = gu + gv. A subvector space V ⊆ U is called a
subrepresentation if it is closed under the action of G, that is, g · v ∈ V for
all g ∈ G and v ∈ V . A non-zero G-representation U is called irreducible
if its only subrepresentations are 0 and U . We denote the set of irreducible
representations of G (up to isomorphism) by Irr G. It is well known that every
G-representation is a finite direct sum of irreducible representations and that
the number of different irreducible G-representations (up to isomorphism) is the
number of conjugacy classes of G. We denote the trivial representation of any
group G by trG . Recall that if V is a G-representations, then V ∗ = Hom(V, C)
is also a G-representation with operation (g · ϕ)(v) = ϕ(g −1 v). Let U and V be
G-representations. The inner tensor product U ⊗ V is again a G-representation
with action defined by g · (u ⊗ v) = gu ⊗ gv and extending linearly. Now,
assume that U1 and U2 are G1 and G2 -representations respectively. The outer
tensor product U1 ⊗ U2 of U1 and U2 is the (G1 × G2 )-representation where
(g1 , g2 )·(u1 ⊗u2 ) = (g1 u1 )⊗(g2 u2 ). Although the two types of tensor product can
4
be distinguished by the context we prefer using different notation for outer tensor
product, denoting it by ⊠. Likewise, the simple tensors of U ⊠V will by denoted
by u ⊠ v. It is well known that Irr(G1 × G2 ) = {U ⊠ V | U ∈ Irr G1 , V ∈ Irr G2 }.
Another simple observation will be important.
Lemma 2.3. Let U1 and V1 (U2 and V2 ) be G1 (respectively, G2 )-representations.
Then
(U1 ⊠ U2 ) ⊗ (V1 ⊠ V2 ) ∼
= (U1 ⊗ V1 ) ⊠ (U2 ⊗ V2 )
as (G1 × G2 )-representations.
Proof. Define T : (U1 ⊠ U2 ) ⊗ (V1 ⊠ V2 ) → (U1 ⊗ V1 ) ⊠ (U2 ⊗ V2 ) by
T ((u1 ⊠ u2 ) ⊗ (v1 ⊠ v2 )) = (u1 ⊗ v1 ) ⊠ (u2 ⊗ v2 )
which clearly extends to a vector space isomorphism and also
T ((g1 , g2 ) · ((u1 ⊠ u2 ) ⊗ (v1 ⊠ v2 ))) = T ((g1 u1 ⊠ g2 u2 ) ⊗ (g1 v1 ⊠ g2 v2 ))
= (g1 u1 ⊗ g1 v1 ) ⊠ (g2 u2 ⊗ g2 v2 )
= (g1 , g2 ) · T ((u1 ⊠ u2 ) ⊗ (v1 ⊠ v2 ))
as required.
The character χU of the G-representation (U, ρ) is the function χU : G → C
defined by χU (g) = trace(ρ(g)). Recall that the multiplicity U ∈ Irr G as an
irreducible constituent in some G-representation V is given by the inner product
hχU , χV i =
1 X
χU (g)χV (g).
|G|
g∈G
Recall also that χV ∗ (g) = χV (g) and χU⊠V ((g1 , g2 )) = χU (g1 )χV (g2 ). In order
to simplify notation, we will usually omit the χ and write U also for the character
of U . Hence the above inner product will be written as
hU, V i =
1 X
U (g)V (g).
|G|
g∈G
5
2.3
Restriction and induction
Let (U, ρ) be a G-representation and let H ≤ G be a subgroup. The restriction
G
of (U, ρ) to H denoted (ResG
H U, ResH ρ) is an H-representation defined by
ResG
H ρ(h)(u) = ρ(h)(u)
that is, restricting the homomorphism to the subgroup H. Note that dim ResG
HU
= dim U and if U is an irreducible G-representation then ResG
H U does not have
to be an irreducible H-representation. Let (U, ρ) be an H-representation, the
G
induction to G denoted (IndG
H U, IndH ρ) is the tensor product
IndG
H U = CG ⊗ U
CH
where the G action is given by
g · (s ⊗ u) = (gs) ⊗ u
where s ∈ CG and u ∈ U . However, we will usually use the following more
concrete description. Choose S = {s1 , . . . , sl } to be representatives of the left
cosets of H in G (where l = [G : H]). Note that any element g ∈ G can be
written in a unique way as g = si h where si ∈ S and h ∈ H. Every element of
IndG
H U is a formal sum of the form
α1 (s1 , u1 ) + . . . + αl (sl , ul )
where ui ∈ U and αi ∈ C. In other words, as a vector space IndG
H U is
l
M
U,
i=1
that is, l copies of U . The action is defined on elements of the form (si , u) by
g · (si , u) = (sj , h · u)
where sj and h are unique such that gsi = sj h. The required action is given by
extending linearly. Note that dim IndG
H U = [G : H] dim U . It is important to
G
mention that the representations IndH U and ResG
H V depend not only on the
groups G and H but also on the specific embedding of H into G. Hence we
will have to give the specific embeddings when discussing these representations.
Both induction and restriction are transitive and additive, that is, if K ≤ H ≤ G
6
then
G
H
∼
IndG
H IndK U = IndK U,
G
G
∼
IndG
H (U ⊕ V ) = IndH U ⊕ IndH V
and
G
G
∼
ResH
K ResH U = ResK U,
G
G
∼
ResG
H (U ⊕ V ) = ResH U ⊕ ResH V.
For restriction this is a trivial statement and for induction the proof is [3, Propositions 1.1.10 and 1.1.11]. An important fact that relates induction to restriction
is the following one (for proof see [3, Corollary 1.1.20 ]).
Theorem 2.4 (Frobenius reciprocity). Let H ≤ G and let U and V be G and
H-representations respectively. Then the multiplicity of V in ResG
H U equals the
multiplicity of U in IndG
HV.
Using characters, Frobenius reciprocity can be written as the following equality
G
hIndG
H V, U i = hV, ResH U i.
2.4
Representations of the symmetric group
We will recall some elementary facts regarding the representation theory of the
symmetric group. More details can be found in [6, 12]. Recall that an integer
composition of n is a tuple λ = [λ1 , . . . , λk ] of non-negative integers such that
λ1 + · · · + λk = n while an integer partition of n (denoted λ ⊢ n) is an integer
composition such that λ1 ≥ λ2 ≥ · · · ≥ λk > 0. From now on, when dealing
with a partition λ we will write its elements in superscript λ = [λ1 , . . . , λk ]
because we want to reserve the subscript for multipartitions. Note that 0 has
one partition, namely the empty partition, denoted by ∅. We can associate to
any partition λ a graphical description called a Young diagram, which is a table
with λi boxes in its i-th row. For instance, the Young diagram associated to
the partition [3, 3, 2, 1] of 9 is:
We will identify the two notions and regard integer partition and Young diagram
as synonyms. It is well known that irreducible representations of Sn are indexed
by integer partitions of n. We denote the irreducible representation associated
to the partition λ (also called its Specht module) by S λ . Explicit description of
S λ can be found in [12, Section 2.3]. It will be often convenient to draw the
diagram λ instead of writing S λ . For instance we may write
7
⊕
instead of: S λ ⊕ S δ for partitions λ = [3] and δ = [2, 1].
We will now describe several branching rules for Sn . Here the advantage of
using Young diagrams becomes clear. Recall that we can think of Sn as the
group of all permutations of {1, . . . , n + 1} that leave n + 1 fixed. Hence, we can
view Sn as a subgroup of Sn+1 . We call this the standard embedding of Sn into
Sn+1 . In this case the branching rules are well known and very natural (proof
can be found in [12, Section 2.8]).
Theorem 2.5 (Classical branching rules). Let λ ⊢ n be a Young diagram.
1. Denote by Y + (λ) the set of Young diagrams obtained from λ by adding
one box. Then
M
S
IndSnn+1 S λ =
Sγ .
γ∈Y + (λ)
2. Similarly, denote by Y − (λ) the set of Young diagrams obtained from λ by
removing one box. Then
ResSSnn−1 S λ =
M
Sγ .
γ∈Y − (λ)
Example 2.6. Let λ =
then
IndSS43 S λ =
⊕
⊕
and
ResSS32 S λ =
⊕
.
We now turn our attention to the Littlewood-Richardson branching rule. If
we identify Sk (Sr ) with the group of all permutations of {1, . . . , k + r} that
leave {k + 1, . . . , k + r} (respectively, {1, . . . , k}) fixed we can view Sk × Sr as a
subgroup of Sk+r . Given λ ⊢ k and δ ⊢ r the Littlewood–Richardson rule gives
S
(S λ ⊠ S δ ) into irreducible Sk+r -representations.
the decomposition of IndSk+r
k ×Sr
8
In other words, if we write this decomposition as
M
S
(S λ ⊠ S δ ) =
IndSk+r
k ×Sr
cγλ,δ S γ
γ⊢(k+r)
it gives a combinatorial interpretation for the coefficients cγλ,δ (called the Littlewood–Richardson coefficients). The aim of this paper is to generalize the
classical branching rules and the Littlewood-Richardson rule to the group F ≀ Sn
for any finite group F . Although the details of the Littlewood-Richardson rule
for Sn will not be essential in the sequel, we give them here for the sake of
completeness. For this we have to introduce some more notions. First we generalize the notion of a Young diagram. For k ≤ n, let λ = (λ1 , · · · , λr ) ⊢ k
and γ = (γ 1 , · · · , γ s ) ⊢ n be partitions such that λi ≤ γ i for every 1 ≤ i ≤ r.
The skew diagram γ/λ is the diagram obtained by erasing the diagram λ from
the diagram γ. For instance if λ = [2, 1] and γ = [4, 3, 1] then γ/λ is the skew
diagram
.
A skew tableau is a skew diagram whose boxes are filled with numbers. We call
the original diagram the shape of the tableau. Let t be a skew tableau with n
boxes such that the number of boxes with entry i is δ i . The content of t is the
composition δ = [δ 1 , . . . , δ l ]. We say that a skew tableau is semi-standard if its
columns are increasing and its rows are non-decreasing. For instance
1 1
2 3
(2.1)
2
is a semi-standard skew tableau of shape [4, 3, 1]/[2, 1] with content [2, 2, 1]. The
row word of a skew tableau t is the string of numbers obtained by reading the
entries of t from right to left and top to bottom. For instance, the row word of
tableau 2.1 is 11322. A string of numbers is called a lattice permutation if for
every prefix of this string and for every number i, there are no less occurrences
of i than occurrences of i + 1. For instance, the string 11322 is not a lattice
permutation since the prefix 113 contains one 3 and no 2’s. Now we can state
the Littlewood-Richardson rule (for proof see [6, Theorem 2.8.13]).
Theorem 2.7. The Littlewood-Richardson coefficient cγλ,δ is the number of
9
semi-standard skew tableaux of shape γ/λ with content δ whose row word is
a lattice permutation.
Example 2.8. If λ = [2, 1], δ = [3, 2] and γ = [4, 3, 1] then cγλ,δ = 2 since there
are two skew tableaux with the required properties. These are:
1 1
2 2
1 1
1 2
1
2
Remark 2.9. Note that for many values of γ we have cγλ,δ = 0, that is, S γ is not
S
an irreducible constituent of IndSk+r
(S λ ⊠ S δ ). For instance, this happens for
k ×Sr
every γ such that γ i < λi for some i.
Remark 2.10. Note that the classical branching rule for induction can be deduced from the Littlewood-Richardson rule if we set r = 1.
3
Representation theory of F ≀ Sn
The goal of this section is to describe the irreducible representations of F ≀ Sn .
We follow the approach of [3], but we introduce different notation that will be
more convenient to our purpose. We also prove some technical lemmas that will
be of later use.
3.1
Inflation
Definition 3.1. Let G be a finite group and let N E G be a normal subgroup.
Let (U, ρ) be a representation of G/N . We denote by (U , ρ) the G-representation
defined as follows. As a vector space, U = U , and the G-action is
ρ(g)(u) = ρ(gN )(u) ∀u ∈ U.
Following [3] we call U the inflation of U .
Note that dim U = dim U and if U is an irreducible G/N -representation then
U is an irreducible G-representation as well.
The specific case we will be interested in is the following. If F and G are finite
groups, then F X E F ≀X G and (F ≀X G)/F X ∼
= G so any G-representation U
can be inflated into an (F ≀X G)-representation with action
(f, g) · u = gu.
10
Lemma 3.2. Let G1 and G2 be groups acting on the disjoint sets X1 and X2
respectively and let U1 and U2 be G1 and G2 -representations. Note that Ui is
an (F ≀Xi Gi )-representation and U1 ⊠ U2 is an (F ≀X (G1 × G2 ))-representation
˙ 2 ). Then
(where X = X1 ∪X
U1 ⊠ U2 ∼
= U1 ⊠ U2
as (F ≀X1 G1 ) × (F ≀X2 G2 ) ∼
= F ≀X (G1 × G2 )-representations.
Proof. As vector spaces both representations are spanned by elements of the
form u1 ⊠ u2 for ui ∈ Ui . An element ((f1 , g1 ), (f2 , g2 )) ∈ F ≀X1 G1 × F ≀X2 G2
acts in both representations by
((f1 , g1 ), (f2 , g2 )) · (u1 ⊠ u2 ) = g1 u1 ⊠ g2 u2 .
3.2
Conjugacy and extensions
Let N E G be a normal subgroup. Then we can define an action of G on the
set Irr N . For every (U, ρ) ∈ Irr N we define g · (U, ρ) = (g U,g ρ) where
g
U = U,
g
ρ(n) = ρ(g −1 ng)
for all g ∈ G and n ∈ N .
In the case of F X E F ≀ G this action has a simple form. Recall that the set of
irreducible representations of F X is
Irr F X = {( ⊠ Ux , ⊠ ρx ) | (Ux , ρx ) ∈ Irr F }.
x∈X
x∈X
Lemma 3.3 ([3, Lemma 2.4.1]). Denote (U, ρ) = ( ⊠ Ux , ⊠ ρx ) ∈ Irr F X then
x∈X
x∈X
(f, g) · U = ⊠ Ug−1 x
x∈X
for all (f, g) ∈ F ≀ G.
Definition 3.4. Assume that H ≤ G satisfies that (Uhx , ρhx ) ∼
= (Ux , ρx ), for
every x ∈ X and h ∈ H. In other words, F ≀ H is a subgroup of the stabilizer of
11
U . Define an (F ≀ H)-representation (ExH U, ExH ρ) in the following way. As a
vector space, ExH U = U and the group action is:
ExH ρ(f, h)( ⊠ ux ) = ⊠ ρh−1 x (f (x))(uh−1 x ) = ⊠ ρx (f (x))(uh−1 x ).
x∈X
x∈X
x∈X
(ExH U, ExH ρ) is called the extension of U with respect to H. When no ambie , ρe) instead of (ExH U, ExH ρ).
guity arises we write (U
e , ρe) is indeed an (F ≀ H)-representation.
Lemma 3.5. (U
Proof. We remark that this proof is essentially [3, Lemma 2.4.3]. Given (f1 , h1 ),(f2 , h2 ) ∈
F ≀ H what we need to prove is:
(3.1)
ρe((f1 , h1 ) · (f2 , h2 ))( ⊠ ux ) = ρe((f1 , h1 ))(e
ρ((f2 , h2 ))( ⊠ ux )).
x∈X
x∈X
The left hand side of Equation (3.1) is:
ρe((f1 , h1 ) · (f2 , h2 ))( ⊠ ux ) = ρe((f1 (h1 ∗ f2 ), h1 h2 ))( ⊠ ux ) =
x∈X
x∈X
= ⊠ ρh−1 h−1 x ((f1 (h1 ∗ f2 ))(x))uh−1 h−1 x .
x∈X
2
1
2
1
Since ρhx = ρx for every h ∈ H this equals
⊠ ρx ((f1 (h1 ∗ f2 ))(x))uh−1 h−1 x = ⊠ ρx (f1 (x)f2 (h1−1 x))uh−1 h−1 x .
x∈X
2
x∈X
1
2
1
The right hand side of Equation (3.1) is
ρe((f1 , h1 ))(ρ̃((f2 , h2 ))( ⊠ ux )) = ρe((f1 , h1 ))( ⊠ ρh−1 x (f2 (x))(uh−1 x )).
x∈X
x∈X
2
2
Again, since ρh−1 x = ρx the last expression equals
2
ρe((f1 , h1 ))( ⊠ ρx (f2 (x))(uh−1 x )) = ⊠ ρh−1 x (f1 (x))(ρh−1 x (f2 (h−1
1 x))(uh−1 h−1 x ))
x∈X
2
x∈X
= ⊠
x∈X
1
1
2
ρx (f1 (x))(ρx (f2 (h1−1 x))(uh−1 h−1 x ))
2
1
= ⊠ ρx (f1 (x)f2 (h1−1 x))uh−1 h−1 x .
x∈X
So we get the desired equality.
12
2
1
1
Remark 3.6. Let H ≤ K ≤ G and let U = ⊠ Ux be an F X -representation
x∈X
such that F ≀ K is a subgroup of the stabilizer of U . Note that
≀K
ExH U = ResF
F ≀H ExK U.
Lemma 3.7. Let G1 and G2 be groups acting on disjoint sets X and Y respectively. Let U = ⊠ Ux and V = ⊠ Uy be F X and F Y -representations
x∈X
y∈Y
respectively. Assume that H1 ≤ G1 , H2 ≤ G2 are subgroups such that F ≀ H1
and F ≀ H2 are subgroups of the stabilizers of U and V respectively. Then
ExH1 U ⊠ ExH2 V ∼
= ExH1 ×H2 (U ⊠ V )
as F ≀X ∪Y
(H1 × H2 ) ∼
= (F ≀X H1 ) × (F ≀Y H2 )-representations.
˙
Proof. Both vector spaces are spanned by elements of the form ( ⊠ ux ) ⊠
x∈X
( ⊠ uy ) and in both cases the group action is
y∈Y
((f1 , h1 ), (f2 , h2 ))·(( ⊠ ux )⊠( ⊠ uy )) = ( ⊠ f1 (x)·uh−1 x )⊠( ⊠ f2 (y)·uh−1 y ).
x∈X
3.3
y∈Y
x∈X
1
y∈Y
2
Irreducible representations of F ≀ Sn
We now return to the specific case of F ≀ Sn . From now on, fix some indexing
for the set of irreducible representation of F , say Irr F = {U1 , . . . , Ul }. Without
loss of generality, we assume that U1 is the trivial representation of F . Let
1 ≤ ij ≤ l for j = 1, . . . , n and let
n
n
j=1
x=1
(U, ρ) = ( ⊠ Uij , ⊠ ρij )
be an irreducible F n -representation. We define its type to be the integer composition
type(U ) = (n1 , . . . , nl )
such that ni is the number of j ∈ {1, . . . , n} such that Uij ∼
= Ui . Clearly
l
X
ni = n and ni ≥ 0 for every 1 ≤ i ≤ l. It is clear that two irreducible
i=1
representations are in the same orbit of conjugation if and only if they have the
13
same type. Moreover, the stabilizer of U is isomorphic to F ≀ (Sn1 × · · · × Snl ).
Given a specific composition n = (n1 , . . . , nl ) define
Un = U1⊠n1 ⊠ · · · ⊠ Ul⊠nl
that is, the first n1 representations in the product are U1 , the next n2 representations are U2 etc. Clearly, the type of Un is n and the set {Un | n is
an integer composition of n} serve as a set of representatives for the orbits of
conjugation. A tuple Λ = (λ1 , . . . , λl ) such that λi ⊢ ni for every i is called a
multipartition of n with l components. We will also call it a multipartition of
the composition n and denote this by Λ n. We denote by S Λ the (irreducible)
Sn1 × · · · × Snl -representation S Λ = S λ1 ⊠ · · · ⊠ S λl . Finally we can define:
Definition 3.8. Let n = (n1 , . . . , nl ) be an integer composition of n and let
Λ = (λ1 , . . . , λl ) be a multipartition of n. Denote by ΦΛ = Φ(λ1 ,...,λl ) the F ≀ Sn
representation
fn ⊗ S Λ ).
(U
IndF ≀Sn
F ≀(Sn1 ×···×Snl )
Theorem 3.9. [3, Theorem 2.6.1] The set
{ΦΛ | n is some integer composition of n and Λ is a multipartition of n}
is a complete list of the irreducible representations of F ≀Sn . Moreover, ΦΛ ∼
= ΦΛ′
′
if and only if Λ = Λ .
We define a multi-Young diagram to be a tuple of Young diagrams. As we
identify partitions and Young diagrams, we also identify multipartitions and
multi-Young diagrams. Hence, multi-Young diagrams (with l components) index
the irreducible representations of F ≀ Sn . For instance, if Λ = ([2], [2, 1], [1, 1, 1])
then the irreducible representation ΦΛ of (say) S3 ≀ S8 corresponds to the multiYoung diagram
(
,
,
).
Remark 3.10. Note that trF , the trivial representation of F = F ≀S1 , corresponds
to the multi-Young diagram
( , ∅, . . . , ∅).
14
Let λ be some partition of n. We set Φiλ = Φ(∅,...,∅,λ,∅,...,∅) where the nonempty partition is in the i-th position. Note that
⊠n ⊗ S λ .
Φiλ = Ug
i
Remark 3.11. Since {U1 , . . . , Ul } are irreducible representations of F = F ≀ S1
then Ui can be also written as Φi([1]) .
A key observation is the following one.
Proposition 3.12. Let n = (n1 , . . . , nl ) be an integer composition of n and let
Λ = (λ1 , . . . , λl ) be a multipartition of n. The following isomorphism holds:
F ≀S
ΦΛ ∼
= IndF ≀Snn1 ×···×F ≀Sn (Φ1λ1 ⊠ · · · ⊠ Φlλl ).
l
Proof. By definition
≀Sn
ΦΛ = IndF
F ≀(Sn
1 ×···×Snl )
fn ⊗ S Λ )
(U
if we write this more explicitly we get
≀Sn
ΦΛ = IndF
F ≀Sn
1 ×···×F ≀Snl
^
(U1⊠n1 ⊠ · · · ⊠ Ul⊠nl ) ⊗ (S λ1 ⊠ · · · ⊠ S λl )
Using Lemma 3.2 and Lemma 3.7 this equals:
≀Sn
IndF
F ≀Sn
1 ×···×F ≀Snl
^
]
(U1⊠n1 ⊠ · · · ⊠ Ul⊠nl ) ⊗ (S λ1 ⊠ · · · ⊠ S λl ).
Now, using Lemma 2.3 this equals:
≀Sn
IndF
F ≀Sn
1 ×···×F ≀Snl
^
]
(U1⊠n1 ⊗ S λ1 ) ⊠ · · · ⊠ (Ul⊠nl ⊗ S λl )
≀Sn
IndF
F ≀Sn
(Φ1λ1 ⊠ · · · ⊠ Φlλl )
which is precisely
1 ×···×F ≀Snl
as required.
15
4
Littlewood-Richardson rule for F ≀ Sn
In this section we generalize the standard Littlewood-Richardson rule for Sn to
the case of F ≀ Sn . As mentioned above this is a new proof for [5, Theorem 4.7].
Given two integers k and r and two integer compositions
k = (k1 , . . . , kl ),
l
X
ki = k
l
X
ri = r
i=1
r = (r1 , . . . , rl ),
i=1
let Λ = (λ1 , . . . , λl ) and ∆ = (δ1 , . . . , δl ) be multipartitions of k and r respectively. We want to find the decomposition of
F ≀S
(ΦΛ ⊠ Φ∆ )
IndF ≀Sk+r
k ×F ≀Sr
into irreducible representations. In other words, if we write
F ≀S
(ΦΛ ⊠ Φ∆ ) =
IndF ≀Sk+r
k ×F ≀Sr
MM
Γ
CΛ,∆
ΦΓ
n Γ n
where the outer sum is over all integer compositions n of k + r, we want to find
Γ
the coefficients CΛ,∆
. We start with a specific case.
Proposition 4.1. Let λ ⊢ k and δ ⊢ r then
F ≀S
(Φiλ ⊠ Φiδ ) =
IndF ≀Sk+r
k ×F ≀Sr
M
cγλ,δ Φiγ
γ⊢(k+r)
where cγλ,δ is the Littlewood-Richardson coefficient.
Before proving this results we need some lemmas.
Lemma 4.2. Let H ≤ G be a subgroup of G. Let U be an H-representation
then
IndF ≀X G U ∼
= IndG U
H
F ≀X H
as (F ≀X G)-representations.
Proof. Let s1 , . . . , sl be representatives of the H cosets in G. Note that (1F , si )
for i = 1, . . . , l are representatives for the F ≀X H cosets in F ≀X G where 1F is
16
the constant function 1F (x) = 1F . Now, define
F ≀X G
T : IndG
H U → IndF ≀X H U
by
T ((si , u)) = ((1F , si ), u)
and extending linearly. Clearly, T is a vector space isomorphism. Now, given
(f, g) ∈ F ≀X G
T ((f, g) · (si , u)) = T (g(si , u))
= T ((sj , hu))
= ((1F , sj ), hu)
assuming that gsi = sj h for h ∈ H. Note that
(f, g)(1F , si ) = (f (g ∗ 1F ), gsi ) = (f 1F , gsi ) =
= (f, sj h) = (1F , sj )(s−1
j ∗ f, h)
hence
(f, g) · T ((si , u)) = (f, g) · ((1F , si ), u)
= ((1F , sj ), (s−1
j ∗ f, h) · u)
= ((1F , sj ), hu)
so
T ((f, g) · (si , u)) = (f, g) · T ((si , u))
as required.
Lemma 4.3 ([3, Proposition 1.1.15]). Assume H ≤ G and let U (V ) be a G
(respectively H)-representation. Then
G
G
∼
IndG
H (ResH (U ) ⊗ V ) = U ⊗ IndH V.
Lemma 4.4. Let V be an (F ≀X G)-representation. Let H ≤ G and let W be
some H-representation.
17
Then
F ≀X G
F ≀X G
G
∼
IndF
≀X H (ResF ≀X H (V ) ⊗ W ) = V ⊗ IndH W .
Proof. Apply Lemma 4.3 and Lemma 4.2.
Proof of Proposition 4.1. According to the definition
F ≀S
F ≀S
g
⊠k ⊗ S λ ) ⊠ (U
⊠r ⊗ S δ )).
((Ug
(Φiλ ⊠ Φiδ ) = IndF ≀Sk+r
IndF ≀Sk+r
i
i
k ×F ≀Sr
k ×F ≀Sr
Using Lemma 2.3 this equals
F ≀S
g
⊠r
⊠k ⊠ U
λ
δ
((Ug
IndF ≀Sk+r
i ) ⊗ (S ⊠ S ))
i
k ×F ≀Sr
and by Lemma 3.2 and Lemma 3.7 this equals
^
F ≀S
⊠(k+r)
) ⊗ (S λ ⊠ S δ )).
((Ui
IndF ≀Sk+r
k ×F ≀Sr
If we use more precise notation, this is actually
⊠(k+r)
F ≀S
((ExSk ×Sr Ui
IndF ≀Sk+r
k ×F ≀Sr
) ⊗ (S λ ⊠ S δ ))
but according to Remark 3.6 this equals
⊠(k+r)
F ≀S
F ≀S
((ResF ≀(Sk+r
ExSk+r Ui
IndF ≀Sk+r
k ×F ≀Sr
k ×Sr )
) ⊗ (S λ ⊠ S δ )).
Returning to imprecise notation, this is
^
F ≀S
F ≀S
⊠(k+r)
)) ⊗ (S λ ⊠ S δ ))
(Ui
((ResF ≀Sk+r
IndF ≀Sk+r
k ×F ≀Sr
k ×F ≀Sr
where now
^
⊠(k+r)
Ui
is an (F ≀ Sk+r )-representation.
By Lemma 4.4 this equals
^
S
⊠(k+r)
(S λ ⊠ S δ )
⊗ IndSk+r
Ui
k ×Sr
18
but according to the standard Littlewood-Richardson rule
S
(S λ ⊠ S δ ) =
IndSk+r
k ×Sr
X
cγλ,δ S γ .
γ⊢(k+r)
Hence our representation equals
^
⊠(k+r)
⊗
Ui
X
cγλ,δ S γ =
γ⊢(k+r)
X
^
⊠(k+r)
⊗ Sγ ) =
cγλ,δ (Ui
γ⊢(k+r)
X
cγλ,δ Φiγ
γ⊢(k+r)
as required.
Now we turn to the general case. Let k = (k1 , . . . , kl ) and r = (r1 , . . . , rl )
be integers compositions of k and r respectively. We denote k + r = (k1 +
r1 , . . . , kl + rl ), an integer composition of k + r.
Theorem 4.5. Let Λ = (λ1 , . . . , λl )
k and ∆ = (δ1 , . . . , δl )
F ≀S
(ΦΛ ⊠ Φ∆ ) =
IndF ≀Sk+r
k ×F ≀Sr
M
r then
Γ
CΛ,∆
ΦΓ
Γ (k+r)
where
Γ
CΛ,∆
=
l
Y
cγλii ,δi .
i=1
Remark 4.6. Note that k + r is the only composition of k + r occurring in this
summation.
Before proving this result we need another technical lemma about induction.
Lemma 4.7. Assume H1 ≤ G1 (H2 ≤ G2 ) and let U1 (respectively, U2 ) be a
representation of H1 (respectively, H2 ). Then
G
G
G1 ×G2
IndH
(U1 ⊠ U2 ) ∼
= IndH11 U1 ⊠ IndH22 U2 .
1 ×H2
Proof. It is more convenient here to use the tensor product definition of induced
G1 ×G2
representation. Define a vector space isomorphism T : IndH
(U1 ⊠ U2 ) →
1 ×H2
G2
1
IndG
H1 U1 ⊠ IndH2 U2 by
T ((s1 , s2 ) ⊗ (u1 ⊠ u2 )) = (s1 ⊗ u1 ) ⊠ (s2 ⊗ u2 )
19
and extending linearly. Now take some (g1 , g2 ) ∈ G1 × G2 and note that
T ((g1 , g2 ) · ((s1 , s2 ) ⊗ (u1 ⊠ u2 ))) = T ((g1 s1 , g2 s2 ) ⊗ (u1 ⊠ u2 ))
= (g1 s1 ⊗ u1 ) ⊠ (g2 s2 ⊗ u2 )
= (g1 · (s1 ⊗ u1 )) ⊠ (g2 · (s2 ⊗ u2 ))
= (g1 , g2 ) · T ((s1 , s2 ) ⊗ (u1 ⊠ u2 )).
as required.
Proof of Theorem 4.5. According to Proposition 3.12, the representation
F ≀S
(ΦΛ ⊠ Φ∆ )
IndF ≀Sk+r
k ×F ≀Sr
equals
F ≀S
≀Sk
(IndF
IndF ≀Sk+r
F ≀Sk
k ×F ≀Sr
1
F ≀Sr
l
1
l
1
×···×F ≀Skl (Φλ1 ⊠· · ·⊠Φλl )⊠IndF ≀Sr1 ×···×F ≀Srl (Φδ1 ⊠· · ·⊠Φδl )).
Using Lemma 4.7 this equals
F ≀S
≀Sk ×F ≀Sr
(IndF
IndF ≀Sk+r
F ≀Sk ×···×F ≀Sk
k ×F ≀Sr
1
l
l
l
1
1
×F ≀Sr1 ×···×F ≀Srl (Φλ1 ⊠· · ·⊠Φλl ⊠Φδ1 ⊠· · ·⊠Φδl ))
which, by transitivity of induction, equals
F ≀S
IndF ≀Sk+r
k ×···×F ≀Sk
1
l
1
×F ≀Sr1 ×···×F ≀Srl (Φλ1
⊠ · · · ⊠ Φlλl ⊠ Φ1δ1 ⊠ · · · ⊠ Φlδl ).
Rearranging we get
F ≀S
IndF ≀Sk+r
k ×F ≀Sr
1
1 ×···×F ≀Skl ×F ≀Srl
((Φ1λ1 ⊠ Φ1δ1 ) ⊠ · · · ⊠ (Φlλl ⊠ Φlδl )).
Again using transitivity we can write this as
F ≀Sk
F ≀S
IndF ≀Sk+r
k +r
1
1
×···×F ≀Skl +rl
+r
×···×F ≀Skl +rl
((Φ1λ1 ⊠Φ1δ1 )⊠· · ·⊠(Φlλl ⊠Φlδl ))
1 ×···×F ≀Skl ×F ≀Srl
IndF ≀Sk1 ×F1 ≀Sr
1
and using Lemma 4.7 this equals
F ≀S
IndF ≀Sk+r
k +r
1
1
F ≀Skl+rl
F ≀Sk1 +r1
1
l
l
1
×···×F ≀Skl +rl (IndF ≀Sk1 ×F ≀Sr1 (Φλ1 ⊠Φδ1 )⊠· · ·⊠IndF ≀Skl ×F ≀Srl (Φλl ⊠Φδl )).
20
According to Proposition 4.1 we get
F ≀S
IndF ≀Sk+r
k +r
1
1
M
cγλ11 ,δ1 Φ1γ1 )
×···×F ≀Skl +rl ((
γ1 ⊢(k1 +r1 )
⊠ ··· ⊠ (
M
cγλll ,δl Φlγl ))
γl ⊢(kl +rl )
which equals
M
γ1 ⊢(k1 +r1 )
···
M
(
l
Y
F ≀S
cγλii ,δi IndF ≀Sk+r
k +r
1
γl ⊢(kl +rl ) i=1
1
1
×···×F ≀Skl +rl (Φγ1
⊠ · · · ⊠ Φ1γl ))
which, according to Proposition 3.12, is precisely
l
M Y
( cγλii ,δi )ΦΓ
Γ k+r i=1
as required.
5
Classical branching rules for F ≀ Sn
In this section we retrieve Pushkarev’s result of “classical” branching rules for
F ≀ Sn [10, Theorem 10]. Let n = (n1 , . . . , nr ) be an integer composition of
n and let Λ = (λ1 , . . . , λl )
n be a multipartition. We want to find the deF ≀S
≀Sn
composition into irreducible representations of IndF ≀Snn+1 ΦΛ and ResF
F ≀Sn−1 ΦΛ .
This is relatively easy using the results of the previous section. We start with
induction.
Theorem 5.1. With notation as above
F ≀S
IndF ≀Snn+1 ΦΛ =
l
M
(dim Ui
i=1
M
Φ(λ1 ,...,γ,...,λl ) )
γ∈Y + (λi )
where γ is in the i-th position of (λ1 , . . . , γ, . . . , λl ).
For the proof of Theorem 5.1 we need the following lemma.
Lemma 5.2. Let U be an H-representation and let K be some group, then
IndK×H
U∼
= CK ⊠ U.
H
Proof. Clearly {(k, 1) | k ∈ K} are representatives of the H ∼
= {1K } × H cosets
21
in K × H. Define T : IndK×H
U → CK ⊠ U by
H
T (((k, 1), u)) = k ⊠ u
which is clearly a vector space isomorphism and note that
T ((k ′ , h′ ) · ((k, 1), u)) = T (((k ′ k, 1), h′ u))
= k ′ k ⊠ h′ u
= (k ′ , h′ ) · (k ⊠ u)
= (k ′ , h′ ) · T (((k, 1), u))
so T is an isomorphism of (K × H)-representations.
Proof of Theorem 5.1. Noting that F ≀n+1 Sn = F ≀ Sn × F = F ≀ Sn × F ≀ S1
and by transitivity of induction
F ≀S
F ≀S
F ≀Sn ×F
ΦΛ .
IndF ≀Snn+1 ΦΛ = IndF ≀Snn+1
×F ≀S1 IndF ≀Sn
According to Lemma 5.2 this equals
F ≀S
IndF ≀Snn+1
×F ≀S1 ΦΛ ⊠ CF.
It is well-known that the decomposition of CF is
CF =
l
M
(dim Ui · Ui )
i=1
so we obtain
l
l
M
M
F ≀S
F ≀S
(dim Ui IndF ≀Snn+1
dim
U
·
U
))
=
(Φ
⊠
(
IndF ≀Snn+1
i
i
Λ
×F ≀S1 (ΦΛ ⊠ Ui )).
×F ≀S1
i=1
i=1
But Ui = Φi([1]) (see Remark 3.11) so we can write this as
l
M
F ≀S
i
(dim Ui IndF ≀Snn+1
×F ≀S1 (ΦΛ ⊠ Φ([1]) )).
i=1
Using Theorem 4.5 and Remark 2.10 this is precisely the required result.
22
Using Frobenius reciprocity we have the following corollary for restriction.
Corollary 5.3. With notation as above
≀Sn
ResF
F ≀Sn−1 ΦΛ =
l
M
(dim Ui
i=1
M
Φ(λ1 ,...,γ,...,λl ) )
γ∈Y − (λi )
where γ is in the i-th position of (λ1 , . . . , γ, . . . , λl ).
Example 5.4. Let Λ be the multipartition associated to the multi-Young diagram
(
,
,
)
so ΦΛ is an irreducible representation of S3 ≀S8 . Assuming we have indexed Irr S3
such that U1 is the trivial representation, U2 is the standard representation and
U3 is the alternating representation then
≀S9
ΦΛ
IndSS33 ≀S
8
is associated to
(
,
2(
,
,
)⊕(
,
,
) ⊕ 2(
(
6
,
,
,
,
,
) ⊕ 2(
,
)⊕(
,
,
,
)⊕
)⊕
).
Application: The quiver of the category algebra C(F ≀ FIn )
Denote by FIn the category of all injective functions between subsets of {1, . . . , n}.
In this section we apply the Littlewood-Richardson rule for computing the ordinary quiver of the category algebra of F ≀ FIn , the wreath product of a finite
group F with FIn . In the next two sections we give some preliminary background on the wreath product of a group with a category and on quivers. In
section 6.3 we give the description of the quiver.
23
6.1
The wreath product of a group with a category
All categories in this paper are finite. Hence we can regard a category C as a
set of objects, denoted C 0 , and a set of morphisms, denoted C 1 . If a, b ∈ C 0 then
C(a, b) is the set of morphisms from a to b. Let g ∈ C(a, b) and g ′ ∈ C(c, d) be
two morphisms. Recall that the composition g ′ · g is defined if and only if b = c
and we denote this fact by by ∃g ′ · g. A category D is called a subcategory of C if
it obtained from C by removing objects and morphisms. D is a full subcategory
if D(a, b) = C(a, b) for every a, b ∈ D0 . Let F be a finite group, let C be a finite
category and let H : C → Set be a functor from C to the category of finite sets.
Define a new category D in the following way. The set of objects is the same
as the set of objects of C, that is, D0 = C 0 . Given two objects a, b ∈ D0 , the
hom-set D(a, b) is {(f, g) | f ∈ F H(a) , g ∈ C(a, b)} where F H(a) is the set of all
functions f : H(a) → F . So we can write a specific morphism as (f, g). Now,
given two morphisms (f, g) ∈ D(a, b) and (f ′ , g ′ ) ∈ D(b, c) the composition is
(f ′ , g ′ ) · (f, g) = ((f ′ (H(g))) · f, g ′ g)
where · is componentwise multiplication of functions in F H(a) .
Definition 6.1. The category D defined above is called the wreath product of
F and C with respect to H and it is denoted by F ≀H C.
Since monoids are categories with one object, Definition 6.1 is also a definition
for the wreath product of a group G with a monoid M . In this case the functor
F is just an action of M on the left of some set X. Hence M acts on the right
of F X in the following way. Given g ∈ M and f ∈ F X the function f ∗ g is
defined by
(f ∗ g)(x) = f (g · x).
The wreath product F ≀X M is then just the right semidirect product F X ⋊ M .
Remark 6.2. One may note that if M is a group then Definition 6.1 does not
coincide with Definition 2.1. However, we will immediately prove that the two
ways to define a wreath product of groups are isomorphic. We have to use
different definition for wreath product in this section because Definition 2.1
does not generalize well to monoids and categories.
In the next lemma we denote by F wrX G the wreath product of Definition 6.1
which is apriory different from F ≀X G of Definition 2.1.
24
Lemma 6.3. Let F and G be finite groups such that G acts on the left X. Then
F wrX G ∼
= F ≀X G.
Proof. In this proof we denote by ∗1 (∗2 ) the left (right) action of G on F X as
in Definition 2.1 (respectively, Definition 6.1). Define T : F ≀X G → F wrX G by
T (f, g) = (f ∗2 g, g).
Clearly, T has an inverse
T −1 (f, g) = (f ∗2 g −1 , g).
Moreover, note that
T ((f, g) · (f ′ , g ′ )) = T (f · (g ∗1 f ′ ), gg ′ ) = T (f · (f ′ ∗2 g −1 ), gg ′ )
= ((f ∗2 (gg ′ )) · (f ′ ∗2 g ′ ), gg ′ )
while
T ((f, g)) · T ((f ′ , g ′ )) = (f ∗2 g, g) · (f ′ ∗2 g ′ , g ′ ) = ((f ∗2 (gg ′ )) · (f ′ ∗2 g ′ ), gg ′ )
so T is also a group homomorphism as required.
6.2
The ordinary quiver of an EI-category algebra
Recall that a unital C-algebra is a unital ring A that is also a vector space over
C such that c(ab) = (ca)b = a(cb) for all c ∈ C and a, b ∈ A. The algebras
that are of interest for us in this section are category algebras. Let D be a
finite category. The category algebra CD is the C-vector space with basis the
morphisms of the category, that is, all formal linear combinations
{c1 g1 + . . . + ck gk | ci ∈ C, gi ∈ D1 }
with multiplication being linear extension of
g′ · g =
g ′ g
0
25
∃g ′ · g
otherwise
.
A quiver is a non-directed graph where multiple edges and loops are permitted.
The (ordinary) quiver Q of an algebra A is a quiver that contains information
about the algebra’s representations. The exact definition is as follows. The vertices of Q are in a one-to-one correspondence with the set Irr A of all irreducible
representations of A (up to isomorphism). Given two irreducible representations
U and V the number of edges (more often called arrows) from U to V is
dim Ext1 (U, V ).
For the sake of simplicity, if Q is the quiver of the algebra of D we will call it
simply the quiver of D.
When considering quivers of categories we can restrict our discussion to a special
kind of categories. Two categories C and D are called equivalent if there are
functors F : C → D and G : D → C such that F G ∼
= 1D and GF ∼
= 1C where ∼
= is
natural isomorphism of functors. It is well known that C and D are equivalent
if and only if there is a fully faithful and essentially surjective functor from C to
D. If C and D are equivalent categories then they have the same quiver (since
their algebras are Morita equivalent, see [14, Proposition 2.2]). A category C is
called skeletal if no two objects of C are isomorphic. Note that any category C
is equivalent to some (unique) skeletal category called its skeleton. The skeleton
of C is the full subcategory having one object from every isomorphism class of
C. So we can restrict ourselves to discussing skeletal categories.
There is a special kind of categories whose quiver has a more concrete description. This description was discovered independently by Li [7] and by Margolis
and Steinberg [9]. For explaining it we need more definitions from category
theory. A category D is called an EI-category if every endomorphism is an isomorphism. In other words, every endomorphism monoid D(a, a) of this category
is a group. A morphism g ∈ D1 of an EI-category is called irreducible if it is not
an isomorphism but whenever g = g ′ g ′′ , either g ′ or g ′′ is an isomorphism. The
set of irreducible morphisms from a to b is denoted IRR A(a, b). The quiver of
skeletal EI-categories is described in the following theorem, which is [7, Theorem 4.7] or [9, Theorem 6.13] for the case of the field of complex numbers.
Theorem 6.4. Let D be a finite skeletal EI-category and denote by Q the
quiver of D. Let C IRR D(a, b) denote the C-vector space spanned by the set
26
IRR D(a, b). It is also an D(b, b) × D(a, a)-representation according to
(h′ , h) · g = h′ gh−1
for (h′ , h) ∈ D(b, b) × D(a, a) and g ∈ IRR D(a, b). Then
1. The vertex set of Q is
G
Irr D(a, a).
a∈D 0
2. If V ∈ Irr(D(a, a)) and U ∈ Irr(D(b, b)), then the number of arrows from
V to U is the multiplicity of U ⊠ V ∗ as an irreducible constituent of the
D(b, b) × D(a, a)-representation C IRR D(a, b).
6.3
The quiver of the category F ≀ FIn
As mentioned above, we denote by FIn the category of all injective functions
between subsets of {1, . . . , n}. In other words, the objects of FIn are subsets
of {1, . . . , n} and given two objects A and B the hom-set FIn (A, B) contains
all the injective functions from A to B. Note that ∅ is an initial object of this
category, that is, for every A ⊆ {1, . . . , n} there is a unique empty function from
∅ to A. We are interested in the category F ≀H FIn where H : FIn → Set is the
inclusion functor. We will omit the H and denote this category by F ≀ FIn . This
category has a natural description using matrices similar to the description of
F ≀ Sn . We can identify the hom-set F ≀ FIn (A, B) with a set of matrices whose
rows are indexed by elements of B and columns are indexed by elements of A.
The matrix M (f,g) identified with (f, g) ∈ F ≀ FIn (A, B) is defined by
(f,g)
Mi,j
f (j) g(j) = i
=
0
otherwise
where i ∈ B and j ∈ A. g is a total function so M (f,g) has no zero columns.
Moreover, since g is an injective function M (f,g) is column and row monomial,
that is, every row and column contains at most one non-zero element. Hence
F ≀ FIn (A, B) can be identified with the set of all column and row monomial
matrices over F without zero columns where the columns are indexed by A
and the rows are indexed by B. Composition of morphisms then corresponds
′ ′
to matrix multiplication. Note that the multiplication M (f ,g ) · M (f,g) of two
matrices of this form where (f, g) ∈ F ≀ FIn (A, B) and (f ′ , g ′ ) ∈ F ≀ FIn (C, D)
27
is defined if and only if B = C. In other words, multiplication is defined if
′
′
and only if the columns of M (f ,g ) and the rows of M (f,g) are indexed by the
same set. It is easy to see that any endomorphism monoid F ≀ FIn (A, A) of
this category is isomorphic to the group F ≀ SA hence this category is actually
an EI-category. Our goal is to describe the quiver of F ≀ FIn . The case where
F is the trivial group was originaly done in [2, Theorem 8.1.2]. A different
computation of this case using Theorem 6.4 is done in [9, Example 6.15] and
here we merely imitate their method. As explained in the previous section
we can work with the skeleton of F ≀ FIn . For the sake of simplicity we will
denote this skeleton by SF n . It is clear that two objects A and B of F ≀ FIn are
isomorphic if and only if |A| = |B|. Hence we can identify the skeleton SF n with
the full subcategory of F ≀FIn whose objects are the empty set and {1, . . . , k} for
k = 1, . . . , n. So we can identify the objects of SF n with 0, . . . , n. Now the homset SF n (k, r) is identified with all the r × k matrices over F which are column
and row monomial and without zero columns. Composition of morphisms then
corresponds to matrix multiplication as explained above. By Theorem 6.4 we
know that the vertices of the quiver of SF n are in one-to-one correspondence
with irreducible representation of the endomorphism groups, which are F ≀ Sk
for 0 ≤ k ≤ n. In other words:
Corollary 6.5. Let F be a group with l distinct irreducible representations.
The vertices in the quiver of F ≀ FIn can be identified with all the multi-Young
diagrams with k boxes and l components where 0 ≤ k ≤ n.
The next step is to identify the irreducible morphisms of SF n .
Lemma 6.6. The irreducible morphisms of SF n are precisely the morphisms
from k to k + 1 for 0 ≤ k ≤ n − 1. In other words,
SF (k, r)
n
IRR SF n (k, r) =
0
r =k+1
.
otherwise
Proof. It is clear that any morphism in SF n (k, k+1) is irreducible. On the other
hand, take some morphism (f, g) ∈ SF n (k, r) and assume that k+1 < r. Choose
j ∈ {1, . . . , r} not in the image of g. Define inc : {1, . . . , k} → {1, . . . , k + 1}
to be the inclusion function and g ′ : {1, . . . , k + 1} → {1, . . . , r} is the function
28
defined by
g(i) i ≤ k
g ′ (i) =
.
j
i=k+1
It is clear that g ′ and inc are not bijections and that g = g ′ ◦ inc. Denote by
1F the constant function 1F : {1, . . . , k + 1} → F defined by 1F (i) = 1F for
i = 1, . . . , k + 1. Since g ′ and inc are not bijections it is clear that (1F , g ′ ) and
(f, inc) are not isomorphisms. Moreover,(1F , g ′ ) · (f, inc) = (f, g) so (f, g) is not
an irreducible morphism as required.
From Theorem 6.4 and Lemma 6.6 we can immediately deduce the following
corollary.
Corollary 6.7. Let V ∈ Irr F ≀ Sk and U ∈ Irr F ≀ Sr be two vertices in the quiver
of F ≀ FIn such that r 6= k + 1 then there are not arrows from V to U .
It is left to consider the situation where r = k + 1. We have to study the representation C IRR(SF n (k, k + 1)) = C SF n (k, k + 1) under the action described
in Theorem 6.4. This is a permutation representation, i.e., it is a linearization
of the action of F ≀ Sk × F ≀ Sk+1 on the set SF n (k, k + 1) given by
((h′ , π ′ ), (h, π)) · (f, g) = (h′ , π ′ ) · (f, g) · (h, π)−1 .
Lemma 6.8. The above action is transitive.
Proof. Chose some (f, g) ∈ SF n (k, k + 1) and let j ∈ {1, . . . , k + 1} be the only
element not in the image of g. Define π ′ ∈ Sk+1 by
g −1 (i) i 6= j
π ′ (i) =
.
k + 1 i = j
Recall that g is injective so g −1 (i) is well defined if i ∈ im g. It is clear that
π ′ g = inc : {1, . . . , k} → {1, . . . , k + 1}. Now define h′ : {1, . . . , k + 1} → F by
h′ (i) =
(f (g −1 (i)))−1
1
i 6= j
.
i=j
It is easy to see that (h′ , π ′ ) · (f, g) = (1F , inc) hence the action is transitive
(even if we multiply only on the left).
29
It is well-known that if the action of G on some set X is transitive then the
permutation representation CX is IndG
K (trK ) where K = stab(x) is the stabilizer
of some x ∈ X and trK is its trivial representation.
So our representation is also of this form. We want to understand better the
stabilizer of some x ∈ SF n (k, k + 1). It is convenient to choose x = (1F , inc)
and to use the matrix interpretation discussed above.
Lemma 6.9. Choose x = (1F , inc) ∈ SF n (k, k + 1). The stabilizer stab(x) is
isomorphic to (F ≀ Sk ) × F .
Proof. (1F , inc) is identified with the (n + 1) × n matrix with 1 along its main
diagonal and 0 elsewhere.
1
0
(1F , inc) =
0
0
..
.
0
0
0
1
0 ···
0
=
I
0
···
0
.
It is easy to see that given any matrix A ∈ F ≀ Sk if we want some B ∈ F ≀ Sk+1
such that
B
I
0
···
0
A =
I
0 ···
0
then B must be of the form
B=
0
..
.
0
A−1
0
···
0
a
= A−1 ⊕ (a)
where a can be any element of F . Hence
stab((1F , inc)) = {(A ⊕ (a), A) | A ∈ F ≀ Sk ,
a ∈ F} ∼
= (F ≀ Sk ) × F
as required.
Proposition 6.10. Let V ∈ Irr F ≀ Sk and U ∈ Irr F ≀ Sk+1 identified with two
vertices of the quiver. The number of arrows from V to U is the multiplicity of U
30
F ≀S
as an irreducible constituent in the F ≀Sk+1 -representation Ind(F ≀Sk+1
(V ⊠trF ).
k )×F
Proof. Denote K = stab((1F , inc)). According to Theorem 6.4 and the above
discussion, the required number is the multiplicity of U ⊠ V ∗ as an irreducible
F ≀S
×F ≀S
k
constituent in the F ≀ Sk+1 × F ≀ Sk -representation IndK k+1
trK where
trK is the trivial representation of K. Using inner product of characters, and
recalling that we use the same notation for the representation and its character,
this number is
F ≀Sk+1 ×F ≀Sk
hU ⊠ V ∗ , IndK
trK i.
By Frobenius reciprocity this equals
F ≀Sk+1 ×F ≀Sk
hU ⊠ V ∗ , IndK
F ≀Sk+1 ×F ≀Sk
trK i = hResK
Recall that K = {(A ⊕ (a), A) | A ∈ F ≀ Sk ,
F ≀Sk+1 ×F ≀Sk
hResK
1
|K|
(U ⊠ V ∗ ), trK i. =
1
=
|K|
1
=
|K|
(U ⊠ V ∗ ), trK i.
a ∈ F } so
X
U ⊠ V ∗ (A ⊕ (a), A)
X
U (A ⊕ (a))V ∗ (A)
X
U (A ⊕ (a))V (A)
(A⊕(a),A)∈K
(A⊕(a),A)∈K
(A⊕(a),A)∈K
We want to think of the last term as an inner product of (F ≀Sk )×F -representations,
but neither U nor V is an (F ≀ Sk ) × F -representations. However,
F ≀S
U (A ⊕ (a)) = ResF ≀Skk+1
×F U (A ⊕ (a))
and
V (A) = V (A) trF (a) = (V ⊠ trF )(A, a)
as a K ∼
= (F ≀ Sk ) × F representation. Hence
1
|K|
X
U (A ⊕ (a))V (A)
(A⊕(a),A)∈K
equals
31
1
|(F ≀ Sk ) × F |
X
F ≀S
(Res(F ≀Sk+1
U )(A ⊕ (a))(V ⊠ trF )(A, a)
k )×F
(A,a)∈F ≀Sk ×F
which is the inner product
F ≀S
hRes(F ≀Sk+1
U, V ⊠ trF i.
k )×F
Using again Frobenius reciprocity this equals
F ≀Sk+1
hU, IndK
(V ⊠ trF )i
which is precisely the required number.
Clearly K ∼
= (F ≀ Sk ) × F = (F ≀ Sk ) × (F ≀ S1 ). Note also that the embedding
K ֒→ F ≀Sk+1 is precisely the standard embedding of the Littlewood-Richardson
rule. By Theorem 4.5 we can conclude:
Theorem 6.11. The vertices of the quiver of F ≀ FIn are in one-to-one correspondence with multi-Young diagrams with with k boxes and l components where
l = | Irr F | and 0 ≤ k ≤ n. Let k = (k1 , . . . kl ) and r = (r1 , . . . , rl ) be two integer compositions of k and r respectively and let Λ = (λ1 , . . . , λl )
k and
∆ = (δ 1 , . . . , δl )
r be multipartitions of k and r respectively. There can be
no more than one arrow from ΦΛ to Φ∆ . There is an arrow if and only if the
following holds:
• r = k + 1.
• r1 = k1 + 1 and ri = ki for 2 ≤ i ≤ l.
• λ1 is obtained from δ1 by adding one box, and λi = δi for 2 ≤ i ≤ l.
Proof. We have already seen that there are no arrows from ΦΛ to Φ∆ unless
r = k + 1 (Corollary 6.7). If r = k + 1 then Proposition 6.10 implies that the
number of arrows from ΦΛ to Φ∆ is the multiplicity of Φ∆ as an irreducible
F ≀S
F ≀S
k+1
constituent in IndF ≀Skk+1
×F (ΦΛ ⊠ trF ) = IndF ≀Sk ×F ≀S1 (ΦΛ ⊠ trF ). Recall that by
Remark 3.10 the multi-Young diagram corresponds to trF is
( , ∅, . . . , ∅)
32
so the result follows immediately from the Littlewood-Richardson rule for F ≀Sn .
Example 6.12. The quiver of the category S3 ≀ FI2 is given in the following
figure:
(
(∅,
, ∅)
, ∅, ∅)
( , ∅, ∅)
( , ∅, ∅)
( ,
, ∅) ( , ∅,
)
(∅, ∅, ∅)
(∅,
, ∅)
)
(∅,
, ∅)
(∅,
,
(∅, ∅,
) (∅, ∅,
)
(∅, ∅,
)
Clearly, two multipartitions Λ = (λ1 , . . . , λl ) and ∆ = (δ 1 , . . . , δl ) are in the
same connected component if and only if λi = δi for i = 2, . . . , l. Hence connected components can be parametrized by multipartitions of k with l − 1 components where k = 0, . . . , n. Denote by Pl (n) the number of multipartitions of n
with l components. A generating function for this sequence and other formulas
can be found in [1]. The following result is immediate.
Corollary 6.13. Let F be a non-trivial finite group and denote l = | Irr F |.
Then the quiver of F ≀ FIn has
n
X
Pl−1 (k)
k=0
connected components.
Acknowledgements: The author is grateful to Tullio Ceccherini-Silberstein
for examining this work and for his very helpful remarks.
References
[1] George E. Andrews. A survey of multipartitions: congruences and identities. In Surveys in number theory, volume 17 of Dev. Math., pages 1–19.
Springer, New York, 2008.
33
[2] Bridget Helen Brimacombe. The Representation Theory of the Incidence
Algebra of an Inverse Semigroup. ProQuest LLC, Ann Arbor, MI, 2011.
Thesis (Ph.D.)–Carleton University (Canada).
[3] Tullio Ceccherini-Silberstein, Fabio Scarabotti, and Filippo Tolli. Representation theory and harmonic analysis of wreath products of finite groups,
volume 410 of London Mathematical Society Lecture Note Series. Cambridge University Press, Cambridge, 2014.
[4] Thomas Church, Jordan S. Ellenberg, and Benson Farb. FI-modules
and stability for representations of symmetric groups. Duke Math. J.,
164(9):1833–1910, 2015.
[5] Frank Ingram, Naihuan Jing, and Ernie Stitzinger. Wreath product symmetric functions. Int. J. Algebra, 3(1-4):1–19, 2009.
[6] Gordon James and Adalbert Kerber. The representation theory of the symmetric group, volume 16 of Encyclopedia of Mathematics and its Applications. Addison-Wesley Publishing Co., Reading, Mass., 1981. With a
foreword by P. M. Cohn, With an introduction by Gilbert de B. Robinson.
[7] Liping Li. A characterization of finite EI categories with hereditary category
algebras. J. Algebra, 345:213–241, 2011.
[8] Liping Li. Upper bounds of homological invariants of FI_G-modules. arXiv
preprint arXiv:1512.05879, 2015.
[9] Stuart Margolis and Benjamin Steinberg. Quivers of monoids with basic
algebras. Compos. Math., 148(5):1516–1560, 2012.
[10] I. A. Pushkarev. On the theory of representations of the wreath products
of finite groups and symmetric groups. Zap. Nauchn. Sem. S.-Peterburg.
Otdel. Mat. Inst. Steklov. (POMI), 240(Teor. Predst. Din. Sist. Komb. i
Algoritm. Metody. 2):229–244, 294–295, 1997.
[11] Eric Ramos. Homological invariants of FI-modules and FI_G-modules.
arXiv preprint arXiv:1511.03964, 2015.
[12] Bruce E. Sagan. The symmetric group, volume 203 of Graduate Texts in
Mathematics. Springer-Verlag, New York, second edition, 2001. Representations, combinatorial algorithms, and symmetric functions.
34
[13] Steven V Sam and Andrew Snowden. Representations of categories of Gmaps. arXiv preprint arXiv:1410.6054, 2014.
[14] Peter Webb. An introduction to the representations and cohomology of
categories. In Group representation theory, pages 149–173. EPFL Press,
Lausanne, 2007.
35
| 4math.GR |
HOMOGENEOUS LOCALLY NILPOTENT DERIVATIONS OF
NON-FACTORIAL TRINOMIAL ALGEBRAS
arXiv:1710.10610v1 [math.AG] 29 Oct 2017
YULIA ZAITSEVA
Abstract. We describe homogeneous locally nilpotent derivations of the algebra of regular functions for a class of affine trinomial hypersurfaces. This class comprises all nonfactorial trinomial hypersurfaces.
1. Introduction
Let K be an algebraically closed field of characteristic zero and R an algebra
over K. A derivation of R is a K-linear map δ : R → R satisfying the Leibniz rule:
δ(f g) = δ(f )g + f δ(g) for all f, g ∈ R. A derivation δ is called locally nilpotent if for
every f ∈ R there is an m ∈ Z>0 such that δ m (f ) = 0.
Let X be an irreducible affine algebraic variety over K and Ga = (K, +) the additive
group of the ground field. Locally nilpotent derivations of the algebra K[X] are known
to be in one-to-one correspondence with regular Ga -actions on the variety X, see e.g. [6,
Section 1.5].
Suppose that R is graded by a finitely generated abelian group K:
M
R=
Rw .
w∈K
A derivation δ : R → R is said to be homogeneous if it maps homogeneous elements to homogeneous ones. In such a case, there exists an element deg δ ∈ K satisfying δ(Rw ) ⊆ Rw+deg δ
for all w ∈ K. The element deg δ ∈ K is called the degree of δ.
Recall that an affine algebraic group is called a quasitorus if it is isomorphic to a direct
product of a torus and a finite abelian group. Suppose that a quasitorus H acts on the
variety X. Such an action corresponds to a grading on the algebra K[X] by the group of
characters K of the quasitorus H:
M
K[X] =
K[X]w , where K[X]w = {f | h ◦ f = w(h)f ∀h ∈ H}.
w∈K
It is easily shown that a locally nilpotent derivation of K[X] is homogeneous with respect
to this grading if and only if the quasitorus H normalizes the corresponding Ga -action
on X. A description of homogeneous locally nilpotent derivations enables us to describe
the automorphism group of an algebraic variety, see e.g. [3, Theorem 5.5], [2, Theorem 5.5].
Fix positive integers n0 , n1 , n2 and let n = n0 + n1 + n2 . For each i = 0, 1, 2, fix a tuple
lin
li = (lij | j = 1, . . . , ni ) of positive integers lij and define a monomial Tili = Ti1li1 . . . Tini i
in the polynomial algebra K[Tij , i = 0, 1, 2, j = 1, . . . , ni ]. By a trinomial we mean a
polynomial
g = T0l0 + T1l1 + T2l2 .
2010 Mathematics Subject Classification. Primary 13N15, 14R20; Secondary 13A50, 14J50.
Key words and phrases. Affine hypersurface, torus action, graded algebra, derivation.
1
2
YULIA ZAITSEVA
A trinomial hypersurface X(g) is the zero set {g = 0} in the affine space An . By R(g) =
K[X(g)] we denote the algebra of regular functions on X(g).
Our motivation to study trinomials comes from toric geometry. Consider an effective
action T × X → X of an algebraic torus T on an irreducible variety X. The complexity of
such an action is the codimension of a general T -orbit in X. It equals dim X − dim T .
Actions of complexity zero are actions with an open T -orbit. A normal variety admitting
a torus action with an open orbit is called a toric variety. If X is a toric (not necessary affine)
variety with the acting torus T , then Ga -actions on X normalized by T can be described
in terms of Demazure roots of the fan corresponding to X; see [5], [15, Section 3.4] for the
original approach and [13, 4, 3] for generalizations.
Let T × X → X be a torus action of complexity one. A description of Ga -actions on X
normalized by T in terms of proper polyhedral divisors may be found in [14] and [13]. It is
an interesting problem to find their explicit form in concrete cases.
The study of toric varieties is related to binomials, see e.g. [16, Chapter 4]. At the
same time, Cox rings establish a close relation between torus actions of complexity one and
trinomials, see [11, 10, 9, 3, 8]. In particular, any trinomial hypersurface admits a torus
action of complexity one.
In this paper we study locally nilpotent derivations of R(g) that are homogeneous with
respect to the ”finest” grading. This grading is defined in Construction 1. The ”finest” grading corresponds to an effective action of a quasitorus H on X(g). The neutral component
of the quasitorus H is a torus of dimension n − 2.LIt acts on X(g) with complexity one.
The weight monoid of a graded algebra R =
Rw is the set S = {w ∈ K | Rw 6= 0}.
w∈K
The weight cone of R is the convex cone ω in the rational vector space KQ = K ⊗Z Q
generated by the weight monoid S. A homogeneous derivation is said to be primitive if its
degree does not lie in the weight cone ω ⊆ KQ . It is clear that every primitive derivation
is locally nilpotent. The converse is false, see Example 7. An explicit description of all
primitive derivations of R(g) that are homogeneous with respect to the ”finest” grading
is given in [3, Theorem 4.3]1. Such derivations have the form hδC,β , where δC,β are some
special primitive derivations of R(g) (see [3, Construction 4.2] or Construction 2 below),
and h is a homogeneous element in the kernel of δC,β . We call derivations of the form hδC,β
elementary.
We expect that all homogeneous locally nilpotent derivations of trinomial algebras are
elementary (see Conjecture 1 in Section 5). We verify this conjecture in Theorem 1 for some
types of trinomials. In particular, the conjecture is proved for all non-factorial trinomial
hypersurfaces (Corollary 1). Corollary 2 gives a criterion for existence of homogeneous
locally nilpotent derivations of trinomial algebras. It enables us to give a new proof of
criteria for rigidity of factorial trinomial hypersurfaces, which was proved earlier in [1,
Theorem 1] (see Corollary 3).
The author is grateful to her supervisor Ivan Arzhantsev for posing the problem and
permanent support and to Sergey Gaifullin for useful discussions and comments.
2. Preliminaries
In this section the ”finest” deg-grading and elementary derivations of a trinomial algebra
are defined.
1In
[3], a more general class of affine varieties defined by a system of trinomials is studied, see [3,
Construction 3.1] for details.
HOMOGENEOUS DERIVATIONS OF TRINOMIAL ALGEBRAS
3
Construction 1. Fix positive integers n0 , n1 , n2 and let n = n0 + n1 + n2 . For each
i = 0, 1, 2, fix a tuple of positive integers li = (lij | j = 1, . . . , ni ) and define a monomial
lin
Tili = Ti1li1 . . . Tini i ∈ K[Tij , i = 0, 1, 2, j = 1, . . . , ni ]. By a trinomial we mean a polynomial
of the form
g = T0l0 + T1l1 + T2l2 .
A trinomial hypersurface X(g) is the zero set {g = 0} in the affine space An . It can
be checked that the polynomial g is irreducible, hence the algebra R(g) := K[Tij ] / (g)
of regular functions on X(g) has no zero divisors. We call such algebras R(g) trinomial.
We use the same notation for elements of K[Tij ] and their projections to R(g).
Following [9], we build a 2 × n matrix L from the trinomial g as follows:
−l0 l1 0
L=
.
−l0 0 l2
Let L∗ be the transpose of L. Denote by K the factor group K = Zn / Im L∗ and by
Q : Zn → K the projection. Let eij ∈ Zn , i = 0, 1, 2, j = 1, . . . , ni , be the canonical basis
vectors. The equalities
deg Tij = Q(eij )
(1)
define a K-grading on the algebra K[Tij ].
Since g is a homogeneous polynomial of degree
µ = li1 Q(ei1 ) + . . . + lini Q(eini )
for some µ ∈ K and any i = 0, 1, 2, it follows that equalities (1) define a K-grading
on R(g) = K[Tij ] / (g) as well. By ”deg” we denote the degree with respect to this grading.
−1 −3 3 0
3
3
2
Example 1. Let g = T01 T02 + T11 + T21 , i.e., L =
.
−1 −3 0 2
Since the
can be reduced by integer elementary row and column operations to
matrix L
1 0 0 0
the form
, it follows that the grading group K = Z4 / Im L∗ is isomorphic
0 1 0 0
to Z2 . The grading can be given explicitly via
−3
1
0
0
deg T01 =
, deg T02 =
, deg T11 =
, deg T21 =
.
3
1
2
3
−1
−1
1
1
0
2
Example 2. Let g = T01 T02 + T11 T12 + T21
and L =
.
−1 −1 0 0 2
The matrix L
can be reduced by elementary row and column operations to the form
1 0 0 0 0
, hence the group K = Z5 / Im L∗ is isomorphic to Z3 . We can define the
0 1 0 0 0
grading explicitly via
0
1
0
2
2
deg T01 = 0 , deg T02 = 2 , deg T11 = 2 , deg T12 = 0 , deg T21 = 1 .
0
1
−1
1
−1
−2 2 0
2
2
2
Example 3. Let g = T01
+ T11
+ T21
and L =
.
−2 0 2
2 0 0
Since L can be reduced to the form
, it follows that K = Z3 / Im L∗ is iso0 2 0
morphic to Z ⊕ Z2 ⊕ Z2 . Denote Z2 = {[0]2 , [1]2 }; the grading can be given explicitly
4
YULIA ZAITSEVA
via
deg T01
1
= [0]2 ,
[0]2
deg T11
1
= [1]2 ,
[0]2
deg T21
1
= [0]2 .
[1]2
The following lemma shows that the constructed deg-grading is the ”finest” grading of
the algebra R(g) such that all generators Tij are homogeneous.
Lemma 1. Let deg be the K-grading on the algebra R(g) defined in Construction 1 and
d be any K-grading
b
b such that all generators Tij are
deg
on R(g) by some abelian group K
d = ψ ◦ deg for some homomorphism ψ : K → K.
b In particular, any
homogeneous. Then deg
derivation that is homogeneous with respect to deg-grading is homogeneous with respect to
d
deg-grading
too.
d Tij determine a well-defined grading on the algebra R(g) if
Proof. The weights w
bij = deg
d
and only if the polynomial g is homogeneous with respect to the deg-grading,
i.e., the sums
b
li1 w
bi1 + . . . + lini w
bini ∈ K are the same for all i = 0, 1, 2. Factorization of Zn = heij i
∗
by Im L provides these equalities for the images of eij in K. Any other grading on R(g)
such that generators Tij are homogeneous is obtained by further factorization of the group
K = Zn / Im L∗ , and this factorization is the required map ψ.
The following construction is given in [3] and is described below
in the case of trinomial
−1 1
hypersurfaces (in notation of [3] that is r = 2, A = 10 −1
,
g
= gI for I = {0, 1, 2},
0
R(g) = R(A, P0 )).
Construction 2. Let us define a derivation δC,β of R(g). The input data are
• a sequence C = (c0 , c1 , c2 ), where ci ∈ Z, 1 6 ci 6 ni ;
• a vector β = (β0 , β1 , β2 ) such that βi ∈ K, β0 + β1 + β2 = 0.
It is clear that for β 6= 0 as above either all entries βi differ from zero or there is a unique i0
with βi0 = 0. According to these cases, we put further conditions and define:
(i) if βi 6= 0 for all i = 0, 1, 2 and there is at most one i1 with li1 ci1 > 1, then we set
Q lk
∂Tk
βi
, j = ci ,
∂Tkck
δC,β (Tij ) =
k6=i
0,
j 6= ci ;
(ii) if βi0 = 0 for a unique i0 and there is at most one i1 with i1 6= i0 and li1 ci1 > 1, then
l
βi Q ∂Tkk , j = ci ,
∂Tkck
δC,β (Tij ) =
k6=i,i0
0,
j 6= ci .
These assignments define a map δC,β on generators Tij . It can be extended uniquely to
a derivation on K[Tij ] by Leibniz rule. Clearly we have δC,β (g) = 0, whence the constructed
map induces a well-defined derivation of the factor algebra R(g).
Lemma 2. Every derivation δC,β of the algebra R(g) is primitive and locally nilpotent.
The proof is given in [3, Construction 4.2].
Let h ∈ R(g) be a homogeneous element in the kernel of a derivation δC,β . The derivation hδC,β is also locally nilpotent.
HOMOGENEOUS DERIVATIONS OF TRINOMIAL ALGEBRAS
5
Definition. We say that a derivation of a trinomial algebra R(g) is elementary if it has the
form hδC,β , where h is a homogeneous element in the kernel of δC,β . In addition, elementary
derivations of Type I are elementary derivations with δC,β corresponding to case (i) in
Construction 2; derivations of Type II are elementary derivations with δC,β corresponding
to case (ii).
3
3
2
Example 4. Suppose g = T01 T02
+ T11
+ T21
(see Example 1). Let us find all elementary
derivations of the algebra R(g). Note that there is a unique variable T01 in g with exponent 1. Hence only case (ii) with i0 6= 0, C = (1, 1, 1), is possible. Thus we have two cases
for elementary derivations of Type II:
(a) i0 = 1, that is β = (β0 , 0, −β0 ) for some β0 ∈ K. Then
3
δC,β (T01 ) = 2β0 T21 , δC,β (T02 ) = 0, δC,β (T11 ) = 0, δC,β (T21 ) = −β0 T02
,
∂
∂
3
i.e., δC,β = 2β0 T21
− β0 T02
, h ∈ Ker δC,β ;
∂T01
∂T21
(b) i0 = 2, that is β = (β0 , −β0 , 0) for some β0 ∈ K. Then
3
2
δC,β (T01 ) = 3β0 T11
, δC,β (T02 ) = 0, δC,β (T11 ) = −β0 T02
,
∂
∂
3
2
− β0 T02
, h ∈ Ker δC,β .
i.e., δC,β = 3β0 T11
∂T01
∂T11
δC,β (T21 ) = 0,
2
Example 5. Let g = T01 T02 +T11 T12 +T21
(see Example 2). There is a unique variable Tij in
the trinomial with exponent lij such that lij > 1. Hence the algebra R(g) admits elementary
derivations of both types: we can take any sequence C = (c0 , c1 , 1), c0 , c1 ∈ {1, 2}, and any
vector β = (β0 , β1 , β2 ) under the condition β0 + β1 + β2 = 0. Let us give an example of
elementary derivation of Type I:
∂
∂
∂
+ T01 T21
− T01 T11
.
δC,β = T11 T21
∂T02
∂T12
∂T21
2
2
2
Example 6. Let g = T01
+ T11
+ T21
(see Example 3); all exponents in this trinomial are
greater than 1. Hence there exists no elementary derivation.
3. Auxiliary lemmas
We use notation introduced in the previous section. Let us recall some definitions. Let K
be an abelian group and R a K-graded algebra. By a K-prime element of R we mean a
homogeneous nonzero nonunit f ∈ R such that f | gh with homogeneous g, h ∈ R implies
f | g or f | h. We say that R is factorially K-graded if every nonzero homogeneous
nonunit of R is a product of K-primes. It is clear that this decomposition is unique up to
permutation of factors and multiplication by units.
Lemma 3. With the notation of Construction 1, the following statements hold.
(a) The generators Tij are pairwise nonassociated K-prime elements of R(g).
(b) The algebra R(g) is factorially K-graded.
This was proved in [9, Proposition 2.2(i) and Theorem 1.1(i)].
The following two lemmas include basic properties of locally nilpotent derivations.
Lemma 4. Suppose δ : R → R is a locally nilpotent derivation of a domain R and f, g ∈ R.
(a) If f | δ(f ), then δ(f ) = 0.
(b) The derivation f δ is locally nilpotent if and only if f ∈ Ker δ holds.
(c) If f | δ(g) and g | δ(f ), then δ(f ) = 0 or δ(g) = 0.
6
YULIA ZAITSEVA
(d) If δ =
P
δi , where all derivations δi are homogeneous with respect to some
m6i6n
Z-grading on R, δm , δn 6= 0, then δm and δn are locally nilpotent.
The proof can be found, for example, in [6, Principles 5, 7, 14 and Corollary 1.20].
Lemma 5. Let R be a finitely generated Zk -graded domain. If there exists a nonzero locally
nilpotent derivation of R, then R admits a nonzero homogeneous locally nilpotent derivation.
Proof. Denote given locally nilpotent derivation on R by δ (0) . Consider a Z-grading on the
algebra R induced by the first component in Zk . Since R is finitely generated, we have
P (0)
(0)
δ (0) =
δi for some homogeneous derivations δi . It follows from Lemma 4(d) that
(1)
m6i6n
(0)
δn is a
δ =
nonzero locally nilpotent derivation that is homogeneous with respect to the
Z-grading by the first component of Zk . Applying Lemma 4(d) k − 1 times to Z-gradings
by other components of Zk gives a derivation δ (k) that is homogeneous with respect to the
grading by all components of Zk , i.e., with respect to the Zk -grading.
The following lemma is proved in [7, Lemma 3.4]. For convenience of the reader we give
a short proof below.
Lemma 6. Let δ be a deg-homogeneous locally nilpotent derivation of an algebra R(g).
Then there exists at most one variable Tij in every monomial Tili such that δ(Tij ) 6= 0.
Proof. Assume the converse. Then there is a monomial with at least two variables not
belonging to the kernel of the derivation. We can assume that these variables are T01 and
T02 in T0l0 .
Consider the following grading on the algebra K[Tij ]:
Tij
d Tij
deg
T01
T02
l02
−l01
T03 . . . T0n0 T11 . . . T1n1 T21 . . . T2n2
0
...
0
0
...
0
0
...
0
The trinomial g is homogeneous (of degree 0) with respect to this grading, therefore the
d
deg-grading
is a well-defined grading on the factor algebra R(g).
d
By Lemma 1, it follows that the derivation δ is deg-homogeneous.
Then we have the
following two cases.
d δ > 0. Then deg
d δ(T01 ) = deg
d δ + deg
d T01 > 0. Note that T01 is a unique variable
1) deg
with a positive degree. Hence every monomial in δ(T01 ) includes T01 and therefore T01
divides δ(T01 ). But δ(T01 ) 6= 0 by assumption. This contradicts Lemma 4(a).
d δ 6 0. Then deg
d δ(T02 ) = deg
d δ + deg
d T02 < 0. In the same way, T02 divides δ(T02 )
2) deg
and this contradicts Lemma 4(a).
4. Main results
The following proposition and its proof are variations on [3, Theorem 4.3]. The distinctions are the following. We consider only the case of trinomial hypersurfaces, and in [3,
Theorem 4.3] varieties defined by a system of trinomials are studied. At the same time, we
need only the assumption that images of monomials are proportional, and do not use the
primitivity of derivations. The latter condition is stronger, since according to [3, Proposition 3.5], the dimension of the homogeneous component of degree w is equal to 1 whenever
w − deg g does not lie in the weight cone.
HOMOGENEOUS DERIVATIONS OF TRINOMIAL ALGEBRAS
7
Proposition 1. Let δ : R(g) → R(g) be a deg-homogeneous locally nilpotent derivation.
Suppose that δ(T0l0 ), δ(T1l1 ), and δ(T2l2 ) lie in a subspace of R(g) of dimension 1. Then the
derivation δ is elementary.
Proof. According to Lemma 6, there is at most one variable Tij in each monomial Tili such
that δ(Tij ) 6= 0. Let K = {i | ∃ci : δ(Tici ) 6= 0}.
Let us prove that if δ(Tici ) 6= 0 and δ(Tkck ) 6= 0, then lici = 1 or lkck = 1. Assume the
converse, so lici 6= 1 and lkck 6= 1. By the above, we obtain
δ(Tili )
∂Tili
= δ(Tici )
.
∂Tici
(2)
Since lici 6= 1, we have Tici | ∂Tili /∂Tici , therefore Tici | δ(Tili ). Similarly, Tkck | δ(Tklk ). By
assumption, δ(Tili ) and δ(Tklk ) are proportional, hence Tkck | δ(Tili ) and Tici | δ(Tklk ). At the
same time, according to (2), δ(Tili ) is a product of δ(Tici ) and variables that are not equal
to Tkck , whence by Lemma 3(b) we have Tkck | δ(Tici ). For the same reason, Tici | δ(Tkck ),
which contradicts Lemma 4(c).
Thus lici > 1 holds for at most one i ∈ K.
The elements δ(Tili ) are proportional, that is there exists an f ∈ R(g) such that
δ(Tili ) ∈ Kf for any i. It follows from equality (2) that ∂Tili /∂Tici divides f for all i ∈ K.
Q
∂Tili /∂Tici divides f . Denote
The application of Lemma 3(b) yields that the product
i∈K
this ratio by h. From (2) we have
Y ∂T li
i
,
f =h
∂T
ici
i∈K
δ(Tici ) =
βi h
0,
Q
k∈K\{i}
l
∂Tkk
,
∂Tkck
i ∈ K,
i∈
/ K.
Let βi = 0 for all i ∈
/ K. Let us show that the sum of all βi is equal to 0. Note that g
should divide
!
l
li
l
k
i
Y ∂T
X
X
Y ∂T lk
X
∂T
∂Ti
i
k
k
βi h
=
=h
βi
δ(g) =
δ(Tici )
∂T
∂T
∂T
∂T
ici
kck
ici
kck
i∈K
i∈K
k∈K
i∈K
k∈K\{i}
in the algebra K[Tij ], since δ is a well-defined
derivation of the algebra R(g). But the
Q
lk
trinomial g and the monomial
∂Tk /∂Tkck are coprime in the factorial algebra K[Tij ].
P k∈K
P
P
Hence g divides the sum
βi , whence
βi =
βi = 0.
i∈K
i∈K
i
Now let C = (c0 , c1 , c2 ) be any sequence completing the ci , where i ∈ K. Then we have
δ = hδC,β . The fact that h belongs to the kernel of δC,β follows from Lemma 4(b).
We come to the main result of the paper.
Theorem 1. Let R(g) be a trinomial algebra. Suppose that there is at most one monomial
in g including a variable with exponent 1, that is (li1 j1 = li2 j2 = 1) ⇒ (i1 = i2 ). Then every
deg-homogeneous locally nilpotent derivation of the algebra R(g) is elementary of Type II.
Proof. Let us denote a deg-homogeneous locally nilpotent derivation of the algebra R(g)
by δ. By Lemma 6, there is at most one variable Tij in every monomial Tili such that
δ(Tij ) 6= 0. Let K = {i | ∃ci : δ(Tici ) 6= 0}.
1) Fix any i ∈ K such that lici > 1. We claim that Tici divides δ(Tkck ) for any k ∈ K \ {i}.
8
YULIA ZAITSEVA
Denote by Zm = {[0]m , [1]m , . . . , [m − 1]m } the cyclic group of order m. Consider the
following Zlici-grading on R(g):
Tks
d Tks
deg
T01
T02
...
Tici
...
T2n2
[0]lici [0]lici . . . [1]lici . . . [0]lici
It is well defined since lici [1]lici = [0]lici , i.e., all monomials in g have the same degree [0]lici .
d
By Lemma 1, the derivation δ is deg-homogeneous.
d δ(Tic ) 6= [0]l then Tic divides δ(Tic ). Indeed, this inequality implies that the
If deg
i
ici
i
i
degrees of all monomials in δ(Tici ) do not equal [0]lici , and it is possible only if every
monomial includes Tici (since Tici is a unique variable with a nonzero degree).
On the other hand, the fact that Tici divides δ(Tici ) together with δ(Tici ) 6= 0 contradicts
d δ(Tic ) = [0]l . Whence deg
d δ = [lic − 1]l 6= [0]l . In particular,
Lemma 4(a). Thus deg
i
ici
i
ici
ici
d
deg δ(Tkc ) 6= [0]l for k ∈ K \ {i}, and therefore Tic divides δ(Tkc ).
k
ici
i
k
2) Let us show that |K| 6 2. Assume the converse. By assumptions of the theorem,
|{i | lici = 1}| 6 1, whence |{i ∈ K | lici > 1}| > 2. Take any distinct i, k from the
latter set. According to 1), Tici divides δ(Tkck ) and Tkck divides δ(Tici ), which contradicts
Lemma 4(c).
3) Thus |K| 6 2. Then there are at most two nonzero δ(Tili ). On the other hand,
δ(T0l0 ) + δ(T1l1 ) + δ(T2l2 ) = δ(g) = 0 in the algebra R(g). Consequently, δ(T0l0 ), δ(T1l1 ),
and δ(T2l2 ) lie in a subspace of dimension 1. According to Proposition 1, the derivation δ
is elementary. In addition, there are at most two nonzero δ(Tili ). This implies that δ is
elementary of Type II.
If a trinomial contains a linear term (i.e., ni li1 = 1 for some i = 0, 1, 2), then the
corresponding trinomial hypersurface is isomorphic to the affine space Kn−1 . Further we
assume that ni li1 > 1 for all i = 0, 1, 2. According to [9, Theorem 1.1(ii)], in this case the
following statement holds.
Proposition 2. The following conditions are equivalent:
(a) the algebra R(g) is factorial;
(b) the group K is torsion free;
(c) the numbers di := gcd(li1 , . . . , lini ) are pairwise coprime.
This proposition enables us to prove some consequences of Theorem 1.
Corollary 1. If an algebra R(g) is non-factorial, then any deg-homogeneous locally nilpotent derivation of R(g) is elementary of Type II.
Proof. Let R(g) be a trinomial algebra that does not satisfy the conditions of Theorem 1.
It suffices to show that R(g) is factorial.
By assumption, g has at least two monomials including variables with exponent 1. Therefore at least two of d0 , d1 , d2 are equal to 1, whence the numbers d0 , d1 , d2 are pairwise
coprime. By Proposition 2, R(g) is factorial.
Corollary 2. The algebra R(g) admits a nonzero deg-homogeneous locally nilpotent derivation if and only if lij = 1 for some i = 0, 1, 2, j = 1, . . . , ni .
Proof. If the condition lij = 1 holds for some pair (i, j), then there exists a derivation of
the form δC,β (see Construction 2, case (ii)). It is deg-homogeneous and locally nilpotent.
HOMOGENEOUS DERIVATIONS OF TRINOMIAL ALGEBRAS
9
Let us prove the inverse implication by contradiction. Let lij > 2 for all i = 0, 1, 2,
j = 1, . . . , ni . Together with Theorem 1 this implies that all homogeneous locally nilpotent
derivations of R(g) are of the form hδC,β (where h ∈ Ker δC,β ). However it follows from
Construction 2 that there is no derivation of the form δC,β under the condition lij > 2 for
all i = 0, 1, 2, j = 1, . . . , ni . This contradiction proves the corollary.
An affine variety is called rigid if its algebra of regular functions admits no nonzero
locally nilpotent derivation. Geometrically, this means that the variety admits no nontrivial Ga -action. The automorphism group of a rigid trinomial variety is described in [2,
Theorem 5.5].
Corollary 2 enables us to obtain a new proof of the result proved earlier in [1, Theorem 1].
Corollary 3. A factorial trinomial hypersurface X(g) is rigid if and only if lij > 2 for any
i = 0, 1, 2, j = 1, . . . , ni .
Proof. If the condition lij > 2 does not hold for some (i, j), then there exists a derivation
of the form δC,β (see Construction 2, case (ii)). This implies the ”only if” part.
Let us prove the ”if” part. According to Corollary 2, there exists no nonzero deghomogeneous locally nilpotent derivation of the algebra R(g). By Proposition 2 together
with factoriality we have that the group K is torsion free. Then by Lemma 5 the nonexistence of nonzero deg-homogeneous locally nilpotent derivations implies the nonexistence
of any nonzero locally nilpotent derivations of R(g). This means rigidity.
3
3
2
Example 7. Consider g = T01 T02
+T11
+T21
(see Examples 1 and 4). This trinomial satisfies
conditions of Theorem 1, whence all deg-homogeneous locally nilpotent derivations δ of R(g)
are divided into two classes:
∂
∂
3
(a) δ = 2hβ0 T21
− hβ0 T02
, where h ∈ Ker δC,β , C = (1, 1, 1), β = (β0 , 0, −β0 );
∂T01
∂T21
∂
∂
3
2
− hβ0 T02
, where h ∈ Ker δC,β , C = (1, 1, 1), β = (β0 , −β0 , 0).
(b) δ = 3hβ0 T11
∂T01
∂T11
Some of them are not primitive, i.e., deg δ = deg hδC,β does not lie in the weight cone ω
for some h ∈ Ker δC,β . Let us consider an example. The weight monoid is generated by
vectors
−3
1
0
0
deg T01 =
, deg T02 =
, deg T11 =
, deg T21 =
,
3
1
2
3
u
whence the weight cone ω is the angle {−u 6 v 6 u}, where deg =
(see Figure 1).
v
3
k
In case (a), we have deg δC,β =
. Since for any k ∈ Z>0 the polynomial h = T11
0
k
belongs to the kernel
of δC,β
the derivation
T11
δC,β is homogeneous and locally nilpotent.
,
3
3
0
=
and lies in the weight cone for k > 2.
+
Its degree equals k
0
2k
2
2
2
2
Example 8. Let g = T01
+ T11
+ T21
(see Examples 3 and 6). According to Corollary 2,
there exists no deg-homogeneous locally nilpotent derivation of R(g).
Nevertheless the algebra R(g) admits a nonzero locally nilpotent derivation. For example,
consider the following derivation δ:
δ(T01 ) = iT21 ,
δ(T11 ) = −T21 ,
δ(T21 ) = −iT01 + T11 ,
where i ∈ K, i2 = −1. The derivation δ is locally nilpotent, since δ(−iT01 + T11 ) = 0,
δ 2 (T21 ) = 0, δ 3 (T01 ) = 0.
10
YULIA ZAITSEVA
v
deg T01
2
deg T11
δC,β
ω
deg T21
deg T11
deg T11 δC,β
deg T02
deg δC,β
u
Figure 1. The weight cone in Example 7
5. Open questions
In this section we discuss several open questions related to locally nilpotent derivations
of trinomial algebras.
Conjecture 1. All deg-homogeneous locally nilpotent derivations of a trinomial algebra
are elementary.
In the proof of Theorem 1 we apply Proposition 1 to confirm Conjecture 1 for trinomial
algebras R(g) corresponding to trinomials g under the following condition: there exists
at most one monomial in g including variables with exponent 1. We would like to apply
Proposition 1 in the remaining case too, that is, when there exist at least two monomials
in g including variables with exponent 1. In this case, by Proposition 2, the algebra R(g)
is factorial, the group K is torsion free, and a quasitorus, whose action corresponds to the
deg-grading, is a torus of dimension n − 2. In [12], all deg-homogeneous locally nilpotent
derivations for several trinomial hypersurfaces are described. These hypersurfaces do not
satisfy the conditions of Theorem 1. In particular, the following statements are proved in
[12, Theorem 3.22 and Theorem 3.24].
Example 9 (compare with Example 5). Any deg-homogeneous locally nilpotent derivation
2
of the algebra R(g), where g = T01 T02 + T11 T12 + T21
, has the form
∂
∂
p
k l
λT0i T1j T21 T1j
− T0i
or
∂T1ī
∂T1j̄
∂
α0 + α1
∂
∂
k l
p
+ α1 T0i T21
−
,
T0i T1j
T0i T1j (α1 T01 T02 − α0 T11 T12 ) α0 T1j T21
∂T0ī
∂T1j̄
2
∂T21
where k, l, p ∈ Z>0 , {i, ī} = {j, j̄} = {1, 2}, α0 , α1 , λ ∈ K, α0 + α1 6= 0.
Example 10. Any deg-homogeneous locally nilpotent derivation of the algebra R(g), where
g = T01 T02 + T11 T12 + T21 T22 , has the form
∂
∂
∂
k0 k1 k2
p
T0i0 T1i1 T2i2 (α2 T11 T12 − α1 T21 T22 ) α0 T1i1 T2i2
+ α1 T0i0 T2i2
+ α2 T0i0 T1i1
,
∂T0ī0
∂T1ī1
∂T2ī2
where k0 , k1 , k2 , p ∈ Z>0 , {i0 , ī0 } = {i1 , ī1 } = {i2 , ī2 } = {1, 2}, α0 , α1 , α2 ∈ K, (α0 , α1 , α2 ) 6=
(0, 0, 0), and α0 + α1 + α2 = 0.
HOMOGENEOUS DERIVATIONS OF TRINOMIAL ALGEBRAS
11
One can check that any derivation δ from these examples is elementary, so Conjecture 1
holds in these cases. Moreover, δ(T0l0 ), δ(T1l1 ), and δ(T2l2 ) lie in a subspace of dimension 1,
hence Conjecture 1 can be confirmed in these cases by Proposition 1 as well.
Besides the ”finest” deg-grading on the algebra R(g) by the group K one can consider
another grading on R(g) by the torsion free component K0 of K. Notice that the condition
to be homogeneous with respect to the deg-grading by the group K is more restrictive
than the condition to be homogeneous with respect to the grading by K0 . According to
Proposition 2, the deg-grading by K coincides with the grading by K0 if and only if the
algebra R(g) is factorial.
2
2
2
Example 11. Let g = T01
+ T11
+ T21
(see Examples 3, 6, and 8). According to Example 3,
we have K = Z + Z2 + Z2 . Therefore K0 = Z and the degree with respect to the grading
by K0 is equal to the standard degree of a polynomial (the sum of exponents). Hence the
derivation δ from Example 8 is homogeneous with respect to the grading by K0 . At the
same time, according to Example 8, there is no nonzero locally nilpotent derivation that is
homogeneous with respect to the K-grading.
If the algebra R(g) admits no nonzero K0 -homogeneous locally nilpotent derivation, then
by Lemma 5 R(g) admits no nonzero locally nilpotent derivation, i.e., the variety X(g) is
rigid.
Problem 1. Describe all locally nilpotent derivations of a trinomial algebra R(g) that are
homogeneous with respect to the K0 -grading.
Whereas the K-grading corresponds geometrically to the action of a quasitorus on the
trinomial hypersurface, the K0 -grading corresponds to the action of its neutral component,
i.e., of the maximal torus. This means that Problem 1 asks for a description of Ga -actions
normalized by the maximal torus, that is, of root subgroups in the automorphism group of
a trinomial hypersurface.
Problem 2. Describe all (not necessary homogeneous) locally nilpotent derivations of a
trinomial algebra R(g).
References
[1] Ivan Arzhantsev. On rigidity of factorial trinomial hypersurfaces. Int. J. Algebra Comput. 26 (2016),
no. 5, 1061–1070
[2] Ivan Arzhantsev and Sergey Gaifullin. The automorphism group of a rigid affine variety. Math.
Nachrichten 290 (2017), no. 5-6, 662–671
[3] Ivan Arzhantsev, Jürgen Hausen, Elaine Herppich, and Alvaro Liendo. The automorphism group of a
variety with torus action of complexity one. Moscow Math. J. 14 (2014), no. 3, 429–471
[4] Ivan Arzhantsev and Alvaro Liendo. Polyhedral divisors and SL2 -actions on affine T-varieties. Michigan
Math. J. 61 (2012), no. 4, 731–762
[5] Michel Demazure. Sous-groupes algébriques de rang maximum du groupe de Cremona. Ann. Sci. École
Norm. Sup. 3 (1970), 507–588
[6] Gene Freudenburg. Algebraic theory of locally nilpotent derivations. Encyclopaedia Math. Sci., vol. 136,
Springer, Berlin, 2006
[7] Sergey Gaifullin. Automorphisms of Danielewski varieties. arXiv:1709.09237, 22 pages
[8] Jürgen Hausen and Milena Wrobel. Non-complete rational T -varieties of complexity one. Math.
Nachrichten 290 (2017), no. 5-6, 815–826
[9] Jürgen Hausen and Elaine Herppich. Factorially graded rings of complexity one. In: Torsors, étale
homotopy and applications to rational points. London Math. Soc. Lecture Note Series 405, 414–428,
2013
12
YULIA ZAITSEVA
[10] Jürgen Hausen, Elaine Herppich, and Hendrik Süß. Multigraded factorial rings and Fano varieties with
torus action. Documenta Math. 16 (2011), 71–109
[11] Jürgen Hausen and Hendrik Süß. The Cox ring of an algebraic variety with torus action. Advances
Math. 225 (2010), no. 2, 977–1012
[12] Polina Kotenkova. Torus actions and locally nilpotent derivations. PhD Thesis, Moscow State University, Moscow, 2014
[13] Alvaro Liendo. Affine T-varieties of complexity one and locally nilpotent derivations. Transform.
Groups 15 (2010), no. 2, 398–425
[14] Alvaro Liendo. Ga -actions of fiber type on affine T-varieties. J. Algebra 324 (2010), no. 12, 3653–3665
[15] Tadao Oda. Convex bodies and algebraic geometry: an introduction to the theory of toric varieties. A
Series of Modern Surveys in Math. 15, Springer, Berlin, 1988
[16] Bernd Sturmfels. Gröbner bases and convex polytopes. University Lecture Series, vol. 8, AMS, Providence, Rhode Island, 1996
Lomonosov Moscow State University, Faculty of Mechanics and Mathematics, Department of Higher Algebra, Leninskie Gory 1, Moscow, 119991 Russia
E-mail address: yuliazaitseva@gmail.com
| 0math.AC |
IEEE TRANSACTIONS ON COMMUNICATIONS 2018
1
arXiv:1802.08676v1 [quant-ph] 23 Feb 2018
A Quantum-Search-Aided Dynamic Programming
Framework for Pareto Optimal Routing in Wireless
Multihop Networks
Dimitrios Alanis, Student Member, IEEE, Panagiotis Botsinis, Member, IEEE, Zunaira Babar, Hung Viet
Nguyen, Member, IEEE, Daryus Chandra, Student Member, IEEE, Soon Xin Ng, Senior Member, IEEE,
and Lajos Hanzo, Fellow, IEEE
Abstract—Wireless Multihop Networks (WMHNs) have to
strike a trade-off among diverse and often conflicting Qualityof-Service (QoS) requirements. The resultant solutions may be
included by the Pareto Front under the concept of Pareto Optimality. However, the problem of finding all the Pareto-optimal
routes in WMHNs is classified as NP-hard, since the number of
legitimate routes increases exponentially, as the nodes proliferate.
Quantum Computing offers an attractive framework of rendering
the Pareto-optimal routing problem tractable. In this context,
a pair of quantum-assisted algorithms have been proposed,
namely the Non-Dominated Quantum Optimization (NDQO) and
the Non-Dominated Quantum Iterative Optimization
(NDQIO).
√
However, their complexity is proportional to N , where N
corresponds to the total number of legitimate routes, thus still
failing to find the solutions in “polynomial time”. As a remedy,
we devise a dynamic programming framework and propose the
so-called Evolutionary Quantum Pareto Optimization (EQPO)
algorithm. We analytically characterize the complexity imposed
by the EQPO algorithm and demonstrate that it succeeds in
solving the Pareto-optimal routing problem in polynomial time.
Finally, we demonstrate by simulations that the EQPO algorithm
achieves a complexity reduction, which is at least an order of
magnitude, when compared to its predecessors, albeit at the cost
of a modest heuristic accuracy reduction.
Index Terms—Quantum Computing, NDQIO, NDQO, Dynamic Programming, Pareto Optimality, Routing.
L IST OF ACRONYMS
BBHT-QSA
BF
BER
CDP
CF(E)
DHA
DN
EQPO
HP
MODQO
Boyer-Brassard-Høyer-Tapp
Quantum
Search Algorithm
Brute Force
Bit Error Ratio
Classical Dynamic Programming
Cost Function (Evaluation)
Durr-Høyer Algorithm
Destination Node
Evolutionary Quantum Pareto Optimization
Hardware Parallelism
Multi-Objective Decomposition Quantum
Optimization
The authors are with the School of Electronics and Computer Science, University of Southampton, Southampton, SO17 1BJ, UK (email:
{da1d16, pb1y14, zb2g10, hvn08r, dc2n14, sxn, lh}@ecs.soton.ac.uk).
The financial support of the EPSRC under the grant EP/L018659/1, that
of the European Research Council, Advanced Fellow Grant and that of the
Royal Society’s Wolfson Research Merit Award is gratefully acknowledged.
Additionally, the authors acknowledge the use of the IRIDIS High Performance Computing Facility, and associated support services at the University
of Southampton, in the completion of this work. The research data of this
paper can be found at http://doi.org/10.5258/SOTON/D0402.
MO-ACO
NDQO
(P-)NDQIO
NSGA-II
OF
OPF-SR
(O)PF
QoS
QP
RN
SN
UV
UF
WMHN
Multi-Objective Ant Colony Optimization
Non-Dominated Quantum Optimization
(Preinitialized) Non-Dominated Quantum Iterative Optimization
Non-dominated Sorting Genetic Algorithm II
Objective Function
Optimal Pareto Front Self-Repair
(Optimal) Pareto Front
Quality-of-Service
Quantum Parallelism
Relay Node
Source Node
Utility Vector
Utility Function
Wireless MultiHop Network
I. I NTRODUCTION
T
HE concept of Wireless Multihop Networks (WMHN)
[1] enables the communication of remote nodes by forwarding the transmitted packets through a cloud of mobile
relays. Naturally, the specific choice of the relays plays a
significant role in the performance of WMHNs [2], thus
bringing their routing optimization in the limelight. Explicitly,
optimal routing relies on a fragile balance of diverse and often
conflicting Quality-of-Service (QoS) requirements [3], such
as the route’s overall Bit-Error-Ratio (BER) or Packet Loss
Ratio (PLR), its total power consumption, its end-to-end delay,
the route’s achievable rate, the entire system’s sum-rate and
its “lifetime” [4].
For the sake of taking into account multiple QoS requirements, several studies consider single-component Objective Functions (OF) as their optimization objectives. In this
context, the metric of Network Lifetime (NL) [4], [5] has
been utilized, which involves the routes’ power consumption
in conjunction with the nodes’ battery levels. Additionally,
the so-called Network Utility (NU) [6] also constitutes a
meritorious single-component optimization OF. Apart from
the aforementioned QoS requirements, NU also takes into
account the routes’ achievable rate [7]. In conjunction with the
construction of aggregate functions, the authors of [8], [9] also
incorporate QoS as constraints, thus providing a more holistic
view of the routing problem. In this context, Banirazi et al. [9]
optimized an aggregate function of the Dirichlet routing cost as
2
well as the average network delay at specific operating points
that maximize the network throughput.
The beneficial properties of dynamic programming [10]
have been exploited for the sake of identifying the optimal
routes, while relying on single-component aggregate functions.
In this context, Dijkstra’s algorithm [11]–[13] has been employed, since it is capable of approaching the optimal routes
at the cost of imposing a complexity on the order of O(E 3 ),
where E corresponds to the number of edges in the network’s
graph. Additionally, the appropriately modified Viterbi decoding algorithm [14], [15] has also been utilized for solving
single-component routing optimization problems, where the
route exploration process can be viewed as a trellis graph
and thus the routing problem is transformed into a decoding
problem. Explicitly, this transformation is reminiscent of the
famous Bellman-Ford algorithm [16].
The aforementioned approaches fail to identify the potential discrepancies among the QoS requirements, but they
can be unified by the concept of Pareto Optimality [17].
However, the search-space of multi-component optimization is
inevitably expanded due to combining the single-component
OFs. Furthermore, the complexity is proportional to O(N 2 ),
where N corresponds to the total number of eligible routes.
Additionally, since N increases exponentially as the relay
nodes proliferate [18], the Pareto-optimal routing problem
is classified as Non-deterministic Polynomial hard (NP-hard)
[19]. This escalating complexity can be partially mitigated
by identifying a single Pareto-optimal solution. For instance,
Gurakan et al. [20] conceived an optimal iterative routing
scheme for identifying a single Pareto-optimal solution in
terms of the sum rate and the energy consumption of wireless
energy-transfer-enabled networks. However, in our application
we are primarily interested in identifying the entire set of
Pareto-optimal solution, since it provides fruitful insights into
the underlying trade-offs [17]. In this context, multi-objective
evolutionary algorithms [18], [21], [22] have been employed
for addressing the escalating complexity. In particular, Yetgin
et al. [18] used both the Non-dominated Sorting Genetic
Algorithm II (NSGA-II) and the Multi-Objective Differential
Evolution Algorithm (MODE) for optimizing the transmission
routes in terms of their end-to-end delay and power dissipation.
While considering a similar context, Camelo et al. [21] invoked the NSGA-II for optimizing the same QoS requirements
for both the ubiquitous Voice over Internet Protocol (VoIP)
and for file transfer. Additionally, the so-called Multi-Objective
Ant Colony Optimization (MO-ACO) algorithm [23] has been
employed in [19] for the sake of addressing the multi-objective
routing problem in WMHNs.
Quantum computing provides a powerful framework [24]–
[26] for the sake of rendering Pareto-optimal routing problems
tractable by exploiting the so-called Quantum Parallelism (QP)
[27]. Explicitly, in [28] Quantum Annealing [29], has been
invoked for the sake of optimizing the activation of the
wireless links in wireless networks, while maintaining the
maximum throughput and minimum interference as well as
providing a substantial complexity reduction w.r.t. its classical
counterpart, namely simulated annealing. In terms of Pareto
optimal routing using universal quantum computing [27],
IEEE TRANSACTIONS ON COMMUNICATIONS 2018
the so-called Non-Dominated Quantum Optimization (NDQO)
algorithm proposed in [19] succeeded in identifying the entire
set of Pareto-optimal routes at the
√ expense of a complexity,
which is on the order of O(N N ), relying on QP. As an
improvement, the so-called Non-Dominated Quantum Iterative Optimization (NDQIO) algorithm was proposed in [30].
Explicitly, the NDQIO algorithm is also capable of identifying
the entire set of Pareto-optimal routes, while imposing a
1
parallel complexity and a sequential √
complexity defined
√ in
2
[30], which is on the order of O(NOPF N ) and O(NOPF N ),
respectively, by relying on the beneficial synergy between QP
and Hardware Parallelism (HP). Note that NOPF corresponds
to the number of Pareto-optimal routes.
Despite the substantial complexity reduction offered both
by the NDQO and the NDQIO algorithms, the multi-objective
problem still remains intractable, when the network comprises
an excessively high number of nodes due to the escalating
complexity. Explicitly, √
Zalka [31] has demonstrated that the
complexity order of O( N ) is the minimum possible, as long
as the database values are uncorrelated. By contrast, when the
formation of the Pareto-optimal route-combinations becomes
correlated owing to socially-aware networking [32], a further
complexity reduction can be achieved. Based on this specific
observation, we will design a novel algorithm, namely the
Evolutionary Quantum Pareto Optimization (EQPO), in order
to exploit the correlations exhibited by the individual Paretooptimal routes by appropriately constructing trellis graphs that
guide the search process in the same fashion as in Viterbi
decoding. Furthermore, we will also exploit the synergies
between QP and HP for the sake of achieving an additional
complexity reduction by considering as low a fraction of the
database entries as possible, while still guaranteeing a nearfull-search-based performance.
Our contributions are summarized as follows:
1) In Section III, we develop a novel multi-objective dynamic programming framework for generating potentially Pareto-optimal routes relying on the correlations
of the specific links constituting the Pareto-optimal
routes, hence substantially reducing the total number
of routes considered. Explicitly, this framework is a
multi-objective extension of the popular single-objective
Bellman-Ford algorithm.
2) In Section IV, we propose a novel quantum-assisted
algorithm, namely the Evolutionary Quantum Pareto
Optimization algorithm, which jointly exploits our novel
dynamic programming framework as well as the synergies between the QP and the HP for the sake of solving
the multi-objective routing problem of WMHNs.
3) In Section V, we also characterize the performance
versus complexity of the EQPO algorithm and demonstrate that it achieves both a parallel and a sequential
complexity reduction of at least an order of magnitude
for a 9-node WMHN, when compared to that of the
NDQIO algorithm.
1 We define the parallel complexity as the complexity imposed while taking
into account the degree of parallelism. By contrast, the sequential complexity
does not consider any kind of parallelism. In [30], they are referred to as
normalized execution time and normalized power consumption, respectively.
D. ALANIS ET AL.: A QUANTUM-AIDED DYNAMIC PROGRAMMING FRAMEWORK FOR PARETO OPTIMAL ROUTING IN WMHNS
The rest of this paper is organized as follows. In Section II,
we will briefly discuss the specifics of the network model
considered in our case study. In Section III, we will present a
dynamic programming framework, which is optimal in terms
of its heuristic accuracy. In Section IV, we will relax the
optimal framework of Section II for the sake of striking a
better accuracy versus complexity trade-off with the aid of
our EQPO algorithm. Subsequently, in Section V-A we will
analytically characterize the EQPO algorithm’s complexity and
in Section V-B we will evaluate its performance.
II. N ETWORK S PECIFICATIONS
In the context of this treatise, the model of the networks
considered both in [19] and in [30] has been adopted. To
elaborate further, the WMHN considered is a fully connected
network and it consists of a single Source Node (SN), a single
Destination Node (DN) and a cloud of Relay Nodes (RN).
The SN and the DN are located in the opposite corners of
a (100×100) m2 square-block area, which is the WMHN
coverage area considered. By contrast, the RNs are considered
to be roaming within the coverage area having random locations, which obey the uniform distribution within the WMHN
coverage area. A WMHN topology is exemplified in Fig. 1
for a WMNH consisting of 5 nodes in total. Additionally,
a cluster-head equipped with a quantum computer, which is
responsible for collecting all the required WMHN information,
such as the nodes’ geolocations and their interference levels, is
considered to be present at the DN side. Therefore, we should
point out that this treatise is focused on a centralized protocol.
3
overall Bit Error Ratio (BER). Before delving into the UFs,
let us define a legitimate route of our WMHN consisting of
Nnodes nodes, as xr = [SN, . . . , DN], which contains each
RN only once for the sake of limiting the total number N
of legitimate routes, while at the same time avoiding routes
associated with excessive power consumption and delay. Note
that we have associated the SN and the DN with the node
indices 1 and Nnodes , respectively, in the context of this treatise.
Additionally, these legitimate routes are mapped to a specific
index x under lexicographic ordering using Lehmer Encoding2
[33]. The route’s overall delay D(x) is considered as one
of our UFs, which is quantified in terms of the number of
hops established by the route. This is formally formulated as
follows:
D(x) = |xr | − 1,
(1)
where the operator |·| corresponds to the number of nodes
along the route xr including the SN and DN. Moving on to the
x-th route’s overall power consumption P (x), it is proportional
to the sum of path-losses incurred by each of the individual
links constituting the route. Explicitly, the path-loss LdB (i, j)
quantified in dB for a single link between the i-th and the j-th
nodes is equal to [30]:
4πdi,j
LdB (i, j) = PT x,ij − PRx,ij = 10α log10
, (2)
λc
where α corresponds to the path-loss exponent, di,j is the
distance between the two nodes and λc denotes the carrier’s
wavelength. In our case-study we have set α = 3 and
λc ≃ 0.125 m corresponding to a frequency of fc = 2.4 GHz.
Consequently, the second UF is formulated as follows:
|xr |−1
L(x) =
X
(i)
10LdB(xr
,xr(i+1) )/10
.
(3)
i=1
Figure 1: Exemplified WMHN topology associated with 5
nodes. The presence of a cluster-head in possession of a
quantum computer is considered at the DN side as in [19]
and in [30]. The interference levels experienced by each node
are presented in the legend.
Based on the network information gathered, the WMHN
cluster-head has to identify the optimal routes emerging from
the SN to the DN based on certain Utility Functions (UF).
Similar to [19] and [30], we have jointly taken into account
the route’s overall delay, its overall power consumption and its
Moving on to the final UF, namely the BER, let us first
elaborate on the interference levels experienced by the nodes.
In our specific scenario, there is only a single pair of source
and destination nodes, resulting in a single route being active. Additionally, we have assumed that the WMHN has
a sufficient number of orthogonal spreading codes and subcarriers for the sake of efficiently separating the routes as
in [32]. In this context, there is no interference stemming
from the WMHN itself; however, we have assumed that a
sufficiently high number of users access the channel, hence
the resultant interference can be treated as Additive White
Gaussian Noise (AWGN), owing to the Central Limit Theorem
(CLT) [34]. Therefore, the interference is modeled by a
random Gaussian process, with its mean set to -90 dBm and its
standard deviation set to 10 dB, while the transmission power
is set to PT x = 20 dBm. Additionally, the nodes transmit
their messages using the uncoded QPSK scheme [35] over
uncorrelated Rayleigh fading channels and utilize Decode-andForward relaying [36] for forwarding the respective messages.
Based on these assumptions, we can readily use the closedform BER performance of the adopted scheme versus the
2 Lehmer Encoding maps a specific permutation to an index in the factoradic
basis [33].
4
IEEE TRANSACTIONS ON COMMUNICATIONS 2018
received Signal-to-Noise Ratio (SNR), while the overall route’s
BER Pe (x) can be calculated using the following recursive
formula [19]:
Pe,tot = Pe,1 + Pe,2 − 2Pe,1 Pe,2 ,
(4)
which corresponds to the output BER Pe,tot of a two-stage
Binary Symmetric Channel (BSC) [19], where Pe,1 and Pe,2
represent the BER associated with the first and the second
stage, respectively.
Having described the UFs considered, let us now proceed by
defining our optimization problem. Explicitly, we will jointly
consider the UFs in the form of a Utility Vector (UV) f (x),
which is defined as follows:
f (x) = [Pe (x), L(x), D(x)] ,
(5)
where D(x) and L(x) correspond to the x-th route’s delay
and power consumption defined in Eqs. (1) and (3), while
Pe (x) denotes the x-th route’s end-to-end BER, which is
recursively evaluated using Eq. (4). Explicitly, we opt for
jointly minimizing the entire set of UFs considered by the
UV of Eq. (5). Therefore, for the evaluation of the fitness of
the UVs we will utilize the concept of Pareto Optimality3 [17],
which is encapsulated by Definitions 1 and 2.
Definition 1. Pareto Dominance [17]: A particular route xi
associated with the UV f (xi ) = [f1 (xi ), . . . , fK (xi )], where
K is the number of the UFs considered, is said to strongly
dominate another route xj associated with the UV f (xj ) =
[f1 (xj ), . . . , fK (xj )], denoted by f (xi ) ≻ f (xj ), iff we have
fk (xi ) < fk (xj ), ∀k ∈ {1, . . . , K}. Equivalently, the route
xi is said to weakly dominate another route xj , denoted by
f (xi ) f (xj ), iff we have fk (xi ) ≤ fk (xj ), ∀k ∈ {1, . . . , K}
and ∃k ′ ∈ {1, . . . , K}, so that we have fk′ (xi ) < fk′ (xj ).
Definition 2. Pareto Optimality [17]: A particular route xi
associated with the UV f (x1 ) is Pareto-optimal, iff there is
no route that dominates xi , i.e. we have ∄xj so that f (xj ) ≻
f (xi ) is satisfied. Equivalently, the route xi is strongly Paretooptimal iff there is no route that weakly dominates xi , i.e. we
have ∄xj , so that f (xi ) f (xj ) is satisfied.
Explicitly, Definition 1 provides us with the criterion for
evaluating the fitness of a specific route with respect to another
reference route, while Definition 2 outlines the condition
of the specific route’s optimality. Based on the number of
routes dominating a specific route, it is possible to group the
routes into the so-called Pareto Fronts (PF). Explicitly, the
PF comprises the Pareto-optimal routes, which are dominated
by no other routes according to Definition 2, which is often
referred to as the Optimal Pareto Front (OPF).
In our application, our aim is to identify the entire set
of weakly Pareto-optimal routes for the sake of gaining
insight into the routing trade-offs associated with the UFs
considered. Naturally, for the sake of identifying a specific
route as Pareto-optimal we have to perform precisely (N − 1)
Pareto-dominance comparisons, where N corresponds to the
total number of legitimate routes. Therefore, the complexity
3 The readers should refer to [32] for a more detailed tutorial on Pareto
optimality.
imposed by the exhaustive search aiming for identifying the
entire set of routes belonging to the OPF is on the order of
O(N 2 ). Explicitly, the total number N of legitimate routes increases exponentially as the number Nnodes of nodes increases
[19], hence rendering the multi-objective routing problem as
NP-hard. Thus sophisticated methods are required for finding
all of the solutions.
Let us now proceed by elaborating on our novel dynamic
framework designed for efficiently exploring the search space.
III. M UTLI -O BJECTIVE ROUTING DYNAMIC
P ROGRAMMING F RAMEWORK
Before delving into the analysis of our multi-objective dynamic programming framework, which is specifically tailored
for our routing problem, we will express each of the UFs
considered in the UV of Eq. (5) as a weighted sum of the
specific UFs associated with the individual links comprised
by a particular route. Explicitly, the power consumption has
already been expressed in this form based on Eq. (3). As for
the delay, which we have defined as the number of hops, it
may be redefined as follows:
D(x) =
|xr |−1
X
i=1
1 − δx(i) ,x(i+1) ,
r
(6)
r
where δi,j corresponds to the Kronecker delta function [37],
while xr and x represent the route and its associated index,
respectively. As for the route’s overall BER, the recursive
formula of Eq. (4) may be approximated as follows:
|xr |−1
|xr |−1
Pe (x) =
X
i=1
Pe,x(i) ,x(i+1) −ǫ(x) ≈
r
r
X
Pe,x(i) ,x(i+1) , (7)
r
r
i=1
where Pe,k,l represents the BER of the specific link established between the k-th and the l-th nodes, while ǫ(x) is the
approximation error, which is on the order of:
r |−1 |xr |−1
|xX
X
Pe,x(j) ,x(i+1) Pe,x(j) ,x(j+1) .
ǫ(x) = O
r
r
r
r
i=1
(8)
j=1
j 6= i
Since the sum of the products of all the links’ BER will
be several orders of magnitude lower than their sum, the
approximation error of Eq. (7) may be deemed to be negligible.
Having expressed the UFs considered as a weighted sum
of the UFs associated with their links, we may now proceed
by exploiting this specific property for the sake of achieving a
further complexity reduction. In fact, it is possible to transform
our composite multi-objective routing problem into a series of
smaller subproblems, thus arriving at a dynamic programming
structure. This transformation is performed with the aid of
Definition 3 in conjunction with Proposition 1.
Definition 3. A specific route x = {SN →R̄i →DN } is said
(j)
to generate another route xg by inserting the single RN Rj
node between the previous RN and the DN. Explicitly, the
(j)
(j)
resultant route xg is xg = {SN →R̄i →Rj →DN }, ∀j ∈
{1, . . . , Nnodes − 2}.
D. ALANIS ET AL.: A QUANTUM-AIDED DYNAMIC PROGRAMMING FRAMEWORK FOR PARETO OPTIMAL ROUTING IN WMHNS
Proposition 1. Let us consider a specific route
x
=
{SN →R̄i →DN } associated with the UV
f (x) = [f1 (x), . . . , fK (x)] and its sub-route x′ = {SN →R̄i }
associated with the UV f (x′ ) = [f1 (x′ ), . . . , fK (x′ )]. Let us
assume furthermore that each component fk (x) of the UV
associated with the route x has a positive value and that
it can be expressed as a sum of the respective UFs of its
individual links xi,i+1 , i.e. we have:
|x|−1
fk (x) =
X
fk (xi,i+1 ),
(9)
i=1
with fk (xi,i+1 ) > 0, ∀ k, i, x : k ∈ {1, ..., K}, i ∈
{1, ..., |x| − 1}, x ∈ S, where K and S are the number
of optimization objectives and the set of legitimate routes,
respectively. The route x cannot generate any Pareto-optimal
routes using the rule of Definition 3 if there is a route
xd = {SN →R̄j →DN } from the SN to the DN associated
with R̄j 6= R̄i that weakly dominates the sub-route x′ , i.e. if
we have ∃xd ∈ S : f (xd ) f (x′ ). The respective proof is
presented in Appendix A.
Explicitly, Proposition 1 guarantees that a specific route
x = {SN →R̄i →DN } comprised by the sub-route x′ =
{SN →R̄i } cannot generate Pareto-optimal routes by adding
an intermediate RN to x between its last RN and the DN, if
the sub-route x′ is weakly dominated by any of the legitimate
routes. Explicitly, should its sub-route x′ be sub-optimal, the
respective route x will be sub-optimal as well, since we have
∃xd ∈ S : f (xd ) f (x′ ) ≻ f (x), based on Proposition 1.
Note that the opposite of this statement does not apply, since
there exist sub-optimal routes, whose sub-routes are indeed
Pareto-optimal.
This specific property can be exploited for the sake of reducing the search-space size required for identifying the entire
OPF. To elaborate further, we can devise an irregular trellis
graph [38] for the sake of guiding the search space exploration,
as portrayed in Fig. 2 for the 5-node WMHN of Fig. 1. Note
however that this specific trellis graph is different from those
used for channel coding in [38], since in the latter we only
have as many legitimate paths as many legitimate symbols. By
contrast, here all transitions represent legitimate routes in our
scenario. Additionally, we rely on Definition 3 for the sake of
determining the possible trellis-node transitions. For instance,
observe in Fig. 2 that a trellis-path emerging from the trellisnode associated with the generator route {1 → 2 → 5} is
only capable of visiting the nodes associated with the routes
{1 → 2 → 3 → 5} and {1 → 2 → 4 → 5}, since a single
RN is inserted before the DN into the generator route based
on Definition 3. Moving on to the next trellis stages, during
the i-th trellis stage the following three steps are carried out:
gen
1) Surviving Routes: The set S(i)
of generated routes are
surv
constructed based on the set S(i−1) of surviving routes of the
previous stage and relying on Definition 3.
OPF
2) Pareto-Optimal Routes: The set S(i)
of Pareto-optimal
routes is identified based on the following optimization prob-
5
lem:
OPF
S(i)
=
argmin
{f (x)},
gen
OPF
∪S(i−1)
x∈S(i)
gen
s.t. ∄j ∈ S(i)
OPF
∪ S(i−1)
: f (j) ≻ f (x).
(10)
Note that the optimization problem of Eq. (10) considers the
gen
joint search space constituted by the all the routes S(i)
of
the i-th trellis stage as well as by the Pareto-optimal routes
OPF
S(i−1)
of the previous stage. Using recursion, we can readily
OPF
observe that the Pareto-optimal routes S(i−1)
of the previous
stage contain the Pareto-optimal routes across all stages up to
the (i−1)-st stage. This property is beneficial for our dynamic
programming framework, since it eliminates the need for
backwards propagation, thus only requiring the employment
of a feed-forward method for the identification of the entire
OPF.
surv
3) Surviving Routes: The set S(i)
of surviving routes is
identified based on the following optimization problem:
surv
S(i)
=
argmin{f (x)},
gen
x∈S(i)
gen
OPF
s.t. ∄j ∈ S(i)
∪ S(i−1)
: f (j) f (x′ ).
(11)
where x′ corresponds to the particular sub-route of x, having
all the links of x, except for the last hop, as detailed in
Proposition 1.
The optimization process proceeds to the next trellis stage
as long as either there exist surviving routes, i.e. we have
surv
S(i)
6= ∅, or if the maximum affordable number of trellis
stages - which is equal to the maximum number of hops of
the legitimate routes - has not been exhausted. Otherwise,
the optimization process terminates by exporting the hitherto
identified OPF.
Let us now proceed by elaborating on the route exploration
process using the 5-node WMHN example of Fig 1. Its
respective trellis is portrayed in Fig 2, while the routes’ and
their respective sub-route’s UVs are shown in Table I. Initially,
gen
the optimization process considers the set S(1) of routes, which
is constituted by all the legitimate routes having a single and
two hops, namely the routes {1 → 5}, {1 → 2 → 5},
{1 → 3 → 5} and {1 → 4 → 5}, as portrayed in the
1st trellis stage of Fig. 2. Based on Table I, all the routes
considered are Pareto-optimal and thus the respective set is
gen
OPF
equal to S(1)
= S(1)
. Subsequently, the set of surviving nodes
is constructed. Explicitly, the direct route is not considered
in this case, since its inclusion leads to the generation of
routes, which have already been processed. Observe in Table I
that all the routes constituted by 2 hops have Pareto optimal
sub-routes and hence the set of surviving routes becomes
surv
S(1)
= [{1 → 2 → 5}, {1 → 3 → 5}, {1 → 4 → 5}].
surv
After the identification of the set of surviving routes S(1)
,
gen
nd
the set S(2) of routes generated in the 2 trellis stage is
created by including an appropriate RN right before the
DN, as annotated with the aid of black arrows in Fig 2.
Naturally, since all the routes constituted by two hops
have been identified as being Pareto-optimal, the entire set
of routes having three hops is visited by the trellis-paths
OPF
in the 2nd trellis stage, as seen in Fig. 2. The set S(1)
st
of Pareto-optimal routes of the 1 trellis stage is then
6
IEEE TRANSACTIONS ON COMMUNICATIONS 2018
Figure 2: Irregular trellis graph designed for guided search-space exploration for the 5-node WMHN of Fig. 1 using the
optimal dynamic programming framework, encapsulated by Definition 3 and Proposition 1. Note that the UVs of each route
are presented in Table I.
gen
concatenated to the set S(2)
of the routes generated in the
nd
OPF
2 trellis stage and the set S(2)
of Pareto-optimal routes
OPF
is identified. After this operation, the latter is set to S(2)
=
[{1 → 5}, {1 → 2 → 5}, {1 → 3 → 5}, {1 → 4 → 5} ,
{1 → 3 → 2 → 5}],
hence
including
the
route
{1 → 3 → 2 → 5} to the OPF, as denoted with the
aid of the bold rectangle in Fig. 2. The surviving routes of
the 2nd trellis stage are then identified using the optimization
problem of Eq. (11). Explicitly, they constitute the set
surv
S(2)
= {1 → 3 → 2 → 5}, {1 → 3 → 4 → 5}, {1 → 4 →
2 → 5}, {1 → 4 → 3 → 5}, as it may be verified by Table I
and denoted with the aid of the gray-filled nodes of Fig. 2.
In the presence of surviving nodes, the optimization process
proceeds with the final trellis stage; however, in this case the
routes {1 → 2 → 3 → 4 → 5} and {1 → 2 → 4 → 3 → 5}
are not considered, since their generators do not have Paretooptimal sub-routes. This is portrayed in Fig. 2 with the aid
both of the gray dashed arrows and of the gray dashed nodes.
gen
Hence, the set S(3)
= {1 → 3 → 2 → 4 → 5}, {1 → 3 →
4 → 2 → 5}, {1 → 4 → 2 → 4 → 5}, {1 → 4 → 3 →
OPF
2 → 5} is generated. The set S(2)
is then concatenated to
that of the routes generated in the final trellis stage and the
OPF
final set S(3)
of Pareto-optimal routes is identified. Explicitly,
the latter is identical to the respective set of the 2nd trellis
stage, since none of the routes generated in the final stage
is Pareto-optimal, as verified by Table I. Additionally, since
we have reached the final stage, the set of surviving routes is
not identified and the process exits by exporting the hitherto
observed OPF.
In a nutshell, this route exploration process succeeds in
transforming the multi-objective routing problem into a series
of significantly less complex sub-problems, each corresponding to a single trellis stage, hence inheriting the structure of
dynamic programming problems [10]. Note that the metricaccumulation, which is typical in dynamic programming problems, is constituted by the update of the Pareto-optimal routes.
Note that this dynamic programming framework is optimal in
terms of its efficacy in identifying the entire OPF, just like
the exhaustive search method. Primarily, this is a benefit of
Proposition 1, which excludes the routes that are incapable of
generating Pareto-optimal routes during the next trellis stages.
IV. E VOLUTIONARY Q UANTUM PARETO O PTIMIZATION
In Section III, we introduced a novel dynamic programming
framework for the sake of guiding the search process in
identifying the Pareto-optimal routes, thus effectively reducing
the complexity. In this section, we exploit this framework
and further improve it with the aid of our EQPO algorithm.
More specifically, we have relaxed the dynamic programming
framework of Section III for the sake of striking a better
accuracy versus complexity trade-off. Additionally, we have
improved the quantum-assisted process of [30] for identifying
D. ALANIS ET AL.: A QUANTUM-AIDED DYNAMIC PROGRAMMING FRAMEWORK FOR PARETO OPTIMAL ROUTING IN WMHNS
7
Table I: Utility Vectors of the legitimate routes and of their respective sub-routes for the 5-node WMHN topology of Fig. 1.
Route x
{1 5}
{1 2 5}
{1 3 5}
{1 4 5}
{1 2 3 5}
{1 2 4 5}
{1 3 2 5}
{1 3 4 5}
{1 4 2 5}
{1 4 3 5}
{1 2 3 4 5}
{1 2 4 3 5}
{1 3 2 4 5}
{1 3 4 2 5}
{1 4 2 3 5}
{1 4 3 2 5}
[4.52
[2.52
[2.35
[1.43
[9.49
[1.91
[1.36
[1.29
[1.42
[1.46
[1.36
[1.94
[1.90
[1.28
[1.49
[1.45
Route UV
10−4 , 74.15, 1]
10−4 , 73.10, 2]
10−4 , 70.89, 2]
10−2 , 71.76, 2]
10−4 , 76.09, 3]
10−2 , 75.72, 3]
10−4 , 69.55, 3]
10−2 , 71.74, 3]
10−2 , 71.19, 3]
10−2 , 73.50, 3]
10−2 , 76.36, 4]
10−2 , 76.50, 4]
10−2 , 74.13, 4]
10−2 , 71.18, 4]
10−2 , 75.23, 4]
10−2 , 72.82, 4]
Sub-route UV
[∞, ∞, ∞]
[2.52 10−4 , 73.10, 1]
[3.13 10−5 , 57.30, 1]
[1.41 10−2 , 67.50, 1]
[7.45 10−4 , 74.61, 2]
[1.89 10−2 , 74.46, 2]
[1.36 10−4 , 69.54, 2]
[1.28 10−2 , 67.46, 2]
[1.42 10−2 , 71.19, 2]
[1.44 10−2 , 70.27, 2]
[1.34 10−2 , 75.30, 3]
[1.92 10−2 , 75.18, 3]
[1.88 10−2 , 72.18, 3]
[1.28 10−2 , 71.17, 3]
[1.47 10−2 , 73.35, 3]
[1.45 10−2 , 72.81, 3]
the Pareto-optimal routes, so that it becomes capable of “remembering” the OPF identified in the previous trellis stages.
We will refer to this improved quantum-assisted process as the
Preinitialized-NDQIO (P-NDQIO) algorithm. In this context,
the P-NDQIO and the EQPO algorithms are presented in
Sections IV-A and IV-B, respectively. Let us now proceed by
presenting the P-NDQIO algorithm.
Optimal Sub-route
X
X
X
X
X
X
X
X
X
Algorithm 1 Preinitialized Non-Dominated Quantum Iterative
Optimization Algorithm (P-NDQIO)
1:
2:
3:
4:
A. Preinitialized NDQIO algorithm
5:
The P-NDQIO algorithm, which is formally stated in Alg. 1,
is the main technique of memorization [10], thus providing a
significant complexity reduction by remembering and propagating the OPF identified across the previous trellis stages
to the next ones. Its memorization is performed in Step 1 of
Alg. 1, where the OPF of the current trellis stage is initialized
to that of the previous stage. Subsequently, the P-NDQIO
algorithm performs its iterations, looking for Pareto-optimal
routes in Steps 2-14 of Alg. 1.
During each iteration, which results in identifying a single
Pareto-optimal route, the P-NDQIO algorithm first invokes
the so-called Boyer-Brassard-Hoyer-Tapp Quantum Search
Algorithm (BBHT-QSA) [25] for the sake of identifying routes
that are not dominated by any of the routes belonging to
the hitherto identified OPF. We refer to this process as the
Backward BBHT-QSA (BW-BBHT-QSA) process [30]. If an
invalid route-solution - i.e. a route that is indeed dominated
by the OPF identified so far - is output by the BBHT-QSA,
the P-NDQIO algorithm concludes that the entire OPF has
been identified. However, since the BBHT-QSA exhibits a low
probability of failing to identify a valid solution4, the BWBBHT-QSA step is repeated for an additional iteration in order
to ensure the detection of the entire OPF, as seen in Steps 12
and 14 of Alg. 1. Otherwise, should a valid route-solution be
identified by the BW-BBHT-QSA step, this specific route is
6:
4 We define a valid route-solution as the specific route that satisfies the
condition in Step 5 of Alg. 1
Optimal Route
X
X
X
X
7:
8:
9:
10:
11:
12:
13:
14:
15:
OPF
OPF
Initialize the OPF to S(i)
← S(i−1)
.
repeat
T ← 0.
Invoke the BBHT-QSA of [30, Alg. 1] searching for
gen
routes in S(i)
that are not dominated by any of the
OPF
routes of S(i) and output xs .
OPF
if f (j) ⊁ f (xs ), ∀j ∈ S(i)
then
repeat
Set j ← xs .
Invoke the BBHT-QSA of [30, Alg. 1] searching
gen
for routes in S(i)
that dominate the route j and
output xs .
until f (xs ) ⊁ f (j).
OPF
Discard the routes from S(i)
that are dominated by
the route j and append it to the OPF.
else
Set T ← T + 1.
end if
until T = 2
OPF
Export the S(i)
and exit.
classified as “potentially” being Pareto-optimal. Consequently,
the P-NDQIO algorithm invokes the so-called BBHT-QSA
chain process [19], [30] in Steps 6-9 of Alg. 1. Explicitly, the
output of the BW-BBHT-QSA is set as the initial reference
solution in Step 7 of Alg. 1 and a BBHT-QSA process is
activated in Step 8 of Alg. 1, which searches for routes
that dominate the reference one. If a route that dominates
the reference one is found, the reference route is updated
to the BBHT-QSA output and a new BBHT-QSA process is
activated. Naturally, the activation of the BBHT-QSA process
is repeated until a particular route is output by the BBHT-QSA
that does not dominate the reference route, thus indicating that
8
IEEE TRANSACTIONS ON COMMUNICATIONS 2018
the reference route is Pareto-optimal. Subsequently, the ParetoOPF
optimal routes of the set S(i)
are checked as to whether they
are dominated by the reference route, so that they are removed
OPF
and the reference route is then included in S(i)
, as seen in
Step 10 of Alg. 1. Explicitly, this check, which is referred to as
the OPF Self-Repair (OPF-SR) process in [30], provides the
EQPO algorithm with resilience against including sub-optimal
routes in the early trellis stages due to the limited number of
generated routes, hence preventing their propagation to the
later stages.
Both the BW-BBHT-QSA process and the BBHT-QSA
chains are parts of the original NDQIO algorithm; thus, the PNDQIO algorithm employs quantum circuits that are identical
to those of the NDQIO algorithm. Therefore, the motivated
readers may refer to [30] for extended discussions.
B. EQPO algorithm
The dynamic framework introduced in Section III, albeit
optimal in terms of its capability of identifying the entire OPF,
it may impose an excessive complexity quantified in terms of
the number of dominance comparisons required for solving the
optimization problem of Eq. (11). To elaborate further, as the
number of UFs considered increases, the number of surviving
routes is increased due to the differences among the UFs.
This in turn leads to the proliferation of the number of routes
generated per trellis stage. However, only a relatively small
fraction of the surviving route-population leads eventually to
generating Pareto-optimal routes in the next trellis stages.
Therefore, the employment of the optimal dynamic framework
presented in Section III imposes a significant complexity
overhead for the sake of ensuring the detection of the entire set
of Pareto-optimal routes. Consequently, a performance versus
complexity trade-off has to be struck for the sake of mitigating
this complexity overhead. In fact, this specific balance is struck
in the context of the EQPO algorithm by jointly relying on
Relaxations 1 and 2.
Relaxation 1. A route can only generate optimal routes
based on Definition 3, if it is Pareto-optimal. This is formally
formulated as follows:
OPF
OPF
surv
.
− S(i−1)
, S(i)
S(i)
(12)
surv
of the surviving routes
Relaxation 1 restricts the set S(i)
at the end of the i-th trellis stage to the set of the newlydiscovered Pareto-optimal routes at this specific trellis stage.
This relaxation provides beneficial complexity reduction, since
surv
it makes the identification both of the set S(i)
of surviving
OPF
routes and of the set S(i) of Pareto-optimal routes possible
by simply solving the optimization problem of Eq. (10).
Explicitly, Proposition 1 does not conflict with Relaxation 1,
since the Pareto-optimal routes are guaranteed to have Paretooptimal sub-routes. This is justified by the fact that the subroutes dominate their routes due to the absence of the final hop,
which results in increasing all the UFs considered. Thus, since
there exist no route from the SN to the DN dominating the
route identified, there exist no routes dominating the respective
sub-route either. However, the complexity reduction offered
by Relaxation 1 comes at the price of reduced accuracy,
since sub-optimal routes having Parero-optimal sub-routes do
exist, which might potentially lead to the generation of Paretooptimal routes in the next trellis stages. This specific limitation
is mitigated with the aid of Relaxation 2.
Relaxation 2. For the sake of facilitating the identification of
all Pareto-optimal routes, Definition 3 is relaxed as follows:
(j,k)
a specific route x is said to generate another route xg
by
inserting the single RN Rj between the k-th and the (k + 1)-st
nodes.
gen
of generated routes, which
Relaxation 2 extends the set S(i)
surv
are created by the set S(i−1) of surviving routes of the previous
trellis stage. This is realized by replacing a single direct link
established either by two RNs or by an RN and the DN with
an indirect link involving an appropriate RN as an intermediate relay. Naturally, this specific modification enhances the
heuristic accuracy of the EQPO algorithm, since it allows the
generation of additional routes, thus acting similarly to the
mutation operation of genetic algorithms [39].
Algorithm 2 Evolutionary Quantum Pareto Optimization
(EQPO) Algorithm.
1:
2:
3:
4:
5:
6:
7:
8:
9:
gen
gen
gen
OPF
surv
Set S(0) ← {SN → DN }, S(0)
← S(0) , S(0)
← S(0) ,
i ← 0.
repeat
Set i ← i + 1.
gen
surv
Generate the set of routes S(i)
from the set S(i−1)
based
on Relaxation 2 by appropriately inserting a single RN
between two intermediate nodes.
gen
gen
OPF
← S(i)
∪ S(n−1)
.
Set S(i)
gen
Invoke the P-NDQIO algorithm of Alg. 1 in the set S(i)
OPF
OPF
and initialize the identified OPF to S(n)
← S(n−1)
.
surv
OPF
OPF
Set S(i) ← S(n) − S(n−1) .
surv
= 0 or i = Nnodes − 1
until S(i)
OPF
Export the OPF S(n)
and terminate.
Let us now proceed by elaborating on the specifics of the
EQPO algorithm, which is formally presented in Alg. 2. To
elaborate further, in Step 1 of Alg. 2 the EQPO algorithm
initializes the set of routes generated, the Pareto-optimal routes
as well as the surviving routes to the direct route, i.e. to the
route {SN → DN }. It then proceeds with the trellis stages
using Steps 2-8 of Alg. 2. During each trellis stage, the set
gen
S(i)
of generated routes is constructed in Step 4 of Alg. 2
relying on Relaxation 2. Upon applying Relaxations 1 and 2
in the trellis of Fig. 2 results in the trellis of Fig. 3.
OPF
This set is then concatenated with the set S(i−1)
of Paretooptimal routes identified in the previous stage. Subsequently,
the P-NDQIO algorithm is invoked in Step 6 of Alg. 2 for
OPF
the sake of identifying the set S(i)
of Pareto-optimal routes
gen
surv
from the set S(i) . Then, the set S(i) of surviving routes is
determined in Step 7 of Alg. 2, relying on Relaxation 1.
More specifically, the steps carried out as part of the EQPO
algorithm’s dynamic programming framework during a single
trellis stage are listed as follows:
D. ALANIS ET AL.: A QUANTUM-AIDED DYNAMIC PROGRAMMING FRAMEWORK FOR PARETO OPTIMAL ROUTING IN WMHNS
9
Figure 3: Irregular trellis graph designed for guided search-space exploration for the 5-node WMHN of Fig. 1 using the EQPO
algorithm’s dynamic programming framework, encapsulated by Relaxations 1 and 2. Note that the UVs of each route are
presented in Table I
gen
1) Route Generation: EQPO creates the set S(i) of routes
surv
based on the set S(i−1)
of surviving routes from the previous trellis stage using Relaxation 2, as seen in Step 4
of Alg. 2. For instance, observe in Fig. 3 that the route
{1 → 2 → 5} is capable of generating 4 routes, namely
the routes {1 → 2 → 3 → 5}, {1 → 2 → 4 → 5},
{1 → 3 → 2 → 5}, {1 → 4 → 2 → 5}. By contrast,
Definition 3 allows the generation of only the first two routes,
as portrayed in Fig. 2. Additionally, in contrast to the optimal
dynamic programming framework of Section III, each route of
the current trellis stage in Fig. 3 can be generated by multiple
surviving routes of the previous stage. This specific feature of
Relaxation 2 enhances the heuristic accuracy, since it enables
the generation of potentially Pareto-optimal routes, which have
suboptimal constructors and hence would be disregarded based
on Relaxation 1.
2) Pareto-Optimal and Surviving Routes: Following the
gen
construction of the set S(i)
of the routes generated, the EQPO
algorithm invokes the P-NDQIO algorithm of Section IV-A
in Step 6 of Alg. 2 in order to search for new Paretogen
optimal routes belonging to the set S(i)
. However, based on
Definition 2, the optimality of the route depends on the set
OPF
of eligible routes considered. Consequently, the OPF S(i−1)
hitherto identified across all the previous trellis stages has to
gen
be concatenated with S(i)
in Step 5 of Alg. 2, thus ensuring
that the routes identified as optimal by the P-NDQIO algorithm
are indeed Pareto-optimal with respect to the entire set of
OPF
legitimate routes. Note that the set S(i)
contains the Paretooptimal routes across all trellis stages all the way up to the
i-th one, as in the optimal dynamic programming framework
of Section III. Consequently, using Relaxation 1 the Paretooptimal routes identified at the current trellis stage are considered as surviving routes. Note that the Pareto-optimal routes
identified throughout the previous stages are not taken into
account, since they would generate routes already processed
during the previous trellis stages.
The EQPO algorithm continues processing the trellis stages
either until it reaches a trellis stage having no surviving paths
or when the maximum affordable number of trellis stages
is exhausted, in a similar fashion to the optimal dynamic
programming framework of Section III.
Let us now highlight the differences between the trellises of
Figs. 2 and 3 considering the 5-node example of Fig. 1. Note
that the same annotation is used in Fig. 3 as that of Fig. 2
Explicitly, based on Eq. (12), the EQPO algorithm classified
the specific routes, which are Pareto-optimal as being “ParetoOptimal” and those that have been generated in the current
stage as “Visited & Surviving”. Hence in contrast to Fig. 2,
they are equivalent in Fig. 3. Similar to the optimal dynamic
programming framework of Section III, the EQPO algorithm
gen
of generated routes to the set of the
initializes the set S(1)
legitimate routes having either single or two hops, as portrayed
in the 1st trellis stage of Fig. 3. Based on Table I, all the
routes having two hops are Pareto-optimal and thus the EQPO
algorithm classifies them as the surviving routes of the 1st
trellis stage, as seen in Fig. 3. Similar to Fig. 2, the EQPO
algorithm’s trellis paths visit the entire set of routes having
three hops and then the algorithm identifies the route {1 →
10
IEEE TRANSACTIONS ON COMMUNICATIONS 2018
3 → 2 → 5} as Pareto-optimal with the aid of the P-NDQIO
algorithm. Consequently, this specific route is deemed to be
the sole surviving route in Fig 3. This is in contrast to Fig. 2,
where three more routes have been identified as surviving ones.
Recall from Fig. 2 that these routes do not lead to Paretooptimal routes in the last trellis stage. This in turn results in
the EQPO algorithm visiting one less route in the 3rd trellis
stage, i.e. not considering the sub-optimal route {1 → 4 →
2 → 3 → 5} as potentially Pareto-optimal.
V. C OMPLEXITY VERSUS ACCURACY D ISCUSSIONS
In this section, we will characterize the complexity imposed
by the EQPO Alg. presented in Alg. 2 and evaluate its heuristic
accuracy versus the complexity invested. Additionally, note
that since we had no quantum computer at our disposal, the
simulations of the QSAs were carried out using a classical
cluster. Explicitly, since the quantum oracle gate O [27]
calculates in parallel the UF vectors of all the legitimate routes
in the QD, they were pre-calculated. We note that this results
in an actual complexity higher than that of the full-search
method. Therefore, the employment of the quantum algorithms
in a quantum computer is essential for observing a complexity
reduction as a benefit of the QP. Hence, in our simulations, we
have made the assumption of employing a quantum computer
and we count the total number of O activations for quantifying
the EQPO’s complexity. This number would be the same for
both classical and quantum implementations. Note that in the
x
x
,
following analysis we will use the notation N(i)
≡ S(i)
x
x
where N(i) corresponds to the cardinality of the set S(i) .
Furthermore, our simulation results have been averaged over
108 runs. During each run we have randomly generated the
node’s locations as well as the interference levels experienced
by them with the aid of the respective distributions mentioned
in Section II. We have ensured that each run is uncorrelated
with the rest of the runs.
Let us now proceed by analytically characterizing the complexity imposed by our proposed algorithm.
A. Complexity
We will first characterize the complexity imposed by the
EQPO algorithm’s dynamic progamming framework, when
the exhaustive search is employed instead of the P-NDQIO
algorithm in Step 6 of Alg. 2. We will refer to this method
as the Classical Dynamic Programming (CDP) method and
we will use it as a benchmarker for assessing the complexity
reduction offered by the QP.
Prior to characterizing the EQPO algorithm and the CDP
surv
method we will analyze the the orders of the number N(i)
of
OPF
the surviving routes and of the number N(i) of the Paretooptimal routes identified across the first i trellis stages. As far
OPF
as the number N(i)
of the Pareto-optimal routes identified
across the first i trellis stages is concerned, the trellis graph
guiding the search for Pareto-optimal routes identifies more
Pareto-optimal routes, as it proceeds through more trellis
stages. Explicitly, its order can be formally expressed as
follows:
OPF
O(N(i)
) = O(ai NOPF ) = O(NOPF ), ∀i ∈ {1, . . . , Nnodes −1},
(13)
where ai corresponds to the fraction of the OPF identified by
the first i trellis stages. Naturally, this fraction ai approaches
unity as the number i of trellis stages moves closer to the
maximum number of hops.
surv
Moving on to the number N(i)
of the surviving routes
at the i-stage, it is equal to the number of Pareto-optimal
routes identified at the i-th trellis stage, based on Relaxation 1.
surv
OPF
is a fraction of the total number N(i)
of
Explicitly, N(i)
the Pareto-optimal routes identified across the first i trellis
surv
OPF
with bi ≤ 1
= bi N(i)
stages. Hence, we have N(i)
surv
∀i ∈ {1, . . . , Nnodes − 1}, since the set S(i) of Pareto-optimal
OPF
routes at the i-th trellis stage is included in the set S(i)
of Pareto-optimal routes identified at the first i trellis stages.
surv
) as follows:
Therefore we can evaluate the order O(N(i)
(13)
surv
OPF
O(N(i)
) = O(bi N(i)
) = O(bi ai NOPF ) = O(NOPF ). (14)
Consequently, in Eqs. (13) and (14), we have upper bounded
surv
the order O(N(i)
) of the number of surviving routes at the
OPF
i-th stage as well as the order O(N(i)
) of the number of
Pareto-optimal routes identified at the first i stages by the order
O(NOPF ) of the total number of Pareto-optimal routes, i.e. we
surv
OPF
have O(N(i)
) = O(N(i)
) = O(NOPF ). Naturally, Eq. (13)
and (14) will facilitate the complexity analysis, since they
render the aforementioned orders independent of the index i
of the trellis stages. Let us now proceed by characterizing the
complexity imposded by the CDP method.
1) CDP method’s complexity: Let us assume that there is a
gen
total of N(i)
generated routes arriving at the i-th trellis stage.
These particular routes are generated by the specific Paretooptimal routes identified at the previous trellis stage, which are
surv
N(i−1)
in total. Based on the aforementioned assumptions, the
number of generated routes arriving at the i-th trellis stage is
formulated as follows:
i
h
gen
surv
surv
Nnodes i ,
(Nnodes − 1 − i)i = O N(i−1)
N(i)
= N(i−1)
(14)
= O [NOPF Nnodes i] .
(15)
Since the set of Pareto-optimal routes of the previous trellis
stage are concatenated to the set of generated routes in Step 5
of Alg. 2, the total number of routes considered at the i-th
trellis stage is given by:
gen
OPF
routes
= N(i)
+ N(i−1)
N(i)
(13),(15)
=
O [NOPF Nnodes i] .
(16)
routes 2
Additionally, the CDP method performs O[(N(i)
) ] dominance comparisons, which we will refer to as the Cost
Function Evaluation (CFE), since each generated route has
to be compared to all of the routes considered. Therefore, the
total complexity imposed by the CDP method across all trellis
stages may be quantified in terms of the number of dominance
D. ALANIS ET AL.: A QUANTUM-AIDED DYNAMIC PROGRAMMING FRAMEWORK FOR PARETO OPTIMAL ROUTING IN WMHNS
comparisons, which is formulated as follows:
Nnodes
2
X−1
2
5
routes
= O(NOPF
Nnodes
),
O N(i)
LCDP =
(17)
i=1
where we have exploited the property
of the sum of squared
Pn
numbers [37], where we have i=1 i2 = O(n3 ).
2) EQPO algorithm’s complexity: Moving on to the EQPO
algorithm’s complexity analysis, the P-NDQIO algorithm is
activated once per trellis stage, based on Alg. 2. Note that
we will classify the complexity imposed by the P-NDQIO
into two different domains, namely that of the parallel and
that of the sequential complexity. To elaborate further, the
P-NDQIO algorithm also exploits the synergies between QP
and HP, which was utilized by the NDQIO algorithm of [30].
Explicitly, the parallel complexity, which is termed as “normalized execution time” in [30], is defined as the number of
dominance comparisons, when taking into account the degree
of HP. Therefore, it may be deemed to be commensurate with
the algorithm’s actual normalized execution time. By contrast,
the sequential complexity, which is termed as “normalized
power consumption” in [30], is defined as the total number
of dominance comparisons, without considering the potential
degree of HP. Hence, this specific complexity may be deemed
to be commensurate with the algorithm’s normalized power
consumption, as elaborated in [30] as well.
Let us now proceed by characterizing the complexity of
the individual sub-processes of the P-NDQIO process. During
each trellis stage, the P-NDQIO algorithm activates its BWBBHT-QSA step. This step invokes the BBHT-QSA once;
however, since the quantum circuits of the original NDQIO
algorithm are utilized, each activation of the quantum oracle,
namely the operator UG in [30, Fig. 8], compares each of the
generated routes to all the routes comprising the OPF identified
so far. Since this set of comparisons is carried out in parallel,
OPF
a single activation imposes a single CFE and N(i)
CFEs in
the parallel and sequential domains, respectively. Note that the
surv
BW-BBHT-QSA process will be activated (N(i)
+ 2) times
during a single trellis stage, since we opted for repeating this
step for an additional iteration, when the BBHT-QSA fails
to identify a valid route. Therefore, the parallel and sequential complexity imposed by the BW-BBHT-QSA process are
quantified as follows:
routes
surv
),
+ 2)LBBHT (N(i)
= (N(i)
LBW,P
(i)
p
= O(NOPF NOPF Nnodes i),
(18)
(19)
chains, it has been demonstrated in [30] that the complexity
imposed by a single of BBHT-QSA chain - which leads to the
identification of a single Pareto-optimal route - is identical
to that of the so-called Durr-Hoyer Algorithm
(DHA) [26],
√
namely on the order of LDHA (N ) = O( N ) in terms of the
number of quantum oracle gate activations. As for the latter,
the Ug′ quantum operator of [30, Fig. 7] has been utilized,
which implements a dominance comparison. Explicitly, each
activation of this operator imposes a parallel complexity of
1/K CFEs and a sequential complexity of a single CFE,
owing to the parallel implementation of the UF comparisons.
Therefore, the parallel and sequential complexity imposed by
the BBHT-QSA chains are quantified as follows:
Lchain,P
(i)
Lchain,S
(i)
X
(20)
(21)
surv
Recall that the term N(i)
in Eqs. (18) and (20) corresponds to
the number of Pareto-optimal routes identified . Additionally,
for the calculation of the orders of complexity we have relied
on the fact that the BBHT-QSA
has a complexity on the order
√
of LBBHT (N ) = O( N ) as demonstrated both in [30] and in
[25]. Moving on to the complexity imposed by the BBHT-QSA
=
surv
routes
N(i)
LDHA (N(i)
),
(22)
(23)
(24)
(25)
N surv
LSR,P
(i)
(i)
1 X
OPF
2
(j + N(i−1)
) = O(NOPF
),
=
K j=1
(26)
surv
N(i)
LSR,S
=
(i)
X
2
OPF
) = O(NOPF
).
(j + N(i−1)
(27)
j=1
Recall from Eqs. (19), (21), (23), (25), (26) and (27) that
surv
we used Eqs. (13) and (14), where we have O(N(i)
) =
OPF
O(N(i) ) = O(NOPF ) with NOPF corresponding to the total
number of Pareto-optimal routes. Let us now proceed with
the evaluation of the total parallel and sequential complexities
of the EQPO algorithm. In the worst-case scenario the EQPO
algorithm will process (Nnodes −1) trellis stages, corresponding
to the maximum possible number of hops, whilst visiting each
node at most once. Thus, the total parallel and sequential
complexities imposed by the EQPO algorithm are quantified
as follows:
LP
EQP O =
=
j=0
surv
routes
+ N(i)
LBBHT (N(i)
),
p
2
= O(NOPF NOPF Nnodes i).
routes
LDHA (N(i)
),
K
p
= O(NOPF NOPF Nnodes i),
=
Finally, as for the OPF-SR dominance comparisons of Step 10
of Alg. 1, the parallel and sequential complexity imposed by
this process are quantified as follows:
LSEQP O =
OPF
routes
)+
(j + N(i−1)
) LBBHT (N(i)
surv
N(i)
p
= O(NOPF NOPF Nnodes i).
surv
N(i)
LBW,S
=
(i)
11
=
Nnodes
X−1
+ LSR,P
,
LBW,P
+ Lchain,P
(i)
(i)
(i)
i=1
3/2 2
),
O(NOPF Nnodes
Nnodes
−1
X
LBW,S
+
(i)
i=1
5/2 2
).
O(NOPF Nnodes
(28)
(29)
+ LSR,S
Lchain,S
(i)
(i) ,
(30)
(31)
Note that in Eqs. (29) and (31) we have exploited the
specific
of the sum of square roots, where we
√
Pnproperty
have
i
=
O(n3/2 ) [37]. Observe from Eqs. (17)
i=1
and (29) that the EQPO algorithm achieves a parallel complexity reduction against
√ the CDP method by a factor on
3
the order of O(Nnodes
NOPF ). Additionally, the respective
sequential complexity reduction is by a factor on the order
12
√
3
of O(Nnodes
/ NOPF ), based on Eqs. (17) and (31). Hence,
the EQPO imposes a lower sequential complexity √
than the
3
CDP method, as long as we have O(Nnodes
) > O( NOPF ).
As far as the EQPO algorithm’s predecessors are concerned,
it has been proven in [30] that the NDQO algorithm imposes
identical parallel and
√ sequential complexities, which are on
the order of O(N N ). By contrast, the NDQIO algorithm
imposes a parallel and √a sequential complexity,
which are
√
2
N ), respectively,
on the order of O(NOPF N ) and O(NOPF
where N corresponds to the total number of legitimate routes.
Consequently, the complexity imposed by both the√NDQO
and the NDQIO algorithms is proportional to O( N ) in
both domains, yielding an exponential increase in the order
of complexity as the number nodes increases. By contrast,
both the EQPO algorithm and the CDP method exhibit a
complexity order similar to polynomial scaling, since its has
been demonstrated in [30, Fig. 11] that the total number NOPF
of Pareto-optimal routes increases at a significantly lower rate
than that of the total number N of routes.
Let us now proceed by presenting the average parallel
and the average sequential complexity imposed both by the
EQPO algorithm and by the CDP method, which are shown
in Figs. 4a and 4b, respectively. We will compare the complexities imposed by the aforementioned algorithms to those
of the Brute-Force (BF) method as well as to those of the
EQPO algorithm’s predecessors, namely the NDQO and the
NDQIO algorithms. The aforementioned methods consider the
entire set of legitimate routes, hence they have no database
correlation exploitation capabilities. Additionally, the NDQO
algorithm and the BF method do not employ any HP scheme,
thus their respective parallel and sequential complexities are
identical. As far as the average complexity of the CDP method
is concerned, observe in Figs. 4a and 4b that it requires a
higher number of CFEs than the BF method for WMHNs
having less than 8 nodes. This parallel complexity overhead
is justified by the fact that the number NOPF of Paretooptimal routes w.r.t. the total number N of legitimate routes
is relatively high. This in turn yields an increase in the
fraction of trellis nodes that are classified as survivors, hence
leading to more dominance comparisons. However, this trend
is reversed for WMHNs having more than 7 nodes, where
the CDP method exhibits a complexity reduction compared
to the BF method. More specifically, for WMHNs constituted
by 9 nodes, this complexity reduction is close to an order of
magnitude. Still referring to 9-node WMHNs, the CDP method
imposes a slightly higher parallel complexity than that of the
NDQO algorithm, while it matches the sequential complexity
of the NDQIO algorithm for the same 9-node network, based
on Figs. 4a and 4b, respectively.
Moving on to the average parallel complexity of the EQPO
algorithm, observe in Fig. 4a that the EQPO algorithm imposes
fewer CFEs than the rest of the algorithms considered for
WHMNs having more than 5 nodes. Explicitly, this complexity
reduction becomes more substantial, as the number of nodes
increases, reaching a parallel complexity reduction of almost
an order of magnitude for 9-node WMHNs, when compared
to the NDQIO algorithm, which is capable of exploiting the
HP as well. As for its sequential complexity, observe in
IEEE TRANSACTIONS ON COMMUNICATIONS 2018
Fig. 4b that the EQPO algorithm imposes more CFEs than
the rest of the algorithms for WMHNs having less than 7
nodes. This may be justified by the relatively small number
of surviving routes, which does not allow the QP to excel
by providing beneficial complexity reduction. However, this
trend is reversed for WMHNs having more than 6 nodes,
where the number of surviving routes becomes higher. More
specifically for 9-node WMHNs, the EQPO algorithm begins to impose a sequential complexity reduction w.r.t. all
the remaining algorithms considered. Additionally, observe
in Figs. 4a and 4b that the EQPO algorithm’s complexity
increases with a much lower gradient, as the number of nodes
increases, when compared to the full-search-based algorithms,
namely to the BF method as well as to the NDQO and the
NDQIO algorithms. Explicitly, this is justified by the “almost
polynomial” order of complexity, as demonstrated in Eqs. (29)
and (31).
B. Accuracy
Having elaborated on the complexity imposed by the EQPO
let us now proceed by discussing its heuristic accuracy. Since
our design target is to identify the entire set of Pareto-optimal
routes, we will evaluate the EQPO algorithm’s accuracy versus
the complexity imposed in terms of two metrics, namely that
of the average Pareto distance E[Pd ] and that of the average
Pareto complection E[C]. The same set of metrics have been
considered in [30] for the evaluation of NDQIO algorithm’s
accuracy as well. To elaborate further, the Pareto distance of
a particular route is defined as the probability of this specific
route being dominated by the rest of the legitimate routes.
Explicitly, given a set of Pareto-optimal routes identified
by the EQPO algorithm, their average Pareto distance is a
characteristic of the OPF, since it provides insights into the
proximity of the exported OPF to the true OPF. Naturally, a
Pareto distance having a value of E[Pd ] = 0 implies that the
OPF identified by the EQPO is fully constituted by true Paretooptimal routes. By contrast, the average Pareto completion is
defined as the specific fraction of the solutions on the true OPF
identified by the EQPO. Therefore, our goal is to achieve a
Pareto completion as close to E[C] = 1 as possible.
Having defined the performance metrics, let us now present
the performance versus complexity results of the EQPO algorithm, which are shown in Fig. 5 for 7-node WMHNs.
The reason we have evaluated the aforementioned metrics
for 7-node WMHNs is for the sake of comparison to the
methods analyzed in [19] as well as in [30]. Apart from the
NDQO and NDQIO algorithms, we have used as benchmarks
two additional classical evolutionary algorithms5, namely the
NSGA-II and the MO-ACO. Using the same convention as in
[19] and [30], we have set the number of individuals equal
to the number of generations and we have matched the total
parallel complexity imposed by these classical algorithms to
that of the NDQO algorithm, since the NDQO algorithm
appears to impose the highest parallel complexity, based on
Fig. 4a. As for their total sequential complexity we have set
5 The readers should refer to [19] and to [18] for a detailed description of
the MO-ACO and the NSGA-II, respectively.
D. ALANIS ET AL.: A QUANTUM-AIDED DYNAMIC PROGRAMMING FRAMEWORK FOR PARETO OPTIMAL ROUTING IN WMHNS
107
107
NDQIO Alg.
NDQO Alg.
BF Method
CDP Method
EQPO Alg.
105
106
Total Number of CFEs
106
Total Number of CFEs
13
104
103
102
10
105
104
103
NDQIO Alg.
NDQO Alg.
BF Method
CDP Method
EQPO Alg.
102
10
1
1
4
5
6
7
8
9
Number of Nodes, Nnodes
(a) Parallel Complexity
4
5
6
7
8
9
Number of Nodes, Nnodes
(b) Sequential Complexity
Figure 4: EQPO Alg. (a) parallel and (b) sequential complexity quantified in terms of the number of CFEs. The results have
been averaged over 108 runs.
it to that of the NDQIO algorithm. Consequently, we have
considered employing 19 individuals over 19 generations for
the parallel complexity matching and 29 individuals over 29
generations for the sequential complexity matching for both
the NSGA-II and the MO-ACO algorithm.
Let us now proceed by elaborating on the average Pareto
distance exhibited by 7-node WMHNs versus the parallel
complexity invested, as portrayed in Fig. 5a. Observe in this
figure that the EQPO algorithm performs optimally – in the
sense that no suboptimal routes are included in the OPF –
for about 130 CFEs and then exhibits an error floor around
6 · 10−6 . Similar trends are observed for the classical NSGAII and for MO-ACO algorithm as well as for the quantumassisted NDQO algorithm; the classical benchmark algorithms
both exhibit an error floor around 10−3 , while the respective
NDQO algorithm’s error floor is around 7 · 10−9. By contrast,
the NDQIO algorithm initially has an error floor of about
3 · 10−5 , which then decays to infinitesimally low levels,
when more CFEs are invested owing to its OPF-SR process
[30]. This specific trend is visible in Fig. 5a, where the
NDQIO algorithm outperforms the NDQO technique in terms
of their E[Pd ] beyond 8842 CFEs in the sequential complexity
domain. Additionally, the NDQIO algorithm begins to exhibit
a lower E[Pd ] than that of the EQPO algorithm after 498 and
2932 CFEs in the parallel and sequential domains, respectively.
Let us now provide some further insights into the significance of the aforementioned error floors. Explicitly, a
particular route is considered suboptimal, if there exists even
just a single route dominating it, i.e. if it has a Pareto distance
higher than or equal to Pdth = 1/N , where N corresponds
to the total number of legitimate routes. This threshold is
visually portrayed with the aid of the dashed and dotted
horizontal lines in Figs. 5a and 5b. Hence, we can normalize
the results w.r.t. this threshold for exporting the probability of
a specific route becoming suboptimal. Consequently, EQPO
algorithm’s error floor is translated into a probability of a
specific route being suboptimal, which is equal to 0.2%, while
the respective probability of the NDQO algorithm is equal
to 2 · 10−6 . Additionally, the respective probabilities of the
classical benchmark algorithms are about 33% and 3.3%, when
parallel and sequential complexity are considered, respectively.
Consequently, the EQPO algorithm’s probability of opting for
a suboptimal route may be regarded as negligible.
The evaluation of the average Pareto completion probability
versus the parallel and the sequential complexity are shown in
Figs. 5c and 5d. Note that the subplots inside these figures
portray the portion of unidentified true Pareto-optimal routes,
as encapsulated by the expression of 1 − E[C]. Explicitly,
we will utilize this metric for assessing the error floor w.r.t.
the E[C], which may not be visible from the main plots.
Additionally, note that we examined both E[Pd ] and E[C]
versus the parallel and sequential complexity imposed up to
the maximum value observed by the EQPO algorithm. As far
as the EQPO algorithm’s average Pareto completion versus the
parallel complexity is concerned, observe in Fig. 5c that the
EQPO is capable of identifying a higher portion of the true
OPF, when compared to the rest of the algorithms examined,
while considering the same number of CFEs in the parallel
complexity domain. Explicitly, the EQPO algorithm succeeds
in identifying almost the entire set of Parero-optimal routes,
since it is only incapable of identifying as few as 0.1% of
the entire true OPF. This error floor is reached after 1301 and
14651 CFEs in the parallel and sequential complexity domains,
respectively, as it can be verified by Figs. 5c and 5d.
By contrast, this trend is not echoed in the sequential
complexity domain. To elaborate further, observe in Fig. 5b
that the EQPO algorithm remains more efficient than its
classical counterparts. On the other hand, while it is indeed
more efficient than the NDQO algorithm up to a complexity
budget of 2147 sequential CFEs, it identifies less Paretooptimal routes than the NDQO algorithm. The same trend
is observed for the NDQIO algorithm as well for a com-
14
IEEE TRANSACTIONS ON COMMUNICATIONS 2018
10−1
1/N
10−3
10−3
10−5
10−5
10−7
EQPO
NDQIO
NDQO
NSGA-II
MO-ACO
10−9
10−11
10−13
0
260
520
780
1040
E[Pd ]
E[Pd ]
10−1
1/N
10−7
EQPO
NDQIO
NDQO
NSGA-II
MO-ACO
10−9
10−11
10−13
1300
0
Parallel Complexity in CFEs
2930
5860
8790
11720
14650
Sequential Complexity in CFEs
(a)
(b)
1
1
5
5
EQPO
NDQIO
NDQO
NSGA-II
MO-ACO
5
10−2
260
10
10−1
EQPO
NDQIO
NDQO
NSGA-II
MO-ACO
5
10−4
2
0
2
520
0
325
780
650
975
1040
1300
1300
Parallel Complexity in CFEs
(c)
1 − E[C]
10−1
E[C]
2
1 − E[C]
E[C]
1
1
−1
10−2
10−3
10−4
0
3362 6724 10086 13448
2
0
2930
5860
8790
11720
14650
Sequential Complexity in CFEs
(d)
Figure 5: EQPO algorithm performance in terms of its Pareto dinstance (a,b) and its Pareto Completion (c,d) versus the parallel
complexity (a,c) and the sequential complexity (b,d) for required 7-node WHMNs. The results have been averaged over 108
runs.
plexity budget of 4794 sequential CFEs. Nevertheless, this
discrepancy between the parallel and sequential complexity is
expected to be decreased, as the number of nodes increases.
This is justified by the fact that the EQPO algorithm imposes
a lower sequential complexity as the nodes proliferate, as seen
in Fig. 4b.
Last but not least, the results portrayed on Fig. 5 rely on
the intelligent central node having perfect knowledge both of
the nodes’ geo-locations and of the interference power levels
experienced by them. This fundamental assumption, albeit
impractical, provides us with the upper bound of the achievable
performance of the routing schemes considered. Explicitly,
despite its impractical nature, it facilitates a fair comparison
of the EQPO algorithm to its predecessors in terms of their
complexity and heuristic accuracy, which is the main focus
of this treatise. Intuitively, a practical network information
update process would result in both approximated and outdated
network information, thus degrading the results of Fig. 5, while
maintaining the complexity per routing routing optimization
at a similar order. Note that we plan on characterizing these
imperfections and conceive a practical network information
update scheme in our future work.
VI. C ONCLUSIONS
In this treatise we have exploited the correlations in the
formation of the Pareto-optimal routes for the sake of achieving a routing complexity reduction. In this context, we have
first developed an optimal dynamic programming framework,
which transforms the multi-objective routing problem into
a decoding problem. However, this optimal framework imposes a high complexity. For this reason, we relaxed the
aforementioned framework and proposed the EQPO algorithm,
which is empowered by the P-NDQIO algorithm and thus
jointly exploits the synergies between the QP and the HP
along with the potential correlation in the formation of the
Pareto-optimal routes. We then analytically characterized the
complexity imposed by the EQPO algorithm showed that it
is capable of solving the multi-objective routing problem in
near-polynomial time. In fact, the EQPO achieved a parallel
complexity reduction of almost an order of magnitude and
a sequential complexity reduction by a factor of 3 for 9node WMHNs. Finally, we demonstrated with the aid of
simulations that this complexity reduction only imposes an
almost negligible error, which was found to be 0.2% and 0.1%
in terms of the average Pareto distance and the average Pareto
completion probability for 7-node WMHNs.
D. ALANIS ET AL.: A QUANTUM-AIDED DYNAMIC PROGRAMMING FRAMEWORK FOR PARETO OPTIMAL ROUTING IN WMHNS
A PPENDIX A
P ROOF OF P ROPOSITION 1
(j)
Proof. Let
us
consider
the
route
xg
=
{SN →R̄i →Rj →DN } generated by the route x. Based
on Eq. (9), the UFs associated with this specific route are
equal to:
fk (x(j)
g ) = fk (SN →R̄i ) + fk (R̄i →Rj ) + fk (Rj →DN ).
(32)
Additionally, the sub-route x′ is associated with the following
UFs
fk (x′ ) = fk (SN →R̄i ).
(33)
Since we have fk (x) > 0, ∀x from Eq. (9), the sub-route
(j)
x′ strongly dominates the route xg based on Eqs. (32) and
(j)
′
(33), i.e. we have f (x ) ≻ f (xg ). Since now there is a specific
route xd from the SN to DN that weakly dominates the subroute x′ , i.e. we have f (xd ) f (x′ ), the route xd strongly
(j)
dominates the route xg as well, yielding:
f (xd ) f (x′ ) ≻ f (x(j)
g ),
f (xd ) ≻
(34)
f (x(j)
g )
(35)
(j)
Consequently, based on Eq. (35) the route xg cannot be
Pareto-optimal, since it is strongly dominated by the route
xd .
R EFERENCES
[1] Z. Jiao, B. Zhang, C. Li, and H. T. Mouftah, “Backpressure-based
routing and scheduling protocols for wireless multihop networks: A
survey,” IEEE Wireless Communications, vol. 23, pp. 102–110, February
2016.
[2] B. Alawieh, Y. Zhang, C. Assi, and H. Mouftah, “Improving spatial
reuse in multihop wireless networks - a survey,” IEEE Communications
Surveys Tutorials, vol. 11, pp. 71–91, Aug 2009.
[3] Y. Chen, S. Zhang, S. Xu, and G. Y. Li, “Fundamental trade-offs on
green wireless networks,” IEEE Communications Magazine, vol. 49,
pp. 30–37, June 2011.
[4] H. Yetgin, K. T. K. Cheung, M. El-Hajjar, and L. Hanzo, “Networklifetime maximization of wireless sensor networks,” IEEE Access, vol. 3,
pp. 2191–2226, 2015.
[5] A. Abdulla, H. Nishiyama, J. Yang, N. Ansari, and N. Kato, “Hymn:
A novel hybrid multi-hop routing algorithm to improve the longevity of
wsns,” IEEE Transactions on Wireless Communications, vol. 11, no. 7,
pp. 2531–2541, 2012.
[6] L. Tan, Z. Zhu, F. Ge, and N. Xiong, “Utility maximization resource
allocation in wireless networks: Methods and algorithms,” IEEE Transactions on Systems, Man, and Cybernetics: Systems, vol. 45, pp. 1018–
1034, July 2015.
[7] Y. Shi, Y. T. Hou, and H. Sherali, “Cross-layer optimization for data rate
utility problem in UWB-based ad hoc networks,” IEEE Transactions on
Mobile Computing, vol. 7, pp. 764–777, June 2008.
[8] M. Caleffi, I. F. Akyildiz, and L. Paura, “Opera: Optimal routing metric
for cognitive radio ad hoc networks,” IEEE Transactions on Wireless
Communications, vol. 11, pp. 2884–2894, August 2012.
[9] R. Banirazi, E. Jonckheere, and B. Krishnamachari, “Heat-Diffusion:
Pareto optimal dynamic routing for time-varying wireless networks,” in
IEEE INFOCOM 2014 - IEEE Conference on Computer Communications, pp. 325–333, April 2014.
[10] S. Dasgupta, C. H. Papadimitriou, and U. Vazirani, Algorithms.
McGraw-Hill, Inc., 2006.
[11] D. Ramirez and B. Aazhang, “Optimal routing and power allocation for
wireless networks with imperfect full-duplex nodes,” IEEE Transactions
on Wireless Communications, vol. 12, pp. 4692–4704, September 2013.
[12] J. Zuo, C. Dong, H. V. Nguyen, S. X. Ng, L. L. Yang, and L. Hanzo,
“Cross-layer aided energy-efficient opportunistic routing in ad hoc
networks,” IEEE Transactions on Communications, vol. 62, pp. 522–
535, February 2014.
15
[13] C. Luo, S. Guo, S. Guo, L. T. Yang, G. Min, and X. Xie, “Green
communication in energy renewable wireless mesh networks: Routing,
rate control, and power allocation,” IEEE Transactions on Parallel and
Distributed Systems, vol. 25, pp. 3211–3220, Dec 2014.
[14] Q. You, Y. Li, M. S. Rahman, and Z. Chen, “A near optimal routing
scheme for multi-hop relay networks based on Viterbi algorithm,”
in 2012 IEEE International Conference on Communications (ICC),
pp. 4531–4536, June 2012.
[15] Y. Wang, M. Z. Bocus, and J. P. Coon, “Dynamic programming for route
selection in multihop fixed gain amplify-and-forward relay networks,”
IEEE Communications Letters, vol. 17, pp. 932–935, May 2013.
[16] J. Yao, S. Feng, X. Zhou, and Y. Liu, “Secure routing in multihop
wireless ad-hoc networks with decode-and-forward relaying,” IEEE
Transactions on Communications, vol. 64, pp. 753–764, Feb 2016.
[17] K. Deb, “Multi-objective optimization,” in Search Methodologies: Introductory Tutorials in Optimization and Decision Support Techniques
(E. K. Burke and G. Kendall, eds.), pp. 403–449, Boston, MA: Springer
US, 2014.
[18] H. Yetgin, K. Cheung, and L. Hanzo, “Multi-objective routing optimization using evolutionary algorithms,” in IEEE Wireless Communications
and Networking Conference (WCNC), pp. 3030–3034, 2012.
[19] D. Alanis, P. Botsinis, S. X. Ng, and L. Hanzo, “Quantum-assisted
routing optimization for self-organizing networks,” IEEE Access, vol. 2,
pp. 614–632, 2014.
[20] B. Gurakan, O. Ozel, and S. Ulukus, “Optimal energy and data routing
in networks with energy cooperation,” IEEE Transactions on Wireless
Communications, vol. 15, pp. 857–870, Feb 2016.
[21] M. Camelo, C. Omaa, and H. Castro, “QoS routing algorithm based on
multi-objective optimization for wireless mesh networks,” in 2010 IEEE
Latin-American Conference on Communications (LATINCOM), pp. 1–6,
2010.
[22] F. Martins, E. Carrano, E. Wanner, R. H. C. Takahashi, and G. Mateus,
“A hybrid multiobjective evolutionary approach for improving the performance of wireless sensor networks,” IEEE Sensors Journal, vol. 11,
no. 3, pp. 545–554, 2011.
[23] M. Lopez-Ibanez and T. Stutzle, “The automatic design of multiobjective
ant colony optimization algorithms,” IEEE Transactions on Evolutionary
Computation, vol. 16, pp. 861–875, Dec 2012.
[24] L. K. Grover, “A fast quantum mechanical algorithm for database
search,” in Proceedings of the twenty-eighth annual ACM symposium
on Theory of computing, pp. 212–219, ACM, 1996.
[25] M. Boyer, G. Brassard, P. Hyer, and A. Tapp, “Tight bounds on quantum
searching,” Fortschritte der Physik, vol. 46, no. 4-5, pp. 493–505, 1998.
[26] C. Durr and P. Høyer, “A quantum algorithm for finding the minimum,”
arXiv preprint quant-ph/9607014, 1996.
[27] M. A. Nielsen and I. L. Chuang, Quantum computation and quantum
information. Cambridge university press, 2010.
[28] C. Wang, H. Chen, and E. Jonckheere, “Quantum versus simulated
annealing in wireless interference network optimization,” Scientific Reports, vol. 6, no. 25797, 2016.
[29] C. Wang, E. Jonckheere, and T. Brun, “Differential geometric treewidth
estimation in adiabatic quantum computation,” Quantum Information
Processing, vol. 15, pp. 3951–3966, Oct 2016.
[30] D. Alanis, P. Botsinis, Z. Babar, S. X. Ng, and L. Hanzo, “Nondominated quantum iterative routing optimization for wireless multihop
networks,” IEEE Access, vol. 3, pp. 1704–1728, 2015.
[31] C. Zalka, “Grover’s quantum searching algorithm is optimal,” Physical
Review A, vol. 60, no. 4, p. 2746, 1999.
[32] D. Alanis, J. Hu, P. Botsinis, Z. Babar, S. X. Ng, and L. Hanzo,
“Quantum-assisted joint multi-objective routing and load balancing for
socially-aware networks,” IEEE Access, vol. 4, pp. 9993–10028, 2016.
[33] D. H. Lehmer, “Teaching combinatorial tricks to a computer,” in Proc.
Sympos. Appl. Math. Combinatorial Analysis, vol. 10, pp. 179–193,
1960.
[34] R. Steele and L. Hanzo, Mobile radio communications: second and third
generation cellular and WATM systems. Wiley - IEEE, J. Wiley, 1999.
[35] L. Hanzo, S. X. Ng, W. Webb, and T. Keller, Quadrature amplitude
modulation: From basics to adaptive trellis-coded, turbo-equalised and
space-time coded OFDM, CDMA and MC-CDMA systems. IEEE PressJohn Wiley, 2004.
[36] S. Yang, X. Xu, D. Alanis, S. X. Ng, and L. Hanzo, “Is the lowcomplexity mobile relay aided FFR-DAS capable of outperforming the
high-complexity CoMP?,” IEEE Transactions on Vehicular Technology,
vol. 65, pp. 2154–2169, April 2016.
[37] M. Abramowitz and I. A. Stegun, Handbook of mathematical functions:
with formulas, graphs, and mathematical tables, vol. 55. Courier
Corporation, 1964.
16
[38] W. Zhang, M. F. Brejza, T. Wang, R. G. Maunder, and L. Hanzo, “Irregular trellis for the near-capacity unary error correction coding of symbol
values from an infinite set,” IEEE Transactions on Communications,
vol. 63, pp. 5073–5088, Dec 2015.
[39] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, “A fast and elitist
multiobjective genetic algorithm: NSGA-II,” IEEE Transactions on
Evolutionary Computation, vol. 6, no. 2, pp. 182–197, 2002.
IEEE TRANSACTIONS ON COMMUNICATIONS 2018
| 8cs.DS |
1
arXiv:1708.03183v1 [cs.CE] 10 Aug 2017
Automated Tiling of Unstructured Mesh Computations with
Application to Seismological Modelling
FABIO LUPORINI, Imperial College London
MICHAEL LANGE, Imperial College London
CHRISTIAN T. JACOBS, University of Southampton
GERARD J. GORMAN, Imperial College London
J. RAMANUJAM, Louisiana State University
PAUL H. J. KELLY, Imperial College London
Sparse tiling is a technique to fuse loops that access common data, thus increasing data locality. Unlike
traditional loop fusion or blocking, the loops may have different iteration spaces and access shared datasets
through indirect memory accesses, such as A[map[i]] – hence the name “sparse”. One notable example of
such loops arises in discontinuous-Galerkin finite element methods, because of the computation of numerical
integrals over different domains (e.g., cells, facets). The major challenge with sparse tiling is implementation –
not only is it cumbersome to understand and synthesize, but it is also onerous to maintain and generalize, as it
requires a complete rewrite of the bulk of the numerical computation. In this article, we propose an approach
to extend the applicability of sparse tiling based on raising the level of abstraction. Through a sequence
of compiler passes, the mathematical specification of a problem is progressively lowered, and eventually
sparse-tiled C for-loops are generated. Besides automation, we advance the state-of-the-art by introducing:
a revisited, more efficient sparse tiling algorithm; support for distributed-memory parallelism; a range of
fine-grained optimizations for increased run-time performance; implementation in a publicly-available library,
SLOPE; and an in-depth study of the performance impact in Seigen, a real-world elastic wave equation solver
for seismological problems, which shows speed-ups up to 1.28× on a platform consisting of 896 Intel Broadwell
cores.
CCS Concepts: • Computing methodologies → Parallel algorithms; • Mathematics of computing →
Mathematical software performance; • Software and its engineering → Compilers; Domain specific
languages;
Additional Key Words and Phrases: Finite element method, unstructured mesh, compiler, performance optimization, loop fusion, loop tiling, sparse tiling
ACM Reference format:
Fabio Luporini, Michael Lange, Christian T. Jacobs, Gerard J. Gorman, J. Ramanujam, and Paul H. J. Kelly. 2017.
Automated Tiling of Unstructured Mesh Computations with Application to Seismological Modelling. 1, 1,
Article 1 (August 2017), 29 pages.
DOI: 10.1145/nnnnnnn.nnnnnnn
This work was supported by the Engineering and Physical Sciences Research Council through grants EP/I00677X/1,
EP/L000407/1, EP/I012036/1], by the Imperial College London Department of Computing, and by the Imperial College
London Intel Parallel Computing Centre (IPCC). The authors would like to thank the HPC Service Support team at Imperial
College London for their help with the Helen cluster. The authors would also like to thank Gheorghe-Teodor Bercea,
Lawrence Mitchell, and David Ham for their suggestions during the development of this project.
Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee
provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the
full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored.
Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires
prior specific permission and/or a fee. Request permissions from permissions@acm.org.
© 2017 Copyright held by the owner/author(s). Publication rights licensed to ACM. XXXX-XXXX/2017/8-ART1 $15.00
DOI: 10.1145/nnnnnnn.nnnnnnn
, Vol. 1, No. 1, Article 1. Publication date: August 2017.
1:2
1
F. Luporini et al.
INTRODUCTION
In many unstructured mesh applications, for example those approximating the solution of partial
differential equations (PDEs) using the finite volume or the finite element method, sequences of
numerical operators accessing common fields need to be evaluated. Usually, these operators are
implemented by iterating over sets of mesh elements and computing a kernel in each element.
In languages such as C or Fortran, the resulting sequence of loops is typically characterized
by heterogeneous iteration spaces and accesses to shared datasets (reads, writes, increments)
through indirect pointers, like A[map[i]]. One notable example of such operators/loops arises
in discontinuous-Galerkin finite element methods, in which numerical integrals over different
domains (e.g., cells, facets) are evaluated; here, A could represent a discrete function, whereas map
could store connectivity information (e.g., from mesh elements to degrees of freedom). In this
article, we devise compiler theory and technology to automate a sophisticated version of sparse
tiling, a technique to maximize data locality when accessing shared fields (like the A and map arrays
in the earlier example), which consists of fusing a sequence of loops by grouping iterations such
that all data dependencies are honored. The goal is to improve the overall application performance
with minimal disruption (none, if possible) to the source code.
Three motivating real-world applications for this work are Hydra, Volna and Seigen. Hydra [30]
is a finite-volume computational fluid dynamics application used at Rolls Royce for the simulation
of next-generation components of jet engines. Volna [11] is a finite-volume computational fluid
dynamics application for the modelling of tsunami waves. Seigen [17] aims to solve the elastic wave
equation using the discontinuous Galerkin finite element method for seismic exploration purposes.
All these applications are characterized by the presence of a time-stepping loop, in which several
loops over the mesh (thirty-three in Hydra, ten in Volna, twenty-five in Seigen) are repeatedly
executed. These loops are characterized by the irregular dependence structure mentioned earlier,
with for example indirect increments in one loop (e.g., A[m[i]] += f(...)) followed by indirect
reads in one of the subsequent loops (e.g., b = g(A[n[j]])). The performance achievable by Seigen
through sparse tiling will extensively be studied in Section 7.
Although our work is general in nature, we are particularly interested in supporting increasingly
sophisticated seismological problems that will be developed on top of Seigen. This has led to the
following strategic decisions:
Automation, but no interest in legacy codes Sparse tiling is an “extreme optimization”.
An implementation in a low level language requires a great deal of effort, as a thoughtful restructuring of the application is necessary. In common with many other low level
transformations, it also makes the source code impenetrable, affecting maintenance and
extensibility. We therefore aim for a fully automated system based on domain-specific
languages (DSLs), which abstracts sparse tiling through a simple interface (i.e., a single
construct to define a scope of fusible loops) and a tiny set of parameters for performance
tuning (e.g., the tile size). We are not interested in automating sparse tiling in legacy codes,
in which the key computational aspects (e.g., mesh iteration, distributed-memory parallelism) are usually hidden for software modularity, thus making such a transformation
almost impossible.
Unstructured meshes require mixed static/dynamic analysis Unstructured meshes are
often used to discretize the computational domain, since they allow for an accurate representation of complex geometries. Their connectivity is stored by means of adjacency lists
(or equivalent data structure), which leads to indirect memory accesses within the loop
, Vol. 1, No. 1, Article 1. Publication date: August 2017.
Automated Tiling of Unstructured Mesh Computations
1:3
nests. Indirections break static analysis, thus making purely compiler-based approaches
insufficient. Runtime data dependence analysis is essential for sparse tiling, so integration
of compiler and run-time tracking algorithms becomes necessary.
Realistic datasets not fitting in a single node Real-world simulations often operate on
terabytes of data, hence execution on multi-node systems is often required. We have
extended the original sparse tiling algorithm to enable distributed-memory parallelism.
Sparse tiling does not change the semantics of a numerical method – only the order in which
some iterations are executed. Therefore, if most sections of a PDE solver suffer from computational
boundedness and standard optimizations such as vectorization have already been applied, then
sparse tiling, which targets memory-boundedness, will only provide marginal benefits (if any).
Likewise, if a global reduction is present in between two loops, then there is no way for sparse
tiling to be applied, unless the numerical method itself is rethought. This is regardless of whether
the reduction is explicit (e.g., the first loop updates a global variable that is read by the second loop)
or implicit (i.e., within an external function, as occurs for example in most implicit finite element
solvers). These are probably the two greatest limitations of the technique; otherwise, sparse tiling
may provide substantial performance benefits.
The rest of the article is structured as follows: in Section 2 we present the abstraction on
which sparse tiling relies. We then show, in Section 3, examples of how the algorithm works on
shared- and distributed-memory systems. This is followed by the formalization of the algorithms
(Sections 4, 5) and the implementation of the compiler that automates sparse tiling (Section 6). The
experimentation is described in Section 7. A discussion on the limitations of the algorithms and the
future work that we expect to carry out in the years to come conclude the article.
2
THE LOOP CHAIN ABSTRACTION FOR UNSTRUCTURED MESH APPLICATIONS
The loop chain is an abstraction introduced in [21]. Informally, a loop chain is a sequence of loops
with no global synchronization points, with attached some extra information to enable run-time
data dependence analysis – necessary since indirect memory accesses inhibit common static approaches to loop optimization. The idea is to replace static with dynamic analysis, exploiting the
information carried by a loop chain. Loop chains must somehow be added to or automatically
derived (e.g., exploiting a DSL) from the input code. A loop chain will then be used by an inspector/executor scheme [32]. The inspector is an algorithm performing data dependence analysis using
the information carried by the loop chain, which eventually produces a sparse tiling schedule. This
schedule is used by the executor, a piece of code semantically equivalent to the original sequence of
loops (i.e., computing the same result) executing the various loop iterations in a different order.
Before diving into the description of the loop chain abstraction, it is worth observing two aspects.
• The inspection phase introduces an overhead. In many scientific computations, the data
dependence pattern is static – or, more informally, “the topology does not change over
time”. This means that the inspection cost may be amortized over multiple iterations of the
executor. If instead the mesh changes over time (e.g., in case of adaptive mesh refinement),
a new inspection must be performed.
• To adopt sparse tiling in a code there are two options. One possibility is to provide a library
and leave the application specialists with the burden of carrying out the implementation
(re-implementation in case of legacy code). A more promising alternative consists of raising
the level of abstraction: programs can be written using a DSL; loop chain, inspector, and
executor can then be automatically derived at the level of the intermediate representation.
As we shall see in Section 6, the tools developed in this article enable both approaches,
though our primary interest is in the automated approach (i.e., via DSLs).
, Vol. 1, No. 1, Article 1. Publication date: August 2017.
Process 0
core
non-exec
exec
owned
F. Luporini et al.
owned
exec
non-exec
1:4
Process 1
core
Fig. 1. Two partitions of a mesh distributed to two neighboring processes. The core region includes all
iterations that can be processed without reading halo data. The owned iterations can be processed only by
reading halo data. Exec halo is the set of iterations that are redundantly executed because they indirectly
access owned iterations. Finally, the non-exec halo region includes off-process iterations that only need to
be read to compute the exec halo iterations. Matching colors across the two processes represent identical
subsets of iterations in the non-partitioned mesh. The image was inspired by an example in [27].
These points will be further elaborated in later sections.
The loop chain abstraction was originally defined as follows:
• A loop chain L consists of n loops, L 0 , L 1 , ..., Ln−1 . There are no global synchronization
points in between the loops. Although there may be dependencies between successive
loops in the chain, the execution order of a loop’s iterations does not influence the result.
• D is a set of disjoint m data spaces, D 0 , D 1 , ..., Dm−1 . Each loop accesses (reads from, writes
to) a subset of these data spaces. An access can be either direct (e.g., A[i]) or indirect (e.g.,
A[map(i)]).
• R Ll →Dd (i) and WLl →Dd (i) are access relations for a loop Ll over a data space Dd ∈ D. They
indicate which locations in the data space Dd an iteration i ∈ Ll reads from and writes to,
respectively. A loop chain must provide all access relations for all loops. For example, if Ll
writes to the array A as A[B(i)] = f(...), then the loop chain will have to provide an
access relation B Ll →A (i).
We here refine this definition, and specialize it for unstructured mesh applications. This allows
the introduction of new concepts, which will be used to extend the sparse tiling algorithm presented
in [34]. Some terminology and ideas are inspired by the programming model of OP2, a library for
unstructured mesh applications [14] used to implement the already mentioned Hydra code.
• A loop chain L consists of n loops, L 0 , L 1 , ..., Ln−1 . There are no global synchronization
points in between the loops. Although there may be dependencies between successive
loops in the chain, the execution order of a loop’s iterations does not influence the result.
• S is a set of m disjoint sets, S 0 , S 1 , ..., Sm−1 . Sets are used to represent iteration and data
spaces. Possible sets are the topological entities in the mesh or the degrees of freedom
associated with a function.
A set S is logically split into three contiguous regions: core (S c ), boundary (S b ), and
non-exec (S ne ). Given a process P and a set S:
S c : the iterations of S that are owned by P.
S b : the boundary region can be seen as the union of two sub-regions, owned (S owned )
and exec (S ex ec ). As shown in Figure 1, S owned are iterations that belong to P which
, Vol. 1, No. 1, Article 1. Publication date: August 2017.
Automated Tiling of Unstructured Mesh Computations
1:5
are redundantly executed by some other processes; S ex ec are iterations from other processes which are redundantly executed by P. We will see that redundant computation
preserves atomic execution – a property that enables executing tiles without the need
for synchronization.
S ne : these are iterations of other processes that are communicated to P because they
need be read to correctly compute S b .
A set is uniquely identified by a name and the sizes of its three regions. For example, the
notation S = (vertices, C, B, N ) defines the vertices set, which comprises C elements
in the core region (iterations 0 to C − 1), B elements in the boundary region (iterations C to
C + B − 1), and N elements in the non-exec region (iterations C + B to C + B + N − 1).
• The depth is an integer indicating the extent of the boundary region of a set. This constant
is the same for all sets in S.
• M is a set of k maps, M 0 , M 1 , ..., Mk −1 . A map of arity a is a vector-valued function M :
Si → S j0 × S j1 × ... × S ja−1 that connects each element of Si to one or more elements in
S j . For example, if a triangular cell c is connected to three vertices v 0 , v 1 , v 2 , we have
M(c) = [v 0 , v 1 , v 2 ].
• A loop Li over the iteration space S is associated with d descriptors, D 0 , D 1 , ..., Dd −1 . A
descriptor D is a 2-tuple D = <M, mode>. M is either a map from S to some other sets
or the special placeholder ⊥, which indicates that memory accesses are direct to some
data associated with S itself. mode is one of [r , w, i], meaning that a memory access is
respectively of type read, write or increment.
There are a few crucial differences in this refined definition for the unstructured mesh case.
One of them is the presence of sets in place of data spaces. In unstructured mesh applications, a
loop tends to access multiple data spaces associated with the same set. The idea is to rely on sets,
rather than data spaces, to perform data dependence analysis. This can significantly improve the
inspection cost, because typically |S| << |D|. Another crucial difference is the characterization
of sets into the three regions core, boundary and non-exec. As we shall see, this separation is
essential to enable distributed-memory parallelism. The extent of the boundary regions is captured
by the depth of the loop chain. Informally, the depth tells how many extra “strips” of elements
are provided by the neighboring processes. This allows some redundant computation along the
partition boundary and also limits the depth of the loop chain (i.e., how many loops can be fused).
The role of the parameter depth will be clear by the end of Section 5.
3
LOOP CHAIN, INSPECTION AND EXECUTION EXAMPLES
Using the example in Figure 2a, we describe the actions performed by our sparse tiling inspector.
The inspector takes as input the loop chain illustrated in Figure 2b. We show two variants, for
shared-memory and distributed-memory parallelism. The value of the variable mode in Figure 2b
determines the variant to be executed.
3.1
Overview
The inspector starts with partitioning the iteration space of a seed loop, for example L 0 . Partitions
are used to initialize tiles: the iterations of L 0 falling in Pi – or, in other words, the edges in partition
Pi – are assigned to the tile Ti . Figure 3 displays the situation after the initial partitioning of L 0 for
a given input mesh. There are four partitions, two of which (P0 and P3 ) are not connected through
any edge or cell. These four partitions correspond to four tiles, [T0 , T1 , T2 , T3 ], with Pi = Ti .
As detailed in the next two sections, the inspection proceeds by populating Ti with iterations
from L 1 and L 2 . The challenge of this task is guaranteeing that all data dependencies – read after
, Vol. 1, No. 1, Article 1. Publication date: August 2017.
1:6
F. Luporini et al.
for t = 0 to T {
inspector = init_inspector (...) ;
// L 0 : loop over edges
for e = 0 to E {
x = X + e;
tmp_0 = tmp + edges2vertices [e + 0];
tmp_1 = tmp + edges2vertices [e + 1];
kernel1 (x , tmp_0, tmp_1) ;
}
// Three sets , edges , cells , and vertices
E = set ( inspector , " edges " , core_edges ,
boundary_edges , nonexec_edges , ...) ;
C = set ( inspector , " cells " , core_cells ,
boundary_cells , nonexec_cells , ...) ;
V = set ( inspector , " verts " , core_verts ,
boundary_verts , nonexec_verts , ...) ;
// L 1 : loop over cells
for c = 0 to C {
res = R + c;
tmp_0 = tmp + cells2vertices [c + 0];
tmp_1 = tmp + cells2vertices [c + 1];
tmp_2 = tmp + cells2vertices [c + 2];
kernel2 ( res , tmp_0 , tmp_1 , tmp_2 );
}
// Two maps , from edges to vertices and from cells
to vertices
e2vMap = map ( inspector , E , V , edges2verts , ...) ;
c2vMap = map ( inspector , C , V , cells2verts , ...) ;
// The loop chain comprises three loops
// Each loop has some descriptors
loop ( inspector , E , { ⊥ , "r"}, { e2vMap , "i" }) ;
loop ( inspector , C , { ⊥ , "r"}, { c2vMap , "i" }) ;
loop ( inspector , E , { e2vMap , "i" }) ;
// L 2 : loop over edges
for e = 0 to E {
tmp_0 = tmp + edges2vertices [e + 0];
tmp_1 = tmp + edges2vertices [e + 1];
kernel3 ( tmp_0, tmp_1) ;
}
// Now can run the inspector
return inspection ( mode , inspector , tile_size , ...)
;
}
(a) Example sequence of sparse-tilable loops.
(b) Loop chain for the example program.
Fig. 2. On the left, a “toy” program used as running example in Section 3 to illustrate the loop chain abstraction
and show how the sparse tiling algorithms (inspection, execution) work. Note that all parameters passed to the
kernels are pointers. On the right, a code snippet showing the loop chain corresponding to the program on the
left. The syntax is very close to the actual API of SLOPE, the sparse tiling library that we have implemented,
described in Section 6.
P1
P3
P0
P2
Fig. 3. Partitioning of the seed loop. The vertices are illustrated to make the connectivity of the mesh clear,
although they do not belong to any partition yet.
write, write after read, write after write – are honored. The output of the inspector is eventually
passed to the executor. The inspection carries sufficient information for computing sets of tiles in
parallel. Ti is always executed by a single thread/process and the execution is atomic; that is, it does
not require communication with other threads/processes. When executing Ti , first all iterations
from L 0 are executed, then all iterations from L 1 and finally those from L 2 .
, Vol. 1, No. 1, Article 1. Publication date: August 2017.
Automated Tiling of Unstructured Mesh Computations
1:7
C0
V1
V0
C1
C2
V2
(a) A snapshot of the mesh after tiling L 0 .
B
B
B
B
B
B
B
B
B
G
R
B
R
R
R
R
R
R
R
R
R
R
R
G
R
R
R
B
G
R
R
B
B
G
B
B
B
B
R
R
B
B
B
R
R
B
R
R
R
(b) The vertices are written by L 0 , so a projection must
be computed before tiling L 1 . Here, the projection is
represented as colored vertices.
R
R
R
R
(c) A snapshot of the mesh after tiling L 1 .
(d) A snapshot of the mesh after tiling L 2 .
Fig. 4. Four passes of the inspection algorithm for shared-memory parallelism.
3.2
Inspection for Shared-Memory Parallelism
Similarly to OP2, to achieve shared-memory parallelism we use coloring. Two tiles that are given
the same color can be executed in parallel by different threads. Two tiles can have the same color
if they are not connected, because this ensures the absence of race conditions through indirect
memory accesses during parallel execution. In the example we can use three colors: red (R), green
(G), and blue (B). T0 and T3 are not connected, so they are assigned the same color. The colored
tiles are shown in Figure 4a. In the following, with the notation Tic we indicate that the i-th tile has
color c.
To populate [T0G , T1B , T2R , T3G ] with iterations from L 1 and L 2 , we first have to establish a total
ordering for the execution of partitions with different colors. Here, we assume the following order:
green (G), blue (B), and red (R). This implies, for instance, that all iterations assigned to T1B must
be executed before all iterations assigned to T2R . By “all iterations” we mean the iterations from L 0
(determined by the seed partitioning) as well as the iterations that will later be assigned from tiling
L 1 and L 2 . We assign integer positive numbers to colors to reflect their ordering, where a smaller
number means higher execution priority. We can assign, for example, 0 to green, 1 to blue, and 2 to
red.
To schedule the iterations of L 1 to [T0G , T1B , T2R , T3G ], we need to compute a projection for any
write or local reduction performed by L 0 . The projection required by L 0 is a function ϕ : V → T
mapping the vertices in V – as indirectly incremented during the execution of L 0 , see Figure 2a –
to a tile Tic ∈ T. Consider the vertex v 0 in Figure 4b. v 0 has 7 incident edges, 2 of which belong to
T0G , while the remaining 5 to T1B . Since we established that G ≺ B, v 0 can only be read after T1B has
finished executing the iterations from L 0 (i.e., the 5 incident blue edges). We express this condition
by setting ϕ(v 0 ) = T1B . Observe that we can compute ϕ by iterating over V and, for each vertex,
applying the maximum function (MAX) to the color of the adjacent edges.
We now use ϕ to schedule L 1 , a loop over cells, to the tiles. Consider again v 0 and the adjacent cells [c 0 , c 1 , c 2 ] in Figure 4b. These three cells have in common the fact that they are
adjacent to both green and blue vertices. For c 1 , and similarly for the other cells, we compute
MAX(ϕ(v 0 ), ϕ(v 1 ), ϕ(v 2 )) = MAX(B, G, G) = B = 1. This establishes that c 1 must be assigned
, Vol. 1, No. 1, Article 1. Publication date: August 2017.
1:8
F. Luporini et al.
to T1B , because otherwise (c 1 assigned instead to T0G ) a read to v 0 would occur before the last
increment from T1B took place. Indeed, we recall that the execution order, for correctness, must
be “all iterations from [L 0 , L 1 , L 2 ] in the green tiles before all iterations from [L 0 , L 1 , L 2 ] in the blue
tiles”. The scheduling of L 1 to tiles is displayed in Figure 4c.
To schedule L 2 to [T0G , T1B , T2R , T3G ] we employ a similar process. Vertices are again written by
L 1 , so a new projection over V will be necessary. Figure 4d shows the output of this last phase.
B
B
B
B
B
B
B
B
G
G
G
G
G
G
(a) After tiling L 0
R
B
G
G
R
R
G
G
G
G
G
G
R
G
G
R
R
R
G
G
G
R
G
G
G
G
G
G
G
G
R
R
R
R
R
R
(b) After tiling L 1
G
R
R
(c) After tiling L 2
Fig. 5. Tiling the program in Figure 2a for shared-memory parallelism can lead to conflicts. Here, the two
green tiles eventually become adjacent, creating race conditions.
Conflicting Colors. It is worth noting how T2R “consumed” the frontier elements of all other tiles
every time a new loop was scheduled. Tiling a loop chain consisting of k loops has the effect of
expanding the frontier of a tile of at most k vertices. With this in mind, we re-inspect the loop
chain of the running example, although this time employing a different execution order – blue (B),
red (R), and green (G) – and a different seed partitioning. Figure 5 shows that, by applying the same
procedure described in this section, T0G and T3G will eventually become adjacent. This violates the
precondition that tiles can be given the same color, and thus run in parallel, as long as they are not
adjacent. Race conditions during the execution of iterations belonging to L 2 are now possible. This
problem will be solved in Section 5.
3.3
Inspection for Distributed-Memory Parallelism
In the case of distributed-memory parallelism, the mesh is partitioned and distributed to a set of
processes. Neighboring processes typically exchange (MPI) messages before executing a loop L j .
A message includes all “dirty” dataset values required by L j modified by any Lk , with Lk ≺ L j . In
the running example, L 0 writes to vertices, so a subset of values associated with border vertices
must be communicated prior to the execution of L 1 . To apply sparse tiling, the idea is to push all
communications to the beginning of the loop chain: as we shall see, this increases the amount
of data to be communicated at a time, but also reduces the number of synchronizations (only 1
synchronization between each pair of neighboring processes per loop chain execution).
From Section 2 it is known that, in a loop chain, a set is logically split into three regions, core,
boundary, and non-exec. The boundary tiles, which originate from the seed partitioning of the
boundary region, will include all iterations that cannot be executed until the communications have
terminated. The procedure described for shared-memory parallelism – now performed individually
by each process on a partition of the input mesh – is modified as follows:
(1) The core region of the seed loop L 0 is partitioned into tiles. Unless aiming for a mixed
distributed/shared-memory scheme, there is no need to assign identical colors to unconnected tiles, as a process will execute its own tiles sequentially. Colors are assigned
increasingly, with Ti given color i. As long as tiles with contiguous ID are also physically
, Vol. 1, No. 1, Article 1. Publication date: August 2017.
Automated Tiling of Unstructured Mesh Computations
1:9
contiguous in the mesh, this assignment retains memory access locality when “jumping”
from executing Ti to Ti+1 .
(2) The same process is applied to the boundary region. Thus, a situation in which a tile includes
iterations from both the core and the boundary regions is prevented by construction. Further,
all tiles within the boundary region are assigned colors higher than those used for the core
tiles. This constrains the execution order: no boundary tiles will be executed until all core
tiles are computed.
(3) We map the whole non-exec region of L 0 to a single special tile, Tne . This tile has the highest
color and will actually never be executed.
boundary
Process 0
core
owned
boundary
non-exec
exec
exec
non-exec
owned
core
Process 1
Fig. 6. A snapshot of the two mesh partitions on Process 0 and Process 1 after inspecting the seed loop
L 0 for distributed-memory parallelism. On each process, there are five tiles in total: two in the core region
(green and violet), two in the boundary region (red and light blue), and Tne . The boundary tiles can safely
cross the owned and exec sub-regions (i.e., the private local iterations and the iterations to be redundantly
computed, respectively). However, no tile can include iterations from both the core and the boundary regions.
boundary
Process 0
core
owned
non-exec
exec
non-exec
boundary
exec
owned
core
Process 1
Fig. 7. A snapshot of the two mesh partitions on Process 0 and Process 1 at the end of the inspection
for distributed-memory parallelism. Tne expands over the boundary region, which minimizes the amount of
redundant computation to be performed. At the end of the execution phase, the orange edges will contain
“dirty values”, but correctness is not affected as the exec region only includes off-process data. The boundary
tiles expand over the core region: this is essential for correctness since none of the red and blue entities from
[L 0 , L 1 , L 2 ] can be executed until the MPI communications have terminated.
From this point on, the inspection proceeds as in the case of shared-memory parallelism. The
application of the MAX function when scheduling L 1 and L 2 makes higher color tiles (i.e., those
having lower priority) “expand over” lower color ones.
, Vol. 1, No. 1, Article 1. Publication date: August 2017.
1:10
F. Luporini et al.
In Figure 6, a mesh is partitioned over two processes and a possible seed partitioning and tiling
of L 0 illustrated. We observe that the two boundary tiles (the red and light blue ones) will expand
over the core tiles as L 1 and L 2 are tiled, which eventually results in the scheduling illustrated in
Figure 7. Roughly speaking, if a loop chain consists of n loops and, on each process, n − 1 extra
layers of iterations are provided (the exec regions in Figure 6), then all boundary tiles are correctly
computed.
The schedule produced by the inspector is subsequently used by the executor. On each process, the
executor starts with triggering the MPI communications required for the computation of boundary
tiles. All core tiles are then computed, since no data from the boundary region is necessary. Hence,
computation is overlapped with communication. As all core tiles are computed and the MPI
communications terminated, the boundary tiles can finally be computed.
Efficiency considerations. The underlying hypothesis is that the increase in data locality will
outweigh the overhead induced by the redundant computation and by the bigger volume of data
exchanged. This is motivated by several facts: (i) the loops being memory-bound; (ii) the core region
being much larger than the boundary region; (iii) the amount of redundant computation being
minimized through the special tile Tne , which progressively expands over the boundary region,
thus avoiding unnecessary calculations.
4
DATA DEPENDENCY ANALYSIS
The loop chain abstraction, described in Section 2, provides the information to construct an
inspector capable of analyzing data dependencies and thus build a legal sparse tiling schedule.
The dependencies between two different loops may be of type flow (read after write), anti (write
after read), or output (write after write). Further, as elaborated below, there may be “reduction
dependencies” between iterations of the same loop. In the following, assume that loop L x , having
iteration space S x , always comes before loop Ly , having iteration space Sy , in the loop chain. Let us
identify a descriptor of a loop L with mmode
S i →S j : this simply indicates that the loop L i has iteration
space Si and uses a map m to write/read/increment elements (respectively, mode ∈ {w, r , i}) in the
space S j .
The flow dependencies can then be enumerated by considering pairs of points (i® and j®) in the
iteration spaces of the two loops L x and Ly :
r
®
®
{i® → j® | i® ∈ S x ∧ j® ∈ Sy ∧ mw
S →S (i ) ∩ m S →S (j ) , ∅}.
x
z
y
z
Anti and output dependencies are defined in a similar way. The anti dependencies for all pairs of
loops L x and Ly are:
®
{i® → j® | i® ∈ S x ∧ j® ∈ Sy ∧ mrS →S (i®) ∩ mw
S →S (j ) , ∅}.
x
z
y
z
While the output dependencies between loops L x and Ly are:
w
®
{i® → j® | i® ∈ S x ∧ j® ∈ Sy ∧ mw
S →S (i ) ∩ m S
x
z
y →S z
(j®) , ∅}.
In essence, there is a storage-related data dependence between two iterations from different loops
(and therefore between the tiles they are placed in) when one of those iterations writes to a data
element and the other iteration reads from or writes to the same data element.
There are local reductions, or “reduction dependencies” between two or more iterations of the
same loop when those iterations “increment” the same location(s); that is, when they read, modify
with a commutative and associative operator, and write to the same location(s). The reduction
dependencies in L x are:
{i® → j® | i® ∈ S x ∧ j® ∈ S x ∧ miS →S (i®) ∩ miS →S (j®) , ∅}.
x
, Vol. 1, No. 1, Article 1. Publication date: August 2017.
z
x
z
Automated Tiling of Unstructured Mesh Computations
1:11
A reduction dependency between two iterations within the same loop indicates that those two
iterations must be executed atomically with respect to each other.
As seen in the example in Section 3, our inspector algorithm handles data dependencies, including
those between non-adjacent loops, by tracking projections. In the next section we formalize how
projections are constructed and used.
5
ALGORITHMS
The pseudo-code for the sparse tiling inspector is shown in Algorithm 1. Given a loop chain and a
“seed” tile size, the algorithm produces a schedule suitable for mixed distributed/shared-memory
parallelism. This schedule – in essence, a set of populated tiles – is used by the executor to perform
the sparse-tiled computation. The executor pseudo-code is displayed in Algorithm 2. In the next
two sections, we elaborate on the main steps of these algorithms. The notation is summarized in
Table 1.
ALGORITHM 1: The inspection algorithm
ALGORITHM 2: The executor algorithm
Input: The loop chain L = [L 0, L 1, ..., L n−1 ], a tile size t s
Output: A set of tiles T, populated with iterations from L
Input: A set of tiles T
Result: Execute the loop chain
seed ← 0;
Φ ← ∅;
C ←⊥;
σs e ed , T ← partition(S s e ed , t s );
seed_map ← find_map(S s e ed , L);
conflicts ← false;
Tc , Tb ← group_tiles_by_region(T);
do
color(T, seed_map);
for j = 1 to n − 1 do
project(L j−1 , σ j−1 , Φ, C );
σ j ← tile(L j , Φ);
assign(σ j , T);
end
if has_conflicts(C) then
conflicts ← true;
add_fake_connection(seed_map, C);
end
while conflicts;
compute_local_maps(T);
return T
start_MPI_comm();
foreach color do
foreach T ∈ Tc s.t. T.color == color do
execute_tile(T);
end
end
end_MPI_comm();
foreach color do
foreach T ∈ Tb s.t. T.color == color do
execute_tile(T);
end
end
ALGORITHM 3: Projection of a tiled loop
ALGORITHM 4: Building a tiling function
Input: A loop L j , a tiling function σ j , the projections set Φ, the conflicts
matrix C
Result: Update Φ and C
Input: A loop L j , the projections set Φ
Output: The tiling function σ j
foreach D ∈ L j .descriptors do
if D.map == ⊥ then
Φ = Φ ∪ σi ;
else
inverse_map ← map_invert(D.map);
S t , S j , values, offset ← inverse_map;
ϕ S t ←⊥;
for e = 0 to S t .size do
for k = offset[e] to offset[e + 1] do
Tl as t = T[values[k ]];
max_color ← MAX(ϕ S t [e].color, Tl as t .color);
if max_color , ϕ S t [e].color then
ϕ S t [e] ← Tl as t ;
end
end
end
update(C, T, ϕ S t );
Φ = Φ ∪ ϕSt ;
end
end
foreach D ∈ L j .descriptors do
if D.map == ⊥ then
σ j ← Φ[S j ];
else
arity ← D.map.arity;
ϕ S ← Φ[D . map .S j ];
for e = 0 to S j .size do
σ j [e] ← T⊥ ;
for k = 0 to arity do
adjacent_tile ← ϕ S [D . map . values[e ∗ arity + k ]];
max_color ← MAX(σ j [e].color,
adjacent_tile.color);
if max_color , σ j [e].color then
σ j [e] ← adjacent_tile;
end
end
end
end
end
return σ j
σ j ←⊥;
, Vol. 1, No. 1, Article 1. Publication date: August 2017.
1:12
F. Luporini et al.
Symbol
L
Lj
Sj
S jc , S jb , S jne
S
D
r, w, i
T
T[i]
ϕS
Φ
σj
ts
Meaning
The loop chain
The j -th loop in L
The iteration space of L j
The core, boundary, and non-exec regions of S j
A generic set in L
A descriptor of a loop
Possible values for D .mode
The set of all tiles
Accessing the i -th tile
A projection ϕ S : S → T
The set of all available projections
A tiling function σ j : S j → T for L j
seed tile size
Table 1. Summary of the notation used throughout the section.
5.1
Inspector
Choice of the seed loop. The seed loop Lseed is used to initialize the tiles. Theoretically, any
loop in the chain can be chosen as seed. Supporting distributed-memory parallelism, however, is
cumbersome if Lseed , L 0 . This is because more general schemes for partitioning and coloring
would be needed to ensure that no iterations in any S bj are assigned to a core tile. A limitation of
our inspector algorithm in the case of distributed-memory parallelism is that it must be Lseed = L 0 .
In the special case in which there is no need to distinguish between core and boundary tiles
(because a program is executed on a single shared-memory system), Lseed can be chosen arbitrarily.
If we however pick Lseed in the middle of the loop chain (L 0 ≺ ... ≺ Lseed ≺ ...), a mechanism
for constructing tiles in the reverse direction (“backwards”), from Lseed towards L 0 , is necessary.
In [34], we propose two “symmetric” algorithms to solve this problem, forward tiling and backward
tiling, with the latter using the MIN function in place of MAX when computing projections. For
ease of exposition, and since in the fundamental case of distributed-memory parallelism we are
imposing Lseed = L 0 , we here neglect this distinction1 .
Tile initialization. Let ts be the user-specified average tile size. The algorithm starts with partitionc
ing Sseed
into m subsets {P0 , P1 , ..., Pm−1 } such that |Pi | = ts (except possibly for Pm−1 ), Pi ∩ P j = ∅,
c
c
and ∪m−1
i=0 Pi = S seed . Among all possible legal partitionings, we choose the one that splits S seed
into blocks of ts contiguous iterations, with P 0 = {0, ..., ts − 1}, P 1 = {ts, ..., 2ts − 1}, and so on. We
b
analogously partition Sseed
into k subsets. We create m + k + 1 tiles, one for each of these partitions
ne
c ,T b , ...,T b
and one extra tile for Sseed . We therefore have T = {T0c , ...,Tm−1
,T ne }.
m
m+k −1 m+k
A tile Ti has four fields, as summarized in Table 2.
• The region is used by the executor to schedule tiles in a given order. This field is set right
c
b
after the partitioning of Lseed , as a tile (by construction) exclusively belongs to Sseed
, Sseed
,
ne
or Sseed .
• The iteration lists contain the iterations in L thatTi will have to execute. There is one iteration
list [Ti ]j for each L j ∈ L. At this stage of the inspection we have [Ti ]seed = [Ti ]0 = Pi ,
whereas still [Ti ]j = ∅ for j = 1, ..., n.
• Local maps may be used for performance optimization by the executor in place of the global
maps provided through the loop chain; this will be discussed in more detail in Section 7
and in the Supplementary Materials.
• The color gives a tile a scheduling priority. If shared-memory parallelism is requested,
adjacent tiles are given different colors (the adjacency relation is determined through the
maps available in L). Otherwise, colors are assigned in increasing order (i.e., Ti is given
1 The algorithm implemented in the library presented in Section 6.1 supports backwards tiling for shared-memory parallelism.
, Vol. 1, No. 1, Article 1. Publication date: August 2017.
Automated Tiling of Unstructured Mesh Computations
1:13
color i). The boundary tiles are always given colors higher than that of core tiles; the
non-exec tile has the highest color. The assignment of colors is carried by the function
color in Algorithm 1.
Field
region
iterations lists
local maps
color
Possible values
core, boundary, non-exec
one list of iterations [Ti ] j for each L j ∈ L
one list of local maps for each L j ∈ L; one local map for each map used in L j
an integer representing the execution priority
Table 2. The tile data structure.
Populating tiles by tracking data dependencies. To schedule a loop to tiles we use projections. A
projection is a function ϕ S : S → T. Initially, the projections set Φ is empty. Each time a loop is tiled,
Φ may be added some new projections or old projections may be updated. Φ, and consequently the
tiling functions for all loops in L, are derived incrementally (within the loop at line 1 in Algorithm 1)
starting from σseed : Sseed → T, the tiling function of Lseed . In the following, we discuss in detail
how projections and tiling functions are constructed.
Deriving a projection from a tiling function. Algorithm 3 takes as input (the descriptors of) L j
and its tiling function σ j : S j → T to update Φ. The algorithm also updates the conflicts matrix
C ∈ Nm×m , which indicates whether two tiles having the same color will become adjacent once
L j+1 is tiled.
A projection tells what tile a set element logically belongs to at a given point of the inspection. A
new projection ϕ S is needed if the elements of S are written by a loop. Let us consider the non-trivial
case in which writes or increments occur indirectly through a map M : S j → S t0 × S t1 × ... × S ta−1 .
To compute ϕ St , we first determine the inverse map (an example is shown in Figure 8). Then, we
iterate over all elements of S t and, for each e ∈ S t , we determine the last tile that writes to e, say
Tl ast . This is accomplished by applying the MAX function over the color of the tiles accessing e.
We finally simply set ϕ St [e] = Tl ast .
9
map:
3
7
9
1
1
3
inverse
map:
7
1
1
1
offset [3]
offset [7]
offset [9]
Fig. 8. Representation of an inverse map. The original map shows that the triangular cell 1 is adjacent to
three vertices, namely 3, 7, and 9. The inverse map associates vertices to cells. Since the mesh is unstructured,
different vertices can be incident to a different number of cells. The array offset determines the distance
between two consecutive vertices in the inverse map. For instance, all entries in the inverse map between
offset[3] and offset[4] are cells incident to vertex 3 (in the illustrated case, the first of these is cell 1).
Deriving a tiling function from the available projections. Using Φ, we compute σ j as described in
Algorithm 4. The algorithm is similar to the projection of a tiled loop, with the main difference
being that now we use Φ to schedule iterations correctly. Finally, σ j is inverted and the iterations
are added to the corresponding iteration lists [Ti ]j , for all Ti ∈ T.
, Vol. 1, No. 1, Article 1. Publication date: August 2017.
1:14
F. Luporini et al.
Detection of conflicts. If C indicates the presence of at least one conflict, say between Ti 1 and Ti 2 ,
we add a “fake connection” between these two tiles and loop back to the coloring stage. Ti 1 and Ti 2
are now connected, so they will be assigned different colors.
On the history of the algorithm. The first algorithm for generalized sparse tiling inspection was
introduced in [34]. In this section, a new, enhanced version of that algorithm has been presented.
In essence, the major differences are: (i) support for distributed-memory parallelism; (ii) use of
coloring instead of a task graph for tile scheduling; (iii) speculative inspection with backtracking if
a coloring conflict is detected; (iv) use of sets, instead of datasets, for data dependency analysis; (v)
use of inverse maps for parallelization of the projection and tiling routines; (vi) computation of
local maps. Most of these changes contributed to reduce the inspection cost, as discussed later.
5.2
Executor
The sparse tiling executor is illustrated in Algorithm 2. It consists of four main phases: (i) exchange of
halo regions amongst neighboring processes through non-blocking communications; (ii) execution
of core tiles (in overlap with communication); (iii) wait for the termination of the communications;
(iv) execution of boundary tiles.
As explained in Sections 3.3, a sufficiently deep halo region enables correct computation of
the boundary tiles. Further, tiles are executed atomically, meaning that all iterations in a tile are
computed without ever synchronizing with other processes. The depth of the boundary region,
which affects the amount of off-process data to be redundantly computed, increases with the number
n of loops to be fused. In the example in Figure 6, there are n = 3 loops, and three “strips” of extra
vertices are necessary for correctly computing the fused loops without tile-to-tile synchronizations.
We recall from Section 2 that the depth of the loop chain indicates the extent of the boundary
region. This parameter imposes a limit to the number of fusible loops. If L includes more loops
than the available boundary region – that is, if n > depth – then L will have to be split into
shorter loop chains, to be fused individually. As we shall see (Section 6), in our inspector/executor
implementation the depth is controlled by the Firedrake’s DMPlex module.
6
IMPLEMENTATION: SLOPE, PYOP2, AND FIREDRAKE
The implementation of automated sparse tiling is distributed over three open-source software
modules.
Firedrake An established framework for the automated solution of PDEs through the finite
element method [28].
PyOP2 A module used by Firedrake to apply numerical kernels over sets of mesh components.
Parallelism is handled at this level.
SLOPE A library for writing inspector/executor schemes, with primary focus on sparse tiling.
PyOP2 uses SLOPE to apply sparse tiling to loop chains.
The SLOPE library is a side-product of the present research. The interplay amongst Firedrake,
PyOP2 and SLOPE is outlined in Figure 9 and discussed in more detail in the following sections.
6.1
SLOPE: a Library for Sparse Tiling Irregular Computations
SLOPE is an open source software that provides an interface to build loop chains and to express
inspector/executor schemes for sparse tiling2 .
The loop chain abstraction implemented by SLOPE has been formalized in Section 2. In essence, a
loop chain comprises some sets (including the separation into core, boundary, and non-exec regions),
2 SLOPE
is available at https://github.com/coneoproject/SLOPE
, Vol. 1, No. 1, Article 1. Publication date: August 2017.
Automated Tiling of Unstructured Mesh Computations
1A
UFL
specification
1B
PyOP2
program
loop _chain
loop _chain
Firedrake
PyOP2
2
1:15
1C
3
p
y
t
h
o
n
C program
Slope
4
Fig. 9. Sparse tiling in the Firedrake-PyOP2-SLOPE framework. There are three ways of sparse tiling a loop
chain: decorating a Firedrake program (1A), decorating a sequence of loops in a PyOP2 program (1B), writing
both the loop chain and the inspector/executor codes explicitly in C through calls to SLOPE (1C). Both (1A)
and (1B) use the loop_chain interface (details in Section 6.2). The kernels generated within a loop_chain
are pre-processed in PyOP2 (2) and forwarded to SLOPE through its Python interface (3). SLOPE now has
access to the loop chain, so it can generate an inspector/executor scheme and return it to PyOP2 (4). The
inspector is compiled and executed. The result, a schedule (i.e., the output of Algorithm 1), is cached and
used as input to the executor. Each time the same loop_chain is encountered in a Firedrake/PyOP2 program,
the corresponding schedule is reused.
maps between sets, and a sequence of loops. Each loop has one or more descriptors specifying
what and how different sets are accessed. The example in Figure 2b illustrates the interface exposed
by SLOPE. It implements the algorithms in Section 5. Further, it provides additional features to
estimate the effectiveness and to verify the correctness of sparse tiling, including generation of
VTK files (suitable for Paraview [2]), to visualize the partitioning of the mesh into colored tiles, as
well as insightful inspection summaries (showing, for example, number and average size of tiles,
total number of colors used, time spent in critical inspection phases).
In the case of shared-memory parallelism, the following sections of code are parallelized through
OpenMP:
Inspection The projection and tiling algorithms; in particular, the loops over set elements in
Algorithm 3 and Algorithm 4).
Execution The computation of same colored tiles; that is, the two loops over tiles in Algorithm 2.
6.2
PyOP2: Lazy Evaluation and Interfaces
PyOP2 is a Python library offering abstractions to model an unstructured mesh – in terms of sets
(e.g. vertices, edges), maps between sets (e.g., a map from edges to vertices to express the mesh
topology), and datasets associating data to each set element (e.g. 3D coordinates to each vertex)
– and applying numerical kernels to sets of entities [28]. In this section, we focus on the three
relevant contributions to PyOP2 made through our work: (i) the interface to identify loop chains;
(ii) the lazy evaluation mechanism that allows loop chains to be built; (iii) the interaction with
SLOPE to automatically build and execute inspector/executor schemes.
To apply sparse tiling to a sequence of loops, the loop_chain interface was added to PyOP2. This
interface, exemplified in Figure 10, is also exposed to the higher layers, for example in Firedrake. In
the listing, the name uniquely identifies a loop chain. Other parameters (most of them optional) are
useful for performance evaluation and performance tuning. Amongst them, the most important are
the tile_size and the fusion_scheme. The tile_size specifies the initial average size for the
seed partitions. The fusion_scheme allows to specify how to break a long sequence of loops into
smaller loop chains, which makes it possible to experiment with a full set of sparse tiling strategies
without having to modify the source code.
, Vol. 1, No. 1, Article 1. Publication date: August 2017.
1:16
F. Luporini et al.
with loop_chain ( name , tile_size , fusion_scheme , ...) :
< some PyOP2 parallel loops are expressed / generated here >
Fig. 10. The loop_chain interface in PyOP2.
PyOP2 exploits lazy evaluation of parallel loops to generate an inspector/executor scheme.
The parallel loops encountered during the program execution – or, analogously, those generated
through Firedrake – are pushed into a queue, instead of being executed immediately. The sequence
of parallel loops in the queue is called the trace. If a dataset f needs to be read, for example because
a user wants to inspect its values or a global linear algebra operation is performed, then the trace is
traversed – from the most recent parallel loop to the oldest one – and a new sub-trace produced.
The sub-trace includes all parallel loops that must be executed to evaluate f correctly. The sub-trace
can then be executed or further pre-processed.
All loops in a trace that were created within a loop_chain scope are sparse tiling candidates. In
detail, the interaction between PyOP2 and SLOPE is as follows:
(1) Figure 10 shows that a loop_chain defines a new scope. As this scope is entered, a stamp
s 1 of the trace is generated. This happens “behind the scenes”, because the loop_chain is
a Python context manager, which can execute pre-specified routines prior and after the
execution of the body. As the loop_chain’s scope is exited, a new stamp s 2 of the trace is
computed. All parallel loops in the trace generated between s 1 and s 2 are placed into a
sub-trace for pre-processing.
(2) The pre-processing consists of two steps: (i) “simple” fusion – consecutive parallel loops
iterating over the same iteration space that do not present indirect data dependencies are
merged; (ii) generation of a loop chain representation for SLOPE.
(3) In (ii), PyOP2 inspects the sequence of loops and translates all relevant data structures (sets,
maps, loops) into a format suitable for the SLOPE’s Python interface. C code implementing
an inspector for the loops in the loop_chain is returned by SLOPE. PyOP2 compiles and
executes this code, which results in an inspection for the loop chain.
(4) A “software cache” mapping loop_chains to inspections is used. This whole process needs
therefore be executed only once for a given loop_chain.
(5) The executor is built in an analogous way to the inspector.
6.3
Firedrake/DMPlex: the S-depth Mechanism for Extended Halo Regions
Firedrake uses DMPlex [23] to handle meshes. DMPlex is responsible for partitioning, distributing
over multiple processes, and locally reordering a mesh. The MPI parallelization is therefore managed
through Firedrake/DMPlex.
During the start-up phase, each MPI process receives a contiguous partition of the original
mesh from DMPlex. The required PyOP2 sets, which can represent either topological components
(e.g., cells, vertices) or function spaces, are created. As intuitively shown in Figure 1, these sets
distinguish between multiple regions: core, owned, exec, and non-exec. Firedrake initializes the
four regions exploiting the information provided by DMPlex.
To support the loop chain abstraction, Firedrake must be able to allocate arbitrarily deep halo
regions. Both Firedrake and DMPlex have been extended to support this feature [19]. A parameter
called S-depth (the name has historical origins, see for instance [7]) regulates the extent of the halo
regions. A value S-depth = n indicates the presence of n strips of off-process data elements in each
set. The default value is S-depth = 1, which enables computation-communication overlap when
, Vol. 1, No. 1, Article 1. Publication date: August 2017.
Automated Tiling of Unstructured Mesh Computations
1:17
executing a single loop at the price of a small amount of redundant computation along partition
boundaries. This is the default execution model in Firedrake.
7
7.1
PERFORMANCE EVALUATION
The Seigen Computational Framework
Seigen is a seismological modelling framework capable of solving the elastic wave equation on
unstructured meshes. Exploiting the well-known velocity-stress formulation [39], the seismic model
is expressible as two first-order linear PDEs, which we refer to as velocity and stress. These
governing equations are discretized in space through the discontinuous-Galerkin finite element
method. The evolution over time is obtained by using a fourth-order explicit leapfrog scheme based
on a truncated Taylor series expansion of the velocity and stress fields. The particular choice of
spatial and temporal discretizations has been shown to be non-dissipative [8]. More details can
be found in [17]. Seigen, which is built on top of Firedrake, is part of OPESCI, an ecosystem of
software for seismic imaging based on automated code generation [15].
Seigen has a set of test cases, which differ in various aspects, such as the initial conditions of the
system and the propagation of waves. However, they are all based upon the same seismological
model; from a computational viewpoint, this means that, in a time step, the same sequence of loops
is executed. In the following, we focus on the explosive_source test case (see the work by [12]
for background details).
7.2
Implementation and Validation
In a time loop iteration, eight linear systems need to be solved, four from velocity and four from
stress. Each solve consists of three macro-steps: assembling a global matrix A; assembling a global
vector b; computing x in the system Ax = b. There are two global “mass” matrices, one for velocity
and one for stress. Both matrices are time invariant, so they are assembled before entering the
time loop, and block-diagonal, as a consequence of the spatial discretization employed (a block
belongs to an element in the mesh). The inverse of a block-diagonal matrix is again block-diagonal
and is determined by computing the inverse of each block. The solution of the linear system Ax = b,
expressible as x = bA−1 , can therefore be evaluated by looping over the mesh and computing a
“small” matrix-vector product in each element, where the matrix is a block in A−1 . Assembling
the global vectors boils down to executing a set of loops over mesh entities, particularly over
cells, interior facets, and exterior facets. Overall, twenty-five loops are executed in a time loop
iteration. Thanks to the hierarchy of “software caches” employed by Firedrake, the translation from
mathematical syntax into loops is only performed once.
Introducing sparse tiling into Seigen was relatively straightforward. Three steps were required:
(i) embedding the time stepping loop in a loop_chain context (see Section 6.2), (ii) propagating
user input relevant for performance tuning, (iii) creating a set of fusion schemes. A fusion scheme
establishes which sub-sequences of loops within a loop_chain will be fused and the respective seed
tile sizes. If no fusion schemes were specified, all of the twenty-five loops would be fused using a
default tile size. As we shall see, operating with a set of small loop chains and heterogeneous tile
sizes is often more effective than fusing long sequences of loops.
Seigen has several mechanisms to validate the correctness of the seismological model and the
test cases. The numerical results of all code versions (with and without tiling) were checked and
compared. Paraview was also used to verify the simulation output. Snapshots of the simulation
output are displayed in Figure 11.
, Vol. 1, No. 1, Article 1. Publication date: August 2017.
1:18
F. Luporini et al.
(a) Right after the explosion.
(b) Wave propagation.
(c) Final snapshot.
Fig. 11. Three phases of the explosive_source test case in Seigen, following an explosion at a point source.
7.3
Computational Analysis of the Loops
We here discuss computational aspects of the twenty-five fusible loops. The following considerations
derive from an analytical study of the data movement in the loop chain, extensive profiling through
the Intel VTune Amplifier tool [16], and roofline models (available in [17]).
Our initial hypothesis was that Seigen would have benefited from sparse tiling. Not only does
data reuse arise within single loops (e.g., by accessing vertex coordinates from adjacent cells), but
also across consecutive loops, through indirect data dependencies. This seemed to make Seigen a
natural fit for sparse tiling. The eight “solver” loops perform matrix-vector multiplications in each
mesh element. It is well established that linear algebra operations of this kind are memory-bound.
Four of these loops arise from velocity, the others from stress. There is significant data reuse
amongst the four velocity loops and amongst the four stress loops, since the same blocks in the
global inverse matrices are accessed. We therefore hypothesized performance gains if these loops
were fused through sparse tiling.
We also observed that because of the particular mesh employed, the exterior facet loops, which
implement the boundary conditions of the variational problems, have negligible cost. However, the
cells and facets loops do have significant cost, and data reuse across them arises. Six loops iterate
over the interior mesh facets to evaluate facet integrals, which ensure the propagation of information
between adjacent cells in discontinuous-Galerkin methods. The operational intensity of these loops
is much lower than that of cell loops, and memory-boundedness is generally expected. Consecutive
facet and cell integral loops share fields, which creates cross-loop data reuse opportunities, thus
strengthening the hypothesis about the potential of sparse tiling in Seigen.
All computational kernels generated in Seigen are optimized through COFFEE [24], a system
used in Firedrake that, in essence, (i) minimizes the operation count by restructuring expressions
and loop nests, and (ii) maximizes auto-vectorization opportunities by applying transformations
such as array padding and by enforcing data alignment.
7.4
Setup and Reproducibility
There are two parameters that we can vary in explosive_source: the polynomial order of the
method, q, and the input mesh. We test the spectrum q ∈ {1, 2, 3, 4}. To test higher polynomial
orders, changes to both the spatial and temporal discretizations would be necessary. For the spatial
discretization, the most obvious choice would be tensor product function spaces, which at the
moment of writing is still unavailable in Firedrake. We use as input a two-dimensional rectangular
domain of fixed size 300×150 tessellated with unstructured triangles (explosive_source only
supports two-dimension domains); to increase the number of elements in the domain, thus allowing
to weak scale, we vary the mesh spacing h.
The generality of the sparse tiling algorithms, the flexibility of the loop_chain interface, and
the S-depth mechanism made it possible to experiment with a variety of fusion schemes without
, Vol. 1, No. 1, Article 1. Publication date: August 2017.
Automated Tiling of Unstructured Mesh Computations
1:19
changes to the source code. Five fusion schemes were devised, based on the following criteria: (i)
amount of data reuse, (ii) amount of redundant computation over the boundary region, (iii) memory
footprint (the larger, the smaller the tile size to fit in cache). The fusion schemes are summarized
in Table 4. The full specification, along with the seed tile size for each sub loop chain, is available
at [46].
The experimentation was conducted on two platforms, whose specification is reported in Table 3.
The two platforms, Erebus (the “development machine”) and Helen (a Broadwell-based architecture
in the Helen cluster [33]) were idle and exclusive access had been obtained when the runs were
performed. Support for shared-memory parallelism is discontinued in Firedrake, so only distributedmemory parallelism with 1 MPI process per physical core was tested. The MPI processes were
pinned to cores. The hyperthreading technology was tried, but found to be generally non-profitable.
The code used for running the experiments was archived with the Zenodo data repository service:
Firedrake [42], PETSc [43], PETSc4py [44], FIAT [41], UFL [49], TSFC [48], PyOP2 [45], COFFEE [40],
SLOPE [47], and Seigen [46].
System
Node
DRAM
Erebus
1x4-core
Intel I7-2600 3.4GHz
Helen
2x14-core
Intel Xeon E5-2680 v4 2.40GHz
16 GB
128 GB (node)
Cache hierarchy
L1=32KB, L2=256KB, L3=8MB
L1=32KB, L2=256KB,
L3/socket=35MB
Compilers
Intel icc 16.0.2
Intel icc 16.0.3
Compiler flags
-O3 -xHost -ip
-O3 -xHost -ip
MPI version
Open MPI 1.6.5
SGI MPT 2.14
Table 3. Systems specification.
Fusion
scheme
Number of
loop chains
fs1
3
fs2
8
fs3
6
fs4
3
fs5
2
Criterion
Fuse costly cells
and facets loops
More aggressive
than fs1
fs2, include all
solver loops
More aggressive
than fs3
velocity, stress
S-depth
2
2
2
3
4
Table 4. Fusion schemes summary.
In the following, for each experiment, we collect three measurements.
Overall completion time – OT Used to compute the maximum application speed-up when
sparse tiling is applied.
Average compute time – ACT Sparse tiling impacts the kernel execution time by increasing
data locality. Communication is also influenced, especially in aggressive fusion schemes:
the rounds of communication decrease, while the data volume exchanged may increase.
ACT isolates the gain due to increased data locality from (i) the communication cost and (ii)
any other action performed in Firedrake (executing Python code) between the invocation
of kernels. This value is averaged across the processes.
Average compute and communication time – ACCT As opposed to ACT, the communication cost is also included in ACCT. By comparing ACCT and ACT, the communication
overhead can be derived.
As we shall see, all of these metrics will be essential for a complete understanding of the sparse
tiling performance.
To collect OT, ACT and ACCT, the following configuration was adopted. All experiments were
executed with “warm cache”; that is, with all kernels retrieved directly from the Firedrake’s software
cache of compiled kernels, so code generation and compilation times are not counted. All of the
non-tiled explosive_source tests were repeated three times. The minimum times are reported
(negligible variance). The cost of global matrix assembly – an operation that takes place before
entering the time loop – is not included in OT. Firedrake needs to be extended to assemble blockdiagonal matrices directly into vectors (an entry in the vector would represent a matrix block).
Currently, this is instead obtained in two steps: first, by assembling into a CSR matrix; then, by
, Vol. 1, No. 1, Article 1. Publication date: August 2017.
1:20
F. Luporini et al.
h
1.0
1.2
q
OT original (s)
OT tiled (s)
ΩOT
Ω ACT
Ω ACCT
1
2
3
4
1
2
3
4
687
1453
3570
7057
419
870
1937
4110
596
1200
2847
5827
365
715
1549
3390
1.15
1.21
1.25
1.21
1.15
1.22
1.25
1.21
1.17
1.25
1.28
1.22
1.17
1.26
1.28
1.23
1.16
1.25
1.27
1.22
1.16
1.26
1.27
1.22
Table 5. OT, ACT and ACCT on Erebus, with 1 MPI process per core.
explicitly copying the diagonal into a vector (a Python operation). The assembly per se never
takes more than 3 seconds, so it was reasonable to exclude this temporary overhead from our
timing. The inspection cost due to sparse tiling is included in OT, and its overhead will be discussed
appropriately in a later section. Extra costs were minimized: no check-pointing, only two I/O
sessions (at the beginning and at the end of the computation), and minimal logging. The time loop
has a fixed duration, while the time step depends on the mesh spacing h to satisfy the CourantFriedrichs-Lewy necessary condition (i.e., CFL limit) for the numerical convergence. In essence,
finer meshes require proportionately smaller time steps to ensure convergence.
We now proceed with discussing the achieved results. Below, a generic instance of explosive_source
optimized with sparse tiling will be referred to as a “tiled version”, otherwise the term “original
version” will be used.
7.5
Single-node Experimentation
Hundreds of single-node experiments were executed on Erebus, which was easily accessible in
exclusive mode and much quicker to use for VTune profiling. The rationale of these experiments
was to assess how sparse tiling impacts the application performance by improving data locality.
We only show parallel runs at maximum capacity (1 MPI process per physical core); the benefits
of sparse tiling in sequential runs tend to be negligible (if any), because (i) the memory access
latency is only marginally affected when a large proportion of bandwidth is available to a single
process, (ii) hardware prefetching is impaired by the small iteration space of tiles, (iii) translation
lookaside buffer (TLB) misses are more frequent due to tile expansion. Points (ii) and (iii) will be
elaborated upon.
Table 5 reports execution times and speed-ups, indicated with the symbol Ω, of the tiled version
over the original version for the three metrics OT, ACT and ACCT. We report the best speed-up
obtained after varying a number of parameters (tile size, fusion scheme and other optimizations
discussed below).
The parameters that we empirically varied to obtain these results were: (i) the fusion scheme,
fsX , X ∈ {1, 2, 3, 4, 5}, see Table 4; (ii) the seed tile size – for each fs and q, up to four different
values chosen to maximize the likelihood of fitting in L2 or L3 cache, were tried3 .
Further, a smaller experimentation varying (i) type of indirection maps (local or global, see
Section 5) and (ii) tile shape (through different mesh partitioning algorithms), as well as introducing
(iii) software prefetching and (iv) extended boundary region (to minimize redundant computation)
led to the conclusion that these optimizations may either improve or worsen the execution time, in
ways that are too difficult to predict beforehand. We therefore decided to exclude these parameters
from the full search space. This simplifies the analysis that will follow and also allowed the
3A
less extensive experimentation with “more adverse” tile sizes showed that: (i) a very small value causes dramatic
slow-downs (up to 8× slower than the original versions); (ii) larger values cause proportionately greater performance drops.
, Vol. 1, No. 1, Article 1. Publication date: August 2017.
fs4
fs5
140
250 320
Tile size factor
[q = 1, h = 1.2]
fs1
fs2
fs3
400
fs4
fs5
140
250 320
Tile size factor
400
1.30
1.25
1.20
1.15
1.10
1.05
1.00
1.30
1.25
1.20
1.15
1.10
1.05
1.00
[q = 2, h = 1.0]
fs1
fs2
fs3
fs4
fs5
70
140 200
Tile size factor
[q = 2, h = 1.2]
fs1
fs2
fs3
300
fs4
fs5
70
140 200
Tile size factor
300
1.30
1.25
1.20
1.15
1.10
1.05
1.00
1.30
1.25
1.20
1.15
1.10
1.05
1.00
[q = 3, h = 1.0]
fs1
fs2
fs3
fs4
fs5
ACT speed-up
fs3
1:21
45
60
Tile size factor
[q = 3, h = 1.2]
fs1
fs2
fs3
75
fs4
fs5
ACT speed-up
fs2
ACT speed-up
fs1
ACT speed-up
1.30
1.25
1.20
1.15
1.10
1.05
1.00
[q = 1, h = 1.0]
ACT speed-up
1.30
1.25
1.20
1.15
1.10
1.05
1.00
ACT speed-up
ACT speed-up
ACT speed-up
Automated Tiling of Unstructured Mesh Computations
45
60
Tile size factor
75
1.30
1.25
1.20
1.15
1.10
1.05
1.00
1.30
1.25
1.20
1.15
1.10
1.05
1.00
[q = 4, h = 1.0]
fs1
20
fs2
fs3
45
Tile size factor
[q = 4, h = 1.2]
fs1
20
fs4
fs2
fs3
45
Tile size factor
fs5
70
fs4
fs5
70
Fig. 12. Search space exploration on Erebus, with h ∈ {1.0, 1.2} and q ∈ {1, 2, 3, 4}. Each plot shows the
average compute time (ACT) speed-up achieved by multiple tiled versions over the original (non-tiled) version.
The seed tile size of a loop chain in an fs, in terms of seed loop iterations, is the product of the “tile size
factor” (x-axis) and a pre-established multiplier (an integer in the range [1, 4]; full list available at [46]).
execution of the whole test suite in less than five days. The interested reader is invited to refer to
the Supplementary Materials attached to this article for more information.
In Figure 12, we summarize the results of the search space exploration. A plot shows the ACT
speed-ups achieved by multiple tiled versions over the non-tiled version for given q and h. In these
single-node experiments, the ACCT trend was basically identical (as one can infer from Table 5),
with variations in speed-up smaller than 1%.
PyOP2 was enhanced with a loop chain analyzer (LCA) capable of estimating the best- and
worst-case tile memory footprint, as well as the percentage of data reuse ideally achievable4 . We use
this tool, as well as Intel VTune, to explain the results shown in Figure 12. We make the following
observations.
• fs, unsurprisingly, is the parameter having largest impact on the ACT. By influencing the
fraction of data reuse convertible into data locality, the amount of redundant computation
and the data volume exchanged, fusion schemes play a fundamental role in sparse tiling.
This makes automation much more than a desired feature: without any changes to the
source code, multiple sparse tiling strategies could be studied and tuned. Automation is one
of our major contributions, and this performance exploration justifies the implementation
effort.
• There is a non-trivial relationship between ACT, q and fs. The aggressive fusion schemes
are more effective with high q – that is, with larger memory footprints – while they tend
to be less efficient, or even deleterious, when q is low. The extreme case is fs5, which fuses
two long sequences of loops (twelve and thirteen loops each). In Figure 12 (Erebus), fs5 is
never a winning choice, although the difference between fs3/fs4 and fs5 decreases as q
grows. If this trend continued with q > 4, then the gain from sparse tiling could become
increasingly larger.
4 It
is part of our future plans to integrate this tool with SLOPE, where the effects of tile expansion can be taken into account
to provide better estimates.
, Vol. 1, No. 1, Article 1. Publication date: August 2017.
1:22
F. Luporini et al.
• A non-aggressive scheme fuses only a few small subsets of loops. As discussed in later
sections, these fusion schemes, despite affording larger tile sizes than the more aggressive
ones (due to the smaller memory footprint), suffer from limited cross-loop data reuse. For
fs1, LCA determines that the percentage of reusable data in the three fused loop chains
decreases from 25% (q = 1) to 13% (q = 4). The drop is exacerbated by the fact that no
reuse can be exploited for the maps. Not only are these ideal values, but also a significant
number of loops are left outside of loop chains. The combination of these factors motivate
the lack of substantial speed-ups. With fs2, a larger proportion of loops are fused and the
amount of shared data increases. The peak ideal reuse in a loop chain reaches 54%, which
translates into better ACTs. A similar growth in data reuse can be appreciated in more
aggressive fusion schemes, with a peak of 61% in one of the fs5’s loop chains. Nevertheless,
the performance of fs5 is usually worse than fs4. As we shall clarify in Section 7.8, this
is mainly due to the excessive memory footprint, which in turn leads to very small tiles.
Speculatively, we tried running a sixth fusion scheme: a single loop chain including all of
the 25 loops in a time iteration. In spite of an ideal data reuse of about 70%, the ACT was
always significantly higher than all other schemes.
• Figure 12 shows the ACT for a “good selection” of tile size candidates. Our approach was as
follows. We took a very small set of problem instances and we tried a large range of seed
tile sizes. Very small tile sizes caused dramatic slow-downs, mostly because of ineffective
hardware prefetching and TLB misses. Tile sizes larger than a fraction of L3 cache also led
to increasingly higher ACTs. If we consider q = 4 in Figure 12, we observe that the ACT of
fs2, fs3, and fs4 grows when the initial number of iterations in a tile is as big as 70. Here,
LCA shows that the tile memory footprint is, in multiple loop chains, higher than 3MB,
with a peak of 6MB in fs3. This exceeds the proportion of L3 cache that a process owns
(on average), which explains the performance drop.
7.6
Multi-node Performance
Weak-scaling experiments were carried out on thirty-two nodes split over two racks in the Helen
cluster at Imperial College London [33, 38]. The node architecture as well as the employed software
stack are summarized in Table 3. The rationale of these experiments was to assess whether the
change in communication pattern and amount of redundant computation caused by sparse tiling
could affect the run-time performance.
For each q in the usual range [1−4], fs3 generally resulted in the best performance improvements,
due to its trade-off between gained data locality and restrained redundant computation (whose
effect obviously worsen as q grows). Figure 13 summarizes the ACCT speed-ups achieved by the
best tiled version (i.e., the one found by empirically varying the tile size, with same tile size factor
as in Figure 12) over the original version. The weak scaling trend is remarkable, with only a small
drop in the case q = 1 when switching from one rack (448 cores) to two racks (896 cores), which
disappears as soon as the per-process workload becomes significant. For instance, with q = 2, each
process already computes over more than 150k degrees of freedom for the velocity and stress fields.
The peak speed-up over the original version, 1.28×, is obtained in the test case q = 3 when running
with 448 processes.
7.7 Negligible Inspection Overhead
As explained in Section 7.4, the inspection cost was included in OT. In this section, we quantify
this overhead in a representative problem instance. In all other problem instances, the overhead
was either significantly smaller than or essentially identical (for reasons discussed below) to that
, Vol. 1, No. 1, Article 1. Publication date: August 2017.
Automated Tiling of Unstructured Mesh Computations
Weak scaling on Helen
q=1 q=2 q=3
1.4
ACCT speed-up (original / tiling)
1:23
q=4
1.2
1.0
0.8
0.6
0.4
0.2
0.0
28
(119K)
56
(240K)
112
(477K)
224
(953K)
Number of processes
(Mesh elements)
448
(1.91M)
896
(3.80M)
Fig. 13. Weak scaling performance of Seigen’s explosive_source on the Helen cluster. The ACCT speed-up
is relative to a single process. Results for q ∈ {1, 2, 3, 4} are displayed. The simulations ran for 1000 timesteps.
reported here. We consider explosive_source on Erebus with h = 1.0, q = 4, and fs5. With this
configuration, the time step was ∆t = 481 · 10−6 (we recall from Section 7.4 that ∆t is a function
of h). Given the simulation duration T = 2.5, in this test case 5198 time steps were performed. A
time step took on average 1.15 seconds. In each time step, twenty-five loops, fused as specified by
fs5, are executed. We know that in fs5 there are two sub loop chains, which respectively consist
of thirteen and twelve loops. To inspect these two loop chains, 1.4 and 1.3 seconds were needed
(average across the four MPI ranks, with negligible variance). Roughly 98% of the inspection time
was due to the projection and tiling functions, while only 0.2% was spent on the tile initialization
phase (see Section 5). These proportions are consistent across other fusion schemes and test cases.
After 200 time steps (less than 4% of the total) the inspection overhead was already close to 1%.
Consequently, the inspection cost, in this test case, was eventually negligible. The inspection cost
increases with the number of fused loops, which motivates the choice of fs5 for this analysis.
7.8 On the Main Performance Limiting Factor
The tile structure, namely its shape and size, is the key factor affecting sparse tiling in Seigen.
The seed loop partitioning and the mesh numbering determine the tile structure. The simplest
way of creating tiles consists of “chunking” the seed iteration space every ts iterations, with ts
being the initial tile size (see Section 5). Hence, the chunk partitioning inherits the original mesh
numbering. In Firedrake, and therefore in Seigen, meshes are renumbered during initialization
applying the reverse Cuthill-McKee (RCM) algorithm. Using chunk partitioning on top of an RCMrenumbered mesh has the effect of producing thin, rectangular tiles, as displayed in Figure 14a.
This dramatically affects tile expansion, as a large proportion of elements will lie on the tile border.
There are potential solutions to this problem. The most promising would consist of using a Hilbert
curve, rather than RCM, to renumber the mesh. This would lead to more regular polygonal tiles
when applying chunk partitioning. Figures 14b and 14c show the tile structure that we would
ideally want in Seigen, from a Hilbert-renumbered mesh produced outside of Firedrake. As later
discussed, introducing support for Hilbert curves in Firedrake is part of our future work.
The memory footprint of a tile grows quite rapidly with the number of fused loops. In particular,
the matrices accessed in the velocity and stress solver loops have considerable size. The larger
the memory footprint, the smaller ts for the tile to fit in some level of cache. Allocating small tiles
, Vol. 1, No. 1, Article 1. Publication date: August 2017.
1:24
F. Luporini et al.
(a) chunk partitioning in Seigen.
(b) Ideal Hilbert partitioning.
(c) Zoomed in Hilbert partitioning.
Fig. 14. Representation of seed iteration space partitioning strategies via SLOPE and Paraview.
has unfortunately multiple implications. First, the proportion of iterations lying on the border grows,
which worsen the tile expansion phenomenon discussed, thus affecting data locality. Secondly, a
small ts impairs hardware prefetching, since the virtual address streams become more irregular.
Finally, using small tiles implies that a proportionately larger number of tiles are needed to “cover”
the iteration space; in the case of shared-memory parallelism, this increases the probability of
coloring conflicts, which result in higher inspection costs.
8
RELATED WORK
The data dependence analysis that we have developed in this article is based on the loop chain
abstraction, which was originally presented in [21]. This abstraction is sufficiently general to
capture data dependencies in programs structured as arbitrary sequences of loops, particularly to
create inspector/executor schemes for many unstructured mesh application. Inspector/executor
strategies were first formalized by [32]. They have been used to exploit data reuse and to expose
shared-memory parallelism in several studies [9, 10, 20, 36].
Sparse tiling is a technique based upon inspection/execution. The term was coined by [36, 37] in
the context of the Gauss-Seidel algorithm and also used in [35] in the Moldyn benchmark. However,
the technique was initially proposed by [10] to parallelize computations over unstructured meshes,
taking the name of unstructured cache blocking. In this work, the mesh was initially partitioned; the
partitioning represented the tiling in the first sweep over the mesh. Tiles would then shrink by one
layer of vertices for each iteration of the loop. This shrinking represented what parts of the mesh
could be accessed in later iterations of the outer loop without communicating with the processes
executing other tiles. The unstructured cache blocking technique also needed to execute a serial
clean-up tile at the end of the computation. [1] also developed an algorithm very similar to sparse
tiling, to parallelize Gauss-Seidel computations. The main difference between [36, 37] and [10] was
that in the former work the tiles fully covered the iteration space, so a sequential clean-up phase at
the end could be avoided. All these approaches were either specific to individual benchmarks or
not capable of scheduling across heterogeneous loops (e.g., one over cells and another over degrees
of freedom). These limitations had been addressed in [34].
The automated code generation technique presented in [29] examines the data affinity among
loops and performs partitioning with the goal of minimizing inter-process communication, while
maintaining load balancing. This technique supports unstructured mesh applications (being based
on an inspector/executor strategy) and targets distributed memory systems, although it does not
exploit the loop chain abstraction and does not introduce any sort of loop reordering transformation.
Automated code generation techniques based on polyhedral compilers have been applied to
structured grid benchmarks or proxy applications [5]. However, there has been very little effort in
providing evidence that these tools can be effective in real-world applications. Time-loop diamond
, Vol. 1, No. 1, Article 1. Publication date: August 2017.
Automated Tiling of Unstructured Mesh Computations
1:25
tiling was applied in [4] to a proxy application, but experimentation was limited to shared-memory
parallelism. In [31], instead, an approach based on raising the level of abstraction, similar to the
one presented in this paper, is described and evaluated. The experimentation is conducted using
realistic stencil codes ported to the OPS library. The main difference with respect to our work is
the focus on structured grids (i.e., different types of numerical methods are targeted).
In structured codes, multiple layers of halo, or “ghost” elements, are often used to reduce
communication [3]. Overlapped tiling exploits the very same idea: trading communication for
redundant computation along the boundary [50]. Several works tackle overlapped tiling within
single regular loop nests (mostly stencil-based computations), for example [6, 22, 25]. Techniques
known as “communication avoiding” [9, 26] also fall in this category. To the best of our knowledge,
overlapped tiling for unstructured mesh applications has only been studied analytically, by [13].
Further, we are not aware of any prior techniques for automation.
9
FUTURE WORK
Using a Hilbert curve numbering will lead to dramatically better tile shapes, thus mitigating the
performance penalties due to tile expansion, TLB misses and hardware prefetching described in
Section 7. This extension is at the top of our future work priorities.
Shared-memory parallelism was not as carefully tested as distributed-memory parallelism. First
of all, we would like to replace the current OpenMP implementation in SLOPE with the MPI Shared
Memory (SHM) model introduced in MPI-3. Not only does a unified programming model provide
significant benefits in terms of maintainability and complexity, but the performance may also be
greater as suggested by recent developments in the PETSc community. Secondly, some extra work
would be required for a fair comparison of this new hybrid MPI+MPI programming model with
and without sparse tiling.
The experimentation was carried out on a number of “conventional” Intel Xeon architectures;
we aim to repeat the experimentation on the Intel’s Knights Landing soon.
Finally, a cost model for automatic derivation of fusion schemes and tile sizes is still missing.
10 CONCLUSIONS
Sparse tiling aims to turn the data reuse in a sequence of loops into data locality. In this article,
three main problems have been addressed: the specialization of sparse tiling to unstructured mesh
applications via a revisited loop chain abstraction, automation via DSLs, effective support for
shared- and distributed-memory parallelism. The major strength of this work lies in the fact that
all algorithmic and technological contributions presented derive from an in-depth study of realistic
application needs. The performance issues we found through Seigen would never have been exposed
by a set of simplistic benchmarks, as often used in the literature. Further experimentation will be
necessary when 3D domains and high-order discretizations will be supported by Seigen. In essence,
the performance experimentation shows systematic speed-ups, in the range 1.10×-1.30×. This is
hopefully improvable by switching to Hilbert curve numberings and by exploiting shared memory
through a suitable paradigm. Finally, our opinion is that sparse tiling is an “extreme” optimization:
at least for unstructured mesh application, it is unlikely that it will lead to speed-ups in the order
of magnitudes. However, through automation via DSLs, and with suitable optimization and tuning,
it may still play a key role in improving the performance of real-world computations.
, Vol. 1, No. 1, Article 1. Publication date: August 2017.
1:26
11
11.1
F. Luporini et al.
SUPPLEMENTARY MATERIALS
Summary of the Optimizations Attempted in Seigen
A number of potential optimizations were attempted when applying sparse tiling to Seigen. The
experimentation with these optimizations was, however, inconclusive. Although performance
improvements were observed in various problem instances, in a significant number of other cases
either a detrimental effect or no benefits were noticed at all. Below we briefly discuss the impact of
four different execution strategies: (i) use of global maps, (ii) variation in tile shape, (iii) software
prefetching, (iv) use of an extended boundary region to minimize redundant computation.
Global and local maps Algorithm 1 computes so called local maps to avoid an extra level
of indirection in the executor. Although no data reuse is available for the local maps (each
fused loop has its own local maps), there might be benefits from improved hardware
prefetching and memory latency. We compared the use of global and local maps (i.e., the
former are normally constructed by Firedrake and provided in the loop chain specification,
the latter are computed by SLOPE), but no definitive conclusion could be drawn, as both
performance improvements and deteriorations within a 5% range were observed.
Tile shape The seed loop partitioning and the mesh numbering determine the tile shape.
The simplest way of creating tiles consists of “chunking” the seed iteration space every
ts iterations, with ts being the initial tile size (see Section 5.1). The chunk partitioning
inherits the original mesh numbering. In Firedrake, and therefore in Seigen, meshes are
renumbered during initialization applying the reverse Cuthill-McKee (RCM) algorithm.
Using chunk partitioning on top of an RCM-renumbered mesh has the effect of producing
thin, rectangular tiles. This affects tile expansion, as a large proportion of elements lie on
the tile border. There are potential solutions to this problem. The most promising would
consist of using a Hilbert curve, rather than RCM, to renumber the mesh. This would lead
to more regular polygonal tiles when applying chunk partitioning. Unfortunately, Hilbert
curves are at the moment not supported in Firedrake. An alternative consists of partitioning
the seed iteration space with a library like METIS [18] before applying RCM. However,
we discovered that this approach too was not exempt from side effects. The main cause
was increased translation lookaside buffer (TLB) misses, which occur whenever the CPU
cannot retrieve the mapping to the physical page corresponding to a given virtual page.
Since the page table has a hierarchical structure, handling a TLB miss usually requires
multiple accesses to memory. Hence, TLB misses are much more costly than cache misses.
Sparse tiling increases the TLB miss/hit ratio because of the fragmented streams of virtual
addresses. This is evident (and more pronounced) when the tile size is small, in which
case a TLB miss is quite likely to occur when jumping to executing a new loop. This
problem is exacerbated by the metis partitioning (in contrast to chunk), which leads to
irregular tile shapes. Here, tile expansion may eventually incorporate iterations living in
completely different virtual pages. VTune experimentation with q = 1 and q = 2 versions
of explosive_source showed that chunk- and metis-based sparse tiling suffer from an
increase in TLB misses of roughly 16% and 35%, respectively. To mitigate this issue, we
explored the possibility of using larger virtual pages through Linux’s Transparent Huge
Pages mechanism, which was enabled to automatically allocate memory in virtual pages
of 2MB (instead of the default 4KB) – as long as the base array addresses were properly
aligned. However, no significant differences were observed, and a deeper investigation is
still necessary.
, Vol. 1, No. 1, Article 1. Publication date: August 2017.
Automated Tiling of Unstructured Mesh Computations
1:27
Software prefetching In a loop, there is usually more than a single stream of memory accesses amenable to hardware prefetching (e.g., accesses to the indirection maps; direct
accesses to data values; indirect accesses to data values if the mesh has a good numbering).
Sparse tiling, unfortunately, impairs hardware prefetching for two reasons: (i) the virtual
addresses streams are considerably shorter; (ii) tile expansion introduces irregularities in
these streams. Software prefetching can be used together with hardware prefetching to
minimize memory stalls. PyOP2 and SLOPE have been extended to emit intrinsics instructions to prefetch the iteration i’s maps and data values while executing the iteration i − k
at distance k. No compelling evidence that this further transformation could systematically
improve the performance was found.
Extended boundary region The special non-exec tile Tne (see Sections 3 and 5) reduces the
amount of redundant computation in long loop chains by expanding over boundary tiles.
There are two ways of creating Tne : either an extra layer of data is added to the boundary
region (e.g., see Figure 6) or during inspection, by searching for mesh boundaries. The
current implementation only supports the first option. Manually deriving Tne would be not
only algorithmically complex, but also potentially very expensive.
REFERENCES
[1] M. F. Adams and J. Demmel. 1999. Parallel Multigrid Solver Algorithms and Implementations for 3D Unstructured
Finite Element Problem. In Proceedings of SC99. Portland, Oregon.
[2] Utkarsh Ayachit. 2015. The ParaView Guide (Full Color Version): A Parallel Visualization Application (paraview 4.3 ed.).
Kitware, Incorporated. http://www.amazon.com/exec/obidos/redirect?tag=citeulike07-20&path=ASIN/1930934300
[3] Federico Bassetti, Kei Davis, and Daniel J. Quinlan. 1998. Optimizing Transformations of Stencil Operations for Parallel
Object-Oriented Scientific Frameworks on Cache-Based Architectures. In Proceedings of the Second International
Symposium on Computing in Object-Oriented Parallel Environments (ISCOPE ’98). Springer-Verlag, London, UK, UK,
107–118. http://dl.acm.org/citation.cfm?id=646894.709707
[4] Uday Bondhugula, Vinayaka Bandishti, Albert Cohen, Guillain Potron, and Nicolas Vasilache. 2014. Tiling and
Optimizing Time-iterated Computations on Periodic Domains. In Proceedings of the 23rd International Conference on
Parallel Architectures and Compilation (PACT ’14). ACM, New York, NY, USA, 39–50. DOI:https://doi.org/10.1145/
2628071.2628106
[5] Uday Bondhugula, Albert Hartono, J. Ramanujam, and P. Sadayappan. 2008. A Practical Automatic Polyhedral
Parallelizer and Locality Optimizer. In Proceedings of the 2008 ACM SIGPLAN Conference on Programming Language
Design and Implementation (PLDI ’08). ACM, New York, NY, USA, 101–113. DOI:https://doi.org/10.1145/1375581.
1375595
[6] Li Chen, Zhao-Qing Zhang, and Xiao-Bing Feng. 2002. Redundant computation partition on distributed-memory
systems. In Algorithms and Architectures for Parallel Processing, 2002. Proceedings. Fifth International Conference on.
252–260. DOI:https://doi.org/10.1109/ICAPP.2002.1173583
[7] A. T. Chronopoulos. 1991. s-Step Iterative Methods for (Non)Symmetric (In)Definite Linear Systems. SIAM J. Numer.
Anal. 28, 6 (1991), 1776–1789. DOI:https://doi.org/10.1137/0728088 arXiv:http://dx.doi.org/10.1137/0728088
[8] Sarah Delcourte, Loula Fezoui, and Nathalie Glinsky-Olivier. 2009. A high-order Discontinuous Galerkin method for
the seismic wave propagation. ESAIM: Proceedings 27 (2009), 70–89. DOI:https://doi.org/10.1051/proc/2009020
[9] James Demmel, Mark Hoemmen, Marghoob Mohiyuddin, and Katherine Yelick. 2008. Avoiding Communication in
Sparse Matrix Computations. In Proceedings of International Parallel and Distributed Processing Symposium (IPDPS).
IEEE Computer Society.
[10] Craig C. Douglas, Jonathan Hu, Markus Kowarschik, Ulrich Rüde, and Christian Weiß. 2000. Cache Optimization for
Structured and Unstructured Grid Multigrid. Electronic Transaction on Numerical Analysis (February 2000), 21–40.
[11] Denys Dutykh, Raphaël Poncet, and Frédéric Dias. 2011. The VOLNA code for the numerical modeling of tsunami
waves: Generation, propagation and inundation. European Journal of Mechanics - B/Fluids 30, 6 (2011), 598 – 615.
Special Issue: Nearshore Hydrodynamics.
[12] W. W. Garvin. 1956. Exact Transient Solution of the Buried Line Source Problem. Proceedings of the Royal Society of
London, Series A 234, 1199 (1956). DOI:https://doi.org/10.1098/rspa.1956.0055
[13] M. B. Giles, G. R. Mudalige, C. Bertolli, P. H. J. Kelly, E. Laszlo, and I. Reguly. 2012. An Analytical Study of Loop
Tiling for a Large-Scale Unstructured Mesh Application. In Proceedings of the 2012 SC Companion: High Performance
, Vol. 1, No. 1, Article 1. Publication date: August 2017.
1:28
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
F. Luporini et al.
Computing, Networking Storage and Analysis (SCC ’12). IEEE Computer Society, Washington, DC, USA, 477–482. DOI:
https://doi.org/10.1109/SC.Companion.2012.68
M. B. Giles, G. R. Mudalige, Z. Sharif, G. Markall, and P. H. J. Kelly. 2011. Performance Analysis of the OP2 Framework
on Many-core Architectures. SIGMETRICS Perform. Eval. Rev. 38, 4 (March 2011), 9–15. DOI:https://doi.org/10.1145/
1964218.1964221
Gerard Gorman, Marcos de Aguiar, David Ham, Felix Herrmann, Christian Jacobs, Paul Kelly, Michael Lange, Chris
Pain, Matthew Piggott, Spencer Sherwin, Felippe Vieira Zacarias, Mike Warner, Fabio Luporini, and Navjot Kukreja.
2015. OPESCI project web page. http://www.opesci.org. (2015).
Intel Corporation. 2016. Intel VTune Performance Analyzer. software.intel.com/en-us/intel-vtune-amplifier-xe. (2016).
Christian T. Jacobs, Michael Lange, Fabio Luporini, and Gerard J. Gorman. 2017. Application of code generation to
high-order seismic modelling with the discontinuous Galerkin finite element method. (2017). In preparation.
George Karypis and Vipin Kumar. 2011. MeTis: Unstructured Graph Partitioning and Sparse Matrix Ordering System,
Version 5.0. http://www.cs.umn.edu/~metis. (2011).
Matthew G. Knepley, Michael Lange, and Gerard Gorman. 2015. Unstructured Overlapping Mesh Distribution in
Parallel. Submitted to ACM Transactions on Mathematical Software (TOMS) (2015). arXiv:1506.06194 http://arxiv.org/
abs/1506.06194
Christopher D. Krieger and Michelle Mills Strout. 2012. Executing Optimized Irregular Applications Using Task Graphs
Within Existing Parallel Models. In Proceedings of the Second Workshop on Irregular Applications: Architectures and
Algorithms (I A3 ) held in conjunction with SC12.
Christopher D. Krieger, Michelle Mills Strout, Catherine Olschanowsky, Andrew Stone, Stephen Guzik, Xinfeng
Gao, Carlo Bertolli, Paul Kelly, Gihan Mudalige, Brian Van Straalen, and Sam Williams. 2013. Loop Chaining: A
Programming Abstraction For Balancing Locality and Parallelism. In Proceedings of the 18th International Workshop on
High-Level Parallel Programming Models and Supportive Environments (HIPS). Boston, Massachusetts, USA.
Sriram Krishnamoorthy, Muthu Baskaran, Uday Bondhugula, J. Ramanujam, Atanas Rountev, and P Sadayappan.
2007. Effective Automatic Parallelization of Stencil Computations. SIGPLAN Not. 42, 6 (June 2007), 235–244. DOI:
https://doi.org/10.1145/1273442.1250761
Michael Lange, Lawrence Mitchell, Matthew G. Knepley, and Gerard J. Gorman. 2016. Efficient mesh management in
Firedrake using PETSc-DMPlex. SIAM Journal on Scientific Computing 38, 5 (2016), S143–S155. DOI:https://doi.org/10.
1137/15M1026092 arXiv:cs.MS/1506.07749
Fabio Luporini, Ana Lucia Varbanescu, Florian Rathgeber, Gheorghe-Teodor Bercea, J. Ramanujam, David A. Ham,
and Paul H. J. Kelly. 2015. Cross-Loop Optimization of Arithmetic Intensity for Finite Element Local Assembly. ACM
Trans. Archit. Code Optim. 11, 4, Article 57 (Jan. 2015), 25 pages. DOI:https://doi.org/10.1145/2687415
Jiayuan Meng and Kevin Skadron. 2009. Performance Modeling and Automatic Ghost Zone Optimization for Iterative
Stencil Loops on GPUs. In Proceedings of the 23rd International Conference on Supercomputing (ICS ’09). ACM, New
York, NY, USA, 256–265. DOI:https://doi.org/10.1145/1542275.1542313
Marghoob Mohiyuddin, Mark Hoemmen, James Demmel, and Katherine Yelick. 2009. Minimizing communication
in sparse matrix solvers. In Proceedings of the Conference on High Performance Computing Networking, Storage and
Analysis (SC ’09). ACM, Article 36, 12 pages. DOI:https://doi.org/10.1145/1654059.1654096
Florian Rathgeber. 2014. Productive and efficient computational science through domain-specific abstractions. Ph.D.
Dissertation. Imperial College London.
Florian Rathgeber, David A. Ham, Lawrence Mitchell, Michael Lange, Fabio Luporini, Andrew T. T. Mcrae, GheorgheTeodor Bercea, Graham R. Markall, and Paul H. J. Kelly. 2016. Firedrake: Automating the Finite Element Method by
Composing Abstractions. ACM Trans. Math. Softw. 43, 3, Article 24 (Dec. 2016), 27 pages. DOI:https://doi.org/10.1145/
2998441
Mahesh Ravishankar, John Eisenlohr, Louis-Noël Pouchet, J. Ramanujam, Atanas Rountev, and P. Sadayappan. 2012.
Code generation for parallel execution of a class of irregular loops on distributed memory systems. In Proc. Intl. Conf.
on High Perf. Comp., Net., Sto. & Anal. Article 72, 11 pages.
I. Z. Reguly, G. R. Mudalige, C. Bertolli, M. B. Giles, A. Betts, P. H. J. Kelly, and D. Radford. 2016. Acceleration of a
Full-Scale Industrial CFD Application with OP2. IEEE Transactions on Parallel and Distributed Systems 27, 5 (May 2016),
1265–1278. DOI:https://doi.org/10.1109/TPDS.2015.2453972
István Z Reguly, Gihan R Mudalige, and Mike B Giles. 2017. Loop Tiling in Large-Scale Stencil Codes at Run-time with
OPS. arXiv preprint arXiv:1704.00693 (2017).
Joel H. Salz, Ravi Mirchandaney, and Kay Crowley. 1991. Run-Time Parallelization and Scheduling of Loops. IEEE
Trans. Comput. 40, 5 (1991), 603–612.
Imperial College High Performance Computing Service. 2017. The cx2/Helen cluster. (2017). DOI:https://doi.org/10.
14469/hpc/2232
, Vol. 1, No. 1, Article 1. Publication date: August 2017.
Automated Tiling of Unstructured Mesh Computations
1:29
[34] M.M. Strout, F. Luporini, C.D. Krieger, C. Bertolli, G.-T. Bercea, C. Olschanowsky, J. Ramanujam, and P.H.J. Kelly. 2014.
Generalizing Run-Time Tiling with the Loop Chain Abstraction. In Parallel and Distributed Processing Symposium,
2014 IEEE 28th International. 1136–1145. DOI:https://doi.org/10.1109/IPDPS.2014.118
[35] Michelle Mills Strout, Larry Carter, and Jeanne Ferrante. 2003. Compile-time Composition of Run-time Data and
Iteration Reorderings. In Proc. ACM SIGPLAN Conf. Prog. Lang. Des. & Impl. (PLDI). ACM, New York, NY, USA.
[36] Michelle Mills Strout, Larry Carter, Jeanne Ferrante, Jonathan Freeman, and Barbara Kreaseck. 2002. Combining
Performance Aspects of Irregular Gauss-Seidel via Sparse Tiling. In Proceedings of the 15th Workshop on Languages
and Compilers for Parallel Computing (LCPC). Springer.
[37] Michelle Mills Strout, Larry Carter, Jeanne Ferrante, and Barbara Kreaseck. 2004. Sparse Tiling for Stationary Iterative
Methods. International Journal of High Performance Computing Applications 18, 1 (February 2004), 95–114.
[38] TOP500. 2016. cx2/Helen cluster specification in the TOP500 ranking. https://www.top500.org/system/178845. (2016).
[39] Jean Virieux. 1986.
P-SV wave propagation in heterogeneous media; velocity-stress finitedifference method.
Geophysics 51, 4 (1986), 889–901.
DOI:https://doi.org/10.1190/1.1442147
arXiv:http://geophysics.geoscienceworld.org/content/51/4/889.full.pdf
[40] Zenodo/COFFEE. 2017. coneoproject/COFFEE: A Compiler for Fast Expression Evaluation. (July 2017). DOI:https:
//doi.org/10.5281/zenodo.836678
[41] Zenodo/FIAT. 2017. firedrakeproject/fiat: The Finite Element Automated Tabulator. (July 2017). DOI:https://doi.org/
10.5281/zenodo.836679
[42] Zenodo/Firedrake. 2017. firedrakeproject/firedrake: an automated finite element system. (July 2017). DOI:https:
//doi.org/10.5281/zenodo.836680
[43] Zenodo/PETSc. 2017. firedrakeproject/petsc: Portable, Extensible Toolkit for Scientific Computation. (July 2017). DOI:
https://doi.org/10.5281/zenodo.836685
[44] Zenodo/PETSc4Py. 2017. firedrakeproject/petsc4py: The Python interface to PETSc. (July 2017). DOI:https://doi.org/
10.5281/zenodo.836684
[45] Zenodo/PyOP2. 2017. OP2/PyOP2: Framework for performance-portable parallel computations on unstructured
meshes. (July 2017). DOI:https://doi.org/10.5281/zenodo.836688
[46] Zenodo/Seigen. 2017. OPESCI/Seigen: An elastic wave equation solver for seismological problems based on the finite
element method . (Aug. 2017). DOI:https://doi.org/10.5281/zenodo.840000
[47] Zenodo/SLOPE. 2017. coneoproject/SLOPE: A run-time system to tile irregular loops. (July 2017). DOI:https:
//doi.org/10.5281/zenodo.836738
[48] Zenodo/TSFC. 2017. firedrakeproject/tsfc: The Two Stage Form Compiler firedrakeproject/tsfc: The Two Stage Form
Compiler. (July 2017). DOI:https://doi.org/10.5281/zenodo.836677
[49] Zenodo/UFL. 2017. firedrakeproject/ufl: The Unified Form Language. (July 2017). DOI:https://doi.org/10.5281/zenodo.
836683
[50] Xing Zhou, Jean-Pierre Giacalone, María Jesús Garzarán, Robert H. Kuhn, Yang Ni, and David Padua. 2012. Hierarchical
Overlapped Tiling. In Proceedings of the Tenth International Symposium on Code Generation and Optimization (CGO
’12). ACM, New York, NY, USA, 207–218. DOI:https://doi.org/10.1145/2259016.2259044
, Vol. 1, No. 1, Article 1. Publication date: August 2017.
| 5cs.CE |
A Decentralized Mechanism for Computing Competitive Equilibria in
Deregulated Electricity Markets∗
arXiv:1603.03083v2 [math.OC] 23 Mar 2016
Erik Miehling and Demosthenis Teneketzis†
Abstract— With the increased level of distributed generation
and demand response comes the need for associated mechanisms that can perform well in the face of increasingly complex deregulated energy market structures. Using Lagrangian
duality theory, we develop a decentralized market mechanism
that ensures that, under the guidance of a market operator, selfinterested market participants:generation companies (GenCos),
distribution companies (DistCos), and transmission companies
(TransCos), reach a competitive equilibrium. We show that even
in the presence of informational asymmetries and nonlinearities (such as power losses and transmission constraints), the
resulting competitive equilibrium is Pareto efficient.
Index Terms— Power systems, Optimization, Decentralized
control, Network analysis and control, Duality theory
I. INTRODUCTION
From the introduction of the Public Utilities Regulatory
Policies Act (PURPA) in 1978 to the establishment of the
Energy Policy Act in 1992, the deregulation of electricity
markets in the United States has grown continuously, primarily under the appeal of increased technological competition
and innovation. Today, despite cases of market manipulation
(such as the California electricity crisis in 2000-2001), many
large electricity markets are, at least in some capacity,
deregulated. This transition has been centered around the
formation of specialized firms for generation, transmission,
and distribution, to name a few, with markets typically
consisting of the following companies [1] (termed market
participants): generation companies (GenCos) who produce
and sell power, transmission companies (TransCos) who own
the transmission assets and are responsible for transmitting
power across the grid, and distribution companies (DistCos)
who own the distribution networks and are tasked with
buying power from GenCos and distributing it to consumers.
The primary goal in an energy market is determining an
outcome that is not only economically optimal (that is, it is
Pareto efficient [2], [3]) but also satisfies the physical constraints of the system. Achieving this goal in a deregulated
energy market is complicated by many factors. First, market
participants possess informational asymmetries arising from
private cost information and localized system knowledge.
Additionally, there are power losses within the network,
limits on transmission lines, contingency considerations, lack
of efficient power storage capabilities, and the underlying,
fundamental rule that power flow is dictated by the laws of
physics (Kirchhoff’s laws). These physical laws result in a
*This work was supported by NSF grant CNS-1238962.
† E. Miehling and D. Teneketzis are with the Department of Electrical
Engineering and Computer Science, University of Michigan, Ann Arbor,
MI, USA, 48109. E-mail: miehling@umich.edu.
phenomenon in power systems termed loop flow, creating
externalities in the market for electricity [4].
Centralized market mechanisms are traditionally the approach used for determining the optimal, feasible outcome of
the market [5]. Under these approaches a centralized market
operator receives bids from the market participants, in the
form of cost/benefit functions and technical constraints, and
solves a large-scale centralized optimization problem to determine the market clearing outcome. This outcome consists
of a physically feasible operating point as well as a vector
of bus-specific power prices termed locational marginal
prices (LMPs). Unfortunately, centralized mechanisms suffer
from some drawbacks. First, reporting cost and technical
information raises privacy concerns for market participants.
Also, as systems grow in size, the centralized optimization
problem can become prohibitively large. This is exacerbated
by the recent surge in distributed generation and demand
side participation [6], increasing the dimensionality and
complexity of the problem further and potentially making
centralized mechanisms computationally intractable.
In hopes of avoiding these drawbacks, we introduce a decentralized market mechanism which achieves the economically optimal outcome, honoring the informational asymmetries of the problem and considering important nonlinearities
of the system (such as power losses and limits on transmission lines). The electricity market model consists of multiple
market participants, DistCos, GenCos, and TransCos, and
a single market operator. Our model allows for the consumption centers of each DistCo and the production centers
of each GenCo to be distributed across the network. For
example, a given GenCo could own generators at multiple
buses in the network (a portfolio of plants). Additionally,
our model allows for the ownership of transmission lines in
the system to be partitioned among multiple TransCos. The
market operator is responsible for obtaining a market clearing
outcome. The process of achieving this market clearing
outcome, termed a decentralized market mechanism, is based
on principles from Lagrangian duality theory, specifically
making use of the dual decomposition method [7]. The
mechanism, which we refer to as the pricing process, consists
of an iterative price response and price update procedure. All
market participants are assumed to act in a self-optimizing
manner, that is, given the current LMPs they adjust their
decision variables in order to maximize their financial surplus
subject to their own local physical and operational constraints. This allows, for instance, for DistCos to exercise
flexible demand participation for the elastic component of
their total demand and for GenCos to self-dispatch. DistCos
and GenCos optimize independently, reporting their surplusmaximizing consumption and production profiles, respectively. TransCos partake in a cooperative message exchange
process to reach an operating point that induces power
flows that maximize their surpluses for transmitting power
along their respective lines. The optimizers are sent to the
market operator who is responsible for updating the LMPs
in such a way that the self-interested behavior of market
participants leads to an outcome that is physically feasible.
This outcome, when coupled with the associated set of
LMPs, forms a competitive equilibrium [2], [8], which we
show is Pareto efficient. Under relatively weak conditions
(a convex DC approximation and edge-wise positive sums
of LMPs), the market participants’ optimization problems
are convex and the pricing process converges. The pricing
process avoids the need for market participants to reveal
sensitive information, and additionally, the mechanism scales
much more effectively than its centralized counterpart.
The body of research concerning deregulated electricity
markets is vast. We focus on papers most similar to ours,
primarily including works that develop decentralized market
mechanisms (under perfect competition) using Lagrangian
duality techniques. Duality theory allows one to solve the
computationally simpler dual problem; however, this can
result in a non-zero duality gap in general. The authors
of [8], [9] construct a market model consisting of GenCos,
DistCos, and a single TransCo, considering a fully nonlinear
AC power flow model. Their decentralized mechanism, based
on a dual approach, is conjectured to converge to a zero duality gap solution under profit optimality and a convexifying
market rule (a restriction of market participants’ behavior).
Lavaei and Sojoudi [10] consider a competitive energy market setting with GenCos, DistCos, and an ISO under the AC
power flow model (using an SDP reformulation). Assuming
positive LMPs, the authors are able to show convergence to
a zero duality gap solution under the assumption of either:
a radial network, or, in the case of a mesh network, the
existence of a phase-shifter for each network cycle. In the
absence of phase-shifters, a zero duality gap can be ensured
if loads are allowed to be over-satisfied (discarding extra
power). Similar mechanisms have been applied in the context
of the unit commitment problem (e.g. [11], [12], [13]) and
demand response exchange markets (see [14] and [6]).
The contributions of this paper are twofold: 1) Modeling
generality: Our model allows for the ownership of power
system assets to be partitioned among the market participants. This allows for each DistCo and GenCo to own
multiple units that are distributed across the network (existing
literature assumes that each participant owns a single unit [8],
[9], [10]). Our model also allows for ownership of lines to
be partitioned among multiple TransCos ([8], [9] consider
a single TransCo); 2) Convergence to a zero duality gap
solution: Existing models contain nonlinearities that either
preclude convergence guarantees ([8], [9]) or require strong
sufficient conditions [10]. Our convex power flow approximation allows for power losses to be well-approximated
while (under natural conditions) ensuring convergence.
II. ENERGY MARKET MODEL
We consider a network of N buses, denoted by the set N ,
connected by transmission lines, denoted by the undirected
edge-set E. Each edge, {n, m} ∈ E, has a line limit
Knm = Kmn > 0 and an admittance Ynm = Gnm + iBnm
which consists of a conductance Gnm = Gmn > 0 and
a susceptance Bnm = Bmn > 0. We set Knm = 0 and
Ynm = 0 + i0 for {n, m} 6∈ E. Each bus has an associated
voltage angle, denoted by θn , with the vector of all angles
(termed the operating point) denoted by θ = {θn }n∈N .
In addition to the market operator (MO), the market
model in our paper contains three types of agents (market
participants): DistCos, denoted by the set D = {1, . . . , D};
GenCos, denoted by G = {1, . . . , G}; and TransCos, denoted
by T = {1, . . . , T }. Each DistCo i ∈ D owns consumption
units, consisting of elastic loads at buses NDi e ⊆ N and
inelastic loads at buses NDi s ⊆ N . The elastic
and inelastic
load profiles of DistCo i ∈ D are ei = ein n∈N i and
De
si = sin n∈N i , respectively, where ein ∈ [ein , ēin ] is the
Ds
elastic demand and sin ≥ 0 is the (given) inelastic demand
of DistCo i’s consumption unit at bus n. Each GenCo i ∈ G
owns generation units at buses NGi ⊆ N. The real power
injection profile of GenCo i ∈ G is pi = pin n∈N i where
G
pin ∈ [pin , p̄in ] is the injection of GenCo i’s generation unit at
bus n. Lastly, each TransCo i ∈ T owns a set of transmission
lines E i with ownership of lines in the system partitioned
among TransCos, that is, E 1 ∪· · ·∪ E T = E and E i ∩E j = ∅,
i 6= j. Each edge-set E i has an associated set of buses NTi
defined as the endpoints of the edges in E i . The associated
voltage angle profile of TransCo i ∈ T is θ i = θn n∈N i .
T
A sample network can be seen in Fig. 1.
1
p22
2
4
K14
p11
p21
K12
p14
e14
K34
K45
s12
3
K23
5
e15
e25
s25
Fig. 1. A sample 5-bus network. GenCo i = 1 owns generator units at buses
NG1 = {1, 4} corresponding to an injection vector p1 = (p11 , p14 ). GenCo
i = 2 has generator units at buses NG2 = {1, 2}, p2 = (p21 , p22 ); DistCo
i = 1 has elastic loads at buses 4 and 5, e1 = (e14 , e15 ), and an inelastic load
1 = {4, 5}, N 1 = {2}; and lastly, DistCo
at bus 2, s1 = s12 , thus ND
e
Ds
2 = N 2 = {5},
i = 2 has both an elastic and inelastic load at bus ND
e
Ds
2
2
2
2
thus e = e
,
s
=
s
.
Bus
3
is
a
zero-injection
bus.
TransCo
i = 1 owns
5
5
lines E 1 = {1, 2}, {1,4}, {2, 3} thus NT1 = {1, 2, 3, 4} and TransCo
i = 2 owns lines E 2 = {3, 4}, {4, 5} so NT2 = {3, 4, 5}.
For later convenience, we also define NTi,j := NTi ∩NTj as
the set of shared buses between two TransCos’ edge-sets E i
and E j and Tn := {i ∈ T | n ∈ NTi } as the set of TransCos
that own lines that are connected to bus n.
The load and generation profiles of DistCos and GenCos
have associated utilities and costs, respectively. For an elastic
load profile ei the aggregate utility
of
P (benefit) function
DistCo i is defined as ui ei := n∈N i e uin ein , where
D
uin (ein ) is the benefit associated with elastic demand level ein .
Similarly, GenCo
i’sPaggregate cost function (total generation
cost) is ci pi := n∈N i cin pin where cin (pin ) represents
G
the cost for producing real power pin .
A. Knowledge model
We now describe the knowledge model, that is, what
each of the power system entities knows about the system.
Each DistCo i ∈ D possesses private information regarding
their utility functions {uin }n∈NDi e and any bounds on the
elastic load level ei = {ein }n∈NDi e , ēi = {ēin }n∈NDi e .
Each GenCo i ∈ G possesses private information regarding
their cost functions {cin }n∈NGi and production bounds pi =
{pin }n∈NGi , p̄i = {p̄in }n∈NGi . Each TransCo i ∈ T knows
the connectivity of their region of the network, (NTi , E i ), as
well as the admittances of the corresponding lines, Ynm for
{n, m} ∈ E i . TransCos also possess private information of
the line limits of their transmission lines, Knm , {n, m} ∈ E i .
Each DistCo i ∈ D knows the inelastic demands at its buses,
{sin }n∈NDi , whereas the MO is assumed to know all inelastic
demand levels. Furthermore, the MO knows the location of
all DistCo and GenCo units, the network connectivity, and
the admittances of all transmission lines in the network.
B. Model Assumptions
We make five core assumptions regarding our model:
Assumption 1 (convex DC approximation): We propose a
power flow approximation which represents power flow as
a convex function of the angle difference. To begin the
derivation, recall that by the AC power flow equations [15]
the real power flowing from bus n to bus m is Pnm =
Gnm Vn2 −Gnm Vn Vm cos(θn−θm )+Bnm Vn Vm sin(θn−θm ).
We set voltage magnitudes to 1 p.u., Vn = 1 ∀n ∈ N , and
assume that voltage angle differences, θn − θm , are small
(similar to the DC approximation). Using the second-order
small angle approximations, sin(θn − θm ) ≈ θn − θm and
cos(θn −θm ) ≈ 1− 12 (θn −θm )2 , we can write the expression
for the power flow from bus n to bus m as a convex function
of the angle difference, θn −θm . The resulting approximation,
which we term the convex DC approximation, dictates that
the flow of power on line (n, m) is
1
g(θnm ) := Bnm (θn − θm ) + Gnm (θn − θm )2 . (1)
2
The above approximation maintains the asymmetry of the
power flow equations, g(θnm ) 6= −g(θmn ), and consequently
allows for power losses to be considered (unlike with the DC
power flow approximation). The real power losses along line
{n, m}, Lnm = Pnm + Pmn , are approximated by Lnm ≈
Gnm (θn − θm )2 . For notational convenience, we split Eq.
(1) into a DC component, ḡ(θnm ) := Bnm (θn − θm ), and a
(convex) loss component, g̃(θnm ) := 12 Gnm (θn − θm )2 .
Assumption 2 (slack buses): Denote the set of slack buses
by NS . We require that each TransCo has exactly one slack
bus, that is, NTi ∩ NS contains one element for all i ∈ T .
Slack buses serve solely as angle references, that is, θn = 0
for all n ∈ NS .
Assumption 3 (strong convexity): We require that all DistCo
utility functions uin are strongly concave and all GenCo cost
functions cin are strongly convex (this condition is equivalent
to strict convexity if the functions are quadratic). See [16]
for justification of the convexity assumption.
Assumption 4 (positive edge-wise sums of prices): We
require that all edge-wise sums of locational marginal prices
are positive. That is, λn + λm > 0 for all {n, m} ∈ E.1 Note
that this allows λn < 0 for some n.
Assumption 5 (price-taking behavior): We assume that the
agents (market participants) are price-taking, that is, they
assume that the price will remain unchanged if they change
their response. This assumption requires that no single agent
is large enough (apart from the MO) to influence the price;
an assumption which is reasonable in our market model as
the number of agents in the system increases.
III. MAXIMIZING SOCIAL WELFARE
We are interested in determining the set of variables,
consisting of DistCo elastic demand levels {ei }i∈D , GenCo
real power injection levels {pi }i∈G , and an operating point θ,
such that the social welfare is maximized, subject to physical
and operational constraints. It is known from microeconomic
theory that maximizing the social welfare results in a Pareto
efficient outcome [2]. The single time-period problem can be
formally stated as Problem (P) below.
X
X i i
max i
J(x) :=
ui ei −
c p
(P)
i
x=({e }i∈D ,{p }i∈G ,θ)
i∈D
i∈G
(P.i)
p ≤ p ≤ p̄ , i ∈ G
(P.ii)
s.t. p − (e + s) = f θ
i
i
i
i
i
i
e ≤ e ≤ ē , i ∈ D
g θnm ≤ Knm , g θmn ≤ Kmn , {n, m} ∈ E
(P.iii)
θn = 0, n ∈ NS
(P.vi)
θnm ≤ θnm ≤ θ̄nm , {n, m} ∈ E
θn ∈ [−π, π], n ∈ N
(P.iv)
(P.v)
(P.vii)
The objective function of Problem (P), J(x), represents the
social welfare and is defined as the total utility to DistCos
minus the total cost to GenCos.
The constraints of problem (P) arise from both physical
laws and the operational requirements of the power system
and the agents. The first constraint (P.i), termed the power
balance equation, takes the form
p − (e + s) = f θ
(2)
P
i
where p = (p1 , p2 , . . . , pN ), with pn =
i∈G pn , is the
net generation vector and (e + s) is the net demand vector
1 Note: We are not enforcing this as a constraint in our problem, rather we
are only considering topologies where this assumption is naturally satisfied.
consisting of two components, the elastic demand vector
e = (e1 , e2 , . . . , eN ) and the (fixed) inelastic
demand vector
P
i
sP = (s1 , s2 , . . . , sN ) (with
e
=
e
=
n
i∈D n and sn
i
s
).
The
vector
f
θ
=
f
θ
,
f
θ
,
.
.
.
,
f
θ
1
2
N
i∈D n
denotes the power injections induced by the operating point
θ, where the injection
P at bus n is defined by the convex
function fn θ = m∈N g θnm , where g(θnm ) represents
the power flow from bus n to m defined in Eq. (1); notice that
g(θnm ) is zero if {n, m} 6∈ E. Constraint (P.i) simply states
that the injections due to physical laws, f (θ), must agree with
the net generation and demand at every bus. Constraints (P.ii)
and (P.iii) reflect the fact that GenCos/DistCos have bounds
on the amount of power they are able to produce/consume.
Transmission constraints on the amount of power flowing on
each line, constraint (P.iv), stability constraints on the voltage
angle difference, (P.v), and slack references, (P.vi), are also
imposed. The last constraint, (P.vii), is a technical condition
that ensures that the voltage angles are well-defined. We
group constraints (P.ii)-(vii) into a set denoted by X. It is
clear that X is convex since it is the intersection of halfspaces and convex inequality constraints. Lastly, we assume
that Problem (P) is feasible.
There are some fundamental difficulties in obtaining a
solution to Problem (P). First, the problem is nonconvex
due the presence of the nonlinear power balance equation.
Furthermore, by the discussion in Section II-A, no single
entity in the system has the information required to obtain
a solution to Problem (P). The remainder of the paper will
focus on obtaining a solution to Problem (P).
IV. SURPLUSES & COMPETITIVE EQUILIBRIA
The notion of a competitive equilibrium will be of central
importance in obtaining a solution to Problem (P). Before
we formally define a competitive equilibrium in the context
of our problem, we need to discuss some aspects related to
the Lagrangian dual function of Problem (P).
A partial Lagrangian of Problem (P) is formed by dualizing the power balance equation through the vector of
dual variables, λ, where each component λn represents
the locational marginal price of power at bus n. Denoting
i
i
the vector of variables
by x = ({e }i∈D , {p }i∈G , θ), and
defining h(x) := f θ − p + e + s, the Lagrangian is
L x, λ := J(x) − λ> h(x)
X
X
=
ui ei − ci pi − λ> f θ − p + e + s
i∈D
=
i∈G
X X
i∈D n∈N i e
X X i i
uin ein −
cn pn
i∈G n∈NGi
D
!
−
X
n∈N
λn
fn (θ) −
X
i∈G
pin
+
X
ein
=
i∈D
i∈D
X
+
sin
X
X
uin ein − λn ein −
λn sin
i
n∈ND
e
i
n∈ND
s
+
X
i∈G
X
X
λn fn (θ). (3)
λn pin − cin pin −
n∈N
n∈NGi
Due to the structure of the Lagrangian, Eq. (3), the evaluation
the dual function, defined as φ(λ) = maxx∈X L x, λ ,
is greatly simplified via separable optimizations.
φ(λ) = max L x, λ
x∈X
X
X
X
i i
i
i
u
e
−
λ
e
−
λ
s
=
max
n
n
n n
n
n
i
i
i∈D
+
e ∈E
X
i∈G
i
n∈ND
e
p ∈P
+ max
θ∈Θ
i
n∈ND
s
X
i
i
i
λn pn − cn pn
max
i
i
−
n∈NGi
X
λn fn (θ)
(4)
n∈N
where the constraint sets are Ei = {ei |ei ≤ ei ≤ ēi }, Pi =
{pi |pi ≤ pi ≤ p̄i }, and Θ = {(θ 1 , . . . , θ T ) ∈ Θ1 × · · · ×
ΘT : θni = θnj , n ∈ NTi,j , i, j ∈ T } with each TransCo’s
feasible set defined as
Θi := θ i g θnm ≤ Knm , g θmn ≤ Kmn , {n, m} ∈ E i ;
θnm ≤ θnm ≤ θ̄nm , {n, m} ∈ E i ;
θn = 0, n ∈ NTi ∩ NS ; θn ∈ [−π, π], n ∈ NTi .
For later reference, the dual problem of Problem (P) is simply
min φ(λ).
(D)
λ
A. Agent Surplus Functions
The arguments of the maximizations in Eq. (4) represent
surplus functions of the agents. This follows from the fact
that the dual variables, λ, of the power balance equation
represent locational marginal prices. The surplus for DistCo
i ∈ D for a given demand profile (ei , si ) at price λ is equal
to the utility obtained from ei minus the cost of total demand
(sum of elastic and inelastic demand), defined as
X
X
i
ΨD
(ei , λ) :=
uin ein − λn ein −
λn sin .
i
n∈ND
e
i
n∈ND
s
The surplus of each GenCo i ∈ G is equal to the payment it
receives for producing power minus the generation cost,
X
ΨGi (pi , λ) :=
λn pin − cin pin .
n∈NGi
TransCos receive a surplus for facilitating power flow across
the network. Congestion and losses in transmission lines
creates different valuations for power across the network
(represented by LMPs) and results in a discrepancy between
the payments received from DistCos and the payments made
to GenCos. This creates a surplus (possibly negative) for
transmitting power from GenCos to DistCos, termed the
merchandizing surplus. Under the convex DC approximation
(assumption 1), the total merchandizing surplus (argument of
the last maximization term in Eq. (4)) can be shown to be
X
X
ΨT (θ, λ) = −
λn fn (θ) = −
λn g(θnm )
n∈N
~
(n,m)∈E
=
1 X
(λm − λn )ḡ(θnm ) − (λn + λm )g̃(θnm ) (5)
2
~
(n,m)∈E
where E~ is the directed edge-set and is defined as the set that
contains the pair (n, m) and (m, n) for
every edge {n,
m} ∈
E. The quantity λm − λn ḡ θnm − λn + λm g̃ θnm
is the merchandizing surplus for enabling flow between
buses n and m at the price
vector λ. Notice that the first
term, λm − λn ḡ θnm , is the familiar expression for the
merchandizing surplus under the DC approximation [17].
The second term, − λn + λm g̃ θnm , arises from the fact
that we are considering losses in our model (see assump. 1).
Since the ownership of transmission lines is partitioned
among TransCos we can separate the total merchandizing
surplus into each TransCo’s merchandizing surplus as
1 X
(λm − λn )ḡ(θnm )
ΨTi (θ i , λ) =
2
i
~
(n,m)∈E
− (λn + λm )g̃(θnm ) . (6)
We show that, via a message exchange process described in
Section V-A.3, TransCos communicate to obtain the angle
profile which maximizes ΨT (θ, λ) over θ ∈ Θ at price λ.
B. Competitive Equilibria in Energy Markets
parallel to obtain the optimal profiles for the current price
λt , denoted by {ei (λt )}i∈D and {pi (λt )}i∈G , respectively.
TransCos partake in a message exchange process (due to coupling of merchandizing surplus functions) in order to obtain
the operating point which maximizes the total merchandizing
surplus at the current price, denoted by θ(λt ). The MO uses
these maximizers to update the price in such a way as to
enforce feasibility (condition (i) of Def. 1). A block diagram
outlining the method can be seen in Fig. 2.
DistCos
D {ei ( t )}i2D
GenCos
MO
Price
Update
G {pi ( t )}i2G
t
t+1
t
TransCos
t+1
T
✓(
t
)
t
x( )
We can now define the concept of a competitive equilibrium in the context of our energy market model. The
definition builds upon the one found in [8].
Definition 1 (Competitive Equilibrium): A
competitive
equilibrium is defined as the tuple ({êi }i∈D , {p̂i }i∈G , θ̂, λ̂)
such that
(i)
fn (θ̂) − p̂n + ên + sn < ε for all n ∈ N , ε > 0
i
(ei , λ̂) s.t. ei ∈ Ei , ∀ i ∈ D
(ii) êi (λ̂) maximizes ΨD
p̂i (λ̂) maximizes ΨGi (pi , λ̂) s.t. pi ∈ Pi , ∀ i ∈ G
θ̂ i (λ̂) maximizes ΨTi (θ i , λ̂) s.t. θ i ∈ Θi , ∀ i ∈ T
The above definition states that a competitive equilibrium
must not only satisfy the power balance equation (condition
(i)) but also result in maximum surplus for all DistCos,
GenCos, and TransCos (condition (ii)).
V. SOLUTION METHODOLOGY
Throughout the remainder of the paper we describe a
procedure in which the MO and agents interact in order
to obtain a globally optimal solution x∗ to the nonconvex
primal problem (P). The procedure is based on the dual
decomposition method; an iterative method that first involves
the evaluation of the dual function for a given set of
dual variables (prices), followed by an update of the dual
variables.
In the context of the electricity market model in this
paper, the evaluation of the dual function is performed in
a distributed fashion by the agents. In fact, maximization of
surpluses by the agents corresponds exactly to the evaluation
of the dual function. DistCos and GenCos maximize in
Fig. 2. The pricing process: Given the current price vector λt , DistCo’s
and GenCo’s update the respective components of the consumption profiles
{ei (λt )}i∈D and generation profile {pi (λt )}i∈G , in parallel. TransCo’s
participate in a message exchange process to reach an angle profile agreement θ(λt ). The MO then updates the price to λt+1 using the responses
x(λt ) = ({ei (λt )}i∈D , {pi (λt )}i∈G , θ(λt )) (outlined in Section V-B).
A. Price Response - Dual Function Evaluation
The first step of the iterative process involves evaluation
of the dual function for a given current price vector λt . This
is achieved via the following agent surplus maximizations.
1) DistCo Optimizations: Each DistCo, i ∈ D, wishes to
specify the elastic demand level ei in order to maximize its
surplus from buying power (both elastic and inelastic) at the
current price λt . Each DistCo i ∈ D solves
i
ei (λt ) = argmax ΨD
(ei , λt )
ei ∈Ei
(PDi )
By assumption 3, each uin is strictly concave and therefore
the maximizer ei (λt ) of Problem (PDi ) is unique for each i.
2) GenCo Optimizations: Each GenCo, i ∈ G, wishes
to specify the injection levels pi in order to maximize its
surplus from selling power at λt . Each GenCo i ∈ G solves
pi (λt ) = argmax ΨGi (pi , λt )
pi ∈Pi
(PGi )
Again, by assumption 3, the maximizer pi (λt ) of Problem
(PGi ) is unique for each i ∈ G.
3) TransCo Optimizations: Each TransCo, i ∈ T , aims
to specify their voltage angle profile θ i ∈ Θi such that the
induced flows maximize their merchandizing surplus at the
current price, ΨTi (θ i , λt ). Doing so is complicated by the
fact that there exist buses that are shared between one or
more TransCos, that is, NTi,j 6= ∅ for neighboring i, j ∈ T .
The presence of these shared buses creates coupling between
the merchandizing surplus functions of distinct TransCos.
As a result, all neighboring TransCos must negotiate the
angle value of their shared buses. Arriving at a systemwide agreement for the shared buses, with each TransCo
maximizing their own merchandizing surplus, results in a
maximization of the total merchandizing surplus (achieving
the value of the last term in Eq. (4)). The angle profile
agreement is achieved via a message exchange process that is
based on the ADMM algorithm [18] in which neighboring
TransCos iteratively exchange the voltage angle values of
their shared buses.
To make use of the ADMM algorithm, it is necessary to
write the problem of maximizing the total merchandizing
surplus, maxθ∈Θ ΨT (θ, λ), as the equivalent problem
X
t
max
Ψ
(θ,
λ
)
=
ΨTi (θ i , λt )
(PT )
T
i
{θ }i∈T ,z
subject to
i∈T
i
i
θ ∈ Θ ,i ∈ T
θ i − zi = 0, i ∈ T
where z ∈ RN is the global variable representing the systemwide angle profile θ and zi = {zn }n∈NTi is the relevant
component of z corresponding to TransCo i’s angle profile.
We associate a set of dual variables, yi = {yn }n∈NTi , with
each of the consensus constraints θ i − zi = 0, i ∈ T . The
consensus constraints enforce the angle profiles of TransCos
to agree. Defining primal and dual residual norms [18] as
r(k) := θ 1,(k) − z1,(k) , . . . , θ T,(k) − zT,(k)
(7)
(k)
1,(k)
1,(k−1)
T,(k)
T,(k−1)
s := −ρ z
−z
,...,z
−z
(8)
the TransCo message exchange process is given by Alg. 1.
Algorithm 1: TransCo Message Exchange Process
Initialize k = 0, yi,(0) = 0 for all i ∈ T , z(0) = 0, ρ > 0;
while ¬(||r(k) ||2 < εprimal and ||s(k) ||2 < εdual ) do
for i ∈ T do (parallel optimization and broadcast)
TransCo i solves:
θ i,(k+1) (λt ) = argmax ΨTi (θ i , λt )
2
i,(k+1)
n
Lemma 1 and the convergence result of the ADMM in [18]
(p.17) lead to the following corollary.
Corollary 1: Alg. 1 generates iterates {z(k) } that converge to the unique solution θ(λt ) of maxθ∈Θ ΨT (θ, λt ).
All of the maximizers for the current price λt are then
broadcast to the MO, as in Fig. 2, and the price is updated.
B. Price Update
The MO receives all of the maximizers from the agents,
x(λt ) = ({ei (λt )}i∈D , {pi (λt )}i∈G , θ(λt )), for the current
price λt and uses them to compute an updated price λt+1 .
The price is updated in such a way as to iteratively enforce
the power balance equation (see cond. (i) of Def. 1). Before
defining the price update, we state the following result.
Lemma 2: The Hessian of the Lagrangian is negative
definite, that is, ∇2xx L(x, λ) ≺ 0 for all x.
Proof: See appendix B.
As a consequence of the strong concavity of the Lagrangian,
the dual function, Eq. (4), is unique and its derivative exists.
The gradient of the dual function is (see Thm. 6.3.3 of [19])
∇λ φ(λ) = h(x(λ))>
(9)
where h(x) was defined at the beginning of Section IV. As
a result, solving the power balance equation is equivalent to
finding where the gradient of the dual function vanishes. The
price update follows a gradient descent method
λt+1 = λt − αt ∇λ φ(λt ) = λt − αt h(x(λt ))
= λt − αt f θ(λt ) − p(λt ) + e(λt ) + s
(10)
with step-size αt , net generationPprofile p(λt ) =
(p1 (λt ), . . . , pN (λt )), with pn (λt ) = i∈G pin (λt ), and net
elastic demand
e(λt ) = (e1 (λt ), . . . , eN (λt )), with
P profile
t
i
t
t
en (λ ) =
i∈D en (λ ). The injection term f θ(λ ) is
computed from the angle profile θ(λt ) (from Sec. V-A.3).
A. Convergence Properties of the Pricing Process
Broadcast {θn
}n∈N i,j to neighboring j ∈ T ;
T
for i ∈ T do (parallel average and dual variable update)
1 X j,(k+1)
(k+1)
Average: zn
=
θn
, ∀n ∈ NTi
|Tn | j∈T
Update: yi,(k+1) = yi,(k) + ρ(θ i,(k+1) − zi,(k+1) )
Update residuals: set r(k+1) , s(k+1) via Eq.’s (7), (8);
Update counter: k ← k + 1;
Lemma 1: The merchandizing surplus function ΨTi (θ i , λ)
is strongly concave in θ i for all i ∈ T .
Proof: See appendix A.
VI. ANALYSIS
θ i ∈Θi
>
ρ
− yi,(k) (θ i − zi,(k) )−
θ i − zi,(k)
2
By assumption 2, each TransCo i has a slack bus in its
set of buses NTi , and the following lemma holds.
We begin by showing that the dual function is Lipschitz
continuous (this follows from the fact that a function with a
bounded derivative is Lipschitz). By Eq. (9), the gradient of
>
the dual function satisfies
∇λ φ(λ) = h(x(λ)) . Noting that
kh(x(λ))k
= kf θ(λ) − p(λ) + e(λ) + sk and the fact that
f θ(λ) , p(λ), and e(λ) are all bounded, there exists some
M < ∞ such that k∇λ φ(λ)k = kh(x(λ))k ≤ M . Thus the
dual function is Lipschitz continuous.
The dual function φ(λ) is a convex function of λ. It can
be shown through standard arguments that, for a sufficiently
small step-size, gradient descent applied to a convex function
generates iterates satisfying
Pt
kλ0 − λ∗ k2 + s=0 αs2 k∇φ(λs )k2
φ(λt ) − φ∗ ≤
Pt
2 s=0 αs
where φ∗ denotes a minimum of φ. In order
P∞ to 2ensure
convergence,
one
must
choose
α
such
that
t
s=0 αs < ∞
P∞
s
and
α
=
∞.
Noting
that
k∇φ(λ
)k
≤
M for all
s=0 s
s, we have φ(λt ) → φ∗ . Selecting a step-size of the form
αt = β/t, β > 0, ensures that the pricing process converges
to a minimizer λ∗ of the dual function φ(λ), solving the
dual problem (D).2 Since the dual problem is unconstrained,
∇φ(λ)|λ=λ∗ = 0, and, again by Eq. (9), h(x(λ∗ )) = 0.
B. Duality Gap
In this section, we demonstrate that the converged dual
solution λ∗ results in a zero duality gap with Problem (P).
Theorem 1: The pricing process described in Section V
generates a competitive eq. (x∗ , λ∗ ), where x∗ = x(λ∗ ) is
a globally optimal solution to Problem (P).
Proof: Consider Problem (Q), defined as
max {G(ω) : r(ω) = (r1 (ω), . . . , rM (ω)) = 0}.
ω∈Ω⊆RW
(Q)
Also, consider the following definition.
Definition 2 (Global ω-max, ν-min saddle point [20]):
A point (ω̂, ν̂) is a global ω-max, ν-min saddle point for
the Lagrangian M(ω, ν) = G(ω) − ν > r(ω) if and only if
M(ω, ν̂) ≤ M(ω̂, ν̂) ≤ M(ω̂, ν) ∀ ω ∈ Ω, ν ∈ RM .
The proof proceeds in two steps: (i) We prove a general result
demonstrating that if (ω̂, ν̂) is a global ω-max, ν-min saddle
point for the Lagrangian M then ω̂ is the global optimum
for the problem (Q) (similar to the proof found in [20]); (ii)
We show that the pricing process generates a global x-max,
λ-min saddle point for the Lagrangian L of Problem (P).
Part (i): Assuming that (ω̂, ν̂) is a global ω-max, ν-min
saddle point, we have M(ω̂, ν̂) ≤ M(ω̂, ν) for all ν. Thus
G(ω̂) −
M
X
m=1
ν̂m rm (ω̂) ≤ G(ω̂) −
M
X
νm rm (ω̂).
(11)
m=1
Let there exist an index m0 such that rm0 (ω̂) > 0, then we
can choose νm0 0 such that Eq. (11) is violated. Similarly,
let there exist an index m00 such that rm00 (ω̂) < 0, we can
violate Eq. (11) by choosing νm00 0. Thus rm (ω̂) = 0 for
all m and therefore ω̂ is feasible for Problem (Q).
Since (ω̂, ν̂) is a global ω-max, ν-min saddle point, we
also have
for all ω ∈ Ω. Thus
PMM(ω, ν̂) ≤ M(ω̂, ν̂) P
M
G(ω)− m=1 ν̂m rm (ω) ≤ G(ω̂)− m=1 ν̂m rm (ω̂). Since
r(ω) = 0 for every feasible ω, we have that G(ω) ≤ G(ω̂)
everywhere on {ω|r(ω) = 0, ω ∈ Ω} and thus ω̂ is optimal
for Problem (Q).
Part (ii): Let x∗ = x(λ∗ ) = argmaxx∈X L(x, λ∗ ), where
λ∗ is the converged price vector obtained from the pricing
2 Prices must satisfy assumption 4 at each iteration t in order to ensure
the TransCo subproblems are convex. This can simply be achieved through
choice of a sufficiently positive λ0 .
process (Eq. 10). The profile x∗ is returned from the agents
when provided with the price λ∗ (via the optimizations
in Section V-A). By assumption 3 and Corollary 1, x∗ is
unique. Notice that (x∗ , λ∗ ) is a competitive equilibrium
by Def. 1. Also, due to the concavity of L (Lemma 2),
φ(λ∗ ) = L(x∗ , λ∗ ) ≥ L(x, λ∗ ) for all x ∈ X. We know that
∇φ(λ)|λ=λ∗ = 0 and thus, by Eq. (9), h(x(λ∗ )) = h(x∗ ) =
0. Consequently, L(x∗ , λ∗ ) = J(x∗ ) − (λ∗ )> h(x∗ ) =
J(x∗ ) = J(x∗ ) − λ> h(x∗ ) = L(x∗ , λ) for all λ. In
summary L(x, λ∗ ) ≤ L(x∗ , λ∗ ) = L(x∗ , λ) for all x ∈
X, λ and thus (x∗ , λ∗ ) is a global x-max, λ-min saddle
point for the Lagrangian L(x, λ).
From parts (i) and (ii), we conclude that the pricing
process generates the pair (x∗ , λ∗ ) where x∗ is a globally
optimal solution to Problem (P).
In summary, the pricing process generates the competitive
equilibrium (x∗ , λ∗ ) which results in a globally optimal
solution x∗ to the (nonconvex) social welfare maximization
problem (P). Consequently, x∗ is a Pareto efficient outcome.
VII. NUMERICAL EXAMPLE
We demonstrate the performance of the pricing process
on a modified version of the IEEE 14 bus test system.
The ownership of generators in the modified system is split
among three GenCos with p1 = (p11 , p12 ), p2 = (p23 , p26 ), and
p3 = (p38 ). The network also consists of seven DistCos with
e1 = (e12 , e13 ), e2 = (e23 , e24 ), e3 = (e35 ), e4 = (e46 , e411 , e412 ),
e5 = (e59 , e510 ), e6 = (e612 , e613 ), e7 = (e714 ) and inelastic
demands (in MW) s12 = 15, s35 = 10, s412 = 15, s510 = 10,
s614 = 15. The
ownership of lines is split among two
TransCos, E 1 = {1,2},{1,5},{2,3},{2,4},{2,5},{3,4},{4,
5},{4,7},{5,6} , E 2 = {4,9},{6,11},{6,12},{6,13},{7,8},
{7,9},{9,10},{9,14},{10,11},{12,13},{13,14} with slack
bus NS = {6}. Parameters for the TransCo message
exchange process (Alg. 1) are ρ = 0.21, εprimal = 5×10-5 ,
εdual = 5×10-6 . Fig. 3 demonstrates the convergence of the
process.
It is evident from Fig. 3(b) that the pricing process
generates a solution where the power balance equation is
satisfied (condition (i) of Def. 1). At the corresponding
prices, agents report their surplus-maximizing responses, satisfying condition (ii) of Def. 1. By Theorem 1, the resulting
competitive equilibrium is Pareto efficient.
VIII. CONCLUSIONS & FUTURE WORK
We have developed an electricity market model and an
associated decentralized market mechanism that, under natural assumptions, ensures convergence to a Pareto efficient
market (competitive) equilibrium. The market model includes
multiple DistCos, GenCos, and (cooperative) TransCos all
of which are assumed to be surplus-maximizing given the
current set of LMPs. A market operator updates LMPs via a
gradient method in order to achieve an operating point that
satisfies the power balance equations and consequently clears
the market.
t=1
Voltage angles
2
t=2
t=3
t=4
t = 14
t = 24
1
0
-1
(corresponding to GenCo units), respectively. By assumption
3, we have U, C ≺ 0. Similar to the proof of Lemma 1,
matrix M can be shown to be an irreducibly diagonally
dominant matrix with a negative diagonal and thus, again by
Prop. 2.2.20 of [22], M ≺ 0. Since ∇2xx L is block-diagonal
with each block negative definite, we conclude ∇2xx L ≺ 0.
R EFERENCES
-2
1
123 1
245 1
271 1
103
1 101
1
98
Algorithm 1 cycle iterations, k
(a)
Bus power mismatch
40
30
20
10
0
-10
-20
-30
-40
1
2
3
4
5
6
7
8
9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
Pricing process iterations, t
(b)
Fig. 3. Convergence of pricing process: (a) TransCos reach an angle
agreement for each price vector λt via Alg. 1 (each negotiation cycle
corresponds to a price vector); (b) The power mismatch at each bus n,
hn (x(λt )) = fn (θ(λt )) − pn (λt ) + en (λt ) + sn , converges to zero.
APPENDIX
A. Proof of Lemma 1
i
λ), P
see Eq. (6), with
Let v = θ i . The Hessian of
ΨT (v,
2
i
respect to v is ∇vv ΨT v, λ = − 12 ∇2vv
~i (λn +
(n,m)∈E
λm )g̃(vnm ) , where the first order terms do not enter into
the expression. Define ι = NTi as the (ordered) set of bus
indices of TransCo i and define
X
−
(λιj + λιl )Gιj ιl if j = k
{j,l}∈E i
Ajk =
−(λιj + λιk )Gιj ιk
if {j, k} ∈ E i
0
if {j, k} 6∈ E i
for each j, k = 1, . . . , |ι| = |NTi |. By assumption 2, there
exists an index s ∈ ι that corresponds to a slack bus. The
Hessian ∇2 ΨTi is defined as matrix A with the sth row
and column removed. Consequently, ∇2 ΨTi belongs to the
class of irreducibly diagonally dominant matrices, known to
be non-singular (see Theorem 6.2.27 of [21]). To see this,
note that the Hessian is diagonally dominant for all rows.
Additionally, it is strictly diagonally dominant in rows that
correspond to buses that are immediately connected to a slack
bus. By assumption 4, the diagonal elements of the Hessian
are negative and by Prop. 2.2.20 of [22], ∇2 ΨTi ≺ 0.
B. Proof of Lemma 2
The Hessian of the Lagrangian, denoted P
by ∇2xx L, is a
i
square,
block-diagonal
matrix
of
dimension
i∈D |ND e | +
P
i
|N
|
+
|N
\
N
|.
It
consists
of
three
blocks,
U,
C,
S
G
i∈G
and M, where U and C are diagonal matrices consisting of
elements (uin )00 (corresponding to DistCo units) and −(cin )00
[1] R. D. Christie and A. Bose, “Load frequency control issues in
power system operations after deregulation,” IEEE Trans. Power Syst.,
vol. 11, no. 3, pp. 1191–1200, 1996.
[2] A. Mas-Colell, M. Whinston, and J. Green, Microeconomic Theory.
Oxford University Press, 1995.
[3] D. Kirschen and G. Strbac, Fundamentals of Power System Economics.
John Wiley & Sons, 2004.
[4] H. P. Chao and S. Peck, “A market mechanism for electric power
transmission,” J. Regul. Econ., vol. 10, pp. 25–59, 1996.
[5] S. Stoft, Power System Economics: Designing Markets for Electricity,
ser. IEEE Press. Wiley, 2002.
[6] D. Papadaskalopoulos and G. Strbac, “Decentralized participation of
flexible demand in electricity markets - Part I: Market mechanism,”
IEEE Trans. Power Syst., vol. 28, no. 4, pp. 3658–3666, 2013.
[7] D. P. Bertsekas, Nonlinear programming. Athena scientific, 1999.
[8] A. Motto, F. Galiana, A. Conejo, and M. Huneault, “On Walrasian
equilibrium for pool-based electricity markets,” IEEE Trans. Power
Syst., vol. 17, no. 3, pp. 774–781, 2002.
[9] F. D. Galiana, A. L. Motto, A. J. Conejo, and M. Huneault, “Decentralized nodal-price self-dispatch and unit commitment,” in The
Next Generation of Electric Power Unit Commitment Models, ser.
International Series in Operations Research and Management Science,
B. F. Hobbs, M. H. Rothkopf, R. P. O’Neill, and H. P. Chao, Eds.
Springer US, 2002, vol. 36, pp. 271–292.
[10] J. Lavaei and S. Sojoudi, “Competitive equilibria in electricity markets
with nonlinearities,” in ACC, 2012, pp. 3081–3088.
[11] F. Zhuang and F. D. Galiana, “Towards a more rigorous and practical
unit commitment by Lagrangian relaxation,” IEEE Trans. Power Syst.,
vol. 3, no. 2, pp. 763–773, 1988.
[12] J. F. Bard, “Short-term scheduling of thermal-electric generators using
Lagrangian relaxation,” Oper. Res., vol. 36, no. 5, pp. 756–766, 1988.
[13] W. Ongsakul and N. Petcharaks, “Unit commitment by enhanced
adaptive Lagrangian relaxation,” IEEE Trans. Power Syst., vol. 19,
no. 1, pp. 620–628, 2004.
[14] D. T. Nguyen, M. Negnevitsky, and M. De Groot, “Walrasian market
clearing for demand response exchange,” IEEE Trans. Power Syst.,
vol. 27, no. 1, pp. 535–544, 2012.
[15] O. Elgerd, Electric Energy Systems Theory: An Introduction.
McGraw-Hill, 1973.
[16] B. Stott, O. Alsac, and A. Monticelli, “Security analysis and optimization,” Proc. IEEE, vol. 75, no. 12, pp. 1623–1644, 1987.
[17] F. Wu, P. Varaiya, P. Spiller, and S. Oren, “Folk theorems on
transmission access: Proofs and counterexamples,” J. Regul. Econ.,
vol. 10, no. 1, pp. 5–23, 1996.
[18] S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein, “Distributed
optimization and statistical learning via the alternating direction
method of multipliers,” Foundations and Trends R in Machine Learning, vol. 3, no. 1, pp. 1–122, 2011.
[19] M. S. Bazaraa, H. D. Sherali, and C. M. Shetty, Nonlinear programming: theory and algorithms. John Wiley & Sons, 2013.
[20] P. B. Morgan, An Explanation of Constrained Optimization for
Economists. University of Toronto Press, 2015.
[21] R. Horn and C. Johnson, Matrix analysis. Cambridge University
Press, 1985.
[22] R. Cottle, J. Pang, and R. Stone, The Linear Complementarity Problem, ser. Classics in Applied Mathematics. SIAM, Philadelphia, PA,
1992.
| 3cs.SY |
In-Datacenter Performance Analysis of a Tensor Processing UnitTM
Norman P. Jouppi, Cliff Young, Nishant Patil, David Patterson, Gaurav Agrawal, Raminder Bajwa, Sarah Bates,
Suresh Bhatia, Nan Boden, Al Borchers, Rick Boyle, Pierre-luc Cantin, Clifford Chao, Chris Clark, Jeremy Coriell,
Mike Daley, Matt Dau, Jeffrey Dean, Ben Gelb, Tara Vazir Ghaemmaghami, Rajendra Gottipati, William Gulland,
Robert Hagmann, C. Richard Ho, Doug Hogberg, John Hu, Robert Hundt, Dan Hurt, Julian Ibarz, Aaron Jaffey, Alek
Jaworski, Alexander Kaplan, Harshit Khaitan, Daniel Killebrew, Andy Koch, Naveen Kumar, Steve Lacy, James Laudon,
James Law, Diemthu Le, Chris Leary, Zhuyuan Liu, Kyle Lucke, Alan Lundin, Gordon MacKean, Adriana Maggiore,
Maire Mahony, Kieran Miller, Rahul Nagarajan, Ravi Narayanaswami, Ray Ni, Kathy Nix, Thomas Norrie, Mark Omernick,
Narayana Penukonda, Andy Phelps, Jonathan Ross, Matt Ross, Amir Salek, Emad Samadiani, Chris Severn, Gregory Sizikov,
Matthew Snelham, Jed Souter, Dan Steinberg, Andy Swing, Mercedes Tan, Gregory Thorson, Bo Tian, Horia Toma,
Erick Tuttle, Vijay Vasudevan, Richard Walter, Walter Wang, Eric Wilcox, and Doe Hyun Yoon
Google, Inc., Mountain View, CA USA
Email: {jouppi, cliffy, nishantpatil, davidpatterson} @google.com
To appear at the 44th International Symposium on Computer Architecture (ISCA), Toronto, Canada, June 26, 2017.
Abstract
Many architects believe that major improvements in cost-energy-performance must now come from domain-specific
hardware. This paper evaluates a custom ASIC—called a Tensor Processing
Unit (TPU)— deployed in datacenters
since 2015 that accelerates the inference phase of neural networks (NN). The heart of the TPU is a 65,536 8-bit MAC
matrix multiply unit that offers a peak throughput of 92 TeraOps/second (TOPS) and a large (28 MiB)
software-managed on-chip memory. The TPU’s deterministic execution model is a better match to the 99th-percentile
response-time requirement of our NN applications than are the time-varying optimizations of CPUs and GPUs
(caches, out-of-order execution, multithreading, multiprocessing, prefetching, …) that help average throughput more
than guaranteed latency. The lack of such features helps explain why, despite having myriad MACs and a big
memory, the TPU is relatively small and low power. We compare the TPU to a server-class Intel Haswell CPU and an
Nvidia K80 GPU, which are contemporaries deployed in the same datacenters. Our workload, written in the high-level
TensorFlow framework, uses production NN applications (MLPs, CNNs, and LSTMs) that represent 95% of our
datacenters’ NN inference demand. Despite low utilization for some applications, the TPU is on average about 15X 30X faster than its contemporary GPU or CPU, with TOPS/Watt about 30X - 80X higher. Moreover, using the GPU’s
GDDR5 memory in the TPU would triple achieved TOPS and raise TOPS/Watt to nearly 70X the GPU and 200X the
CPU.
Index terms–DNN, MLP, CNN, RNN, LSTM, neural network, domain-specific architecture, accelerator
1. Introduction to Neural Networks
The synergy between the large data sets in the cloud and the numerous computers that power it has enabled a renaissance in
machine learning. In particular, deep neural networks (DNNs) have led to breakthroughs such as reducing word error rates in
speech recognition by 30% over traditional approaches, which was the biggest gain in 20 years [Dea16]; cutting the error rate
in an image recognition competition since 2011 from 26% to 3.5% [Kri12] [Sze15] [He16]; and beating a human champion at
Go [Sil16]. Unlike some hardware targets, DNNs are applicable to a wide range of problems, so we can reuse a DNN-specific
ASIC for solutions in speech, vision, language, translation, search ranking, and many more.
Neural networks (NN) target brain-like functionality and are based on a simple artificial neuron: a nonlinear function
(such as max(0, value)) of a weighted sum of the inputs. These artificial neurons are collected into layers, with the
outputs of one layer becoming the inputs of the next one in the sequence. The “deep” part of DNN comes from going beyond
a few layers, as the large data sets in the cloud allowed more accurate models to be built by using extra and larger layers to
capture higher levels of patterns or concepts, and GPUs provided enough computing to develop them.
The
two phases of NN are called
training (or learning) and inference (or prediction), and they refer to development
versus production. The developer chooses the number of layers and the type of NN, and training determines the weights.
Virtually all training today is in floating point, which is one reason GPUs have been so popular. A step called quantization
transforms floating-point numbers into narrow integers—often just 8 bits—which are usually good enough for inference.
Eight-bit integer multiplies can be 6X less energy and 6X less area than IEEE 754 16-bit floating-point multiplies, and the
1
advantage for integer addition is 13X in energy and 38X in area [Dal16].
Three kinds of NNs are popular today:
1. Multi-Layer Perceptrons (MLP): Each new layer is a set of nonlinear functions of weighted sum of all outputs (fully
connected) from a prior one, which reuses the weights.
2. Convolutional Neural Networks (CNN): Each ensuing layer is a set of of nonlinear functions of weighted sums of
spatially nearby subsets of outputs from the prior layer, which also reuses the weights.
3. Recurrent Neural Networks (RNN): Each subsequent layer is a collection of nonlinear functions of weighted sums of
outputs and the previous state. The most popular RNN is Long Short-Term Memory (LSTM). The art of the LSTM is
in deciding what to forget and what to pass on as state to the next layer. The weights are reused across time steps.
Table 1 shows two examples of each of the three types of NNs—which represent 95% of NN inference workload in our
datacenters—that we use as benchmarks. Typically written in TensorFlow [Aba16], they are surprisingly short: just 100 to
1500 lines of code. Our benchmarks are small pieces of larger applications that run on the host server, which can be thousands
to millions of lines of C++ code. The applications are typically user-facing, which leads to rigid response-time limits.
Each model needs between 5M and 100M weights (9th column of Table 1), which can take a lot of time and energy to
access. To amortize
the access costs, the same weights are reused across a batch of independent examples during inference or
training, which improves performance.
This
paper describes and measures the Tensor Processing Unit (TPU) and compares its performance and power for
inference to its contemporary CPUs and GPUs. Here is a preview of the highlights:
● Inference apps usually emphasize response-time over throughput since they are often user-facing.
● Due to latency limits, the K80 GPU is underutilized for inference, and is just a little faster than a Haswell CPU.
● Despite having a much smaller and lower power chip, the TPU has 25 times as many MACs and 3.5 times as much
on-chip memory as the K80 GPU.
● The TPU is about 15X - 30X faster at inference than the K80 GPU and the Haswell CPU.
● Four of the six NN apps are memory-bandwidth limited on the TPU; if the TPU were revised to have the same
memory system as the K80 GPU, it would be about 30X - 50X faster than the GPU and CPU.
● The performance/Watt of the TPU is 30X - 80X that of contemporary products; the revised TPU with K80 memory
would be 70X - 200X better.
● While most architects have been accelerating CNNs, they represent just 5% of our datacenter workload.
Name
LOC
MLP0
MLP1
LSTM0
LSTM1
CNN0
CNN1
100
1000
1000
1500
1000
1000
FC
5
4
24
37
4
Layers
Nonlinear
TPU Ops / TPU Batch % of Deployed
Weights
function
Weight Byte
Size
TPUs in July 2016
Conv Vector Pool Total
5
ReLU
20M
200
200
61%
4
ReLU
5M
168
168
34
58 sigmoid, tanh
52M
64
64
29%
19
56 sigmoid, tanh
34M
96
96
16
16
ReLU
8M
2888
8
5%
72
13
89
ReLU
100M
1750
32
Table 1. Six NN applications (two per NN type) that represent 95% of the TPU’s workload. The columns are the NN name; the number of
lines of code; the types and number of layers in the NN (FC is fully connected, Conv is convolution, Vector is self-explanatory, Pool is
pooling, which does nonlinear downsizing on the TPU; and TPU application popularity in July 2016. One DNN is RankBrain [Cla15]; one
LSTM is a subset of GNM Translate [Wu16]; one CNN is Inception; and the other CNN is DeepMind AlphaGo [Sil16][Jou15].
2. TPU Origin, Architecture, and Implementation
Starting as early as 2006, we discussed deploying GPUs, FPGAs, or custom ASICs in our datacenters. We concluded that the
few applications that could run on special hardware could be done virtually for free using the excess capacity of our large
datacenters, and it’s hard to improve on free. The conversation changed in 2013 when a projection where people use voice
search for 3 minutes a day using speech recognition DNNs would require our datacenters to double to meet computation
demands, which would be very expensive to satisfy with conventional CPUs. Thus, we started a high-priority project to
quickly produce a custom ASIC for inference (and bought off-the-shelf GPUs for training). The goal was to improve
cost-performance by 10X over GPUs. Given this mandate, the TPU was designed, verified [Ste15], built, and deployed in
datacenters in just 15 months. (Space limits the amount and the level of detail on the TPU in this paper; see [Ros15a],
[Ros15b], [Ros15c], [Ros15d], [Tho15], and [You15] for more.)
Rather than be tightly integrated with a CPU, to reduce the chances of delaying deployment, the TPU was designed to be
a coprocessor on the PCIe I/O bus, allowing it to plug into existing servers just as a GPU does. Moreover, to simplify
hardware design and debugging, the host server sends TPU instructions for it to execute rather than fetching them itself.
2
Hence, the TPU is closer in spirit to an FPU (floating-point unit) coprocessor than it is to a GPU.
Figure 1. TPU Block Diagram. The main computation part is the
yellow Matrix Multiply unit in the upper right hand corner. Its inputs
are the blue Weight FIFO and the blue Unified Buffer (UB) and its
output is the blue Accumulators (Acc). The yellow Activation Unit
performs the nonlinear functions on the Acc, which go to the UB.
Figure 2. Floor Plan of TPU die. The shading follows Figure 1.
The light (blue) data buffers are 37% of the die, the light (yellow)
compute is 30%, the medium (green) I/O is 10%, and the dark
(red) control is just 2%. Control is much larger (and much more
difficult to design) in a CPU or GPU
The goal was to run whole inference models in the TPU to reduce interactions with the host CPU and to be flexible
enough to match the NN needs of 2015 and beyond, instead of just what was required for 2013 NNs. Figure 1 shows the block
diagram of the TPU.
The TPU instructions are sent from the host over the PCIe Gen3 x16 bus into an instruction buffer. The internal blocks
are typically connected together by 256-byte-wide paths. Starting in the upper-right corner, the Matrix Multiply Unit is the
heart of the TPU. It contains 256x256 MACs that can perform 8-bit multiply-and-adds on signed or unsigned integers. The
16-bit products are collected in the 4 MiB of 32-bit A
ccumulators below the matrix unit. The 4MiB represents 4096,
256-element, 32-bit accumulators. The matrix unit produces one 256-element partial sum per clock cycle. We picked 4096 by
first noting that the operations per byte need to reach peak performance (roofline knee in Section 4) is ~1350, so we rounded
that up to 2048 and then duplicated it so that the compiler could use double buffering while running at peak performance.
When using a mix of 8-bit weights and 16-bit activations (or vice versa), the Matrix Unit computes at half-speed, and it
computes at a quarter-speed when both are 16 bits. It reads and writes 256 values per clock cycle and can perform either a
matrix multiply or a convolution. The matrix unit holds one 64 KiB tile of weights plus one for double-buffering (to hide the
256 cycles it takes to shift a tile in). This unit is designed for dense matrices. Sparse architectural support was omitted for
time-to-deploy reasons. Sparsity will have high priority in future designs.
The weights for the matrix unit are staged through an on-chip Weight FIFO that reads from an off-chip 8 GiB DRAM
called Weight Memory (for inference, weights are read-only; 8 GiB supports many simultaneously active models). The weight
FIFO is four tiles deep. The intermediate results are held in the 24 MiB on-chip Unified Buffer, which can serve as inputs to
the Matrix Unit. A programmable DMA controller transfers data to or from CPU Host memory and the Unified Buffer.
Figure 2 shows the floor plan of the TPU die. The 24 MiB Unified Buffer is almost a third of the die and the Matrix
Multiply Unit is a quarter, so the datapath is nearly two-thirds of the die. The 24 MiB size was picked in part to match the
pitch of the Matrix Unit on the die and, given the short development schedule, in part to simplify the compiler (see Section 7).
Control is just 2%. Figure 3 shows the TPU on its printed circuit card, which inserts into existing servers like an SATA disk.
As instructions are sent over the relatively slow PCIe bus, TPU instructions follow the CISC tradition, including a repeat
field. The average clock cycles per instruction (CPI) of these CISC instructions is typically 10 to 20. It has about a dozen
instructions overall, but these five are the key ones:
1. Read_Host_Memory reads data from the CPU host memory into the Unified Buffer (UB).
2. Read_Weights reads weights from Weight Memory into the Weight FIFO as input to the Matrix Unit.
3. MatrixMultiply/Convolve causes the Matrix Unit to perform a matrix multiply or a convolution from the
Unified Buffer into the Accumulators. A matrix operation takes a variable-sized B*256 input, multiplies it by a
256x256 constant weight input, and produces a B*256 output, taking B pipelined cycles to complete.
3
4.
Activate performs the nonlinear function of the artificial neuron, with options for ReLU, Sigmoid, and so on. Its
inputs are the Accumulators, and its output is the Unified Buffer. It can also perform the pooling operations needed
for convolutions using the dedicated hardware on the die, as it is connected to nonlinear function logic.
5. Write_Host_Memory writes data from the Unified Buffer into the CPU host memory.
The other instructions are alternate host memory read/write, set configuration, two versions of synchronization, interrupt host,
debug-tag, nop, and halt. The CISC MatrixMultiply instruction is 12 bytes, of which 3 are Unified Buffer address; 2 are
accumulator address; 4 are length (sometimes 2 dimensions for convolutions); and the rest are opcode and flags.
The philosophy of the TPU microarchitecture is to keep the matrix unit busy. It uses a 4-stage pipeline for these CISC
instructions, where each instruction executes in a separate stage. The plan was to hide the execution of the other instructions
by overlapping their execution with the MatrixMultiply instruction. Toward that end, the Read_Weights instruction
follows the decoupled-access/execute philosophy [Smi82], in that it can complete after sending its address but before the
weight is fetched from Weight Memory. The matrix unit will stall if the input activation or weight data is not ready.
We don’t have clean pipeline overlap diagrams, because our CISC instructions can occupy a station for thousands of
clock cycles, unlike the traditional RISC pipeline with one clock cycle per stage. Interesting cases occur when the activations
for one network layer must complete before the matrix multiplications of the next layer can begin; we see a “delay slot,”
where the matrix unit waits for explicit synchronization before safely reading from the Unified Buffer.
As reading a large SRAM uses much more power than arithmetic, the matrix unit uses systolic execution to save energy
by reducing reads and writes of the Unified Buffer [Kun80][Ram91][Ovt15b]. It relies on data from different directions
arriving at cells in an array at regular intervals where they are combined. Figure 4 shows that data flows in from the left, and
the weights are loaded from the top. A given 256-element multiply-accumulate operation moves through the matrix as a
diagonal wavefront. The weights are preloaded, and take effect with the advancing wave alongside the first data of a new
block. Control and data are pipelined to give the illusion that the 256 inputs are read at once, and that they instantly update
one location of each of 256 accumulators. From a correctness perspective, software is unaware of the systolic nature of the
matrix unit, but for performance, it does worry about the latency of the unit.
The TPU software stack had to be compatible with those developed for CPUs and GPUs so that applications could be
ported quickly to the TPU. The portion of the application run on the TPU is typically written in TensorFlow and is compiled
into an API that can run on GPUs or TPUs [Lar16]. Like GPUs, the TPU stack is split into a User Space Driver and a Kernel
Driver. The Kernel Driver is lightweight and handles only memory management and interrupts. It is designed for long-term
stability. The User Space driver changes frequently. It sets up and controls TPU execution, reformats data into TPU order,
translates API calls into TPU instructions, and turns them into an application binary. The User Space driver compiles a model
the first time it is evaluated, caching the program image and writing the weight image into the TPU’s weight memory; the
second and following evaluations run at full speed. The TPU runs most models completely from inputs to outputs,
maximizing the ratio of TPU compute time to I/O time. Computation is often done one layer at a time, with overlapped
execution allowing the matrix multiply unit to hide most non-critical-path operations.
Figure 3. TPU Printed Circuit Board. It can be inserted in the slot
for an SATA disk in a server, but the card uses PCIe Gen3 x16.
Figure 4. Systolic data flow of the Matrix Multiply Unit. Software
has the illusion that each 256B input is read at once, and they instantly
update one location of each of 256 accumulator RAMs.
4
Model
Die
Measured
TOPS/s
On-Chip
mm2 nm MHz TDP
GB/s
Dies
Memory
Idle Busy 8b FP
Haswell
662 22 2300 145W 41W 145W 2.6 1.3 51
E5-2699 v3
NVIDIA K80
561 28 560 150W 25W 98W -- 2.8 160
(2 dies/card)
TPU
NA* 28 700 75W 28W 40W 92
--
34
51 MiB
2
8 MiB
8
28 MiB
4
Benchmarked Servers
DRAM Size
256 GiB
256 GiB (host)
+ 12 GiB x 8
256 GiB (host)
+ 8 GiB x 4
TDP
Measured
Idle Busy
504W 159W 455W
1838W 357W 991W
861W 290W 384W
Table 2. Benchmarked servers use Haswell CPUs, K80 GPUs, and TPUs. Haswell has 18 cores, and the K80 has 13 SMX processors.
Figure 10 has measured power. The low-power TPU allows for better rack-level density than the high-power GPU. The 8 GiB DRAM per
TPU is Weight Memory. GPU Boost mode is not used (Sec. 8). SECDEC and no Boost mode reduce K80 bandwidth from 240 to 160. No
Boost mode and single die vs. dual die performance reduces K80 peak TOPS from 8.7 to 2.8. (*The TPU die is ≤ half the Haswell die size.)
3. CPU, GPU, and TPU Platforms
The six production applications in Table 1 are our workload for this paper. As mentioned above, these six are representative
of 95% of TPU use in our datacenters. Ironically, deploying and measuring popular small DNNs like AlexNet or VGG is
difficult on production machines. However, one of our CNNs derives from Inception V2, which is widely used.
The benchmark platforms are server-class computers that were available in 2015 when the TPUs were deployed. This
restriction meant that they must include at least SECDED protection of internal SRAM as well as external DRAM memory
like the TPU, which excludes some choices such as the Nvidia Maxwell GPU. For our company to purchase and deploy them,
they also had to be sensibly configured machines, and not awkward artifacts assembled solely to win benchmarks.
Table 2 lists our choices. The traditional CPU server is represented by an 18-core, dual-socket Haswell processor from
Intel. This platform is also the host server for GPUs or TPUs. Haswell was fabbed in an Intel 22nm process. Both the CPU
and GPU are very large dies: about 600 mm2!
The 2.3 GHz CPU clock rate doesn’t include Turbo mode because it seldom occurs in our datacenters for NN apps.
Haswell has different clock rates depending on whether programs use AVX instructions, which our NN apps often use. The
higher clock rate of Turbo mode (for programs that avoid AVX) occurs when they don’t use all their cores. Thus, another
reason Turbo mode is rare in our datacenters is that our apps typically do use all the cores, plus they can run other datacenter
jobs to fill any idle cores.
The GPU accelerator is the Nvidia K80. Each K80 card contains two dies and offers SECDED on internal memory and
DRAM. Nvidia states that the “K80 Accelerator dramatically lowers datacenter cost by delivering application performance
with fewer, more powerful servers”[Nvi16]. NN researchers frequently used K80s in 2015, and they were chosen for new
cloud-based GPU offerings as recently as September 2016 [Bar16].
Up to eight K80 dies can be installed in four cards on this server, which is the configuration we benchmark. It offers a
Boost mode to increase clock rate as high as 875 MHz. Turbo mode in Haswell is controlled by hardware and so can operate
in short bursts before the chip temperature rises significantly. However, Boost mode is under the control of a software driver
[Nvi15], and thus lasts at least hundreds of milliseconds. Hence, power and cooling would have to be provisioned for the K80
as if it were essentially always running in Boost mode, for otherwise the chips could get too hot. For this platform, enabling
Boost mode would force us to reduce the number of K80 cards, which would hurt total cost of ownership. Thus, Boost mode
is disabled. This restriction reduces peak bandwidth and TOPS (see Table 2 caption); Sec. 8 examines if it were enabled.
As the number of dies per benchmarked server varies between 2 to 8, we usually show results below normalized per die
(Figures 5-8, Figures 10-11, and Tables 3, 4, and 6), but we occasionally show whole systems (Figure 9). We hope this
distinction is clear.
4. Performance: Rooflines, Response-Time, and Throughput
To illustrate the performance of the six apps on the three processors, we adapt the Roofline Performance model from
high-performance computing (HPC) [Wil09]. This simple visual model is not perfect, yet it offers insights on the causes of
performance bottlenecks. The assumption behind the model is that applications don’t fit in on-chip caches, so they are either
computation-limited or memory bandwidth-limited. For HPC, the Y-axis is performance in floating-point operations per
second, thus the peak computation rate forms the “flat” part of the roofline. The X-axis is operational intensity, measured as
floating-point operations per DRAM byte accessed. Memory bandwidth is bytes per second, which turns into the “slanted”
part of the roofline since (FLOPS/sec)/ (FLOPS/Byte) = Bytes/sec. Without sufficient operational intensity, a program is
memory bandwidth-bound and lives under the slanted part of the roofline.
5
Figure 5. TPU (die) roofline. Its ridge point is far to the right at 1350 operations per byte of weight memory fetched.
The gap between the actual operations per second of an application and the ceiling directly above it shows the potential
benefit of further performance tuning while leaving operational intensity untouched; of course, optimizations that increase
operational intensity (such as cache blocking) may yield even greater benefit.
To use the Roofline model for the TPU, when NN applications are quantized, we first replace floating-point operations
with integer operations. As weights do not normally fit in on-chip memory for NN applications, the second change is to
redefine operational intensity to be integer operations per byte of weights read (see the tenth column of Table 1).
Figure 5 shows the Roofline model for a single TPU die on log-log scales. The TPU has a long “slanted” part of its
roofline, where operational intensity means that performance is limited by memory bandwidth rather than by peak compute.
Five of the six applications are happily bumping their heads against the ceiling: the MLPs and LSTMs are memory bound,
and CNNs are computation bound. CNN1, despite a very high operational intensity, is running at only 14.1 TOPS while
CNN0 runs at 86 TOPS.
Table 3 explains what happened with CNN1, based on the performance counters that give us partial visibility into TPU
operation. The TPU spends less than half of its cycles performing matrix operations for CNN1 (column 7, row 1). On each of
those active cycles, only about half of the 65,536 MACs hold useful weights because some layers in CNN1 have shallow
feature depths. About 35% of cycles are spent waiting for weights to load from memory into the matrix unit, which occurs
during the 4 fully connected layers that run at an operational intensity of just 32 (see the last Fallacy in Section 8). This leaves
roughly 19% of cycles not explained by the matrix-related counters. Because of overlapped execution on the TPU, we do not
have exact accounting for those cycles, but we can see that 23% of cycles have stalls for RAW dependences in the pipeline,
and 1% are spent stalled for input over the PCIe bus.
6
Application
Array active cycles
Useful MACs in 64K matrix (% peak)
Unused MACs
Weight stall cycles
Weight shift cycles
Non-matrix cycles
RAW stalls
Input data stalls
TeraOps/sec (92 Peak)
MLP0
12.7%
12.5%
0.3%
53.9%
15.9%
17.5%
3.3%
6.1%
12.3
MLP1 LSTM0 LSTM1 CNN0 CNN1 Mean Row
10.6%
8.2% 10.5% 78.2% 46.2% 28% 1
9.4%
8.2%
6.3% 78.2% 22.5% 23% 2
1.2%
0.0%
4.2% 0.0% 23.7%
5% 3
44.2%
58.1% 62.1% 0.0% 28.1% 43% 4
13.4%
15.8% 17.1% 0.0% 7.0% 12% 5
31.9%
17.9% 10.3% 21.8% 18.7% 20% 6
8.4%
14.6% 10.6% 3.5% 22.8% 11% 7
8.8%
5.1%
2.4% 3.4% 0.6%
4% 8
9.7
3.7
2.8
86.0 14.1 21.4
9
Table 3. Factors limiting TPU performance of the NN workload based on hardware performance counters. Rows 1, 4, 5, and 6 total 100%
and are based on measurements of activity of the matrix unit. Rows 2 and 3 further break down the fraction of 64K weights in the matrix
unit that hold useful weights on active cycles. Our counters cannot exactly explain the time when the matrix unit is idle in row 6; rows 7 and
8 show counters for two possible reasons, including RAW pipeline hazards and PCIe input stalls. Row 9 (TOPS) is based on measurements
of production code while the other rows are based on performance-counter measurements, so they are not perfectly consistent. Host server
overhead is excluded here. The MLPs and LSTMs are memory-bandwidth limited but CNNs are not. CNN1 results are explained in the text.
Figure 6. Intel Haswell CPU (die) roofline with its ridge point at 13 operations/byte, which is much further left than in Figure. 5.
LSTM0 and MLP1 are faster on Haswell than on the K80, but it is vice versa for the other DNNs.
Figure 7. NVIDIA K80 GPU die Roofline. The much higher memory bandwidth moves the ridge point to 9 operations per weight byte,
which is even further left than in Figure 6. The DNNs are much lower than their Roofline because of response time limits (see Table 4).
7
Figures 6 and 7 show rooflines for a single Haswell die and for a single K80 die. The six NN applications are generally
further below their ceilings than was the TPU in Figure 5. Response time is the reason. Many of these NN applications are
parts of end-user-facing services. Researchers have demonstrated that small increases in response time cause customers to use
a service less [Sch09]. Hence, while training may not have hard response time deadlines, inference usually does. That is,
inference prefers latency over throughput [Pat04].
Table 4 illustrates the impact of the 99th-percentile response time limit of 7 ms for MLP0 on Haswell and the K80
[Dea13], which was required by the application developer. (The inferences per second and 7 ms latency include the server
host time as well as the accelerator time.) They operate at 42% and 37%, respectively, of the highest throughput achievable
for MLP0 if the response time limit was relaxed. Thus, while CPUs and GPUs have potentially much higher throughput, it’s
wasted if they don’t meet the response time limit. These bounds affect the TPU as well, but at 80% in Table 4 it is operating
much closer to its highest MLP0 throughput. As compared to CPUs and GPUs, the single-threaded TPU has none of the
sophisticated microarchitectural features that consume transistors and energy to improve the average case but not the
99th-percentile case: no caches, branch prediction, out-of-order execution, multiprocessing, speculative prefetching, address
coalescing, multithreading, context switching, and so forth. Minimalism is a virtue of domain-specific processors.
Table 3 shows TPU performance, but it doesn’t account for host server time, which can be divided into running the host
share of the application and talking to the TPU. Table 5 lists the second part, but the first part is hard. Queueing theory shows
that long input queues raise throughput—by ensuring that the computer is never idle—but stretch response time. Thus, most
applications keep their input queues empty. Alas, we can’t measure when the TPU is idle since it is waiting for the CPU to do
its portion of the application or because the CPU is also idle due to an empty input queue.
Table 6 gives the bottom line of relative inference performance per die including the host server overhead for the two
accelerators versus the CPU. The next-to-last column shows the geometric mean of the relative performance for the six NN
applications, which suggests the K80 die is 1.1X the speed of a Haswell die, that the TPU die is 14.5 times as fast, and thus
the TPU die is 13.2 times as fast as the GPU die. Figure 8 shows their relative speeds visually.
Recall that architects use the geometric mean when they don’t know the actual mix of programs that will be run [Hen18].
For this study, however, we do know the mix (Table 1). The weighted mean in the last column of Table 6 using the actual mix
increases the GPU to 1.9X and the TPU to 29.2X, so the TPU die is now 15.3 times as fast as the GPU die.
Type Batch 99th% Response Inf/s (IPS)
CPU 16
7.2 ms
5,482
CPU 64
21.3 ms
13,194
GPU 16
6.7 ms
13,461
GPU 64
8.3 ms
36,465
TPU 200
7.0 ms
225,000
TPU 250
10.0 ms
280,000
% Max IPS
42%
100%
37%
100%
80%
100%
Table 4. 99-th% response time and per die throughput (IPS) for MLP0 as batch size varies for MLP0. The longest allowable latency is 7
ms. For the GPU and TPU, the maximum MLP0 throughput is limited by the host server overhead. Larger batch sizes increase throughput,
but as the text explains, their longer response times exceed the limit, so CPUs and GPUs must use less-efficient, smaller batch sizes (16 vs.
200).
MLP0 MLP1 LSTM0
21%
76%
11%
LSTM1
20%
CNN0
51%
CNN1
14%
Table 5. Time for host CPU to interact with the TPU expressed as percent of TPU execution time (from TPU performance counters). This
fraction is the time the CPU and TPU are communicating over the PCIe bus, not including the time the CPU is doing a portion of the
application but not interacting with the TPU. As the text explains, it’s hard for the TPU to measure if the CPU is idle or working on the app.
Type
GPU
TPU
Ratio
MLP0 MLP1 LSTM0 LSTM1 CNN0 CNN1 GM
2.5
0.3
0.4
1.2
1.6
2.7
1.1
41.0 18.5
3.5
1.2
40.3 71.0 14.5
16.7 60.0
8.0
1.0
25.4 26.3 13.2
WM
1.9
29.2
15.3
Table 6. K80 GPU die and TPU die performance relative to CPU for the NN workload. GM and WM are geometric and weighted mean
(using the actual mix of the six apps in Table 1). Relative performance for the GPU and TPU includes host server overhead. MLPs and
CNNs perform well on the TPU. Table 4 explains that the TPU can have larger batch sizes and still meet the time limits, increasing
operations per byte (Table 1) or, equivalently, reducing memory accesses per operation. Also, CNNs by their nature have greater weight
reuse and thus higher operations per byte. Thus, the lower memory bandwidth of the TPU doesn’t significantly hurt CNN performance.
8
Figure 8. Figures 5-7 combined into a single log-log graph. Stars are for the TPU, triangles are for the K80, and circles are for Haswell. All
TPU stars are at or above the other 2 rooflines.
Figure 9. Relative performance/Watt (TDP) of GPU server (blue bar) and TPU server (red bar) to CPU server, and TPU server to GPU
server (orange bar). TPU’ is an improved TPU (Sec. 7). The green bar shows its ratio to the CPU server and the lavender bar shows its
relation to the GPU server. Total includes host server power, but incremental doesn’t. GM and WM are the geometric and weighted means.
9
5. Cost-Performance, TCO, and Performance/Watt
When buying computers by the thousands, cost-performance trumps performance. The best cost metric in a datacenter is total
cost of ownership (TCO). The actual price we pay for thousands of chips depends on negotiations between the companies
involved. For business reasons, we can’t publish such price information or data that might let them be deduced. However,
power is correlated with TCO, and we can publish Watts per server, so we use performance/Watt as our proxy for
performance/TCO in this paper. In this section, we compare whole servers rather than single dies, which Table 2 lists in the
“Benchmarked Server” columns.
Figure 9 shows the geometric and weighted mean performance/Watt for the K80 GPU and TPU relative to the Haswell
CPU. We present two different calculations of performance/Watt. The first (“total”) includes the power consumed by the host
CPU server when calculating performance/Watt for the GPU and TPU. The second (“incremental”) subtracts the host CPU
server power from the GPU and TPU beforehand.
For total-performance/Watt, the K80 server is 1.2 - 2.1X Haswell. For incremental-performance/Watt, when Haswell
server power is omitted, the K80 server is 1.7- 2.9X.
The TPU server has 17 to 34 times better total-performance/Watt than Haswell, which makes the TPU server 14 to 16
times the performance/Watt of the K80 server. The relative incremental-performance/Watt—which was our company’s
justification for a custom ASIC—is 41 to 83 for the TPU, which lifts the TPU to 25 to 29 times the performance/Watt of the
GPU.
6. Energy Proportionality
Thermal Design Power (TDP) affects the cost of provisioning power, as you must supply sufficient power and cooling when
hardware is at full power. However, the cost of electricity is based on the average consumed as the workload varies during the
day. [Bar07] found that servers are 100% busy less than 10% of the time and advocated energy proportionality: servers
should consume power proportional to the amount of work performed. The estimate of power consumed in the prior section is
based on the fraction of the TDP that has been seen in our datacenters.
We measured performance and power as the offered workload utilization varies from 0% to 100%, collected in buckets of
10% delta of workload [Lan09]. Figure 10 shows server power divided by the number of dies per server for the three chips by
varying CNN0’s workload. We plot incremental (K80 and TPU) as well as total power (K80+Haswell/4 and
TPU+Haswell/2) for the GPU and TPU. Note that all were given the same batch sizes.
We see that the TPU has the lowest power—118W per die total (TPU+Haswell/2) and 40W per die incremental (TPU in
Fig. 10)— but it has poor energy proportionality: at 10% load, the TPU uses 88% of the power it uses at 100%. (The short
design schedule prevented inclusion of many energy-saving features.) Not surprisingly, Haswell is the best at energy
proportionality of the group: it uses 56% of the power at 10% load as it does at 100%. The K80 is closer to the CPU than the
TPU, using 66% of the full load power at 10% workload. LSTM1, which is not computation bound, performs similarly: at
10% load the CPU uses 47% of full power, the GPU uses 78%, and the TPU uses 94%.
What happens to the server power usage when running CNN0 if it becomes a host to accelerators? When the GPU and
TPU are at 100% load, the CPU server uses 52% of full power for the GPU and 69% for the TPU. (The CPU does more work
for the TPU because it is running so much faster than the GPU.) Consequently, the Haswell server plus four TPUs use <20%
additional power but run CNN0 80 times faster than the Haswell server alone (4 TPUs vs 2 CPUs).
7. Evaluation of Alternative TPU Designs
Like an FPU, the TPU coprocessor has a relatively easy microarchitecture to evaluate, so we created a performance model for
our six applications. Table 7 shows the differences between the model results and the hardware performance counters, which
average below 10%. We then modeled performance as we varied the memory bandwidth, the clock rate and number of
accumulators, and the matrix multiply unit size.
Figure 11 shows the mean performance sensitivity of TPU die as we scale these parameters over the range for 0.25x to
4x. It plots weighted means, but the geometric means look similar. In addition to evaluating the impact of only raising clock
rates (clock in Figure 11), we also plot a design (clock+) where the clock rate is increased and the number of accumulators is
correspondingly scaled so the compiler can keep more memory references in flight. Likewise, we plot matrix unit expansion if
we increase the number of accumulators with the square of the rise in one dimension (matrix+), since the number of
multipliers in the matrix grows in both dimensions, as well as just increasing the matrix unit alone (matrix).
MLP0
6.8%
MLP1
10.9%
LSTM0
7.7%
LSTM1
5.4%
CNN0
8.2%
CNN1
11.2%
Table 7. Difference in clock cycles between the TPU hardware performance counters and the TPU performance model. The average is 8%.
10
Figure 10. Watts/die for CNN0 as target platform utilization varies from 0% to 100%. The Total GPU and TPU power are the red and
orange lines and their Incremental power are the green and purple lines. (The blue line is power for the Haswell CPU, which by definition
is Total power.) A server has 2 CPUs and 8 GPUs or 4 TPUs, so we normalize power by dividing by 2, 8, and 4, respectively.
Figure 11. Weighted mean TPU performance as metrics scale from 0.25x to 4x: memory bandwidth, clock rate + accumulators, clock rate,
matrix unit dimension + accumulators, and matrix unit dimension. The weighted mean makes it hard see to contributions of individual
DNNs, but MLPs and LSTMs improve 3X with 4X memory bandwidth, but get nothing from a higher clock. For CNNs it’s vice versa; 2X
for 4X clock, but get little benefit from faster memory. A bigger matrix multiply unit doesn’t help any DNN.
11
First, increasing memory bandwidth (memory) has the biggest impact: performance improves 3X on average when
memory increases 4X. Second, clock rate has little benefit on average with or without more accumulators. The reason is the
MLPs and LSTMs are memory bound but only the CNNs are compute bound. While hard to see in Figure 11, since it shows
only the weighted mean of all six DNNs, increasing the clock rate by 4X has almost no impact on MLPs and LSTMs but
improves performance of CNNs by about 2X. Third, the average performance in Figure 11 slightly degrades when the matrix
unit expands from 256x256 to 512x512 for all apps, whether or not they get more accumulators. The issue is analogous to
internal fragmentation of large pages, only worse since it’s in two dimensions. Consider the 600x600 matrix used in LSTM1.
With a 256x256 matrix unit, it takes 9 steps to tile 600x600, for a total of 18 us of time. The larger 512x512 unit requires only
four steps, but each step takes four times longer, for 32 us of time. Our CISC instructions are long, so decode is insignificant
and does not hide the overhead of loading from the DRAM.
Table 8 shows the utilization of the 24 MiB Unified Buffer, which was initially sized to allow MLPs to run at batch sizes
up to 2048. We recently improved the storage allocator for the Unified Buffer, which reduces the memory needed for the
largest of the six applications to 14 MiB. For the first 18 months of deployment, the TPU used its full capacity while the new
allocator was being developed. Now the extra capacity adds margin for adopting bigger models.
We next used the performance model to evaluate a hypothetical TPU die (TPU’) that might have been designed in the
same process technology if we had more than 15 months. More aggressive logic synthesis and block design might have
increased the clock rate by 50%. Designing an interface circuit for GDDR5 memory, as in the K80, would improve Weight
Memory bandwidth by more than a factor of five, shifting its roofline ridge point from 1350 to 250. As Figure 11 shows,
increasing clock rate to 1050 MHz but not helping memory makes almost no change. If we left the clock at 700 MHz but used
GDDR5 for Weight Memory, the geometric mean increase jumps to an impressive 2.6 and the weighted mean to 3.9. Doing
both raises the geometric mean (2.9) but not the weighted mean, so TPU’ just has faster memory.
Figure 11 does not include host server time. We used Table 5 to calculate time for the host server interaction overhead for
the TPU. Adding that same extra time drops TPU’ means from 2.6 to 1.9 and 3.9 to 3.2. This change is both optimistic, since
it doesn’t include CPU time to run its share of the app, and pessimistic, as we likely would aggressively tune the host code
given a 3X faster TPU’.
Replacing just the DDR3 Weight Memory with the equivalent GDDR5 memory requires doubling the number of
memory channels to four. This improvement would expand die size by about 10%. However, higher memory bandwidth
reduces pressure on the Unified Buffer, so reducing the Unified Buffer to 14 MiB could gain back 10% in area. GDDR5
would also increase the TPU system power budget from 861 Watts to about 900 Watts, as there are 4 TPUs per server.
Figure 9 above shows the relative total-performance/Watt/die of TPU’ leaps to 31X - 86X over Haswell and 25X - 41X
over the K80. The incremental metric soars to an amazing 69X - 196X over Haswell and 42X - 68X over the K80.
MLP0
11.0
MLP1
2.3
LSTM0
4.8
LSTM1
4.5
CNN0
1.5
CNN1
13.9
Table 8. Maximum MiB of the 24 MiB Unified Buffer used per NN app. A 14 MiB Unified Buffer is sufficient for these apps.
8. Discussion
This section follows the fallacy and pitfall with rebuttal style of [Hen18].
● Fallacy: NN inference applications in datacenters value throughput as much as response time.
We were surprised that our developers had strong response-time demands, as some suggested in 2014 that batch sizes would
be large enough for the TPU to reach peak performance or that latency requirements wouldn’t be as tight. One driving
application was off-line image processing, and the intuition was that if interactive services also wanted TPUs, most of them
would just accumulate larger batches. Even the developers of one application in 2014 that cared about response time
(LSTM1) said the limit was 10 ms in 2014, but shrank it to 7 ms when they actually ported it to the TPU. The unexpected
desire for TPUs by many such services combined with the impact on and preference for low response time changed the
equation, with application writers often opting for reduced latency over waiting for bigger batches to accumulate. Fortunately,
the TPU has a simple and repeatable execution model to help meet the response-time targets of interactive services and such
high peak throughput that even small batch sizes result in higher performance than contemporary CPUs and GPUs.
● Fallacy: The K80 GPU architecture is a good match to NN inference.
GPUs have traditionally been seen as high-throughput architectures that rely on high-bandwidth DRAM and thousands of
threads to achieve their goals. This perspective helps explain why the K80 is only a little faster at inference than Haswell and
much slower than the TPU. Successors to the K80 will surely include optimizations to improve peak inference performance,
but given their throughput-oriented architectural approach, it may be more challenging for GPUs to meet the strict latency
limits. And as Section 7 shows, there is plenty of headroom to improve the TPU, so it’s not an easy target.
12
● Pitfall: Architects have neglected important NN tasks.
We are pleased by the attention that the architecture community is paying to NN: 15% of the papers at ISCA 2016 were on
hardware accelerators for NN [Alb16] [Che16a][Chi16][Han16][Kim16][LiK16][Liu16][Rea16] [Sha16]! Alas, all nine
papers looked at CNNs, and only two mentioned other NNs. CNNs are more complex than MLPs and prominent in NN
competitions [Rus15], which might explain their allure, but they are only about 5% of our datacenter NN workload. While
CNNs may be common in edge devices, the volume of convolutional models hasn’t yet caught up with MLPs and LSTMs in
the datacenter. We hope that architects try to accelerate MLPs and LSTMs with at least as much gusto.
● Pitfall: For NN hardware, Inferences Per Second (IPS) is an inaccurate summary performance metric.
Our results show that IPS is a poor overall performance summary for NN hardware, as it’s simply the inverse of the
complexity of the typical inference in the application (e.g., the number, size, and type of NN layers). For example, the TPU
runs the 4-layer MLP1 at 360,000 IPS but the 89-layer CNN1 at only 4,700 IPS, so TPU IPS vary by 75X! Thus, using IPS as
the single-speed summary is even more misleading for NN accelerators than MIPS or FLOPS are for regular processors
[Hen18], so IPS should be even more disparaged. To compare NN machines better, we need a benchmark suite written at a
high-level to port it to the wide variety of NN architectures. Fathom is a promising new attempt at such a benchmark suite
[Ado16].
● Fallacy: The K80 GPU results would be much better if Boost mode were enabled.
Setting aside the negative impact of K80 Boost mode on TCO (Section 3), we measured it on LSTM1. Boost mode increased
the clock rate by a factor of up to 1.6—from 560 to 875 MHz—which increased performance by 1.4X, but it also raised power
by 1.3X. The net gain in performance/Watt is 1.1X, and thus for LSTM1, boost mode would have a minor impact on our
energy-speed analysis.
● Fallacy: CPU and GPU results would be comparable to the TPU if we used them more efficiently or compared to
newer versions.
We originally had 8-bit results for just one DNN on the CPU, due to the significant work to use AVX2 integer support
efficiently. The benefit was ~3.5X. It was less confusing (and less space) to present all CPU results in floating point, rather
than having one exception, with its own roofline. If all DNNs had similar speedup, performance/Watt ratio would drop from
41-83X to 12-24X. The new 16-nm, 1.5GHz, 250W P40 datacenter GPU can perform 47 Tera 8-bit ops/sec, but was
unavailable in early 2015, so isn’t contemporary with our three platforms. We also can’t know the fraction of P40 peak
delivered within our rigid time bounds. If we compared newer chips, Section 7 shows that we could triple performance of the
28-nm, 0.7GHz, 40W TPU just by using the K80’s GDDR5 memory (at a cost of an additional 10W).
● Pitfall: Performance counters added as an afterthought for NN hardware.
The TPU has 106 performance counters, and if anything we would like a few more (see Table 3). The raison d'etre for NN
accelerators is performance, and it is way too early in their evolution to have good intuition about what is going on.
● Fallacy: After two years of software tuning, the only path left to increase TPU performance is hardware upgrades.
The performance of CNN1 on the TPU could improve if developers and compiler writers did more work to match CNN1 to
the TPU hardware. For example, developers could reorganize the applications to aggregate multiple short batches out of the
convolution layers into a single, deeper batch (from 32 to 128) for the four fully connected layers. Such a single layer would
improve utilization of the matrix unit (Table 3). As CNN1 currently runs more than 70 times faster on the TPU than the CPU,
the CNN1 developers are already very happy, so it’s not clear whether or when such optimizations would be performed.
9. Related Work
Two survey articles document that custom NN ASICs go back at least 25 years [Ien96][Asa02]. For example, CNAPS chips
contained a 64 SIMD array of 16-bit by 8-bit multipliers, and several CNAPS chips could be connected together with a
sequencer [Ham90]. The Synapse-1 system was based on a custom systolic multiply-accumulate chip called the MA-16,
which performed sixteen 16-bit multiplies at a time [Ram91]. The system concatenated several MA-16 chips together and had
custom hardware to do activation functions.
Twenty-five SPERT-II workstations, accelerated by the T0 custom ASIC, were deployed starting in 1995 to do both NN
training and inference for speech recognition [Asa98]. The 40-Mhz T0 added vector instructions to the MIPS instruction set
architecture. The eight-lane vector unit could produce up to sixteen 32-bit arithmetic results per clock cycle based on 8-bit and
16-bit inputs, making it 25 times faster at inference and 20 times faster at training than a SPARC-20 workstation. They found
that 16 bits were insufficient for training, so they used two 16-bit words instead, which doubled training time. To overcome
that drawback, they introduced “bunches” (batches) of 32 to 1000 data sets to reduce time spent updating weights, which
made it faster than training with one word but no batches.
The more recent DianNao family of NN architectures minimizes memory accesses both on the chip and to external
DRAM by having efficient architectural support for the memory access patterns that appear in NN applications [Keu16]
13
[Che16a]. All use 16-bit integer operations and all designs dove down to layout, but no chips were fabricated. The original
DianNao uses an array of 64 16-bit integer multiply-accumulate units with 44 KB of on-chip memory and is estimated to be 3
mm2 (65 nm), to run at 1 GHz, and to consume 0.5W [Che14a]. Most of this energy went to DRAM accesses for weights, so
one successor DaDianNao (“big computer”) includes eDRAM to keep 36 MiB of weights on chip [Che14b]. The goal was to
have enough memory in a multichip system to avoid external DRAM accesses. The follow-on PuDianNao (“general
computer”) is aimed at more traditional machine learning algorithms beyond DNNs, such as support vector machines [Liu15].
Another offshoot is ShiDianNao (“vision computer”) aimed at CNNs, which avoids DRAM accesses by connecting the
accelerator directly to the sensor [Du15].
The Convolution Engine is also focused on CNNs for image processing [Qad13]. This design deploys 64 10-bit
multiply-accumulator units and customizes a Tensilica processor estimated to run at 800 MHz in 45 nm. It is projected to be
8X to 15X more energy-area efficient than an SIMD processor, and within 2X to 3X of custom hardware designed just for a
specific kernel.
The Fathom benchmark paper seemingly reports results contradictory to ours, with the GPU running inference much
faster than the CPU [Ado16]. However, their CPU and GPU are not server-class, the CPU has only four cores, the
applications do not use the CPU’s AVX instructions, and there is no response-time cutoff (see Table 4) [Bro16].
Catapult is the most widely deployed example of using reconfigurability to support DNNs, which many have proposed
[Far09][Cha10][Far11][Pee13][Cav15][Zha15]. They chose FPGAs over GPUs to reduce power as well as the risk that
latency-sensitive applications wouldn’t map well to GPUs. FPGAs can also be re-purposed, such as for search, compression,
and network interface cards [Put15]. The TPU project actually began with FPGAs, but we abandoned them when we saw that
the FPGAs of that time were not competitive in performance compared to the GPUs of that time, and the TPU could be much
lower power than GPUs while being as fast or faster, giving it potentially significant benefits over both of FPGAs and GPUs.
Although first published in 2014 [Put14], Catapult is a TPU contemporary since it deployed 28-nm Stratix V FPGAs into
datacenters concurrently with the TPU in 2015. Catapult has a 200 MHz clock, 3,926 18-bit MACs, 5 MiB of on-chip
memory, 11 GB/s memory bandwidth, and uses 25 Watts. The TPU has a 700 MHz clock, 65,536 8-bit MACs, 28 MiB, 34
GB/s, and typically uses 40 Watts. A revised version of Catapult uses newer FPGAs and was deployed at larger scale in 2016
[Cau 16].
Catapult V1 runs CNNs—using a systolic matrix multiplier—2.3X as fast as a 2.1 GHz, 16-core, dual-socket server
[Ovt15a]. Using the next generation of FPGAs (14-nm Arria 10) of Catapult V2, performance might go up to 7X, and perhaps
even 17X with more careful floorplanning [Ovt15b]. Although it’s apples versus oranges, a current TPU die runs its CNNs
40X to 70X versus a somewhat faster server (Tables 2 and 6). Perhaps the biggest difference is that to get the best
performance the user must write long programs in the low-level hardware-design-language Verilog [Met16][Put16] versus
writing short programs using the high-level TensorFlow framework. That is, reprogrammability comes from software for the
TPU rather than from firmware for the FPGA.
Recent research, which appeared after the TPU was deployed, accelerates DNNs by optimizing the cases when weights
and data are very small or zero. Our tight schedule precluded such optimizations in the TPU, but we saw the same opportunity
in our studies. The Efficient Inference Engine is based on a first pass that reduces the number of weights by about a factor of
10 [Han15] as a separate step by filtering out very small values and then uses Huffman encoding to shrink the data even
further to improve inference performance [Han16]. Cnvlutin [Alb16] avoids multiplications when an activation input is
zero—which it is 44% of the time, presumably in part due to ReLU nonlinear function that transforms negative values to
zero—to improve performance by an average 1.4 times.
Eyeriss is a novel, low-power dataflow architecture that takes advantage of zeros by run-length encoding data to reduce
the memory footprint and saves power by avoiding computations when an input is zero [Che16a]. Using Eyeriss terminology,
a TPU convolutional layer maps C and M to the rows and columns of the matrix unit, taking HWN cycles to perform one
pass. With high C/M, it takes RS passes to process the layer; for low C/M, a number of techniques reduce passes and improve
utilization. (More can be found in the online references [Ros15a][Ros15b][Ros15c][Ros15f][Tho15][You15]).
Minerva is a co-design system that crosses algorithm, architecture, and circuit disciplines to reduce power by 8X in part
by pruning activation data with small values and in part by quantizing the data [Rea16]. [Gup15] looks at 16-bit fixed-point
arithmetic for training instead of for inference. Others leverage the lower precision of DNN calculations by utilizing analog
circuits during the computation to improve energy and performance [LiK16] [Sha16]. By tailoring an instruction set to DNNs,
Cambricon reduces code size [Liu16]. Recent work looked at processor-in-memory architectures for NNs [Chi16][Kim16].
Comparing the TPU to some of these architectures:
● [Che14a] DMAs data from DRAM to input and weight buffers. They are read by the 3-stage pipelined NFU that
performs multiplies, adds, and non-linear-functions; the results go to the output buffer, and then to DRAM. The NFU
has no storage and isn’t systolic.
14
[Gup15] appears to stream both matrix inputs while storing partial sums in the systolic array; the TPU stores the
weight matrix tile while streaming the other input and the pre-activation partial sums. The TPU doesn’t support
stochastic rounding.
● [Zha15] is built out of computation units equivalent to a 4x2 version of the TPU matrix unit. In an ASIC, the wiring
cost of the crossbars that connect input and output buffers to these compute engines would be significant. We are
surprised that we didn’t see architectural support for additional reductions to combine results from compute engines
in [Zha15].
All three of [Gup15][Che14a][Zha15] store activations in DRAM during computation; the TPU’s Unified Buffer is sized so
that no DRAM spilling or reloading happens during normal operation.
●
10. Conclusion
Despite living on an I/O bus and having relatively little memory bandwidth that limits utilization of the TPU—four of the six
NN applications are memory-bound—a small fraction of a big number can nonetheless be relatively large, as the Roofline
performance model demonstrates. This result suggests a “Cornucopia Corollary” to Amdahl’s Law: low utilization of a huge,
cheap resource can still deliver high, cost-effective performance.
The TPU leverages the order-of-magnitude reduction in energy and area of 8-bit integer systolic matrix multipliers over
32-bit floating-point datapaths of a K80 GPU to pack 25 times as many MACs (65,536 8-bit vs. 2,496 32-bit) and 3.5 times
the on-chip memory (28 MiB vs. 8 MiB) while using less than half the power of the K80 in a relatively small die. This larger
memory helps increase the operational intensity of applications to let them utilize the abundant MACs even more fully.
We found that despite a recent emphasis on CNNs in the architecture community, they constitute only about 5% of the
representative NN workload for our datacenters, which suggests more attention should be paid to MLPs and LSTMs.
Repeating history, it’s similar to when many architects concentrated on floating-point performance when most mainstream
workloads turned out to be dominated by integer operations.
We observed that inferences per second (IPS) is more a function of the NN than of the underlying hardware, and so IPS is
an even worse single performance metric for NN processors than MIPS and MFLOPS are for CPUs and GPUs.
We also learned that inference applications have serious response-time bounds because they are often part of user facing
applications, thus NN architectures need to perform well when coping with 99th-percentile latency deadlines. While the K80
may excel at training, on average it is just a little faster than Haswell at inference for our workload, perhaps because of its
emphasis on throughput rather than latency; that conflicts with the strict response-time deadlines of our inference applications.
The TPU die leverages its advantage in MACs and on-chip memory to run short programs written using the domainspecific TensorFlow framework 15 times as fast as the K80 GPU die, resulting in a performance/Watt advantage of 29 times,
which is correlated with performance/total cost of ownership. Compared to the Haswell CPU die, the corresponding ratios are
29 and 83. While future CPUs and GPUs will surely run inference faster, a redesigned TPU using circa 2015 GPU memory
would go two to three times as fast and boost the performance/Watt advantage of nearly 70 over the K80 and 200 over
Haswell.
In summary, the TPU succeeded because of the large—but not too large—matrix multiply unit; the substantial softwarecontrolled on-chip memory; the ability to run whole inference models to reduce dependence on host CPU; a single-threaded,
deterministic execution model that proved to be a good match to 99th-percentile response time limits; enough flexibility to
match the NNs of 2017 as well as of 2013; the omission of general-purpose features that enabled a small and low power die
despite the larger datapath and memory; the use of 8-bit integers by the quantized applications; and that applications were
written using TensorFlow, which made it easy to port them to the TPU at high-performance rather than them having to be
rewritten to run well on the very different TPU hardware.
Order-of-magnitude differences between commercial products are rare in computer architecture, which may lead to the
TPU becoming an archetype for domain-specific architectures. We expect that many will build successors that will raise the
bar even higher.
Acknowledgements
We thank the leadership of our company for recognizing the need for a TPU and for providing the resources to build,
distribute, evaluate, and publish. Special thanks go to Luiz Barroso and James Laudon for helping start the project. It takes a
village to design, verify, and implement the hardware and software of a system like a TPU and to manufacture, deploy, and
use it at scale, which is why there are many authors. (All authors but David Patterson worked on the TPU; he joined in 2016.)
The first four authors did the bulk of the evaluation in this paper, which is why they are in front, with the rest in alphabetical
order. Jouppi was also the senior architect of the whole project; the reward for his yeoman’s work is being the lead author.
15
References
[Aba16]
[Alb16]
[Ado16]
[Asa02]
[Asa98]
[Bar07]
[Bar16]
[Bro16]
[Cau 16]
[Cav15]
[Cha10]
[Che14a]
[Che14b]
[Che16a]
[Che16b]
[Chi16]
[Cla15]
[Dal16]
[Dea13]
[Dea16]
[Du15]
[Far09]
[Far11]
[Gup15]
[Ham90]
[Han15]
[Han16]
[He16]
[Hen18]
[Höl09]
[Ien96]
[Int16]
[Jou16]
[Keu16]
[Kim16]
[Kri12]
[Kun80]
[Lan09]
[Lar16]
Abadi, M., Agarwal, A., Barham, P., Brevdo, E., Chen, Z., Citro, C., Corrado, G.S., Davis, A., Dean, J., Devin, M. and Ghemawat, S., 2016.
Tensorflow: Large-scale machine learning on heterogeneous distributed systems. arXiv preprint arXiv:1603.04467.
Albericio, J., Judd, P., Hetherington, T., Aamodt, T., Jerger, N.E. and Moshovos, A., 2016 Cnvlutin: Ineffectual-Neuron-Free Deep Neural
Network Computing. Proceedings of the 43rd International Symposium on Computer Architecture.
Adolf, R., Rama, S., Reagen, B., Wei, G.Y. and Brooks, D., 2016, September. Fathom: reference workloads for modern deep learning methods.
IEEE International Symposium on Workload Characterization (IISWC).
Asanović, K. 2002. Programmable Neurocomputing, in The Handbook of Brain Theory and Neural Networks: Second Edition, M. A. Arbib (Ed.),
MIT Press, ISBN 0-262-01197-2, November 2002. https://people.eecs.berkeley.edu/~krste/papers/neurocomputing.pdf
Asanović, K. 1998. Asanović, K., Beck, Johnson, J., Wawrzynek, J., Kingsbury, B. and Morgan, N., November 1998. Training Neural Networks
with Spert-II. Chapter 11 in Parallel Architectures for Artificial Networks: Paradigms and Implementations, N. Sundararajan and P. Saratchandran
(Eds.), IEEE Computer Society Press, ISBN 0-8186-8399-6. https://people.eecs.berkeley.edu/~krste/papers/annbook.pdf
Barroso, L.A. and Hölzle, U., 2007. The case for energy-proportional computing. IEEE Computer, vol. 40.
Barr, J. September 29, 2016, New P2 Instance Type for Amazon EC2 – Up to 16 GPUs.
https://aws.amazon.com/blogs/aws/new-p2-instance-type-for-amazon-ec2-up-to-16-gpus/
Brooks, D. November 4, 2016. Private communication.
Caulfield, A.M., Chung, E.S., Putnam, A., Haselman, H.A.J.F.M., Humphrey, S.H.M., Daniel, P.K.J.Y.K., Ovtcharov, L.T.M.K., Lanka,
M.P.L.W.S. and Burger, D.C.D., 2016. A Cloud-Scale Acceleration Architecture. MICRO conference.
Cavigelli, L., Gschwend, D., Mayer, C., Willi, S., Muheim, B. and Benini, L., 2015, May. Origami: A convolutional network accelerator.
Proceedings of the 25th edition on Great Lakes Symposium on VLSI.
Chakradhar, S., Sankaradas, M., Jakkula, V. and Cadambi, S., 2010, June. A dynamically configurable coprocessor for convolutional neural
networks. Proceedings of the 37th International Symposium on Computer Architecture.
Chen, T., Du, Z., Sun, N., Wang, J., Wu, C., Chen, Y. and Temam, O., 2014. Diannao: A small-footprint high-throughput accelerator for
ubiquitous machine-learning. Proceedings of ASPLOS.
Chen, Y., Luo, T., Liu, S., Zhang, S., He, L., Wang, J., Li, L., Chen, T., Xu, Z., Sun, N. and Temam, O., 2014, December. Dadiannao: A
machine-learning supercomputer. Proceedings of the 47th Annual International Symposium on Microarchitecture.
Chen, Y.H., Emer, J. and Sze, V., 2016. Eyeriss: A Spatial Architecture for Energy-Efficient Dataflow for Convolutional Neural Networks.
Proceedings of the 43rd International Symposium on Computer Architecture.
Chen, Y., Chen, T.,Xu, Z., Sun, N., and Teman, O., 2016. DianNao Family: Energy-Efficient Hardware Accelerators for Machine Learning,
Research Highlight, Communications of the ACM, 59(11).
Chi, P., Li, S., Qi, Z., Gu, P., Xu, C., Zhang, T., Zhao, J., Liu, Y., Wang, Y. and Xie, Y., 2016. PRIME: A Novel Processing-In-Memory
Architecture for Neural Network Computation in ReRAM-based Main Memory. Proceedings of the 43rd International Symposium on Computer
Architecture.
Clark, J. October 26, 2015, Google Turning Its Lucrative Web Search Over to AI Machines. Bloomberg Technology, www.bloomberg.com.
Dally, W. February 9, 2016. High Performance Hardware for Machine Learning, Cadence ENN Summit.
Dean, J. and Barroso, L.A., 2013. The tail at scale. Communications of the ACM, 56(2).
Dean, J. July 7, 2016 Large-Scale Deep Learning with TensorFlow for Building Intelligent Systems, ACM Webinar.
Du, Z., Fasthuber, R., Chen, T., Ienne, P., Li, L., Luo, T., Feng, X., Chen, Y. and Temam, O., 2015, June. ShiDianNao: shifting vision processing
closer to the sensor. Proceedings of the 42nd International Symposium on Computer Architecture.
Farabet, C., Poulet, C., Han, J.Y. and LeCun, Y., 2009, August. Cnp: An FPGA-based processor for convolutional networks. 2009 International
Conference on Field Programmable Logic and Applications.
Farabet, C., Martini, B., Corda, B., Akselrod, P., Culurciello, E. and LeCun, Y., 2011, June. Neuflow: A runtime reconfigurable dataflow
processor for vision. In CVRP 2011 Workshops.
Gupta, S., Agrawal, A., Gopalakrishnan, K. and Narayanan, P., 2015, July. Deep Learning with Limited Numerical Precision. ICML.
Hammerstrom, D., 1990, June. A VLSI architecture for high-performance, low-cost, on-chip learning. 1990 IJCNN International Joint Conference
on Neural Networks.
Han, S.; Pool, J.; Tran, J.; and Dally, W., 2015. Learning both weights and connections for efficient neural networks. In Advances in Neural
Information Processing Systems.
Han, S., Liu, X., Mao, H., Pu, J., Pedram, A., Horowitz, M.A. and Dally, W.J., 2016. EIE: efficient inference engine on compressed deep neural
network. Proceedings of the 43rd International Symposium on Computer Architecture.
He, K., Zhang, X., Ren, S. and Sun, J., 2016. Identity mappings in deep residual networks. Also in arXiv preprint arXiv:1603.05027.
Hennessy, J.L. and Patterson, D.A., 2018. Computer architecture: a quantitative approach, 6th edition, Elsevier.
Hölzle, U. and Barroso, L., 2009. The datacenter as a computer. Morgan and Claypool.
Ienne, P., Cornu, T. and Kuhn, G., 1996. Special-purpose digital hardware for neural networks: An architectural survey. Journal of VLSI signal
processing systems for signal, image and video technology, 13(1).
Intel, 2016, Intel® Xeon® Processor E5-4669 v3, http://ark.intel.com/products/85766/Intel-Xeon-Processor-E5-4669-v3-45M-Cache-2_10-GHz
Jouppi, N. May 18, 2016. Google supercharges machine learning tasks with TPU custom chip. https://cloudplatform.googleblog.com
Keutzer, K., 2016. If I could only design one circuit…: technical perspective. Communications of the ACM, 59(11),.
Kim, D., Kung, J.H., Chai, S., Yalamanchili, S. and Mukhopadhyay, S., 2016. Neurocube: A Programmable Digital Neuromorphic Architecture
with High-Density 3D Memory. Proceedings of the 43rd International Symposium on Computer Architecture.
Krizhevsky, A., Sutskever, I. and Hinton, G., 2012. Imagenet classification with deep convolutional neural networks. Advances in neural
information processing systems.
Kung, H.T. and Leiserson, C.E., 1980. Algorithms for VLSI processor arrays. Introduction to VLSI systems.
Lange, K.D., 2009. Identifying shades of green: The SPECpower benchmarks. IEEE Computer, 42(3).
Larabel, M. March 10, 2016, Google Looks To Open Up StreamExecutor To Make GPGPU Programming Easier, Phoronix,
https://www.phoronix.com/scan.php?page=news_item&px=Google-StreamExec-Parallel.
16
[LiK16]
[Liu15]
[Liu16]
[Met16]
[Nvi15]
[Nvi16]
[Ovt15a]
[Ovt15b]
[Pat04]
[Pee13]
[Put14]
[Put15]
[Put16]
[Qad13]
[Ram91]
[Rea16]
[Ros15a]
[Ros15b]
[Ros15c]
[Ros15d]
[Rus15]
[Sch09]
[Sha16]
[Sil16]
[Smi82]
[Ste15]
[Sze15]
[Tho15]
[Wil09]
[Wu16]
[You15]
[Zha15]
LiKamWa, R., Hou, Y., Gao, J., Polansky, M. and Zhong, L., 2016. RedEye: Analog ConvNet Image Sensor Architecture for Continuous Mobile
Vision. Proceedings of the 43rd International Symposium on Computer Architecture.
Liu, D., Chen, T., Liu, S., Zhou, J., Zhou, S., Teman, O., Feng, X., Zhou, X. and Chen, Y., 2015, March. Pudiannao: A polyvalent machine
learning accelerator. Proceedings of the 42nd International Symposium on Computer Architecture.
Liu, S., Du, Z.D., Tao, J.H., Han, D., Luo, T., Xie, Y., Chen, Y. and Chen, T., 2016. Cambricon: An instruction set architecture for neural
networks. Proceedings of the 43rd International Symposium on Computer Architecture.
Metz, C. September 26, 2016, Microsoft Bets Its Future On A Reprogrammable Computer Chip, Wired Magazine,
https://www.wired.com/2016/09/microsoft-bets-future-chip-reprogram-fly/
Nvidia, January 2015. Tesla K80 GPU Accelerator. Board Specification
https://images.nvidia.com/content/pdf/kepler/Tesla-K80-BoardSpec-07317-001-v05.pdf.
Nvidia, 2016. Tesla GPU Accelerators For Servers. http://www.nvidia.com/object/tesla-servers.html.
Ovtcharov, K., Ruwase, O., Kim, J.Y., Fowers, J., Strauss, K. and Chung, E.S., February 2, 2015. Accelerating deep convolutional neural
networks using specialized hardware. Microsoft Research Whitepaper.
https://www.microsoft.com/en-us/research/publication/accelerating-deep-convolutional-neural-networks-using-specialized-hardware/
Ovtcharov, K., Ruwase, O., Kim, J.Y., Fowers, J., Strauss, K. and Chung, E.S., 2015, August. Toward accelerating deep learning at scale using
specialized hardware in the datacenter. 2015 IEEE Hot Chips 27 Symposium.
Patterson, D.A., 2004. Latency lags bandwith. Communications of the ACM, 47(10).
Peemen, M., Setio, A.A., Mesman, B. and Corporaal, H., 2013, October. Memory-centric accelerator design for convolutional neural networks. In
2013 IEEE 31st International Conference on Computer Design (ICCD).
Putnam, A., Caulfield, A.M., Chung, E.S., Chiou, D., Constantinides, K., Demme, J., Esmaeilzadeh, H., Fowers, J., Gopal, G.P., Gray, J.,
Haselman, M., Hauck, S., Heil, S., Hormati, A., Kim, J-Y., Lanka, S., Larus, J., Peterson, E., Pope, S ., Smith, A., Thong, J., Xiao, P.Y., Burger,
D. 2014, June. A reconfigurable fabric for accelerating large-scale datacenter services. 41st International Symposium on Computer Architecture.
Putnam, A., Caulfield, A.M., Chung, E.S., Chiou, D., Constantinides, K., Demme, J., Esmaeilzadeh, H., Fowers, J., Gopal, G.P., Gray, J.,
Haselman, M., Hauck, S., Heil, S., Hormati, A., Kim, J-Y., Lanka, S., Larus, J., Peterson, E., Pope, S ., Smith, A., Thong, J., Xiao, P.Y., Burger,
D. 2015. A Reconfigurable Fabric for Accelerating Large-Scale Datacenter Services. IEEE Micro, 35(3).
Putnam, A., Caulfield, A.M., Chung, E.S., Chiou, D., Constantinides, K., Demme, J., Esmaeilzadeh, H., Fowers, J., Gopal, G.P., Gray, J.,
Haselman, M., Hauck, S., Heil, S., Hormati, A., Kim, J-Y., Lanka, S., Larus, J., Peterson, E., Pope, S ., Smith, A., Thong, J., Xiao, P.Y., Burger,
D. 2016. A Reconfigurable Fabric for Accelerating Large-Scale Datacenter Services. Communications of the ACM.
Qadeer, W., Hameed, R., Shacham, O., Venkatesan, P., Kozyrakis, C. and Horowitz, M.A., 2013, June. Convolution engine: balancing efficiency
& flexibility in specialized computing. Proceedings of the 40th International Symposium on Computer Architecture.
Ramacher, U., Beichter, J., Raab, W., Anlauf, J., Bruels, N., Hachmann, U. and Wesseling, M., 1991. Design of a 1st Generation Neurocomputer.
In VLSI design of Neural Networks. Springer US.
Reagen, B., Whatmough, P., Adolf, R., Rama, S., Lee, H., Lee, S.K., Hernández-Lobato, J.M., Wei, G.Y. and Brooks, D., 2016. Minerva:
Enabling low-power, highly-accurate deep neural network accelerators. Proceedings of the 43rd International Symposium on Computer
Architecture.
Ross, J., Jouppi, N., Phelps, A., Young, C., Norrie, T., Thorson, G., Luu, D., 2015. Neural Network Processor, Patent Application No. 62/164,931.
Ross, J., Phelps, A., 2015. Computing Convolutions Using a Neural Network Processor, , Patent Application No. 62/164,902.
Ross, J., 2015. Prefetching Weights for a Neural Network Processor, Patent Application No. 62/164,981.
Ross, J., Thorson, G., 2015. Rotating Data for Neural Network Computations,Patent Application No. 62/164,908.
Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M. and Berg, A.C., 2015.
Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3).
Schurman, E. and Brutlag, J., 2009, June. The user and business impact of server delays, additional bytes, and HTTP chunking in web search. In
Velocity Web Performance and Operations Conference.
Shafiee, A., Nag, A., Muralimanohar, N., Balasubramonian, R., Strachan, J.P., Hu, M., Williams, R.S. and Srikumar, V., 2016. ISAAC: A
Convolutional Neural Network Accelerator with In-Situ Analog Arithmetic in Crossbars. Proceedings of the 43rd International Symposium on
Computer Architecture.
Silver, D., Huang, A., Maddison, C.J., Guez, A., Sifre, L., Van Den Driessche, G., Schrittwieser, J., Antonoglou, I., Panneershelvam, V., Lanctot,
M. and Dieleman, S., 2016. Mastering the game of Go with deep neural networks and tree search. Nature, 529(7587).
Smith, J.E., 1982, April. Decoupled access/execute computer architectures. Proceedings of the 11th International Symposium on Computer
Architecture.
Steinberg, D., 2015. Full-Chip Simulations, Keys to Success. Proceedings of the Synopsys Users Group (SNUG) Silicon Valley 2015.
Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V. and Rabinovich, A., 2015. Going deeper with
convolutions. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.
Thorson, G., Clark, C., Luu, D., 2015. Vector Computation Unit in a Neural Network Processor, Patent Application No. 62/165,022.
Williams, S., Waterman, A. and Patterson, D., 2009. Roofline: an insightful visual performance model for multicore architectures.
Communications of the ACM.
Wu, Y., Schuster, M., Chen, Z., Le, Q., Norouzi, M., Macherey, W., Krikun, M., Cao, Y., Gao, Q., Macherey, K., Klingner, J., Shah, A., Johnson,
M., Liu, X., Kaiser, Ł., Gouws, S., Kato, Y., Kudo, T., Kazawa, H., Stevens, K., Kurian, G., Patil, N., Wang, W., Young, C., Smith, J., Riesa, J.,
Rudnick, A., Vinyals, O., Corrado, G., Hughes, M., and Dean, J. September 26, 2016, Google's Neural Machine Translation System: Bridging the
Gap between Human and Machine Translation, http://arxiv.org/abs/1609.08144.
Young, C., 2015. Batch Processing in a Neural Network Processor, Patent Application No. 62/165,020.
Zhang, C., Li, P., Sun, G., Guan, Y., Xiao, B. and Cong, J., 2015, February. Optimizing FPGA-based accelerator design for deep convolutional
neural networks. Proceedings of the 2015 ACM/SIGDA International Symposium on Field-Programmable Gate Arrays.
17
| 9cs.NE |
Object Recognition by Using Multi-level Feature Point Extraction
Yang Cheng, Timeo Dubois
arXiv:1710.10522v1 [cs.CV] 28 Oct 2017
1
Abstract
In this paper, we present a novel approach for object recognition in real-time by employing multilevel feature analysis and demonstrate the practicality of adapting feature extraction into a Naive
Bayesian classification framework that enables simple, efficient, and robust performance. We also
show the proposed method scales well as the number of level-classes grows. To effectively understand the patches surrounding a keypoint, the trained classifier uses hundreds of simple binary
features and models class posterior probabilities. In addition, the classification process is computationally cheap under the assumed independence between arbitrary sets of features. Even though
for some particular scenarios, this assumption can be invalid. We demonstrate that the efficient
classifier nevertheless performs remarkably well on image datasets with a large variation in the
illumination environment and image capture perspectives. The experiment results show consistent
accuracy can be achieved on many challenging dataset while offer interactive speed for large resolution images. The method demonstrates promising results that outperform the state-of-the-art
methods on pattern recognition.
1
2
Introduction
In the literature, one common idea image feature extraction is to focus on finding strong features
that is robust enough to deal with perspective changes, lighting variations, such as SIFT local
patterns detection and combines the output of the classifiers. Plagemann et al.6 create a novel
interest point detector for catching body components from depth images.
Identifying textured patches surrounding keypoints across images acquired under widely
varying poses and lightning conditions is at the heart of many Computer Vision algorithms 1 .
The resulting correspondences can be used to register different views of the same scene, extract
3D shape information, or track objects across video frames. Correspondences also play a major
role in object category recognition and image retrieval applications 2 .
Because of its strength to fractional impediments and computational proficiency, acknowledgment of picture patches removed around distinguished key focuses is critical for some vision
issues. Therefore, two fundamental classes of methodologies have been produced to accomplish
power to point of view and lighting changes 3 . The primary family depends on nearby descriptors intended to be invariant, or possibly hearty, to particular classes of distortions 4 . An inferior
depends on measurable learning procedures to figure a probabilistic model of the fix.
2
3
A Semi-naive Bayesian approach to patch recognition
Picture patches can be perceived on the premise of extremely straightforward and arbitrarily picked
double tests that are gathered into choice trees and recursively parcel the space of all conceivable
patches. By and by, no single tree is sufficiently discriminative when there are many classes. In
any case, utilizing various trees and averaging their votes yields great outcomes in light of the fact
that everyone parcels the component space in an unexpected way 5 .
Formulation of Feature Combination Given the patch surrounding a key point detected in an
image, our task is to assign it to the most likely class. Let Ci, I= 1, H Be the set of classes and
let fj, j= 1, N Be the set of binary features that will be calculated over the patch we are trying to
classify. Formally, we are looking for
Ci = argC imaxP (C = ci |f 1, f 2, ..., fN )
Where C is a random variable that represents the class. Bayes’s formula yields
P (C = ci |f 1, f 2, ..., fN ) =
c)P (C = ci )
p(f 1, f 2, ..., fN )
[As the denominator is simply a scaling factor, it can be reduced to]
3
Ci = argC imaxP (C = ci |f 1, f 2, ..., fN )
fj =
1if I(dj ,1 ) < I(dj ,2 )
0otherwise
Where I represents the image patch 7 . (As the features specified are pretty simple, so N 300.
So)
P (f 1, f 2, ....., fN |C = ci ) =
N
Y
P (fj |C = ci )
j=1
4
Comparison With Randomized Trees
As shown in Figures 3 and 4, Ferns can be considered as simplified trees. To compare RTs and
Ferns, we experimented with the three images of Figure 5.
Greeneries vary from trees in two vital regards: The probabilities are increased in a NaiveBayesian manner as opposed to being arrived at the midpoint of and the various leveled structure
is supplanted by a level one.
The preparation set is acquired by arbitrarily twisting pictures of Figure 5. To play out these
4
examinations, we speak to relative picture disfigurements as 22 networks of the shape.
R0 R− ∅diag(λ1, λ2)R∅
Where diag (?1, ?2) is a corner to corner 2 2 framework and R? speaks to a revolution of
point ? . Both to prepare and to test our greeneries, we distorted the first pictures utilizing such
misshapenness registered by arbitrarily picking ? and f in the [0: 2p] territory and ?1 and ?2 in the
5
[0.6: 1.5] territory. Fig. 6 delineates patches encompassing individual intrigue focuses first in the
first pictures and after that in the distorted ones 7 . We utilized 30 irregular relative distortions for
each level of pivot to create 10800 pictures.
For the most part the test set is acquired by creating a different arrangement of 1000 pictures
in a similar relative twisting extent and including clamor. In Figure 7, we plot the outcomes as an
element of the quantity of trees or Ferns being utilized.
We first note that utilizing either level Fern or progressive tree structures does not influence
6
the acknowledgment rate, which was not out of the ordinary as the components are taken totally
at irregular. Moreover as the scramble plots of Figure 7 appear, for the Naive-Bayesian blend the
acknowledgment rate on individual twisted pictures never falls beneath a worthy rate.
5
Experiments
It is hard to play out a totally reasonable speed examination between our Ferns and SIFT for a
few reasons
16
. Filter reuses moderate information from the key guide extraction toward register
canonic scale and introductions and the descriptors, while greeneries can depend on a minimal
effort key-point extraction 8 . Ferns vs SIFT to detect 3D objects So far we have considered that
the key-focuses lie on a planar protest and assessed the heartiness of Ferns concerning viewpoint
impacts. This rearranges preparing as a solitary view is adequate and the known 2D geometry
can be utilized to figure ground truth correspondences. However most protests have really three dimensional appearance, which suggests that self-impediments and complex enlightenments impacts
must be considered to effectively assess the execution of any key-point coordinating calculation 9 .
Figure 11 shows such pictures traversing a 70? camera revolution go. We utilized this picture
database to assess the execution of Ferns for an assortment of 3D items. We look at our outcomes
against the SIFT identifier/descriptor match which has been found to perform extremely well on
this database. The key-focuses and the descriptors are registered utilizing an indistinguishable
programming from some time recently. We acquired the ground truth by utilizing simply geometric
strategies, which is conceivable on the grounds that the cameras and the turn table are adjusted
7
10
. The underlying correspondences are acquired by utilizing the trifocal geometry between the
top/base cameras in the inside view and each other camera as outlined by Figure 12
15
. We then
reproduce the 3D focuses for each such correspondence in the base/focus camera arrange casing
and utilize these to frame the underlying tracks that traverse the-35?/+35?rotation territory around
a focal view. At last to build power against spurious tracks framed by anomalies, we take out
tracks covering less than 30percent of the perspectives and the rest of the tracks shape the ground
8
truth for the assessment 13 , which is free of exceptions. Test ground truth information is portrayed
by Figure 13, which demonstrates the mind boggling varieties in fix appearance incited by the 3D
structure of the items 11 .
6
Panorama and 3D Scene Annotation
With the current expansion of cell phones with huge handling power, there has been a surge of
enthusiasm for building genuine applications that can consequently explain the photographs and
give helpful data about spots of intrigue. We have tried Ferns on two such applications, comment
of display scenes and parts of a verifiable working with 3D structure
14
. Both applications run
easily at casing rate utilizing a standard portable PC and an of the rack web camera. By applying
standard improvements for implanted equipment, we have ported this execution onto a cell phone
that keeps running at a couple outlines for every second.
9
7
Conclusion
We have introduced an intense strategy for picture fix acknowledgment that performs well even
within the sight of extreme viewpoint twisting. A key part of our approach is the Naive-Bayesian
blend of classifiers that obviously beats the averaging of probabilities we utilized as a part of prior
work. We have demonstrated that such a credulous mix methodology is a beneficial option when
the particular issue is not excessively delicate to the suggested autonomy suspicions.
8
References
1. Thomas Mrwald, Markus Liu, ”Self-monitoring to improve robustness of 3D object tracking
for robotics”, Robotics and Biomimetics (ROBIO) 2011 IEEE International Conference on, pp.
10
2830-2837, 2013.
2. Jun-Sik Kim, Makoto Lee, Joo Xavier, Pedro Aguiar, Takeo Kanade, ”6D pose estimation of
textureless shiny objects using random ferns for bin-picking”, Intelligent Robots and Systems
(IROS) 2012 IEEE/RSJ International Conference on, pp. 3334-3341, 2002, ISSN 2153-0858.
3. J. Shen and J. Yang, “Automatic human animation for non-humanoid 3d characters,” International Conference on Computer-Aided Design and Computer Graphics (CAD/Graphics), pp.
220-221, 2015.
4. Wei Xiong, Shue Ching Chia, Yue Chen, Weimin Huang, Jiayin Zhou, Yufeng Zhou, Wilson
Gao, Kae Jack Tay, Henry Ho, ”Prostate boundary segment extraction using cascaded shape regression and optimal surface detection”, Engineering in Medicine and Biology Society (EMBC)
2014 36th Annual International Conference of the IEEE, pp. 2886-2889, 2010, ISSN 1557170X.
5. Dominik Sibbing, Bastian Liu, Leif Kobbelt, ”SIFT-Realistic Rendering”, 3DTV-Conference
2013 International Conference on, pp. 56-63, 2015.
6. Shu Wang, Mei Xie, ”Iris matching using ferns classifier”, Wavelet Active Media Technology and Information Processing (ICWAMTIP) 2012 International Conference on, pp. 109-112,
2014.
7. Ronny Hnsch, Xiaohong Zhao, ”Machine-learning based detection of corresponding interest
points in optical and SAR images”, Geoscience and Remote Sensing Symposium (IGARSS)
2014 IEEE International, pp. 1492-1495, 2016, ISSN 2153-7003.
11
8. Shaoguo Zhang, Yiyi Wei, Chunhong Pan, ”BB-Homography: Joint Binary Features and Bipartite Graph Matching for Homography Estimation”, Circuits and Systems for Video Technology
IEEE Transactions on, vol. 25, pp. 239-250, 2011, ISSN 1051-8215.
9. J. Shen and S. C. S. Cheung, “Layer Depth Denoising and Completion for Structured-Light
RGB-D Cameras,” IEEE Conference on Computer Vision and Pattern Recognition, pp. 11871194, 2013.
10. Hideaki Uchiyama, Jia Li, Eric Marchand, ”Toward augmenting everything: Detecting and
tracking geometrical features on planar objects”, Mixed and Augmented Reality (ISMAR) 2011
10th IEEE International Symposium on, pp. 17-25, 2015.
11. J. Shen, P. C. Su, S. c. S. Cheung, J. Zhao, “Virtual mirror rendering with stationary rgb-d
cameras and stored 3-d background,” IEEE Transactions on Image Processing, vol. 22, no. 9,
pp. 3433-3448, 2013.
12. Taiki Fuji, Toshio Moriya, ”Furniture layout AR application using floor plans based on planar
object tracking”, RO-MAN 2012 IEEE, pp. 670-675, 2002, ISSN 1944-9445.
13. Yasue Mitsukura, Wei Xiong, Yue Wang, Shue Ching Chia, Wenyu Chen, Jia Du, Ying Gu,
Victor Ter Shen Kow, ”CHORD: Cascaded and a contrario method for hole crack detection”,
Image Processing (ICIP) 2015 IEEE International Conference on, pp. 3300-3304, 2011.
14. Victor Eruhimov, Y. Konishi, C. Beltran, V. Murino, A. Del Bue, ”Fast 6D pose from a single
RGB image using Cascaded Forests Templates”, Intelligent Robots and Systems (IROS) 2016
IEEE/RSJ International Conference on, pp. 4062-4069, 2010, ISSN 2153-0866.
12
15. E. Muoz, Wim Meeussen, ”Outlet detection and pose estimation for robot continuous operation”, Intelligent Robots and Systems (IROS) 2011 IEEE/RSJ International Conference on, pp.
2941-2946, 2010, ISSN 2153-0858.
16. J. Shen and W. Tan, “Image-based indoor place-finder using image to plane matching,” IEEE
International Conference on Multimedia and Expo, pp. 1-6, 2013.
13
| 7cs.IT |
Linear Regression with Sparsely Permuted Data
arXiv:1710.06030v2 [math.ST] 15 Nov 2017
Martin Slawski
Department of Statistics
George Mason University
Fairfax, VA 22030, USA
mslawsk3@gmu.edu
Emanuel Ben-David
Center for Statistical Research Methodology
U.S. Census Bureau
Suitland, MD 20746, USA
emanuel.ben.david@census.gov
Abstract
In regression analysis of multivariate data, it is tacitly assumed that response and
predictor variables in each observed response-predictor pair correspond to the same
entity or unit. In this paper, we consider the situation of “permuted data” in which this
basic correspondence has been lost. Several recent papers have considered this situation
without further assumptions on the underlying permutation. In applications, the latter
is often to known to have additional structure that can be leveraged. Specifically, we
herein consider the common scenario of ”sparsely permuted data” in which only a
small fraction of the data is affected by a mismatch between response and predictors.
However, an adverse effect already observed for sparsely permuted data is that the
least squares estimator as well as other estimators not accounting for such partial
mismatch are inconsistent. One approach studied in detail herein is to treat permuted
data as outliers which motivates the use of robust regression formulations to estimate
the regression parameter. The resulting estimate can subsequently be used to recover
the permutation. A notable benefit of the proposed approach is its computational
simplicity given the general lack of procedures for the above problem that are both
statistically sound and computationally appealing.
1
Introduction
A largely unquestioned assumption in regression analysis with response-predictor pairs
{(yi , xi )}ni=1 is that each yi corresponds to the same statistical unit as xi . In this paper,
we consider the situation where the identifiers of the predictors (or equivalently those of
the responses) are subject to an unknown permutation so that correspondence between yi
and xi may not be be taken for granted. We refer to this situation as “permuted data”
respectively “sparsely permuted data” when the permutation only affects a small fraction
of all response-predictor pairs. Restoring the original correspondence between responses
and predictors may not be achievable in practice for both computational and statistical
reasons, but may also not be required for consistent estimation of regression parameters.
Conversely, if the primary interest concerns recovering the permutation itself, which is the
case in entity resolution, the regression model can facilitate that task.
1.1
Background and Motivation
Large organizations that own or have access to multiple data sources regularly rely on data
integration for conducting large-scale scientific projects. Available datasets are gathered
or produced at different points of time and independently of one another. The main
reason for combining datasets is that no single dataset contains all variables of interest
that pertain the research questions. Data collected from a single source are often limited
and do not have all the variables needed for statistical analysis. Limited budget, time
and resources prevent each particular agency from collecting a comprehensive dataset. In
sparse_permutation_arxiv-v1.tex, version: November 17, 2017
2
●
●
●
●
●
●●
● ●
●
●●
●
●
●
●
●
●
●
● ●●● ●
●
● ●
●
●
● ●
● ●●●●
● ●
●●
● ●●●
●●
●●
●●
●●
●
●
●
●● ●
●
●● ● ●
●
●●
●
●
●
●
●
●
●
●
●● ● ●
●
●
●●
●
●
● ● ●
● ●
● ●● ●
●
●
●
●
●
●
●
●
●
●
●
●● ●
●●●
●
●
●
●●●
●
●
●
●
● ●●
●● ●
● ●
●
● ●●●●
●●
●
●● ●
●
●
●
●
●
●
●
●
●●
●●●
●
●
●●
●●●
●
●
●
●
●
●
● ●●●● ●●
●
●
●●
●● ●●
● ●●●
●
●
●
● ●
●
●●
●
●
●
● ● ● ●●
●
●
●
●
●
●●
●
●
●
●●
●
●●●
●●
●
●
●
●
● ●
●●
●
●
●
●
● ●
●
●
●●
●
●●● ●
● ●
●
● ●
●
●
●
●● ● ● ●
●●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
● ●● ● ●●●
●● ● ●
●●
●
●
●
●
● ●
●●
●●● ● ● ●●
●●● ● ●
●
●
●
●
●● ●
●
●
●●● ●
●
●
●
●
●
●
●
●
● ●●
● ●●
● ●
●
●
● ● ●
●● ●
●
●
●
●●
●
●
●
●
●
●● ● ●
●●
●
●
●
●
●●●
●●●
●●
●
●
●
●
●
●●
●
●●
● ●
●
● ●●
●
●
●
●●
●
●
●● ●
●
●
●
●
●
●
●
●
●●●●● ●
●
●
● ●●
●
●
●
●●
●
●
●
●●
●
●
●
●
● ●
●
● ●
●
●
●
●
●
● ●
●
●
●
●●
●
●
● ● ●
●
●●
●
●
●
●
●
●
3
●
●
2
●
0
●
●
●
−4
0
2
● ●
● ● ●
●●
●
● ● ●●
●
●
●
● ● ●●●●
●● ●
● ● ● ●●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●
● ●
●
● ●
● ●
●
●
● ●●
●
●
●●
●
●
●
●
●●
●
●
●●●●● ●
●
●●
●●●●●
●
●
●● ●● ●
●
●●
●● ●
●●
●●
●●
●
● ● ● ●● ● ●● ●● ● ●
●● ● ●
●
●
● ●
●
●
● ●
●
●●●●●
●
● ● ●●● ●
●
●
●●
●
●
●
●
●
●
● ● ● ●●● ●
● ●● ●
●●
● ●● ●
●
●● ●
●●
●
●
●
●
●
●●
●
●
●
●●
●●●●●
●
●
●
●
●●
●
●
●
●●●●●
●●
●
●●
●●
●●
●
●●●●● ●●
●●
●
●
●
●
●
●●
●●●●●●●●●●
● ●●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●● ●
●
●
●●● ●
●●
●
●●
●
●
●
●
●●
●
●● ● ●●
●
●
●●
●
●
●
●●
●
●● ●
●●
● ●
●●
●●
●
●● ●
●
● ●●
●●●
●
● ● ●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●● ●● ●
●
●●
● ●
●
● ●
●
●●●
● ●● ●
●●
●
●●
● ●●
●●
● ●● ●●
●●
● ●●
●●
●
●●
●
●
● ● ●●
● ● ●●
●
●● ●● ●
●
●
●●●● ●
●
●
●●●
●
●
●
● ●●●●
●
●●
●● ●
● ●●
●
●●●
●
●
● ● ●
●
● ●
●●
●
● ●● ● ●
●●
●
●
● ● ●
●
●
●
●●
●
● ● ●
●
●
−3
−2
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
−1
Residuals
1
●
●
−2
−2
0
●
−4
Permuted response variables in the merged data
y (original) y (merged)
●
original data
merged data
−4
−3
−2
Response variables as observed in the original data
Fitted
−1
0
1
..
.
−0.96
−3.45
0.28
−3.06
−0.33
−2.98
−0.53
−1.35
..
.
..
.
−3.45
−0.96
−1.86
−0.33
−3.02
−0.53
−2.98
−3.73
..
.
Figure 1: Illustration of the effect of mismatchings in the response variable on linear
regression with merged data. For this analysis, a subset of the El Niño data set (cf. §3.2)
was artificially broken into two parts which were subsequently merged based on a nonunique identifier. Left: Original response vs. response observed after data merging. Middle:
Residuals vs. fitted values in linear regression when using the original (dots) and merged
data set (triangles), respectively. The circles and triangles colored in red correspond to the
excerpt of the response variable depicted on the right hand side.
most situations, however, relevant datasets from other sources are already available. For
example, if a demographic survey does not have all the relevant variables for a particular
research project, other existing surveys or administrative data can be used to include those
missing variables.
Record linkage, or entity resolution, is an essential task in data integration. The task is
to identify which records in different datasets belong to the same entity. In this context, the
term ”entity” is to be understood in a broad sense, and may refer to customers, tax payers
or patients, etc. Record linkage has a long history of uses in large enterprises, government
agencies and health care institutions. A business register consisting of names, addresses,
and other identifying information such as total financial receipts might be constructed
from tax and employment data bases; a survey of retail establishments or agricultural
establishments might combine results from an area frame and a list frame. In this case,
units from the area frame would need to be identified in the list frame [52].
Generally, due to the lack of unique entity identifiers in data files, record matching is
based on methods of probabilistic record linkage [24] that score similarity between common
quasi-identifiers. For example, when the data files contain health information of individuals,
the quasi-identifiers can be first and last names, addresses, dates of birth. Spelling errors,
typographical variation and missing values in records inhibit exact matching and lead to
matching errors: mismatches and missed-matches. Even low matching error rates can
lead to selection bias and pervasive outliers, particularly when the record linkage process
includes or excludes particular types of entities and their attributes [10]. Selection bias
and outliers contaminate subsequent statistical analysis. To reduce bias, it is of interest to
develop statistical methods that can alleviate the adverse effects of matching errors.
In this paper, we focus on the problem of linear regression with sparsely permuted data
as it arises frequently when linking survey records to external data. The canonical example
is regression estimation of a study variable y on auxiliary variables x that reside in another
data source such as master data or administrative records. The matching error, determined
by the fraction of mismatches and missed-matches, is expected to be only a small fraction
of the sample size n [52]. The paper [47] considers two such situations. To model the
2
(y1 , x1 )
..
.
(yn , xn )
(y1 , xϕ(1) )
..
.
(yn , xϕ(n) )
(y1 , xϕ(1) )
..
.
(yn , xϕ(n) )
(y, Π∗ X)
(Π∗> y, X)
data with true correspondence
(y, X)
observed data
Figure 2: Regression with lost correspondence between response and predictors.
energy economy properly, an economist might need company-specific data on the fuel and
feedstocks used by companies that are only available from Agency A and corresponding
microdata on the goods produced for companies that is only available from Agency B. To
model the health of individuals in society, a demographer or health science policy worker
might need individual-specific information on those receiving social benefits from Agencies
B1, B2, and B3, corresponding income information from Agency I, and information on
health services from Agencies H1 and H2.
Regression with permuted data arises in other applications as well. The papers [3, 39]
provide extensive lists including multi-target tracking in radar systems [43] and pose and
correspondence estimation in computer vision [19].
1.2
Problem statement
To setup the problem more concretely, suppose the data (y1 , x1 ), . . . , (yn , xn ) is obtained
from matching two data files A and B, where for each i, yi ∈ R resides in file A and
xi ∈ Rd resides in file B. Since the process of record linkage is not error-free, some xi
may have been paired up with a non-correspondent yi . If the number of such mismatches
is known to be less than or equal to k ≤ n, then there is an unknown permutation ϕ on
[n] := {1, . . . , n} so that: ϕ moves at most k of the indices, and (y1 , xϕ(1) ), . . . , (yn , xϕ(n) )
are independent realizations from the classical linear regression model
y = x> β ∗ + , where ∼ N (0, σ 2 ) and x ⊥⊥ .
(1)
Let Π∗ and Π∗> bethe matrix representations of ϕ and its inverse, respectively, and let
X > = x1 · · · xd . Eq. (1) readily implies that
y = Π∗ Xβ ∗ + , where y = (y1 , . . . , yn )> and = (1 , . . . , n )>
⇔ Π∗> y = Xβ ∗ + Π∗> .
(2)
For simplicity, since the distribution of Π∗> and is the same, we do not distinguish
between Π∗> and in the sequel and write in place of Π∗> .
The main question to be studied here is how the parameters Π∗ and β ∗ can be accurately
and efficiently estimated from the partially mismatched pairs (yi , xi ). Efficiency here refers
to computational complexity of the procedure that outputs the desired estimates. Let us
consider the number of mismatches which can be expressed as dH (Π∗ , In ) := |{i : Π∗ii = 0}|,
the Hamming distance between Π∗ and the identity matrix In . We can naturally estimate
Π∗ and β ∗ by a least-squares approach as:
b = argmin kΠXβ − yk2
b β)
(Π,
2
Π,β
(3)
subject to dH (Π, In ) ≤ k,
3
where Π and β run over Pn , the set of all permutation matrices in Rn×n , and Rd , respectively. A permutation respectively its matrix representation Π is said to be k-sparse if
dH (Π, In ) ≤ k. We note that [39] have shown that unless d = 1 the optimization problem
(3) is NP-hard for k = n. The same is true for any k that is defined as a fraction of n.
Notation. We here gather some notation frequently used in the present paper. For a
positive integer m, Im denotes the m × m identity matrix, and Sm−1 denotes the unit
sphere in Rm . We write |S| for the cardinality of a set S. The complement of S with
respect to a given base set depending on the context is denoted by S c . For a matrix
A, kAk2 = σmax (A) denotes its spectral norm respectively maximum singular value, and
range(A) denotes the column space of A. For an index set I ⊆ [m] := {1, . . . , m} and
v ∈ Rm , vI denotes the subvector of v corresponding to I. We write a ∨ b = max{a, b}.
Positive constants are denoted by C, c, c1 etc. We make use of the usual Big-O notation
in terms of O, o, Ω and Θ.
1.3
Prior work
Linear regression with linked data files is a common scenario in which traditional methods
of statistical data analyses are error-prone. Early work in [37] recognizes the adverse
effect of matching error on the regression analysis of linked datasets. They show that
as a consequence of matching error, the ordinary least squares estimator βbols is generally
not an unbiased estimator of β ∗ . In the paper [37], the process of matching is regarded
as random. Letting zi denote the response that is actually in correspondence to xi , and
letting yi denote the response that has been matched to xi , the model in [37] assumes that
for i = 1, . . . , n,
(
zi with probability qii
yi =
(4)
zj with probability qij , i 6= j.
Assuming that these probabilities can be estimated [34, 35, 37, 46, 47] focus on constructing
an estimator that is unbiased or less biased than βbols . In fact, in [35] the matrix Q = (qij ) is
used to construct an unbiased estimator of β ∗ . The papers [29, 30, 49] propose to estimate
β ∗ using a Bayesian procedure. The main shortcoming of these approaches are that they
rely on the assumption that the doubly stochastic matrix Q is known or can be accurately
estimated from the data. In addition, while achieving reduction in bias, the proposed
estimators may still be inconsistent and may have large mean squared errors.
The classical papers [20, 21, 22] and the later note [53] consider the situation of permuted data under the term “broken sample”. A broken sample is a sample of (yi , xi )-pairs
that up to some permutation of the {xi } (or equivalently the {yi }) are generated from their
joint distribution. In other words, each component of (yi , xi ) is observed separately (as if it
were generated from its marginal distribution), with possibly different orders for each component. Assuming that the {(yi , xi )} are generated from a bi-variate normal distribution
up to a permutation, these papers discuss recovering the permutation or estimating the correlation parameter of the underlying bivariate normal distribution. In [5, 15], the authors
discuss whether that parameter can be consistently estimated from a broken sample.
In recent years, computer science and engineering have witnessed a surge of interest
in regression analysis of permuted data resulting from problems in multi-target tracking,
statistical seriation [25], and unlabeled compressed sensing [50], to mention just a few.
The papers [39, 40, 50] are particularly important as they provide rigorous results on
fundamental questions associated with the problem. We shall refer to some of these results
in the subsequent sections in more detail. While the paper [17] is not concerned with a
4
regression setting, it provides a detailed analysis of the problem of finding correct matches
between two sets of objects in the presence of noise, which bears some relation to the
problem discussed in §2.4 below.
1.4
Summary of Contributions
The analysis in this paper concerns estimators of Π∗ and β ∗ given model (1) under the
i.i.d.
additional assumption xi ∼ N (0, Id ), i ∈ [n], as also assumed in the recent work [39]. It
i.i.d.
is not hard to extend the results to the case where xi ∼ N (0, Σ), i ∈ [n], where Σ is a
d × d symmetric positive definite matrix.
We first provide a negative result concerning the least squares approach (3) if k = n,
i.e., if no constraint is imposed on the permutation Π. It is shown that optimizing over
all possible choices of Π leads to overfitting in the sense that if β ∗ = 0, one still has
b 2 = Ω(1) with high probability. This result complements another one of a similar spirit
kβk
2
in the recent paper [3] who show that for d = 1, the least squares estimator (3) converges
almost surely to a limit different from β ∗ . Our result also aligns well with a minimax lower
bound in [40].
Altogether, these negative results provide additional justification to consider the regime
of sparse permutations with k n. We demonstrate a bound on the estimation error
kβb − β ∗ k2 with βb as in (3). Specifically, the bound implies that the error vanishes as
both d/n and k log(n/k)/n → 0. In view of the fact that the optimization problem (3)
is not computationally tractable, we consider a convex relaxation that takes the form
of a robust regression estimation problem as it has been considered before in different
contexts [26, 38, 48], with the permuted observations here being in correspondence to gross
errors. The robust regression formulation can be reduced to a specific sparse regression
problem in an underdetermined setting with n − d samples and n parameters, one for each
observation in a given sample that is affected by partial mismatches between the {xi }ni=1
and the {yi }ni=1 . Our analysis of the robust regression problem applies generically beyond
the specific setting of sparsely permuted data. Prior works [26, 38] have considered a more
general version of the problem in which β ∗ is assumed to be sparse as well, thereby being
able to cover the regime n < d, but it is not clear whether the results in [26, 38] can be
specialized to match those of the present paper for the traditional n > d case.
While the robust regression formulation gives rise to an error bound for estimating
∗
β that is comparable to that of the computationally hard formulation (3), it does not
immediately yield an estimator for the permutation Π∗ . We address this issue by adopting
a two-stage approach that uses an accurate estimator θb of β ∗ to match the fitted values
n
bn
{x>
i θ}i=1 to the responses {yi }i=1 . This reduces to simple sorting operations, thereby
avoiding the computational challenges associated with problem (3). We show that our
approach recovers the underlying permutation under qualitatively the same condition as
in [39] which is considerably more stringent in terms of required signal-to-noise ratio than
what is required for accurate estimation of the regression coefficients β ∗ . We complement
our result with a comparable lower bound on the signal-to-noise ratio that is required for
permutation recovery even if β ∗ itself is known.
As already pointed out in [39], the Hamming ball constraint in (3) does not substantially
change the fundamental statistical limits of permutation recovery. However, that constraint
does help in that it gives rise to a computationally efficient estimator of Π∗ , whereas the
statistical achievability result in [39] is for the computationally hard estimator (3).
5
Π∗
Theorem 3
β∗
Π∗ Xβ ∗ − Xβ ∗
Theorem 1
Theorem 2
Lemma 2 (B)
Lemma 6 (D)
S ∗ = {i : Π∗ii 6= 1}
Corollary 1
Figure 3: Schematic overview on targets in the setting of linear regression with sparsely
permuted data and pointers to corresponding results in the paper. The directions of
indicate that if the object on the left hand side is known or can be accurately estimated,
it immediately yields or simplifies estimation of the object on the right hand side.
2
Main results
This section is structured as follows. We start by showing that the estimator βb in (3) lacks
statistical consistency if k = n. This sets the stage for considering the scenario k n. We
first focus on recovery of β ∗ , treating Π∗ as a nuisance parameter. We then present and
analyze a post-processing step for recovering Π∗ given an accurate estimate of β ∗ .
In the sequel, unless stated otherwise, we assume Gaussian design for the {xi }ni=1 :
(G) :
i.i.d.
xi ∼ N (0, Id ), i ∈ [n].
i.i.d.
Our results generalize to the case xi ∼ N (0, Σ), i ∈ [n], for a symmetric positive definite
matrix Σ ∈ Rd×d in a straightforward manner by defining a new regression parameter
Σ1/2 β ∗ . An estimator of that parameter (as discussed in the sequel) and an estimator of Σ
can then be combined to form an estimator of β ∗ . Note that estimation of Σ is not affected
by the presence of an unknown permutation.
2.1
Least squares estimation of (Π∗ , β ∗ ) without additional constraints:
a negative result
Let us consider problem (3) for k = n, i.e., no further constraints are imposed on the
b It turns out that in this case, we cannot hope for βb to be a good estimator
b β).
solution (Π,
∗
of β . As can be seen from the proof of the following proposition, complete freedom in
choosing Π to fit the data results into over-adaptation to noise even if β ∗ is low-dimensional
– in fact, the phenomenon already manifests itself for d = 1.
Proposition 1. Let β ∗ = 0 and Π∗ = In , i.e., y = and consider the estimator (3) with
k = n. Then there exist constants c, C > 0 such that with probability at least 1−C exp(−cn)
b 2≥
kβk
2
n
σ2
.
2n + d 32π 2
b 2 will be bounded away from zero by a constant
In other words, in a pure noise setting, kβk
(assuming d = O(n)). A result of a similar flavor is shown in [3]. For d = 1, they show
that the estimator βb converges almost surely to a limit different from β ∗ as n → ∞, and
they derive an explicit expression for that limit. Our result here is more of a qualitative
nature, but it is non-asymptotic and not limited to d = 1.
Both the result in [3] as well as Proposition 1 raise the question whether there exist
alternative estimators that do significantly better in a regime where Π∗ can be an arbitrary
6
element of Pn . Theorem 1 in [40] indicates that the answer is negative: they show that for
e one has
e β),
any estimator (Π,
h
i
e βe − Π∗ Xβ ∗ k2 ≥ cσ 2 .
sup
E n1 kΠX
2
Π∗ ∈Pn , β ∗ ∈Rd
While the above lower bound concerns estimation of Π∗ Xβ ∗ rather than β ∗ , accurate insample prediction (or synonymously denoising), i.e., recovery of Π∗ Xβ ∗ , is typically easier
than recovery of β ∗ . Along this direction, another negative result is shown in [32]. They
demonstrate that a lower bound on the signal-to-noise ratio (SNR)
SNR = kβ ∗ k22 /σ 2
(5)
of the order Ω(d/ log log n) is required for any estimator of β ∗ to achieve a non-trivial
expected relative estimation error1 . As shown in [39], the condition SNR > nc for c large
b of (3) to recover Π∗ , in which case kβb − β ∗ k2 scales
enough is sufficient for the solution Π
2
as O(d/n) as in the usual regression setting in the absence of an unknown permutation.
2.2
Least squares estimation of β ∗ if Π∗ is k-sparse, k n
In summary, the previous section points to the fact that we cannot hope for accurate
estimation of β ∗ without additional assumptions on Π∗ and/or the SNR of the problem.
As motivated in the introduction, we henceforth turn our attention to the case of Π∗ being
k-sparse, with k “significantly smaller” than n. The allowable range of k is addressed in
our analysis presented below.
We start by fixing notation. For 0 ≤ k ≤ n, let us introduce the shorthand
Pn,k = {Π ∈ Pn : dH (Π, In ) ≤ k}
for the constraint set of Π in problem (3). Moreover, for a compact and symmetric2 set
K ⊂ Rn , its Gaussian width is defined by
w(K) = Eg∼N (0,In ) sup | hg, xi | .
(6)
x∈K
While originating in geometric analysis, the Gaussian width is a measure of complexity that
has been increasingly adopted in the analysis of high-dimensional linear inverse problems
[13, 16, 42] in connection with Gordon’s ”Escape Through a Mesh Theorem” [27], which
is the key component in the proof of Theorem 1 below as well. In our setting, we use the
Gaussian width (6) in conjunction with the set
o
[ n
T =
range(Π − Π∗> ) ∩ Sn−1 ,
(7)
Π∈Pn,k
where we recall that Π∗> ∈ Pn,k is the inverse of Π∗ . Let k·k0 denote the `0 -”norm”, i.e.,
the number of non-zero entries of a vector. A simple observation is that for any Π ∈ Pn,k
and any v ∈ Rn , it holds that
kΠv − Π∗> vk0 = k(Π − In )v − (Π∗> − In )vk0 ≤ k(Π − In )vk0 + k(Π∗> − In )vk0 ≤ 2k.
1
2
Non-trivial here refers to a relative estimation error of lower order than that of the estimator θb ≡ 0.
A set S ⊂ Rn is called symmetric if x ∈ S implies that −x ∈ S.
7
As a result, T ⊆ B0 (2k; n) ∩ Sn−1 , where for 0 ≤ r ≤ n, the set B0 (r; n) = {v ∈ Rn :
kvk0 ≤ r} denotes the `0 -”ball” of radius r in Rn . By well-known results (cf. [41], Lemma
2.3), we hence have that
p
w(T ) ≤ w(B0 (2k; n) ∩ Sn−1 ) ≤ 3.5 2k log(en/2k).
(8)
Moreover, it is not hard to show that w(T ) ≥ w(B0 (n − k; k) ∩ Sn−1 ), i.e., there is a lower
bound on w(T ) of the same order. After these preparations, we are in position to state
the following result.
Theorem 1. Consider
optimization
h
i problem (3) and for positive integers m, denote νm =
m
1/2
Eg∼N (0,Im ) [kgk2 ] ∈ (m+1)1/2 , m
and let ε ∈ (0, 1/2) be a number such that
νn−d −
2
ε
νn ≥
w(T ).
1−ε
1−ε
(9)
If furthermore n > 9 ∨ 4d, then with probability ≥ 1 − 72 exp(−cε n) − 2 exp − 12 (d ∨ log n) ,
!
r
√ −2
σ
5(d
∨
log
n)
2(1
+
2)ε
{w(T
)
∨
log
n}
∗
q
√
kβb − β k2 ≤
+
.
(10)
n
n
4d∨log n
1−
n
We start our interpretation of Theorem 1 by inspecting condition (9) which imposes a limit
on how large d and k can be in relation to n. Roughly speaking, the requirements are n =
Ω(d) and n = Ω(w2 (T )). In light of (8), the latter condition becomes n = Ω(k log(n/k)).
Specifically, let us fix ε = 1/4 and d = αn for α ∈ (0, 14 ), then (9) essentially evaluates as
2
8
√
n≥
w2 (T ).
3( 1 − α − 1/3)
The error bound (10) consists of two parts. The first part equals the error one would
have if the permutation Π∗ were known in advance and is thus inevitable. The second
term is a bound on the excess error incurred for not knowing Π∗ . That term is well
controlled as long as long as the fraction of permuted observations is small relative to n. A
b 2 . Under
crucial intermediate step in the proof of Theorem 1 is a bound on kΠ∗> y − Πyk
b identifies
an additional condition on the SNR, we may deduce from that bound that Π
∗
∗
∗
the “support” S = {i : Πii 6= 1} of Π . We may then re-fit with the corresponding
observations left out, to achieve a smaller error in estimating β ∗ .
Corollary 1. For any δ ∈ (0, 1), under the conditions of Theorem 1, if it additionally
holds that
√
2(1 + 2)2 ε−4 k 2 {w(T ) ∨ log n}2
SNR >
,
n − k ≥ 9 ∨ 4d,
·
δ2
n
the following events hold with probability at least 1 − δ − 2 exp(− 21 (d ∨ log n)):
r
σ
5(d ∨ log n)
bb
∗
∗
b
b
q
S := {i : Πii 6= 1} = S ,
kβ − β k2 ≤
.
n−k
n
1 − 4d∨log
n−k
b
b
where βb denotes the ordinary least squares estimator based on data {(yi , xi ) : i ∈ [n] \ S}.
Under the conditions of the corollary, as long as n − k = Ω(n), the error rate in estimating
β ∗ is of the same order as if Π∗ were known in advance; the second term in (10) gets
eliminated. It is important to note that support recovery (i.e., {Sb = S ∗ }) is a weaker
b
result than permutation recovery (i.e., {Π∗ = Π}).
As discussed in more detail in §2.4
below, the latter requires a considerably more stringent condition on the SNR than what
is required in Corollary 1.
8
2.3
Convex relaxation
While the approach of the previous section has appealing statistical properties, its computational hardness asks for computationally efficient alternatives with similar guarantees.
As long as Π∗ is treated as a nuisance parameter, we may eliminate it in the following way.
Introducing f ∗ = Π∗ Xβ ∗ − Xβ ∗ = (Π∗ − In )Xβ ∗ , model (2) can be re-expressed as
y = Xβ ∗ + f ∗ + .
This prompts the optimization problem
ky − Xβ − f k22
[
subject to f ∈
range(Π − In ).
min
β∈Rd ,f ∈Rn
Π∈Pn,k
A first relaxation is given by
min
β∈Rd ,f ∈Rn
ky − Xβ − f k22
(11)
subject to kf k0 ≤ k.
P
We note in passing that one could additionally impose the constraint ni=1 fi = 1. However, it turns out that its addition does not yield significant statistical benefits, and it is
thus omitted. Relaxation (11) is still not convex, but following the standard approach of
replacing the `0 -norm by the `1 -norm, we end up with the convex optimization problem
min
β∈Rd ,f ∈Rn
ky − Xβ − f k22
(12)
subject to kf k1 ≤ b.
Since it tends to be difficult to choose b appropriately in practice, it is more convenient to
√
work with the Lagrangian form of (12). After re-parameterizing e = f / n3
min
β∈Rd ,e∈Rn
√
1
ky − Xβ − nek22 + λkek1 ,
n
λ > 0.
(13)
Formulation (13) and variants thereof have been used in robust regression and signal recovery with gross corruptions (e.g., [26, 36, 38, 48]). In fact, (13) is equivalent to employing
Huber’s loss [33] instead of squared loss, cf. [48] and the references therein. The connection
to robust regression comes naturally as observations with mismatch between x and y are
likely to induce severe errors in model fitting beyond the usual noise, and hence take the
role of outliers. Indeed, this reasoning could have been used to motivate (13) right away
instead of via the sequence of relaxations stated above. Formulation (13) is related to least
absolute deviation regression
min ky − Xβk1
(14)
β∈Rd
in that (14) is obtained from (13) in the limit λ → 0 and the additional constraint y =
√
Xβ + ne. In that sense, (14) can be seen as the counterpart to (13) in a noiseless setup.
Problem (14) has been analyzed under assumption (G) in the landmark papers of [14, 45]
on the classical error correcting problem in coding theory.
Theorem 2 below provides an upper bound on the `2 -error of the estimator of β ∗
resulting as the optimal β for (13).
3
This reparameterization is merely for technical reasons
9
e ee) be a minimizer of (13) with λ = 4(1 + M )σ
Theorem 2. Let (β,
M > 0. There exist constants c1 , c2 , ε > 0 so that if
k ≤ c1
p
2 log(n)/n for some
n−d
,
log(n/k)
2
the following holds with probability ≥ 1−2 exp(−c2 (n−d))−2n−M −2 exp − 21 (d ∨ log n) :
!
r
r
5(d
∨
log
n)
2k
log
n
σ
n
q
.
(15)
kβe − β ∗ k2 ≤
+ 48(1 + M )
ε−1
n
n
−
d
n
4d∨log n
1−
n
Comparing the upper bounds (10) and (15) of the original problem and its relaxation,
respectively, we observe a close agreement given that w(T ) = Θ(k log(n/k)). Apart from
the slight change of the order in the second term, only constants differ. Similarly to
Corollary 1, we may consider a two-stage procedure in order to further improve upon (15):
given ee and a suitable threshold t, we let Set = {i : |e
ei | ≥ t} and obtain a plain least squares
e
e With k assumed to be known, we may
fit based on data {(yi , xi ) : i ∈ [n] \ Set } yielding β.
take t as the k-th largest largest entry of ee in absolute magnitude. The formal analysis is
similar to Corollary 1 and is hence omitted.
2.4
Estimation of Π∗ given an estimator of β ∗
Having discussed estimation of β ∗ , we now come back to the problem of estimating Π∗ . As
indicated above, the latter problem turns out to be significantly more challenging than the
former problem. In the sequel, we study the question of how the availability of an accurate
estimate of β ∗ can be leveraged to construct an estimator of Π∗ that is computationally
feasible including the case d > 1. As mentioned earlier, except for d = 1, the optimization
problem in Eq. (3) is in general NP-hard. When d = 1, that is there is only one predictor
b is determined as a minimizer of the optimization problem
in the regression model, Π
max {hΠX, yi , hΠX, −yi}
subject to dH (Π, In ) ≤ k.
Π∈Pn
(16)
For k = n, maximizing each term inside the curly brackets is a specific instance of a linear
assignment problem [12], a class of problems that can be solved in polynomial time despite
their combinatorial nature. In fact, it easy to see that
max hΠX, yi =
Π∈Pn
n
X
X(i) y(i) ,
(17)
i=1
where X(i) and y(i) denote the i-th order statistic of X and y, respectively, i ∈ [n]. Hence,
for k = n, problem (16) reduces to two vanilla sorting operations. At this point, it is not
well understood yet whether the Hamming constraint dH (Π, In ) ≤ k for general k causes
problem (16) to be NP-hard again. So far, we are not aware of any computationally efficient
algorithm for general k.
Note that when β ∗ is known, computing the least squares estimator of Π∗ reduces to
solving precisely one of the two optimization problem encountered in (16) for d = 1:
max hΠXβ ∗ , yi
Π∈Pn
subject to dH (Π, In ) ≤ k.
(18)
At this point, a natural idea is to replace β ∗ by an accurate and computationally feasible
estimator like the one discussed in §2.3. From a computational point of view, this already
10
constitutes a simplification as (18) reduces to an integer linear program; while problems
in this class are still NP-hard in general, problem (18) can be considered as much more
benign than the original problem (3) which belongs to the class of quadratic assignment
problems notorious for their computational hardness. Due to recent advances in integer
programming that have meanwhile been taken advantage of for a series of other statistical
problems [7, 8], it turns out that problem (18) is practically feasible at least for n in the
order of a few thousands. For large n, we instead recommend estimating the support of Π∗
and then solve the unconstrained problem (17) restricted to observations in the estimated
support. Formally, denote by Se an estimator of the support of Π∗ and let ySe and XSe
e
be the sub-vector of y and the row submatrix of X corresponding to observations in S,
∗
e
respectively. We then estimate Π by Π defined by
!
ee e
Π
S,S
e =
ΠX
X,
(19)
In−|S|
e
where
e eX =
Π
S
XSe
,
XSec
e e e = argmax ΠX e, y e .
and Π
S,S
S
S
Π∈P|S|
e
We now turn to the statistical limits of permutation recovery, including a lower bound on
the SNR in the idealized case with known β ∗ . For simplicity, the theorem stated below is
for the case k = n, but it generalizes to k < n conditional on having {Se = S ∗ } in that all
expressions in n below would get replaced by k.
b
Theorem 3. For b ∈ Rd , consider Π(b)
= argmaxΠ∈Pn hΠXb, yi.
(a) Let δ, ∆, η > 0, and let θb be an estimator s.t. the event {kX θb − Xβ ∗ k∞ ≤ σ∆} holds
with probability at least 1 − η. If
kβ ∗ k22
n(n − 1) 2
n2 (n − 1)2
b 6= Π∗ ) ≤ 2δ + η.
b θ)
∆
+
2
log
>
, then P(Π(
σ2
4δ 2 π
δ
(b) Suppose that n ≥ 5. If
kβ ∗ k22
b ∗ ) 6= Π∗ ) ≥ .35.
< Cn2 , for some constant C > 0, then P(Π(β
σ2
e 4 ) is a
If β ∗ is known, specializing part (a) to the case ∆ = 0 asserts that SNR = Ω(n
sufficient condition for permutation recovery. This is to be compared to a result in [39]
b in (3) recovers Π∗ if SNR = Ω(nc ), where the constant c > 0
stating that the estimator Π
is not specified. Next, let us now turn to the case where β ∗ is substituted by an estimator
b Using standard concentration arguments, one shows that
θ.
p
√
kX θb − Xβ ∗ k∞ ≤ max kxi k2 kθb − β ∗ k2 ≤ C( d + log n)kθb − β ∗ k2
1≤i≤n
holds with high probability. Specifically, considering the estimator βe in §2.3, Theorem 2
√
then yields kX βe − Xβ ∗ k∞ ≤ Cσ n for k small enough, with high probability. Inserting
√
e 5 ). We stress again
∆ = C n into part (a) then results into the requirement SNR = Ω(n
e
b
e
that as opposed to Π, the estimator Π(β) is computationally appealing as it is obtained
from a quadratic program (13) and subsequent sorting.
Finally, part (b) provides evidence that the condition SNR = Ω(n2 ) is necessary for
permutation recovery. While the result in (b) concerns a specific estimator, there does not
seem to be much indication for the existence of a substantially better estimator. In [39], it
is shown that SNR = Ω(n8/9 ) is necessary for any estimator.
11
3
Numerical results
3.1
Simulated data
Below, we present the results of two synthetic data experiments that are meant to serve
as illustration of the developments in the previous section. In the first set of experiments, we generate n = 200 observations from model (2) under (G) with d = 10,
σ ∈ {.01, .02, .05, .1, .2, .5, 1}, and k/n ∈ {.01, .02, .05, .1, .15, .2, . . . , .5}, where the support
of Π∗ is selected uniformly at random. The parameter β ∗ is generated from the uniform
distribution on Sd−1 . For each pair (σ, k/n), we conduct 100 independent replications.
We compare the following estimators:
(i) The naive estimator (ordinary least squares estimator) of β ∗ that ignores the fact
that a fraction of the data is mismatched; this corresponds to the choice k = 0 in (3).
p
e ee) of §2.3 with the choice λ = 0.2σ log(n)/n.
(ii) The estimator (β,
(iii) A thresholded version of the estimator in (ii) that discards those observations corresponds to the largest k elements of {|e
ei |, i ∈ [n]}, and performs a least squares re-fit
using the remaining n − k observations.
e ee) is computed in CVX [28].
The estimator (β,
The estimator βb in §2.2 is compared to βe in a second set of simulations only for the
case d = 1. In that case, computation of βb reduces to (16) where each of the two inner
optimization problems can be cast as an integer linear program with n2 variables and 2n+1
linear constraints. It turns out that the general purpose routine cplexbilp in IBM CPLEX
[1] is able to solve such problems in only a few seconds for n = 200. Again, such reduction
is limited to the case d = 1.
-2
-4
-6
-8
0
0.1
0.2
0.3
0.4
(iii)
log2(estimation error)
(ii)
log2(estimation error)
log2(estimation error)
(i)
-2
-4
-6
-8
0
0.1
0.2
k/n
0.3
k/n
0.4
-2
-4
-6
-8
0
0.1
0.2
0.3
0.4
k/n
Figure 4: Comparison of the average `2 -errors (log2 -scale) in recovering β ∗ for the approaches (i) to (iii) described above. Each curve corresponds to a different value of the
noise level σ. The black dashed lines correspond to the oracle (ordinary least squares with
knowledge of Π∗ ).
Inspection of Figure 4 shows that the approach (ii) improves dramatically over the naive
estimator (i) as long as the SNR is not too small (for σ = 1 and σ = 0.5, there is no longer a
visible improvement). The results look promising in that the tolerable fraction of permuted
observations can be as much as 0.5 as long as the noise level is small; ”tolerable” here
refers to the fact that the estimation error increases gently with the fraction of permuted
observations as opposed to (i) with a sharp increase in error as k/n moves away from zero.
12
Approach (iii) appears to yield further improvements over (ii) for large SNR. For small
SNR, (ii) and (iii) are not distinguishable. This observation aligns well with Corollary 1.
Figure 5 reveals that for d = 1, the estimator βb performs substantially better than the
-2
-3
-4
-5
-6
-7
-8
-9
-10
-11
0
βe
log2(estimation error)
log2(estimation error)
βb
0.1
0.2
0.3
0.4
-2
-3
-4
-5
-6
-7
-8
-9
-10
-11
0
k/n
0.1
0.2
0.3
0.4
k/n
Figure 5: Comparison of the average `2 -errors (log2 -scale) in recovering β ∗ for the estimators βb and βe for d = 1. Each curve corresponds to a different value of the noise level σ.
The black dashed lines correspond to the oracle (least squares with knowledge of Π∗ ).
e We observe that βb is on par with the oracle for σ < 0.2,
convex formulation yielding β.
being hardly effected by an increase in k/n. That starts changing for σ = 0.2, while for
σ = 0.5, the adverse effect of an increase in k/n becomes clearly visible. At this point, it
is not clear whether similar improvements of βb over βe would be observed for d > 1. The
analysis in §2 only provides (comparable) upper bounds for both approaches, and lower
bounds would be needed to answer that question. The observation made from Figure 5
hence points towards an important open direction.
3.2
Real data
The El Niño data set in the UCI Machine Learning Repository [2] contains oceanographic
and surface meteorological readings taken from a series of buoys positioned throughout
the equatorial Pacific. The data set consists of n = 93, 935 records with the following attributes: buoy identifier, date, location (latitude and longitude), zonal and meridional
wind speeds (zon and mer), relative humidity (humidity), air temperature (air.temp),
sea surface temperature and subsurface temperatures down to a depth of 500 meters
(s.s.temp). The following linear regression model is considered:
air.temp = β0 + βzon · zon + βmer · mer+
+ βhumidity · humidity + βs.s.temp · s.s.temp + .
(20)
The results of least squares regression indicate an excellent fit of this model with an Rsquared of about 0.9. With the goal to mimic the situation of data merging based on nonunique identifiers, the data set is divided into two data sets A and B, with A containing the
response variable and B containing the predictors. The variables latitude and longitude
are maintained in both data sets, and used as quasi-identifiers for record linkage of A and
B with the R package fastLink [23]. Since (latitude, longitude)-pairs are generally
associated with multiple observations, the linkage process is not free of mismatches. The
merged data set in turn follows the format depicted in Figure 2 with a permutation affecting
13
interc
zon
mer
humidity
s.s.temp
RMSE
`2 -dist.
∅-hold-out error
βboracle
βbols
5.15
−.056
−.031
−.022
.844
.509
0
.260 (4.6·10−3 )
6.72
−.037
−.045
−.017
.774
.771
1.57
.276 (4.9·10−3 )
βe
5.74
−.044
−.037
−.016
.806
.773
.59
.267 (4.8·10−3 )
Table 1: Results of the real data analysis. Here, βboracle denotes the least squares estimator
based on the original data set, βbols denotes the naive least squares estimator, and βe denotes
the estimator from formulation (13). The columns labeled zon, . . ., s.s.temp contain the
regression parameter estimates for the respective variable; interc refers to the intercept.
The three rightmost columns contain the root mean square errors (RMSEs) on the original
e the `2 -distance to βboracle , and the average
(βboracle ) respectively the merged data set (βbols , β),
hold-out errors (standard errors in parentheses), respectively.
the correspondence between responses and predictors. As shown in Table 1, least squares
regression with the thus merged data set leads to an increase of the residual sum of squares
by 27%. In addition, the estimates of the regression parameters of the linear model (20)
change noticeably. As an alternative, we consider the approach (13) studied in detail
√
herein, where the parameter λ is chosen as λ = 2cb
σ / n with c = 1.345 and σ
b being a
robust estimator of the residual standard error. More specifically, we use the function
rlm (short for “robust linear model fitting”) in the R package MASS [44] with its default
arguments which is in exact correspondence to the above choice of λ, as follows from the
connection between linear regression with the Huber loss and formulation (13) (cf., e.g.,
Proposition 3.1 in [48]). The figures in Table 1 indicate that this approach provides some
remedy relative to the naive use of the least squares estimator based on the merged data
set. The estimates of the regression parameters from (13) reduce the gap to those obtained
with the original data set by about a factor of one half in terms of `2 -distance. Moreover,
approach (13) also yields a better fit in terms of the mean squared prediction error on a
collection of twenty hold-out sets obtained from leaving out different random subsets of
10% of the observations; in each of those twenty runs, the hold-in data set is split and
subsequently re-linked in the same manner as described above. The reduction in excess
risk (relative to the least squares estimator) achieved by (13) in comparison to the naive
least squares solution is again a factor of roughly one half.
4
Conclusion
Regression problems in which the correspondence between responses and predictors has
been lost can be traced back to classical work in statistics and naturally arises in the context of record linkage. Despite its long history, the problem has seen revived interest lately
and has been analyzed through the lens of modern non-asymptotic statistical theory. An
additional layer of complexity to be dealt with is the computational hardness of (quadratic)
assignment problems. In the present paper, both aspects are taken into account under the
assumption of sparsely permuted data. One central insight of our work is that the availability of an accurate estimator of β ∗ may be helpful in circumventing the computational
barrier associated with permutation recovery. On the other hand, preliminary numerical
results for d = 1 indicate a gap in statistical performance between a presumably hard
formulation capturing precisely the notion of a sparse permutation and a computationally convenient relaxation that introduces a loss of information. This observation deserves
further investigation.
14
Acknowledgments
The second author would like to thank William Winkler for his guidance and Anindya Roy
for helpful discussions related to this paper.
References
[1] IBM ILOG CPLEX Optimization
marketplace/ibm-ilog-cplex.
Studio.
http://www.ibm.com/us-en/
[2] UCI Machine Learning Repository. https://archive.ics.uci.edu/ml/datasets.
html.
[3] A. Abid, A. Poon, and J. Zou, Linear Regression with Shuffled Labels.
arXiv:1705.01342, 2017.
[4] R. Adler and J. Taylor, Random Fields and Geometry, Springer, 2007.
[5] Z. Bai and T. Hsing, The broken sample problem, Probability Theory and Related
Fields, 131 (2005), pp. 528–552.
[6] R. Baraniuk, M. Davenport, R. DeVore, and M. Wakin, A Simple Proof of
the Restricted Isometry Property for Random Matrices, Constructive Approximation,
28 (2006), pp. 253–263.
[7] D. Bertsimas, A. King, and R. Mazumder, Best subset selection via a modern
optimization lens, The Annals of Statistics, 44 (2016), pp. 813–852.
[8] D. Bertsimas and R. Mazumder, Least quantile regression via modern optimization, The Annals of Statistics, 42 (2014), pp. 2494–2525.
[9] P. Bickel, Y. Ritov, and A. Tsybakov, Simultaneous analysis of Lasso and
Dantzig selector, The Annals of Statistics, 37 (2009), pp. 1705–1732.
[10] M. Bohensky, Methodological Developments in Data Linkage, John Wiley & Sons,
Ltd, 2015, ch. Bias in data linkage studies.
[11] S. Boucheron, G. Lugosi, and P. Massart, Concentration Inequalities: A
Nonasymptotic Theory of Independence, Oxford University Press, 2013.
[12] R. Burkard, M. Dell’Amico, and S. Martello, Assignment Problems: Revised
Reprint, SIAM, 2009.
[13] T. Cai, T. Liang, and A. Rakhlin, Geometric inference for general highdimensional linear inverse problems, The Annals of Statistics, 44 (2016), pp. 1536–
1563.
[14] E. Candes and T. Tao, Decoding by Linear Programming, IEEE Transactions on
Information Theory, 51 (2005), pp. 4203–4215.
[15] H.-P. Chan and W.-L. Loh, A file linkage problem of DeGroot and Goel revisited,
Statistica Sinica, 11 (2001), pp. 1031–1045.
[16] V. Chandrasekaran, B. Recht, P. Parrilo, and A. Willsky, The convex
geometry of linear inverse problems, Foundations of Computational Mathematics, 12
(2012), pp. 805–849.
15
[17] O. Collier and A. Dalalyan, Minimax Rates in Permutation Estimation for Feature Matching, Journal of Machine Learning Research, 17 (2016), pp. 1–31.
[18] S. DasGupta, An elementary proof of a theorem of Johnson and Lindenstrauss, Random Structures and Algorithms, 22 (2003), pp. 60–65.
[19] P. David, D. Dementhon, R. Duraiswami, and H. Samet, Softposit: Simultaneous pose and correspondence determination, Int. J. Comput. Vision, 59 (2004),
pp. 259–284.
[20] M. DeGroot, P. Feder, and P. Goel, Matchmaking, The Annals of Mathematical
Statistics, 42 (1971), pp. 578–593.
[21] M. DeGroot and P. Goel, The Matching Problem for Multivariate Normal Data,
Sankhya, Series B, 38 (1976), pp. 14–29.
[22]
, Estimation of the correlation coefficient from a broken random sample, The
Annals of Statistics, 8 (1980), pp. 264–278.
[23] E. Enamorado, B. Eifield, and K. Imai, Fast Probabilistic Record Linkage with
Missing Data. R-package, Version 0.2.0.
[24] I. P. Fellegi and A. B. Sunter, A theory for record linkage, Journal of the American Statistical Association, 64 (1969), pp. 1183–1210.
[25] N. Flammarion, C. Mao, and P. Rigollet, Optimal rates of statistical seriation.
arXiv:1607.02435, 2016.
[26] R. Foygel and L. Mackey, Corrupted Sensing: Novel Guarantees for Separating
Structured Signals, IEEE Transactions on Information Theory, 60 (2014), pp. 1223–
1247.
[27] Y. Gordon, On Milman’s inequality and random subspaces which escape through a
mesh in Rn , Springer Berlin Heidelberg, Berlin, Heidelberg, 1988, pp. 84–106.
[28] M. Grant and S. Boyd, CVX: Matlab software for disciplined convex programming,
version 2.1. http://cvxr.com/cvx, Mar. 2014.
[29] R. Gutman, C. Afendulis, and A. Zaslavsky, A Bayesian Procedure for File
Linking to Analyze End-of-Life Medical Costs, Journal of the American Statistical
Association, 108 (2013), pp. 34–47.
[30] M. H. P. Hof and A. H. Zwinderman, Methods for analyzing data from probabilistic linkage strategies based on partially identifying variables, Statistics in Medicine, 31
(2012), pp. 4231–4242.
[31] D. Hsu, S. Kakade, and T. Zhang, A tail inequality for quadratic forms of subGaussian random vectors, Electronic Communications in Probability, 52 (2012), pp. 1–
6.
[32] D. Hsu, K. Shi, and X. Sun, Linear regression without correspondence.
arXiv:1705.07048, 2017.
[33] P. Huber, Robust Estimation of a Location Parameter, The Annals of Mathematical
Statistics, 53 (1964), pp. 73–101.
16
[34] G. Kim and R. Chambers, Regression analysis under probabilistic multi-linkage,
Statistica Neerlandica, 66 (2012), pp. 64–79.
[35] P. Lahiri and M. D. Larsen, Regression analysis with linked data, Journal of the
American Statistical Association, 100 (2005), pp. 222–230.
[36] J. N. Laska, M. A. Davenport, and R. G. Baraniuk, Exact Signal Recovery
from Sparsely Corrupted Measurements through the Pursuit of Justice, in Asilomar
Conference on Signals, Systems and Computers, 2009, pp. 1556–1560.
[37] J. Neter, S. Maynes, and R. Ramanathan, The effect of mismatching on the
measurement of response error, Journal of the American Statistical Association, 60
(1965), pp. 1005–1027.
[38] N. Nguyen and T. Tran, Robust Lasso with Missing and Grossly Corrupted Observations, IEEE Transactions on Information Theory, 59 (2013), pp. 2036–2058.
[39] A. Pananjady, M. Wainwright, and T. Cortade, Linear Regression with an Unknown Permutation: Statistical and Computational Limits, in Communication, Control and Computing (Allerton), 2016, pp. 417–424.
[40]
, Denoising Linear Models with Permuted Data. arXiv:1704.07461, 2017.
[41] Y. Plan and R. Vershynin, Robust 1-bit compressed sensing and sparse logistic regression: a convex programming approach, IEEE Transactions on Information Theory,
59 (2013), pp. 482–494.
[42]
, The generalized Lasso with non-linear observations, IEEE Transactions on Information Theory, 62 (2016), pp. 1528–1537.
[43] A. B. Poore and S. Gadaleta, Some assignment problems arising from multiple
target tracking, Math. Comput. Model., 43 (2006), pp. 1074–1091.
[44] B. Ripley, B. Venables, D. Bates, K. Hornik, A. Gebhard, and D. Firth,
MASS: Support Functions and Datasets for Venables and Ripley’s MASS. R-package
version 7.3.-47.
[45] M. Rudelson and R. Vershynin, Geometric approach to error correcting codes
and reconstruction of signals, International Mathematical Research Notices, 64 (2005),
pp. 4019–4041.
[46] F. Scheuren and W. Winkler, Regression analysis of data files that are computer
matched I, Survey Methodology, 19 (1993), pp. 39–58.
[47]
, Regression analysis of data files that are computer matched II, Survey Methodology, 23 (1997), pp. 157–165.
[48] Y. She and A. Owen, Outlier Detection Using Nonconvex Penalized Regression,
Journal of the American Statistical Association, 106 (2012), pp. 626–639.
[49] A. Tancredi and B. Liseo, Regression analysis with linked data: problems and
possible solutions, Statistica, 75 (2015), pp. 19–35.
[50] J. Unnikrishnan, S. Haghighatshoar, and M. Vetterli, Unlabeled sensing with
random linear measurements. arXiv:1512.00115, 2015.
17
[51] R. Vershynin, In: Compressed Sensing: Theory and Applications, Cambridge University Press, 2012, ch. ’Introduction to the non-asymptotic analysis of random matrices’.
[52] W. E. Winkler, Matching and record linkage, Wiley Interdisciplinary Reviews: Computational Statistics, 6 (2014), pp. 313–325.
[53] Y. N. Wu, A note on broken sample problem, tech. rep., Department of Statistics,
University of Michigan, 1998.
[54] S. Zhou, Restricted Eigenvalue Conditions on Subgaussian Random Matrices.
arXiv:0912.4045.
A
Proof of Proposition 1
Consider the least squares problem (3) with y = :
mink − ΠXβk22 .
Π,β
Omitting the term not depending on β, an equivalent problem is given by
min −2 hΠ, Xβi + kXβk22 .
Π,β
Re-parameterizing Xβ = u · r, with r ≥ 0 for u ∈ range(X) ∩ Sn−1 , we obtain the optimization problem
min
−2 hΠ, ui r + r2 .
Π,u∈range(X)∩Sn−1 ,r≥0
Observe that for any u and Π, the optimal value of r is found as
rb(u) = max{hΠ, ui , 0}.
b u
In particular, for a minimizer (Π,
b),
b u
b u
rb := rb(Π,
b) = hΠ,
bi
b 2 , and the optimal value of the objective is given by −b
so that rb = kX βk
r2 . We hence have
rb =
max
u∈range(X)∩Sn−1 ,Π
hΠ, ui
≥ max hΠ, Xj /kXj k2 i
1≤j≤d,Π
≥ maxhΠ, X1 i
Π
1
.
kX1 k2
We now bound
maxhΠ, X1 i ≥ hΠ± , X1 i,
Π
where the permutation Π± is defined as follows. Define
n,+ = {i : i > 0},
n,− = n − n,+ ,
nX1 ,+ = {i : Xi1 > 0},
nX1 ,− = n − nX1 ,+ .
Let further
n+ = min{n,+ , nX1 ,+ },
n− = min{n,− , nX1 ,− },
18
n± = n − n+ − n− .
(21)
Denote by
+
i+
1 < . . . < inX,+
the indices of the positive entries of X1 ,
j1+ < . . . < jn+,+
the indices of the positive entries of ,
−
i−
1 < . . . < inX,−
j1− < . . . < jn−,−
the indices of the negative entries of X1 ,
the indices of the negative entries of .
Furthermore, let
±
i±
1 < . . . < in±
j1± < . . . < jn±±
be the indices in
−
+
−
{1, . . . , n} \ ({i+
1 , . . . , in+ } ∪ {i1 , . . . , in− })
respectively {1, . . . , n} \ ({j1+ , . . . , jn++ } ∪ {j1− , . . . , jn−− })
in increasing order. We then construct Π± such that
hΠ± , X1 i =
n+
X
X1i+ j + +
k
n−
X
X1i− j − +
k
k
k
X1ik ± jk ± .
k=1
k=1
k=1
n±
X
All terms inside the first two sums are i.i.d. from the same distribution as |g||h|, where
g and h are independent N (0, 1)-random variables, and accordingly the terms inside the
last summand are i.i.d. from the same distribution as −|g||h|. Conditional on n± , hence
hΠ± , X1 i follows the same distribution as
n−n±
X
n
X
|gi ||hi | −
i=1
|gi ||hi |,
i=n−n± +1
where {gi }ni=1 and {hi }ni=1 are two independent sequences of n i.i.d. random variables from
the N (0, 1)-distribution. The expectation of the above expression (conditional on n± ) is
given by
2
(n − 2n± ) E[|g||h|] = (n − 2n± ) E[|g|]2 = (n − 2n± ) .
π
From Proposition 5.16 in [51], we have the following concentration inequality:
n−n±
n
X
X
1
(|gi ||hi | − E[|g||h|]) −
(|gi ||hi | − E[|g||h|]) ≥ (n − 2n± ) n± ≤ exp(−nc).
P
π
i=1
i=n−n± +1
Hence, conditional on n± , with probability at least 1 − exp(−nc),
n−n±
X
i=1
|gi ||hi | −
n
X
|gi ||hi | ≥
i=n−n± +1
1
(n − 2n± ).
π
Consider the definition of n+ and n− Observe that n,+ , nX1 ,+ , n,− and nX1 ,− are each
binomial random variables with n trials and probability of success 1/2. By repeated application of Hoeffding’s inequality, we hence have
P(n+ ≤ 3n/8) ≤ 2 exp(−n/32),
P(n− ≤ 3n/8) ≤ 2 exp(−n/32).
19
Conditional on the events {n+ > 3n/8} and {n− > 3n/8}, we have n± ≤ 1/4 and thus
with probability at least 1 − C exp(−nc)
n−n±
X
|gi ||hi | −
i=1
n
X
i=n−n± +1
|gi ||hi | ≥
n
.
2π
n
Accordingly, the same applies to the event {hΠ± , X1 i > 2π
}. Lemma 15 in Appendix H
√
yields kX1 k2 ≤ 2 n with probability at least 1 − exp(−n/2). Putting together the pieces,
it follows that with probability at least 1 − C exp(−nc)
√
√
n
b 2 = rb ≥ n ⇒ kβk
b 2≥
kX βk
,
4π
σmax (X)4π
where σmax (X) denotes the
singular value of X. Finally, Lemma 16 in Appendix
√
√ maximum
H yields that σmax (X) ≤ 2n+ d with probability at least 1−exp(−cn), which concludes
the proof.
B
Proof of Theorem 1
b
b and β.
We start with a basic, yet important observation that allows us to decouple Π
⊥
Let PX and PX denote the orthogonal projection on range(X) respectively its orthogonal
complement.
Lemma 1. Consider optimization problem (3). Then, if n ≥ d, with probability one
(
)
> X −1 X > Π
b >y
X
>
⊥
2
>
⊥
2
b ∈ argminkP Πyk ,
b ∈ argminkP Πyk . (22)
Π
βb ∈
, Π
X
2
X
2
n
n
Π∈Pn,k
Π∈Pn,k
Proof. We have
min kΠXβ − yk22 = min kXβ − Πyk22
Π,β
Π,β
n
o
2
2
= min k P⊥
X Πyk2 + kXβ − PX Πyk2 .
Π,β
The second part can always be made equal to zero by choosing β such that Xβ = PX Πy.
It hence suffices to minimize the first part w.r.t. Π and to back-substitute the result into
the second part. Under model (G), X is non-singular with probability one as long as n ≥ d
which yields the second expression in (22).
b > y.
In the next lemma, we bound the `2 -distance between Π∗> y and Π
Lemma 2.
Under the
conditions of Theorem 1, the following holds with probability at least
2
ψ(ε,d/n)νn
7
− exp(−{w(T ) ∨ log n}), where ψ(ε, d/n) is defined in (24) below.
1 − 2 exp −
8
√
2(1 + 2)σε−2 {w(T ) ∨ log n}
>
∗>
b
√
kΠ y − Π yk2 ≤
.
n
b > and the fact that Π∗> is a feasible solution,
Proof. In view of Lemma 1, by definition of Π
we have
⊥ ∗>
2
b> 2
kP⊥
X Π yk2 ≤ kPX Π yk2
∗>
⊥ ∗>
2
⊥ ∗>
2
b>
⇒ kP⊥
X (Π − Π )y + PX Π yk2 ≤ kPX Π yk2
∗>
2
⊥ ∗>
∗>
b>
b>
⇒ kP⊥
X (Π − Π )yk2 ≤ 2|hPX Π y, (Π − Π )yi|.
20
Since
∗>
⊥
∗
⊥
P⊥
X Π y = PX (Xβ + ) = PX ,
we have the implication that
∗>
2
⊥
∗>
b>
b>
kP⊥
X (Π − Π )yk2 ≤ 2|hPX , (Π − Π )yi|.
∗>
b>
Dividing both sides by kP⊥
X (Π − Π )yk2 , we obtain that
⊥
∗>
b > − Π∗> )yk22 kP⊥
b 2
b
b>
k(Π
X ∆k2 ≤ 2|hPX , ∆i|k(Π − Π )yk2 ,
⇒
b :=
∆
⊥
b > − Π∗> )yk2 kP⊥
b 2
b
k(Π
X ∆k2 ≤ 2|hPX , ∆i|.
b > − Π∗> )y
(Π
b > − Π∗> )yk2
k(Π
(23)
i) Lower bounding the left hand side of (23).
b ∈ T as defined in (7). Moreover, note that under (G), the random
Observe that ∆
subspace range(X) ⊆ Rn follows the uniform distribution on the Grassmannian G(n, d).
Hence, applying Lemma 13 with K = T , V = range(X) and thus m = n and l = n − d,
we have for any ε ∈ (0, 1/2) in compliance with (9)
⊥
P(dist(T , range(X)) > ε) = P min kPX ∆k2 > ε
∆∈T
!
7
1 (1 − ε)νn−d − ενn − w(T ) 2
≥ 1 − exp −
2
2
3 + ε + ενn /νn−d
!2
ν
(1 − ε) n−d
7
ψ(ε, d/n)νn2
νn − ε
≥ 1 − exp −
, ψ(ε, d/n) :=
,
ε2
2
8
3 + ε + 1−ε
(24)
where the last inequality follows from condition (9). Conditional on the event {kP⊥
X ∆k2 ≥
ε}, which holds with the probability as stated, the left hand side of (23) is lower bounded
b > − Π∗> )yk2 .
by ε2 k(Π
ii) Upper bounding the right hand side of (23).
Consider t > 0 arbitrary, to be chosen later. We have
⊥
⊥
b
P(|hPX , ∆i| ≥ t) ≤ P sup |hPX , ∆i| > t .
∆∈T
⊥
Now note that and P⊥
X are independent. As a result, when conditioning on PX , we have
from Lemma 14 that
⊥
⊥
⊥
⊥
⊥
⊥
P sup |hPX , ∆i| > t PX ≤ P sup |hPX , ∆i| > E sup |hPX , ∆i| PX + τ PX
∆∈T
∆∈T
∆∈T
⊥
⊥
≤ P sup |hPX , ∆i| > σw(T ) + τ PX
∆∈T
2
2
⊥
⊥
≤ exp(−τ /(2σ )), τ := t − E sup |hPX , ∆i| PX
∆∈T
≥ t − σw(T ).
The second equality is a consequence of the Sudakov-Fernique comparison
√ inequality (Theorem 2.2.3 in [4]) and the fact that kP⊥
k
≤
1.
Choosing
t
=
(1
+
2)σ{w(T ) ∨ log n},
X 2
combining this result with i), and putting together the pieces in (23), the assertion of the
lemma follows.
21
In order to complete the proof of Theorem 1, we need one last lemma whose proof is
deferred to Appendix G.
Lemma 3. Let σmin (·) denote the minimum singular value of a matrix. If n ≥ 9 ∨ 4d, it
holds that
(
)
r
p
X > X −1 X >
√
5(d
∨
log
n)}
4d
∨
log
n
√
q
P σmin (X/ n) < 1 −
∪
≥σ
n
n
n
4d∨log n
1
−
2
n
1
≤ 2 exp − (d ∨ log n) .
2
Equipped with Lemmas 2 and 3, we are in position to conclude Theorem 1. Expanding
the expression for βb in (22), we obtain that
βb =
X >X
n
−1
−1 > ∗>
> −1 > b >
X >X
X Π y
X X
X (Π − Π∗> )y
+
n
n
n
n
> −1 >
>
∗>
b
X X
X ( + (Π − Π )y)
= β∗ +
,
n
n
b >y
X >Π
=
n
and thus
kβb − β ∗ k2 ≤
X>X
n
−1
√
X>X
n
≤
1−
2
n
−1
√
≤
>
X
√
n
>
X
√
n
2
n
σ
q
4d∨log n
n
r
X>X
n
−1
>
X
√
n
+
+
b > − Π∗> )yk2
k(Π
√2
1 √
b>
k(Π
σmin (X/ n)
√
n
− Π∗> )yk2
n
5(d ∨ log n) 2(1 +
+
n
√
2)σε−2 {w(T ) ∨ log n}
√
n
with the stated probability by combining Lemmas 2 and 3.
C
Proof of Corollary 1
Observe that the event {Sb = S ∗ } is implied by the event
∗>
>
b
min∗ |yi − yϕ−1 (i) | > kΠ y − Π yk∞ ,
i∈S
which is in turn implied by
b > yk2 =: γ ,
min∗ |yi − yϕ−1 (i) | > kΠ∗> y − Π
i∈S
We now have
P min∗ |yi − yϕ−1 (i) | ≤ γ
i∈S
≤
X
P |yi − yϕ−1 (i) | ≤ γ
i∈S ∗
≤ k max P (|yi − yj | ≤ γ) .
i6=j
22
!
Note that for any i 6= j, yi − yj ∼ N (0, 2(kβ ∗ k22 + σ 2 )). Invoking Lemma 19, we thus obtain
that for any i 6= j
γ
.
P (|yi − yj | ≤ γ) ≤ p
∗
2(kβ k22 + σ 2 )
It hence follows from the previous two displays that for any δ > 0,
P min∗ |yi − yϕ−1 (i) | ≤ γ ≤ δ
i∈S
if
k2 γ 2
.
2δ 2
Substituting γ by the bound from Lemma 2, the above condition becomes
√
2(1 + 2)2 ε−4 k 2 {w(T ) ∨ log n}2
SNR >
·
,
δ2
n
kβ ∗ k22 >
and the first part of the corollary follows from Lemma 2. Turning to the second part,
consider the least squares estimator using the reduced data set {(xi , yi ), i ∈ [n] \ S ∗ } and
b ∗ ) the corresponding least squares estimator. We then have for any δ > 0
denote by β(S
b ∗ ) − β ∗ k2 ≤ δ) = P(kβ(S
b ∗ ) − β ∗ k2 ≤ δ, Sb = S ∗ ) + P(kβ(S
b ∗ ) − β ∗ k2 ≤ δ, Sb 6= S ∗ ),
P(kβ(S
which implies
b
b ∗ ) − β ∗ k2 ≤ δ) − P(Sb 6= S ∗ ).
P(kβb − β ∗ k2 ≤ δ) ≥ P(kβ(S
It thus remains to control the first probability on the left hand side, which can be done
with arguments similar to those used in the last part of the proof of Theorem 1.
D
Proof of Theorem 2
We start with a Lemma that parallels Lemma 1. The reasoning is analogous, and the proof
is hence omitted.
e ee). Then, if n ≥ d, with
Lemma 4. Consider optimization problem (13) with solution (β,
probability one
(
> X −1 X > (y − √ne
√
1 ⊥
X
e)
2
ee ∈ argmin kPX (y − ne)k2 + λkek1 , βe ∈
,
n
n
n
e
o
√
1
2
ee ∈ argmin kP⊥
(y
−
ne)k
+
λkek
1 .
2
n X
e
For our analysis of the optimization problem of Lemma 4, we need the following crucial
lemma whose proof is provided in Appendix F.
n−d
Lemma 5. There exists constants c1 , c2 , ε > 0 so that if k ≤ c1 log(n/k)
, the following event
holds with probability at least 1 − 2 exp(−c2 (n − d)).
1 ⊥ 2
n−d
2
∗
kP vk ≥ ε
kvk2 , v ∈ C(S , 3) ,
n X 2
n
with S ∗ = {j : Π∗jj 6= 1} and C(S ∗ , 3) defined in (31).
23
√
In correspondence with Lemma 2, we first bound ke
e −e∗ k2 , where e∗ = (Π∗ Xβ ∗ −Xβ ∗ )/ n.
Lemma 6. Under the conditions of Theorem 2, with probability at least 1 − 2 exp(−cε (n −
2
d)) − 2n−M , we have
r
n −1 k log n
∗
ε
.
ke
e − e k2 ≤ 48(1 + M )σ
n−d
n
Proof. We first decompose
√ ∗
⊥
∗
⊥ √ ∗
P⊥
ne + e, e := P⊥
X y = PX (Xβ + ne + ) = PX
X .
According to Lemma 4, since ee is a minimizer, we have that
1 ⊥√ ∗
1
ek1 ≤ ke
kPX n(e − ee) + ek22 + λke
k22 + λke∗ k1 .
n
n
Expanding the square on the left-hand side and re-arranging terms, we obtain that
√
1 ⊥√ ∗
kPX n(e − ee)k22 + λke
ek1 ≤ 2| e∗ − ee, e/ n | + λke∗ k1 ,
n
which implies
√
λke
ek1 ≤ 2ke∗ − eek1 ke
/ nk∞ + λke∗ k1 ,
(25)
and in turn
√
λke
e(S ∗ )c − e∗(S ∗ )c k1 = λke
e(S ∗ )c k1 ≤ 2ke∗ − eek1 ke
/ nk∞ + λ(ke∗ k1 − ke
eS ∗ k1 )
√
∗
∗
≤ 2ke − eek1 ke
/ nk +λkeS ∗ − eeS ∗ k1 .
| {z ∞}
=:λ0
Re-arranging, we arrive at
(λ − 2λ0 )ke
e(S ∗ )c − e∗(S ∗ )c k1 ≤ (λ + 2λ0 )ke∗S ∗ − eeS ∗ k1 .
Conditional on the event Λ = {λ ≥ 4λ0 }, we hence have that
ke
e(S ∗ )c − e∗(S ∗ )c k1 ≤ 3ke
eS ∗ − e∗S ∗ k1 .
(26)
Equipped with this is intermediate result, we go back to (25). We first obtain by rearranging that
1 ⊥√ ∗
n(e − ee)k22 ≤ (2λ0 + λ)ke
e − e∗ k1
kP
n X
≤ (2λ0 + λ)4ke
eS ∗ − e∗S ∗ k1
√
≤ (2λ0 + λ)4 kke
eS ∗ − e∗S ∗ k2 ,
where we have used (26) in the second inequality. In order to lower bound the left hand
side, note that (26) can be written as ee − e∗ ∈ C(S ∗ , 3) according to (31). Conditional on
the event in Lemma 5,
n−d ∗
1 ⊥√ ∗
kP
n(e − ee)k22 ≥ ε
ke − eek22 .
n X
n
Combining this with the previous display, we find that
√
n
ke∗ − eek2 ≤ 12ε−1
λ k.
n−d
p
In view of Lemma 18, the choice λ = 4(1 + M )σ 2 log(n)/n for M ≥ 0 guarantees that
2
the event Λ defined above (26) occurs with probability at least 1 − 2n−M . This completes
the proof of the Lemma.
Using Lemmas 2 and 4, the proof of the theorem is completed along the lines of the last
part of the proof of Theorem 1. For brevity, we omit those steps here.
24
E
Proof of Theorem 3
The proof of part (a) of the theorem relies on two lemmas, which are stated and proved
first.
Lemma 7. Let γ, δ > 0. Consider the event
> ∗
> ∗ 2
2
Aγ = min(xi β − xj β ) ≤ γ .
i<j
(27)
Then P (Aγ ) < δ if
n2 (n − 1)2 2
kβ ∗ k22 >
γ .
4δ 2 π
o
S n
∗ − x> β ∗ )2 ≤ γ 2 . Using the fact that x> β ∗ −
Proof. It is clear that Aγ ⊆ i<j (x>
β
i
j
i
∗
∗ 2
x>
j β ∼ N (0, 2kβ k2 ), we have
∗
[
γ
β
γ
≤ (xi − xj )> √
≤√
−√
P (Aγ ) ≤ P
∗
∗
∗k
2kβ
k
2kβ
k
2kβ
2
2
2
i<j
X
γ
γ
≤
P −√
≤g≤ √
, g ∼ N (0, 1)
2kβ ∗ k2
2kβ ∗ k2
i<j
≤
n(n − 1)
γ
√
∗
2
kβ k2 π
Thus, it suffices to require that
by Lemma 19 in Appendix H.
n(n − 1)
γ
√ < δ. Resolving for kβ ∗ k2 , this condition
∗
2
kβ k2 π
becomes
kβ ∗ k22 >
n2 (n − 1)2 2
γ .
4δ 2 π
The following lemma is of interest in its own when the design matrix is fixed and both β ∗
and σ 2 are known or accurate estimates of these quantities are available. Condition (28)
can then be evaluated explicitly, at least after substituting β ∗ and σ 2 by their respective
estimates. The result is close in spirit to those in the framework in [17].
b ∗ ) 6= Π∗ | X < δ if
Lemma 8. Conditional on X, we have P Π(β
2
n(n − 1)
∗
> ∗
min x>
β
−
x
β
> 4σ 2 log
.
i
j
i<j
δ
(28)
b ∗ ),
Proof. Without loss of generality, we may assume that Π∗ = In . By construction of Π(β
∗
∗
>
∗
>
∗
b
it is clear that Π(β ) 6= Π whenever there exists i 6= j such that (yi −yj )(xi β −xj β ) < 0,
∗
> ∗
that is, the order between x>
i β and xj β is different than that of yi , yj . First, conditioning
on the random design matrix X, we have
o
[n
∗
> ∗
b ∗ ) 6= Π∗ | X) = 2 P
P(Π(β
yi − yj > 0 | x>
i β − xj β < 0
i<j
≤2
X
∗
> ∗
P yi − yj > 0 | x>
β
−
x
β
<
0
i
j
i<j
25
∗
> ∗
> ∗
> ∗
2
Since yi − yj | x>
i β − xj β ∼ N (xi β − xj β , 2σ ), using the usual tail bound for the
Gaussian distribution for each term in the above sum, we obtain
2
> β ∗ − x> β ∗
x
X
i
j
b ∗ ) 6= Π∗ | X) ≤ 2
P(Π(β
exp −
2
4σ
i<j
mini<j
≤ n(n − 1) exp −
∗
x>
i β
−
4σ 2
∗
x>
j β
2
.
b ∗ ) 6= Π∗ | X) < δ if
Hence, for any δ > 0, P(Π(β
2
n(n − 1)
∗
> ∗
> 4σ 2 log
min x>
.
i β − xj β
i<j
δ
b ∗ ) from the
Combining Lemmas 7 and 8, part (a) in Theorem 3 q
follows immediately for Π(β
b ∗ ) 6=
previous lemma by invoking Lemma 7 with γ = 2σ log n(n−1) . We then have P(Π(β
δ
Π∗ ) ≤ P(Aγ ) + (1 − P(Aγ )) δ < 2δ under the conditions of the theorem. The general
version of the theorem results from
>b
> ∗
> ∗
> ∗
b
b
min |x>
i θ − xj θ| ≥ min |xi β − xj β | − 2 max |xi (β − θ)|
i<j
i<j
≥
1≤i≤n
∗
min |x>
i β
i<j
−
∗
x>
j β |
− 2σ∆.
∗
b
conditional on the event {kXβ
The assertion follows from an application
q − X θk∞ ≤ σ∆}.
of Lemma 7 with γ = 2σ
log n(n−1)
+∆ .
δ
The proof of part (b) in Theorem 3 requires another lemma.
Lemma 9. Consider the event Aγ in (27). Then, for all n ≥ 5, P (Aγ ) ≥ 0.85 if
kβ ∗ k22 ≤
1 2 2
n γ .
25π
∗
∗
Proof. Since x>
i β /kβ k2 ∼ N (0, 1), we have for g ∼ N (0, 1)
γ
γ
γ
γ
γ
> ∗
√ ,
p := P − √ < x β < √
= P −√
<g< √
≥
2
2
2kβ ∗ k2
2kβ ∗ k2
kβ ∗ k2 2π
(29)
(30)
whenever γ < √32 kβ ∗ k2 , otherwise p > .86 as can be verified by evaluating the Gaussian
cumulative
distribution functionnumerically.
γ
γ
Let B := − √ < x> β ∗ < √ . For i.i.d. observations x1 , . . . , xn , let
2
2
u :=
n
X
I(xi ∈ B) ∼ binom(n, p),
i=1
26
with p as defined in (30). The event Aγ in (27) occurs whenever at least two xi , xj ∈ B,
i.e., the event {u ≥ 2} ⊆ Aγ . Let us suppose for the moment that p ≥ 3.5/n for n ≥ 5.
Then
P (Aγ ) ≥ P (u ≥ 2)
3.5 n−1
3.5 n
3.5
1−
≥1− 1−
−n
n
n
n
> .85 for n ≥ 5.
Using Eq.(30), we now conclude that P (Aγ ) ≥ .85 whenever
holds if
kβ ∗ k22 ≤
γ
kβ ∗ k2
√
2π
≥ 3.5/n. This
1 2 2
n γ .
25π
We finally turn to the proof of part (b) of the theorem. Define indices i0 , j0 ⊂ [n] by the
∗
> ∗
> ∗
> ∗
relation x>
i0 β − xj0 β = mini<j |xi β − xj β |. For given γ > 0, we have
b ∗ ) 6= Π∗ ) ≥ P yi − yj < 0 | 0 < x> β ∗ − x> β ∗ < γ P 0 < x> β ∗ − x> β ∗ < γ
P(Π(β
i0
j0
i0
j0
0
0
∗
> ∗
> ∗
> ∗
= P x>
β
+
σ
−
x
β
−
σ
<
0
|
0
<
x
β
−
x
β
<
γ
×
i0
j0
i0
j0
i0
j0
∗
> ∗
× P 0 < x>
i0 β − xj0 β < γ
∗
> ∗
≥ P (σi0 − σj0 < −γ) P 0 < x>
β
−
x
β
<
γ
i0
j0
γ
∗
> ∗
P 0 < x>
β
−
x
β
<
γ
≥ P g < −√
i0
j0
2σ
√
√
4 πkβ ∗ k2
4 πkβ ∗ k2
> ∗
> ∗
c ,
c P 0 < xi0 β − xj0 β <
≥P g<− √
n
2σn
√
4 πkβ ∗ k2
by choosing γ =
c, for 0 < c ≤ 1.
n
√
4 πkβ ∗ k2
≥ Φ(−.1)(.85), by Lemma 9, for a choice of c s.t. √
c ≤ .1.
2σn
≥ .35,
Rx
where in the penultimate line, Φ(x) = −∞ √12π exp(−t2 /2) dt denotes the Gaussian cub ∗ ) 6= Π∗ ) ≥ .35 whenever
mulative distribution function. Therefore, P(Π(β
some C > 0. This concludes the proof of part (b) of the theorem.
F
kβ ∗ k22
σ2
<
C
n2
for
Proof of Lemma 5
The proof of Lemma 5 is given at the end of this section. The proof depends on several
lemmas collected from the literature which are stated first.
For 1 ≤ s ≤ p, let J (s) denote the set of all index subsets of [p] of size s.
Definition 1. For J ∈ J (s), and α ∈ [1, ∞), define
C(J, α) = {v ∈ Rp : kvJ c k1 ≤ αkvJ k1 }.
27
(31)
Definition 2. [14] Let A ∈ RN ×p and B0 (s; p) = {v ∈ Rp : kvk0 ≤ s}. Define the s-sparse
minimum and maximum eigenvalues associated with A by
λmin (s) =
kAvk22
,
v∈B0 (s;p)\{0} kvk2
2
min
λmax (s) =
kAvk22
.
v∈B0 (s;p)\{0} kvk2
2
max
Let δ ∈ (0, 1). We say that A is an (s, δ)-restricted isometry if
1 − δ ≤ λmin (s) ≤ λmax (s) ≤ 1 + δ.
Definition 3. [9] Let A ∈ RN ×p . We define the (s, α)-restricted eigenvalue by
φ(s, α) = min
min
J∈J (s) v∈C(J,α)\0
kAvk22
.
kvJ k22
(32)
Lemma 10. [9] Suppose that A is a (3s, δ ∗ /4)-restricted isometry for some δ ∗ ∈ (0, 1).
The (s, 3)-restricted eigenvalue then satisfies φ(s, 3) ≥ 1 − δ ∗ .
Lemma 11. [6] Let A be a random of matrix of size N × p satisfying
P((1 − ε)kvk22 ≤ kAvk22 ≤ (1 + ε)kvk22 ) ≤ 2 exp(−N c0 (ε))
(33)
for any ε ∈ (0, 1) and any v ∈ Rp , where c0 (ε) > 0 is a constant depending only on
ε. Then for any δ ∈ (0, 1), there exist constants c1 , c2 > 0 depending only on δ such
that A is a (q, δ)-restricted isometry for any q ≤ c1 N/ log(p/q) with probability at least
1 − 2 exp(−c2 N ).
Lemma 12. (from Lemma 2.2 in [18]) Let P denote the orthogonal projection on an N dimensional subspace of Rp chosen uniformly at random from the Grassmannian G(p, N ).
Then, for any v ∈ Rp and any ε ∈ (0, 1),
p
P (1 − ε)kvk22 ≤
kP vk22 ≤ (1 + ε)kvk22 ≤ 2 exp(−N (ε2 /4 − ε3 /6)})
N
Proof of Lemma 5: the lemma is essentially implied by the fact that under the stated
conditions, the matrix P⊥
X has its (k, 3)-restricted eigenvalue bounded away from zero by
a constant ε as to be shown below. ”Essentially” refers to the fact that it can be shown
(cf. [54]) that a matrix A with restricted eigenvalue φ(k, 3) obeys
[
1
kAvk22 ≥ φ(k, 3)kvk22
∀v ∈
C(J, 3).
32
J∈J (k)
As sketched above, a lower bound on φ(k, 3) follows if A is a (3k, δ ∗ /4) restricted isometry
for some δ ∗ ∈ (0, 1). That property is in turn satisfied with high probability if A satisfies
(33). Applying Lemma 12 with P = P⊥
X , p = n and N = n − d yields the conclusion.
G
Proof of Lemma 3
> −1 >
X
√
We invoke Lemma 17 conditional on X, with X n X
in the role of A, so that
n
> −2 >
X
√ . Observe that
Γ = √Xn X n X
n
> −1 !
X X
d
1
√ , (ii) kΓk2 = 2
√ ,
(i) tr(Γ) = tr
≤ 2
n
σmin (X/ n)
σmin (X/ n)
v
(34)
u
√
> −2 !
u
p
X
X
d
√ .
(iii)
tr(Γ2 ) = ttr
≤ 2
n
σmin (X/ n)
28
In view of the relations in (34), application of Lemma 17 with t = d ∨ log n and comparison
of terms yields that
!
p
> −1 >
5{d ∨ log n}
X X
X
√
√
P
>σ
X ≤ exp(−d ∨ log n).
n
n
σmin (X/ n)
2
Using Lemma 16 for σmin (X) with the choice t =
H
√
d ∨ log n, the assertion readily follows.
Additional lemmas
We here collect various auxiliary results acting as lemmas in the proofs of our main results.
Lemma 13. (Gordon’s Escape through a Mesh theorem [27]) Let K be a closed
subset of the unit sphere in Rm , let νr = Eg∼N (0,Ir ) [kgk2 ], and let ε > 0. If the Gaussian
width (6) of K obeys w(K) < (1 − ε)νl − ενm , then an (m − l)-dimensional subspace V
drawn uniformly from the Grassmannian G(m, m − l) satisfies
!
1 (1 − ε)νl − ενm − w(K) 2
7
P(dist(K, V ) > ε) ≥ 1 − exp −
.
2
2
3 + ε + ενm /νl
Lemma 14. (Concentration of Gaussian processes [11]) Let K be a closed subset of
the unit sphere in Rm with Gaussian width w(K), and let g ∼ N (0, Im ). Then for all t > 0
P sup | hg, xi | ≥ w(K) + t ≤ exp(−t2 /2).
x∈K
Lemma 15. (Concentration of the norm of Gaussian random vectors [51])
√
Let g ∼ N (0, In ). Then for all t > 0, P(kgk2 ≥ (1 + t) n) ≤ exp(−t2 n/2).
Lemma 16.(Concentration of extreme singular values of Gaussian matrices [51])
Let X be an n × d Gaussian matrix with i.i.d. N (0, 1)-entries. Denote by σmin (X) and
σmax (X) the minimum respectively maximum singular value of X. Then for any t > 0
√
√
√
√
P(σmin (X) ≥ n − d − t) ≤ exp(−t2 /2),
P(σmax (X) ≥ n + d + t) ≤ exp(−t2 /2).
Lemma 17. (Concentration of quadratic forms [31])
Let A be a an m × n matrix, Γ = A> A, and g ∼ N (0, In ). Then for all t > 0
p
P(kAgk22 > tr(Γ) + 2 tr(Γ2 )t + 2kΓk2 t) ≤ exp(−t).
Lemma 18. (Maximal inequality for Gaussian random variables)
Let g1 , . . . , gn be zero-mean Gaussian random variables such that max1≤i≤n E[gi2 ] ≤ σ 2 .
√
2
Then for all M > 0, P(max1≤i≤n |gi | > (1 + M )σ 2 log n) ≤ 2n−M .
Lemma 19. (Small ball probability)
Let g ∼ N (0, 1). Then for all t > 0, P(|g| < t) ≤
29
q
2
π
t < t.
| 10math.ST |
Complexity Analysis of Unsaturated Flow in Heterogeneous Media
Using a Complex Network Approach
Hamed.O.Ghaffari, Mamdou Fall and Erman.Evgin
Dept. Civil Engineering, University of Ottawa, Ottawa, Ontario, Canada
Abstract: In this study, we investigate the complexity of two-phase flow (air/water) in a heterogeneous
soil sample by using a complex network theory .Based on the different similarity measurements (i.e.,
correlation, Euclidean metrics) over the emerged patterns from the evolution of saturation of non-wetting
phase of a multi-heterogeneous soil sample, the emerged complex networks are recognized.
Understanding of the properties of complex networks (such degree distribution, mean path length,
clustering coefficient) can be supposed as a way to analysis of variation of saturation profiles structures
(as the solution of finite element method on the coupled PDEs) where complexity is coming from the
changeable connection and links between assumed nodes. Also, the path of evolution of the supposed
system will be illustrated on the state space of networks either in correlation and Euclidean
measurements. The results of analysis showed in a closed system the designed complex networks
approach to a small world network where the mean path length and clustering coefficient are low and
high, respectively. As another result, the evolution of macro -states of system (such mean velocity of air
or pressure) can be scaled with characteristics of structure complexity of saturation. In other part, we tried
to find a phase transition criterion based on the variation of non-wetting phase velocity profiles over a
network which had been constructed over correlation distance.
Keywords: Unsaturated Flow, Complex Networks, heterogeneity.
1
1. Introduction
Complexity of flow in porous media–especially in multiphase state- is the result of multi
dimensional (geological) heterogeneity of particles where each element of system (or group of
elements) based on nonlinear mutual relation with gravity, capillarity and pressure gradient
shows its behaviour. In analysis of such rich complex behaviour apart of physical simplification
to solve the problem, the possible uncertainties in different format (spontaneously or nonspontaneously) will induce other facet of complexity. This type of complexity has allocated
considerable literature where uncertainty (i.e. stochastic or fuzzy) in direct form construct
stochastic differential equations or fuzzy differential equation while in indirect branch
uncertainty acts basically over collected information from system and gives inferences associated
with in(i.e. fuzzy) or on(i.e. rough set) boundaries of system [1],[17],[19].
Along this complexity, the interaction of particles which their dynamic directly are
governed by forces (agents) exhibits another complexity (absolutely coupling with uncertainty
and controlling equations): structure and topology based complexity. Undoubtedly, the later
complexity has an effective role in recognizing of state of each particle where the interactions of
the (local) states with each other and external forces give practical measurement of the system. In
other world, dynamical superposition of complexity in external forces, uncertainty in behavior
and interaction of particles may give a general evolution path of a system [1, 4]. The point that
must be considered is the role of local interactions of particles (in spatial or /and temporal form)
on the evolution of system so that such locality in different computational based methods (i.e.,
game theory, cellular automata, lattice Boltzmann methods, finite difference solutions, etc) has
been taken in to account. However, the idea behind the mentioned methods (as the basic and
initial idea) is considering the function of immediate neighborhoods[2-3].
Understanding of structures Complexity in variation and dynamic of a physical system
can be evaluated by constructing links (may be with weights) over nodes (as particles or agents).
This interpretation of topology complexity during the last years has opened a new perspective to
traditional graph theory, called complex networks [2]. Picturing, modeling and evaluation in a
simple and intuitive way are some of the discriminated features of complex networks [3, 4].
Based on the mentioned description of the topology structure over space and/or time, possible
2
explicit or implicit of agents /governing equations within or behind of complexity can lead to
networks on differential equations (NODs-NPDs) or linguistic rules such fuzzy rules (NFRs) [4]
.Complex networks have been developed in the several fields of science and engineering for
example social, information, technological, biological and earthquake networks are the main
distinguished networks [5-12]. Regard to the geosciences field, during past decade, some
considerable efforts have been done on the complex earthquake networks (generally recursive
events [7, 10] and climate dynamic networks. The dynamical growth of a thin film surface (and
in a nano-structural way) and evolution of a rock joint (complex aperture network-millimeter
scale) by this approach have been devised [8, 10].
The aim of this study is to analysis of complexity evolution of saturation (of air) profiles in a
closed heterogeneous soil sample (under mildly linear evolution of gas) either in spatial or
temporal way. In this way based on two metric spaces, firstly along each sampled time steps,
networks will be constructed. Characterization of the appeared networks and comparison of
network properties with each other and other distinguished mean properties of system (such
pressure and velocity) will be the next step of this study. In fact we are searching a suitable space
in where emerged properties of networks over one parameter can scaled with other dimension
(attribute) of system. In other word, the question “is there any hidden space so that structure
evolution of one property over that space is related to another attribute of same system”. Also,
we will make state spaces of complex networks where show the path of evolution of system
where are dissimilar for evaluated distances. The organization of this paper is as follow: in the
second part we will introduce the model parameters and accompanied governing equations; the
next section will cover a brief on complex networks properties and the results of the designated
networks on saturation and velocity profiles will be completed this section.
2.
Advection flow in porous media
Generally, the process of displacement of wetting/non-wetting phase is mainly affected by the
properties of the permeable medium, and fluids in single-phase both in homogeneous and in
heterogeneous media. In two-phase immiscible flows, interactions between the permeable
medium and the fluids also affect the fluid flow paths. Because flow dynamics depend on a
combination of conditions such as heterogeneity, moisture content, and chemistry, the resulting
3
transient flow and transport are usually complex. A large number of numerical methods have
been developed to model two-phase flow in heterogeneous media. The finite difference and finite
volume methods are the general frameworks for numerical simulation for the study of fluid flow
in very large problems. It has been proved that the finite difference method, however, is strongly
influenced by the mesh quality and orientation, which makes the method unattractive for
unstructured gridding [13-14].
Finite element methods (in different forms), have been used to model single-phase and
two-phase flow [13-17] in fractured media and heterogonous permeable media with different
capillarity pressures. In this part of study, based on the transition modeling of immiscible twophase flow, the role of permeability, porosity and the width of pore-size distribution in a small
sample of soil using FEMLAB [18] will be evaluated, while non-wetting phase (air) changes
with time, mildly. In all of this study, we assume that the porous medium is a non-deformable
(constant porosity) and that cross-product permeability terms associated with the viscous drag
tensor can be neglected. The general form of the two-fluid flow equations (without source–sink
terms) is described by the two-fluid, volume-averaged momentum and continuity equations [1517]:
qw = −
φ
kw
µw
∂ ( ρ w Sw )
+ ∇.( ρ wq w ) = 0
∂t
q nw = −
φ
[∇Pw + ρ wg ]
k nw
µnw
.[∇pnw + ρ nwg ]
∂ ( ρ nw S nw )
+ ∇.( ρ nw q nw ) = 0
∂t
(1)
(2)
(3)
(4)
In Eqs (1-4), the subscripts w and nw denote the wetting and non-wetting fluids,
respectively; Pi (i = w, nw) , Si , q i ,
g , µ i and k i denote pressure ,degree of saturation relative to
the porosity φ , the flux density vector , the gravitational acceleration vector, dynamic viscosity
, density and the effective permeability tensor, respectively. The effective permeability can be
defined as the relation between intrinsic permeability ( k ) and relative permeability ( k ri ):
ki = kri k .
4
With definition of volumetric fluid content as
θi = φSi we can write:
Sw + Snw = 1
(5)
θw + θnw = φ
Assuming one-dimensional vertical flow and that the wetting fluid is incompressible,
substitution of Eq. (1) into Eq. (2) gives:
∂θw ∂ kw ∂Pw
= (
+ ρw g )
∂t ∂z µw ∂z
(6)
When replacing fluid pressures and capillary pressure ( Pc = Pnw − Pw ) with pressure head
hi =
Pi
K =
ρH2O .g and defining the hydraulic conductivity of fluid i by i
ki ρ H2O . g
µi
, transient flow
of the wetting fluid by considering the capillary capacity ( C w = dθ w
) is described by :
dh
c
Cw (
∂hnw ∂hw
∂
∂h
−
) = K w ( nw + 1)
∂t
∂t
∂z
∂z
(7)
Substitution of Eq.(3) in Eq.(4) yields :
∂( ρnwθnw ) ∂ ρ nwknw ∂Pnw
=
(
+ ρ nw g )
∂t
∂z µnw
∂z
(8)
For air, the density of non –wetting fluid has a dependency to the pressure head:
ρ nw = ρ 0 ,nw + (
ρ 0 ,nw
h0
(9)
) hnw
where ρ0,nw , ρ0,nw are the reference density and pressure head (at atmospheric pressure),
respectively. The ratio of
ρ 0,nw
h0
is defined as the compressibility λ (λ = 1.24 × 10−6 g
cm4
) . Then,
the flow of non-wetting fluid can be concluded:
{(φ − θ w )λ − ρ nwCw }
∂hnw
∂h
∂
∂h
ρ
+ ρ nwC w w = ρ nw K nw ( nw + nw
ρ H 2O
∂t
∂t
∂z
∂z
5
(10)
Equations (7) and (10) for the pre-defined boundary conditions can be solved
simultaneously (coupled PDES). By functional description of the capillary pressure–saturation,
hc(Sw), and permeability functions, ki(Sw), the evolution of wetting and non-wetting phases
distribution can be estimated.. In this study, we use the van Genuchten equation (VG) [21]:
Sew = [1 + (αhc )n ]−m
where Sew denotes the effective saturation of the wetting fluid, Sew = (θ w − θ wr )
where
θwr and θwr
(11)
(θ ws − θ wr )
are the saturated and residual wetting fluid saturation, respectively;
α
,
and n
are fitting parameters, that are inversely proportional to the non-wetting fluid entry pressure
value and the width of pore-size distribution, respectively. We assume that m = 1 − 1 , and that
n
the effective saturation of the non-wetting fluid (Sen) is derived from Sen =1-Sew.
The capillary pressure–saturation function can be considered a static soil property, while
the permeability function is a hydrodynamic property describing the ability of the soil to conduct
a fluid. The basic assumption behind Capillary pressure–permeability prediction models are from
conceptual models of flow in capillary tubes combined with pore-size distribution knowledge
which are derived from the capillary pressure–saturation relationship. A typical representation of
this type of model follows Mualem formulation [22]:
Se dSe
∫
η 0 hc
krw = Sew 1
dSe
∫
0 hc
2
1 dS e
∫
hc
η Se
k rn = (1 − S ew ) 1
dS
∫ e
0 hc
6
(12a)
2
(12b)
Combining the van Genuchten capillary pressure–saturation Eq. (11) with the Mualem
(VGM) model-with introducing new parameter the tortuosity parameter ( η ) gives permeability
functions as follows by [15,22]:
k rw
k rnw
1
k
η
m m
= w = S ew
1
−
(1
−
S
ew )
k
2
1
k
m
= nw = (1 − S ew )η 1 − S ew
k
(13a)
2m
(13b)
Combination Eqs. (8) and (10), the boundary and initial conditions (Fig. 1), and the
constitutive relationships in Eq.(12-13) constitute the mathematical model of the assumed
system. In the upper bound of the geometry (fig. 1), a slight growth of the air pressure in 0 to
1hour (see Fig. 2) will be increased where the lower bound of the sample is in-transferable to
non-wetting phase (Fig. 1). The experiment injects air over the surface of a laboratory column
filled with water and sand. The incoming air (the nonwetting phase for this fluid pair) forces the
water (the wetting phase) toward the outlet at the base of the column. At the inlet, air pressure
increases by steps in time, and no water exits through the column top. In moving to the outlet,
the water passes through a disc that is impermeable to air flow. Neither the air nor the water can
pass through the vertical column walls. The water pressure at the outlet, which changes in time,
corresponds to the height of fluid rise in a receiving buret. The column has a total length of 8.34
cm, a 6-cm radius, and the disk is 0.74 cm thick [15]. Initial non-wetting phase pressure head is
assumed 0.2 (m water) and at the water outlet, the fluid level increases linearly in time from 0 m
to 0.1 m. The employed parameters for the sample come from Lincoln sand (88.6% sand, 9.4%
silt and 2.0% clay). The constant parameters for our model are as follow (see figure 3 for
uncorrelated normal distribution of random parameters):
θr (cm3cm−3 ) = 0.0210
θ s = [θ s ];
2
−9
k (cm ×10 ) = [kij ]
−1
α (cm ) = 0.0189
n = [nij ]
7
in which θs , k and n will follow an uncorrelated Gaussian distribution drawn in Figure 3.
y
qw = 0, pnw = pnw (t )
7.6 cm
top
pnw = pnw (t0 ) = pnw
+ (8.34(cm) − y ) ρ nw g
pw = pw (t0 ) = − ρ w gy
Impermeable layer for air
0.74 cm
qnw = 0; pw = pw (t )
6cm
Figure 1. Schematic representation of boundary and initial conditions [15,18].
8
x
6.59
6.58
Gas Pressure (H)
6.57
6.56
6.55
6.54
6.53
6.52
6.51
6.5
0
0.2
0.4
0.6
0.8
1
Time(h)
Figure 2. Gas evolution transition on a supposed sample
m2
Figure 3. Intrinsic Permeability variation based on a normal distribution
9
250
250
t=0
250
t=0.5 h
250
t=0.8h
200
200
200
200
150
150
150
150
100
100
100
100
50
50
50
50
50
100
150
200
250
50
100
150
200
250
50
100
150
200
250
t=1 h
50
100
150
200
S_nw
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
Figure 4. Effective saturation of non-wetting phase evolution
(h)
Figure 5. Absolute air-velocity fluctuation (inset: effective saturation of non-wetting phase evolution along time)
10
250
(h)
Figure 6. Air pressure changes vs. time for four randomly selected points
Figures 4-6 represent the result of
simulation with 1857 triangular elements for
saturation of non-wetting phase ,absolute value of air-velocity and pressure of non-wetting
phase evolution, respectively. Based on the later mentioned figure it seems in association with
initial boundary condition a phase transition before 0.1 h can be observed. The point of phase
transition is known as a critical point to include some basic terminology. Usually two types of
phase transitions are distinguished, first order and second order. First order designates phase
transition where the macroscopic states change in a discontinuous way upon passage through the
critical point, and second order indicates phase transitions where the states change in a
continuous way [23]. When a system passes through a second order transition; it may be left at
that point. At this point the system is like a pencil balanced on its end (only small variation may
be observed). We cannot tell which direction it will fall, but a small perturbation can send it
falling in a certain direction. In our case, the phase transition can be pronounced around 0.074 h
where overall signals of the variables states show an indented behaviour change.
11
Frequency of normalized velocities
Normalized total velocity
Figure 7. frequency distribution of the total velocity over time in log-log coordinate
Depicting of normalized (between 0 and 1) velocity frequency (Fig.7) shows the general overall
probability patterns of locally scaled- velocity. Roughly speaking, It can be followed that the
frequency of displacements rates of non-wetting phase is well approximated by a truncated
power-law [23, 24]: N ( v ) = ( v + v0 ) − β exp( − v ) ; where v0 = 0.05; β =1.5;κ = 0.85 . This equation
κ
suggests that air particles motion follows approximately a truncated Levy flight (Levy
processes), indicating that, despite the diversity of their travel history rate, the scaled rate of
displacement of non-wetting phase particles follow reproducible patterns. The main property of
the levy walk procedures is obeying from the infinite variance despite of the Gaussian process
which has a finite variance. However, extension of Levy flight to truncated Levy flight transfers
the infinite variance to finite variance [25-26].
12
3. Complex networks
As we mentioned in the previous section, the emerged patterns are following relatively
complex forms. Such complex forms come out in the signals and time series formats. Recently,
one of the methods for analysis of signals (information patterns) which has been considered is
employing of complex networks over the time-space (spatial/temporal). Different methodologies
have been proposed to capture the signal evolution based on the networks formation [31].
A network (graph) consists of nodes and edges connecting them. To set up a network, we
consider each profile (in y-direction- saturation or pressure of non wetting phase) as a node. To
make edge between two nodes, a relation should be defined. Several similarity or metric spaces
has been proposed for a construction of a proper network. The main point in selection of each
space is to explore the explicit or implicit hidden relation among different distributed elements of
a system. In this study we will use p-value criterion (a matrix for testing the hypothesis of no
correlation) related to correlation measurement and Euclidean distance over the non-wetting
saturation profiles. For each pair of signals (profiles) Vi and V j , containing L elements (pixels) the
correlation coefficient can be written as [30-31]:
L
∑ [V (k )− ≺ V ]. V ( k )− ≺ V
i
Cij =
i
j
j
(14)
k =1
L
L
∑ [V ( k )− ≺ V ] . ∑ [V (k )− ≺ V ]
2
i
i
2
i
k =1
i
k =1
L
∑V (k )
i
where ≺ Vi =
k =1
L
. Obviously,
Cij is restricted to the −1 ≤ Cij ≤ 1 , where Cij = 1,0 and −1
are related to perfect correlations, no correlations and perfect anti-correlations, respectively.
Each p-value is the probability of getting a correlation as large as the observed value by random
chance, when the true correlation is zero. The p-value is computed by transforming the
correlation to create a t- statistic having N-2 degrees of freedom, where N shows number of
profiles (samples) [27]. If p is small-here less than p ≤ ξ =0.05- then the correlation Cij is
significant. Euclidean distance is given as:
13
L
d ij =
∑ (V (k ) − V (k ))
i
2
(15)
j
k =1
Selection of threshold ( ξ ) is a challengeable discussion that can be seen from different
view. Choosing of such constant value may be associated with the current accuracy at data
accumulation where after a maximum threshold the system loses its dominant order. In fact,
there is not any unique way in selection of constant value, however, preserving of general
patterns of evolution must be considered while the hidden patterns (in our study come out from
the solution of two-PDEs based on a two-phase flow constitutive model) can be related to the
several characters of a networks . These characters can express different facets of the relations,
connectivity, assortivity (hubness), centrality, grouping and other properties of nodes and/or
edges. Generally, it seems obtaining stable patterns of evolution (not absolute) over a variation of
ξ can give a suitable and reasonable formed network. For the Euclidean distance we set
d ij ≥ ξ = 0.7 d ijmax . Consider with this definition we are investigating dissimilarity of profiles over
the metric space while for former metric the variations of profiles related to each other are
measured.
Let us introduce some properties of the networks: clustering coefficient (C) , the degree
distribution ( P (k ) ) and average path length ( L ). The clustering coefficient describes the
degree to which k neighbors of a particular node are connected to each other. Our mean about
neighbors is the connected nodes to the particular node. To better understanding of this
concept the question “are my friends also friends of each other?” can be used. In fact
clustering coefficient shows the collaboration between the connected nodes to one. Assume
the i
th
node to have k i neighboring nodes. There can exist at most k i (k i −1) / 2 edges
between the neighbors (local complete graph). Define c i as the ratio
ci =
actual
number
of
edges
between the neighbors
k i ( k i − 1) / 2
of
the
i th
node
(16)
Then, the clustering coefficient is given by the average of c i over all the nodes in the
network [7,10]:
14
C =
For k i ≤ 1 we define C
1
N
N
∑c .
i
(17)
i =1
≡ 0 . The closer C is to one the larger is the interconnectedness of the
network. The connectivity distribution (or degree distribution), P (k ) is the probability of finding
nodes with k edges in a network. In large networks, there will always be some fluctuations in the
degree distribution. The large fluctuations from the average value ( < k > ) refers to the highly
heterogeneous networks while homogeneous networks display low fluctuations [6]. The average
(characteristic) path length L is the mean length of the shortest paths connecting any two nodes
on the graph. The shortest path between a pair (i, j) of nodes in a network can be assumed as
their geodesic distance gij , with a mean geodesic distance L given as below [11]:
L=
2
∑ gij ,
N ( N − 1) i < j
(18)
where gij is the geodesic distance (shortest distance) between node i and j, and N is the number
of nodes. We will use a well known algorithm in finding the shortest paths presented by Dijkstra
[28]. Based on the mentioned characteristics of networks two lower and upper boards of
networks can be recognized: regular networks and random networks (or Erd˝os-Renyi networks
[5-6]). Regular networks have a high clustering coefficient (C ≈ 3/4) and a long average path
length. Random networks (construction based on random connection of nodes) have a low
clustering coefficient and the shortest possible average path length. However Watts and Strogatz
[30] introduced a new type of networks with high clustering coefficient and small (much smaller
than the regular ones) average path length (this is called SW property).
To extraction of saturation networks we set up X=273*Y=273 points on the constrained
area (only upper part-see figure 1). During the evolution of system, in 10 time-points using
aforementioned method, complex networks along Y-direction and on the non-wetting phase were
elicited. Consider that the numbers of evaluated points for construction of networks are smaller
and bigger than the number of (finite) elements and heterogeneous elements which give the
solution. This implies a limited correlation around each granule (node) with neighbourhood
nodes. This can be followed by the evolution of adjacency matrix visualization where for
15
correlation measurement and Euclidean distance have been depicted at figure 8. Appearance of a
thin layer of connections (disconnections for Euclidean distance) around diagonal of matrix
along successive time steps suggests a correlation around each node. It seems that the variation
of the emerged network over the p-value measurement is insignificant whereas networks based
on dissimilarity of Euclidean metric shows much higher variations. Evolution of number of
edges, mean clustering coefficient and average path length can be followed in Figures 9 -11,
respectively. As one can ensue the pattern and general trend of clustering coefficient and edges
are similar for each case and average path length shows an opposite variation with the later
properties. Then it can be interpreted, in mean view, increasing or decreasing of edges rate is
near to forming or decaying of triangles (clustering coefficient). Also, high clustering coefficient
and low average path length (especially for Euclidean measurement at last time steps) can be
distinguished (property of small-world networks) [6, 27]. Notice the fluctuation of properties of
saturation networks are not as well as the change of saturation (Fig.5), but is so similar to
absolute value of air velocity. Roughly speaking, one can infer from comparison of velocity and
L , < ci > , K
that
maximum
velocity
due
to
numerical
solution
coincides
with
=0.005−0.01
t =0.005−0.01
< ci >tcorrelation
, Kcorrelation
. It is noteworthy before t<0.02 the properties of two distances are
reveres each other but after this step the general trends are same so that we can write:
max{K , C , L}Euclidean = t{0.08, 0.08, 0.02}
min{K , C , L}Euclidean = t{0.02, 0.02,0.08}
max{K , C , L}p-value = t{0.01,0.01, 0.02}
.
min{K , C , L}p-value = t{0.02, 0.02, 0.01}
Consider that maximum (minimum) value of K and C are adapted to minimum
(maximum) L and vice versa. If we consider t=0.02 as transition point as has been depicted in
figure 5; this gives the point (or stage) with lowest stability which may indicate lowest
correlation (t=0.02). The temporal evolution of clustering coefficient and edges over Euclidean
distance so approve this stage is a much more chaotic step where after this point the patterns of
spatial and temporal characters of structure complexity approximately reach a quasi stationary
step. To understand the path evolution of saturation profiles, we construct state diagram of
k i − ci for each case (Fig.14). For correlation distance (Fig.14 a) nodes generally (without
16
considering time effects) lies from higher clustering coefficient to a congested area where in this
area (can be called equilibrium place) we have 100 < kicore < 120;0.4 < cicore < 0.45 .
1
1
250
0.9
t=0
250
0.9
t=0.08 h
200
200
0.7
0.7
0.6
150
0.5
0.4
0.3
0.3
0.2
50
100
150
200
250
Correlation Distance
250
0.7
0.7
0.6
0.6
150
150
0.5
0.5
0.4
100
0.4
100
0.3
0.3
0.2
50
0.2
50
0.2
50
0.1
0.1
50
0.8
200
200
0.4
100
0.9
t=0.08 h
0.8
0.6
150
0.5
100
250
0.9
t=0
0.8
0.8
1
1
250
0.1
0.1
50
0
100
150
200
250
0
50
100
150
200
250
t=1 h
250
0
50
t=1h
100
1
150
200
250
0
0.9
Euclidean Distance
0.8
200
200
150
150
0.7
0.6
0.5
a)
b)
100
0.4
100
0.3
0.2
50
50
0.1
50
100
150
200
250
50
100
150
200
250
0
Figure 8. Evolution of adjacency matrix visualization of non-wetting phase saturation along time: based on a) p-value of
Correlation measurements and b) Euclidean Distance
(h)
Figure 9. Variations of number of edges in complex saturation networks vs. time: based on a) Euclidean Distance and b) pvalue of Correlation measurements
17
0.7
0.45
0.65
0.6
b) Correlation Distance
0.445
a) Euclidean Distance
0.55
<c>
< C>
0.44
0.5
0.45
0.435
0.4
0.43
0.35
0
0.1
0.2
0.3
0.4
0.5
t(h)
0.6
0.7
0.8
0.9
1
0
0.2
0.4
0.6
0.8
1
t
Figure 10. Variations of mean clustering coefficient in complex saturation networks vs. time: based on a) Euclidean Distance
and b) p-value of Correlation measurements
1.625
1.95
1.9
b) Correlation Distance
1.62
a) Euclidean Distance
1.615
1.8
mean geodesic distance
m ean geodesic distance
1.85
1.75
1.7
1.65
1.6
1.61
1.605
1.6
1.55
1.595
1.5
0
0.1
0.2
0.3
0.4
0.5
t(h)
0.6
0.7
0.8
0.9
1.59
1
0
0.1
0.2
0.3
0.4
0.5
t(h)
0.6
0.7
0.8
0.9
1
Figure 11. Evolution of mean geodesic distance (mean length of the shortest paths) of complex saturation networks along
time: a) Euclidean Distance and b) p-value of Correlation measurements
18
c
i
1
t=0
0.5
0
1
t=0.005
0.5
t=0.01
0.5
t=0.02
0.5
t=0.04
0.5
0
1
0
1
0
1
0
1
t=0.08
0.5
t=0.1
0.5
0
1
0
1
t=0.5
0.5
t=0.8
0.5
t=1
0.5
0
1
0
1
0
0
0.5
1
0
t=0
0.5
1
0
t=0.005
Figure 12. Temporal change of
0.5
1
0
t=0.01
0.5
1
0
t=0.02
0.5
1
0
t=0.04
0.5
1
0
t=0.08
0.5
1
0
t=0.1
0.5
1
0
t=0.5
0.5
1
t=0.8
0
0.5
1
t=1
c i related to different reference point s in association with Euclidean Distance
ki
400
t=0
200
0
400
t=0.005
200
t=0.01
200
0
400
0
200
t=0.02
100
0
400
t=0.04
200
0
400
t=0.08
200
0
400
t=0.1
200
0
400
t=0.5
t=0.8
200
0
400
200
0
400
200
0
0
200
t=0
400 0
200
400 0
t=0.005
Figure 13. Temporal change of
200
400 0
t=0.01
100
200 0
t=0.02
200
400 0
t=0.04
200
400 0
t=0.08
200
400 0
t=0.1
200
t=0.5
400 0
200
400 0
t=0.8
200
t=1
k i related to different reference points (time) in association with Euclidean Distance
19
400
data1
data2
data3
data4
data5
data6
data7
data8
data9
data10
-0.3
ci
ci
ci
c(i)
10
-1
10
-0.4
10
a) Correlation Distance
data1
data2
data3
data4
data5
data6
data7
data8
data9
data10
2
1
10
k(i)
k
b) Euclidean Distance
10
i
Figure 14. Diagram of structure complexity space over two variables of networks
2
k
10
kii
k i and c i : a) p-value of Correlation, b)
Euclidean Distance-data 1 to 10 are the sampling values within the time steps.
Then, roughly speaking, the path evolution of saturation profiles in correlation
measurement shows a narrow variation over the complexity structure space so that at stationary
state the core of space can be imagined as a point of absorption. It must be considered that the
discrimination of time steps over this evolutionary grape shape path is difficult. Despite of the
former metric, Euclidean distance can give irregularities over time and space (Fig14b). Figure
14.b shows in t=0.02 there is distinguished chaotic separately different from overall way of
change of patterns which is matched with pervious results. Also, the way of evolution can be
estimated by: ci ≈ 2.61ki−0.32 . One may modify this relation with a truncated power law. In another
view and with refer to our results drawn in Figure 14; we can conclude that the correlated and
dissimilar patterns of saturation profiles (at a closed system) in a proper space of networks may
be assumed as synchronizing oscillators. In other word, based on anharmonic oscillators’ theory
and the governing non linear differential equations we can write (for Figure 14):
d 2 ci
dc
= −ω 2 sin ci (t ) − α i + f cos ωt ,
dt
dt
in which ω , α and f are constant values.
Notice that with increasing of edges the communicability of system on this space is
decreasing which shows the dissimilarity patterns tend to growth over all on nodes rather that the
20
neighbourhoods of nodes. So, comparison with the distribution of edges indicates that nodes with
high correlation and dissimilarities are decaying as well as triangles densities around each node
(Figs.15&16). This implies that low dissimilar patterns (small k i ) have higher communicability.
Let us investigate the temporal structure complexity. Two immediate solution can be presented
in first situation matrix of connections along evolution of time is extracted. In this case spatial
connection of nodes is replaced with temporal evolution of same node with same node at next (or
pervious) time steps or other nodes. Consider a reference time step ( t ref . ) then we can make the
following matrix among t ref . , t ref . + ∆ t :
ref .
M
t ref .,t ref . +∆t
t +∆t
. . .
ref .
= t . mi .
; mi ∈ {0,1}; i = 1,.., N
. . . N × N
(19)
where mi is the connection state of element i (position i) due to one relation along a time
interval. So, a multiple array over different time references can cover completely both spatial (
∆t = 0 ) and temporal evolution of structure complexity such:
M (t0 ,t0 +∆t )
.
B =
.
M (t0 ,t0 + n×∆t )
.
; n = 1 + t f − t0
∆t
.
( t ,t + n×∆t )
n×n
. . M f f
. .
. .
. .
M
( t f ,t f +∆t )
(20)
in which n is the time granulation. Consider for ∆t = 0 ,we can see only evolution of
spatial complexity on
the assumed time and in collective way as is usual in networks
terminology: [ B ]1× n . The mentioned method can give general complexity patterns of time - space
and interwoven structures, explicitly. Analysis of the attributes and properties of the B is out of
the scope of this study. Second case is an implicit (more qualitatively) view to time-based
complexity. In this state, the spatial characteristics ( k i , c i ) of the system at several time steps
will be compared (see figures 12 and 13).
In another procedure we construct air-velocity networks. The procedure is as well as
correlation distance, only the number of nodes has decreased to 93. Three distinguished
21
behaviours due to the nodes and edges variation are recognized, approximately (Fig.17 a,
Fig.18a): 1) decreasing of nodes where the numbers of edges are approximately fixed (t= [0.01,
0.02]) compatibility in variations and 3) vertices are constant while sensitive variation in links
can be perused (t=[0.3,1]). The behaviour of emerged network especially after (t>0.2) can be
estimated by the properties of small-world networks. The shape of the degree distribution is
similar to that of a random graph. It has a pronounced peak at ≺ k t >0.2 ≈ 8 while the topology of
the network is relatively homogeneous and all nodes having approximately the same number of
edges (Fig.18 b).
Another point can be inferred from the plotting of the reverse of mean clustering coefficient so
that in semi-logarithmic coordinate (Fig17.c) it follows a phase-transition (first order) function
that can be followed by a sigmoid function [22]:
≺ 1/ c ∝ (1 + e
where β and
Lnt
− β (1+1+δ Lnt )
)
(21)
δ are the regulator parameters which determine the declining rate. Then, the reverse
of mean clustering coefficient for velocity networks gives an acceptable phase change criteria.
-3
x 10
0.035
z data
fit 3
a) Correlation
Distance
0.03
z data
fit 1
b) Euclidean
Distance
7
PDF
6
0.025
5
i
Density
PDF 0.02
0.015
4
3
0.01
2
0.005
1
0
0
60
70
80
90
100
110
Data
120
130
140
150
0
50
100
150
ki
ki
200
250
k
Figure15. Edge distribution of complex saturation networks: based on a) p-value of Correlation measurements and b)
Euclidean Distance
22
N (ci )
Figure 16. Evolution of mean clustering coefficient distribution along time: based on a) p-value of Correlation measurements
and b) Euclidean Distance (inset: probability density function at 0.6<t <1 h)
1600
0.7
a)
1400
b)
0.6
1200
0.5
1000
< C>
edges
0.4
800
0.3
600
0.2
400
0.1
200
0
0
0.1
0.2
0.3
0.4
0.5
time
0.6
0.7
0.8
0.9
0
1
0
0.1
0.2
0.3
0.4
0.5
time
0.6
0.7
0.8
0.9
1
45
c)
40
35
30
1./C
25
20
15
10
5
0
-3
10
-2
-1
10
10
0
10
time
Figure 17. a) Evolution of total number of edges for air velocity networks , b) Mean cluster coefficient variation and c) inverse
of Mean cluster coefficient variation
23
Sampled time steps [0,1]
N(k)
Number of Nodes
b)
a)
Sampled time steps [0,1]
k
Figure 18.a) Number of nodes-time ;b) frequency of edges evolution along time for velocity (of air) networks
4. Conclusion
Analysis of two phase flow especially in heterogeneous media, has allocated a noticeable
research literature in porous media area. To take in to account the complexity due to
heterogeneity in different parameters which are characterizing the general reaction (s) of the field
(s), different methodologies have been proposed. The aim of this study was to investigate the
appeared complexity of two-phase flow (air/water) in a heterogeneous soil. The supposed filed
was under the small range of air pressure variation. By considering the capillary pressuresaturation, permeability functions, as constitutive relations and governing equations (obtained
separately for evolution of the model’s wetting and non-wetting phase parameters) the unknown
parameters can be estimated. In this way, using finite element method the role of multiheterogeneity the successive change of different variables (such relative permeability, saturation,
capillary pressure etc) was analysed. Based on the different similarity measurements (i.e., correlation,
Euclidean metrics) over the properties of the emerged patterns of the complex networks were recognized.
In this way, the path of evolution of the supposed system was illustrated based on the state space of
networks either in correlation and Euclidean measurements. The results of analysis showed in a closed
system the designed complex networks approach to small world network property where the mean path
length and clustering coefficient are low and high. As another result, we tried to scale the evolution of
macro -states of system (such mean velocity of air or pressure) with characteristics of structure
complexity of saturation.
24
5. References
1.
Doering R.C.1991.Modeling Complex Systems: Stochastic processes, Stochastic Differential Equations, and Fokker-Planck
Equations; 1990 Lectures in complex systems, SFI Studies in the science of complexity.
2.
Arenas A., Diaz-Guilera A., Kurths J., Moreno Y., Zhou C.2008. Synchronization in complex networks .Physics Reports
469,93-153
3.
Dorogovetev S.N., Goltsev A.V.2008.Crtitical phenomena in complex networks .Review Modern Physics.Vol.80.OctDec.2008
4.
Boccaletti S.,Latora V., Moreno Y.,Chavez M.,Hwang D-U.2006.Complex Networks : Structure and Dynamics. Physics
Reports 424,175-308
5.
Newman M. E. J. 2003.The structure and function of complex networks, SIAM Review; 45(2): 167- 256.
6.
Albert R., Barabasi A.-L.2002. Statistical mechanics of complex networks. Review of Modern Physics; 74, 47–97.
7.
Abe S, Suzuki N.2006. Complex network description of seismicity. Nonlin process Geophys;13:145-150
8.
Ghaffari OH., Sharifzadeh M.,Fall M.2009. Analysis of Aperture Evolution in a Rock Joint Using a Complex Network
Approach. International Journal of Rock Mechanics and Mining Sciences, In Press, Corrected Proof, Available online 5
April 2009.
9.
Valentini L Perugini D, Poli G.2007. The small-world topology of rock fracture networks. Physica A; 377:323–328.
10. Baiesi M, Paczuski M. 2004.Scale free networks of earthquakes and aftershocks. Physical review E 2004; 69 (2): 066106.1066106.8
11. Tsonis, A. A., and P. J. Roebber, 2003: The architecture of the climate network. Physica, 333A, 497–504.
12. Yuksel M,Karabacak T, Guclu H. 2007.Networking behavior in thin film and nanostructure growth dynamics . 2nd IEEE
International Conference on Nano-Networks . Available at <http://arxiv.org/abs/0711.4745v1>
13. Hoteit H, Firoozabadi A. 2008. Numerical modeling of two-phase flow in heterogeneous permeable media with different
capillarity pressures; Advances in Water Resources 31, 56–73.
14. Lunati I, Jenny P. 2006. Multiscale finite-volume method for compressible multiphase flow in porous media. J Comput
Phys;216(2): 616–36
15. Chen J., Hopmans J.W. & Grismer M.E. 1999. Parameter estimation of two-fluid capillary pressure–saturation and
permeability functions; Advances in Water Resources Vol. 22, No. 5, 479–493.
16.
Kueper B. H., Frind E. O. 1991.Two-Phase Flow in Heterogeneous Porous Media Model Development; Water Resources
Research, vol. 27, 6, p 1049-1057.
17. Zhang D .2002. Stochastic methods for flow in porous media: coping with uncertainties. Academic, San Diego, ISBN 012–
7796215, p 350.
18. COMSOL Multiphysics. 2008b. COMSOL Multiphysics Modeling Guide: Version 3.5 Stockholm, Sweden: COMSOL AB.
19. Ghanem G., Dham S. Stochastic finite element analysis for multiphase flow in heterogeneous media. Transport Porous
Media 32:239–262, 1998.
20. Parker, J.C., R.J. Lenhard, and T. Kuppusany. 1987. A parametric model for constitutive properties governing multiphase
flow in porous media. Water Resour. Research 23:618:624
25
21. Van Genuchten, M.Th. 1980. A closed-form equation for predicting the hydraulic conductivity of unsaturated soils. Soil Sci.
Soc. Amer. J. 44:892-898.
22. Mualem, Y. 1976. A new model for predicting the hydraulic permeability of unsaturated porous media. Water Resources
Research 12:513-522.
23. Millonas M.M. Swarms, phase transitions, and collective intelligence, 1993: http://arxiv.org/abs/adap-org/9306002.
24. González M. C., Hidalgo C. A., Barabási A.-L. 2008. Understanding individual human mobility patterns, Nature 453, 779782.
25. Klafter J., Zumofen G., Shlesinger M.F. 1995. Levy Flights and Related Topics in Physics. Lecture Notes in Physics, vol.
450, Springer, Berlin, p. 196.
26. Mantegna, R. N. & Stanley, H. E. 1994. Stochastic process with ultraslow convergence to a Gaussian: the truncated Levy
flight. Phys. Rev. Lett. 73, 2946–2949.
27. Watts DJ, Strogatz SH. Collective dynamics of small-world networks. Nature 1998; 393:440-442.
28. E.W. Dijkstra, Numerische Math. 1, 269 (1959)
29. Massey, F. J.1951. The Kolmogorov-Smirnov Test for Goodness of Fit, Journal of the American Statistical Association,
Vol. 46, No. 253, pp. 68–78.
30. Zhongke Gao and Ningde Jin.2009. Flow-pattern identification and nonlinear dynamics of gas-liquid two-phase flow in
complex networks; Physical Review E 79, 066303.
31. Zhang, J., Small, M.2006. Complex network from pseudoperiodic time series: Topology versus dynamics. Physical Review
Letters 96(238701).
26
| 5cs.CE |
arXiv:1707.06883v3 [math.AG] 23 Mar 2018
IS THE AFFINE SPACE DETERMINED BY ITS
AUTOMORPHISM GROUP?
HANSPETER KRAFT, ANDRIY REGETA,
AND IMMANUEL VAN SANTEN (BORN STAMPFLI)
Abstract. In this note we study the problem of characterizing the
complex affine space An via its automorphism group. We prove the following. Let X be an irreducible quasi-projective n-dimensional variety
such that Aut(X) and Aut(An ) are isomorphic as abstract groups. If X
is either quasi-affine and toric or X is smooth with Euler characteristic
χ(X) 6= 0 and finite Picard group Pic(X), then X is isomorphic to An .
The main ingredient is the following result. Let X be a smooth irreducible quasi-projective variety of dimension n with finite Pic(X). If X
admits a faithful (Z/pZ)n -action for a prime p and χ(X) is not divisible
by p, then the identity component of the centralizer CentAut(X) ((Z/pZ)n )
is a torus.
1. Introduction
In 1872, Felix Klein suggested as part of his Erlangen Programm to study
geometrical objects through their symmetries. In the spirit of this program
it is natural to ask to which extent a geometrical object is determined by its
automorphism group. For example, a smooth manifold, a symplectic manifold or a contact manifold is determined by its automorphism group, see
[Fil82, Ryb95, Ryb02].
We work over the field of complex numbers C. For a variety X we denote
by Aut(X) the group of regular automorphisms of X. As the automorphism
group of a variety is usually quite small, it almost never determines the
variety. However, if Aut(X) is large, like for the affine space An , this might
be true. Our guiding question is the following.
Question. Let X be a variety. Assume that Aut(X) is isomorphic to the
group Aut(An ). Does this imply that X is isomorphic to the affine space An ?
This question cannot have a positive answer for all varieties X. For example, Aut(An ) and Aut(An × V ) are isomorphic for any complete variety
V with a trivial automorphism group. Similarly, the automorphism group
of An does not change if one forms the disjoint union with a variety with a
trivial automorphism group. Thus we have to impose certain assumptions on
X. Moreover, we assume that n ≥ 1, since there exist many affine varieties
with a trivial automorphism group.
2010 Mathematics Subject Classification. 14R10, 14R20, 14J50, 22F50.
1
2
H. KRAFT, A. REGETA, AND I. VAN SANTEN
In [Kra17], the first author studies the problem of characterizing the affine
space An by its automorphism group regarded as a so-called ind-group. It is
shown that if X is a connected affine variety such that Aut(X) and Aut(An )
are isomorphic as ind-groups, then X and An are isomorphic as varieties.
For some generalizations of this result we refer to [Reg17].
In dimension two, a generalization of our guiding question is studied in
[LRU17]. For an irreducible normal affine surface X it is shown that if
Aut(X) is isomorphic to Aut(Y ) for an affine toric surface Y , then X is
isomorphic to Y .
In order to state our main result, let us introduce the following notation.
For a variety X we denote by χ(X) the Euler characteristic and by Pic(X)
the Picard group.
Main Theorem. Let X be an irreducible quasi-projective variety such that
Aut(X) ≃ Aut(An ). Then X ≃ An if one of the following conditions holds.
(1) X is smooth, χ(X) 6= 0, Pic(X) is finite, and dim X ≤ n;
(2) X is toric, quasi-affine, and dim X ≥ n.
As a direct application of this result we get that Aut(An \S) and Aut(An )
are non-isomorphic as abstract groups for every non-empty closed subset S
in An with Euler characteristic χ(S) 6= 1.
∼
Let us give an outline of the proof. Let θ : Aut(An ) → Aut(X) be an
isomorphism. First, we prove that if a maximal torus of Aut(An ) is mapped
onto an algebraic group via θ and X is quasi-affine, then X ≃ An (see
Proposition 23). Our main result to achieve this condition is the following.
Theorem 1. Let X and Y be irreducible quasi-projective varieties such that
dim Y ≤ dim X =: n. Assume that the following conditions are satisfied:
(1) X is quasi-affine and toric;
(2) Y is smooth, χ(Y ) 6= 0, and Pic(Y ) is finite.
∼
If θ : Aut(X) → Aut(Y ) is an isomorphism, then dim Y = n, and for each
n-dimensional torus T ⊆ Aut(X), the identity component of the image θ(T )◦
is a closed torus of dimension n in Aut(Y ). Furthermore, Y is quasi-affine.
For the definition of the topology on Aut(X), the definition of the identity
component G◦ of a subgroup G ⊆ Aut(X) and the definition of algebraic
subgroups of Aut(X) we refer to section 2.2.
For the proof of Theorem 1 we first remark that every torus T ⊆ Aut(X)
of maximal dimension n = dim X is self-centralizing (Lemma 10). For any
prime p the torus T contains a unique subgroup µp which is isomorphic
to (Z/pZ)n . In particular, T ⊆ CentAut(X) (µp ), and thus the image of T
under an isomorphism θ : Aut(X) → Aut(Y ) is mapped to a subgroup of
the centralizer of θ(µp ) inside Aut(Y ). Our strategy is then to prove that
the identity component of the centralizer CentAut(Y ) (θ(µp )) is an algebraic
group. Our main result in this direction is the following generalization of
[KS13, Proposition 3.4].
IS THE AFFINE SPACE DETERMINED BY ITS AUTOMORPHISM GROUP?
3
Theorem 2. Let X be a smooth, irreducible, quasi-projective variety of dimension n with finite Picard group Pic(X). Assume that X is endowed with a
faithful (Z/pZ)n -action for some prime p which does not divide χ(X). Then
the centralizer G := CentAut(X) ((Z/pZ)n ) is a closed subgroup of Aut(X),
and the identity component G◦ is a closed torus of dimension ≤ n.
For the proof of Theorem 2 we use the fact that p does not divide χ(X)
to find a fixed point of the (Z/pZ)n -action on X (Proposition 16), and the
n
smoothness of X to show that the fixed point variety X (Z/pZ) is finite.
Acknowledgements. We would like to thank Alvaro Liendo for fruitful discussions.
2. Preliminary results
Throughout this note we work over the field C of complex numbers. A
variety will be a reduced separated scheme of finite type over C.
2.1. Quasi-affine varieties. Let us recall some well-known results about
quasi-affine varieties. The first lemma is known for affine varieties and can
be reduced to this case by using open affine coverings.
Lemma 3. Let X, Y be varieties. Then the natural homomorphism
O(X) ⊗C O(Y ) → O(X × Y )
is an isomorphism of C-algebras.
Lemma 4. Let X be a quasi-affine variety. Then the canonical morphism
η : X → Spec O(X) is a dominant open immersion of schemes.
Proof. Let f : Spec O(X) → A1 be a morphism which vanishes on η(X).
Since f can be understood as a regular function f˜ on X, we get the following
commutative diagram
1
: AO
t
f
t
t
tt
t
t
X η / Spec O(X)
t
f˜ ttt
which shows that f˜ = 0. This implies that η is dominant.
We have an open immersion ι : X → Y where Y is affine, and therefore a
decomposition of ι
ι: X
η
/ Spec O(X)
α
/Y .
In particular, η is injective. For any nonzero f ∈ O(Y ) such that Yf ⊆ ι(X)
∼
we see that ι induces an isomorphism Xι∗ (f ) → Yf , hence the composition
ι′ : Xα∗ (f )
η′
/ Spec O(X)α∗ (f ) = α−1 (Yf )
α′
/ Yf
4
H. KRAFT, A. REGETA, AND I. VAN SANTEN
is an isomorphism where η ′ and α′ are the restrictions of η and α respectively.
Since ι′ is an isomorphism Xα∗ (f ) is affine and thus η ′ is an isomorphism,
because O(X)α∗ (f ) = O(Xα∗ (f ) ). Therefore, η is a local isomorphism, hence
an open immersion.
Lemma 5. Let X be a quasi-affine variety and Y a variety. Then every
morphism Y × X → X extends uniquely to a morphism Y × Spec O(X) →
Spec O(X) via X → Spec O(X). In particular, every regular action of an
algebraic group on X extends to a regular action on Spec O(X).
Proof. We can assume that Y is affine. By Lemma 3 we have O(Y × X) =
O(Y ) ⊗C O(X). Hence Y × X → X induces a homomorphism of C-algebras
O(X) → O(Y ) ⊗C O(X) which in turn gives the desired extension Y ×
Spec O(X) → Spec O(X).
2.2. Algebraic structure on the group of automorphisms. In this
subsection, we recall some basic results about the automorphism group
Aut(X) of a variety X. The survey [Bla16] and the article [Ram64] will
serve as references. Recall that a morphism ν : A → Aut(X) is a map from
a variety A to Aut(X) such that the associated map
ν̃ : A × X → X ,
(a, x) 7→ ν(a)(x)
is a morphism of varieties. We get a topology on Aut(X) by declaring a subset F ⊂ Aut(X) to be closed, if the preimage ν −1 (F ) under every morphism
ν : A → Aut(X) is closed in A. Similarly, a morphism ν = (ν1 , ν2 ) : A →
Aut(X) × Aut(X) is a map from a variety A into Aut(X) × Aut(X) such
that ν1 and ν2 are morphisms. Thus we get analogously as before a topology
on Aut(X) × Aut(X). Note that for morphisms ν, ν1 , ν2 : A → Aut(X) the
following maps are again morphisms
A → Aut(X) , a 7→ ν1 (a) ◦ ν2 (a)
A → Aut(X) , a 7→ ν(a)−1
and that ν −1 (∆) is closed in A where ∆ ⊂ Aut(X) × Aut(X) denotes the
diagonal. From these properties, one can deduce that Aut(X) behaves a bit
like an algebraic group:
Lemma 6. For any variety X, the maps
Aut(X) × Aut(X) → Aut(X) , (ϕ1 , ϕ2 ) 7→ ϕ1 ◦ ϕ2
Aut(X) → Aut(X) , ϕ 7→ ϕ−1
are continuous and the diagonal ∆ is closed in Aut(X) × Aut(X).
Example 1. For any set S ⊆ Aut(X) the centralizer Cent(S) is a closed
subgroup of Aut(X). This is a consequence of Lemma 6.
For a subset S ⊆ Aut(X) its dimension is defined by
there exists a variety A of dimension d and an indim S := sup d
.
jective morphism ν : A → Aut(X) with image in S
IS THE AFFINE SPACE DETERMINED BY ITS AUTOMORPHISM GROUP?
5
The following lemma generalizes the classical dimension estimate to morphisms A → Aut(X).
Lemma 7. If ν : A → Aut(X) is a morphism, then dim ν(A) ≤ dim A.
Proof. Let η : B → Aut(X) be an injective morphism such that η(B) ⊆
ν(A). The statement follows if we prove dim B ≤ dim A. Since η is injective,
there exist points x1 , . . . , xk ∈ X such that the map
η′ : B → X
· · × X} ,
| × ·{z
b 7→ (η(b)(x1 ), . . . , η(b)(xk ))
n-times
is injective, see e.g. [Ram64, Lemma 1]. Let
ν′ : A → X
· · × X} ,
| × ·{z
a 7→ (ν(a)(x1 ), . . . , ν(a)(xk )) .
n-times
Since η(B) ⊆ ν(A), we get η ′ (B) ⊆ ν ′ (A) and thus dim B = dim η ′ (B) ≤
dim ν ′ (A) ≤ dim A.
For a subgroup G ⊆ Aut(X), the identity component G◦ ⊆ G is defined
by
G◦ =
g∈G
there exists an irreducible variety A and a morphism
ν : A → Aut(X) with image in G such that g, e ∈ ν(A)
We call a subgroup G ⊆ Aut(X) connected if G = G◦ . In the next proposition, we list several properties of the identity component of a subgroup of
Aut(X).
Proposition 8. Let X be a variety and let G ⊆ Aut(X) be a subgroup.
Then the following holds:
(1) G◦ is a normal subgroup of G;
(2) The cosets of G◦ in G are the equivalence classes under:
there exists an irreducible variety A
and a morphism ν : A → Aut(X)
g1 ∼ g2 : ⇐⇒
with image in G such that g1 , g2 ∈ ν(A) ;
(3) For each morphism ν : A → Aut(X) with image in G, the preimage
ν −1 (G◦ ) is closed in A. In particular, if G is closed in Aut(X), then
G◦ is closed in Aut(X);
(4) If X is quasi-projective and G is closed in Aut(X), then the index
of G◦ in G is countable.
Proof. (1): One can directly see, that G◦ is a normal subgroup of G.
(2): We first show that “ ∼ ” defines an equivalence relation on G. Reflexivity and symmetry are obvious. For proving the transitivity, let g ∼ h
and h ∼ k. By definition there exist irreducible varieties A, B, morphisms
ν : A → Aut(X), η : B → Aut(X) with image in G and a1 , a2 ∈ A, b1 , b2 ∈ B
such that ν(a1 ) = g, ν(a2 ) = h, η(b1 ) = h, η(b2 ) = k. Then
A × B → Aut(X) ,
(a, b) 7→ ν(a) ◦ h−1 ◦ η(b)
.
6
H. KRAFT, A. REGETA, AND I. VAN SANTEN
is a morphism with image in G that maps (a1 , b1 ) onto g and (a2 , b2 ) onto k.
Thus g ∼ k, which proves the transitivity. In particular, G◦ is the equivalence
class with respect to ∼ which contains the neutral element e. This implies
the statement
(3): Let
k
[
Bi = ν −1 (G◦ ) ⊆ A
i=1
be the decomposition of the closure of ν −1 (G◦ ) into irreducible components
B1 , . . . , Bk . Thus Bi ∩ ν −1 (G◦ ) is non-empty. Since ν has image in G it
follows from the transitivity of “ ∼ ” that ν(Bi ) ⊆ G◦ . Thus Bi ⊆ ν −1 (G◦ )
for all i. Hence ν −1 (G◦ ) is closed in A.
(4): Let ν : A → Aut(X) be a morphism. Since ν −1 (G) ⊆ A is closed, it
has only finitely many irreducible components. This implies that its image
ν(A) meets only finitely many cosets of G◦ in G. The claim follows if we
show that there exist countably many morphisms of varieties into Aut(X)
whose images cover Aut(X). We will show this.
Since X is quasi-projective, there exists a projective variety X and an
open embedding X ⊆ X. For each polynomial p ∈ Q[x] we denote by Hilbp
the Hilbert scheme of X × X associated to the Hilbert polynomial p, and
denote by U p ⊆ Hilbp × X × X the universal family, which is by definition
flat over Hilbp . By [Gro95, Theorem 3.2], Hilbp is a projective scheme over
C. For i = 1, 2 consider the following morphisms
qi : (Hilbp × X × X) ∩ U p → Hilbp × X ,
(h, x1 , x2 ) 7→ (h, xi )
p
which are defined over Hilb . By [Gro66, Proposition 9.6.1], the points h ∈ H
where the restriction
qi |{h} : ({h} × X × X) ∩ U p → {h} × X
is an isomorphism, form a constructible subset S p of Hilbp . Now choose
locally closed subsets Sjp , j = 1, . . . , kp of Hilbp that cover S p . We equip
each Sjp with the underlying reduced scheme structure of Hilbp . Note that
(Hilbp × X × X) ∩ U p and Hilbp × X are both flat over Hilbp . Therefore,
we can apply [Gro71, Proposition 5.7] and we get that qi restricts to an
isomorphism over Sjp . Thus for each j we get a morphism of varieties
Sjp
(q1 |S p )−1
×X
j
/ (S p × X × X) ∩ U p
j
q2 |S p
j
/ Sp × X
j
/X
which defines a morphism Sjp → Aut(X). For each automorphism ϕ in
Aut(X), the closure in X × X of the graph Γϕ ⊆ X × X defines a (closed)
point in the Hilbert scheme Hilbp for a certain rational polynomial p, which
belongs to S p . Thus the images of the morphisms Sjp → Aut(X) cover
Aut(X). Since there are only countably many rational polynomials, the claim
follows.
IS THE AFFINE SPACE DETERMINED BY ITS AUTOMORPHISM GROUP?
7
We say that G is an algebraic subgroup of Aut(X) if there exists a morphism ν : H → Aut(X) of an algebraic group H onto G which is a homomorphism of groups.
The next result gives a criterion for a subgroup of Aut(X) to be algebraic.
The main argument is due to Ramanujam [Ram64].
Theorem 9. Let X be an irreducible variety and let G ⊆ Aut(X) be a
subgroup. Then the following statements are equivalent:
(1) G is an algebraic subgroup of Aut(X);
(2) there exists a morphism of a variety into Aut(X) with image G;
(3) dim G is finite and G◦ has finite index in G;
(4) there is a unique structure of an algebraic group on G such that for
each irreducible variety A we have a bijective correspondence
morphisms of
morphisms A → Aut(X)
1:1
←→
varieties A → G
with image in G
given by f 7→ ι ◦ f where ι : G → Aut(X) denotes the canonical
inclusion.
Proof. The implication (1) ⇒ (2) is obvious.
Assume that (2) holds, i.e. there is a morphism η : A → Aut(X) with
image equal to G. By Lemma 7 we get dim G ≤ dim A and hence dim G
is finite. Since A is a variety and thus has only finitely many irreducible
components, it follows from Proposition 8 (2) that G◦ has finite index in G.
This proves (2) ⇒ (3).
The implication (3) ⇒ (4) is done in [Ram64, Theorem p.26] in case
G = G◦ for irreducible A. Thus in the general case, G◦ carries the structure
of an algebraic group with the required property. Since G◦ has finite index
in G we obtain a unique structure of an algebraic group on G extending the
given structure on G◦ . It remains to see that the required property holds
for G. Note that the canonical inclusion ι : G → Aut(X) is a morphism and
thus each morphism of varieties A → G yields a morphism A → Aut(X)
by composing with ι. For the reverse, let ν : A → Aut(X) be a morphism
with image in G. Since A is irreducible, by Proposition 8 (2) there is g ∈ G
such that the image of ν lies in gG◦ . Note that the composition of ν with
θg−1 : Aut(X) → Aut(X), ϕ 7→ g−1 ◦ ϕ is a morphism with image in G◦ .
Thus θg−1 ◦ ν corresponds to a morphism A → G◦ of varieties. Using that
G → G, h 7→ gh is an isomorphism of varieties, we get that ν corresponds
to a morphism A → G of varieties. This proves (3) ⇒ (4).
The implication (4) ⇒ (1) is obvious.
2.3. Ingredients from toric geometry. Recall that a toric variety is a
normal irreducible variety X together with a regular faithful action of a
torus of dimension dim X. For details concerning toric varieties we refer to
[Ful93]. The first two lemmas are certainly well-known; for the convenience
of the reader we present for both a short proof.
8
H. KRAFT, A. REGETA, AND I. VAN SANTEN
Lemma 10. Let X be a toric variety, and let T be a torus of dimension
dim X that acts faithfully on X. Then the centralizer of T in Aut(X) is
equal to T . In particular, the image of T in Aut(X) is closed.
Proof. Let g ∈ Aut(X) such that gt = tg for all t ∈ T . By definition, there is
an open dense T -orbit in X, say U . Since gU ∩ U is non-empty, there exists
x ∈ U such that gx ∈ U . Using that U = T x we find t0 ∈ T with gx = t0 x.
Thus for each t ∈ T we get
gtx = tgx = tt0 x = t0 tx .
Using that U = T x is dense in X, we get g = t0 .
Lemma 11. Let X be a toric variety. Then the coordinate ring O(X) is
finitely generated and integrally closed.
Proof. This follows from [Kno93]. Here is a short direct proof. Let N be the
lattice of one-parameter groups corresponding to the torus that acts on X
and let Σ be the fan in NQ = N ⊗Z Q that corresponds to X. Now, O(X) is
generated as a C-algebra by the intersection of the finitely generated semigroups σ ∨ ∩ M , where M = Hom(N, Z) is the dual lattice to N , σ is a
cone in Σ and σ ∨ denotes the dual cone of σ inside MQ = M ⊗Z Q. As
the intersection of the semi-groups σ ∨ ∩ M is the intersection of the convex
rational polyhedral cone ∩σ∈Σ σ ∨ with M , it is a finitely generated semigroup
(Gordon’s Lemma, see e.g. [Ful93, Proposition 1 in §1.2]). This proves the
first claim.
T Since X is normal, every local ring OX,x is integrally closed, and O(X) =
x∈X OX,x . Hence O(X) is also integrally closed.
The next proposition is based on the study of homogeneous Ga -actions on
affine toric varieties in [Lie10]. Recall that a group action ν : G → Aut(X)
on a toric variety is called homogeneous if the torus normalizes the image
ν(G). Note that for any homogeneous Ga -action ν there is a well-defined
character χ : T → Gm , defined by the formula
t ν(s) t−1 = ν(χ(t) · s) for t ∈ T, s ∈ C.
Proposition 12. Let X be a n-dimensional quasi-affine toric variety. If X
is not a torus, then there exist homogeneous Ga -actions
η1 , . . . , ηn : Ga × X → X
such that the corresponding characters χ1 , . . . , χn are linearly independent.
The proof needs some preparation. Denote by Y the spectrum of O(X).
By Lemma 11, the variety Y is normal, and the faithful torus action on X
extends uniquely to a faithful torus action on Y , by Lemma 5.
The following notation is taken from [Lie10]. Let N be a lattice of rank n,
M = Hom(N, Z) be its dual lattice and NQ = N ⊗Z Q, MQ = M ⊗Z Q. Thus,
we have a natural pairing MQ × NQ → Q, (m, n) 7→ hm, ni. Let σ ⊂ NQ
IS THE AFFINE SPACE DETERMINED BY ITS AUTOMORPHISM GROUP?
9
∨ be the
be the strongly convex polyhedral cone that describes Y and let σM
∨
intersection of the dual cone σ in MQ with M . Thus Y = Spec R, where
M
∨
Cχm ⊆ C[M ] .
R := C[σM
]=
∨
m∈σM
For each extremal ray ρ ⊂ σ, denote by ρ⊥ the elements u ∈ MQ with
∨ and let
hu, vi = 0 for all v ∈ ρ. Moreover, let τM = ρ⊥ ∩ σM
∨
∨
∨
, e + m ∈ σM
for all m ∈ σM
\ τM } .
Sρ = { e ∈ M | e 6∈ σM
By [Lie10, Remark 2.5] we have Sρ 6= ∅, and e + m ∈ Sρ for all e ∈ Sρ
and all m ∈ τM . Let us recall the description of the homogeneous locally
nilpotent derivations on R.
Proposition 13 ([Lie10, Lemma 2.6, Theorem 2.7]). Let ρ be an extremal
ray in σ and let e ∈ Sρ . Then
∂ρ,e : R → R ,
χm 7→ hm, ρiχe+m
is a homogeneous locally nilpotent derivation of degree e, and every homogeneous locally nilpotent derivation of R is a constant multiple of some ∂ρ,e .
Proof of Proposition 12. Since X is not a torus, Y is also not a torus. Thus
σ contains extremal rays, say ρ1 , . . . , ρk and k ≥ 1. Recall that associated
to these extremal rays, there exist torus-invariant divisors V (ρ1 ), . . . , V (ρk )
in Y . Again, since X is not a torus, one of these divisors does intersect X.
Let us assume that ρ = ρ1 is an extremal ray such that V (ρ) ∩ X is nonempty. Then using the orbit-cone correspondence, one can see that Y \ X
S
is contained in the union Z = ki=2 V (ρi ), see [Ful93, §3.1]. Let e ∈ Sρ be
fixed. We claim that the Ga -action on Y associated to the locally
S nilpotent
derivation ∂ρ,e+m′ of Proposition 13 fixes Z for all m′ ∈ τM \ i≥2 ρ⊥
i .
S
′
′
Let us fix m ∈ τM with hm , vi > 0 for all v ∈ i≥2 ρi . Note that the
fixed point set of the Ga -action on Y corresponding to ∂ρ,e+m′ is the zero
set of the ideal generated by the image of ∂ρ,e+m′ . The divisor V (ρi ) is the
zero set of the kernel of the canonical C-algebra surjection
m
χ , if m ∈ ρ⊥
m
⊥
∨
∨
i
,
pi : C[σM ] → C[σM ∩ ρi ] , χ 7→
0,
otherwise
see [Ful93, §3.1]. Thus we have to prove that for all i = 2, . . . , k the composition
∂ρ,e+m′
pi
∨
∨
∨
C[σM
] −→ C[σM
] −→ C[σM
∩ ρ⊥
i ]
∨ , we
is the zero map. Since, by definition, ∂ρ,e+m′ vanishes on τM = ρ⊥ ∩ σM
∨
have only to show that for all m ∈ σM \ τM the following holds:
he + m′ + m, vi > 0
for all v ∈ ρi , i = 2, . . . , k.
This is satisfied, because hm′ , vi > 0 and he + m, vi ≥ 0 (note that e ∈ Sρ
∨ ). This proves the claim.
implies e + m ∈ σM
10
H. KRAFT, A. REGETA, AND I. VAN SANTEN
Since
spans a hyperplane in M and e 6∈ τM , we can choose m′1 , . . . , m′n ∈
S τM ⊥
τM \ i≥2 ρi such that e + m′1 , . . . , e + m′n are linearly independent in MQ .
Hence, the homogeneous locally nilpotent derivations
∂ρ,e+m′i ,
i = 1, . . . , n
define Ga -actions on Y that fix Z and thus restrict to Ga -actions on X.
′
Moreover, the character of ∂ρ,e+m′i is χi = χe+mi . In particular, χ1 , . . . , χn
are linearly independent, finishing the proof of Proposition 12.
2.4. Some topological ingredients. For the convenience of the reader,
we insert the following well-known statement.
Lemma 14. For a complex variety X, the rational singular (co)homology
groups are finitely generated.
Proof. Using the universal coefficient theorem for cohomology, it is enough to
prove this for the homology groups. If X is affine, then X has the homotopy
type of a finite CW-complex (see [Kar79] or [HM97, Theorem 1.1]), and
hence all homology groups are finitely generated. Since every variety can be
covered by finitely many open affine subvarieties and since intersections of
open affine subvarieties are again affine, the lemma follows from the MayerVietoris exact sequence.
For a variety X, the Euler characteristic is defined by
X
χ(X) =
(−1)i dimQ H i (X, Q) ,
i≥0
where H i (X, Q) denotes the i-th singular cohomology group with rational
coefficients. We will use the following properties of the Euler characteristic,
see [KP85, Appendix].
Lemma 15. The Euler characteristic has the following properties.
(1) If X is a variety and Y ⊆ X a closed subvariety, then χ(X) =
χ(Y ) + χ(X \ Y ).
(2) If X → Y is a fiber bundle which is locally trivial in the étale topology
with fiber F , then χ(X) = χ(Y )χ(F ).
2.5. Results on the fixed point variety. If G is a group, acting on a
variety X, then we denote by X G the fixed point variety of X by G.
The first result gives us a criterion for the existence of fixed points for a
p-group action.
Proposition 16. Let G be a finite p-group for a prime p and let X be
a quasi-projective G-variety. If p does not divide the Euler characteristic
χ(X), then the fixed point variety X G is non-empty.
Proof. Assume that X G is empty, i.e. every G-orbit has cardinality pk for
some k > 0. We prove by induction on the dimension of X that p divides
χ(X). Since X is quasi-projective, the same is true for the smooth locus X sm
IS THE AFFINE SPACE DETERMINED BY ITS AUTOMORPHISM GROUP?
11
and thus the geometric quotient π : X sm → X sm /G exists, see [ByB02, Theorem 4.3.1]. By generic smoothness [Har77, Corollary 10.7, Chp. III] there
exists an open dense subset U in X/G such that q := π|π−1 (U ) : π −1 (U ) → U
is étale. Since π : X → X/G is finite, q is also finite and thus q is an étale
locally trivial fiber bundle, see [Gro71, Corollaire 5.3]. Since each fiber of q
is a G-orbit, it follows by Lemma 15 that the Euler characteristic of π −1 (U )
is divisible by p. By assumption, G acts without fixed point on X \ π −1 (U )
and thus by induction hypothesis, χ(X \ π −1 (U )) is divisible by p. Using
χ(X) = χ(π −1 (U )) + χ(X \ π −1 (U ))
(see Lemma 15) we get that p divides χ(X).
The second result is a consequence of a theorem of Fogarty [Fog73].
Proposition 17. Let G be a reductive group acting on a variety X. Assume
that X is smooth at some point x ∈ X G . Then X G is smooth at x and the
tangent space satisfies Tx (X G ) = (Tx X)G .
Proof. Let us denote by X (G) ⊆ X the largest closed subscheme which
is fixed under G, see [Fog73, §2]) for details. It then follows that X G =
(X (G) )red . For x ∈ X we denoteLby Cx X the tangent cone in x, i.e. Cx X =
is the associated graded algeSpec gr OX,x where gr OX,x := i≥0 mix /mi+1
x
bra with respect to the maximal ideal mx ⊆ OX,x . By definition, there is
a closed immersion µx : Cx X ֒→ Tx X, and X is smooth at x if and only if
µx is an isomorphism. If x ∈ X G is a fixed point we obtain the following
commutative diagram of schemes where all morphisms are closed immersions
Cx (X (G) )
⊆
/ Tx (X (G) )
⊆
⊆
(Cx X)(G)
⊆
/ (Tx X)(G) .
It is shown in [Fog73, Theorem 5.2] that for a reductive group G we have
Cx (X (G) ) = (Cx X)(G) . If X is smooth at x we get (Cx X)(G) = (Tx X)(G) .
Hence all morphisms in the diagram above are isomorphisms. In particular
X (G) is smooth at x and thus X G = (X (G) )red is smooth at x. Moreover, we
get Tx (X G ) = (Tx X)G .
Remark 2. Assume that (Z/pZ)n acts faithfully on a smooth quasi-projective
variety X. If p does not divide χ(X), then dim X ≥ n.
In fact, by Proposition 16 there is a fixed point x ∈ X, and the action
of (Z/pZ)n on the tangent space Tx X is faithful [KS13, Lemma 2.2], hence
n ≤ dim Tx X = dim X.
12
H. KRAFT, A. REGETA, AND I. VAN SANTEN
3. Proof of Theorem 1 and Theorem 2
Let us introduce the following terminology for this section. Let X be a
variety and let M ⊆ Aut(X) be a subset. A map η : M → Z into a variety
Z is called regular if for every morphism ν : A → Aut(X) with image in M ,
the composition η ◦ ν : A → Z is a morphism of varieties.
3.1. Semi-invariant functions.
Lemma 18. Let X be an irreducible normal variety, and let f ∈ O(X)
be a non-constant function such that the zero set Z := VX (f ) ⊂ X is an
irreducible hypersurface. Let G ⊆ Aut(X) be a connected subgroup which
stabilizes Z. Then we have the following.
(1) The function f is a G-semi-invariant with character χ : G → C∗ ,
i.e. f (gx) = χ(g)−1 · f (x) for all x ∈ X and g ∈ G.
(2) The character χ : G → C∗ is a regular map.
For the proof we need the following description of the invertible functions
on a product variety due to Rosenlicht [Ros61, Theorem 2]. We denote for
any variety X the group of invertible functions on X by O(X)∗ .
Lemma 19. Let X1 , X2 be irreducible varieties. Then O(X1 × X2 )∗ =
O(X1 )∗ · O(X2 )∗ .
Proof of Lemma 18. (1) Since X is normal, the local ring R = OX,Z is a
discrete valuation ring. Let m be the maximal ideal of R. By assumption,
f R = mk for some k > 0. Since m is stable under G, the same is true for
mk . Hence, for every g ∈ G, there exists a unit rg ∈ R∗ such that gf = rg · f
in R. Since f and gf have no zeroes in X \ Z, it follows that rg is regular
and nonzero in X \ Z. Moreover, the open set where rg ∈ R is defined and
nonzero meets Z, hence rg is a regular invertible function on X. Consider
the map
χ : G → O(X)∗ , g 7→ rg .
For all x ∈ X \ Z, g ∈ G we get f (gx) = χ(g)(x)−1 f (x), and f (gx), f (x)
are both nonzero. Since for each morphism ν : A → Aut(X) with image in
G, the map ν̃ : A × X → X, (a, x) 7→ ν(a)(x) is a morphism, we see that
A × (X \ Z) → C∗ , (a, x) 7→ χ(ν((a))(x) = f (x) · f (ν̃(a, x))−1
is a morphism. If A is irreducible, then by Lemma 19 there exist invertible
functions q ∈ O(A)∗ and p ∈ O(X \ Z)∗ such that χ(ν((a))(x) = q(a)p(x).
If, moreover, ν(a0 ) = e ∈ G for some a0 ∈ A, then
1 = re (x) = χ(ν(a0 ))(x) = q(a0 )p(x) for all x ∈ X \ Z ,
i.e. p is a constant invertible function. In this case, the composition χ ◦
ν : A 7→ O(X)∗ has image in C∗ . Since G is connected, this implies that the
whole image of χ lies in C∗ .
IS THE AFFINE SPACE DETERMINED BY ITS AUTOMORPHISM GROUP?
13
(2) Choose x0 ∈ X \ Z. As before, for each morphism ν : A → Aut(X)
with image in G, the map
A → C∗ ,
a 7→ χ(ν(a)) = f (x0 ) · f (ν(a)(x0 ))−1
is also a morphism.
Lemma 20. Let X be an irreducible normal variety, and let G ⊆ Aut(X)
be a connected subgroup. Assume that f1 , . . . , fn ∈ O(X) have the following
properties.
(1) T
Zi := VX (fi ), i = 1, . . . , n, are irreducible G-invariant hypersurfaces;
(2) i Zi contains an isolated point.
If χi : G → C∗ is the character of fi (see Lemma 18), then
χ := (χ1 , . . . , χn ) : G → (C∗ )n
is a regular homomorphism with finite kernel.
Proof. Let G act on An by
g(a1 , . . . , an ) := (χ1 (g)−1 · a1 , . . . , χn (g)−1 · an ).
Then the map f := (f1 , . . . , fn ) : X → An is G-equivariant.
Let Y ⊆ An be
T
−1
the closure of f (X). By assumption, f (0) = i Zi contains an isolated
point, hence f : X → Y has finite degree, i.e. the field extension C(X) ⊃
C(Y ) is finite. This implies that the kernel K of χ : G → (C∗ )n is finite,
because K embeds into AutC(Y ) (C(X)). By Lemma 18, χ is regular.
3.2. Another centralizer result. For an irreducible normal variety X, we
denote by CH1 (X) the first Chow group, i.e. the free group of integral Weil
divisors modulo linear equivalence (see [Har77, §6, Chp. II]).
Proposition 21. Let X be an irreducible normal variety of dimension n
such that CH1 (X) is finite. Assume that for a prime p the group (Z/pZ)n
acts faithfully on X with a (not necessarily unique) fixed point x0 which is a
smooth point of X. Then G := CentAut(X) ((Z/pZ)n ) is a closed subgroup of
Aut(X) and the identity component G◦ is a closed torus of dimension ≤ n.
Proof. By [KS13, Lemma 2.2] we get a faithful representation of (Z/pZ)n on
Tx0 X, and thus we can find generators σ1 , . . . , σn such that (Tx0 X)σi ⊂ Tx0 X
n
is a hyperplane for each i, and that (Tx0 X)(Z/pZ) = 0. By Proposition 17,
the hypersurface X σi ⊂ X is smooth at x0 , with tangent space Tx0 (X σi ) =
(Tx0 X)σi . Hence there is a unique irreducible hypersurface Zi ⊆ X that
is contained in X σi and contains x0 ; thus Zi is G◦ -stable. Moreover,
since
T
n
(Tx0 X)(Z/pZ) = 0, it follows that x0 is an isolated point of i Zi . Since a
multiple of Zi is zero in CH1 (X), there exist G◦ -semi-invariant functions
fi ∈ O(X) such that VX (fi ) = Zi (Lemma 18), and the corresponding
characters χi define a regular homomorphism
χ = (χ1 , . . . , χn ) : G◦ → (C∗ )n
14
H. KRAFT, A. REGETA, AND I. VAN SANTEN
with a finite kernel (Lemma 20). It follows that dim G◦ ≤ n. Indeed, if
ν : A → Aut(X) is an injective morphism with image in G◦ , then χ ◦ ν : A →
(C∗ )n is a morphism with finite fibers, and so dim A ≤ n. This implies, by
Theorem 9, that G◦ ⊆ Aut(X) is an algebraic subgroup and that χ is a
homomorphism of algebraic groups with a finite kernel. Hence G◦ is a torus.
Since G is closed in Aut(X) the same holds for G◦ , see Proposition 8.
3.3. Proof of Theorem 2. Now we can prove Theorem 2 which has the
same conclusion as the proposition above, but under different assumptions.
We have to show that the assumptions of Proposition 21 are satisfied. Since
X is smooth, it follows that CH1 (X) ≃ Pic(X) is finite. Proposition 16
n
implies that the fixed point variety X (Z/pZ) ⊆ X is non-empty. Now the
claims follow from Proposition 21.
3.4. Images of maximal tori under group isomorphisms.
Proposition 22. Let X and Y be irreducible quasi-projective varieties such
that dim Y ≤ dim X =: n. Assume that the following conditions are satisfied:
(1) X is quasi-affine and toric;
(2) Y is smooth, χ(Y ) 6= 0, and Pic(Y ) is finite.
∼
If θ : Aut(X) → Aut(Y ) is an isomorphism, then dim Y = n, and θ(T )◦
is a closed torus of dimension n in Aut(Y ) for each n-dimensional torus
T ⊆ Aut(X).
Proof. Let θ : Aut(X) → Aut(Y ) be an isomorphism. Since χ(Y ) 6= 0 it
follows that there is a prime p that does not divide χ(Y ).
Let n = dim X and denote by T ⊂ Aut(X) a torus of dimension n.
We have that T = CentAut(X) (T ) (Lemma 10), and thus θ(T ) is a closed
subgroup of Aut(Y ). Let µp ⊂ T be the subgroup generated by the elements
of order p, and let G := CentAut(Y ) (θ(µp )) which is closed in Aut(X). Note
that θ(T ) ⊆ G and that dim Y = n by Remark 2. Now Theorem 2 implies
that G◦ is a closed torus of dimension ≤ n in Aut(Y ), and by Proposition 8
and Theorem 9, θ(T )◦ is a closed connected algebraic subgroup of G◦ .
In order to prove that dim θ(T )◦ ≥ n we construct closed subgroups {1} =
T0 ⊂ T1 ⊂ T2 ⊂ · · · ⊂ Tn = T with the following properties:
(i) dim Ti = i for all i;
(ii) θ(Ti ) is closed in θ(T ) for all i.
It then follows that θ(Ti )◦ is a connected algebraic subgroup of θ(T )◦ . Since
the index of θ(Ti )◦ in θ(Ti ) is countable (Proposition 8), but the index of
Ti in Ti+1 is not countable, we see that dim θ(Ti+1 )◦ > dim θ(Ti )◦ , and so
dim θ(T )◦ ≥ n.
(a) Assume first that X is a torus. Then Aut(X) contains a copy of the
symmetric groups Sn , and we can find cyclic permutations τi ∈ Aut(X) such
that Ti := CentT (τi ) is a closed subtorus of dimension i, and Ti ⊂ Ti+1 for
all 0 < i < n. It then follows that θ(Ti ) = Centθ(T ) (θ(τi )) is closed in θ(T ),
and we are done.
IS THE AFFINE SPACE DETERMINED BY ITS AUTOMORPHISM GROUP?
15
(b) Now assume that X is not a torus. By Proposition 12 there exist
one-dimensional unipotent subgroups U1 , . . . , Un of Aut(X) normalized by
T such that the corresponding characters χ1 , . . . , χn : T → C∗ are linearly
independent. Since
ker(χi ) = {t ∈ T | t ◦ ui ◦ t−1 = ui for all ui ∈ Ui } = CentT (Ui )
it follows that
Ti :=
n−i
\
ker(χk ) = CentT (U1 ∪ · · · ∪ Un−i ) ⊆ T
k=1
is a closed algebraic subgroup of T of dimension i. It follows that the image
θ(Ti ) = Centθ(T ) (θ(U1 ) ∪ · · · ∪ θ(Un )) is closed in θ(T ), and the claim follows
also in this case.
3.5. Proof of Theorem 1. Using Proposition 22, it is enough to show
that a smooth toric variety Y with finite (and hence trivial) Picard group
is quasi-affine.
For proving this, let Σ ⊂ NQ = N ⊗Z Q be the fan that describes Y where
N is a lattice of rank n. Let N ′ ⊆ N be the sublattice spanned by Σ ∩ N
and let Y ′ be the toric variety corresponding to the fan Σ in NQ′ = N ′ ⊗Z Q.
It follows from [Ful93, p. 29] that
Y ≃ Y ′ × (C∗ )k
where k = rank N/N ′ . Thus Y ′ is a smooth toric variety with trivial Picard
group. Hence it is enough to prove that Y ′ is quasi-affine and therefore we
can assume k = 0, i.e. Σ spans NQ . By [Ful93, Proposition in §3.4] we get
0 = rank Pic(Y ) = d − n
where d is the number of edges in Σ. Let σ ⊂ NQ be the convex cone spanned
by the edges of Σ and let σ ∨ denote the dual cone of σ in MQ = M ⊗Z Q
where M = Hom(N, Z). Since d = n, the edges of Σ are linearly independent
in NQ and thus σ is a simplex. From the inclusion of the cones of Σ in σ
we get a morphism f : Y → Spec C[σ ∨ ∩ M ] by [Ful93, §1.4], and since each
cone in Σ is a face of σ it is locally an open immersion (see [Ful93, Lemma in
§1.3]). This implies that f is quasi-finite and birational and thus by Zariski’s
Main Theorem [Gro61, Corollaire 4.4.9] it is an open immersion.
4. Proof of the Main Theorem
4.1. A first characterization.
∼
Proposition 23. Let X be an irreducible quasi-affine variety. If Aut(An ) →
Aut(X) is an isomorphism that maps an n-dimensional torus in Aut(An )
to an algebraic subgroup, then X ≃ An as a variety.
16
H. KRAFT, A. REGETA, AND I. VAN SANTEN
Proof. Since all n-dimensional tori in Aut(An ) are conjugate (see [BB66]),
all n-dimensional tori are sent to algebraic subgroups of Aut(X) via θ. The
standard maximal torus T in Aut(An ) acts via conjugation on the subgroup
of standard translations Tr ⊂ Aut(An ) with a dense orbit O ⊂ T and thus
we get Tr = O ◦ O.
This implies that S := θ(T ) acts on U := θ(Tr) via conjugation and we
get U = θ(O) ◦ θ(O). Hence, for fixed u0 ∈ θ(O) ⊂ U the morphism
S × S → Aut(X) ,
−1
(s1 , s2 ) 7→ s1 ◦ u0 ◦ s−1
1 ◦ s2 ◦ u0 ◦ s2
has image equal to U . Now it follows from Theorem 9 that U is a closed
(commutative) algebraic subgroup of Aut(X) with no nontrivial element of
finite order, hence a unipotent subgroup.
We claim that U has no non-constant invariants on X. Indeed, let ρ : Ga ×
X → X be the Ga -action on X coming from a nontrivial element of U . If
f ∈ O(X)U is a U -invariant, then it is easy to see that
(∗)
ρf (s, x) := ρ(f (x) · s, x)
is a Ga -action commuting with U . Since U is self-centralizing, we see that
ρf (s) ∈ U for all s ∈ Ga . Moreover, formula (∗) shows that for every finite dimensional subspace V ⊂ O(X)U the map V → U , f 7→ ρf (1), is a morphism
which is injective. Indeed, ρf (1) = ρh (1) implies that ρ(f (x), x) = ρ(h(x), x)
for all x ∈ X, hence f (x) = h(x) for all x ∈ X \ X ρ . It follows that O(X)U
is finite-dimensional. Since X is irreducible, O(X)U is an integral domain
and hence equal to C, as claimed.
Since X is irreducible and quasi-affine, the unipotent group U has a dense
orbit which is closed, and so X is isomorphic to an affine space Am . Since m
is the maximal number such that there exists a faithful action of (Z/2Z)m
on Am (see Remark 2), we finally get m = n.
If X is an affine variety, then X has the structure of a so-called affine indgroup, see e.g. [Kum02, Sta13, FK17] for more details. The following result
is a special case of [Kra17, Theorem 1.1]. It is an immediate consequence of
Proposition 23 above, because a homomorphism of affine ind-groups sends
algebraic groups to algebraic groups.
Corollary 24. Let X be an irreducible affine variety. If there is an isomorphism Aut(X) ≃ Aut(An ) of affine ind-groups, then X ≃ An as a variety.
Corollary 25. Let X be a smooth, irreducible quasi-projective variety such
that χ(X) 6= 0 and Pic(X) is finite. If there is an isomorphism Aut(An ) ≃
Aut(X) of abstract groups and if dim X ≤ n, then X ≃ An as a variety.
∼
Proof. Theorem 1 shows that for an isomorphism θ : Aut(An ) → Aut(X)
and any n-dimensional torus T ⊆ Aut(An ), the identity component of the
image S := θ(T )◦ is a closed torus of dimension n in Aut(X), dim X = n, and
∼
X is quasi-affine. Thus we can apply Theorem 1 to θ −1 : Aut(X) → Aut(An )
IS THE AFFINE SPACE DETERMINED BY ITS AUTOMORPHISM GROUP?
17
and get that θ −1 (S)◦ is a closed torus of dimension n in Aut(An ). Since
θ −1 (S)◦ ⊆ θ −1 (S) ⊆ T ,
it follows that θ −1 (S) = T , i.e. θ(T ) = S is a closed n-dimensional torus
in Aut(X). The assumptions of Proposition 23 are now satisfied for the
∼
isomorphism θ : Aut(An ) → Aut(X), and the claim follows.
4.2. Proof of the Main Theorem. If (1) holds, i.e. X is a smooth, irreducible, quasi-projective variety of dimension ≤ n such that χ(X) 6= 0 and
Pic(X) is finite, then the claim follows from Corollary 25.
Now assume that the assumptions (2) are satisfied, i.e., that X is quasiaffine and toric of dimension ≥ n. Let T ⊆ Aut(X) be a torus of maxi∼
mal dimension. We can apply Theorem 1 to an isomorphism θ : Aut(X) →
n
◦
n
Aut(A ) and find that S := θ(T ) ⊂ Aut(A ) is a closed torus of dimension
n. Since the index of the standard n-dimensional torus in its normalizer in
Aut(An ) has finite index and since all n-dimensional tori in Aut(An ) are
conjugate (see [BB66]), it follows that S has finite index in θ(T ). Hence
θ −1 (S) has finite index in T . Since T is a divisible group, θ −1 (S) = T is
an algebraic group. Thus we can apply Proposition 23 to the isomorphism
∼
θ −1 : Aut(An ) → Aut(X) and find that X ≃ An as a variety.
References
A. Bialynicki-Birula, Remarks on the action of an algebraic torus on kn , Bull.
Acad. Polon. Sci. Sér. Sci. Math. Astronom. Phys. 14 (1966), 177–181. 4.1, 4.2
[Bla16] Jérémy Blanc, Algebraic structures of groups of birational transformations, Proceedings of Symposia in Pure Mathematics, to appear (2016),
https://algebra.dmi.unibas.ch/blanc/articles/structurebir.pdf. 2.2
[ByB02] Andrzej Bial ynicki Birula, Quotients by actions of groups, Algebraic quotients.
Torus actions and cohomology. The adjoint representation and the adjoint action, Encyclopaedia Math. Sci., vol. 131, Springer, Berlin, 2002, pp. 1–82. 2.5
[Fil82]
R. P. Filipkiewicz, Isomorphisms between diffeomorphism groups, Ergodic Theory Dynamical Systems 2 (1982), no. 2, 159–171 (1983). 1
[FK17] Jean-Philippe Furter and Hanspeter Kraft, On the geometry of automorphism
groups of affine varieties, in preparation, 2017. 4.1
[Fog73] John Fogarty, Fixed point schemes, Amer. J. Math. 95 (1973), 35–51. 2.5, 2.5
[Ful93] William Fulton, Introduction to toric varieties, Annals of Mathematics Studies,
vol. 131, Princeton University Press, Princeton, NJ, 1993, The William H. Roever
Lectures in Geometry. 2.3, 2.3, 2.3, 3.5
[Gro61] Alexander Grothendieck, Éléments de géométrie algébrique. III. Étude cohomologique des faisceaux cohérents. I, Inst. Hautes Études Sci. Publ. Math. (1961),
no. 11, 167. 3.5
[Gro66]
, Éléments de géométrie algébrique. IV. Étude locale des schémas et des
morphismes de schémas. III, Inst. Hautes Études Sci. Publ. Math. (1966), no. 28,
255. 2.2
[Gro71]
, Revêtements étales et groupe fondamental, Lecture Notes in Mathematics, Vol. 224, Springer-Verlag, Berlin-New York, 1971, Séminaire de
Géométrie Algébrique du Bois Marie 1960–1961 (SGA 1), Dirigé par Alexandre Grothendieck. Augmenté de deux exposés de M. Raynaud. 2.2, 2.5
[BB66]
18
H. KRAFT, A. REGETA, AND I. VAN SANTEN
[Gro95]
[Har77]
[HM97]
[Kar79]
[Kno93]
[KP85]
[Kra17]
[KS13]
[Kum02]
[Lie10]
[LRU17]
[Ram64]
[Reg17]
[Ros61]
[Ryb95]
[Ryb02]
[Sta13]
, Techniques de construction et théorèmes d’existence en géométrie
algébrique. IV. Les schémas de Hilbert, Séminaire Bourbaki, Vol. 6, Soc. Math.
France, Paris, 1995, pp. Exp. No. 221, 249–276. 2.2
Robin Hartshorne, Algebraic geometry, Springer-Verlag, New York-Heidelberg,
1977, Graduate Texts in Mathematics, No. 52. 2.5, 3.2
Helmut A. Hamm and Nicolae Mihalache, Deformation retracts of Stein spaces,
Math. Ann. 308 (1997), no. 2, 333–345. 2.4
K. K. Karčjauskas, Homotopy properties of algebraic sets, Zap. Nauchn. Sem.
Leningrad. Otdel. Mat. Inst. Steklov. (LOMI) 83 (1979), 67–72, 103, Studies in
topology, III. 2.4
Friedrich Knop, Über Hilberts vierzehntes Problem für Varietäten mit Kompliziertheit eins, Math. Z. 213 (1993), no. 1, 33–36. 2.3
Hanspeter Kraft and Vladimir L. Popov, Semisimple group actions on the threedimensional affine space are linear, Comment. Math. Helv. 60 (1985), no. 3,
466–479. 2.4
Hanspeter Kraft, Automorphism groups of affine varieties and a characterization of affine n-space, Trans. Moscow Math. Soc., to appear (2017),
http://kraftadmin.wixsite.com/hpkraft. 1, 4.1
Hanspeter Kraft and Immanuel Stampfli, On automorphisms of the affine Cremona group, Ann. Inst. Fourier 63 (2013), no. 3, 1137–1148. 1, 2, 3.2
Shrawan Kumar, Kac-Moody groups, their flag varieties and representation theory, Progress in Mathematics, vol. 204, Birkhäuser Boston, Inc., Boston, MA,
2002. 4.1
Alvaro Liendo, Affine T-varieties of complexity one and locally nilpotent derivations, Transform. Groups 15 (2010), no. 2, 389–425. 2.3, 2.3, 13
Alvaro Liendo, Andriy Regeta, and Christian Urech, Characterisation of affine surfaces and smooth Danielewski surfaces, in preparation,
http://andriyregeta.wixsite.com/homepage, 2017. 1
C. P. Ramanujam, A note on automorphism groups of algebraic varieties, Math.
Ann. 156 (1964), 25–33. 2.2, 2.2, 2.2, 2.2
Andriy Regeta, Characterization of n-dimensional normal affine SLn -varieties,
https://arxiv.org/abs/1702.01173, 2017. 1
Maxwell Rosenlicht, Toroidal algebraic groups, Proc. Amer. Math. Soc. 12
(1961), 984–988. 3.1
Tomasz Rybicki, Isomorphisms between groups of diffeomorphisms, Proc. Amer.
Math. Soc. 123 (1995), no. 1, 303–310. 1
, Isomorphisms between groups of homeomorphisms, Geom. Dedicata 93
(2002), 71–76. 1
Immanuel Stampfli, Contributions to automorphisms of affine spaces.,
http://edoc.unibas.ch/diss/DissB_10504, 2013. 4.1
Departement Mathematik und Informatik, Universität Basel,
Spiegelgasse 1, CH-4051 Basel, Switzerland
E-mail address: Hanspeter.Kraft@unibas.ch
Mathematisches Institut, Universität zu Köln,
Weyertal 86-90, 50931 Köln, Germany
E-mail address: andriyregeta@gmail.com
Departement Mathematik und Informatik, Universität Basel,
Spiegelgasse 1, CH-4051 Basel, Switzerland
E-mail address: immanuel.van.santen@math.ch
| 4math.GR |
arXiv:1612.02734v2 [cs.LG] 22 Dec 2017
Learning in the Machine:
Random Backpropagation and the Deep
Learning Channel
Pierre Baldi1,∗, Peter Sadowski1, and Zhiqin Lu2
Abstract
Abstract: Random backpropagation (RBP) is a variant of the
backpropagation algorithm for training neural networks, where the
transpose of the forward matrices are replaced by fixed random matrices in the calculation of the weight updates. It is remarkable both
because of its effectiveness, in spite of using random matrices to communicate error information, and because it completely removes the
taxing requirement of maintaining symmetric weights in a physical
neural system. To better understand random backpropagation, we
first connect it to the notions of local learning and learning channels.
Through this connection, we derive several alternatives to RBP, including skipped RBP (SRPB), adaptive RBP (ARBP), sparse RBP,
and their combinations (e.g. ASRBP) and analyze their computational complexity. We then study their behavior through simulations
using the MNIST and CIFAR-10 bechnmark datasets. These simulations show that most of these variants work robustly, almost as well
as backpropagation, and that multiplication by the derivatives of the
activation functions is important. As a follow-up, we study also the
low-end of the number of bits required to communicate error information over the learning channel. We then provide partial intuitive
explanations for some of the remarkable properties of RBP and its
variations. Finally, we prove several mathematical results, including
the convergence to fixed points of linear chains of arbitrary length, the
Corresponding author.1 Department of Computer Science, University of California,
Irvine. 2 Department of Mathematics, University of California, Irvine.
∗
1
convergence to fixed points of linear autoencoders with decorrelated
data, the long-term existence of solutions for linear systems with a
single hidden layer and convergence in special cases, and the convergence to fixed points of non-linear chains, when the derivative of the
activation functions is included.
1
Introduction
Over the years, the question of biological plausibility of the backpropagation
algorithm, implementing stochastic gradient descent in neural networks, has
been raised several times. The question has gained further relevance due to
the numerous successes achieved by backpropagation in a variety of problems
ranging from computer vision [21, 31, 30, 14] to speech recognition [12] in
engineering, and from high energy physics [7, 26] to biology [8, 32, 1] in the
natural sciences, as well to recent results on the optimality of backpropagation [6]. There are however, several well known issues facing biological neural
networks in relation to backpropagation, these include: (1) the continuous
real-valued nature of the gradient information and its ability to change sign,
violating Dale’s Law; (2) the need for some kind of teacher’s signal to provide
targets; (3) the need for implementing all the linear operations involved in
backpropagation; (4) the need for multiplying the backpropagated signal by
the derivatives of the forward activations each time a layer is traversed; (5)
the need for precise alternation between forward and backward passes; and
(6) the complex geometry of biological neurons and the problem of transmitting error signals with precision down to individual synapses. However, perhaps the most formidable obstacle is that the standard backpropagation algorithm requires propagating error signals backwards using synaptic weights
that are identical to the corresponding forward weights. Furthermore, a related problem that has not been sufficiently recognized, is that this weight
symmetry must be maintained at all times during learning, and not just during early neural development. It is hard to imagine mechanisms by which
biological neurons could both create and maintain such perfect symmetry.
However, recent simulations [24] surprisingly indicate that such symmetry
may not be required after all, and that in fact backpropagation works more
or less as well when random weights are used to backpropagate the errors.
Our general goal here is to investigate backpropagation with random weights
and better understand why it works.
2
The foundation for better understanding random backpropagation (RBP)
is provided by the concepts of local learning and deep learning channels introduced in [6]. Thus we begin by introducing the notations and connecting
RBP to these concepts. In turn, this leads to the derivation of several alternatives to RBP, which we study through simulations on well known benchmark
datasets before proceeding with more formal analyses.
2
Setting, Notations, and the Learning Channel
Throughout this paper, we consider layered feedforward neural networks and
supervised learning tasks. We will denote such an architecture by
A[N0 , . . . , Nh , . . . , NL ]
(1)
where N0 is the size of the input layer, Nh is the size of hidden layer h,
and NL is the size of the output layer. We assume that the layers are fully
connected and let wijh denote the weight connecting neuron j in layer h − 1
to neuron i in layer h. The output Oih of neuron i in layer h is computed by:
X
Oih = fih (Sih ) where Sih =
wijh Ojh−1
(2)
j
The transfer functions fih are usually the same for most neurons, with typical exceptions for the output layer, and usually are monotonic increasing
functions. The most typical functions used in artificial neural networks are
the: identity, logistic, hyperbolic tangent, rectified linear, and softmax.
We assume that there is a training set of M examples consisting of input
and output-target pairs (I(t), T (t)), with t = 1, . . . , M. Ii (t) refers to the
i-th component of the t-th input training example, and similarly for the
target Ti (t). In addition, there is an error function E to be minimized by the
learning process. In general we will asssume standard error functions such as
the squared error in the case of regression and identity transfer functions in
the output layer, or relative entropy in the case of classification with logistic
(single class) or softmax (multi-class) units in the output layer, although this
is not an essential point.
While we focus on supervised learning, it is worth noting that several “unsupervised” learning algorithms for neural networks (e.g. autoencoders, neu-
3
ral autoregressive distribution estimators, generative adversarial networks)
come with output targets and thus fall into the framework used here.
2.1
Standard Backpropagation (BP)
Standard backpropagation implements gradient descent on E, and can be
applied in a stochastic fashion on-line (or in mini batches) or in batch form,
by summing or averaging over all training examples. For a single example,
omitting the t index for simplicity, the standard backpropagation learning
rule is easily obtained by applying the chain rule and given by:
∆wijh = −η
∂E
= ηBih Ojh−1
∂wijh
(3)
where η is the learning rate, Ojh−1 is the presynaptic activity, and Bih is
the backpropagated error. Using the chain rule, it is easy to see that the
backpropagated error satisfies the recurrence relation:
Bih =
X
∂E
h+1
h ′
Bkh+1 wki
=
(f
)
i
∂Sih
k
(4)
with the boundary condition:
BiL =
∂Ei
= Ti − OiL
∂SiL
(5)
Thus in short the errors are propagated backwards in an essentially linear
fashion using the transpose of the forward matrices, hence the symmetry
of the weights, with a multiplication by the derivative of the corresponding
forward activations every time a layer is traversed.
2.2
Standard Random Backpropagation (RBP)
Standard random backpropagation operates exactly like backpropagation except that the weights used in the backward pass are completely random and
fixed. Thus the learning rule becomes:
∆wijh = ηRih Ojh−1
(6)
where the randomly back-propagated error satisfies the recurrence relation:
4
Rih = (fih )′
X
Rkh+1 ch+1
ki
(7)
k
and the weights ch+1
are random and fixed. The boundary condition at the
ki
top remains the same:
RiL =
∂Ei
= Ti − OiL
L
∂Si
(8)
Thus in RBP the weights in the top layer of the architecture are updated by
gradient descent, identically to the BP case.
2.3
The Critical Equations
Within the supervised learning framework considered here, the goal is to find
an optimal set of weights wijh . The equations that the weights must satisfy
at any critical point are simply:
X
∂E
=
Bih (t)Ojh−1(t) = 0
∂wijh
t
(9)
Thus in general the optimal weights must depend on both the input and the
targets, as well as the other weights in the network. And learning can be
viewed as a lossy storage procedure for transferring the information contained
in the training set into the weights of the architecture.
The critical Equation 9 shows that all the necessary forward information
about the inputs and the lower weights leading up to layer h − 1 is subsumed
by the term Ojh−1(t). Thus in this framework a separate channel for communicating information about the inputs to the deep weights is not necessary.
Thus here we focus on the feedback information about the targets, contained
in the term Bih (t) which, in a physical neural system, must be transmitted
through a dedicated channel.
Note that Bih (t) depends on the output O L (t), the target T (t), as well as
all the weights in the layers above h in the fully connected case (otherwise
just those weight which are on a path from unit i in layer h to the output
units), and in two ways: through O L (t) and through the backpropagation
process. In addition, Bih (t) depends also on all the upper derivatives, i.e. the
derivatives of the activations functions for all the neurons above unit i in layer
h in the fully connected case (otherwise just those derivatives which are on a
5
path from unit i in layer h to the output units). Thus in general, in a solution
of the critical equations, the weights wijh must depend on Ojh−1, the outputs,
the targets, the upper weights, and the upper derivatives. Backpropagation
shows that it is sufficient for the weights to depend on Ojh−1, T − O, the
upper weights, and the upper derivatives.
2.4
Local Learning
Ultimately, for optimal learning, all the information required to reach a critical point of E must appear in the learning rule of the deep weights. In a
physical neural system, learning rules must also be local [6], in the sense
that they can only involve variables that are available locally in both space
and time, although for simplicity here we will focus only on locality in space.
Thus typically, in the present formalism, a local learning rule for a deep layer
must be of the form
∆wijh = F (Oih , Ojh−1, wijh )
(10)
∆wijL = F (Ti , OiL , OjL−1, wijL )
(11)
and
assuming that the targets are local variables for the top layer. Among other
things, this allows one to organize and stratify learning rules, for instance by
considering polynomial learning rules of degree one, two, and so forth.
Deep local learning is the term we use to describe the use of local learning
in all the adaptive layers of a feedforward architecture. Note that Hebbian
learning [15] is a form of local learning, and deep local learning has been
proposed for instance by Fukushima [10] to train the neocognitron architecture, essentially a feed forward convolutional neural network inspired by the
earlier neurophysiological work of Hubel and Wiesel [18]. However, in deep
local learning, information about the targets is not propagated to the deep
layers and therefore in general deep local learning cannot find solutions of
the critical equations, and thus cannot succeed at learning complex functions
[6].
6
2.5
The Deep Learning Channel
From the critical equations, any optimal neural network learning algorithm
must be capable of communicating some information about the outputs, the
targets, and the upper weights to the deep weights and, in a physical neural
system, a communication channel [28, 27] must exist to communicate this
information. This is the deep learning channel, or learning channel in short
[6], which can be studied using tools from information and complexity theory. In physical systems the learning channel must correspond to a physical
channel and this leads to important considerations regarding its nature, for
instance whether it uses the forward connections in the reverse direction or
a different set of connections. Here, we focus primarily on how information
is coded and sent over this channel.
In general, the information about the outputs and the targets communicated through this channel to wijh is denoted by Iijh (T, O L ). Although
backpropagation propagates this information from the top layer to the deep
layers in a staged way, this is not necessary and Iijh (T, O L ) could be sent
directly to the deep layer h somehow skipping all the layers above. This
observation leads immediately to the skipped variant of RBP described in
the next section. It is also important to note that in principle this inforl
mation should have the form Iijh (T, O L , wrs
for l > h, f ′ (Srl ) for l ≥ h).
However standard backpropagation shows that it is possible to send the
same information to all the synapses impinging onto the same neuron, and
thus it is possible to learn with a simpler type of information of the form
l
Iih (T, O L , wrs
for l > h, f ′ (Srl ) for l ≥ h) targeting the postsynaptic neuron i.
This class of algorithms or channels is what we call deep targets algorithms,
as they are equivalent to providing a target for each deep neuron. Furthermore, backpropagation shows that all the necessary information about the
outputs and the targets is contained in the term T − O L so that we only need
l
Iih (T − O L , wrs
for l ≥ h, f ′ (Srl ) for l > h). Standard backpropagation uses
information about the upper weights in two ways: (1)through the output O L
which appears in the error terms T − O L ; and through the backpropagation
process itself. Random backpropagation crucially shows that the information
about the upper weights contained in the backpropagation process is not necessary. Thus ultimately we can focus exclusively on information which has
l
the simple form: Iih (T − O L , rrs
for l ≥ h, f ′ (Srl ) for l ≥ h), where r denotes
a set of fixed random weights.
Thus, using the learning channel, we are interested in local learning rules
7
of the form:
l
∆wijh = F (Oih, Ojh−1, wijh , Iih (T − O L , rrs
for l ≥ h, f ′(Srl ) for l ≥ h))
(12)
In fact, here we shall focus exclusively on learning rules with the multiplicative form:
l
∆wijh = ηIih (T − O L , rrs
for l ≥ h, f ′ (Srl ) for l ≥ h)Ojh−1
(13)
corresponding to a product of the presynaptic activity with some kind of
backpropagated error information, with standard BP and RBP as a special cases. Obvious important questions, for which we will seek full or parl
tial answers, include: (1) what kinds of forms can Iih (T − O L , rrs
for l ≥
′
l
h, f (Sr ) for l ≥ h) take (as we shall see there are multiple possibilities)?
(2) what are the corresponding tradeoffs among these forms, for instance in
terms of computational complexity or information transmission? and (3) are
the upper derivatives necessary and why?
3
Random Backpropagation Algorithms and
Their Computational Complexity
We are going to focus on algorithms where the information required for the
deep weight updates Iih (T − O L , f ′ (Srl ) for l ≥ h) is produced essentially
through a linear process whereby the vector T (t) − O(t), computed in the
output layer, is processed through linear operations, i.e. additions and multiplications by constants (which can include multiplication by the upper derivatives). Standard backpropagation is such an algorithm, but there are many
other possible ones. We are interested in the case where the matrices are
random. However, even within this restricted setting, there are several possibilities, depending for instance on: (1) whether the information is progressively propagated through the layers (as in the case of BP), or broadcasted
directly to the deep layers; (2) whether multiplication by the derivatives of
the forward activations is included or not; and (3) the properties of the matrices in the learning channel (e.g. sparse vs dense). This leads to several
new algorithms. Here we will use the following notations:
• BP= (standard) backpropagation.
8
• RBP= random backpropagation, where the transpose of the feedforward matrices are replaced by random matrices.
• SRBP = skipped random backpropagation, where the backpropagated
signal arriving onto layer h is given by C h (T −O) with a random matrix
C h directly connecting the output layer L to layer h, and this for each
layer h.
• ARBP = adaptive random backpropagation, where the matrices in
the learning channel are initialized randomly, and then progressively
adapted during learning using the product of the corresponding forward and backward signals, so that ∆clrs = ηRsl+1 Orl , where R denotes
the randomly backpropagated error. In this case, the forward channel
becomes the learning channel for the backward weights.
• ASRBP = adaptive skipped random backpropagation, which combines
adaptation with skipped random backpropagation.
• The default for each algorithm involves the multiplication at each layer
by the derivative of the forward activation functions. The variants
where this multiplication is omitted will be denoted by: “(no f’)”.
• The default for each algorithm involves dense random matrices, generated for instance by sampling from a normalized Gaussian for each
weight. But one can consider also the case of random ±1 (or (0,1))
binary matrices, or other distributions, including sparse versions of the
above.
• As we shall see, using random weights that have the same sign as
the forward weights is not essential, but can lead to improvements in
speed and stability. Thus we will use the word “congruent weights”
to describe this case. Note that with fixed random matrices in the
learning channel initialized congruently, congruence can be lost during
learning when the sign of a forward weight changes.
SRBP is introduced both for information theoretic reasons– what happens if the error information is communicated directly?–and because it may
facilitate the mathematical analyses since it avoids the backpropagation process. However, in one of the next sections, we will also show empirically that
SRBP is a viable learning algorithm, which in practice can work even better
9
than RBP. Importantly, these simulation results suggest that when learning the synaptic weight wijh the information about all the upper derivatives
(f ′ (Srl ) for l ≥ h)) is not needed. However the immediate (l = h) derivative
f ′ (Sih ) is needed.
Note this suggests yet another possible algorithm, skipped backropagation
(SBP). In this case, for each training example and at each epoch, the matrix
used in the feedback channel is the product of the corresponding transposed
forward matrices, ignoring multiplication by the derivative of the forward
transfer functions in all the layers above the layer under consideration. Multiplication by the derivative of the forward transfer functions is applied to
the layer under consideration. Another possibility is to have a combination
of RBP and SRBP in the learning channel, implemented by a combination of
long-ranged connections carrying SRBP signals with short-range connections
carrying a backpropagation procedure, when no long-range signals are available. This may be relevant for biology since combinations of long-ranged and
short-ranged feedback connections is common in biological neural systems.
In general, in the case of linear networks, f ′ = 1 and therefore including
or excluding derivative terms makes no difference. Furthermore, for any
linear architecture A[N, . . . , N, . . . , N] where all the layers have the same
size, then RBP is equivalent to SRBP. However, if the layers do not have the
same size, then the layer sizes introduce rank constraints on the information
that is backpropagated through RBP that may differ from the information
propagated through SRBP. In both the linear and non-linear cases, for any
network of depth 3 (L = 3), RBP is equivalent to SRBP, since there is only
one random matrix.
Additional variations can be obtained by using dropout, or multiple sets of
random matrices, in the learning channel, for instance for averaging purposes.
Another variation in the skipped case is cascading, i.e. allowing backward
matrices in the learning channel between all pairs of layers. Note that the
notion of cascading increases the number of weights and computations, yet
it is still interesting from an exploratory and robustness point of view.
3.1
Computational Complexity Considerations
The number of computations required to send error information over the
learning channel is a fundamental quantity which, however, depends on the
computational model used and the cost associated with various operations.
Obviously, everything else being equal, the computational cost of BP and
10
RBP are basically the same since they differ only by the value of the weights
being used. However more subtle differences can appear with some of the
other algorithms, such as SRBP.
To illustrate this, consider an architecture A[N0 , . . . , Nh , . . . , NL ], fully
connected, and let W be the total number of weights. In general, the primary
cost of BP is the multiplication of each synaptic weight by the corresponding
signal in the backward pass. Thus it is easy to see that the bulk of the
operations required for BP to compute the backpropagated signals scale like
O(W ) (in fact Θ(W )) with:
W = N0 × N1 + N1 × N2 . . . + NL−1 × NL =
L−1
X
Nk Nk+1
(14)
k=0
Note that whether biases are added separately or, equivalently, implemented
by adding a unit clamped to one to each layer, does not change the scaling.
Likewise, adding the costs associated with the sums computed by each neuron
and the multiplications by the derivatives of the activation functions does not
change the scaling, as long as these operations have costs that are within a
constant multiplicative factor of the cost for multiplications of signals by
synaptic weights.
As already mentioned, the scaling for RBP is obviously the same, just
using different matrices. However the corresponding term for SRBP is given
by
′
W = NL × N1 + NL × N2 . . . NL × NL−1 = NL
k=L−1
X
Nk
(15)
k=1
In this sense, the computational complexity of BP and SRBP is identical if
all the layers have the same size, but it can be significantly different otherwise, especially taking into consideration the tapering off associated with
most architectures used in practice. In a classification problem, for instance,
NL = 1 and all the random matrices in SRBP have rank 1, and W ′ scales
like the total number of neurons, rather than the total number of forward
connections. Thus, provided it leads to effective learning, SRBP could lead
to computational savings in a digital computer. However, in a physical neural system, in spite of these savings, the scaling complexity of BP and SRBP
could end up being the same. This is because in a physical neural system,
once the backpropagated signal has reached neuron i in layer h it still has to
11
be communicated to the synapse. A physical model would have to specify the
cost of such communication. Assuming one unit cost, both BP and SRBP
would require Θ(W ) operations across the entire architecture. Finally, a full
analysis in a physical system would have to take into account also costs associated with wiring, and possibly differential costs between long and short
wires as, for instance, SRBP requires longer wires than standard BP or RBP.
4
Algorithm Simulations
In this section, we simulate the various algorithms using standard benchmark
datasets. The primary focus is not on achieving state-of-the-art results, but
rather on better understanding these new algorithms and where they break
down. The results are summarized in Table 1 at the end.
4.1
MNIST
Several learning algorithms were first compared on the MNIST [22] classification task. The neural network architecture consisted of 784 inputs,
four fully-connected hidden layers of 100 tanh units, followed by 10 softmax
output units. Weights were initialized by sampling from a scaled normal
distribution [11]. Training was performed for 100 epochs using mini-batches
of size 100 with an initial learning rate of 0.1, decaying by a factor of 10−6
after each update, and no momentum. In Figure 1, the performance of each
algorithm is shown on both the training set (60,000 examples) and test set
(10,000 examples). Results for the adaptive versions of the random propagation algorithms are shown in Figure 2, and results for the sparse versions
are shown in Figure 3.
The main conclusion is that the general concept of RBP is very robust
and works almost as well as BP. Performance is unaffected or degrades gracefully when the the random backwards weights are initialized from different
distributions or even change during training. The skipped versions of the algorithms seem to work slightly better than the non-skipped versions. Finally,
it can be used with different neuron activation functions, though multiplying
by the derivative of the activations seem to play an important role.
12
1.00
BP
SBP
RBP
SRBP
Top layer only
BP (no-f©)
SBP (no-f©)
RBP (no-f©)
SRBP (no-f©)
Train Accuracy
0.95
0.90
0.85
0.80
0.75
0.70
0
20
40
60
80
100
Epoch
1.00
BP
SBP
RBP
SRBP
Top layer only
BP (no-f©)
SBP (no-f©)
RBP (no-f©)
SRBP (no-f©)
Test Accuracy
0.95
0.90
0.85
0.80
0.75
0.70
0
20
40
60
80
100
Epoch
Figure 1: MNIST training (upper) and test (lower) accuracy, as a function
of epoch, for nine different learning algorithms: backpropagation (BP), skip
BP (SBP), random BP (RBP), skip random BP (SRBP), the version of each
algorithm in which the error signal is not multiplied by the derivative of the
post-synaptic transfer function (no-f’), and the case where only the top layer
is trained while the lower layer weights are fixed (Top Layer Only). Note
that these algorithms differ only in how they backpropagate error signals to
the lower layers; the top layer is always updated according to the typical
gradient descent rule. Models are trained five times with different weight
initializations; the trajectory of the mean is shown here.
13
1.00
BP
RBP
SRBP
ARBP
ASRBP
Train Accuracy
0.98
0.96
0.94
0.92
0.90
0
20
40
60
80
100
Epoch
1.00
BP
RBP
SRBP
ARBP
ASRBP
Test Accuracy
0.98
0.96
0.94
0.92
0.90
0
20
40
60
80
100
Epoch
Figure 2: MNIST training (upper) and test (lower) accuracy, as a function of
training epoch, for the adaptive versions of the RBP algorithm (ARBP) and
SRBP algorithm (ASRBP). In these simulations, adaption slightly improves
the performance of SRBP and speeds up training. For the ARBP algorithm,
the learning rate was reduced by a factor of 0.1 in these experiments to keep
the weights from growing too quickly. Models are trained five times with
different weight initializations; the trajectory of the mean is shown here.
14
1.00
BP
RBP
RBP Sparse-8
RBP Sparse-2
RBP Sparse-1
SRBP
SRBP Sparse-8
SRBP Sparse-2
SRBP Sparse-1
Top layer only
Train Accuracy
0.95
0.90
0.85
0.80
0.75
0.70
0
20
40
60
80
100
Epoch
1.00
BP
RBP
RBP Sparse-8
RBP Sparse-2
RBP Sparse-1
SRBP
SRBP Sparse-8
SRBP Sparse-2
SRBP Sparse-1
Top layer only
Test Accuracy
0.95
0.90
0.85
0.80
0.75
0.70
0
20
40
60
80
100
Epoch
Figure 3: MNIST training (upper) and test (lower) accuracy, as a function
of training epoch, for the sparse versions of the RBP and SRBP algorithms.
Experiments are run with different levels of sparsity by controlling the expected number n of non-zero connections sent from one neuron to any other
layer it is connected to in the backward learning channel. The random backpropagation matrix connecting any two layers is created by sampling each
entry using a (0,1) Bernoulli distribution, where each element is 1 with probability p = n/(fan − in) and 0 otherwise. For example, in SRBP (Sparse-1),
each of the 10 softmax outputs sends a non-zero (hence with a weight equal
to 1) connection to an average of one neuron in each of the hidden layers.
We compare to the (Normal) versions of RBP and SRBP, where the elements
of these matrices are initialized from a standard Normal distribution scaled
in the same way as the forward weight matrices [11]. Models are trained
five times with different weight initializations; the trajectory of the mean is
shown here.
15
1.00
Final accuracy
0.98
0.96
0.94
0.92
0.90
Train
Test
0
20
40
60
80
100
n
Figure 4: MNIST post-training accuracy for the sparse versions of the SRBP
algorithm. For extreme values of n, sparse SRBP fails: for n = 0, all the
backward weights are set to zero and no error signals are sent; for n = 100 all
the backward weights are set to 1, and all the neurons in a given layer receive
the same error signal. The performance of the algorithm is surprisingly robust
in between these extremes. For sparse √
RBP (not shown), the backward
weights should be scaled by a factor of 1/ n to avoid an exponential growth
in the error signals of the lower layers.
4.2
Additional MNIST Experiments
In addition to the experiments presented above, the following observations
were made by training on MNIST with other variations of these algorithms:
1. If the matrices of the learning channel in RBP are randomly changed
at each stochastic mini-batch update, sampled from a distribution with
mean 0, performance is poor and similar to training only the top layer.
2. If the matrices of the learning channel in RBP are randomly changed
at each stochastic mini-batch update, but each backwards weight is
16
constrained to have the same sign as the corresponding forward weight,
then training error goes to 0%. This is the sign-concordance algorithm
explored by Liao, et al. [23].
3. If the elements of the matrices of the learning channel in RBP or SRBP
are sampled from a uniform or normal distribution with non-zero mean,
performance is unchanged. This is also consistent with the sparsity
experiments above, where the means of the sampling distributions are
not zero.
4. Updates to a deep layer with RBP or SRBP appear to require updates
in the precedent layers in the learning channel. If we fix the weights in
layer h, while updating the rest of the layers with SRBP, performance
is often worse than if we fix layers l ≤ h.
5. If we remove the magnitude information from the SRBP updates, keeping only the sign, performance is better than the Top Layer Only algorithm, but not as good as SRBP. This is further explored in the next
section.
6. If we remove the sign information from the SRBP updates, keeping
only the absolute value, things do not work at all.
7. If a different random backward weight is used to send an error signal
to each individual weight, rather than to a hidden neuron which then
updates all it’s incoming weights, things do not work at all.
8. The RBP learning rules work with different transfer functions as well,
including linear, logistic, and ReLU (rectified linear) units.
4.3
CIFAR-10
To further test the validity of these results, we performed similar simulations
with a convolutional architecture on the CIFAR-10 dataset [20]. The specific
architecture was based on previous work [16], and consisted of 3 sets of
convolution and max-pooling layers, followed by a densely-connected layer of
1024 tanh units, then a softmax output layer. The input consists of 32-by-32
pixel 3-channel images; each convolution layer consists of 64 tanh channels
with 5x5 kernel shape and 1x1 strides; max-pooling layers have 3x3 receptive
fields and 2x2 strides. All weights were initialized by sampling from a scaled
17
normal distribution [11], and updated using stochastic gradient descent on
mini-batches of size 128 and a momentum of 0.9. The learning rate started
at 0.01 and decreased by a factor of 10−5 after each update. During training,
the training images are randomly translated up to 10% in either direction,
horizontally and vertically, and flipped horizontally with probability p = 0.5.
Examples of results obtained with these 2D convolutional architectures
are shown in Figures 5 and 6. Overall they are very similar to those obtained
on the MNIST dataset.
18
1.0
BP
SBP
RBP
SRBP
Top layer only
BP (no-f©)
SBP (no-f©)
RBP (no-f©)
SRBP (no-f©)
Train Accuracy
0.8
0.6
0.4
0.2
0.0
0
20
40
60
80
100
Epoch
1.0
BP
SBP
RBP
SRBP
Top layer only
BP (no-f©)
SBP (no-f©)
RBP (no-f©)
SRBP (no-f©)
Test Accuracy
0.8
0.6
0.4
0.2
0.0
0
20
40
60
80
100
Epoch
Figure 5: CIFAR-10 training (upper) and test (lower) accuracy, as a function
of training epoch, for nine different learning algorithms: backpropagation
(BP), skip BP (SBP), random BP (RBP), skip random BP (SRBP), the
version of each algorithm in which the error signal is not multiplied by the
derivative of the post-synaptic transfer function (no-f’), and the case where
only the top layer is trained while the lower layer weights are fixed (Top Layer
Only). Models are trained five times with different weight initializations; the
trajectory of the mean is shown here.
19
1.0
BP
RBP
RBP Sparse-8
RBP Sparse-2
RBP Sparse-1
SRBP
SRBP Sparse-8
SRBP Sparse-2
SRBP Sparse-1
Top layer only
Train Accuracy
0.8
0.6
0.4
0.2
0.0
0
20
40
60
80
100
Epoch
1.0
BP
RBP
RBP Sparse-8
RBP Sparse-2
RBP Sparse-1
SRBP
SRBP Sparse-8
SRBP Sparse-2
SRBP Sparse-1
Top layer only
Test Accuracy
0.8
0.6
0.4
0.2
0.0
0
20
40
60
80
100
Epoch
Figure 6: CIFAR-10 training (upper) and test (lower) accuracy for the sparse
versions of the RBP and SRBP algorithms. Experiments are run with different levels of sparsity by controlling the expected number n of non-zero
connections sent from one neuron to any other layer it is connected to in the
backward learning channel. The random backpropagation matrix connecting
any two layers is created by sampling each entry using a (0,1) Bernoulli distribution, where each element is 1 with probability p = n/(fan − in) and 0
otherwise. We compare to the (Normal) versions of RBP and SRBP, where
the elements of these matrices are initialized from a standard Normal distribution scaled in the same way as the forward weight matrices [11]. Models
are trained five times with different weight initializations; the trajectory of
the mean is shown here.
20
BP
RBP
MNIST Baseline
97.9 (0.1) 97.2 (0.1)
′
No-f
89.9 (0.3) 88.3 (1.1)
Adaptive
97.3 (0.1)
Sparse-8
96.0 (0.4)
Sparse-2
96.3 (0.5)
Sparse-1
90.3 (1.1)
Quantized error 5-bit
97.6
95.4
Quantized error 3-bit
96.5
92.5
Quantized error 1-bit
94.6
89.8
Quantized update 5-bit
95.2
94.0
Quantized update 3-bit
96.5
91.0
Quantized update 1-bit
92.5
9.6
LC Dropout 10%
97.7
96.5
LC Dropout 20%
97.8
96.7
LC Dropout 50%
97.7
96.7
CIFAR-10 Baseline
83.4 (0.2) 70.2 (1.1)
′
No-f
54.8 (3.6) 32.7 (6.2)
Sparse-8
46.3 (4.3)
Sparse-2
62.9 (0.9)
Sparse-1
56.7 (2.6)
SRBP
Top layer only
97.2 (0.2)
84.7 (0.7)
88.4 (0.7)
97.3 (0.1)
96.9 (0.1)
95.8 (0.2)
94.6 (0.6)
95.1
93.2
91.6
93.3
92.2
90.7
97.1
97.2
97.1
72.7 (0.8)
47.9 (0.4)
39.9 (3.9)
70.9 (0.7)
65.7 (1.9)
62.6 (1.8)
Table 1: Summary of experimental results showing the final test accuracy (in
percentages) for the RBP algorithms after 100 epochs of training on MNIST
and CIFAR-10. For the experiments in this section, training was repeated
five times with different weight initializations; in these cases the mean is
provided, with the sample standard deviation in parentheses. Also included
are the quantization results from Section 5, and the experiments applying
dropout to the learning channel from Section 6.
5
5.1
Bit Precision in the Learning Channel
Low-Precision Error Signals
In the following experiment, we investigate the nature of the learning channel
by quantizing the error signals in the BP, RBP, and SRBP algorithms. This is
21
distinct from other work that uses quantization to reduce computation [17] or
memory [13] costs. Quantization is not applied to the forward activations or
weights; quantization is only applied to the backpropagated signal received by
each hidden neuron, Jih (T −O L ), where each weight update after quantization
is given by
∆wijh = Iih (T − O L ) × Ojh−1
= Quantize
Jih (T
(16)
L
−O ) ×
(fih )′
×
Ojh−1
(17)
where (fih )′ is the derivative of the activation function and
Iih (T − O L ) = Jih (T − O L ) × (fih )′
(18)
in the non-quantized update. We define the quantization formula used here
as
Quantizeα,bits (x) = α × sign(x) × 2round(clip(log2| α |,
x
−bits+1, 0))
(19)
where bits is the number of bits needed to represent 2bits possible values and
α is a scale factor such that the quantized values fall in the range [−α, α].
Note that this definition is identical to the quantization function defined in
Hubara, et al. [17], except that this definition is more general in that α is
not constrained to be a power of 2.
In BP and RBP, the quantization occurs before the error signal is backpropagated to previous layers, so the quantization errors accumulate. In
experiments, we used a fixed scale parameter α = 2−3 and varied the bit
width bits. Figure 7 shows that the performance degrades gracefully as the
precision of the error signal decreases to small values; for larger values, e.g.
bits = 10, the performance is indistinguishable from the unquantized updates
with 32-bit floats.
22
1.00
BP
BP 5-bits
BP 3-bits
BP 1-bit
SRBP
SRBP 5-bits
SRBP 3-bits
SRBP 1-bit
RBP
RBP 5-bits
RBP 3-bits
RBP 1-bit
0.98
Train Accuracy
0.96
0.94
0.92
0.90
0.88
0.86
0
20
40
60
80
100
Epoch
1.00
BP
BP 5-bits
BP 3-bits
BP 1-bit
SRBP
SRBP 5-bits
SRBP 3-bits
SRBP 1-bit
RBP
RBP 5-bits
RBP 3-bits
RBP 1-bit
0.98
Test Accuracy
0.96
0.94
0.92
0.90
0.88
0.86
0
20
40
60
80
100
Epoch
Figure 7: MNIST training (upper) and test (lower) accuracy, as a function
of training epoch, for the sparse versions of the RBP and SRBP algorithms.
Experiments are run with different levels of quantization of the error signal
by controlling the bitwidth bits, according to the formula given in the text
(Equation 19).
23
5.2
Low-Precision Weight Updates
The idea of using low-precision weight updates is not new [25], and Liao, et
al. [23] recently explored the use of low-precision updates with RBP. In the
following experiment, we investigate the robustness of both RBP and SRBP
to low-precision weight updates by controlling the degree of quantization.
Equation 19 is again used for quantization, with the scale factor reduced to
α = 2−6 since weight updates need to be small. The quantization is applied
after the error signals have been backpropagated to all the hidden layers,
but before summing over the minibatch; as in the previous experiments, we
use minibatch updates of size 100, a non-decaying learning rate of 0.1, and
no momentum term (Figure 8). The main conclusion is that even very lowprecision updates to the weights can be used to train an MNIST classifier to
90% accuracy, and that low-precision weight updates appear to degrade the
performance of BP, RBP, and SRBP in roughly the same way.
24
1.00
BP
BP 5-bits
BP 3-bits
BP 1-bit
SRBP
SRBP 5-bits
SRBP 3-bits
SRBP 1-bit
RBP
RBP 5-bits
RBP 3-bits
RBP 1-bit
0.98
Train Accuracy
0.96
0.94
0.92
0.90
0.88
0.86
0
20
40
60
80
100
Epoch
1.00
BP
BP 5-bits
BP 3-bits
BP 1-bit
SRBP
SRBP 5-bits
SRBP 3-bits
SRBP 1-bit
RBP
RBP 5-bits
RBP 3-bits
RBP 1-bit
0.98
Test Accuracy
0.96
0.94
0.92
0.90
0.88
0.86
0
20
40
60
80
100
Epoch
Figure 8: MNIST training (upper) and test (lower) accuracy, as a function
of training epoch, for the sparse versions of the RBP and SRBP algorithms.
Experiment are carried with different levels of quantization of the weight
updates by controlling the bitwidth bits, according to the formula given in
the text (Equation 19). Quantization is applied to each example-specific
update, before summing the updates within a minibatch.
25
6
Observations
In this section, we provide a number of simple observations that provide some
intuition for some of the previous simulation results and why RBP and some
of its variations may work. Some of these observations are focused on SRBP
which in general is easier to study than standard RBP.
Fact 1: In all these RBPs algorithms, the L-layer at the top with parameters wijL follows the gradient, as it is trained just like BP, since there are no
random feedback weights used for learning in the top layer. In other words,
BP=RBP=SRBP for the top layer.
Fact 2: For a given input, if the sign of T −O is changed, all the weights updates are changed in the opposite direction. This is true of all the algorithms
considered here–BP, RBP, and their variants–even when the derivatives of
the activations are included.
Fact 3: In all RBP algorithms, if T − O = 0 (online or in batch mode) then
for all the weights ∆wijh = 0 (on line or in batch mode).
Fact 4: Congruence of weights is not necessary. However it can be helpful
sometimes and speed up learning. This can easily be seen in simple cases.
For instance, consider a linear or non-linear A[N0 , N1 , 1] architecture with
coherent weights, and denote by a the weights in the bottom layer, by b the
weights in the top layer, and by c the weights in the learning channel. Then,
for all variants of RBP, all the weights updates are in the same direction as
the gradient. This is obvious for the top layer (Fact 1 above). For the first
layer of weights, the changes are given by ∆wij1 = η(T − O)ci Ij , which is very
similar to the change produced by gradient descent ∆1ij = η(T − O)bi Ij since
ci and bi are assumed to be coherent. So while the dynamics of the lower layer
is not exactly in the gradient direction, it is always in the same orthant as
the gradient and thus downhill with respect to the error function. Additional
examples showing the positive but not necessary effect of coherence are given
in Section 7.
Fact 5: SRBP seems to perform well showing that the upper derivatives
are not needed. However the derivative of the corresponding layer seem
to matter. In general, for the activation functions considered here, these
derivatives tend to be between 0 and 1. Thus learning is attenuated for
neurons that are saturated. So an ingredient that seems to matter is to let
the synapses of neurons that are not saturated change more than the synapses
of neurons that are saturated (f ′ close to 0).
Fact 6: Consider a multi-class classification problem, such as MNIST. All
26
the elements in the same class tend to receive the same backpropagated
signal and tend to move in unison. For instance, consider the the beginning
of learning, with small random weights in the forward network. Then all the
images will tend to produce a more or less uniform output vector similar to
(0.1, 0.1, . . . , 0.1). Thus all the images in the “0” class will tend to produce
a more or less uniform error vector similar to (0.9, −0.1, . . . , −0.1). All the
images in the “1” class will tend to produce a more or less uniform error
vector similar to (−0.1, 0.9, . . . , −0.1), which is essentially orthogonal to the
previous error vector, and so forth. In other words, the 10 classes can be
associated with 10 roughly orthogonal error vectors. When these vectors are
multiplied by a fixed random matrix, as in SRBP, they will tend to produce
10 approximately orthogonal vectors in the corresponding hidden layer. Thus
the backpropagated error signals tend to be similar within one digit class,
and orthogonal across different digit classes. At the beginning of learning,
we can expect roughly half of them (5 digits out of 10 in the MNIST case)
to be in the same direction as BP.
Thus, in conclusion, an intuitive picture of why RBP may work is that:
(1) the random weights introduce a fixed coupling between the learning dynamics of the forward weights (see also mathematical analyses below); (2)
the top layer of weights always follows gradient descent and stirs the learning
dynamic in the right direction; and (3) the learning dynamic tends to cluster
inputs associated with the same response and move them away from other
similar clusters. Next we discuss a possible connection to dropout.
6.1
Connections to Dropout
Dropout [16, 5] is a very different training algorithm which, however, is also
based on using some form of randomness. Here we explore some possible
connections to RBP.
First observe that the BP equations can be viewed as a form of dropout
averaging equations, in the sense that, for a fixed example, they compute
the ensemble average activity of all the units in the learning channel. The
ensemble average is taken over all the possible backpropagation networks
where each unit is dropped stochastically, unit i in layer h being dropped
with probability 1 − f ′ (Soh ) [ assuming the derivatives of the transfer functions are always between 0 and 1 inclusively, which is the case for the standard transfer functions, such as the logistic or the rectified linear transfer
functions–otherwise some rescaling is necessary]. Note that in this way the
27
dropout probabilities change with each example and units that are more saturated are more likely to be dropped, consistently with the remark above
that saturated units should learn less.
In this view there are two kinds of noise: (1) choice of the dropout probabilities which vary with each example; (2) the actual dropout procedure.
Consider now adding a third type of noise on all the symmetric weights in
the backward pass in the form
wijh + ξijh
(20)
and assume for now that E(ξijh ) = 0. The distribution of the noise could be
Gaussian for instance, but this is not essential. The important point is that
the noise on a weight is independent of the noise on the other weights, as well
as independent of the dropout noise on the units. Under these assumptions,
as shown in [5], the expected value of the activity of each unit in the backward
pass is exactly given by the standard BP equations and equal to Bih for unit
i in layer h. In other words, standard backpropagation can be viewed as
computing the exact average over all backpropagation processes implemented
on all the stochastic realizations of the backward network under the three
forms of noise described above. Thus we can reverse this argument and
consider that RBP approximates this average or BP by averaging over the
first two kinds of noise, but not the third one where, instead of averaging, a
random realization of the weights is selected and then fixed at all epochs. This
connection suggests other intermediate RBP variants where several samples
of the weights are used, rather than a single one.
Finally, it is possible to use dropout in the backward pass. The forward
pass is robust to dropping out neurons, and in fact the dropout procedure
can be beneficial [16, 5]. Here we apply the dropout procedure to neurons in
the learning channel during the backward pass. The results of simulations
are reported in Figure 9 and confirm that BP, RBP, SRBP, are robust with
respect to dropout.
28
1.00
BP
BP 10%
BP 20%
BP 50%
SRBP
SRBP 10%
SRBP 20%
SRBP 50%
RBP
RBP 10%
RBP 20%
RBP 50%
0.98
Train Accuracy
0.96
0.94
0.92
0.90
0.88
0.86
0
20
40
60
80
100
Epoch
1.00
BP
BP 10%
BP 20%
BP 50%
SRBP
SRBP 10%
SRBP 20%
SRBP 50%
RBP
RBP 10%
RBP 20%
RBP 50%
0.98
Test Accuracy
0.96
0.94
0.92
0.90
0.88
0.86
0
20
40
60
80
100
Epoch
Figure 9: MNIST training (upper) and test (lower) accuracy, as a function of
training epoch, for BP, RBP, and SRBP with different dropout probabilities
in the learning channel: 0% (no dropout), 10%, 20%, and 50%. For dropout
probability p, the error signals that are not dropped out are scaled by 1/(1 −
p). As with dropout in the forward propagation, large dropout probabilities
lead to slower training without hurting final performance.
29
7
7.1
Mathematical Analysis
General Considerations
The general strategy to try to derive more precise mathematical results is to
proceed from simple architectures to more complex architectures, and from
the linear case to the non-linear case. The linear case is more amenable to
analysis and, in this case, RBP and SRBP are equivalent when there is only
one hidden layer, or when all the layers have the same size. Thus we study the
convergence of RBP to optimal solutions in linear architectures of increasing
complexity: A[1, 1, 1], A[1, 1, 1, 1], A[1, 1, . . . , 1], A[1, N, 1] A[N, 1, N], and
then the general A[N0, N1, N2] case with a single hidden layer. This is
followed by the study of a non-linear A[1, 1, 1] case.
For each kind of linear network, under a set of standard assumptions,
one ca derive a set of non-linear–in fact polynomial–autonomous, ordinary
differential equations (ODEs) for the average (or batch) time evolution of the
synaptic weights under the RBP or SRBP algorithm. As soon as there is more
than one variable and the system is non-linear, there is no general theory to
understand the corresponding behavior. In fact, even in two dimensions,
the problem of understanding the upper bound on the number and relative
position of the limit cycles of a system of the form dx/dt = P (x, y) and
dy/dt = Q(x, y), where P and Q are polynomials of degree n is open–in fact
this is Hilbert’s 16-th problem in the field of dynamical systems [29, 19].
When considering the specific systems arising from the RBP/SRBP learning equations, one must first prove that these systems have a long-term solution. Note that polynomial ODEs may not have long-term solutions (e.g.
dx/dt = xα , with x(0) 6= 0, does not have long-term solutions for α > 1).
If the trajectories are bounded, then long-term solutions exist. We are particularly interested in long-term solutions that converge to a fixed point, as
opposed to limit cycles or other behaviors.
A number of interesting cases can be reduced to polynomial differential
equations in one dimension. These can be understood using the following
theorem.
Theorem 1: Let dx/dt = Q(x) = k0 + k1 x + . . . + kn xn be a first order
polynomial differential equation in one dimension of degree n > 1, and let
r1 < r2 . . . < rk (k ≤ n) be the ordered list of distinct real roots of Q (the
fixed points). If x(0) = ri then x(t) = ri and the solution is constant If
30
ri < x(0) < ri+1 then x(t) → ri if Q < 0 in (ri , ri+1 ), and x(t) → ri+1 if
Q > 0 in (ri , ri+1 ). If x(0) < r1 and Q > 0 in the corresponding interval, then
x(t) → r1 . Otherwise, if Q < 0 in the corresponding interval, there is no long
time solution and x(t) diverges to −∞ within a finite horizon. If x(0) > rk
and Q < 0 in the corresponding interval, then x(t) → rk . Otherwise, if Q > 0
in the corresponding interval, there is no long time solution and x(t) diverges
to +∞ within a finite horizon. A necessary and sufficient condition for the
dynamics to always converge to a fixed point is that the degree n be odd, and
the leading coefficient be negative.
Proof: The proof of this theorem is easy and can be visualized by plotting
the function Q.
Finally, in general the matrices in the forward channel are denoted by
A1 , A2 , . . . , and the matrices in the learning channel are denoted by C1 , C2, . . .
Theorems are stated in concise form and additional important facts are contained in the proofs.
7.2
The Simplest Linear Chain: A[1, 1, 1]
Derivation of the System: The simplest case correspond to a linear
A[1, 1, 1] architecture (Figure 10). Let us denote by a1 and a2 the weights
in the first and second layer, and by c1 the random weight of the learning
channel. In this case, we have O(t) = a1 a2 I(t) and the learning equations
are given by:
(
∆a1 = ηc1 (T − O)I = ηc1 (T − a1 a2 I)I
(21)
∆a2 = η(T − O)a1 I = η(T − a1 a2 I)a1 I
When averaged over the training set:
(
E(∆a1 ) = ηc1 E(IT ) − ηc1 a1 a2 E(I 2 ) = ηc1 α − ηc1 a1 a2 β
E(∆a2 ) = ηa1 E(IT ) − ηa21 a2 E(I 2 ) = ηa1 α − ηa21 a2 β
(22)
where α = E(IT ) and β = E(I 2 ). With the proper scaling of the learning
rate (η = ∆t) this leads to the non-linear system of coupled differential
equations for the temporal evolution of a1 and a2 during learning:
(
da1
= αc1 − βc1 a1 a2 = c1 (α − βa1 a2 )
dt
(23)
da2
2
=
αa
−
βa
a
=
a
(α
−
βa
a
)
1
2
1
1
2
1
dt
31
Note that the dynamic of P = a1 a2 is given by:
dP
da2
da1
= a1
+ a2
= (a21 + a2 c1 )(α − βP )
dt
dt
dt
The error is given by:
(24)
1
1
1
1
α2
1
E = E(T − P I)2 = E(T 2 ) + P 2 β − P α = E(T 2 ) + (α − βP )2 −
2
2
2
2
2β
2β
(25)
and:
dE
= −α + βP with
dP
the last equality requires ai 6= 0.
∂E
P
= (−α + βP )
∂ai
ai
(26)
Theorem 2: The system in Equation 23 always converges to a fixed point.
Furthermore, except for trivial cases associated with c1 = 0, starting from
any initial conditions the system converges to a fixed point corresponding to
a global minimum of the quadratic error function. All the fixed points are
located on the hyperbolas given by α − βP = 0 and are global minima of
the error function. All the fixed points are attractors except those that are
interior to a certain parabola. For any starting point, the final fixed point
can be calculated by solving a cubic equation.
Proof: As this is the first example, we first deal with the trivial cases in
detail. For subsequent systems, we will skip the trivial cases entirely.
Trivial Cases: 1) If β = 0 then we must have I = 0 and thus α = 0. As a
result the activity of the input, hidden, and output, neuron will always be 0.
Therefore the weights a1 and a2 will remain constant (da1 /dt = da2 /dt = 0)
and equal to their initial values a1 (0) and a2 (0). The error will also remain
constant, and equal to 0 if and only if T = 0. Thus from now on we can
assume that β > 0.
2) If c1 = 0 then the lower weight a1 never changes and remains equal to its
initial value. If this initial value satisfies a1 (0) = 0, then the activity of the
hidden and output unit remains equal to 0 at all times, and thus a2 remains
constant and equal to its initial value a2 = a2 (0). The error remains constant
and equal to 0 if only if T is always 0. If a1 (0) 6= 0, then the error is a simple
32
quadratic convex function of a2 and since the rule for adjusting a2 is simply
gradient descent, the value of a2 will converge to its optimal value given by:
a2 = α/βa1(0).
General Case: Thus from now on, we can assume that β > 0 and c1 6= 0.
Furthermore, it is easy to check that changing the sign of α corresponds to
a reflection about the a2 -axis. Likewise, changing the sign of c1 corresponds
to a reflection about the origin (i.e. across both the a1 and a2 axis). Thus in
short, it is sufficient to focus on the case where: α > 0, β > 0, and c1 > 0.
In this case, the critical points for a1 and a2 are given by:
P = a1 a2 =
α
E(IT )
=
=0
β
E(I 2 )
(27)
which corresponds to two hyperbolas in the two-dimensional (a1 , a2 ) plane,
in the first and third quadrant for α = E(IT ) > 0. Note that these critical
points do not depend on the feedback weight c1 . All these critical points correspond to global minima of the error function E = 21 E[(T −O)2]. Furthermore,
the critical points of P include also the parabola:
a21 + a2 c1 = 0 or a2 = −a21 /c1
(28)
(Figure 11). These critical points are dependent on the weights in the learning
channel. This parabola intersects the hyperbola a1 a2 = P = α/β at one point
1/3
with coordinates: a1 = (−c1 α/β)1/3 and a2 = (−α2/3 /(c1 β 2/3 ).
In the upper half plane, where a2 and c1 are congruent and both positive,
the dynamics is simple to understand. For instance in the first quadrant
where a1 , a2 , c1 > 0, if α − βP > 0 then da1 /dt > 0, da2 /dt > 0, and
dP/dt > 0 everywhere and therefore the gradient vector flow is directed
towards the hyperbola of critical points. If started in this region, a1 , a2 ,
and P will grow monotonically until a critical point is reached and the error
will decrease monotonically towards a global minimum. If α − βP < 0 then
da1 /dt < 0, da2 /dt < 0, and dP/dt < 0 everywhere and again the vector flow
is directed towards the hyperbola of critical points. If started in this region,
a1 , a2 , and P will decrease monotonically until a critical point is reached and
the error will decrease monotonically towards a global minimum. A similar
situation is observed in the fourth quadrant where a1 < 0 and a2 > 0.
More generally, if a2 and c1 have the same sign, i.e. are congruent as in
BP, then a21 + a2 c1 ≥ 0 and P will increase if α − βP > 0, and decrease
if α − βP < 0. Note however that this is also true in general when c1 is
33
Figure 10: Left: A[1, 1, 1] architecture. The weights a1 and a2 are adjustable,
and the feedback weight c1 is constant. Right: A[1, 1, 1, 1] architecture. The
weights a1 , a2 , and a3 are adjustable, and the feedback weights c1 and c2 are
constant.
small regardless of its sign, relative to a1 and a2 , since in this case it is still
true that a21 + a2 c1 is positive. This remains true even if c1 varies, as long as
it is small. When c1 is small, the dynamics is dominated by the top layer.
34
The lower layer changes slowly and the top layer adapts rapidly so that the
system again converges to a global minimum. When a2 = c1 one recovers the
convergent dynamic of BP, as dP/dt always has the same sign as α − βP .
However, in the lower half plane, the situation is slightly more complicated
(Figure 11).
To solve the dynamics in the general case, from Equation 23 we get:
da2
1 da1
= a1
dt
c1 dt
which gives a2 =
1 2
a
2c1 1
(29)
+ C so that finally:
a2 =
1 2
1 2
a1 + b(0) −
a (0)
2c1
2c1 1
(30)
Given a starting point a1 (0) and a2 (0), the system will follow a trajectory
given by the parabola in Equation 30 until it converges to a critical point
(global optimum) where da1 /dt = da2 /dt = 0. To find the specific critical point to which it converges to, Equations 30 and 27 must be satisfied
simultaneously which leads to the depressed cubic equation:
a31 + (2c1 a2 (0) − a1 (0)2 )a1 − 2
c1 α
=0
β
(31)
which can be solved using the standard formula for the roots of cubic equations. Note that the parabolic trajectories contained in the upper half plane
intersect the critical hyperbola in only one point and therefore the equation
has a single real root. In the lower half plane, the parabolas associated with
the trajectories can intersect the hyperbolas in 1, 2, or 3 distinct points corresponding to 1 real root, 2 real roots (1 being double), and 3 real roots. The
double root corresponds to the point −(c1 α/β)1/3 associated with the intersection of the parabola of Equation 30 with both the hyperbola of critical
points a1 a2 = α/β and the parabola of additional critical points for P given
by Equation 28.
When there are multiple roots, the convergence point of each trajectory is
easily identified by looking at the derivative vector flow (Figure 11). Note on
the figure that all the points on the critical hyperbolas are stable attractors,
except for those in the lower half-plane that satisfy both a1 a2 = α/β and
a2 c1 + a21 < 0. This can be shown by linearizing the system around its
critical points.
35
Linearization Around Critical Points. If we consider a small deviation
a1 + u and a2 + v around a critical point a1 , a2 (satisfying α − βa1 a2 = 0)
and linearize the corresponding system, we get:
(
du
= −βc1 (a2 u + a1 v)
dt
(32)
dv
=
−βa
(a
u
+
a
v)
1
2
1
dt
with a1 a2 = α/β. If we let w = a2 u + a1 v we have:
dw
2
= −β(c1 a2 + a21 )w thus w = w(0)e−β(c1 a2 +a1 )t
(33)
dt
Thus if β(c1 a2 + a21 ) > 0, w converges to zero and a1 , a2 is an attractor.
In particular, this is always the case when c1 is very small, or c1 has the
same sign as a2 . If β(c1 a2 + a21 ) < 0, w diverges to +∞, and corresponds to
unstable critical points as described above. If β(c1a2 +a21 ) = 0, w is constant.
Finally, note that in many cases, for instance for trajectories in the upper half plane, the value of P along the trajectories increases or decreases
monotonically towards the global optimum value. However this is not always
the case and there are trajectories where dP/dt changes sign, but this can
happen only once.
7.3
Adding Depth: the Linear Chain A[1, 1, 1, 1].
Derivation of the System: In the case of a linear A[1, 1, 1, 1] architecture,
for notational simplicity, let us denote by a1 , a2 and a3 the forward weights,
and by c1 and c2 the random weights of the learning channel (note the index
is equal to the target layer). In this case, we have O(t) = a1 a2 a3 I(t) = P I(t).
The learning equations are:
∆a1 = ηc1 (T − O)I = ηc1 (T − a1 a2 a3 I)I
(34)
∆a2 = ηc2 (T − O)a1 I = ηc2 (T − a1 a2 a3 I)a1 I
∆a3 = η(T − O)a1 a2 I = η(T − a1 a2 a3 I)a1 a2 I
When averaged over the training set:
2
E(∆a1 ) = ηc1 E(IT ) − ηc1 P E(I ) = ηc1 α − ηc1 P β
E(∆a2 ) = ηc2 a1 E(IT ) − ηc2 a1 P E(I 2 ) = ηc2 a1 α − ηc2 a1 P β
E(∆a3 ) = ηa1 a2 E(IT ) − ηa1 a2 P E(I 2 ) = ηa1 a2 β − ηa1 a2 P β
36
(35)
3
3.0
2.4
2
1.8
1.2
1
0
0.0
dP
dt
a2
0.6
−0.6
−1
−1.2
−1.8
−2
−2.4
−3
−3
−3.0
−2
−1
0
1
2
3
a1
Figure 11: Vector field for the A[1, 1, 1] linear case with c1 = 1, α = 1, and
β = 1. a1 correspond to the horizontal axis and a2 correspond to the vertical
axis. The critical points correspond to the two hyperbolas, and all critical
points are fixed points and global minima of the error functions. Arrows
are colored according to the value of dP/dt, showing how the critical points
inside the parabola a2 = −a21 /c1 are unstable. All other critical points are
attractors. Reversing the sign of α, leads to a reflection across the a2 -axis;
reversing the sign of c1 , leads to a reflection across both the a1 and a2 axes.
where P = a1 a2 a3 . With the proper scaling of the learning rate (η = ∆t)
this leads to the non-linear system of coupled differential equations for the
temporal evolution of a1 , a2 and a3 during learning:
da1
dt = c1 (α − βP )
da2
(36)
= c2 a1 (α − βP )
dt
da3
= a1 a2 (α − βP )
dt
37
The dynamic of P = a1 a2 a3 is given by:
da3
da1
da2
dP
= a1 a2
+a2 a3
+a1 a3
= (a21 a22 +c1 a2 a3 +c2 a21 a3 )(α−βP ) (37)
dt
dt
dt
dt
Theorem 3: Except for trivial cases (associated with c1 = 0 or c2 = 0),
starting from any initial conditions the system in Equation 36 converges to
a fixed point, corresponding to a global minimum of the quadratic error
function. All the fixed points are located on the hypersurface given by α −
βP = 0 and are global minima of the error function. Along any trajectory,
and for each i, ai+1 is a quadratic function of ai . For any starting point, the
final fixed point can be calculated by solving a polynomial equation of degree
seven.
Proof: If c1 = 0, a1 remains constant and thus we are back to the linear case
of a A[1, 1, 1] architecture where the inputs I are replaced by a1 I. Likewise,
if c2 = 0 a2 remains constant and the problem can again be reduced to the
A[1, 1] case with the proper adjustments. Thus for the rest of this section
we can assume c1 6= 0 and c2 6= 0.
The critical points of the system correspond to α − βP = 0 and do
not depend on the weights in the learning channel. These critical points
correspond to global minima of the error function. These critical points are
also critical points for the product P . Additional critical points for P are
provided by the hypersurface: a21 a22 + c1 a2 a3 + c2 a21 a3 = 0 with (a1 , a2 , a3 ) in
R3 .
The dynamics of the system can be solved by noting that Equation 36
yields:
da2
a1 c2 da1
=
dt
c1 dt
and
da3
a2 da2
=
dt
c2 dt
(38)
As a result:
a2 =
c2 2
a + C1
2c1 1
with C1 = a2 (0) −
c2
a1 (0)2
2c1
(39)
a3 =
1 2
a + C2
2c2 2
with C2 = a3 (0) −
1
a2 (0)2
2c2
(40)
and:
38
Substituting these results in the first equation of the system gives:
da1
1 2
c2 2
a1 + C1 )(
a + C2 )]
= c1 [α − βa1 (
dt
2c1
2c2 2
(41)
1 c2 2
c2 2
da1
a1 + C1 )(
(
a + C1 )2 + C2 )]
= c1 [α − βa1 (
dt
2c1
2c2 2c1 1
(42)
and hence:
In short da1 /dt = Q(a1 ) where Q is a polynomial of degree 7 in a1 . By
expanding and simplifying Equation 42, it is easy to see that the leading
term of Q is negative and given by βc22 /(16c21). Therefore, using Theorem 1,
for any initial conditions a1 (0), a1 (t) converges to a finite fixed point. Since
a2 is a quadratic function of a1 it also converges to a finite fixed point, and
similarly for a3 . Thus in the general case the system always converges to a
global minimum of the error function satisfying α−βP = 0. The hypersurface
a21 a22 + c1 a2 a3 + c2 a21 a3 = 0 depends on c1 , c2 and provides additional critical
points for the product P . It can be shown again by linearization that this
hypersurface separates stable from unstable fixed points.
As in the previous case, small weights and congruent weights can help
learning but are not necessary. In particular, if c1 and c2 are small, or if c1
is small and c2 is congruent (with a3 ), then a21 a22 + c1 a2 a3 + c2 a21 a3 > 0 and
dP/dt has the same sign as α − βP .
7.4
The General Linear Chain: A[1, . . . , 1].
Derivation of the System: The analysis can be extended immediately to
a linear chain architecture A[1, . . . , 1] of arbitrary length (Figure 12). In this
case, let a1 , a2 , . . . , aL denote the forward weights and c1 , . . . , cL−1 denote
the feedback weights. Using the same derivation as in the previous cases and
letting O = P I = a1 a2 . . . aL I gives the system:
∆ai = ηci (T − O)a1a2 . . . ai−1 I
(43)
for i = 1, . . . , L. Taking expectations as usual leads to the set of differential
39
Figure 12: Left: A[1, . . . , 1] architecture. The weights ai are adjustable, and
the feedback weight ci are fixed. The index of each parameter is associated
with the corresponding target layer.
equations:
da1
= c1 (α − βP )
dt
da2
dt = c2 a1 (α − βP )
...
daL−1
= cL−1 a1 a2 . . . aL−2 (α − βP )
dt
daL = a . . . a (α − βP )
dt
1
L−1
or, in more compact form:
40
(44)
k=i−1
Y
dai
= ci
ak (α − βP ) for i = 1, . . . , L
(45)
dt
k=1
Q
with cL = 1. As usual, P = Li=1 ai , α = E(T I), and β = E(I 2 ). A simple
calculation yields:
i−1
L
L
X
X
dP
P dai
ci Y
ak
=
= (α − βP )
P
dt
a dt
ai k=1
i=1 i
i=1
(46)
the last equality requiring ai 6= 0 for every i.
Theorem 4: Except for trivial cases, starting from any initial conditions the
system in Equation 44 converges to a fixed point, corresponding to a global
minimum of the quadratic error function. All the fixed points are located on
the hypersurface given by α − βP = 0 and are global minima of the error
function. Along any trajectory, and for each i, ai+1 is a quadratic function of
ai . For any starting point, the final fixed point can be calculated by solving
a polynomial equation of degree 2L − 1.
Proof: Again, when all the weights in the learning channel are non zero,
the critical points correspond to the curve α − βP = 0. These critical
points are independent of the weights in the learning channel and correspond to global minima of the error function. Additional
P critical
Q points for
the product P = a1 . . . aL are given by the surface Li=1 P acii i−1
k=1 ak = 0.
These critical points are dependent on the weights in the learning channel.
If the cQ
congruent with the respective feedforward weights, then
i are small orQ
P
L
j=k−1
[
a
][c
aj ] > 0 and dP/dt has the same sign as α − βP .
L−k
k=1
i6=k i
j=1
Thus small or congruent weights can help the learning but they are not necessary.
To see the convergence, from Equation 45, we have:
dai
dai+1
= ci+1 ai
(47)
dt
dt
Note that if one the derivatives dai /dt is zero, then they are all zero and thus
there cannot be any limit cycles. Since in the general case all the ci are non
zero, we have:
ci
41
ci+1 2
a +C
(48)
2ci i
showing that there is a quadratic relationship between ai+1 and ai , with no
linear term, for every i. Thus every ai can be expressed as a polynomial
function of a1 of degree 2i−1 , containing only even terms:
ai+1 =
i−1
ai = k0 + k1 a21 + . . . + ki−1 a21
(49)
and:
ki−1 =
ci
ci−1 2 ci−2 4
c3 2i−1
(
)(
) ...(
)
2ci−1 2ci−2 2ci−3
2c2
(50)
By substituting these relationships in the equation for the derivative of a1 ,
we get da1 /dt = Q(a1 ) where Q is a polynomial with an odd degree n given
by:
n = 1 + 2 + 4 + . . . + 2L−1 = 2L − 1
(51)
Furthermore, from Equation 50 it can be seen that leading coefficient is
negative therefore, using Theorem 1, for any set of initial conditions the
system must converge to a finite fixed point. For a given initial condition,
the point of convergence can be solved by looking at the nearby roots of the
polynomial Q of degree n.
Gradient Descent Equations: For comparison, the gradient descent equations are:
dai
P
∂E
= aL . . . ai+1 a1 . . . ai−1 (α − βP ) = (α − βP ) = −
dt
ai
∂ai
(52)
(the equality in the middle requires that ai 6= 0). In this case, the coupling
between neighboring terms is given by:
ai
dai+1
dai
= ai+1
dt
dt
(53)
Solving this equation yields:
da2
da2i
= i+1
dt
dt
or a2i+1 = a2i + C
42
(54)
7.5
Adding Width (Expansive): A[1, N, 1]
Derivation of the System: Consider a linear A[1, N, 1] architecture (Figure 13). For notational simplicity, we let a1 , . . . , aN be the weights in the
lower layer, b1 , . . . , bN be the weights in the upper layer, and c1 , . . . , cN
the
P random weights of the
P learning channel. In this case, we have O(t) =
i ai bi I(t). We let P =
i ai bi . The learning equations are:
(
P
∆ai = ηci (T − O)I = ηci (T − i ai bi I)I
(55)
P
∆bi = η(T − O)ai I = η(T − i ai bi I)ai I
When averaged over the training set:
(
E(∆ai ) = ηci E(IT ) − ηci P E(I 2) = ηci α − ηci P β
E(∆bi ) = ηai E(IT ) − ηai P E(I 2 ) = ηai α − ηai P β
(56)
where α = E(IT ) and β = E(I 2 ). With the proper scaling of the learning
rate (η = ∆t) this leads to the non-linear system of coupled differential
equations for the temporal evolution of ai and bi during learning:
(
dai
= αci − βci P = ci (α − βP )
dt
(57)
dbi
=
αa
−
βa
P
=
a
(α
−
βP
)
i
i
i
dt
P
The dynamic of P = i ai bi is given by:
X dbi
X
dai
dP
=
ai
+ bi
= (α − βP )
[bi ci + a2i ]
dt
dt
dt
i
i
(58)
Theorem 5: Except for trivial cases, starting from any initial conditions the
system in Equation 57 converges to a fixed point, corresponding to a global
minimum of the quadratic error function. All the fixed points are located
on the hyersurface given by α − βP = 0 and are global minima of the error
function. Along any trajectory, each bi is a quadratic polynomial function
of ai . Each ai is an affine function of any other aj .For any starting point,
the final fixed point can be calculated by solving a polynomial differential
equation of degree 3.
Proof: Many of the features found in the linear chain are found again in
this system using similar analyses. In the general case where the weights
43
in the learning channel are non zero, the critical points are given by the
surface α − βP = 0 and correspond to global optima. These critical points
are independent of the weights
critical
P in the learning channel. Additional
P 2
points for the product P = i ai bi are given by the surface i ai + bi ci = 0
which depends on the weights in the learning
If the ci ’s are small,
P channel.
2
or congruent with the respective bi ’s, then i ai + bi ci > 0 and dP/dt has
the same sign as α − βP .
To address the convergence, Equations 57 leads to the vertical coupling
between ai and bi :
ai
dai
dbi
= ci
dt
dt
or bi =
1 2
a + Ci
2ci i
(59)
for each i = 1, . . . , N. Thus the dynamics of the ai variables completely
determines the dynamics of the bi variables, and one only needs to understand
the behavior of the ai variables. In addition to the vertical coupling between
ai and bi , there is an horizontal coupling between the ai variables given again
by Equation 57 resulting in:
ci+1 dai
ci+1
dai+1
=
or ai+1 =
ai + Ki+1
(60)
dt
ci dt
ci
Thus, iterating, all the variables ai can be expressed as affine functions of a1
in the form:
ci
a1 + Ki′ i = 1, . . . , N
(61)
c1
Thus solving the entire system can be reduced to solving for a1 . The differential equation for a1 is of the form da1 /dt = Q(a1 ) where Q is a polynomial of
degree 3. Its leading term, is the leading term of −c1 βP . To find its leading
term we have:
ai =
P =
X
i
ai bi =
X a3
i
+ ci ai
2c
i
i
(62)
and thus the leading term of Q is given by Ka31 where:
N
1
1 c3N
β 1 X 2
1 c32
K = −βc1 [
+
.
.
.
]
=
−
[
c ]
+
2c1 2c2 c31
2cN c31
2 c21 1 i
(63)
Thus the leading term of Q has a negative coefficient, and therefore a1 always
converges to a finite fixed point, and so do all the other variables.
44
Figure 13: Left: Expansive A[1, N, 1] Architecture. Right: Compressive
A[N, 1N] Architecture. In both cases, the parameters ai and bi are adjustable, and the parameters ci are fixed.
7.6
Adding Width (Compressive): A[N, 1, N ]
Derivation of the System: Consider a linear A[N, 1, N] architecture (Figure 13). The on-line learning equations are given by:
(
P
∆ai = η N
k=1 ck (Tk − Ok )Ii
(64)
P
∆bi = η(Ti − Oi ) N
k=1 ak Ik
for i = 1, . . . , N. As usual taking expectations, using matrix notation and a
small learning rate, leads to the system of differential equations:
(
dA
= C(ΣT I − BAΣII )
dt
(65)
dB
t
=
(Σ
−
BAΣ
)A
T
I
II
dt
Here A is an 1 × N matrix, B is an N × 1 matrix, and C is an 1 × N
matrix, and M t denotes the transpose of the matrix M. ΣII = E(II t ) and
ΣT I = E(T I t ) are N × N matrices associated with the data.
Lemma 1: Along the flow of the system defined by Equation 65, the solution
satisfies:
1
CB = ||A||2 + K
2
45
(66)
where K is a constant depending only on the initial values.
Proof: The proof is immediate since:
C
dB
dA t
=
A
dt
dt
or
X dbi X dai
ci
=
ai
dt
dt
i
i
or
X dbi
1 d||A||2
ci
=
dt
2 dt
i
(67)
where ||A||2 = a21 + . . . + a2N . The theorem is obtained by integration.
Theorem 6: In the case of an autoencoder with uncorrelated normalized
data (Equation 68), the system converges to a fixed point satisfying A = βC,
where β is a positive root of a particular cubic equation. At the fixed point,
B = C t /(β||C||2) and the product P = BA converges to C t C/||C||2.
Proof: For an autoencoder with uncorrelated and normalized data (ΣT I =
ΣII = Id). In this case the system can be written as:
(
dA
= C(Id − BA)
dt
(68)
dB
= (Id − BA)At
dt
We define
1
(69)
σ(t) = kAk2 + K
2
and let A0 = A(0). Note that σ(t) ≥ K. We assume that C and A0 are
linearly independent, otherwise the proof is easier. Then we have:
dA
= C − σ(t)A
dt
Therefore the solution A(t) must have the form:
(70)
A(t) = f (t)C + g(t)A0
(71)
f ′ (t) = 1 − σ(t)f (t), f (0) = 0
g ′ (t) = −σ(t)g(t), g(0) = 1
(72)
which yields:
46
or:
g(t) = e−
−
f (t) = e
Rt
0
σ(s)ds
Rt
σ(s)ds
0
Z
t R
r
e
0
σ(s)ds
dr
(73)
0
From the above expressions, we know that both f and g are nonnegative.
We also have
Z t
1
f (t) = g(t)
dr
(74)
0 g(r)
Since σ(t) ≥ K, g(t) is bounded, and thus
Z ∞
1
dr = ∞.
g(r)
0
(75)
By a more general theorem shown in the next section, we know also that
||A|| is bounded and therefore f is also bounded. Using Equation 74, this
implies that that g(t) → 0 as t → ∞. Now we consider again the equation:
f ′ = 1 − σf
(76)
Now consider the cubic equation:
1
1 − ( t2 ||C||2 + K)t = 0
2
For t large enough, since g(t) → 0, we have:
1
σ(t) ≈ f 2 ||V ||2 + K
2
Thus Equation 76 is close to the polynomial differential equation:
(77)
(78)
1
(79)
h′ = 1 − ( h2 ||C||2 + K)h
2
By Theorem 1, this system is always convergent to a positive root of Equation
77, and by comparison the system in Equation 76 must converge as well. This
proves that f (t) → β as t → ∞, and in combination with g(t) → 0 as t → ∞,
shows that A converges to βC. As A converges to a fixed point, the error
function converges to a convex function and B performs gradient descent
on this convex function and thus must also approach a fixed point. By the
47
results in [2, 3], the solution must satisfy BAAt = At . When A = βC this
gives: B = C t /(βCC t ) = C t /(β||C||2). In this case, the product P = BA
converges to the fixed point: C t C/||C||2. The proof can easily be adapted to
the slightly more general case where ΣII is a diagonal matrix.
7.7
The General Linear Case: A[N0, N1, . . . , NL]
Derivation of the System: Although we cannot yet provide a solution
for this case, it is still useful to derive its equations. We assume a general
feedforward linear architecture (Figure 14) A[N0 , N1 , . . . , NL ] with adjustable
forward matrices A1 , . . . , AL and fixed feedback matrices C1 , . . . , CL−1 (and
CL = Id). Each matrix Ai is of size Ni × Ni−1 and,
QL in SRBP, each matrix Ci
is of size Ni × NL . As usual, O(t) = P I(t) = ( i=1 Ai )I(t).
Figure 14: General linear case with an architecture A[N0 , . . . , NL ]. Each
forward matrix Ai is adjustable and of size Ni ×Ni−1 . In SRBP, each feedback
matrices Ci is fixed and of size Ni × NL .
Assuming the same learning rate everywhere, using matrix notation we
have:
48
∆Ai = ηCi (T − O)(Ai−1 . . . A1 I)t = ηCi (T − O)I t At1 . . . Ati−1
(80)
which, after taking averages, leads to the system of differential equations
dAi
= Ci (ΣT I − P ΣII )At1 . . . Ati−1
(81)
dt
with P = AL AL−1 . . . A1 , ΣT I = E(T I t ), and ΣII = E(II t ). ΣT I is a NL ×N0
matrix and ΣII is a N0 × N0 matrix. In the case of an autoencoder, T = I
and therefore ΣT I = ΣII . Equation 81 is true also for i = 1 and i = L
with CL = Id where Id is the identity matrix. These equations establish a
coupling between the layers so that:
dAi+1
= Ci+1 (ΣT I − P ΣII )At1 . . . Ati
dt
When the layers have the same sizes, the coupling can be written as:
(82)
dAi t
dAi t
dAi+1
dAi+1
= Ci−1
Ai or
= Ci+1 Ci−1
A
(83)
dt
dt
dt
dt i
where we can assume that the random matrices Ci are invertible square
matrices.
−1
Ci+1
Gradient Descent Equations: For comparison, the gradient descent equations are given by:
dAi
= Ati+1 . . . AtL (ΣT I − P ΣII )At1 . . . Ati−1
dt
resulting in the coupling:
Ati+1
dAi t
Ai+1
=
A
dt
dt i
(84)
(85)
and, by definition:
∂E
dAi
=−
dt
∂Ai
where E = E(T − P I)2/2.
49
(86)
RBP Equations: Note that in the case of RBP with backward matrices
C1 , . . . , CL−1 , as opposed to SRBP, one has the system of differential equations:
dAi
= Ci . . . CL−1 (ΣT I − P ΣII )At1 . . . Ati−1
(87)
dt
By letting Bi = Ci . . . CL−1 one obtains the SRBP equations however the size
of the layers may impose contraints on the rank of the matrices Bi .
7.8
The General Three-Layer Linear Case A[N0 , N1, N2].
Derivation of the System: Here we let A1 be the N1 × N0 matrix of
weights in the lower layer, A2 be N2 × N1 matrix of weights in the upper
layer, and C1 the N1 × N2 random matrix of weights in the learning channel.
In this case, we have O(t) = A2 A1 I(t) = P I(t)) and ΣII = E(II t ) (N0 ×N0 ),
and ΣT I = E(T I t ) (N2 × N1 ). The learning equations are given by:
(
dA2
= (ΣT I − P ΣII )At1
dt
(88)
dA1
=
C
(Σ
−
P
Σ
)
1
T
I
II
dt
resulting in the coupling:
dA2
dA1 t
=
A
(89)
dt
dt 1
The corresponding gradient descent equations are obtained immediately by
replacing C1 with At2 .
Note that the two-layer linear case corresponds to the classical Least
Square Method which is well understood. The general theory of the threelayer linear case, however, is not well understood. In this section, we take
a significant step towards providing a complete treatment of this case. One
of the main results is that system defined by Equation 88 has long-term
existence, and C1 P = C1 A2 A1 is convergent and thus, in short, the system
is able to learn. However this alone does not imply that the matrix valued
functions A1 (t), A2 (t) are individually convergent. We can prove the latter in
special cases like A[N, 1, N] and A[1, N, 1] studied in the previous sections,
as well as A[2, 2, 2].
We begin with the following theorem.
C1
50
Theorem 7: The general three layer linear system (Equation 88) always has
long-term solutions. Moreover ||A1 || is bounded.
Proof: As in Lemma 1, we have:
d(CA2 )
dA1 t
= C(ΣT I − A2 A1 ΣII )At1 =
A
dt
dt 1
(90)
Thus we have:
dA1 t
dAt
d
d((CA2 ) + (CA2 )t )
=
A1 + A1 1 = (A1 At1 ).
dt
dt
dt
dt
It follows that:
(CA2 ) + (CA2 )t = A1 At1 + C0
(91)
(92)
where C0 is a constant matrix. Let:
f = Tr(A1 At1 ).
(93)
Using Lemma 2 below, we have:
df
dA1 t
= 2Tr(
A ) = 2Tr(CΣT I At1 −CA2 A1 ΣII At1 ) ≤ c3 kA1 k−2Tr(CA2 A1 ΣII At1 ).
dt
dt 1
(94)
Since:
2Tr(CA2 A1 ΣII At1 ) = Tr(CA2 A1 ΣII At1 ) + Tr(A1 ΣII At1 (CA2 )t )
(95)
2Tr(CA2 A1 ΣII At1 ) = Tr(CA2 A1 ΣII At1 ) + Tr((CA2 )t A1 ΣII At1 )
(96)
or:
using Equation 92, we have:
2Tr(CA2 A1 ΣII At1 ) = Tr(A1 At1 A1 ΣII At1 ) + Tr(C0 A1 ΣII At1 )
Using the second inequality in Lemma 2 below, we have:
51
(97)
p
df
1
≤ c3 ||A1 ||+Tr(C0 A1 ΣII At1 )−c1 f 2 ≤ c3 f +c4 f −c1 f 2 ≤ c5 − c1 f 2 (98)
dt
2
for positive constants c1 , . . . , c5 . Since A1 has long-term existence, so does
f . Note that it is not possible for f to be increasing as t → ∞ because if
we had f ′ (t) ≥ √
0, then we would have c5 − 21 c1 f 2 ≥ 0 and thus f would be
√
bounded (f ≤ 2c5 / c1 ). But if f√ is not always increasing, at √
each local
√
√
maximum point of f we have f ≤ 2c5 / c1 , which implies f ≤ 2c5 / c1
everywhere.
Lemma 2: There is a constant c1 > 0 such that
1. f ≥ c1 kA1 k2 ,
2. Tr(A1 At1 A1 ΣII At1 ) ≥ c1 f 2 .
Proof: The first statement is obvious. To prove the second one, we observe
that:
Tr(A1 At1 A1 ΣII At1 ) = Tr(At1 A1 ΣII At1 A1 ) ≥ c2 Tr(At1 A1 At1 A1 ) ≥ c1 f 2
(99)
for some constants c1 , c2 > 0.
To complete the proof of Theorem 7, we must estimate A2 to make sure it
does not diverge at a finite time. Let
1
h = Tr(A2 At2 )
2
(100)
Then:
dh
= Tr((ΣT I −A2 A1 ΣII )At1 At2 ) = Tr(ΣT I At1 At2 )−Tr(A2 A1 ΣII At1 At2 ) (101)
dt
and thus:
dh
≤ Tr(ΣT I At1 At2 )
dt
52
(102)
Since we have shown that ||A1 || is bounded:
√
dh
≤ Tr(ΣT I At1 At2 ) ≤ K||A2 || ≤ K h
dt
for some constant K. As a result, h ≤ K1 t2 + K2 or
p
||A2 || ≤ K1 t2 + K2 ≤ K3 t + K4
(103)
(104)
Since for every t, 1/||A2|| is bounded, the system has long-term solutions.
The main result of this section is as follows.
Theorem 8: [Partial Convergence Theorem] Along the flow of the system
in Equation 88, A1 and C1 A2 are uniformly bounded. Moreover, C1 A2 A1 →
C1 ΣT I Σ−1
II as t → ∞ and:
Z ∞
2
kC1 A2 A1 − C1 ΣT I Σ−1
(105)
II k dt < ∞
0
Proof. Let:
U = C1 (ΣT I − A2 A1 ΣII )Σ−1
II
Then:
dA1
= UΣII ,
dt
(106)
d(C1 A2 )
= UΣII AT1
dt
(107)
It follows that:
d(C1 A2 )
(C1 A2 )T = UΣII AT1 (C1 A2 )T = UΣII (C1 A2 A1 )T
dt
T
d(A1 (C1 ΣT I Σ−1
T
T
II ) )
= UΣII (C1 ΣT I Σ−1
II ) = U(C1 ΣT I )
dt
Thus we have:
and
T
d(A1 (C1 ΣT I Σ−1
d(C1 A2 )
T
II ) )
(C1 A2 ) −
= −UΣII U T ≤ 0
dt
dt
(108)
Here, for two matrices X and Y , we write X ≤ Y if and only if Y − X is a
semi-positive matrix. Let:
T
−1
T
V = (C1 A2 )(C1 A2 )T − A1 (C1 ΣT I Σ−1
II ) − (C1 ΣT I ΣII )A1
53
Then:
dV
≤0
dt
By Theorem 7, there is a lower bound on the matrix V
T
−1
T
V ≥ −A1 (C1 ΣT I Σ−1
II ) − (C1 ΣT I ΣII )A1 ≥ −C
for a constant matrix C. Thus as t → ∞, V = V (t) is convergent. Using the
inequality above, the expression
T
−1
T
(C1 A2 )(C1 A2 )T − A1 (C1 ΣT I Σ−1
II ) − (C1 ΣT I ΣII )A1
(109)
is monotonically decreasing. Since A1 is bounded by Theorem 7, and A2 AT2
is nonnegative, the expression is convergent. In particular, C1 A2 is also
bounded along the flow. By the (108), both A1 and C1 A2 are L2 integrable.
Thus in fact we have pointwise convergence of C1 A2 A1 . Since C1 may not be
full rank, we call it partial convergence. If C1 has full rank (which in general
is the case of interest), then as C1 A2 A1 is convergent, so is A2 A1 .
When does partial convergence imply the convergences of the solution
(A1 (t), A2 (t))? The following result gives a sufficient condition.
Theorem 9: If the set of matrices A1 , A2 satisfying:
C1 A2 A1 = C1 ΣT I Σ−1
II
(C1 A2 ) + (C1 A2 )T − A1 AT1 = K
(110)
C1 A2 AT2 C1T + A1 (C1 ΣT I )T + C1 ΣT I AT1 = L
is discrete, then A1 (t) and C1 A2 (t) are convergent.
Proof. By the proof of Theorem 8, we know that A1 (t), C1 A2 (t) are
bounded, and the limiting points of the pair (A1 (t), C1 A2 (t)) satisfy the relationships in Equation 110. If the set is discrete, then the limit must be be
unique and A1 (t) and C1 A2 (t) converge.
If C1 has full rank, then the system in Equation (88) is convergent, if the
assumptions in Theorem 9 are satisfied. Applying this result to the A[1, N, 1]
and A[N, 1, N] cases, provides alternative proofs for Theorem 3 and Theorem 6. The details are omitted. Beyond these two cases, the algebraic set
54
defined by Equation (110) is quite complicated to study. The first non-trivial
case that can be analyzed corresponds to the A[2, 2, 2] architecture. In this
special case, we we can solve the convergence problem entirely as follows.
For the sake of simplicity, we assume that ΣII = ΣT I = C1 = I. Then
the system associated with Equation (88) can be simplified to:
(
dB
= (I − BA)At
dt
(111)
dA
= (I − BA)
dt
where A(t), B(t) are 2 × 2 matrix functions. By Theorem 7, we know that
B(t)A(t) is convergent. In order to prove that B(t) and A(t) are individually
convergent, we prove the following result.
Theorem 10: Let F be the set of 2 × 2 matrices A, B satisfying the equations:
B + B T − AAT = K
A + AT − BB T = L
AB = I
(112)
where K, L are fixed matrices. Then F is a discrete set and the system
defined by Equation 111 is convergent.
Proof: The proof is somewhat long and technical and thus is given in the
Appendix. It uses basic tools from algebraic geometry.
Theorem 10 provides evidence that in general the algebraic set defined
by Equation (110) might be discrete. Although at this moment we are not
able to prove discreteness in the general case, this is a question of separate
interest in mathematics (real algebraic geometry). The system defined by
Equation (110) is an over-determined system of algebraic equations. For
example, if A(t), B(t) are n × n matrices, and if C is non-singular, then the
system contains n(n + 1) equations with n2 unknowns. One can define the
Koszul complex [9] associated with these equations Using the complex, given
specific matrices C, ΣT I , ΣII , K, L, there is a constructive algorithmic way to
determine whether the set is discrete. If it is, then the corresponding system
of ODE is convergent. 1
1
We thank Professor Vladimir Baranovsky for providing this information.
55
7.9
A Non-Linear Case
As can be expected, the case of non-linear networks is challenging to analyze
mathematically. In the linear case, the transfer functions are the identity
and thus all the derivatives of the transfer functions are equal to 1 and thus
play no role. The simulations reported above provide evidence that in the
non-linear case the derivatives of the activation functions play a role in both
RBP and SRBP. Here we study a very simple non-linear case which provides
some further evidence.
We consider a simple A[1, 1, 1] architecture, with a single power function
non linearity with power µ 6= 1 in the hidden layer, so that O 1 (S) = (S 1 )µ.
The final output neuron is linear O 2 (S 2 ) = S 2 and thus the overall inputoutput relationship is: O = a2 (a1 I)µ . Setting µ to 1/3, for instance, provides
an S-shaped transfer function for the hidden layer, and setting µ = 1 corresponds to the linear case analyzed in a previous section. The weights are a1
and a2 in the forward network, and c1 in the learning channel.
Derivation of the System Without Derivatives:
When no derivatives are included, one obtains:
(
da2
= aµ1 [E(T I µ ) − a2 aµ1 E(µI 2 )] = aµ1 (α − βa2 aµ1 )
dt
da1
= c1 [E(T I) − a2 aµ1 E(I µ+1 )] = c1 (γ − δa2 aµ1 )
dt
(113)
where here α = E(T I µ ), β = E(I 2µ ), γ = E(T I), and δ = E(I µ+1 ). Except
for trivial cases, such a system cannot have fixed points since in general one
cannot have a2 aµ1 = α/β and a2 aµ1 = γ/δ at the same time.
Derivation of the System With Derivatives:
In contrast, when the derivative of the forward activation is included the
system becomes:
(
da2
dt
da1
dt
= aµ1 [E(T I µ ) − a2 aµ1 E(I 2µ )] = aµ1 (α − βa2 aµ1 )
= c1 µaµ−1
E(T I µ ) − a2 c1 µa2µ−1
E(I 2µ ) = aµ−1
c1 µ(α − βa2 aµ1 )
1
1
1
(114)
This leads to the coupling:
a1
da1
da2
= c1 µ
dt
dt
or a2 =
56
a21
+K
2c1 µ
(115)
excluding as usual the trivial cases where c1 = 0 or µ = 0. Here K is a
constant depending only on a1 (0) and a2 (0). The coupling shows that if
da1 /dt = 0 then da2 /dt = 0 and therefore in general limit cycles are not
possible. The critical points are given by the equation:
α − βa2 aµ1 = 0 or a2 =
α
βaµ1
(116)
and do not depend on the weight in the learning channel. Thus, in the nontrivial cases, a2 is an hyperbolic function of aµ1 . It is easy to see, at least in
some cases, that the system converges to a fixed point. For instance, when
α > 0, c1 > 0, µ > 1, and a1 (0) and a2 (0) are small and positive, then
da1 /dt > 0 and da2 /dt > 0 and both derivatives are monotonically increasing and α − βa2 aµ1 decreases monotonically until convergence to a critical
point.Thus in general the system including the derivatives of the forward
activations is simpler and better behaved. In fact, we have a more general
theorem.
Theorem 12: Assume that α > 0, β > 0 c1 > 0, and µ ≥ 1. Then for
any positive initial values a1 (0) ≥ 0 and a2 (0) ≥ 0, the system described by
Equation 114 is convergent to one of the positive roots of the equation for t:
α − β(
t2
+ K)tµ = 0
2c1 µ
(117)
Proof: Using Equation 115, the differential equation for a1 can be rewritten
as:
da1
a21
= µaµ−1
c
(α
−
β(
+ K)aµ1 ) = Q(a1 )
1
1
dt
2c1 µ
(118)
When µ is an integer, Q(a1 ) is a polynomial of odd degree with a leading
coefficient that is negative and therefore, using Theorem 1, the system is
convergent. If µ is not an integer, let r1 < . . . < rk be the positive roots of
the function Q. The proof then proceeds similarly to the proof of Theorem
1. That is this differential equation (Equation 118) is convergent to one of
the (non-negative) roots of Q(t). However, since a1 (0) > 0, a more careful
analysis shows that it is not for a1 to converge to zero. Thus a1 must converge
to a positive root of Equation 117.
57
Gradient Descent Equations: Finally, for comparison, in the case of
gradient descent, the system is given by:
(
da2
dt
da1
dt
= aµ1 [E(T I µ ) − a2 aµ1 E(I 2µ )] = aµ1 (α − βa2 aµ1 )
= a2 µaµ−1
E(T I µ ) − a22 µa2µ−1
E(I 2µ ) = aµ−1
a2 µ(α − βa2 aµ1 )
1
1
1
(119)
Except for trivial cases, the critical points are again given by Equation 116,
and the system always converges to a critical point.
8
Conclusion
Training deep architectures with backpropagation on digital computers is
useful for practical applications, and it has become easier than ever, in part
because of the creation of software packages with automatic differentiation
capabilities. This convenience, however, can be misleading as it hampers
thinking about the constraints of learning in physical neural systems, which
are merely being mimicked on digital computers. Thinking about learning
in physical systems is useful in many ways: it leads to the notion of local learning rules, which in turn identifies two fundamental problems facing
backpropagation in physical systems. First backpropagation is not local, and
thus a learning channel is required to communicate error information from
the output layer to the deep weights. Second, backpropagation requires symmetric weights, a significant challenge for those physical systems that cannot
use the forward channel in the reverse direction, thus requiring a different
pathway to communicate errors to the deep weights.
RBP is one mode for communicating information over the learning channel, that completely bypasses the need for symmetric weights, by using fixed
random weights instead. However RBP is only one possibility among many
other ones for harnessing randomness in the learning channel. Here we have
derived several variants of RBP and studied them through simulations and
mathematical analyses. Additional variants are studied in a followup paper [4] which considers additional symmetry issues such as having a learning
channel with an architecture that is not a symmetric version of the forward
architecture, or having non-linear units in the learning channel that are similar to the non-linear units of the forward architecture.
In combination, the main emerging picture is that the general concept of
RBP is remarkably robust as most of the variants lead to robust learning.
58
RBP and its many variants do not seem to have a practical role in digital
simulations as they often lead to slower learning, but they should be useful
in the future both to better understand biological neural systems, and to
implement new neural physical systems in silicon or other substrates.
Information
Algorithm
l
Iijh = Iijh (T, O, wrs
(l > h), f ′ (l ≥ h))
General Form
l
Iijh = Iih (T, O, wrs
(l > h), f ′ (l ≥ h))
BP (symmetric weights)
l
Iijh = Iih (T − O, wrs
(l > h), f ′ (l ≥ h))
BP (symmetric weights)
h+1
l
Iijh = Iih (T − O, wrs
(l > h + 1), wki
, f ′ (l ≥ h))
l
h
Iijh = Iih (T − O, rrs
(l ≥ h + 1), rki
, f ′ (l ≥ h))
BP (symmetric weights)
RBP (random weights)
h
Iijh = Iih (T − O, rki
, f ′ (l ≥ h))
SRBP (random skipped weights)
h
Iijh = Iih (T − O, rki
, f ′ (l = h))
SRBP (random skipped weights)
Iijh = Iih (F (T − O), f ′(l = h))
F sparse/low-prec./adaptive/non-lin.
Table 2: Postsynaptic information required by deep synapses for optimal
learning. Iijh represents the signal carried by the deep learning channel and
the postsynaptic term in the learning rules considered here. Different algorithms reveal the essential ingredients of this signal and how it can be
simplified. In the last row, the function F can be implemented with sparse
or adaptive matrices, carry low precision signals, or include non-linear transformations in the learning channel (see also [4]).
In supervised learning, the critical equations show that in principle any
deep weights must depend on all the training examples and all the other
weights of the network. Backpropagation shows that it is possible to derive
effective learning rules of the form ∆wijh = ηIijh Ojh−1 where the role of the
lower part of the network is subsumed by the presynaptic activity term Ojh−1
and Iijh is a signal communicated through the deep learning channel that
carries information about the outputs and the targets to the deep synapses.
Here we have studied what kind of information must be carried by the signal
Iijh and how much it can be simplified (Table 2). The main conclusion is that
59
the postynaptic terms must: (1) implement gradient descent for the top layer
(i.e. random weights in the learning channel for the top layer do not work
at all); (2) for any other deep layer h it should be of the form f ′ F (T − O),
wheref ′ represents the derivatives of the activations of the units in layer h
(the derivatives above are not necessary) and F is some reasonable function
of the error T −O. By reasonable, we mean that the function F can be linear,
or a composition of linear propagation with non-linear activation functions,
it can be fixed or slowly varying, and when matrices are involved these can
be random, sparse, etc. As can be expected, it is better if these matrices are
full rank although gracious degradation, as opposed to catastrophic failure,
is observed when these matrices deviate slightly from the full rank case.
The robustness and other properties of these algorithms cry for explanations and more general principles. We have provided both intuitive and
formal explanations for several of these properties. On the mathematical side,
polynomial learning rules in linear networks lead to systems of polynomial
differential equations. We have shown in several cases that the corresponding
ODEs converge to an optimal solution. However these polynomial systems
of ODEs rapidly become complex and, while the results provided are useful,
they are not yet complete, thus providing directions for future research.
Acknowledgement
Work supported in part by NSF grant IIS-1550705 and a Google Faculty
Research Award to PB, and NSF grant DMS-1547878 to ZL. We are also
grateful for a hardware donation from NVDIA Corporation.
Appendix: Proof of Theorem 10
Assume that (A, B) ∈ F . If near (A, B), F is not discrete, then there are
real analytic matrix-valued functions (A(t), B(t)) ∈ F for small t > 0 such
that (A(0), B(0)) = (A, B). Moreover, if we write:
A(t) = A + tE +
t2
t3
F + G + o(t3 )
2
6
(120)
then E 6= 0. We use A′ , A′′ , A′′′ , B ′ , B ′′ , B ′′′ to denote A′ (0), A′′ (0), A′′′ (0), B ′(0),
B ′′ (0), B ′′′ (0), respectively. The general strategy is to prove that E = 0 or,
in the case E 6= 0, to take higher order derivatives to reach a contradiction.
60
It is easy to compute:
B ′ = −A−1 A′ A−1 = −BEB
(121)
By taking the derivative of the first two relations in Equation (112), we have:
− BEB − (BEB)T − EAT − AE T = 0
(122)
X = EAT + BEB
(123)
E + E T + BEBB T + B(BEB)T = 0
Let:
Y = E + BEBB T
Then by the above equations, both X, Y are skew symmetric, and we have
Y AT = X. If Y 6= 0, using an orthogonal transformation and scaling, we
may assume that:
0 −1
(124)
Y =
1 0
Write:
Then:
a b
A=
c d
(125)
−b −d
YA =
a
c
(126)
T
Since X skew-symmetric also, we must have b = c = 0, and a = d. Thus
A = aI for a real number a 6= 0. As a result, we have:
2
1
K = ( − a2 )I L = (2a − 2 )I
(127)
a
a
and (A, B) = (aI, a−1 I). Let (Ã(t), B̃(t)) be the upper triangular matrices
obtained by orthogonal transformation from (A(t), B(t)). Since both K, L
are proportional to the identity, (Ã(t), B̃(t)) ∈ F . Now let us write:
ã b̃
(128)
Ã(t) =
0 d˜
Then the equation B + B T − AAT = K is equivalent to the following system:
61
−1
2ã − (ã2 + b̃2 ) = 2a−1 − a2
2d˜−1 − d˜2 = 2a−1 − a2
−b̃ã−1 d˜−1 − b̃d˜ = 0
(129)
Since t is small, Ã(t) should be sufficiently close to aI. From the second
equation of the system above, we have d˜ = a. If b̃ = 0, then we conclude
from the first equation of the same system that ã = a, and hence Ã(t) = aI.
This implies that (A(t), B(t)) = (A, B). So in this case E = 0.
Things are more complicated when b̃ 6= 0. We first assume that a 6=
−1. In this case, from the third equation of the system above, we have
ã−1 d˜−1 + d˜ = 0. Since we already have d˜ = a 6= 1, for sufficiently small t,
ã = −d˜−2 = −a−2 , which is distinct from a. Thus in this case b̃ must be zero.
If a = −1, then we have d˜ = −1 and ã = −1. Using the first equation of
the system above, we have b̃ = 0 and the again (A(t), B(t)) = (Ã(t), B̃(t)) =
(A, B), and we conclude that E = 0.
From the results above, we know that if Y 6= 0 or if A is proportional to
the identity, near (A, B) ∈ F , there are no other elements in F and thus F
is discrete. When X = Y = 0, it is possible to have E 6= 0. However, we
have the following Lemma:
Lemma 3: If X = Y = 0, and if A 6= −I, then E is not an invertible matrix.
Proof. By contradiction, assume that E is invertible. Then from X = 0, we
have:
− A = EBB T E −1
(130)
By taking determinant on both sides, we get:
det A = det(−A) = det(BB T )
(131)
det A = 1
(132)
Thus we have:
Since A is similar to a negative definite matrix −BB T , the eigenvalues λ1 , λ2
of A are all negative. Since λ1 λ2 = det A = 1, we have:
62
− λ1 − λ2 ≥ 2
(133)
− a − d = −Tr A = Tr (BB T ) = a2 + b2 + c2 + d2
(134)
Using the same matrix representation as in Equation (125), we have:
However:
a2 + b2 + c2 + d2 = (a + d)2 + (b − c)2 − 2 ≥ (a + d)2 − 2 ≥ −a − d, (135)
and the equality is true if and only if b = c and a+d = −2. Since −λ1 −λ2 = 2
and λ1 λ2 = 1, the eigenvalues of A must be −1, −1, which implies b = c = 0.
Thus A = −I which is impossible by our assumption.
Next we consider the remaining case: X = Y = 0, and E is not invertible
(but not equal to zero), and A is not proportional to the identity. In this
case, we have to take up to third order derivatives to reach the conclusion.
By taking derivatives of the first two relations in Equation (112), we get:
P + P T = Q + QT = R + RT = S + S T = 0
(136)
where:
P = −B ′′ + A′′ AT + A′ (A′ )T
Q = A′′ − B ′′ B T − B ′ (B ′ )T
R = −B ′′′ + A′′′ AT + 3A′′ (A′ )T
(137)
S = A′′′ − B ′′′ B T − 3B ′′ (B ′ )T
Similar to the relations between the matrices X, Y , we have:
QAT − P = −B ′ (B ′ )T AT − A′ (A)T
(138)
Since AB = I, we have:
Thus:
(B ′ )T AT = −B T (A′ )T
QAT − P = B ′ B T AT − EAT = −BEBB T AT − EAT = 0
63
(139)
(140)
because X = 0. Since A is not proportional to the identity, then we must
have P = Q = 0 as in the case for X and Y .
The relationship between R, S is more complicated, but can be computed
using the same idea. We first have:
SAT − R = −3B ′′ (B ′ )T AT − 3A′′ (A′ )T
(141)
Using Equation (139) and the fact that P = 0, we have:
SAT − R = −3A′ (A′ )T B T (A′ )T = 3EE T B T E T
(142)
Since E is not invertible and we assume that E 6= 0, we must have:
E = ξη T
(143)
for some column vectors ξ, η. From the fact that Y = 0, we conclude that:
ξη T + Bξη T BB T = 0
(144)
and:
Bξ = −
Thus we compute:
kηk2
ξ
kB T ηk2
(145)
kηk4 · hξ, ηi T
ξξ
kB T ηk2
(146)
kηk4 · hξ, ηi T
ξξ
kB T ηk2
(147)
kηk4 · hξ, ηi −1 T
S ξξ
kB T ηk2
(148)
EE T B T E T = −
and:
SAT − R = −3
If hξ, ηi =
6 0, then S 6= 0. Thus:
AT = S −1 R − 3
For the matrix S −1 ξξ T , both the trace and determinant are zero. So the
eigenvalues are zero. On the other hand, since both S, R are skew-symmetric
matrices, S −1 R is proportional to the identity. As a result, the matrix AT ,
hence A, has two identical eigenvalues. Let λ be an eigenvalue of A, then:
64
Tr (A) = 2λ,
det(A) = λ2
(149)
Taking the trace in the first two relations of Equation(112), we get:
4λ−1 − kAk2 = Tr (K)
4λ − λ−2 kAk2 = Tr (L)
(150)
Thus for fixed K, L, λ and kAk can only assume discrete values. Since t is
small, A(t) = Q(t)AQ(t)T for some orthogonal matrix Q(t). Let us write:
0 −1
λ b
′
(151)
, Q (0) =
A=
−1 0
0 λ
Then E = A′ (0) is equal to:
b 0
E=
0 −b
(152)
By Lemma 3, E is not invertible. Thus b = 0. But if b = 0, then A is
proportional to the identity and this case has been discussed above.
We must still deal with the case hξ, ηi = 0. Without loss of generality,
we may assume that:
0
1
(153)
, η=
ξ=
1
0
By checking the equation AE = −EBB T ,
−2
−d
A=
0
we can conclude that:
0
.
d
(154)
In fact, when t is small, the eigenvalues of A(t) must be −d−2 and d for some
d 6= 0. Again, by taking the trace of the first two relations in Equation (112),
we get:
− 2d2 + 2d−1 − kAk2 = Tr (K);
− 2d−2 + 2d − d2 kAk2 = Tr (L).
(155)
Therefore, d is locally uniquely determined by K, L. Finally, if we write
A(t) = Q(t)AQ(t)T and assume that:
65
0 −1
Q (0) =
,
1 0
(156)
0
d + d−2
.
E=
d + d−2
0
(157)
′
we have:
Since E must be singular, we have d = −1 and hence A = −I. This case has
been covered above and thus the proof of Theorem 10 is complete.
References
[1] F. Agostinelli, N. Ceglia, B. Shahbaba, P. Sassone-Corsi, and P. Baldi.
What time is it? deep learning approaches for circadian rhythms. Bioinformatics, 32(12):i8–i17, 2016.
[2] P. Baldi and K. Hornik. Neural networks and principal component analysis: Learning from examples without local minima. Neural Networks,
2(1):53–58, 1988.
[3] P. Baldi and Z. Lu. Complex-valued autoencoders. Neural Networks,
33:136–147, 2012.
[4] P. Baldi, Z. Lu, and P. Sadowski. Learning in the machine: the symmetries of the deep learning channel. Neural Networks, 95:110–133, 2017.
[5] P. Baldi and P. Sadowski. The dropout learning algorithm. Artificial
Intelligence, 210C:78–122, 2014.
[6] P. Baldi and P. Sadowski. A theory of local learning, the learning channel, and the optimality of backpropagation. Neural Networks, 83:61–74,
2016.
[7] P. Baldi, P. Sadowski, and D. Whiteson. Searching for exotic particles
in high-energy physics with deep learning. Nature Communications, 5,
2014.
[8] P. Di Lena, K. Nagata, and P. Baldi. Deep architectures for protein contact map prediction. Bioinformatics, 28:2449–2457, 2012. doi:
10.1093/bioinformatics/bts475. First published online: July 30, 2012.
66
[9] D. Eisenbud. Commutative algebra, volume 150 of Graduate Texts in
Mathematics. Springer-Verlag, New York, 1995. With a view toward
algebraic geometry.
[10] K. Fukushima. Neocognitron: A self-organizing neural network model
for a mechanism of pattern recognition unaffected by shift in position.
Biological cybernetics, 36(4):193–202, 1980.
[11] X. Glorot and Y. Bengio. Understanding the difficulty of training deep
feedforward neural networks. In In Proceedings of the International Conference on Artificial Intelligence and Statistics (AISTATS10). Society
for Artificial Intelligence and Statistics, 2010.
[12] A. Graves, A.-r. Mohamed, and G. Hinton. Speech recognition with deep
recurrent neural networks. In Acoustics, Speech and Signal Processing
(ICASSP), 2013 IEEE International Conference on, pages 6645–6649.
IEEE, 2013.
[13] S. Han, H. Mao, and W. J. Dally. Deep compression: Compressing deep
neural network with pruning, trained quantization and huffman coding.
CoRR, abs/1510.00149, 2015.
[14] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image
recognition. arXiv preprint arXiv:1512.03385, 2015.
[15] D. Hebb. The organization of behavior: A neurophychological study.
Wiley Interscience, New York, 1949.
[16] G. E. Hinton, N. Srivastava, A. Krizhevsky, I. Sutskever, and R. R.
Salakhutdinov. Improving neural networks by preventing co-adaptation
of feature detectors. arXiv:1207.0580, July 2012.
[17] I. Hubara, M. Courbariaux, D. Soudry, R. El-Yaniv, and Y. Bengio.
Quantized neural networks: Training neural networks with low precision
weights and activations. CoRR, abs/1609.07061, 2016.
[18] D. H. Hubel and T. N. Wiesel. Receptive fields, binocular interaction
and functional architecture in the cat’s visual cortex. The Journal of
physiology, 160(1):106, 1962.
67
[19] Y. Ilyashenko. Centennial history of hilberts 16th problem. Bulletin of
the American Mathematical Society, 39(3):301–354, 2002.
[20] A. Krizhevsky and G. Hinton. Learning multiple layers of features from
tiny images. 2009.
[21] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification
with deep convolutional neural networks. In Advances in neural information processing systems, pages 1097–1105, 2012.
[22] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based
learning applied to document recognition. Proceedings of the IEEE,
86(11):2278–2324, 1998.
[23] Q. Liao, J. Leibo, and T. Poggio. How important is weight symmetry
in backpropagation? In Proceedings of the Thirtieth AAAI Conference
on Artificial Intelligence, pages 1837–1844, 2016.
[24] T. P. Lillicrap, D. Cownden, D. B. Tweed, and C. J. Akerman. Random feedback weights support learning in deep neural networks. arXiv
preprint arXiv:1411.0247, 2014.
[25] M. Riedmiller and H. Braun. A direct adaptive method for faster backpropagation learning: the rprop algorithm. In IEEE International Conference on Neural Networks, pages 586–591 vol.1, 1993.
[26] P. Sadowski, J. Collado, D. Whiteson, and P. Baldi. Deep learning, dark
knowledge, and dark matter. Journal of Machine Learning Research,
Workshop and Conference Proceedings, 42:81–97, 2015.
[27] C. E. Shannon. A mathematical theory of communication (part III).
Bell System Technical Journal, XXVII:623–656, 1948.
[28] C. E. Shannon. A mathematical theory of communication (parts I and
II). Bell System Technical Journal, XXVII:379–423, 1948.
[29] S. Smale. Mathematical problems for the next century. The Mathematical Intelligencer, 20(2):7–15, 1998.
[30] R. K. Srivastava, K. Greff, and J. Schmidhuber. Training very deep
networks. In Advances in Neural Information Processing Systems, pages
2368–2376, 2015.
68
[31] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich. Going deeper with convolutions.
In Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition, pages 1–9, 2015.
[32] J. Zhou and O. G. Troyanskaya. Predicting effects of noncoding variants
with deep learning-based sequence model. Nature methods, 12(10):931–
934, 2015.
69
| 2cs.AI |
1
Phase Retrieval Without
Small-Ball Probability Assumptions
arXiv:1604.07281v3 [cs.IT] 11 Sep 2017
Felix Krahmer
Abstract—In the context of the phase retrieval problem, it is
known that certain natural classes of measurements, such as
Fourier measurements and random Bernoulli measurements, do
not lead to the unique reconstruction of all possible signals,
even in combination with certain practically feasible random
masks. To avoid this difficulty, the analysis is often restricted
to measurement ensembles (or masks) that satisfy a small-ball
probability condition, in order to ensure that the reconstruction
is unique.
This paper shows a complementary result: for random
Bernoulli measurements, there is still a large class of signals
that can be reconstructed uniquely, namely those signals that
are non-peaky. In fact, this result is much more general: it holds
for random measurements sampled from any subgaussian distribution D, without any small-ball conditions. This is demonstrated
in two ways: first, a proof of stability and uniqueness, and second,
a uniform recovery guarantee for the PhaseLift algorithm. In all
of these cases, the number of measurements m approaches the
information-theoretic lower bound.
Finally, for random Bernoulli measurements with erasures, it
is shown that PhaseLift achieves uniform recovery of all signals
(including peaky ones).
Index Terms—Phase retrieval, random measurements,
PhaseLift, inverse problems, reconstruction algorithms,
sampling methods, convex optimization
I. I NTRODUCTION
A. Ambiguities in Phase Retrieval
Phase retrieval is the problem of recovering an unknown
vector x ∈ Cn from measurements of the form
2
yi = |aTi x| + wi
(for i = 1, . . . , m),
(1)
where the vectors ai ∈ Cn are known, and the wi ∈ R
represent additive noise which is unknown. (The name phase
retrieval refers to the fact that the measurements reveal the
magnitudes, but not the phases, of the aTi x.) Phase retrieval
has numerous applications including X-ray crystallography
[1], [2], astronomy [3], ptychography and coherent diffractive
imaging [4], [5], and quantum state tomography [6].
A typical experiment setup is that one places a detector
far from the object being imaged, and the detector records
F. Krahmer is with Research Unit M15, Department of Mathematics,
Technische Universität München (e-mail: felix.krahmer@tum.de).
Y.-K. Liu is with the National Institute of Standards and Technology,
Gaithersburg, MD, USA, and also with the Joint Center for Quantum
Information and Computer Science, University of Maryland (e-mail: yikai.liu@nist.gov).
This paper was presented in part at the 11th International Conference on
Sampling Theory and Applications (SampTA 2015), Washington, DC, May
25–29, 2015.
Copyright 2017 IEEE. Personal use of this material is permitted. However,
permission to use this material for any other purposes must be obtained from
the IEEE by sending a request to pubs-permissions@ieee.org.
Yi-Kai Liu
the intensity of the light field, but not its phase. To a first
approximation, this situation can be described by phaseless
measurements where the ai are Fourier basis vectors. It is
well known that such phaseless measurements can give rise
to ambiguities, in the sense that the solution is not uniquely
determined. These ambiguities may include spatial shifts and
conjugate inversion. Especially in the one dimensional case,
larger classes of ambiguities can arise [7].
One way to ensure solution uniqueness is to consider a setup
with random illuminations, which mathematically corresponds
to multiplying each of the image pixels with a randomly chosen factor [8]. As argued in [8], a feasible setup is to consider
phase modulations, where each of these factors lies on the
unit circle in the complex plane. Solution uniqueness is then
shown in the two-dimensional case, under the assumption that
the image has support of rank 2. These additional assumptions
are necessary to exclude simple counterexamples. Namely, any
configuration of phase modulations yields measurements of the
form ai = (ai1 , ai2 , . . . , ain ) where
|ai1 | = |ai2 | = · · · = |ain | = 1.
(2)
Using such measurements, it is always impossible to distinguish between the vectors x = (1, 0, 0, . . . , 0) and x̃ =
2
2
(0, 1, 0, . . . , 0), since |aTi x| = 1 = |aTi x̃| .
Motivated by the uniqueness results of [8], a number of
follow-up works studied tractable reconstruction algorithms
for phase retrieval with random diffraction patterns (cf. Section I-B below). In contrast to [8], however, these works do
not restrict the signal class, but rather avoid the ambiguity
problems by considering random masks that also vary in
amplitude, which arguably are more difficult to realize in
experiments.
The main goal of this paper is to combine these approaches,
studying recoverability under mild assumptions on the class
of signals, without requiring measurement vectors of varying
amplitude. We focus on the simplest class of measurements
with property (2), namely random Bernoulli measurements,
where the ai are sampled independently and uniformly at
n
random from {1, −1} . We see our results as a proof of
concept that measurements of constant amplitude are tractable
and expect that this will lay the foundation to study the case
of phase modulated Fourier measurement, but we will leave
this case for future work.
For random Bernoulli measurements, we show that a surprisingly large class of vectors x can be recovered: one can
recover all vectors x ∈ Rn that are µ-flat, in the sense that
they satisfy
kxk∞ ≤ µkxk2 ,
(3)
2
for some constant µ ∈ (0, 1) that is independent of the
dimension n. Intuitively, this condition says that x is not too
peaky, in the sense that at most a constant fraction of its mass
is concentrated on any one coordinate.
Our result for random Bernoulli measurements is a special
case of a more general result that applies whenever the
measurement vectors ai are sampled independently at random
from some subgaussian distribution D. In the following sections, we will describe this more general setting, and state our
results in detail.
B. Gaussian and Subgaussian Measurements
The scenario in which the measurement vectors ai are
chosen at random according to certain distributions has been
investigated intensively over the last few years, with the
Gaussian distribution being the paradigmatic example. Two
main viewpoints have been taken, focusing either on stable
uniqueness, or on recovery via computationally tractable algorithms. The first viewpoint asks when x is uniquely determined
from the measurements (1) (up to sign ambiguity and a
small reconstruction error resulting from the noise) [17]. Such
stability results are known in rather general settings, where x
is promised to lie in some known set T ⊂ Rn (for instance, the
set of k-sparse vectors), and one wants to bound the number of
measurements m as a function of some complexity parameter
of the set T .
(Note, however, that these stability results [17] were shown
in the real case, where the signal x and the measurements ai
are in Rn , rather than Cn . For simplicity, in this paper we will
likewise focus on the real case.)
The second viewpoint aims at finding tractable algorithms
with provable recovery guarantees. A well-known example
is PhaseLift, which reduces the problem to one of low-rank
matrix recovery, and then solves a convex relaxation [9], [10],
[11]. In particular, these works show that PhaseLift can recover
any vector x ∈ Cn using m = O(n) noisy measurements.
The initial works from both of these viewpoints were specific to Gaussian random measurement vectors ai . Subsequent
work (mainly from the recovery viewpoint) succeeded in
partially derandomizing these results, using techniques such
as spherical designs [12], [6] and coded diffraction patterns
[13], [14]. In addition, stable uniqueness results have been
shown for ai chosen from subgaussian distributions, subject to
additional assumptions on their small ball probabilities or their
fourth moments [17]. Nonetheless, even these assumptions on
the distribution of the ai ’s are somewhat restrictive.
It is natural to ask whether these stability analyses and recovery guarantees can be extended to the most straightforward
generalization of the Gaussian measurement setup, namely
subgaussian measurement vectors ai sampled from a product
distribution (i.e., each entry aij is sampled independently
from a subgaussian distribution D on R). However, this opens
the door to measurements such as random Bernoulli vectors
n
ai ∈ {1, −1} , where phase retrieval is not always possible.
Previous work on subgaussian phase retrieval has therefore
imposed certain restrictions on the distribution of the ai , such
as small-ball and fourth moment assumptions [17], which
exclude these pathological cases.
C. Our Results
In this paper we show that a large class of vectors x can
be recovered uniquely from subgaussian measurements ai ,
without imposing any additional conditions on the distribution
of the ai . In particular, we show that one can recover all
vectors x ∈ Rn that are µ-flat in the sense of (3), where
µ ∈ (0, 1) is a constant that depends on D, but not on the
dimension n.
In particular, our results apply to Bernoulli measurements
ai . Also, note that the µ-flatness requirement does not rule
out all sparse vectors. For instance, a vector that has support
of size 1/µ2 (i.e., constant size), and that does not have any
unusually large entries, will still satisfy equation (3), and hence
will still be recoverable.
To some extent, our results are analogous to a recent result
on one-bit compressed sensing [18]. Subgaussian measurements also fail in that context, and this issue can be overcome
by restricting to the case of signals which are not too peaky.
However, the techniques used there are somewhat different.
Our results can also be compared with recent work on phase
retrieval using local correlation measurements, which also
imposed a “flatness” condition on the signal x [19]. However,
the flatness condition in that paper is more elaborate than ours,
as it depends not only on the magnitudes of the entries in the
vector x = (x1 , . . . , xn ), but also on their ordering.1
We prove three main results. First, we consider stable
uniqueness, as in [17].2 We show that the results of [17], on
phase retrieval of signals x belonging to some set T ⊂ Rn ,
can be generalized to the setting of subgaussian measurements
ai (with independent entries aij ), provided that all signals
in the set T are µ-flat (for some constant µ). We emphasize
that the ai need not satisfy any small-ball or fourth-moment
assumptions.
In particular, we show that the number of measurements m
scales with the complexity of the set T in the same way as in
[17]. For instance, where [17] showed results on phase retrieval
of k-sparse vectors in Rn , we obtain comparable results on
phase retrieval of µ-flat k-sparse vectors in Rn . (Note that a
vector can be both k-sparse and µ-flat, as long as k ≥ 1/µ2 .)
Second, we prove that the PhaseLift convex program
achieves uniform recovery of all µ-flat vectors in Rn , using
m = O(n) subgaussian measurements, in the presence of
noise.3 This extends the work of Candès and Li, who showed
a similar statement for the recovery of all vectors x ∈ Rn ,
when the ai are Gaussian distributed [11].
Here, uniform recovery means that, with high probability,
a random choice of the ai will allow correct recovery of all
possible vectors x; in contrast, a non-uniform guarantee states
that for any particular vector x, with high probability over the
choice of the ai , x will be recovered correctly. Note that the
1 Essentially, the flatness condition in that paper ensures that the vector
x does not contain long strings of consecutive 0’s. This ensures that, after
recovering different pieces of the vector x from different sets of local
measurements, one can estimate the relative phase-differences among these
pieces, in order to “stitch them together” and recover x.
2 A conference version of this first part has appeared in [15].
3 This second part also extends a conference version, which has appeared
in [16].
3
use of m = O(n) measurements is optimal up to a constant
factor.
Our proof follows a similar approach as Candès and Li
[11], but we encounter some technical differences, since in our
setting the vectors ai do not have the convenient properties of
Gaussian random vectors, and at several points we need to
exploit the µ-flatness of the vector x.
Third, we consider a special class of subgaussian measurements, sometimes called random Bernoulli vectors with
erasures. Roughly speaking,4 these are random vectors ai
whose entries aij are chosen independently from some symmetric distribution on {1, 0, −1}. Intuitively, phase retrieval
works well with these kinds of measurement vectors, because
the presence of zeroes (“erasures”) prevents the pathological
behavior that occurs with Bernoulli vectors. Indeed, this idea
was used previously in work on phase retrieval using coded
diffraction patterns [13], [14].
We show that PhaseLift can recover all vectors in Rn
(without any restriction to non-peaky or µ-flat vectors), using
m = O(n) random Bernoulli measurements with erasures (using a particular choice p = 2/3 for the “erasure probability”).
Our recovery guarantee holds uniformly over all signals in Rn ,
with noisy measurements.
This gives a new example of a class of non-Gaussian
measurements where PhaseLift works nearly as well as it
does in the case of Gaussian measurements. We remark that
while stability and uniqueness in this setting were previously
known (since the aij satisfy Eldar and Mendelson’s fourthmoment conditions) [17], there is hardly any previous work
involving PhaseLift in this situation. (The only exception
we know is [22], which uses Bernoulli measurements with
erasures in an experimental procedure for characterizing linear
optical circuits, which have applications in quantum information processing. Ref. [22] also proves recovery guarantees for
PhaseLift with this particular class of measurements, but using
a different technique from ours.)
Overall, we see our work as a proof of concept that
shows that recovery guarantees are possible for many kinds of
random measurements that were not previously considered to
be suitable. This is useful in certain situations. For instance, the
authors of [22] remark that in their optical devices, Bernoulli
measurements with erasures can be implemented more easily
than Gaussian measurements, because they use fewer levels of
quantization.
An important next step will be to carry over this approach to
more realistic and practical scenarios, such as measurements
using coded diffraction patterns [13], [14]. Previous work in
this area uses measurements that involve random masks that
are generated from a very specific distribution. We think it
is an interesting question whether one can make a tradeoff,
similar to that shown in the present paper, which would allow
more flexibility in the choice of masks, at the expense of
slightly restricting the class of signals.
4 In the precise definition, the a
ij are multiplied by a normalization factor,
so that the distribution has variance 1.
D. Outline of the Paper, and Notation
We introduce some basic definitions in Section II. We state
and prove our results on stable uniqueness in Sections III and
IV. We then present our results on PhaseLift in Sections V
through IX. Finally, we describe some directions for future
work in Section X.
We let [m] denote the set {1, 2, . . . , m}. We write vectors
in boldface, and matrices in boldface capital letters. kxkp
denotes the ℓp norm of a vector x. kXkF and kXk denote
the Frobenius and operator (spectral) norms of a matrix X,
respectively.
II. P RELIMINARIES
In this paper, we will consider phaseless measurements of
the form (1). Following [17], we will consider the real case,
where the signal x and the measurement vectors ai are real
(rather than complex).
Also following [17], we will suppose that the measurement
vectors ai ∈ Rn are sampled independently from some
subgaussian distribution. We recall that a random vector in Rn
is called L-subgaussian if all of its one-dimensional marginals
are L-subgaussian in the following sense:
Definition II.1. (cf. [21]) A real valued random variable X
is subgaussian with parameter L, if for every u ≥ 1, one has
Pr[|X| ≥ Lu] ≤ 2 exp(−u2 /2).
Here we will consider the (more specific) situation where
each ai consists of independent subgaussian entries aij ∈ R,
each sampled from some distribution D.
The main results in [17] concern measurements ai that are
subgaussian and satisfy a small-ball probability assumption:
there exists some constant c > 0 such that, for all vectors
t ∈ Rn and for all ε > 0,
Pr[|aTi t| ≤ εktk2 ] ≤ cε.
(4)
In addition, some results are shown in [17] for measurements
that satisfy a fourth-moment condition: the aij are symmetric,
with variance E(a2ij ) = 1, and fourth moment E(a4ij ) > 1.
In contrast, here we will make no such assumptions on
the ai . In particular, our results will hold for Bernoulli
measurements ai , which are sampled uniformly from the set
n
{1, −1} , and which violate both the small-ball assumption5
and the fourth-moment condition.
Our results will apply to signals x ∈ Rn that are not too
peaky, in the following sense. Let µ ∈ (0, 1) be a constant
that depends on D, but not on the dimension n.
Definition II.2. We say that a vector x ∈ Rn is µ-flat if it
satisfies
kxk∞ ≤ µkxk2 ,
(5)
A set T ⊂ Rn is called µ-flat if all its elements are µ-flat.
5 This
can be seen by setting t = (1, 1, 0, . . . , 0).
4
III. S TABLE U NIQUENESS
Our first main result concerns stable uniqueness, and follows the approach taken by Eldar and Mendelson [17]. That
is, the goal will be to find conditions to ensure that if the
measurements y1 and y2 are close, then the underlying signals
x1 and x2 must also be close (up to sign ambiguity, i.e., either
x1 − x2 or x1 + x2 must be small).
For conciseness, we write the phaseless measurement operation as
y = φ(Ax) + w,
(6)
where A ∈ Rm×n is the matrix whose i’th row is the
measurement vector ai , and φ : Rm → Rm is the function
that maps
2
2
2
φ : (s1 , s2 , . . . , sm ) 7→ (|s1 | , |s2 | , . . . , |sm | ).
(7)
A. The Noise-Free Case
We begin by considering the noise-free case. Here the notion
of stability is formalized in the following definition. One
assumes that the signal x lies in some known set T ⊂ Rn , in
order to address situations where the signal has some known
structure, such as sparsity.
Definition III.1 (Definition 2.3 in [17]). The mapping φ(Ax)
is stable with constant C > 0 in a set T ⊂ Rn if for every
s, t ∈ T ,
kφ(As) − φ(At)k1 ≥ Cks − tk2 ks + tk2 .
One then shows that, if the number of measurements m
is sufficiently large (with respect to certain parameters that
quantify the complexity of the set T ), then the mapping φ(Ax)
is stable. In particular, Eldar and Mendelson [17] proved
results of this type, for several natural choices of the set T ,
where the measurements ai are subgaussian and satisfy smallball or fourth-moment assumptions. They showed that stability
is achieved with a number of measurements m that is only
slightly larger than the information-theoretic lower bound.
We show analogous results, for the same choices of the set T
as in [17], but restricted to µ-flat vectors with constant µ, and
where the measurements ai are subgaussian with independent
coordinates, but without any small-ball or fourth-moment
assumptions. In particular, we consider the set Tµ ⊂ Rn of
all µ-flat vectors, and the set Tµ,k ⊂ Rn of all vectors which
are both k-sparse and µ-flat. We find that stability holds with
a number of measurements m that scales in the same way as
in [17].
Theorem III.2. For every L > 0, there exist constants
1
,
c1 , . . . , c8 for which the following holds. Let 0 < µ < 2√
2
n
and let Tµ,k ⊂ Tµ ⊂ R be as in the preceding paragraph.
Consider a random vector a ∈ Rn with independent Lsubgaussian entries aj with mean zero and unit variance. Let
A ∈ Rm×n be a matrix whose rows are independent copies
of this vector. Then:
n
(a) for u ≥ c1 and m ≥ c2 u3 1−8µ
2 , one has with probability
2
at least 1 − 2 exp(−c3 u n) that the mapping φ(Ax) is
stable with constant c4 (1 − 8µ2 )1/2 in Tµ .
(b) for u ≥ c5 and m ≥ c6 u3 k log(en/k)
1−8µ2 , one has with
probability at least 1−2 exp(−c7 u2 k log(en/k)) that the
mapping φ(Ax) is stable with constant c8 (1 − 8µ2 )1/2
in Tµ,k .
To summarize, in these two instances of the phase retrieval
problem, most assumptions on the distribution of the measurement vectors ai can be dropped if µ-flatness is introduced
as an additional condition on the signal x, while leaving the
other parts of the result unchanged. As we will see, the proof
is quite general, and likely applies to many other instances of
the problem.
B. Proof Outline
We now sketch the proof of Theorem III.2. This uses the
framework introduced in [17], with some technical modifications to take advantage of the µ-flatness property of the signals.
First, we define the complexity parameter ρT,m as follows. We
define T+ and T− via
s−t
T− := { ks−tk
T+ :=
: s, t ∈ T, t 6= −s}
2
s+t
{ ks+tk
2
and then we set ρT,m =
: s, t ∈ T, t 6= s}
√E
m
E = max E sup
+
n
X
E2
m,
where
gi vi , E sup
n
X
gi wi
w∈T+ i=1
v∈T− i=1
with gi independent centered Gaussian random variables of
unit variance.
For technical reasons, we will slightly modify the definition
of the second complexity parameter κ that was used in [17].
Namely, in our definition of κ we restrict to S n−1 , setting for
any v, w ∈ S n−1
κ(v, w) = E |ha, viha, wi|.
(8)
In this modified notation and restricted to our measurement
setup, the main result of [17] for the noiseless case reads as
follows.
Theorem III.3 (Theorem 2.4 in [17]). For every L ≥ 1 and
T ⊂ Rn , there exist constants c1 , c2 , c3 that depend only on
L such that the following holds.
Let a ∈ Rn be a random vector with independent, Lsubgaussian entries with mean zero and unit variance. Consider a matrix A ∈ Rm×n whose rows are independent copies
of this vector.
Then, for u ≥ c1 , with probability ≥ 1 −
2 exp(−c2 u2 min(m, E 2 )), the mapping φ(Ax) is stable in
T with constant
s−t
s+t
C = inf κ( ks−tk
, ks+tk
) − c3 u3 ρT,m .
s,t∈T
2
(9)
2
Thus in addition to bounding ρT,m from above, it suffices
to estimate the infimum of κ over the set
s−t
s+t
T∓ = {( ks−tk
, ks+tk
) : s, t ∈ T, t 6= s, t 6= −s}. (10)
2
2
As it turns out, this refined infimum allows for sharper
bounds when the set under consideration consists of not too
5
peaky vectors. Our technical contribution consists of a lower
bound on κ, which holds for µ-flat sets:
Proposition III.4. For each L > 0 there exists a constant
c > 0 such that the following holds. Consider a random vector
a with independent L-subgaussian entries ai with mean zero
and unit variance. Let T∓ and κ be defined as in equations
1
(10) and (8). Then if T ⊂ Rn is µ-flat for some µ < 2√
, one
2
has
(11)
inf κ(v, w) ≥ c(1 − 8µ2 )1/2 .
A. Proof of Proposition III.5
By the µ-flatness assumption and as v, w ∈ S n−1 , one has
kvk∞ ≤ µ or kwk∞ ≤ µ
and thus
n
X
i=1
vi2 wi2 ≤ µ2 max(kvk22 , kwk22 ) = µ2 .
(13)
Set
Z = ha, viha, wi
(v,w)∈T∓
(14)
Proof of Theorem III.2:: We seek to apply Theorem III.3, and observe that
thus we need to bound the right hand side of (9) from
kZk2L2 = E |ha, viha, wi|2
below. Applying Proposition III.4 yields a lower bound of
n
X
c(1−8µ2 )1/2 for the first summand. For the second summand,
=E
ai aj ak aℓ vi vj wk wℓ
we ignore the µ-flatness, which can only make ρT,m larger
i,j,k,ℓ=1
and hence the bound smaller. The resulting setup is exactly
n
n
n
i
h X
X
X
the same as in [17], so the bounds from Sections 3.3.1 and
a2i a2k vi2 wk2 + a4i vi2 wi2
= E 2 a2i a2j vi vj wi wj +
3.3.2 in [17] directly imply
i=1
i,j=1
i,k=1
q
q
i6=j
i6=k
k log(en/k) k log(en/k)
n
n
n
n
ρTµ ,m . m + m and ρTµ,k ,m .
+
.
X
X
m
m
2
(E a4i − 1)vi2 wi2
vi2 wi2 +
= 1 + 2hv, wi − 2
Noting that in both cases, for our choice of m, the square root
i=1
i=1
is of leading order, (9) yields the result.
≥ 1 − 2µ2 .
(15)
C. The Noisy Case
In [17], also an analysis of the case of phase retrieval
with noise is presented. The results are technically somewhat
more involved. It should be noted, however, that again the
only place where additional assumptions on the measurement
vectors enter is that they ensure a lower bound of κ.
A minor difference from the noise-free case is that in this
s
, t ), s, t ∈ T
framework, one needs a bound on κ( ksk
2 ktk2
s−t
rather than κ( ks−tk
, s+t ) (both in terms of the definition
2 ks+tk2
of κ given above, which is slightly different from the one in
[17]). We obtain the following bound for the noisy case:
Proposition III.5. For each L > 0 there exists a constant
c > 0 such that the following holds. Consider a random vector
a with independent L-subgaussian entries ai with mean zero
and unit variance. Let κ be defined as in equation (8). Then
if v, w ∈ S n−1 and at least one of them is µ-flat for some
µ < √12 , one has
κ(v, w) ≥ c(1 − 2µ2 )1/2 .
(12)
s
t
Taking the infimum over all v = ksk
, w = ktk
for
2
2
s, t ∈ T yields analogous results to those in [17] for the noisy
case with independent measurement entries, where no small
ball probability or moment assumptions are required provided
T is µ-flat. Again, the stability constant has an additional
factor of (1−8µ2)1/2 . To prove the proposition, the ingredients
necessary in addition to the proof in [17] are exactly the same
as in the noise-free case discussed in Theorem III.2 above. We
thus refrain from repeating these details.
IV. L OWER - BOUNDS ON κ
Here we prove the lower-bounds on κ that are used to show
stable uniqueness. We will first prove Proposition III.5, then
use it to show Proposition III.4.
Here the third equality uses that due to the independence
assumption, all summands where an ai appears in first power
have zero mean, so only those terms with two different ai ’s
appearing as a square or just one ai appearing in fourth power
contribute to the sum. The fourth equality uses that the ai ’s
are all unit variance, and in the last inequality we use (13)
as well as the fact that a random variable’s fourth moment
always dominates its variance.
The result now follows tracing exactly the steps of Corollary
3.7 in [17].
B. Proof of Proposition III.4
Consider (v, w) ∈ T∓ . Then by definition, there exist
s−t
s+t
vectors s, t ∈ T such that v = ks−tk
and w = ks+tk
.
2
2
Using the triangle inequality, and the fact that s, t are µ-flat,
we have that:
ks + tk∞ ≤ µ(ksk2 + ktk2 ),
ks − tk∞ ≤ µ(ksk2 + ktk2 ).
(16)
(17)
Also, using the triangle inequality,
ks + tk2 + ks − tk2 ≥ k2sk2 ,
ks + tk2 + ks − tk2 ≥ k2tk2 ,
(18)
(19)
hence
ksk2 + ktk2 ≤ ks + tk2 + ks − tk2
≤ 2 max {ks + tk2 , ks − tk2 }.
(20)
Combining all of the above, we see that at least one of the
following inequalities must hold:
ks + tk∞ ≤ 2µks + tk2 ,
ks − tk∞ ≤ 2µks − tk2 .
(21)
(22)
This shows that at least one of s + t and s − t is 2µ-flat. The
result follows by applying Proposition III.5.
6
V. P HASE L IFT
Our second main result concerns the PhaseLift algorithm
[9], [10], [11]. PhaseLift (in the noiseless case) is based on a
matrix reformulation of the problem in terms of the rank one
matrix X = xxT , namely
find X ∈ Rn×n
such that X 0, rank(X) = 1,
tr(ai aTi X)
= yi
(23)
(∀i = 1, . . . , m).
Then one solves a convex relaxation of this problem:
find X ∈ Rn×n that minimizes tr(X)
such that X 0,
tr(ai aTi X) = yi
(24)
(∀i = 1, . . . , m).
b from which one extracts the leading
This yields a solution X,
eigenvector x̂; one hopes that x̂ ≈ x.
PhaseLift can be modified to handle noise in different ways;
in particular, one can solve the following convex program [11]:
find
s.t.
X ∈ Rn×n that minimizes
m
X
i=1
|tr(ai aTi X) − yi |
X 0.
(25)
A. Recovery of Non-Peaky Signals from Subgaussian Measurements
We assume that each measurement vector ai ∈ Rn is chosen
by sampling its entries aij (j = 1, 2, . . . , n) independently
from some subgaussian distribution D on R. We assume that
D has mean zero and variance 1. Also, we let C4 and Cψ2
denote its fourth moment and ψ2 -norm, respectively:6
E aij = 0,
C4 := E(a4ij ) ≥ 1,
E(a2ij ) = 1,
Cψ2 := kaij kψ2 .
(26)
We are interested in recovering signals that are non-peaky,
or µ-flat in the sense of Definition II.2. We show that, for any
choice of D, there is a constant µ ∈ (0, 1) such that, for all
sufficiently large n, PhaseLift achieves uniform recovery of
all µ-flat vectors in Rn , using m = O(n) measurements, in
the presence of noise.
This extends the order-optimal results of Candès and Li [11]
to a larger class of measurement distributions, at the expense
of a mild non-peakiness restriction on the class of vectors to
be recovered. We emphasize that there is no dependence of µ
on the dimension n.
More precisely, we prove:
Theorem V.1. Consider PhaseLift with noisy measurements,
as shown in equations (1) and (25). Let D be any subgaussian
distribution on R, with mean 0, variance 1, and parameters C4
and Cψ2 as shown in equation (26). Then there exist constants
0 < µ < 1 and κ0 > 1 such that the following holds.
For all n sufficiently large, and for all m ≥ κ0 n, with probability at least 1−e−Ω(n) (over the choice of the measurement
6 For
the definition of the ψ2 -norm, see, e.g., [21].
vectors ai ), PhaseLift achieves uniform recovery of all µ-flat
b
vectors in Rn : for all µ-flat vectors x ∈ Rn , the solution X
to equation (25) obeys
kwk1
,
(27)
m
where k·kF is the Frobenius norm, C0 is a universal constant,
b be the
and w is the noise term in equation (1). If we let x
b then x
b satisfies the bound
leading eigenvector of X,
kwk1
,
(28)
kb
x − eiφ xk2 ≤ C0 min kxk2 ,
mkxk2
b − xxT k ≤ C0
kX
F
for some φ ∈ [0, 2π].
Note that here the success probability is 1 − e−Ω(n) , rather
than 1−e−Ω(m) as in [11]. This is due to a technical difference
in the proof: in our setting, the sampling operator A satisfies
the desired injectivity properties with probability 1 − e−Ω(n) ,
rather than 1 − e−Ω(m) . However, note that the construction of
the dual certificate still succeeds with probability 1 − eΩ(m) ,
as needed in order to use the union bound over Rn . (See the
following section for details.)
B. Recovery of Arbitrary Signals from Bernoulli Measurements with Erasures
Next we consider a special class of subgaussian measurements, namely Bernoulli measurements with erasure probability p ∈ [0, 1]. We let each measurement vector ai ∈ Rn have
iid entries chosen from the following distribution:
√
with probability (1 − p)/2,
1/ 1 − p
(29)
aij = 0
with probability p,
√
−1/ 1 − p with probability (1 − p)/2.
This distribution has the following properties:
E aij = 0,
C4 := E(a4ij ) =
1
1−p ,
E(a2ij ) = 1,
Cψ2 := kaij kψ2 ≤
√1 .
1−p
(30)
We consider the phase retrieval problem using these types of
measurements, for arbitrary signals x ∈ Rn (without any µflatness assumption). The basic question of stable uniqueness
was already settled by Eldar and Mendelson [17]: whenever
p > 0, the distribution of the aij satisfies their fourth-moment
condition, hence stable uniqueness holds. However, no recovery guarantees using PhaseLift were previously known.
Here we show that, in the particular case where p = 2/3,
PhaseLift achieves uniform recovery of all vectors in Rn , using
m = O(n) measurements, in the presence of noise.
Theorem V.2. Consider PhaseLift with noisy measurements,
as shown in equations (1) and (25). Let the ai ∈ Rn be
Bernoulli random vectors with erasure probability p = 2/3, as
shown in equation (29). Then there exists a constant κ0 > 1
such that the following holds.
For all n sufficiently large, and for all m ≥ κ0 n, with
probability at least 1 − e−Ω(n) (over the choice of the ai ),
PhaseLift achieves uniform recovery of all vectors in Rn .
b satisfies
Furthermore, for all x ∈ Rn , the PhaseLift solution X
the same error bounds as in Theorem V.1.
7
VI. P ROOF O UTLINE
We begin by describing the proof of Theorem V.1. (We
will prove Theorem V.2 later, in Section IX.) Our proof uses
the same overall strategy as in previous work on PhaseLift
[9], [10], [11], though new techniques are needed in several
places.
We begin by showing an injectivity
property of the sampling
operator A : X 7→ tr(ai aTi X) i∈[m] , but while this follows
from relatively straightforward arguments in [11], [10], [9], in
our setting we need to use more sophisticated arguments due
to Eldar and Mendelson [17] (including a bound on empirical
processes from [20]).
To characterize the solution of the PhaseLift convex program, we construct a dual certificate Y, which is similar to that
of Candès and Li [11] (see also similar work by Demanet and
Hand [10]). Because the ai are no longer Gaussian distributed,
our proofs are somewhatmore involved: we use various fourthmoment estimates, and large-deviation bounds for sums of
independent subgaussian and sub-exponential random vectors
[21].
A. Injectivity of the sampling operator
We define the sampling operator A : Rn×n → Rm as
follows:
(31)
A(X) = tr(ai aTi X) i∈[m] .
We will prove that A satisfies certain injectivity properties.
First, we need an upper bound on kA(X)k1 , which is a
straightforward generalization of the first half of Lemma 2.1
in [11], and of Lemma 3.1 in [9]:
Lemma VI.1. Let D be as in Theorem V.1. Then there exist
constants C > 0 and c > 0 such that the following holds.
Fix any 0 < δ < 12 , and assume that m ≥ 20(C/δ)2 n.
Let ε > 0 be the positive root of 14 δ = ε2 + ε, that is, ε =
√
1
−cmε2
2 (−1 + 1 + δ). Then with probability at least 1 − 2e
(over the choice of the ai ), the sampling operator A has the
property that:
1
m kA(X)k1
≤ (1 + δ) tr(X),
∀X 0.
(32)
The proof is identical to that of Lemma 3.1 in [9], but using
bounds on the singular values of random matrices whose rows
are independent subgaussian vectors (rather than Gaussian
vectors), as in Theorem 5.39 in [21].
Next we prove a lower bound on kA(X)k1 , which resembles
the second half of Lemma 2.1 in [11], and Lemma 3.2 in [9].
Our lemma differs from these previous works in that it only
applies to matrices X that lie in the tangent spaces T (x0 )
associated with µ-flat vectors x0 , rather than all matrices
X that are symmetric with rank 2. However, this lemma is
sufficient for our purposes.
Formally, fix some µ ∈ (0, 1), and let S(µ) be the set of
all µ-flat vectors:
S(µ) = {x ∈ Rn | kxk∞ ≤ µkxk2 }.
(33)
For any vector x0 ∈ S, define T (x0 ) to be the following
subspace of Rn×n :
T (x0 ) = {X ∈ Rn×n | X = xxT0 + x0 xT , x ∈ Rn }. (34)
We will prove:
Lemma VI.2. Let D be as in Theorem V.1. Then there exist
constants 0 < µ < 1, κ0 > 1, c > 0 and α > 0 such that the
following holds.
For all n sufficiently large, and for all m ≥ κ0 n, with probability at least 1 − e−cn (over the choice of the measurement
vectors ai ), the sampling operator A satisfies:
1
m kA(X)k1
≥ αkXk,
∀x0 ∈ S(µ)\{0},
∀X ∈ T (x0 )\{0}.
(35)
Note that here the constant α > 0 may be quite small, in
contrast to previous work [11] where this constant was close to
1. Because of this difference, we will have to construct a dual
certificate Y that satisfies kYT kF ≤ εT (for a small constant
3
as in [11]. (See Sections
εT > 0), rather than kYT kF ≤ 20
VI-B and VI-C for details.)
The proof of this lemma is different from [11], [9]. It uses
arguments due to Eldar and Mendelson, in particular a PaleyZygmund argument for lower-bounding an expectation value
of the form E |Z| (Lemma 3.6 in [17]), and a bound on the
suprema of certain empirical processes (Theorem 2.8 in [17],
see also [20]). The proof of this lemma is given in Section
VII.
Remark VI.3. Note that Lemma VI.2 crucially relies on
the assumption that x0 is µ-flat. For example, let the ai be
n
random Bernoulli measurements in {1, −1} , choose x0 =
T
T
T
(1, 1, 0, . . . , 0) , and let X = xx0 + x0 x ∈ T (x0 ) where
x = (1, −1, 0, . . . , 0)T . This yields A(X) = 0 (this follows
from (38) below, as for each of the ai , either aTi x or aTi x0
vanishes), so the conclusion of the Lemma does not hold.
B. Approximate dual certificates
Following [11], we will use approximate dual certificates
to characterize the solutions to the PhaseLift convex program.
First, we will show that, for any µ-flat vector x0 , with high
probability (over the choice of the measurement vectors ai ),
there exists an approximate dual certificate Y (for the vector
x0 ). We will then use the union bound over an ǫ-net on the
unit sphere S n−1 in Rn , together with a continuity argument
(which shows that any vector in S n−1 can be approximated
by a vector in the ε-net). This will prove a uniform guarantee:
with high probability (over the choice of the ai ), for all µ-flat
vectors x0 , there is a dual certificate Y.
For the first claim, we will prove a variant of Lemma 2.3
in [11]:
Lemma VI.4. Let D be as in Theorem V.1. Let 0 < εT ⊥ ≤ 1
and 0 < εT ≤ 1 be any constants. Then there exist constants
0 < µ < 1, κ0 > 1, c > 0 and B0 > 0 such that the following
holds.
For all sufficiently large n, and for all m ≥ κ0 n, let x0 ∈
S(µ) be any µ-flat vector, and let T = T (x0 ) be the tangent
space. Then with probability at least 1−e−cm (over the choice
of the measurement vectors ai ), there exists a dual certificate
Y ∈ Rn×n , which has the form Y = A∗ (λ), where kλk∞ ≤
B0
m , and which satisfies
kYT ⊥ + 2IT ⊥ k ≤ εT ⊥ ,
kYT kF ≤ εT .
(36)
8
Here, A∗ : Rm → Rn×n denotes the adjoint of the linear
operator A, YT denotes the projection of Y onto the subspace
T , and YT ⊥ denotes the projection of Y onto the subspace
T ⊥ , which is the orthogonal complement of T in Rn×n .
The proof of this lemma is given in Section VIII. The dual
certificate is constructed in a similar way to [11], but the
analysis is more involved. In particular, unlike the Gaussian
case studied in [11], here the distribution of the vectors ai
is not rotationally invariant. We use vector analogues of Hoeffding’s inequality and Bernstein’s inequality [21], combined
with fourth-moment estimates which depend on the µ-flatness
of the vector x0 .
From the above lemma, one concludes a uniform guarantee
on the existence of dual certificates for all µ-flat vectors x0 .
To see this, one first constructs a dual certificate for all points
x0 in an ǫ-net, via a union bound; second, one shows that this
dual certificate works well for all points x0 , using a continuity
argument. This part of the proof is identical to the proof of
Corollary 2.4 in [11], so we refrain from repeating the details.
comes from the Paley-Zygmund argument from reference [17]
(see equation (45)).
VII. I NJECTIVITY
OF THE SAMPLING OPERATOR
We now prove Lemma VI.2:
For any x0 ∈ S(µ) \ {0}, and any X ∈ T (x0 ) \ {0}, we
can write X = xxT0 + x0 xT . Since all the variables are real,
we can write
m
m
X
X
T
1
1
2
|tr(ai ai X)| = m
|aTi x||aTi x0 |.
m kA(X)k1 = m
i=1
i=1
(38)
Next we reduce to the case where kxk2 = kx0 k2 = 1. Let
S n−1 denote the unit sphere in Rn . For any v ∈ S n−1 , and
any v0 ∈ S(µ) ∩ S n−1 , define
Γ(v, v0 ) :=
1
m
m
X
i=1
|aTi v||aTi v0 |.
(39)
Then we can write
1
m kA(X)k1
C. Combining the pieces
We can characterize the solution of the PhaseLift convex
program, using the injectivity of the sampling operator A, and
the existence of a dual certificate Y. We restate the bound
proved in Section 2.3 in [11] (with more general choices for
the parameters):
Lemma VI.5. Suppose that the sampling operator A satisfies
equations (32) and (35), with parameters δ and α. Let x0
b be the solution of
be any µ-flat vector in Rn , and let X
the PhaseLift convex program with noisy measurements, as
in equations (1) and (25).
Suppose that there exists a dual certificate Y ∈ Rn×n
that satisfies YT ⊥ −IT ⊥ and kYT kF ≤ εT , where
α√
. Also suppose that Y has the form Y = A∗ (λ),
εT < (1+δ)
2
b must satisfy
where kλk ≤ B0 . Then X
∞
m
b − x0 xT k ≤ 2C0
kX
0 F
where C0 = (1 + εT )
α√
(1+δ) 2
− εT
kwk1
,
m
−1
(37)
1
( 1+δ
+ B0 ) + B0 .
The proof of this lemma is identical to that in [11]. By
combining this with the preceding lemmas, we prove Theorem
V.1.
Remark VI.6. Theorem V.1 shows that there exists some
constant µ > 0 such that PhaseLift succeeds in recovering
all µ-flat vectors. The specific value of µ depends on a few
factors. First, in Lemma VI.2, in order to show
√ injectivity of
the sampling operator, we must have µ < 1/ 2 (see equation
(45)). More importantly, in Lemma VI.4,p
in order to construct
the dual certificate, we need to set µ ≤ ε1 /|C4 − 3|, where
1
min {εT ⊥ , εT } (see equations (51) and (52)). The
ε1 = 20
specific values of εT ⊥ and εT are determined by Lemma VI.5:
α√
. The most important
they are εT ⊥ = 1 and εT < (1+δ)
2
factor in determining these values is α, which describes the
injectivity of the sampling operator in Lemma VI.2. This can
be estimated as α ≈ c1 , where c1 is a numerical constant that
x
= 2kxk2 kx0 k2 Γ( kxk
, kxx00k )
2
2
x
, kxx00k )kXk,
≥ Γ( kxk
2
(40)
2
where we used the fact that kXk ≤ kxxT0 k + kx0 xT k ≤
2kxk2 kx0 k2 .
Furthermore, define
κ(v, v0 ) := EA Γ(v, v0 )
(41)
and
δ :=
sup
v∈S n−1 , v0 ∈S(µ)∩S n−1
|Γ(v, v0 ) − EA Γ(v, v0 )|. (42)
Then we have
1
m kA(X)k1
x
≥ κ( kxk
, kxx00k ) − δ kXk.
2
2
(43)
We now proceed to bound κ(v, v0 ), using a version of the
Paley-Zygmund argument from Corollary 3.7 in [17]. We note
that κ(v, v0 ) can be written in terms of a single measurement
vector a, as follows:
κ(v, v0 ) = Ea |aT v||aT v0 | = E |Z|,
T
(44)
T
where Z := (a v)(a v0 ). Using (15), we have that kZkL2 ≥
p
1 − 2µ2 .
Now fix any q > 2. Using the same argument as in Corollary
3.7 in [17], we have that kZkLq ≤ Cq , where Cq < ∞
depends only on the parameter q and the distribution D (and
not on the dimension n). Finally, using Lemma 3.6 in [17],
we get that
p
κ(v, v0 ) = E |Z| ≥ c1 kZkL2 ≥ c1 1 − 2µ2 ,
(45)
where c1 > 0 depends only on the choice of q and the
distribution D (and not on the dimension
n). We assume that
√
µ is a constant satisfying µ < 1/ 2; then κ(v, v0 ) is lowerbounded by a constant that is strictly greater than 0.
Next, we will upper-bound δ, by bounding the supremum
of an empirical process, using Theorem 2.8 in [17] (see also
[20]). Here the empirical process is indexed by two sets T1 =
9
S n−1 and T2 = S(µ) ∩ S n−1 . We define the diameter of a
set T ⊂ Rn to be d(T ) = supt∈T ktk2 . So for the sets T1
and T2 , we have d(T1 ) = d(T2 ) = 1. We define the Gaussian
complexity of a set T ⊂ Rn to be ℓ(T ) = Eg supt∈T |gT t|,
where g is a Gaussian random vector in Rn with mean 0 and
covariance matrix I. The Gaussian complexities
of T1 and T2
√
can be bounded by ℓ(T2 ) ≤ ℓ(T1 ) ≤ n (see, e.g., Section
3.3.1 in [17]).
Now Theorem 2.8 in [17] implies that there exist constants
c1 , c2 > 0 and c3 (which depend on the distribution D, but
not on the dimension n) such that the following holds: for all
u ≥ c1 , with probability at least 1 − 2 exp(−c2 u2 min {m, n})
(over the choice of the measurement vectors ai ), we have that
q
n
n
+m
).
(46)
δ ≤ c3 u 3 ( m
Now set u to be constant, and set m ≥ κ0 n. Then we have
δ ≤ c3 u3 ( √1κ0 + κ10 ). By choosing κ0 large, we can make δ
arbitrarily small.
Now substitute equations (45) and 46 into (43). This completes the proof of Lemma VI.2.
VIII. C ONSTRUCTING
We now prove Lemma VI.4:
Part 1: We first show some properties of the distribution D.
We claim that for any δct > 0, we can define a cutoff radius
Rct > 0 such that, for any x ∈ S n−1 , we have
where c > 0 is an absolute constant. Then set Rct to be:
(49)
We will set δct small enough to satisfy various bounds that
are needed in the proof, in particular, equations (84), (89) and
(110).
We now calculate certain fourth moments of the ai , namely
E[(aTi x0 )4 ], E[(aTi x0 )2 (aTi v)2 ] and E[(aTi x0 )3 (aTi v)] (where
v ∈ Rn is a unit vector orthogonal to x0 ). We show that when
the vector x0 is µ-flat, these moments have approximately the
same values as if ai were a Gaussian random vector. This is
the key fact that allows us to apply the proof techniques from
[11] in this more general setting.
First, let x0 be a “µ-flat” signal, and assume without loss
of generality that kx0 k2 = 1. We consider the 4th moment of
aTi x0 , and we write
E[(aTi x0 )4 ] = E
=
X
n
X
j=1
E[a4ij ]x40j
+3
j
=
X
j
=
X
j
aij x0j
X
4
(C4 −
+3
X
(C4 − 3)x40j + 3.
jk
(50)
x20j x20k
n
min {εT ⊥ , εT }.
(52)
Next, let v ∈ R be a unit vector orthogonal to x0 , i.e.,
kvk2 = 1 and vT x0 = 0. We calculate the following mixed
4th moment of ai :
E[(aTi x0 )2 (aTi v)2 ]
X
E[a4ij ]x20j vj2 +
=
j
X
E[a2ij ] E[a2ik ](x20j vk2 + x0j x0k vj vk + x0j x0k vk vj )
j6=k
= (C4 − 3)
= (C4 − 3)
X
x20j vj2 +
j
X
(x20j vk2 + 2x0j x0k vj vk )
jk
X
x20j vj2
j
X
2
2
+ kx0 k2 kvk2 + 2(xT0 v)2
x20j vj2 + 1.
j
(53)
In the special case where ai is a Gaussian random vector, we
have C4 = 3 and E[(aTi x0 )2 (aTi v)2 ] = 1. In general, we have
2
2
|E[(aTi x0 )2 (aTi v)2 ] − 1| ≤ |C4 − 3| kx0 k∞ kvk2
≤ |C4 − 3| µ2 = ε1 .
(54)
Finally, we calculate another mixed 4th moment of ai :
E[(aTi x0 )3 (aTi v)]
X
X
E[a4ij ]x30j vj + 3
E[a2ij ] E[a2ik ]x20j x0k vk
=
j
= (C4 − 3)
= (C4 − 3)
= (C4 − 3)
X
j6=k
x30j vj
+3
j
X
x20j x0k vk
jk
x30j vj
+
j
X
X
(55)
2
3kx0 k2 (xT0 v)
x30j vj .
j
In the special case where ai is a Gaussian random vector, we
have C4 = 3 and E[(aTi x0 )3 (aTi v)] = 0. In general, we have
2
|E[(aTi x0 )3 (aTi v)]| ≤ |C4 − 3| kx0 k∞ kx0 k2 kvk2
≤ |C4 − 3| µ2 = ε1 .
E[a2ij ] E[a2ik ]x20j x20k
j6=k
3)x40j
1
20
ε1 =
(47)
To see this, use a Hoeffding-type inequality (Prop. 5.10 in
[21]):
(48)
Pr[|aTi x| ≥ Rct ] ≤ e · exp(−cRct2 /Cψ2 2 ),
Rct = ( 1c ln( δect ))1/2 Cψ2 .
Note that we can choose µ > 0 to be small, in order to make
ε1 an arbitrarily small constant. In particular, let us choose µ
small enough so that:
= (C4 − 3)
THE DUAL CERTIFICATE
Pr[|aTi x| ≥ Rct ] ≤ δct .
In the special case where ai is a Gaussian random vector, we
have C4 = 3 and E[(aTi x0 )4 ] = 3. In the general case, we
can use the µ-flatness of x0 to show that E[(aTi x0 )4 ] ≈ 3, as
follows:
|E[(aTi x0 )4 ] − 3| ≤ |C4 − 3| kx0 k2∞ kx0 k22
(51)
≤ |C4 − 3| µ2 =: ε1 .
(56)
Part 2: We now construct the dual certificate Y, following
the same approach as [11]. Without loss of generality, we can
assume kx0 k2 = 1. We construct Y as follows:
Y :=
m
X
i=1
λi ai aTi ,
1 T
λi :=
(ai x0 )2 1[|aTi x0 | ≤ Rct ] − β0 ,
m
(57)
10
where we set
β0 := E[(aT1 x0 )4 1[|aT1 x0 | ≤ Rct ]].
(58)
This choice of β0 ensures that E xT0 Yx0 = 0, which will be
useful later.
Note that β0 ≈ E[(aT1 x0 )4 ] ≈ 3, and note that this implies
1
1
(Rct2 + β0 ) ≤ O( m
). In order to make this rigorous,
|λi | ≤ m
we can bound β0 as follows. First define
=
≤
≤
≤
1
(59)
1
20 εT ⊥ .
1
10 εT ⊥ .
1 X T
(a x0 )2 1[|aTi x0 | ≤ Rct ] ai aTi ,
m i=1 i
m
Y
1 X
β0 ai aTi .
:=
m i=1
(61)
(62)
Part 3: First, we want to show that YT ⊥ + 2IT ⊥ is small.
We can write
(0)
(1)
kYT ⊥ + 2IT ⊥ k ≤ kYT ⊥ − IT ⊥ k + kYT ⊥ − 3IT ⊥ k. (63)
(1)
We begin by showing that YT ⊥ − 3IT ⊥ is small — this is
the second term in (63). This uses a similar argument to [11],
with more general bounds for subgaussian random matrices
[21]. We can write
Y
(1)
=
√C
κ0
≤
1
15 εT ⊥ .
(68)
(1)
kYT ⊥ − 3IT ⊥ k ≤ kYT ⊥ − β0 IT ⊥ k + ε1 + εct
≤ kY(1) − β0 Ik + ε1 + εct
≤
≤
≤
1
15 εT ⊥ β0 + ε1 + εct
1
15 εT ⊥ (3 + ε1 + εct )
1
1
5 εT ⊥ + (1 + 15 )(ε1
+ ε1 + εct
(69)
+ εct )
< (0.31) εT ⊥ .
(0)
m
(1)
(67)
(60)
We now show that Y has the desired properties. To do this,
it is convenient to write Y = Y(0) − Y(1) , where
Y(0) :=
1
km
AT A − Ik ≤ 3 max {2ε0 , (2ε0 )2 } ≤ 6ε0 =
(1)
Combining with equation (51), we see that
|β0 − 3| ≤ ε1 + εct ≤
1
km
AT A − Ik ≤ 3 max { √Cκ0 + ε0 , ( √Cκ0 + ε0 )2 }.
This implies
where we used the Cauchy-Schwarz inequality and standard
properties of subgaussian random variables, and where C0 is
some universal constant (see Lemma 5.9 in [21]). Note that,
by choosing Rct sufficiently large, we can make δct , and hence
εct , an arbitrarily small constant. In particular, let us choose
Rct large enough so that
εct =
1 − C √1κ0 − ε0 ≤ smin ( √Am ) ≤ smax ( √Am ) ≤ 1 + C √1κ0 + ε0 .
(66)
This implies (using Lemma 5.36 in [21]) that
1
εT ⊥ , and set κ0 large enough so that
Now set ε0 = 90
ε0 . Then we have
β01 := E[(aT1 x0 )4 ] − β0
E[(aT1 x0 )4 [|aT1 x0 | > Rct ]]
E[(aT1 x0 )8 ]1/2 Pr[|aT1 x0 | > Rct ]1/2
4 √ 4 1/2
kaT1 x0 kψ2 8 δct
1/2
C02 Cψ4 2 · 64δct =: εct ,
√
We restate this as follows, substituting t = ε0 m: for all
ε0 ≥ 0, with probability ≥ 1 − 2 exp(−cε20 m), A satisfies
T
1
m β0 A A,
Next, we will show that YT ⊥ − IT ⊥ is small — this is the
first term in (63). This requires a more involved argument. We
(0)
can write YT ⊥ in the form:
m
(0)
YT ⊥ =
ξ i := (aTi x0 )1[|aTi x0 | ≤ Rct ](Π0 ai ),
(70)
where Π0 = I − x0 xT0 is the projector onto the subspace
orthogonal to x0 . As in [11], the random variables ξi are
subgaussian, but now there is an added complication, because
the ξ i may have nonzero mean and may not be isotropic. We
will show that the contributions due to the nonzero means and
anisotropy are small, hence the argument from [11] can be
adapted to this situation.
We first shift the ξ i , in order to get centered random
variables ζ i with E ζ i = 0, that is:
ζ i := ξ i − µi ,
We then define
µi := E ξ i .
(71)
m
Z :=
(64)
where A is a random m-by-n matrix whose i’th row is the
vector aTi . Note that E Y(1) = β0 I and E A = 0. Using
standard bounds on the singular values of random matrices
with independent subgaussian rows (Lemma 5.39 in [21]), we
get that there exist constants C > 0 and c > 0 (which depend
only on the subgaussian norm of the rows, given by C4 ), such
that for all t ≥ 0, with probability ≥ 1 − 2 exp(−ct2 ), A
satisfies
√
√
√
√
m − C n − t ≤ smin (A) ≤ smax (A) ≤ m + C n + t,
(65)
where smax (A) and smin (A) denote the largest and smallest
singular values of A.
1 X
ξ ξT ,
m i=1 i i
1 X
ζ ζT .
m i=1 i i
(72)
(0)
We claim, first, that Z is a good approximation to YT ⊥ ;
second, that the ζ i are approximately isotropic, so E Z will
be close to IT ⊥ ; and third, that Z is concentrated around its
expectation E Z. Finally, we will use these claims to upper(0)
bound YT ⊥ − IT ⊥ , as follows:
(0)
(0)
kYT ⊥ − IT ⊥ k ≤ kYT ⊥ − Zk + kZ − E Zk + kE Z − IT ⊥ k
(73)
(0)
We begin by upper-bounding YT ⊥ − Z (the first term in
(73)). The first step is to upper-bound the µi . We write
ξi = θi − ν i ,
(74)
11
where θi = (aTi x0 )(Π0 ai ) and ν i = (aTi x0 )1[|aTi x0 | >
Rct ](Π0 ai ). For any v ∈ S n−1 such that vT x0 = 0, we have
E[vT θ i ] = E[(aTi x0 )(vT ai )] = vT E[ai aTi ]x0 = vT x0 = 0,
(75)
and using equation (54), we have
|vT E ν i | = E (aTi x0 )1[|aTi x0 | > Rct ](vT ai )
≤ E[(aTi x0 )2 (vT ai )2 ]1/2 Pr[|aTi x0 | > Rct ]1/2
p
√
≤ 1 + ε1 δct ,
(76)
√
√
hence we conclude that E θi = 0, kE ν i k2 ≤ 1 + ε1 δct ,
and
p
√
kµi k2 ≤ 1 + ε1 δct .
(77)
(0)
We can now bound YT ⊥ − Z as follows:
(0)
kYT ⊥
− Zk =
=
1
km
m
X
(ζ i µTi
i=1
m
X
1
k( m
1
≤ 2 km
ζ i )µT1
i=1
m
X
i=1
m
X
1
≤ 2 km
+
i=1
µi ζ Ti
+
+
1
µ1 ( m
T
ζ i) +
µ1 µT1 k
2
1 + ε1
p
δct + (1 + ε1 )δct .
(78)
1 Pm
We will bound k m
i=1 ζ i k2 using a Hoeffding-type inequality for subgaussian random vectors (Lemma A.1). Note
that the ζ i are independent, centered random variables taking
values in span(x0 )⊥ ⊂ Rn , and they are subgaussian with
norm
kζ i kψ2 ≤ kξ i kψ2 + kµi kψ2
p
√
≤ Rct CCψ2 + 1 + ε1 δct =: K.
(79)
Using Lemma A.1, we have that there exists a universal
constant c > 0 such that, for all t ≥ 0 and all 0 < ε < 1,
1
Pr[k m
m
X
i=1
ζ i k2 ≥
1
1−ε t]
≤ e · exp n ln( 3ε ) −
ct2
K2 m
. (80)
Setting ε = 12 and t = K(1 − ε), and recalling that m ≥ κ0 n,
we get that:
1
Pr[k m
m
X
i=1
ζ i k2 ≥ K] ≤ e · exp
ln(6)
κ0 m
E Z = E ζ i ζ Ti = E ξ i ξ Ti − µi µTi ,
(85)
E ξ i ξ Ti = E θ i θTi − E θ i ν Ti − E ν i θ Ti + E ν i ν Ti
ζ i k2 kµ1 k2 + kµ1 k2
√
Next, we show that E Z is close to IT ⊥ (corresponding to
the third term in (73)). First we write
2
i=1
ζ i k2
This quantity can be made arbitrarily small, by choosing δct to
be a sufficiently small constant. In particular, we set δct small
enough that
(0)
1
kYT ⊥ − Zk ≤ 10
εT ⊥ .
(84)
and note that kµi µTi k = kµi k2 ≤ (1 + ε1 )δct . Next we write
µi µTi )k
m
X
the definition of K in equation (79), and setting Rct as a
function of δct as specified in equation (49), we get that
p
√
(0)
kYT ⊥ − Zk ≤ 2K 1 + ε1 δct + (1 + ε1 )δct
p
√
≤ 2Rct CCψ2 1 + ε1 δct + 3(1 + ε1 )δct
p
√
≤ 2( 1c ln( δect ))1/2 Cψ2 2 C 1 + ε1 δct + 3(1 + ε1 )δct .
(83)
− 4c m .
ln(6)
κ0
<
By setting κ0 is sufficiently large so that
that:
m
X
1
ζ i k2 ≥ K] ≤ exp(−Ω(m)).
Pr[k m
c
4,
(81)
= E θ i θTi − E ν i ν Ti ,
(86)
using the fact that θ i ν Ti = ν i θTi = ν i ν Ti . For any v ∈ S n−1
such that vT x0 = 0, we have
vT E[θi θTi ]v = E[(aTi x0 )2 (vT ai )2 ] ∈ [1 − ε1 , 1 + ε1 ], (87)
and using the same argument as in equation (59), we have
|vT E[ν i ν Ti ]v| = E (aTi x0 )2 1[|aTi x0 | > Rct ](vT ai )2
≤ E[(aTi x0 )4 (vT ai )4 ]1/2 Pr[|aTi x0 | > Rct ]1/2
1/2
≤ E[(aTi x0 )8 ]1/4 E[(vT ai )8 ]1/4 δct
1/2
≤ C02 Cψ4 2 · 64δct
≤ εct ,
(88)
hence we conclude that kE θi θTi − IT ⊥ k ≤ ε1 and
kE ν i ν Ti k ≤ εct . Combining these pieces, and setting δct
sufficiently small, we get that
kE Z − IT ⊥ k ≤ ε1 + εct + (1 + ε1 )δct
≤
≤
1
10 εT ⊥
1
5 εT ⊥ .
+ (1 + ε1 )δct
(89)
Finally, we will upper-bound Z − E Z (the second term in
(73)). To accomplish this, we will whiten the random variables
ζ i , to make them isotropic. We consider the covariance matrix
of the ζ i , and we compute its spectral decomposition,
E Z = E ζ i ζ Ti =: QDQT .
(90)
we get
Then we transform each ζ i to get an isotropic random variable
ϕi :
ϕi := D−1/2 QT ζ i ,
(91)
(82)
where Q, D and D−1/2 are understood to act on the subspace
orthogonal to x0 . We also define
i=1
Thus,
Pm with probability ≥ 1 − exp(−Ω(m)), we have that
1
km
i=1 ζ i k2 ≤ K. Plugging this into equation (78), using
m
F :=
1 X
ϕ ϕT ,
m i=1 i i
(92)
12
so we have E F = IT ⊥ . We can then upper-bound Z − E Z as
follows:
is bounded by Rct4 , and the second term is the square of a
subgaussian random variable. In particular, we can write
kZ − E Zk = kQD1/2 (F − IT ⊥ )D1/2 QT k
≤ kDk kF − IT ⊥ k
kξi kψ1 ≤ Rct4 + β0 k(aTi x0 )2 kψ1
≤ (1 +
1
5 εT ⊥ ) kF
(93)
− IT ⊥ k,
where we used equation (89).
1 Pm
T
We can now bound F = m
i=1 ϕi ϕi as follows. Note
that the ϕi are subgaussian, with norm
kϕi kψ2 ≤ kD−1/2 kkζ i kψ2
kξ i kψ + kµi kψ2
≤ q2
1 − 15 εT ⊥
√
√
Rct CCψ2 + 1 + ε1 δct
q
≤
,
which is constant with respect to n. We bound F using the
same argument that led to equation (69). We get that there exist
constants C > 0 and c > 0 (which depend only on kϕi kψ2 ),
such that for all ε0 ≥ 0, with probability ≥ 1−2 exp(−cε20 m),
F satisfies
kF − IT ⊥ k ≤ 6ε0 .
(95)
1
30 εT ⊥ .
(96)
≤ (0.24) εT ⊥ .
Combining equations (73), (84), (89) and (96), we see that
with probability ≥ 1 − e−Ω(m) , we have
(0)
1
εT ⊥ + 15 εT ⊥ + (0.24) εT ⊥
kYT ⊥ − IT ⊥ k ≤ 10
= (0.54) εT ⊥ .
(97)
kYT ⊥ + 2IT ⊥ k ≤ (0.31) εT ⊥ + (0.54) εT ⊥ = (0.85) εT ⊥ ,
(98)
as desired.
Part 4: We now want to bound YT . Following [11], we
can write
2
where we used Lemmas 5.14 and 5.24 in [21], and C is
some universal constant. Thus we can use a Bernstein-type
inequality (Cor. 5.17 in [21]) to bound xT0 Yx0 . We get that
there exists a universal constant c > 0 such that, for all ε ≥ 0,
ε
Pr[|xT0 Yx0 | ≥ ε] ≤ 2 exp(−c min { K
2,
2
kYT kF = |xT0 Yx0 | + 2kΠ0 Yx0 k2 ,
(99)
where Π0 = I − x0 xT0 is the projector onto span(x0 )⊥ , and
we assume (without loss of generality) that kx0 k2 = 1. We
will bound each term in equation (99) separately.
For the first term in (99), we use same argument as in [11].
We can write
m
1 X
ξi ,
(100)
xT0 Yx0 =
m i=1
where ξi := (aTi x0 )4 1[|aTi x0 | ≤ Rct ] − β0 (aTi x0 )2 . Note that
E ξi = 0, by our choice of β0 in equation (58). Furthermore,
ξi is a sub-exponential random variable, since the first term
ε
K }m).
(102)
Now set ε = 12 εT . Then, with probability ≥ 1 − exp(−Ω(m)),
we have
|xT0 Yx0 | < 21 εT .
(103)
For the second term in (99), we need to use a different
argument. The proof in [11] uses the fact that aTi x0 and Π0 ai
are independent random variables, when ai is sampled from a
spherical Gaussian distribution; but this no longer holds in our
setting. Instead, we give a more general argument. We write
m
1 X
si − ti ,
Π0 Yx0 =
m i=1
(104)
where si := (aTi x0 )3 1[|aTi x0 | ≤ Rct ](Π0 ai ), and ti :=
β0 (aTi x0 )(Π0 ai ). We note that si is a subgaussian random
vector and ti is a sub-exponential random vector. We then use
a Bernstein-type inequality (for vectors rather than scalars) to
bound Π0 Yx0 .
To make this precise, we define centered random variables
ui = si − ti − E si + E ti , and we write
m
Finally, combining equations (63), (69) and (97), we see
that with probability ≥ 1 − e−Ω(m) ,
2
=: K,
This implies the
kZ − E Zk ≤ 51 εT ⊥ (1 + 15 εT ⊥ )
(101)
≤ Rct4 + 2β0 C 2 Cψ2 2
2
(94)
4
5
In particular, we can set ε0 =
following bound on Z − E Z:
2
≤ Rct4 + 2β0 kaTi x0 kψ2
Π0 Yx0 =
m
1 X
1 X
ui +
(E si − E ti ).
m i=1
m i=1
(105)
We begin by calculating E si and E ti . Let us write si as a
difference of two terms,
si = si0 − si1 ,
(106)
where we define si0 := (aTi x0 )3 (Π0 ai ) and si1 :=
(aTi x0 )3 1[|aTi x0 | > Rct ](Π0 ai ). Now consider any vector
v ∈ S n−1 such that vT x0 = 0. We have the following bounds:
|vT (E si0 )| = |E[(aTi x0 )3 (vT ai )]| ≤ ε1 ,
(107)
where we used equation (56);
|vT (E si1 )| = |E[(aTi x0 )3 (vT ai )1[|aTi x0 | > Rct ]]|
≤ E[(aTi x0 )6 (vT ai )2 ]1/2 Pr[|aTi x0 | > Rct ]1/2
1/2
≤ E[(aTi x0 )12 ]1/4 E[(vT ai )4 ]1/4 δct
√
√
1/2
≤ (kaTi x0 kψ2 12)3 (kvT ai kψ2 4) δct
√
1/2
≤ (CCψ2 )4 (48 3) δct ,
(108)
13
where we used standard properties of subgaussian random
variables, with C a universal constant; and
vT (E ti ) = E[β0 (vT ai )(aTi x0 )]
= β0 vT E[ai aTi ]x0 = β0 vT x0 = 0.
(109)
≤
=
1
20 εT
1
4 εT ,
+ 51 εT
(110)
(111)
Next, we note that si and ti are sub-exponential random
vectors, whose norms are bounded by:
ksi kψ1 = sup kvT si kψ1 ≤ sup kvT si kψ2
v∈S n−1
v∈S n−1
3
T
Rct kv Π0 ai k ≤ Rct3 CCψ2 ,
(118)
Finally we combine equations (105), (110), (111) and
(118) to get a bound on Π0 Yx0 : with probability ≥ 1 −
exp(−Ω(m)), we have
where we chose δct sufficiently small; and we conclude that
E ti = 0.
m
1 X
ui k2 ≥ 14 εT ≤ exp(−Ω(m)).
Pr k m
i=1
Thus we conclude that E si is bounded as follows:
kE si k2 ≤ kE si0 k2 + kE si1 k2
√
1/2
≤ ε1 + (CCψ2 )4 (48 3) δct
Now set t = 14 εT . Then there exists some (sufficiently large)
constant κ0 > 0 such that
kΠ0 Yx0 k2 ≤ 14 εT + 41 εT = 21 εT .
(119)
Combining equations (99), (103) and (119), we get the
following bound on YT : with probability ≥ 1 − e−Ω(m) ,
kYT k2F ≤ 41 ε2T + 2 · 14 ε2T = 43 ε2T .
(120)
This completes the proof.
IX. P ROOF
OF
T HEOREM V.2
(112)
To prove Theorem V.2, we follow the same strategy used
to prove Theorem V.1, as described in the preceding sections.
Now the measurement vectors ai ∈ Rn are Bernoulli random
and
vectors with erasures, which are a special case of the subgauskti kψ1 = sup kvT ti kψ1
sian random vectors considered previously; hence most of the
v∈S n−1
proof goes through in the same way as before.
T
T
(113)
≤ sup 2β0 kai x0 kψ2 kv Π0 ai kψ2
The present situation is different, however, in that the signal
v∈S n−1
x ∈ Rn is no longer assumed to be µ-flat. This affects Lemma
2
≤ 2(3 + ε1 + εct )(CCψ2 ) ,
VI.2 (injectivity of the sampling operator) and Lemma VI.4
where C is a universal constant, and we used standard prop- (construction of the dual certificate). We claim that, if the ai
erties of sub-exponential random variables. 7
are Bernoulli random vectors with erasure probability p =
This implies that the ui are sub-exponential, with norm
2/3, then we can set µ = 1 in both of these lemmas, so that
both
lemmas now apply to all x (as the µ-flatness condition
kui kψ1 ≤ ksi kψ1 + kti kψ1 + kE si k2
is
trivially
satisfied).
≤ Rct3 CCψ2 + 2(3 + ε1 + εct )(CCψ2 )2 + εcolumn1
In the original proofs of these two lemmas, the µ-flatness
=: K.
conditions are used via equations (15), (51), (54) and (56),
(115) which bound certain 4th moments of the ai , involving proP
jections onto vectors v that are µ-flat. We claim that similar
m
1
We will now bound m
i=1 ui , using a Bernstein-type inequality (Lemma A.2). We get that, for any t ≥ 0 and any bounds hold in our new situation (where the ai are Bernoulli
random vectors with erasure probability p = 2/3, and we
0 < ε < 1,
allow the vector v to be arbitrary).
m
1 X
2
To see this, note that when p = 2/3, the aij have mean
1
ui k2 ≥ 1−ε
t ≤ 2 exp n ln( 3ε )−c min { Kt 2 , Kt }m ,
Pr k m
2
E
a
ij = 0, variance E(aij ) = 1 and fourth moment C4 :=
i=1
1
4
(116) E(aij ) = 1−p = 3, which are the same as the moments of
where c > 0 is a universal constant. We set ε = 12 , substitute the Gaussian distribution. Intuitively, this means that the ai
behave like Gaussian random vectors, which are rotationally
t 7→ t(1 − ε), and recall that m ≥ κ0 n; this gives
symmetric in Rn . Thus, there is no pathological behavior in
m
1 X
1
t2
t
ui k2 ≥ t ≤ 2 exp κ0 ln(6)m−c min { 4K 2 , 2K }m the
. 4th moments of the ai , when one projects onto a direction
Pr k m
v that is not µ-flat.
i=1
(117)
To make this rigorous, we can redo the calculation of
equation (15) as follows: for any v, w ∈ S n−1 ,
7 In particular, note that for any two scalar-valued random variables X and
≤ sup
v∈S n−1
Y,
p 1/p
kXY kψ1 = sup
1
p
E[|XY | ]
≤ sup
1
p
E[|X|2p ]1/2p E[|Y |2p ]1/2p
p≥1
p≥1
≤ 2 sup
p≥1
(114)
√1
2p
2p 1/2p
E[|X| ]
≤ 2kXkψ2 kY kψ2 .
sup
q≥1
√1
2q
2q 1/2q
E[|Y | ]
E[(aTi v)2 (aTi w)2 ]
n
X
X
X
vi2 wi2 +
vi2 wj2 + 2
(vi wi )(vj wj )
= C4
i=1
= (C4 − 3)
n
X
i6=j
i6=j
vi2 wi2
kvk22 kwk22
i=1
T
2
= 1 + 2(v w) .
+
T
2
+ 2(v w)
(121)
14
Note that this bound holds for arbitrary unit vectors v and w,
which need not be µ-flat. This bound serves the same purpose
as equation (15). When we set v = w, this bound serves the
same purpose as equation (51). When we choose v and w
that are orthogonal, this bound serves the same purpose as
equation (54).
Finally, we can redo the calculation of equation (56) as
follows: for any v, w ∈ S n−1 ,
E[(aTi v)3 (aTi w)]
X
X
vj3 wj + 3
vj2 vk wk
= C4
j
= (C4 − 3)
= (C4 − 3)
X
j
X
j6=k
vj3 wj
+3
X
vj2 vk wk
(122)
jk
vj3 wj + 3kvk22 (vT w)
j
T
= 3(v w).
Again, this holds for arbitrary unit vectors v and w, which
need not be µ-flat. When we choose v and w that are
orthogonal, this bound serves the same purpose as equation
(56).
Following these changes, the rest of the proof goes through
as before.
X. D ISCUSSION
Broadly speaking, in this paper we have investigated the
power of subgaussian measurements for phase retrieval. It is
known that for certain natural classes of subgaussian measurements, such as Bernoulli random vectors, phase retrieval is
impossible, in that certain signals can never be distinguished
without unambiguity. We have shown that a large class of
signals, namely those that are “non-peaky” (more precisely,
those that are µ-flat), can still be recovered in this setting. This
µ-flatness condition, where µ can be a constant independent
of the dimension n, is surprisingly weak. We have extended
recent results on stable uniqueness [17] and PhaseLift [11] to
this setting.
In addition, we have shown that for one particular example
of a subgaussian measurement distribution, namely Bernoulli
random vectors with erasures, one does not need any µ-flatness
restriction at all: PhaseLift can recover all vectors in Rn , using
m = O(n) measurements. This is close to the informationtheoretic lower-bound.
Our proof is based on a dual certificate argument, and it is an
interesting question whether similar or better results could be
derived using Mendelson’s small ball method [6]. Indeed, the
small ball method is known to yield stronger results in other
scenarios, especially when the measurements are affected by
noise. However, it is not a priori clear how to incorporate our
non-peakiness condition into the framework of the small ball
method. The small ball method analyzes the measurements
corresponding to a matrix X that lies in the descent cone
of the nuclear norm, rather than the tangent space. Thus X
need not be of the form xxT0 + x0 xT , and it may have rank
greater than two. Hence it is difficult to even precisely capture
the µ-flatness assumption in this framework. In our opinion,
formulating and applying the small ball method for non-peaky
signals is a very interesting direction for follow-up work. With
such an approach, it may even be possible to generalize our
strategy to measurements with non-vanishing mean, such as
0/1 Bernoulli measurements, as has been successfully done in
the context of sparse recovery [23].
It would also be interesting to understand better what
conditions on the measurements ai and the signal x are
necessary and sufficient for phase retrieval. For example, Eldar
and Mendelson’s small-ball and fourth-moment conditions
involve the ai , while our µ-flatness condition involves x.
Both of these seem to be special cases of some more general
conditions that involve the ai and x jointly. In particular,
these conditions seem to involve certain fourth moments of
the ai projected onto directions that depend on x, as seen
in equations (15), (51), (54) and (56). Understanding these
conditions may be helpful for generalizing our results to other
types of measurements, and other classes of signals.
Acknowledgements: Our work on this paper was stimulated
by the Oberwolfach mini-workshop Mathematical Physics
meets Sparse Recovery in April 2014, and in parts performed
during the ICERM semester program High-dimensional Approximation in September 2014. We thank Shahar Mendelson
and the anonymous reviewers of the SampTA 2015 conference
for several helpful comments, and Richard Kueng for his
suggestion to consider random Bernoulli measurements with
erasures. Finally, we thank the reviewers of this journal for a
number of insightful comments and perspectives.
Felix Krahmer’s work on this topic was supported by the
German Science Foundation (DFG) in the context of the
Emmy Noether Junior Research Group KR4512/1-1 (RaSenQuaSI) and Project KR4512/2-1 as well as by the German
Israeli Foundation (Grant no. 1266). Contributions to this work
by NIST, an agency of the US government, are not subject to
US copyright law.
R EFERENCES
[1] R. W. Harrison, “Phase problem in crystallography,” JOSA A, 10(5),
pp.1046-1055, 1993.
[2] R. P. Millane, “Phase retrieval in crystallography and optics,” JOSA A,
7(3), pp.394-411, 1990.
[3] J. C. Dainty and J. R. Fienup, “Phase retrieval and image reconstruction
for astronomy,” in Image Recovery: Theory and Application, H. Stark
(ed.), pp.231-275, Academic Press, 1987.
[4] J.M. Rodenburg, “Ptychography and Related Diffractive Imaging Methods,” Advances in Imaging and Electron Physics 150, pp.87-184 (2008).
[5] Y. Shechtman, Y. C. Eldar, O. Cohen, H. N. Chapman, J. Miao and
M. Segev, “Phase retrieval with application to optical imaging,” IEEE
Signal Proc. Mag., 32(3), pp. 87–109, 2015.
[6] R. Kueng, H. Rauhut and U. Terstiege, “Low rank matrix recovery from
rank one measurements,” Appl. Comput. Harmonic Analysis, to appear.
[7] R. Beinert, G. Plonka-Hoch, “Ambiguities in one-dimensional discrete
phase retrieval from Fourier magnitudes,” J. Fourier Anal. Appl. 21(6),
1169–1198, 2015
[8] A. Fannjiang, “Absolute uniqueness in phase retrieval with random
illumination,” Inverse Prob. 28, 075008, 2012.
[9] E. J. Candes, T. Strohmer and V. Voroninski, “PhaseLift: exact and
stable signal recovery from magnitude measurements via convex programming,” Commun. Pure Appl. Math. 66, pp.1241–1274.
[10] L. Demanet and P. Hand, “Stable optimizationless recovery from phaseless linear measurements,” J. Fourier Anal. Appl., 20(1), pp.199-221,
2014.
15
[11] E. J. Candes and X. Li, “Solving quadratic equations via PhaseLift when
this to get |Nε | ≤ ( 3ε )n . Taking the union bound over all
there are about as many equations as unknowns,” Found. Comput. Math.,
w ∈ Nε , we get that
14(5), pp.1017–1026, 2014
[12] D. Gross, F. Krahmer, and R. Kueng, “A partial derandomization of
N
X
phaselift using spherical designs,” J. Fourier Anal. Appl., 21(2), pp.229–
T
ai xi | ≥ t
Pr
∃w
∈
N
s.t.
|w
ε
266, 2015
i=1
[13] E. J. Candes, X. Li and M. Soltanolkotabi, “Phase retrieval from coded
2
diffraction patterns,” Appl. Comput. Harmonic Analysis, 39(2), pp.277–
≤ e · exp n ln( 3ε ) − K 2ctkak2 . (125)
299, 2015.
2
[14] D. Gross, F. Krahmer and R. Kueng, “Improved recovery guarantees for
Finally, a standard argument (similar to Lemma 5.3 in [21])
phase retrieval from coded diffraction patterns,” Appl. Comput. Harmon.
Anal., 42(1), pp.37–64, 2017
shows that, for any vector v ∈ Rn ,
[15] F. Krahmer and Y.-K. Liu, “Phase retrieval without small-ball probabil1
ity assumptions: Stability and uniqueness,” 2015 Intl. Conf. Sampling
max vT x ≤ kvk2 ≤ 1−ε
max vT x.
(126)
x∈Nε
x∈Nε
Theory Appl., pp.411–414.
[16] F. Krahmer and Y.-K. Liu, “Phase retrieval without small-ball probaP
1
Thus, if k N
bility assumptions: Recovery guarantees for phaselift,” 2015 Intl. Conf.
i=1 ai xi k2 ≥ 1−ε t, then there must exist some
P
Sampling Theory Appl., pp.622–626.
w ∈ Nε such that |wT N
i=1 ai xi | ≥ t. Combining this with
[17] Y. C. Eldar and S. Mendelson, “Phase retrieval: stability and recovery
equation
(125)
completes
the
proof.
guarantees,” Appl. Comput. Harmonic Analysis, 36(3), pp.473–494.
[18] A. Ai, A. Lapanowski, Y. Plan, and R. Vershynin. “One-bit compressed
sensing with non-gaussian measurements,” Lin. Alg. Appl. 441, pp.22–
Lemma A.2 (Vector Bernstein inequality). Let x1 , . . . , xN
239, 2014.
be
independent, centered, sub-exponential random variables
[19] M. Iwen, A. Viswanathan and Y. Wang, “Fast Phase Retrieval from
taking values in Rn , and let K = maxi∈[N ] kxi kψ1 . Fix some
Local Correlation Measurements,” arXiv:1501.02377.
[20] S.
Mendelson,
“Oracle
inequalities
and
the
isoa = (a1 , . . . , aN ) ∈ RN . Let 0 < ε < 1 and t ≥ 0. Then
morphic
method,”
manuscript,
2012,
available
at:
http://maths-people.anu.edu.au/∼mendelso/papers/subgaussian-12-01-2012.pdf.
N
X
[21] R. Vershynin, “Introduction to the non-asymptotic analysis of random
1
ai xi k2 ≥ 1−ε
Pr
k
t
matrices,” chapter 5 in Y. Eldar and G. Kutyniok (eds.), Compressed
i=1
Sensing, Theory and Applications, Cambridge Univ. Press, 2012.
2
[22] D. Suess, R. Kueng and D. Gross, “Certifying linear optical circuits
t
≤ 2 exp n ln( 3ε ) − c min { K 2tkak2 , Kkak
} , (127)
via phaseless estimation techniques,” poster presentation, Conference on
∞
2
Quantum Information Processing (QIP 2016), Jan. 10-15, 2016, Banff,
where c > 0 is a universal constant.
Alberta, Canada.
[23] R. Kueng and P. Jung, “Robust Nonnegative Sparse Recovery and the
Proof: We use the same argument as for Lemma A.1, but
Nullspace Property of 0/1 Measurements”, arXiv preprint 1603.07997.
starting from a scalar Bernstein inequality (Prop. 5.16 in [21]).
A PPENDIX
The following variants of the Hoeffding and Bernstein
inequalities, for sums of independent subgaussian and subexponential random vectors, are straight-forward generalizations
of the standard results.
Lemma A.1 (Vector Hoeffding inequality). Let x1 , . . . , xN be
independent, centered, subgaussian random variables taking
values in Rn , and let K = maxi∈[N ] kxi kψ2 . Fix some a =
(a1 , . . . , aN ) ∈ RN . Let 0 < ε < 1 and t ≥ 0. Then
N
X
ai xi k2 ≥
Pr k
i=1
1
1−ε t
2
≤ e·exp n ln( 3ε )− K 2ctkak2 , (123)
2
Felix Krahmer received his BSc in Mathematics from International University Bremen and his MSc as well as his PhD in Mathematics from New York
University under the supervision of Percy Deift and Sinan Güntürk. He was a
postdoctoral fellow in the group of Holger Rauhut at the University of Bonn,
Germany, from 2009-2012. In 2012 he joined the University of Göttingen as an
assistant professor for mathematical data analysis, where he has been awarded
an Emmy Noether Junior Research Group. Since 2015 he has been a tenure
track assistant professor for optimization and data analysis in the department
of mathematics at the Technical University of Munich. His research interests
include mathematical signal and data processing, applied harmonic analysis,
and random matrix theory.
where c > 0 is a universal constant.
Proof: We use the Hoeffding inequality for scalar random
variables, together with a covering argument over the unit
sphere.
For any vector w ∈ S n−1 , we have kwT xi kψ2 ≤ K. Using
the scalar Hoeffding inequality (Prop. 5.10 in [21]), we get
that
N
X
2
ai xi | ≥ t ≤ e · exp − K 2ctkak2 .
Pr |wT
i=1
(124)
2
Using Lemma 5.2 in [21], we know that there is an ε-net for
S n−1 , with respect to the ℓ2 norm, with cardinality (1 + 2ε )n ;
call this set Nε . Note that when 0 < ε < 1, we can simplify
Yi-Kai Liu received his bachelors degree in Mathematics at Princeton University in 2002. He received his PhD in Computer Science at the University of
California, San Diego, in 2007. He was the recipient of an NSF Mathematical
Sciences Postdoctoral Research Fellowship, and a postdoctoral researcher at
Caltech and at the University of California, Berkeley. In 2011 he joined the
US National Institute of Standards and Technology (NIST). In addition, he is
currently a Fellow at the Joint Center for Quantum Information and Computer
Science (QuICS) at the University of Maryland. His research interests include
quantum computation, cryptography, and machine learning.
| 7cs.IT |
ScreenerNet: Learning Self-Paced Curriculum for Deep Neural Networks
Tae-Hoon Kim 1 Jonghyun Choi 2
arXiv:1801.00904v2 [cs.CV] 20 Feb 2018
Abstract
We propose to learn a curriculum or a syllabus for
supervised learning and deep reinforcement learning with deep neural networks by an attachable
deep neural network, called ScreenerNet. Specifically, we learn a weight for each sample by jointly
training the ScreenerNet and the main network in
an end-to-end self-paced fashion. The ScreenerNet neither has sampling bias nor requires to
remember the past learning history. We show
the networks augmented with the ScreenerNet
achieve early convergence with better accuracy
than the state-of-the-art curricular learning methods in extensive experiments using three popular
vision datasets such as MNIST, CIFAR10 and Pascal VOC2012, and a Cart-pole task using Deep
Q-learning. Moreover, the ScreenerNet can extend other curriculum learning methods such as
Prioritized Experience Replay (PER) for further
accuracy improvement.
1. Introduction
Training a machine learning model with chosen training
samples in a certain order improves the speed of learning
and is called Curriculum Learning (Bengio et al., 2010).
The curriculum learning recently gains much attention due
to the difficulty of training deep models for reinforcement
learning (Andrychowicz et al., 2017; Graves et al., 2017;
Sukhbaatar et al., 2017). However, selecting and ordering
samples is a hard-decision process and significantly lowers
the chance of samples to be selected in the later iterations
and changes solution to which the model converges if the
samples are rejected earlier (sampling bias). In addition, the
decision criteria are mostly defined by a set of hand-crafted
rules such as classification error or confidence of the main
network in the majority of previous work (Bengio et al.,
2015; 2009; Graves et al., 2017). Those hand-crafted rules
require additional rules to handle the sampling bias.
To alleviate the sampling bias, we present a scheme to determine a soft decision by a weight value of every training
1
Intel, Korea 2 Allen Institute for Artificial Intelligence
(AI2), Seattle, WA, USA. Correspondence to: Jonghyun Choi
<jonghyunc@allenai.org>.
Replay
Rules
Training
Samples
✓
Discarded
Samples
Soft Curriculum
by weights
0.1
0.9
Loss
0.8
Main Network
ScreenerNet
Joint Training
Figure 1. ScreenerNet learns the valuation of training samples for
a main network. It regresses the sample into its significance for
training; significant if it is hard for the current performance of the
main network, and non-significant if it is easy.
sample for building a curriculum. Moreover the soft decision is learned in a self-paced manner (Kumar et al., 2010;
Jiang et al., 2015; 2014; Supancic & Ramanan, 2013; Ma
et al., 2017) without requiring to memorize the historic loss
values (Schaul et al., 2016). This is a generalization of the
hard decision by considering all samples at every curriculum
update, thus never ignores any samples at any iteration to
maximize the efficacy of the curricular learning, motivated
from the recent work of efficiently exploring spaces of state,
action, and goal through the learning (Andrychowicz et al.,
2017; Sukhbaatar et al., 2017; Jiang et al., 2015) and is
similar to (Hinton, 2007; Schaul et al., 2016; Jiang et al.,
2015). Moreover, to discover the rules for curriculum at
each training iteration that are beyond our intuition, we learn
the curriculum by an attachable neural network along with
the main network being trained. We call the attachable network as ScreenerNet. The ScreenerNet provides the locally
most accurate weights per the main network being trained
in a self-paced manner.
Unlike the many recent approaches that are mostly applied
to deep reinforcement learning problems (Schaul et al.,
2016; Andrychowicz et al., 2017; Fan et al., 2017), we
empirically show that our ScreenerNet can be applied to
various types of networks including convolutional neural
network for visual recognition as well as deep reinforcement
learning. The ScreenerNet not only expedites the convergence of the main network but also improves the accuracy
at the end of the training. Finally, the ScreenerNet can
extend the stochastic sampling based curriculum learning
approaches for further improvement of final accuracy.
ScreenerNet: Learning Self-Paced Curriculum for Deep Neural Networks
2. Related Work
As a pioneering work of curriculum learning, Hinton (Hinton, 2007) introduced an error-based non-uniform sampling
scheme by importance sampling to significantly speed-up
training of a network on a digit classification using MNIST
dataset. The idea of error-based correction has been widely
used as one of the most popular cues in the curriculum learning literature. Since Bengio et al. (Bengio et al., 2009) coin
the term curriculum learning for the method, number of
approaches (Graves et al., 2017; Lee & Grauman, 2011;
Schaul et al., 2016; Lin et al., 2017; Zhou et al., 2017) have
been proposed to sample, weigh, or sort the training examples to improve accuracy and/or expedite the training.
However, these methods require to design a hand-crafted
rule and/or to run the training procedure twice to obtain
accurate estimate of sample valuation. To address the issues,
Jiang et al. propose to estimate the sample valuation at the
same time of training a learning model (Jiang et al., 2015)
by adding a term in the objective function. This is an extension of Kumar et al.’s work on Self-Paced learning (Kumar
et al., 2010) in the curriculum learning literature.
Recently, Schaul et al. (Schaul et al., 2016) proposed
a sampling scheme to increase the replay probability
of training samples that have a high expected learning
progress determined by an optimization loss. They also
proposed a weighted importance sampling method to address the bias of sampling, which is similar to our weighting
scheme. In stochastic gradient-based optimization perspective, Loshchilov and Hutter (Loshchilov & Hutter, 2016)
also proposed to sample mini-batches non-uniformly, called
AdaDelta (Zeiler, 2012) and Adam (Kingma & Ba, 2014).
Most recently, Graves et al. (Graves et al., 2017) proposed
an automatic curriculum learning method for LSTM for
the NLP application. They define a stochastic syllabus by
a non-stationary multi-armed bandit algorithm of getting
a reward signal from each training sample. They defined
mappings from the rates of increase in prediction accuracy
and network complexity to the reward signal.
Koh and Liang (Koh & Liang, 2017) adopted influence
functions from robust statistics to measure the effect of parameter changes or perturbations of the training data such as
pixel values. It was applied to debugging models, detecting
dataset error, and training-set attack. Although the direct
application to curriculum learning is not presented, the authors show a way of predicting the significance of training
samples, which can be useful for learning curriculum.
Similar to our idea of attaching another network to the main
one, self-play between the policy and a task-setting was
proposed by Sukhbaatar et al. (Sukhbaatar et al., 2017). The
task-setter tries to find the simplest tasks that the policy
cannot complete, which makes learning for the policy easy
because the selected task is likely to be only just beyond
the capability of the policy. Similarly, Andrychowicz et
al. (Andrychowicz et al., 2017) recently proposed to learn a
curriculum but with a hindsight replay. They use unshaped
reward signals even when they are sparse and binary, without requiring domain knowledge. These recent approaches
address the sparsity and complexity of solution spaces in
deep reinforcement learning, which is not a main issue for
the supervised learning.
Very recently, Zhou et al. (Zhou et al., 2017) proposed
an adaptive feeding method that classifies an input sample
as easy or hard one in order to forward the input to an
appropriate one of a fast but less accurate neural network and
an accurate but slow network for object detection. However,
they only showed the speed up the inference preserving the
accuracy of the main classifier.
In addition, Jiang et al. (Jiang et al., 2017) propose an attachable network to regularize the network not to overfit to
the data with corrupted labels, called MentorNet. The MentorNet uses feature embedding of the sample by the main
network, referred as StudentNet, and compute a weight for
each sample. It is pretrained using a few steps of loss history
of the StudentNet with LSTM and is attached to the StudentNet while its weights are fixed due to the risk of trivial
solution. Whereas, focusing on faster and better convergence of the main network, our ScreenerNet is an attachable
network that is jointly trained with the main network in an
end-to-end fashion with no pretraining required and with
less chance of falling into a trivial solution.
Unlike the previous approaches, our ScreenerNet has two
benefits: First, it includes all samples to update weight even
though the samples have weight values close to zero. This
benefit is particularly significant in stochastic sampling approaches, if an important sample is assigned a low sampling
priority at the early stage of training, it may not be likely to
be picked again and may not be significantly used for the
training until other samples have low sampling priority as
well. Second, it can learn a direct mapping of a training
sample to its significance, even if the training sample is
unseen, unlike the other memory-based methods.
To the best of our knowledge, ScreenerNet is the first approach of an attachable deep weight regressor to both speed
up the training and improve the accuracy without modifying
the main network. Moreover, it can extend the stochastic
sampling curriculum learning approaches to improve further
by taking the benefits of both methods.
3. Approach
We formulate the problem of building a curriculum by learning the importance of each sample for training the main
network by a scalar weight in a joint learning framework.
ScreenerNet: Learning Self-Paced Curriculum for Deep Neural Networks
Specifically, we define the online importance of a training
sample x as a random variable wx . Since we want to train
the main network better in speed and accuracy, the objective of the curriculum learner maximizes the likelihood by
sample weight wx given the error E between the prediction
of the main network for x and its target. The maximum
cx can be written as follows:
likelihood estimator of wx , w
cx = arg max P (E|wx , Wc )
w
wx
0.8
0.6
wx
= arg max
1
P (wx |E, Wc )P (E|Wc )
,
P (wx |Wc )
(1)
0.4
0.2
where Wc is a parameter of the main network. Since E is a
function of Wc , Equation 1 can be reduced to
0
0
1
0.5
Figure 2. Loss function0.5LS (ex , wx ) with M = 1 without the L1
1 our experiments.
0 for
regularizer. We bound wxwto be in (0, 1)
ex
x
P (wx |E)P (E|Wc )
arg max
.
wx
P (wx |Wc )
(2)
to increase weight wx when error of the main network ex is
high and promotes to decrease weight when the error is low.
The ScreenerNet is a neural network to optimize the objective.
Note that we bound the weight value by the ScreenerNet
to be in (0, 1) by a Sigmoid layer at the end of the ScreenerNet, since its multiplication to the gradient without the
bound may cause overshooting or undershooting of the main
network (Hinton, 2007).
3.1. ScreenerNet
Valuating the exact significance of each training sample to
maximize the final accuracy of the main network is computationally intractable (Graves et al., 2017). The influence
functions in (Koh & Liang, 2017) could be a potential solution to estimate the final accuracy with less computational
burden but still requires significant computational cost at the
initialization of every iteration of training the main network.
Instead of estimating the final accuracy, we propose to simplify the problem of sample-wise significance valuation to
a local optimal policy that predicts the weights of training
samples at the current iteration of the training.
Let wx be a weight of training sample, x, predicted by
ScreenerNet, S. Let LF (F(x), tx ) be an objective function
for the main network F(·) to compute an error between
F(x) and its target label tx . We define an objective function
that ScreenerNet minimizes, LS (·), as follows:
X
x∈X
X
(1 − wx )2 ex + wx2 max(M − ex , 0) + α
kpk1 , (3)
p∈WS
where wx = S(x), ex = LF (F(x), tx ), WS is parameters
of the ScreenerNet S, and α is a balancing hyper-parameter
for the regularizer of the ScreenerNet. X is a set of training
images and M is a margin hyper-parameter. We plot LS
with M = 1 except the L1 regularizer in Figure 2.
As shown in the Figure 2, the objective function is a nonnegative saddle like function with minima at (wx , ex ) =
(0, 0) or (1, 1) and with maxima at (wx , ex ) = (0, 1) or
(1, 0). Thus, the LS promotes the sample weight to be high
when the error (loss) of the sample by the main network is
high and vice versa. Therefore, the ScreenerNet encourages
Optimization. Optimizing the loss of the ScreenerNet augmented network is not trivial as the loss is non convex. Thus,
we employ the block-coordinate descent to optimize the network and depict the gradient path of the subset of variables
in Figure 3. The numbers in the blue boxes in the figure
denotes the order of updating the gradient.
F
ŷ
Main Network
Back-propagation
5
1
3
LF (ŷ, 1)
LF (ŷ, wx )
x
Input Image
S
ScreenerNet
2
ex
4
eweighted
x
LS
wx
7
6
e wx
Back-propagation
Figure 3. Optimization path of a ScreenerNet attached network.
The numbers in blue boxes show the order of gradient update.
By the block coordinate descent, at each iteration of the
epochs in the training phase, we predict the weight, wx of
the training sample x to update the main network F using
the error, ex from x and the weight wx . Then, we update
ScreenerNet using the error, ex . The training procedure
of ScreenerNet augmented deep network is summarized as
Algorithm 1.
Architecture of ScreenerNet. A larger ScreenerNet than
the main network would have larger capacity than the main
network, then it may require more sample to reliably pre-
ScreenerNet: Learning Self-Paced Curriculum for Deep Neural Networks
Algorithm 1 Training a network with ScreenerNet
1: Given: training samples (X, tX ), main network F,
ScreenerNet S
2: Initialize F, S
3: for each iteration of the training do
4:
x ← Sample(X).
. sample a mini-batch from X
5:
wx ← S(x).
. predict an weight of a sample
6:
ŷ ← F(x).
. prediction from the main network
7:
eweighted
←
L
(ŷ,
wx ). . compute a weighted error
F
x
8:
ex ← LF (ŷ, 1).
. compute an error of the sample
9:
UpdateNetwork(F, ew
. train main network
x ).
10:
UpdateNetwork(S, ex ).
. train ScreenerNet
11: end for
dict the significance of the samples. A simple ScreenerNet
would not predict the significance well and likely predicts
it like a uniform distribution. We empirically found that
ScreenerNet whose architecture is similar to that of the
main network or slightly simpler performs the best. Further,
we also try weight sharing with the main network (refer to
Section 4.6 for details).
4. Experiments
4.1. Datasets
We have evaluated our algorithm with three popular
vision datasets; MNIST (Lecun & Cortes, 1992), CIFAR10 (Krizhevsky, 2012) and Pascal VOC 2012 (Everingham et al., 2012), and a Cart-pole example using the
deep Q-learning (van Hasselt et al., 2016; Mnih et al., 2013),
which is one of the most popular tasks in deep reinforcement
learning.
MNIST dataset has 60, 000 images with size of 28×28 in
training and 10, 000 images in testing set. CIFAR10 has
50, 000 images with size of 32×32 in training and 10, 000
images in testing set. Pascal VOC 2012 has 5, 717 images
in training and 5, 823 images in validation set. Both MNIST
and CIFAR10 are widely used for many neural network
training benchmarks. Pascal VOC 2012 is widely used
as one of the most popular visual recognition benchmarks
along with the ImageNet.
For the Cart-pole experiment, we use Cart-pole-v0 in OpenAI Gym (Brockman et al., 2016), which gets 4-dimensional
input of the state of the cart and pole, and has 2 discrete
actions to move left or right.
3.2. Extension with Stochastic Sampling
Since the ScreenerNet is a weight regression network, it
can extend the stochastic sampling approaches such as Prioritized Expereience Replay (PER) (Schaul et al., 2016).
The extension has two benefits; first, it may further improve
the convergence speed and the final accuracy than the single deployment of either methods. Second, it reduces the
computation for the ScreenerNet.
For instance, if we combine the ScreenerNet with the PER.
The PER determines the probability of a training sample to
be selected by
P (x) = P
pα
x
x̃∈X
pα
x̃
,
(4)
where px > 0 is the priority of each sample x, and α
controls how much prioritization is used. When α = 0,
it is equivalent to the uniform sampling. The priority is
defined as:
px = |ex | + ,
(5)
where is a very small constant to prevent from assigning
zero priority to x. To extend the PER or any other sampling
based methods with the ScreenerNet, we can simply predict
weights of x that the PER or other sampling based methods
select. (Refer to the Sec. 4.4)
4.2. Experimental Set-up
We sketch the neural network architecture of the main network and the ScreenerNet used in our experiments on each
dataset and their important hyper-parameters. The detailed
architectures and set-ups are summarized in the supplementary material.
MNIST. We use a simple CNN architecture (consisting of
two convolutional layers, two fully connected layers) for
main network and use a simpler ScreenerNet than the main
network (refer to the dicussion in Sec.4.6). We use minibatches of 64 samples, α = 0.01, and the margin M = 1.0,
respectively.
CIFAR10. We have three architectures for main network;
first, a simple CNN adopted from the PyTorch tutorial, second, its variant with more convolutional layers and last, the
VGG-11 (denoted as ‘small’, ‘mid’ and ‘VGG11’ in Figure 4, respectively). We show the three set-ups to compare
the gain by the ScreenerNet when complexity of the main
network varies. For the loss fuction of the main network,
we used the negative log-likelihood. We use mini-batches of
64 samples and α = 0.01 in Equation 3 for all experiments.
For the margin M , we linearly annealed it from 20 to 5
through 50 epochs, since the error of the main network was
large at the early training iterations.
Pascal VOC 2012. We fine-tune a pre-trained VGG-19 (Simonyan & Zisserman, 2015) by replacing its final layer with
a few layers for both main network and ScreenerNet. We
ScreenerNet: Learning Self-Paced Curriculum for Deep Neural Networks
98.0%
20
40
60
Epoch
80
90.0%
70.0%
60.0%
Baseline: small
Baseline: mid
Baseline: VGG11
50.0%
Baseline
ScreenerNet
100
40.0%
190
185
Accuracy
Accuracy
Accuracy
98.5%
Cartpole
VOC2012 classification
80.0%
99.0%
97.5%
CIFAR10 classification
90.0%
10
20
ScreenerNet: small
ScreenerNet: mid
ScreenerNet: VGG11
30
40
Epoch
Average Reward
MNIST classification
99.5%
80.0%
70.0%
Baseline (train)
Baseline (test)
60.0%
50.0%
10
20
30
Epoch
40
50
175
170
165
ScreenerNet (train)
ScreenerNet (test)
50
180
Baseline (Double DQN)
DDQN + ScreenerNet
160
60
10000
20000
30000
40000
50000
Steps
(a) MNIST
(b) CIFAR10
(c) Pascal VOC 2012
(d) Cart-pole
Figure 4. Comparison of the test accuracy (except (c) showing both the training and test accuracy) of main network (Blue) and the
ScreenerNet augmented network (Red)
used a binary cross entropy loss for multi-label one-vs-all
classification, mini-batches of 32 samples, the margin of
M = 1.0, and α = 0.001. For the evaluation metric, we use
the mean Average Precision (mAP) across all the 20 classes.
We apply data augmentations including random horizontal
flip, rotation and cropping.
In the results for the MINST dataset shown in Figure 4(a), both the baseline and ScreenerNet augmented networks
converge quickly and the learning curves at the early stage
of training are similar. But ScreenerNet improves the final
accuracy though the gain is marginal. We discuss the results
in the second next paragraph.
Cart-pole. We designed a simple CNN for a deep Qnetwork of main network and use the same architecture
for ScreenerNet. We used α = 0.001 and the margin of
M = 2.0. We limit the maximum reward from an episode
to be 200, at which the episode finished.
In the results for the CIFAR10 dataset shown in Figure 4(b), with the simple main network architecture (shown in
tinted solid line), we observed that ScreenerNet yielded
a significant improvement in the learning speed and final
accuracy. With the mid-size main network, the ScreenerNet
shows decent improvement. But with the pretty complicated
main network (VGG11), the ScreenerNet either improves
marginally or performs on par.
We also compare the ScreenerNet with the state-of-the-art
stochastic sampling based curriculum learning method, Prioritized Experience Replay (PER) (Schaul et al., 2016).
For the PER, we used importance-sampling weights wi =
β
1
1
, where β is linearly annealed from 0.4 to 1.0
N · P (i)
until 40, 000 steps then is fixed to 1.0. We set the discount
parameter γ to be 0.99 and the replay memory to be a sliding window memory of size 50, 000. The PER processes
mini-batches of 32 samples (for visual recognition task) or
32 transitions (deep Q-learning) sampled from the memory. In Cart-pole task, a mini-batch is updated at every 32
transitions entering the memory. We use the epsilon greedy
exploration that linearly anneals epsilon value from 1.0 to
0.1 in the first 1, 000 steps. For the evaluation at every
5, 000 step, we disabled the exploration to prevent random
behavior. The evaluation is repeated 100 times to get the
final average reward. The main network for the Double
DQN is interpolated with a smoothness parameter value of
0.01 at every training iteration.
Our evaluation metric is two folds; speed of convergence
in training and the final classification accuracy for the visual recognition tasks (MNIST, CIFAR10, and Pascal VOC
2012) and the average reward per episode for the Cart-pole.
4.3. Faster and Better Convergence by ScreenerNet
We compare the test accuracy curves of the baseline network (main network only) and the ScreenerNet augmented
network (detnoted as ‘ScreenerNet’) in Figure 4 for all four
tasks.
With the result for the MNIST and CIFAR10 dataset, we
argue that when the main network is large enough, curriculum does not matter for the performance as they can simply
memorize all the details of all samples. Even though we
omit the training curve for visual clarity, the VGG11 quickly
overfits to the training set of the CIFAR10 dataset while others (small and mid) less overfit which is a usual scenario
for recognition; a less capable main network for large scale
datasets. Unless the main network is easily overfitting to the
dataset, the ScreenerNet shows decent improvement.
In the results for the Pascal VOC 2012 dataset shown in Figure 4-(c), ScreenerNet helps the main network to improve
the learning speed at early epochs even though VGG-19 is
pre-trained with a much larger dataset (ImageNet). ScreenerNet also improves the accuracy when the networks are
close to convergence. Note that even though we apply extensive data augmentations for training, the ScreenerNet
significantly improves the learning curve with training set
and improve the test accuarcy as well. We believe that
the ScreenerNet helps better training rather than overfitting.
(refer to discussion about Parameter Sharing in Sec.4.6)
In the results for the Cart-pole task, the average reward
obtained at every 5, 000 training step is shown in Figure 4(d). The ScreenerNet augmented network begins with higher
average reward than the baseline Double DQN (DDQN)
and clearly shows the higher overall gain comparing to the
baseline.
ScreenerNet: Learning Self-Paced Curriculum for Deep Neural Networks
VOC2012 classification
98.5%
20
40
60
80
60.0%
50.0%
Baseline
ScreenerNet
PER
98.0%
Accuracy
99.0%
100
40.0%
Baseline
ScreenerNet
PER
10
20
Epoch
185
30
40
80.0%
70.0%
Baseline
ScreenerNet
PER
60.0%
50
Epoch
(a) MNIST
Cartpole
190
90.0%
70.0%
Accuracy
99.5%
Accuracy
CIFAR10 classification
80.0%
Average Reward
MNIST classification
100.0%
50.0%
10
20
30
40
50
Epoch
(b) CIFAR10
(c) Pascal VOC 2012
180
175
170
Baseline (Double DQN)
DDQN + PER
DDQN + ScreenerNet
165
160
60
10000
20000
30000
40000
50000
Cartpole
(d) Cart-pole
Figure 5. Comparison with PER (Green), Baseline (the main network only, Blue) and ScreenerNet augmented Network (Red)
190
Average Reward
185
180
175
Baseline (Double DQN)
DDQN + PER
DDQN + ScreenerNet
DDQN + PER + ScreenerNet
DDQN + ScreenerNet Sampling
170
165
160
10000
20000
30000
40000
50000
Steps
Figure 6. Learning curves for the cart-pole problem by average reward. Extension of combination of ScreenerNet and PER (Orange)
and stochastic sampling using ScreenerNet (Purple)
Compare with Prioritized Experience Replay (PER).
While the PER selects samples by a hard decision, ScreenerNet selects samples by a soft decision by a weight. They
could be both comparable and able to be combined together,
which we will discuss further in Sec. 4.4. To see the benefit
of ScreenerNet over the PER (Schaul et al., 2016), we compare ScreenerNet with PER for all four datasets in Figure 5.
We observe that ScreenerNet shows better learning of the
main network than the PER in all four evaluations.
4.4. Extension with Stochastic Sampling Methods for
Deep Q-Learning
Since the stochastic sampling methods determine whether
the samples to be included in the training or not by a defined
probability, we can extend them by the ScreenerNet to investigate the synergistic effect of both methods. Particularly, we
can extend the ScreenerNet in two ways; first, ScreenerNet
provides weights on the selected samples by the stochastic
sampling methods. Second, the sampling selects the sample
by the ScreenerNet’s weight. We evaluate both scenarios
with the Prioritized Experience Replay (PER) as a choice of
stochastic sampling methods for the deep Q-learning task
using the cart-pole task in Figure 6.
4.4.1. S CREENER N ET W EIGHS S AMPLES S ELECTED BY
THE P RIORITIZED E XPERIENCE R EPLAY (PER)
For the first scenario, we weigh the samples by the ScreenerNet on the samples selected by the PER (denoted as Screen-
erNet + PER (orange solid line) in Figure 6). This extension outperforms the PER (green dotted line) but performs
worse than the ScreenerNet (red dotted line); since ScreenerNet weighs already selected samples, ScreenerNet+PER
performance could be bounded by the performance of the
ScreenerNet.
Interestingly, the ScreenerNet(red dotted line), the PER
(green dotted line) and ScreenerNet+PER (orange solid line)
show similar learning progress; it begins with high accuracy
but falls then gradually increases. Since they all have a
sample weighting scheme, we believe that the methods with
weighting schemes quickly learn the easy solutions in the
beginning of learning (e.g., move the pole slightly from the
initial position) then learn more difficult solutions thus the
rewards drop in the early stage of learning and increase.
4.4.2. PER WITH S CREENER N ET P ROBABILITY
In the PER, the probability that determines the priority of
samples to be selected is defined by the error of the main
network. For the second scenario of the ScreenerNet extension, we replace the probablity px in Equation 5 with the
output of ScreenerNet (px = S(x) + ) and setting α = 1
in Equation 4.
The result is shown by the purple solid line
(DDQN+ScreenerNet Sampling) in Figure 6. Comparing the PER (green dotted line), the PER outperforms
the ScreenerNet Sampling in the beginning of training
where ScreenerNet is not trained enough but eventually the
ScreenerNet probability outperforms the direct error value.
4.5. Qualitative Analysis
Error Analysis with MNIST. To better understand the effect of the ScreenerNet, we investigate failure cases of
ScreenerNet augmented network (ScreenerNet) and the
main network alone (baseline) evaluated for the MNIST
dataset. We present confusion matrices of the baseline and
ScreenerNet in Figure 7 only with failed examples for visual
clarity. We also show qualtitative comparison that the failure cases that either only the ScreenerNet or the baseline
classifies incorrectly in Figure 8.
It is observed that the widely spread confusions of the base-
ScreenerNet: Learning Self-Paced Curriculum for Deep Neural Networks
0
1
0
0
0
0
0
2
0
0
3
0
0
1
0
0
4
0
1
2
4
0
0
0
0
0
0
1
0
1
5
1
0
0
5
0
0
2
1
6
4
2
0
0
1
4
0
7
0
3
9
0
1
0
0
8
3
9
2
2
1
1
2
0
1
0
0
0
0
1
0
0
1
0
0
1
0
0
0
4
0
0
7
7
2
0
0
0
0
0
0
1
3
0
0
6
0
6
3
0
0
0
0
0
2
0
3
2
0
4
5
4
0
0
1
0
0
0
2
0
1
8
0
0
4
5
1
0
0
6
0
0
2
1
0
0
0
0
0
6
2
0
0
1
1
0
0
5
0
3
3
2
0
0
3
7
0
2
6
0
0
0
0
0
0
3
2
2
8
2
0
2
1
0
0
0
1
0
2
9
1
1
0
2
0
2
1
2
0
4
6
2
0
5
2
0
9
8
7
6
5
4
3
2
1
0
1.0
Ground Truth
5
5
3
5
0.6
4
0.4
1
0
2
4
0
4
2
0
0.2
0
Predicted label
Predicted label
0.0
Figure 7. Confusion matrix only with the failed examples for visual clarity. Each number in the cell shows the number of misclassification. Note that the diagonal of the matrix is zero because
success cases are excluded for the visualization. (Left) by the
network without ScreenerNet (Right) by ScreenerNet augmented
network.
995 779 887 660 771 993 006 880 553 998
622 866 944 599 355 311 599
994 995 335 991 665 993 550 774 993 997
711 655 944 866 899 866 944
221 665 335 994 994 990 884 006 772 449
733 055 944 644 355 244 944
993 997 006 002 772 885 772 114 885 556
977 599 599 100 722
(a) Baseline only fails
5
0.8
wx
2
0
1
0
8
9
1
0
8
0
0
7
0
2
6
1
0
5
3
0
4
0
1
3
0
0
8
0
2
0
1
Confusion matrix of ScreenerNet
0
1
2
0
True label
0
0
9
0
0
0
True label
Confusion matrix of Baseline
0
(b) ScreenerNet only fails
Figure 8. Comparison of exclusive failure examples. (a) baseline
only fails but ScreenerNet succeeds and (b) ScreenerNet only fails
but baseline succeeds. The three numbers under each sample image
denote classification from ScreenerNet (in red), ground-truth (in
green), and the baseline (in blue), respectively from left to right.
line are reduced overall by ScreenerNet in Figure 7. But it
also has a few new failures like mis-classification of 4 into 9.
Note that the failure examples of 4 into 9 (Figure 8-(b)’s 944
examples) are extremely challenging. Instead, ScreenerNet
increases the precision of recognizing 4, being more strict
with classifying 1, 7, 8, and 9 into 4. Similarly ScreenerNet
increases recall for 8 at the expense of reduced precision.
Easy and Difficult Training Samples Selected by ScreenerNet. As the learning of ScreenerNet augmented neural
network proceeds, difficult samples should receive high attention in the training procedure by high weight, while easy
samples should receive low attention by low weight. We
plot the tracked weight of some easy and difficult samples
as the learning proceeds in Figure 9-(a) and also present the
samples with highest and lowest weights at the end of the
training in Figure 9-(b) and (c).
The samples with high weights are visually difficult to distinguish, while the ones with low weights are visually distinctive to the other class thus training with these in the later
epoch would not add much value to improve the accuracy.
20
40
60
80
100
Epoch
(a) Weight progress of MNIST
5
3
3
3
3
3
3
2
4
6
0.649 0.544 0.542 0.537 0.522
0.010 0.012 0.012 0.012 0.012
(b) Ends with highest weights
(c) Ends with lowest weights
Figure 9. Weight progress curves and their corresponding training
images in MNIST. In (b) and (c) Green number denotes the ground
truth label and the number below denotes the final weight value.
4.6. Choice of Architecture of ScreenerNet
Complexity of the ScreenerNet. In our experiments, we
use ScreenerNet architectures that are identical to the main
network or slightly simpler than that. An intution behind this
choice is that the ScreenerNet architecture resembling the
main network is expected to have output distribution similar
to that of the main network. Thus, ScreenerNet might serve
as a conjugate prior for the maximum a posteriori estimator
(ScreenerNet augmented network) that outputs the same
distribution of the main network itself.
Empirically, we compare the learning curves of the variously
complicated ScreenerNet architectures in Figure 10-(a); the
simple (Simple SNet), the complex (Complex SNet) and the
one similar to the main network (ScreenerNet). The simple
architecture consists of two convolutional, two maxpool
with ReLU activation and one fully connected layers with a
sigmoid activation. The complex architecture consists of the
pre-trained VGG-19 followed by two fully connected with a
sigmoid activation. We use CIFAR10 dataset for evaluation.
The Simple SNet performs worse than the baseline (the
main network only); it begins with the accuracy as high as
the ScreenerNet but ends up with the lower accuracy than
the baseline. The Complex SNet performs even worse from
the beginning. We argue that the Simple SNet performs
well at the early stage of the training because it is simple
enough to quickly model the behavior of the main network.
In the later epochs, however, the Simple SNet is not strong
enough to learn details of the behavior. Whereas the Complex SNet is expected to be strong to learn the behavior,
but it is not easy for it to quickly catch up predicting its
behavior of the main network and eventually falling into
ScreenerNet: Learning Self-Paced Curriculum for Deep Neural Networks
of mini-batches to update ScreenerNet while weight prediction is done for all training samples. For example, choosing
one of every two samples to update ScreenerNet reduced
around 10−20% of the computation time in our experiment,
with the acceptable degradation of the accuracy. Alternative
approach is to combine PER to ScreenerNet at the expense
of accuracy as discussed in Sec. 4.4.
CIFAR10 classification
80.0%
70.0%
Accuracy
60.0%
50.0%
40.0%
30.0%
Baseline (train)
Simple SNet (train)
Baseline (test)
Complex SNet (train)
ScreenerNet (train)
Simple SNet (test)
ScreenerNet (test)
20.0%
10
20
Complex SNet (test)
30
40
5. Conclusion
50
Epoch
We propose to estimate the significance of the training samples for effective curriculum learning by augmenting a deep
neural network, called ScreenerNet, to the main network and
jointly train them. We demonstrated that the ScreenerNet
helps training deep neural networks both by fast and better
convergence in various tasks including visual classification
and deep reinforcement learning such as deep Q-learning.
Moreover, the ScreenerNet not only outperforms the stateof-the-art sampling based curriculum learning method such
as the Prioritized Experience Replay (PER) (Figure 5) but
also can extend it for further improvement (Figure 6).
(a) Different complexity of ScreenerNet’s.
CIFAR10 classification for shared parameters
70.0%
Accuracy
60.0%
Baseline (test)
50.0%
ScreenerNet (test)
Parameter sharing 1 (train)
Parameter sharing 1 (test)
40.0%
Parameter sharing 2 (train)
Parameter sharing 2 (test)
10
20
30
40
50
Epoch
(b) Parameter of ScreenerNet is shared with the main network.
Figure 10. Accuracies of different ScreenerNet configurations on
CIFAR10 dataset (small architecture).
a bad local minimum. More iterations of optimization or
higher learning rate were helpful but marginally according
to our experiments. The ScreenerNet architecture similar to
the main network empirically performs the best.
Parameter Sharing. We also tried the network parameter
sharing between common layers of the main network and
the ScreenerNet using CIFAR10 dataset. There are two
parameter sharing scenarios; first, ScreenerNet does not
change the parameter of shared layers but fine-tunes the last
FC 1 layer. Second, ScreenerNet also updates the parameter
of the shared layers.
As illustrated in Figure 10-(b), parameter sharing decreases
the accuracy. It implies that as the ScreenerNet does not
learn the same objective that the main network learns, rather
it learns the behavior of the main network; even the low
level signal to the ScreenerNet is different from that of the
main network. We argue that the ScreenerNet helps the
main network to learn faster and better not helps it overfit.
4.7. Computational Cost
ScreenerNet needs to update its weights during training at
the cost of additional computation and memory. Detailed
cost comparison can be found in the supplementary material.
To alleviate the computational burden, one can use a subset
The learning objective of ScreenerNet is not the same as
that of the main network as other work is trying to model.
Instead, the ScreenerNet estimates the probability that the
main network will correctly classify the given sample or
not. The architecture of the ScreenerNet is slightly simpler
than the target network and thus can be trained ahead of it
in terms of training maturity, which we empirically found
to perform the best.
Since the ScreenerNet is not a memory-based model, it can
also be considered as an error estimator of the current state
of the main network for the new sample. Thus, the ScreenerNet can estimate the sample confidence at inference time,
which is particularly useful for the real world reinforcement
learning system, similarly to the adaptive classifier proposed
by (Zhou et al., 2017).
Limitation. Since the objective function of the ScreenerNet
views training samples with large errors as significant ones
to train the main network, it possibly boosts weight values of
mislabeled training samples which may perturb a decision
boundary of the main network.
Future work. We can extend the ScreenerNet to be progressively expanding its complexity from a simple network
to a complex one by adding new layers as the training progresses, similarly to (Wang et al., 2017). Since the newly
added layers to the ScreenerNet might lead the learning to
be unstable, we can use a momentum in updating the weight
as wx = λSold (x) + (1 − λ)Snew (x), where Sold and Snew
are respectively previous and current networks, and λ is a
hyper-parameter in [0, 1].
ScreenerNet: Learning Self-Paced Curriculum for Deep Neural Networks
References
Andrychowicz, Marcin, Wolski, Filip, Ray, Alex, Schneider,
Jonas, Fong, Rachel, Welinder, Peter, McGrew, Bob, Tobin, Josh, Abbeel, Pieter, and Zaremba, Wojciech. Hindsight Experience Replay . CoRR, abs/1707.01495, 2017.
Bengio, S., Weston, J., and Grangier, D. Label Embedding
Trees for Large Multi-Class Tasks. In NIPS, 2010.
Bengio, S., Vinyals, O., Jaitly, N., and Shazeer, N. Scheduled sampling for sequence prediction with recurrent neural networks. In NIPS, 2015.
Bengio, Y., Louradour, Jérôme, Collobert, R., and Weston,
J. Curriculum Learning. In ICML, 2009.
Brockman, Greg, Cheung, Vicki, Pettersson, Ludwig,
Schneider, Jonas, Schulman, John, Tang, Jie, and
Zaremba, Wojciech. OpenAI Gym, 2016.
Kumar, M. Pawan, Packer, B., and Koller, D. Self-Paced
Learning for Latent Variable Models. In NIPS, 2010.
Lecun, Y. and Cortes, C. The MNIST Database of Handwritten Digits. NYU, 1992.
Lee, Y. J. and Grauman, K. Learning the Easy Things First:
Self-Paced Visual Category Discovery. In CVPR, 2011.
Lin, Tsung-Yi, Goyal, P., Girshick, R., He, K., and Dollar, P.
Focal Loss for Dense Object Detection. In CVPR, 2017.
Loshchilov, I. and Hutter, F. Online Batch Selection for
Faster Training of Neural Networks. In Proceedings of
ICLR Workshop, 2016.
Ma, F., Meng, D., Xie, Q., Li, Z., and Dong, X. Self-Paced
Co-training. In ICML, 2017.
Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A.,
Antonoglou, I., Wierstra, D., and Riedmiller, M. Playing
Atari with Deep Reinforcement Learning. In NIPS Deep
Learning Workshop, 2013.
Everingham, M., Van Gool, L., Williams, C. K. I., Winn, J.,
and Zisserman, A. The PASCAL Visual Object Classes
Challenge 2012 (VOC2012) Results. http://www.pascalnetwork.org/challenges/VOC/voc2012/workshop/index.html, Schaul, Tom, Quan, John, Antonoglou, Ioannis, and Silver,
David. Prioritized experience replay. In Proceedings of
2012.
ICLR, 2016.
Fan, Y., Tian, F., Qin, T., Bian, J., and Liu, Tie-Yan. LearnSimonyan, K. and Zisserman, A. Very Deep Convolutional
ing What Data to Learn. arXiv preprint:1702.08635,
Networks for Large-Scale Image Recognition. In ICLR,
2017.
2015.
Graves, Alex, Bellemare, Marc G., Menick, Jacob, Munos,
Sukhbaatar, Sainbayar, Kostrikov, Ilya, Szlam, Arthur, and
Rémi, and Kavukcuoglu, Koray. Automated Curriculum
Fergus, Rob. Intrinsic Motivation and Automatic CurricLearning for Neural Networks. In Proceedings of ICML,
ula via Asymmetric Self-Play. CoRR, abs/1703.05407,
2017.
2017.
Hinton, G. E. To Recognize Shapes, First Learn to Generate
Supancic, J. S. and Ramanan, D. Self-paced learning for
Images. Progress in brain research, 165:535–547, 2007.
long-term tracking. In CVPR, 2013.
Jiang, L., Meng, D., Yu, Shoou-I, Lan, Z., Shan, S., and
van Hasselt, Hado, Guez, Arthur, and Silver, David. Deep
Hauptmann, A. G. Self-Paced Learning with Diversity.
Reinforcement Learning with Double Qlearning. In AAAI,
In NIPS, 2014.
2016.
Jiang, L., Meng, D., Zhao, Q., Shan, S., and Hauptmann,
Wang, Yu-Xiong, Ramanan, D., and Hebert, M. Growing
A. G. Self-Paced Curriculum Learning. In AAAI, 2015.
a Brain: Fine-Tuning by Increasing Model Capacity. In
CVPR,
2017.
Jiang, L., Zhou, Z., Leung, T., Li, Li-Jia, and Fei-Fei, L.
MentorNet: Regularizing Very Deep Neural Networks on
Corrupted Labels. arXiv preprint:1712.05055, 2017.
Zeiler, Matthew D. Adadelta: An adaptive learning rate
method. arXiv preprint:1212.5701, 2012.
Kingma, D. P. and Ba, J. Adam: A Method for Stochastic
Optimization. CoRR, abs/1412.6980, 2014.
Zhou, Hong-Yu, Gao, Bin-Bin, and Wu, Jianxin. Adaptive feeding: Achieving fast and accurate detections by
adaptively combining object detectors. In Proceedings of
ICCV, 2017.
Koh, P. W. and Liang, P. Understanding Black-box Predictions via Influence Functions. arXiv preprint:1703.04730,
2017.
Krizhevsky, Alex. Learning multiple layers of features from
tiny images. University of Toronto, 2012.
ScreenerNet: Learning Self-Paced Curriculum for Deep Neural Networks
A. Architecture of the main network and the
ScreenerNet
For all experiments, we used the same or slightly simpler
network architecture for the ScreenerNetḞor the description of the neural network architectures, we use following
notations: Convk m denotes a convolution layer of k × k
kernel, m output channels, FC m denotes a fully connected
layer of which output dimension is m and MaxPool denotes
2x2 max pooling. Dropout probability is set to 0.5 for all
experiments unless mentioned. Table 1 shows the detailed
architecutres for the main network and the ScreenerNet for
all experiments. Table 2 shows the detailed optimization
parameter sets for all experiments.
Table 3. Computational cost and sizes of networks.
Cost
Main Network
ScreenerNet
Common
# of param.
ms. / sample
# of param.
ms. / sample
batch-size
CIFAR10
62,018
0.32
123,271
0.69
64
MNIST
21,840
0.37
28.017
0.62
64
VOC2012
140,097,492
13.6
279,671,829
26.9
32
Table 1. Architectures of baseline and ScreenerNet
Network
Main Network
MNIST
ScreenerNet
CIFAR10
small
CIFAR10
mid
CIFAR10
VGG11
Pascal
VOC2012
Main Network
ScreenerNet
Main Network
ScreenerNet
Baseline
ScreenerNet
Main Network
ScreenerNet
Main Network
CartPole
ScreenerNet
Architecture
Conv5 10–ReLU–MaxPool–Conv5 20–Dropout–ReLU
–MaxPool–FC 50–ReLU–Dropout–FC 10–LogSoftmax
Conv3 4–ELU–Conv3 8–ELU–Conv3 16–ELU
–Conv3 32–ELU–FC 1
Conv5 6–ReLU–MaxPool–Conv5 16–ReLU–MaxPool
–FC 120–ReLU–FC 84–ReLU–FC 10–LogSoftmax
Conv5 6–ReLU–MaxPool–Conv5 16–ReLU–MaxPool
–FC 120–ReLU–FC 84–ReLU–FC 1–Sigmoid
Conv5 32–ReLU–MaxPool–Conv5 32–ReLU–MaxPool
–FC 400–ReLU–FC 200–ReLU–FC 10–LogSoftmax
Conv5 32–ReLU–MaxPool–Conv5 32–ReLU–MaxPool
–FC 400–ReLU–FC 200–ReLU–FC 1–Sigmoid
[VGG-11 convolutional features network]–Dropout–FC 512
–ReLU–Dropout–FC 512–ReLU–FC 10–LogSoftmax
[VGG-11 convolutional features network]–Dropout–FC 512
–ReLU–Dropout–FC 512–ReLU–FC 1–Sigmoid
[VGG-19 up to FC 4096]–Dropout–FC 128–
BatchNorm–ReLU–Dropout–FC 20–Sigmoid
[VGG-19 up to FC 4096]–FC 64–ReLU–FC 1
–Sigmoid
FC 16–ReLU–FC 16–ReLU–FC 16–ReLU–FC 2
FC 16–ReLU–FC 16–ReLU–FC 16–ReLU–FC 1
–Sigmoid
Table 2. Parameters used for our experiments: optimizer algorithm,
learning rate, batch-size, margin, and regularizer in Equation (3).
Parameters
MMNIST
CIFAR10
(all)
Pascal
VOC2012
CartPole
Main Network
ScreenerNet
Main Network
ScreenerNet
Main Network
ScreenerNet
Main Network
ScreenerNet
Optimizer
SGD
Adam
SGD
Adam
SGD
Adam
Adam
Adam
LR
0.01
0.0001
0.01
0.001
0.001
0.0025
0.001
0.0015
Batch
64
64
32
32
M
α
N/A
1.0
0.01
N/A
20 to 5 0.01
N/A
1.0
0.001
N/A
2.0
0.001
B. Computational Cost
Table 3 shows the comparison of the baselind and its combination with ScreenerNet in terms of required computation
and network size. The training performance is measured for
PyTorch implementations on a single NVIDIA GTX 1080.
C. More Qualitative Results for the Error
Analysis with MNIST
We show the examples that the main network only and the
ScreenerNet augmented network both fail in Figure 11.
727 060 535 282 353 949 535 989 565 646 272
794
272 494 171 757 979 787 737 383 494 353 848
865
949 161 090 202 191 353 161 080 838 797 979
732
060 464 989 272 272 727 717 717 080 353 838
712
898 171 707 282 949 272 272 797 797 656
860
Figure 11. Examples that both ScreenerNet and baseline fail. Left:
Both ScreenerNet and baseline yield the same classification. Right:
ScreenerNet and baseline classify differently.
| 1cs.CV |
Polyphonic Music Generation by Modeling
Temporal Dependencies Using a RNN-DBN
arXiv:1412.7927v1 [cs.LG] 26 Dec 2014
Kratarth Goel, Raunaq Vohra, and J. K. Sahoo
Birla Institute of Technology and Science, Pilani,
Goa, India
kratarthgoel@gmail.com,
ronvohra@gmail.com,
jksahoo@goa.bits-pilani.ac.in
Abstract. In this paper, we propose a generic technique to model temporal dependencies and sequences using a combination of a recurrent neural network and a Deep Belief Network. Our technique, RNN-DBN, is
an amalgamation of the memory state of the RNN that allows it to provide temporal information and a multi-layer DBN that helps in high level
representation of the data. This makes RNN-DBNs ideal for sequence
generation. Further, the use of a DBN in conjunction with the RNN
makes this model capable of significantly more complex data representation than an RBM. We apply this technique to the task of polyphonic
music generation.
Keywords: Deep architectures, recurrent neural networks, music generation, creative machine learning, Deep Belief Networks, generative models
1
Introduction
Creative machine learning is an extremely relevant research area today. Generative models form the basis of creative learning, and for this reason, a high level
of sophistication is required from these models. Also, identifying features in the
subjective fields of art, literature and music is an arduous task, which is only
made more difficult when more elaborate learning is desired. Deep architectures,
therefore, present themselves as an ideal framework for generative models, as
they are inherently stochastic and support increasingly complex representations
with each added layer. Recurrent neural networks (RNNs) have also used with
great success as regards generative models, particularly handwriting generation,
where they have been used to achieve the current state-of-the-art results. The
internal feedback or memory state of these neural networks is what makes them
a suitable technique for sequence generation in tasks like polyphonic music composition.
There have been many attempts to generate polyphonic music in the past,
and a variety of techniques have been applied, some of which will be discussed
here. Matic et al. [1] used neural networks and cellular automata to generate
2
Kratarth Goel, Raunaq Vohra and J. K. Sahoo
melodies. However, this is dependent on a feature set based on emotions and
further requires specialized knowledge of musical theory. A similar situation is
observed with the work done by Maeda and Kajihara [2], who used genetic algorithms for music generation. Elman networks with chaotic inspiration were
used by [3] to compose music. While RNNs are an excellent technique to model
sequences, they used chaotic inspiration as an external input instead of real
stochasticity to compose original music. A Deep Belief Network with a sliding
window mechanism to create jazz melodies was proposed by [4]. However, due
to lack of temporal information, there were many instances of repeated notes
and pitches. Most recently, Boulanger-Lewandowski et al. [5] used RNN-RBMs
for polyphonic music generation and obtained promising results. We propose
a generic technique which is a combination of a RNN and a Deep Belief Network for sequence generation, and apply it to automatic music composition. Our
technique, the RNN-DBN, effectively combines the superior sequence modeling
capabilities of a RNN with the high level data modeling that a DBN enables
to produce rich, complex melodies which do not feature significant repetition.
Moreover, the composition of these melodies does not require any feature selection from the input data. This model is presesnted as a generic technique, i. e.
it does not make any assumptions about the nature of the data. We apply our
technique to a variety of datasets and have achieved excellent results which are
on par with the current state-of-the-art.
The rest of this paper is organized as follows: Section 2 discusses various deep
and neural network architectures that serve as the motivation for our technique,
described in Section 3. We demonstrate the application of RNN-DBNs to the task
of polyphonic music generation in Section 4 and present our results. Section 5
discusses possible future work regarding our technique and concludes the paper.
2
2.1
Preliminaries
Restricted Boltzmann Machines
Restricted Boltzmann machines (RBMs) are energy based models with their
energy function E(v, h) defined as:
E(v, h) = −b′ v − c′ h − h′ W v
(1)
where W represents the weights connecting hidden (h) and visible (v) units
and b, c are the biases of the visible and hidden layers respectively.
This translates directly to the following free energy formula:
X
X
F (v) = −b′ v −
log
ehi (ci +Wi v) .
(2)
i
hi
Because of the specific structure of RBMs, visible and hidden units are conditionally independent given one another. Using this property, we can write:
Y
p(hi |v)
(3)
p(h|v) =
i
Polyphonic Music Generation Using a RNN-DBN
p(v|h) =
Y
p(vj |h).
3
(4)
j
Samples can be obtained from a RBM by performing block Gibbs sampling,
where visible units are sampled simultaneously given fixed values of the hidden
units. Similarly, hidden units are sampled simultaneously given the visible unit
values. A single step in the Markov chain is thus taken as follows,
h(n+1) = σ(W ′ v (n) + c)
v (n+1) = σ(W h(n+1) + b),
(5)
where σ represents the sigmoid function acting on the activations of the
(n + 1)th hidden and visible units. Several algorithms have been devised for
RBMs in order to efficiently sample from p(v, h) during the learning process, the
most effective being the well-known contrastive divergence (CD − k) algorithm
[6].
In the commonly studied case of using binary units (where vj and hi ∈ {0, 1}),
we obtain, from Eqn. (4), a probabilistic version of the activation function:
P (hi = 1|v) = σ(ci + Wi v)
P (vj = 1|h) = σ(bj + Wj′ h)
(6)
The free energy of an RBM with binary units thus further simplifies to:
X
log(1 + e(ci +Wi v) ).
(7)
F (v) = −b′ v −
i
We obtain the following log-likelihood gradients for an RBM with binary
units:
−
∂log p(v)
(i)
= Ev [p(hi |v) · vj ] − vj · σ(Wi · v (i) + ci )
∂Wij
∂log p(v)
−
= Ev [p(hi |v)] − σ(Wi · v (i) )
∂ci
∂log p(v)
(i)
=
Ev [p(vj |h)] − vj
−
∂bj
(8)
where σ(x) = (1 + e−x )−1 is the element-wise logistic sigmoid function.
2.2
Recurrent Neural Network
Recurrent neural networks (RNNs) are a particular family of neural networks
where the network contains one or more feedback connections, so that activation
of a cluster of neurons can flow in a loop. This property allows for the network to
effectively model time series data and learn sequences. An interesting property of
RNNs is that they can be modeled as feedforward neural networks by unfolding
4
Kratarth Goel, Raunaq Vohra and J. K. Sahoo
them over time. RNNs can be trained using the Backpropogation Through Time
(BPTT) technique. If a network training sequence starts at a time instant t0
and ends at time t1 , the total cost function is simply the sum over the standard
error function Esse/ce at each time step,
Etotal =
Xt1
t=t0
Esse/ce (t)
(9)
and the gradient descent weight update contributions for each time step are
given by,
∆wij = −η
∂Etotal (t0 , t1 ) Xt1 ∂Esse/ce (t)
=
t=t0
∂wij
∂wij
(10)
∂E
sse/ce
now have contributions
The partial derivatives of each component ∂w
ij
from multiple instances of each weight wij ∈ {Wv(t−1) h(t−1) , W(ht−1) h(t) } and are
dependent on the inputs and hidden unit activations at previous time instants.
The errors now must be back-propagated through time as well as through the
network.
2.3
Deep Belief Network
RBMs can be stacked and trained greedily to form Deep Belief Networks (DBNs).
DBNs are graphical models which learn to extract a deep hierarchical representation of the training data [7]. They model the joint distribution between observed
vector x and the ℓ hidden layers hk as follows:
1
ℓ
P (x, h , . . . , h ) =
ℓ−2
Y
k=0
k
P (h |h
k+1
!
) P (hℓ−1 , hℓ )
(11)
where x = h0 , P (hk−1 |hk ) is a conditional distribution for the visible units
conditioned on the hidden units of the RBM at level k, and P (hℓ−1 , hℓ ) is the
visible-hidden joint distribution in the top-level RBM.
The principle of greedy layer-wise unsupervised training can be applied to
DBNs with RBMs as the building blocks for each layer [8]. We begin by training
the first layer as an RBM that models the raw input x = h(0) as its visible
layer. Using that first layer, we obtain a representation of the input that will
be used as data for the second layer. Two common solutions exist here, and
the representation can be chosen as the mean activations p(h(1) = 1|h(0) ) or
samples of p(h(1) |h(0) ). Then we train the second layer as an RBM, taking the
transformed data (samples or mean activations) as training examples (for the
visible layer of that RBM). In the same vein, we can continue adding as many
hidden layers as required, while each time propagating upward either samples
or mean values.
Polyphonic Music Generation Using a RNN-DBN
2.4
5
Recurrent Temporal Restricted Boltzmann Machine
The Recurrent Temporal Restricted Boltzmann Machine (RTRBM) [9] is a sequence of conditional RBMs (one at each time instant) whose parameters {btv , bth
, W t } are time-dependent and depend on the sequence history at time t, denoted
by A(t) = {v (τ ) , u(τ ) |τ < t}, where u(t) is the mean-field value of h(t), as seen
in [5]. The RTRBM is formally defined by its joint probability distribution,
P (v (t) , h(t)) =
T
Y
P (v (t) , h(t) |A(t) )
(12)
t=1
where P (v (t) , h(t) |A(t) ) is the joint probability of the tth RBM whose parameters are defined below, from Eqn. (13) and Eqn. (14). While all the parameters
of the RBMs may usually depend on the previous time instants, we will consider
the case where only the biases depend on u(t−1) .
(t)
bh = bh + Wuh u(t−1)
(13)
(t−1)
b(t)
v = bv + Wuv u
(14)
(0)
which gives the RTRBM six parameters,{W, bv , bh , Wuv , Wuh , u }. The more
general scenario is derived in similar fashion. While the hidden units h(t) are binary during inference and sample generation, it is the mean-field value u(t) that
is transmitted to its successors (Eqn. (15)). This important distinction makes
exact inference of the u(t) easy and improves the efficiency of training[9].
(t)
u(t) = σ(Wvh v (t) + bh ) = σ(Wvh v (t) + Wuh u(t−1) + bh )
(15)
Observe that Eqn. (15) is exactly the defining equation of a RNN (defined
in Section 4) with hidden units u(t) .
3
RNN-DBN
The RTRBM can be thought of as a sequence of conditional RBMs whose parameters are the output of a deterministic RNN [5], with the constraint that
the hidden units must describe the conditional distributions and convey temporal information for sequence generation. The use of a single RBM layer greatly
constricts the expressive power of the model as a whole. This constraint can be
lifted by combining a full RNN having distinct hidden units u(t) with a RTRBM
graphical model, replacing the RBM structure with the much more powerful
model of a DBN. We call this model the RNN-DBN.
In general, the parameters of the RNN-DBN are made to depend only on
u(t−1) given by Eqn. (13) and Eqn. (14) along with,
(t)
bh2 = bh2 + Wuh2 u(t−1)
(16)
for the second hidden layer in an RNN-DBN with 2 hidden layers. The joint
probability distribution of the RNN-DBN is also given by Eqn. (12), but with
6
Kratarth Goel, Raunaq Vohra and J. K. Sahoo
u(t) defined arbitrarily, as given by Eqn. (17). For simplicity, we consider the
(t)
(t)
(t)
RNN-DBN parameters to be {Wvh(t) , bv , bh1 , bh2 } for a 2 hidden layer RNNDBN (shown in Fig. 1), i.e. only the biases are variable, and a single-layer RNN,
whose hidden units u(t) are only connected to their direct predecessor u(t−1) and
to v (t) by the relation,
u(t) = σ(Wvu v (t) + Wuu h(t−1) + bu ).
(17)
Fig. 1: A RNN-DBN with 2 hidden layers.
The DBN portion of the RNN-DBN is otherwise exactly the same as any
general DBN. This gives the 2 hidden layer RNN-DBN twelve parameters, {
Wvh1 ,Wh1 h2 , bv , bh1 , bh2 , Wuv , Wuh1 , Wuh2 , u(0) , Wvu , Wuu , bu }.
The training algorithm is based on the following general scheme:
1. Propagate the current values of the hidden units u(t) in the RNN portion of
the graph using Eqn. (17).
2. Calculate the DBN parameters that depend on the u(t) (Eqn. (13), (14) and
(16)) by greedily training layer-by-layer of the DBN, each layer as an RBM
(Train the first layer as an RBM that models the raw input as its visible
layer and use that first layer to obtain a representation of the input that will
be used as data for the second layer and so on).
3. Use CD-k to estimate the log-likelihood gradient (Eq. (8)) with respect to
W , bv and bh for each RBM composing the DBN.
4. Repeat steps 2 and 3 for each layer of the DBN.
(t)
(t)
(t)
5. Propagate the estimated gradient with respect to bv , bh and bh2 backward
through time (BPTT) [10] to obtain the estimated gradient with respect to
the RNN, for the RNN-DBN with 2 hidden layers.
Polyphonic Music Generation Using a RNN-DBN
4
7
Implementation and Results
Table 1: Log-likelihood (LL) for various musical models in the polyphonic music
generation task.
Model
JSB Chorales MuseData
(LL)
(LL)
Random
-61.00
-61.00
RBM
-7.43
-9.56
NADE
-7.19
-10.06
Note N-Gram
-10.26
-7.91
RNN-RBM
-7.27
-9.31
RNN1 (HF)
-8.58
-7.19
RTRBM1
-6.35
-6.35
RNN-RBM1
-6.27
-6.01
RNN-NADE1
-5.83
-6.74
RNN-NADE1 (HF) -5.56
-5.60
RNN-DBN
-5.68
-6.28
Nottingham
(LL)
-61.00
-5.25
-5.48
-4.54
-4.72
-3.89
-2.62
-2.39
-2.91
-2.31
-2.54
PianoMidi.de (LL)
-61.00
-10.17
-10.28
-7.50
-9.89
-7.66
-7.36
-7.09
-7.48
-7.05
-7.15
We demonstrate our technique by applying it to the task of polyphonic music
generation. We used a RNN-DBN with 2 hidden DBN layers - each having 150
binary units - and 150 binary units in the RNN layer. The visible layer has 88
binary units, corresponding to the full range of the piano from A0 to C8. We
implemented our technique on four datasets - JSB Chorales , MuseData 2 , Nottingham 3 and Piano-midi.de. None of the preprocessing techniques mentioned
in [5] have been applied to the data, and only raw data has been given as input
to the RNN-DBN. We evaluate our models qualitatively by generating sample
sequences and quantitatively by using the log-likelihood (LL) as a performance
measure. Results are presented in Table 1 (a more comprehensive list can be
found in [5]).
The results indicate that our technique is on par with the current state-of-theart. We believe that the difference in performance between our technique and the
current best can be attributed to lack of preprocessing. For instance, transposing
the sequences in a common tonality (e.g. C major/minor) and normalizing the
tempo in beats (quarternotes) per minute as preprocessing can have the most
effect on the generative quality of the model. It also helps to have as pretraining, the initialization the Wvh1 , Wh1 h2 , bv , bh1 , bh2 parameters with independent
RBMs with fully shuffled frames (i.e. Wuh1 = Wuh2 = Wuv = Wuu = Wvu = 0).
Initializing the Wuv , Wuu , Wvu , bu parameters of the RNN with the auxiliary
cross-entropy objective via either stochastic gradient descent (SGD) or, preferably, Hessian-free (HF) optimization and subsequently finetuning significantly
1
2
3
These marked results are obtained after various preprocessing, pretraining methods
and optimization techniques described in the last paragraph of this section.
http://www.musedata.org
ifdo.ca/ seymour/nottingham/nottingham.html
8
Kratarth Goel, Raunaq Vohra and J. K. Sahoo
helps the density estimation and prediction performance of RNNs which would
otherwise perform worse than simpler Multilayer Perceptrons [5]. Optimization
techniques like gradient clipping, Nesterov momentum and the use of NADE for
conditional density estimation also improve results.
5
Conclusions and Future Work
We have proposed a generic technique called Recurrent Neural Network-Deep
Belief Network (RNN-DBN) for modeling sequences with generative models and
have demonstrated its successful application to polyphonic music generation. We
used four datasets for evalutaing our technique and have obtained results on par
with the current state-of-the-art. We are currently working on improving the
results this paper, by exploring various pretraining and optimization techniques.
We are also looking at showcasing the versatility of our technique by applying
it to different problem statements.
References
1. Matic, I., Oliveira, A., Cardoso, A.: Automatic melody generation using neural
networks and cellular automata. (Sept 2012) 89–94
2. Maeda, Y., Kajihara, Y.: Rhythm generation method for automatic musical composition using genetic algorithm. (July 2010) 1–7
3. Coca, A., Romero, R., Zhao, L.: Generation of composed musical structures
through recurrent neural networks based on chaotic inspiration. (July 2011) 3220–
3226
4. Bickerman, G., Bosley, S., Swire, P., Keller, R.: Learning to create jazz melodies
using deep belief nets. First International Conference on Computational Creativity
(2010)
5. Boulanger-Lewandowski, N., Bengio, Y., Vincent, P.: Modeling temporal dependencies in high-dimensional sequences: Application to polyphonic music generation
and transcription. In: ICML, icml.cc / Omnipress (2012)
6. Hinton, G.E.: Training products of experts by minimizing contrastive divergence.
Neural Computation 14(8) (2002) 1771–1800
7. Hinton, G.E., Osindero, S.: A fast learning algorithm for deep belief nets. Neural
Computation 18 (2006) 2006
8. Bengio, Y.: Learning deep architectures for ai. Foundations and Trends in Machine
Learning 2(1) (2009) 1–127
9. Sutskever, I., Hinton, G.E., Taylor, G.W.: The recurrent temporal restricted boltzmann machine. (2008) 1601–1608
10. Rumelhart, D., Hinton, G., Williams, R.: Learning representations by backpropagating errors. Nature 323(6088) (1986) 533–536
This figure "fig1.png" is available in "png" format from:
http://arxiv.org/ps/1412.7927v1
| 9cs.NE |
arXiv:1603.00585v1 [cs.NI] 2 Mar 2016
Optimal Scheduling for Incentive WiFi Offloading
under Energy Constraint
Juntao Gao, Minoru Ito
Norio Shiratori
Graduate School of Information Science
Nara Institute of Science and Technology, Nara, Japan
Email: {jtgao, ito}@is.naist.jp
Research Institute of Electrical Communication
Tohoku University, Sendai, Japan
Abstract—WiFi offloading, where mobile device users (e.g.,
smart phone users) transmit packets through WiFi networks
rather than cellular networks, is a promising solution to alleviating the heavy traffic burden of cellular networks due to
data explosion. However, since WiFi networks are intermittently
available, a mobile device user in WiFi offloading usually needs
to wait for WiFi connection and thus experiences longer delay
of packet transmission. To motivate users to participate in WiFi
offloading, cellular network operators give incentives (rewards
like coupons, e-coins) to users who wait for WiFi connection and
transmit packets through WiFi networks.
In this paper, we aim at maximizing users’ rewards while
meeting constraints on queue stability and energy consumption.
However, we face scheduling challenges from random packet
arrivals, intermittent WiFi connection and time varying wireless
link states. To address these challenges, we first formulate
the problem as a stochastic optimization problem. We then
propose an optimal scheduling policy, named Optimal scheduling
Policy under Energy Constraint (OPEC), which makes online
decisions as to when to delay packet transmission to wait for
WiFi connection and which wireless link (WiFi link or cellular
link) to transmit packets on. OPEC automatically adapts to
random packet arrivals and time varying wireless link states,
not requiring a priori knowledge of packet arrival and wireless
link probabilities. As verified by simulations, OPEC scheduling
policy can achieve the maximum rewards while keeping queue
stable and meeting energy consumption constraint.
I. I NTRODUCTION
With the explosion of smart devices (like smart phones,
tablets) in daily life, mountainous data (like high resolution
photos and videos) have been generated every day. Global mobile data traffic has already reached 3.7 exabytes per month at
the end of 2015 and will be 30.6 exabytes per month by 2020,
i.e., eightfold increase, according to the report of Cisco [1].
As a result, cellular networks are overloaded, which degrades
users’ quality of experience (we already experienced long
delay, low throughput during peak hours in crowded downtown
areas [2]). To improve users’ quality of experience, cellular
network operators struggle to increase network connection
speeds. However, deploying more cellular infrastructures (like
base stations) is very expensive and cannot bring reasonable
benefits. Recently, offloading mobile data traffic from cellular
networks to WiFi networks of low cost and high connection
speed (known as WiFi offloading) has been proposed as a
promising solution to alleviating the heavy traffic burden of
cellular networks [2], [3].
)KRR[RGX(9
=O,O)U\KXGMK
=O,O)U\KXGMK
'6
'6
'6
\OJKU
\OJKU
SU\KSKTZ
Fig. 1.
Illustration of participants in WiFi offloading.
Participants in WiFi offloading include cellular network
operators (cellular Base Station (BS)), WiFi Access Point (AP)
owners and mobile device users, as shown in Figure 1. In
such a network, mobile device users leverage loose delay
requirement of delay-tolerant applications (e.g., uploading
photos and videos to iCloud/youtube, synchronizing dropbox)
to postpone packet transmission for utilizing intermittent WiFi
connection. While cellular network operators can benefit from
WiFi offloading, both WiFi AP owners and users may suffer from bad quality of experience: WiFi AP owners need
to sacrifice WiFi bandwidth and energy to offer offloading
services and users need to delay packet transmission to wait
for intermittent WiFi connection. To motivate WiFi AP owners
to provide WiFi offloading services and users to utilize WiFi
offloading services, incentive frameworks have been proposed
recently (see Section II for related work), where cellular
network operators give rewards to WiFi AP owners and users
to compensate for their loss. However, these frameworks focus
on maximizing the utility of network operators (related to
operators’ satisfaction, e.g., minimizing operators’ incentive
cost) rather than mobile device users. Further, they do not
take into account limited battery power of mobile devices.
In this paper, we aim at maximizing users’ rewards in
incentive WiFi offloading under constraint on energy consumption of mobile devices. However, to achieve this goal,
we face scheduling challenges from random packet arrivals,
intermittent WiFi connection and time varying wireless link
states. To address these challenges, we first formulate the
problem as a stochastic optimization problem with constraint
on average energy consumption and queue stability (defined
later). We then propose an optimal scheduling policy, named
Optimal scheduling Policy under Energy Constraint (OPEC),
which makes online decisions as to when to delay packet
transmission to wait for WiFi connection and which wireless
link (WiFi link or cellular link) to transmit packets on.
OPEC automatically adapts to random packet arrivals and time
varying wireless link states, not requiring a priori knowledge
of packet arrival and wireless link probabilities. As verified
by simulations, OPEC scheduling policy can achieve the
maximum rewards while keeping queue stable and meeting
energy consumption constraint.
The rest of this paper is organized as follows. In Section
II, we review related work on incentive WiFi offloading and
scheduling policies under energy constraint. We introduce
network models and assumptions for WiFi offloading system
in Section III. We give a formulation of the optimal scheduling
problem in Section IV and present a detailed description of
OPEC scheduling policy in Section V. Simulation results to
verify OPEC are presented in Section VI and the whole paper
is concluded in Section VII.
II. R ELATED W ORK
To motivate users to delay their packet transmissions for
offloading cellular traffic to complementary networks (DTNs
and WiFi hotspots), Zhuo et al. [4] proposed an incentive
framework named ”Win-Coupon,” in which cellular network
operators provide users incentives (service discount coupon).
Their objective is to minimize the incentive cost of cellular
network operators.
A similar work [5], which also motivates users to offload
cellular traffic to WiFi networks, proposed an incentive mechanism to allow network operators to optimally award users,
but differs from [4] in how network operators negotiate an
offloading contract with users.
Instead of motivating users, Kang et al. [6] aim to motivate
WiFi Access Point (AP) owners to provide data offloading
services for network operators. They proposed an incentive
mechanism, where network operators award WiFi AP owners
based on both the amount of data offloading services and
service quality they provide. This work also focuses on maximizing network operators’ utility.
Iosifidis et al. [7] discussed the challenges of designing
incentive schemes to encourage WiFi AP owners to participate
in offering data offloading services (such networks are called
user-provided networks in [7]). They pointed out that energy
consumption and cost for providing and receiving offloading
services are critical to incentive scheme designs.
However, all above works [4]–[6] focused on maximizing
the utility of network operators rather than users and they
did not consider the issue of energy consumption of mobile
devices, which is a major concern to all battery-powered
mobile devices. Different from [4]–[6], our goal in this paper
is to maximize rewards of users under energy consumption
constraint.
There are also some works considering energy consumption
issues in offloading cellular traffic. Ra et al. [8] studied
the tradeoff between energy consumption and data delay in
WiFi offloading networks. They proposed an algorithm that
can achieve the minimum energy consumption under queue
stability constraint. However, they did not consider incentive
issue and their objective is different from ours.
Al-Kanj et al. [9] investigated how to offload cellular traffic
under energy consumption constraint on mobile devices in
multihop ad hoc networks. Their objective is to minimize the
usage of scarce cellular spectrum, which is different from the
goal of our work.
Recently, Onoue et al. [10] considered reducing data delay
in WiFi offloading networks under energy consumption constraint. The idea is that a user prefetches the data to be used
before he leaves WiFi network coverage. Therefore, the user
can obtain the data immediately when he leaves WiFi coverage
without waiting for next WiFi connection, thus reducing data
delay. However, they did not consider incentive issues.
III. N ETWORK M ODELS
In this section, we introduce network models for WiFi
offloading system concerning network time, packet arrival
process, wireless links, rewards and energy consumption. Key
notations are summarized in Table I.
Network time: In a WiFi offloading network as illustrated
in Figure 1, network time is divided into time slots t of equal
duration, t = 0, 1, 2, · · · . For example, one time slot could be
duration of 20 seconds.
Packet arrival process and queue: Packets arrive at the
mobile device following a general stochastic process {a(t); t ≥
0}, where a(t) is the number of packets arriving at the mobile
device at time slot t. We assume a(t) has finite second
moment for any t, E{a(t)2 } < ∞. All arriving packets will be
stored in a queue of the mobile device awaiting for wireless
transmission. Denote by Q(t) the number of packets in the
queue at the beginning of time slot t.
Wireless links: If a mobile device is connected through a
wireless channel to either a cellular base station (BS) or a
WiFi access point (AP), we say a wireless link is available
for the mobile device. We assume at most n > 0 wireless
links are available for the mobile device every time slot t,
including 1 cellular wireless link and n − 1 WiFi wireless
links. The state of a wireless link at slot t (i.e., how many
packets can be transmitted by that link at slot t) is a random
variable and varies from time slot to time slot, due to factors
like user mobility, interference and wireless channel fading.
Denote the states of n wireless links at slot t by a vector
S(t) = S1 (t), S2 (t), · · · , Sn (t) , where vector element S1 (t)
is the state of the cellular wireless link (i.e., at most S1 (t)
packets can be transmitted by cellular link at slot t) and
Si (t), 2 ≤ i ≤ n, is the state of WiFi wireless link i (at
most Si (t) packets can be transmitted by WiFi link i at slot
t).
TABLE I
K EY NOTATIONS IN W I F I OFFLOADING SYSTEM .
Notation
t
a(t)
Q(t)
Z(t)
n
S(t)
α(t)
D
π
P
απ (t)
pπ (t)
pc
pw
pav
bπ (t)
r π (t)
Description
slotted network time, t = 0, 1, 2, · · ·
the number of packets arriving at the mobile device at time slot t
the number of packets in the queue at the beginning of time slot t, also called queue size
the value of virtual queue at the beginning of time slot t
the maximum number of wireless links available for the mobile device every time slot t
the vector of states of n wireless links at slot t, where element S1 (t) is the state of cellular wireless link
and all other vector elements are the states of WiFi wireless links
a transmission decision made by the mobile device at slot t
the set of all n + 1 possible transmission decisions
a transmission policy
the set of all possible transmission policies for the considered WiFi offloading network
the transmission decision made by policy π at slot t
the energy consumption under transmission policy π at slot t
the energy consumption when transmitting packets by a cellular wireless link at a time slot
the energy consumption when transmitting packets by a WiFi wireless link at a time slot
constraint on average energy consumption
the transmission capacity under transmission policy π at time slot t
the reward given to the user under transmission policy π at slot t
Transmission decision: At each time slot t, the mobile
device makes a transmission decision from the following n+1
decision options:
D1: delaying packet transmission (i.e., not transmitting
packets at slot t),
D2: transmitting packets by the cellular link,
D3: transmitting packets by a WiFi link i, 2 ≤ i ≤ n.
We assume the mobile device can transmit packets by at
most one wireless link at a time slot. Denote by D the set of
all n+1 possible transmission decisions. Denote a transmission
decision at slot t by vector α(t) = α1 (t), α2 (t), · · · , αn (t) ,
where α(t) ∈ D and αi (t) is a binary variable regarding
wireless link i. Then
for D1,
(0, 0, · · · , 0)
(1, 0, · · · , 0)
for D2,
α(t) =
(1)
(0, 0, · · · , 1, · · · , 0)
for D3, αi (t) = 1.
Transmission policy: A transmission policy π is a set of
rules for a user to make transmission decisions at each time
slot t.
For example, a transmission policy π can be one that first
observes the states of all wireless links S(t) at each time slot t
and then makes a random transmission decision from decision
set D.
Denote by απ (t) = απ1 (t), απ2 (t), · · · , απn (t) the transmission decision under transmission policy π at time slot t,
απ (t) ∈ D. Denote by P the set of all possible policies for
the considered WiFi offloading network.
Energy consumption: Denote by pπ (t) the energy consumption under transmission policy π at slot t.
for D1,
0
pc
for D2,
pπ (t) =
(2)
pw
for D3.
Transmission capacity: Denote by bπ (t) the transmission
capacity under transmission policy π at time slot t, i.e., the
maximum number of packets that can be transmitted with
transmission decision απ (t) and link states S(t),
bπ (t) = απ (t)S(t) =
n
X
Si (t)απi (t)
(3)
i=1
We assume bπ (t) has finite second moment for any t under
any transmission policy π ∈ P, i.e., E{bπ (t)2 } < ∞.
Incentive: To encourage mobile device user to offload
cellular traffic, one unit of reward rπ (t) is given to the user
if the user decides to delay packet transmission or transmit
packets by a WiFi link under transmission policy π at slot t,
i.e.,
1
for D1 or D3,
π
r (t) =
(4)
0
for D2.
IV. O PTIMAL S CHEDULING P ROBLEM F ORMULATION
Our goal is to maximize user rewards under energy consumption constraint (defined later) in incentive WiFi offloading
network. In this section, we formulate the problem as a
stochastic optimization problem [11].
We first define how queue Q(t) evolves, when queue Q(t)
is said to be stable and quantities concerning time average
expectations of packet arrivals, energy consumption, transmission capacity and rewards.
Queue dynamics: Under transmission policy π ∈ P, the
queue in the mobile device evolves as follows:
Q(t + 1) = max Q(t) − bπ (t), 0 + a(t)
(5)
Queue stability: The queueing process Q(t) is mean rate
stable [11] if
E |Q(t)|
=0
(6)
lim
t→∞
t
Time average expectations: For transmission policy π ∈
P, we define the following time average expectations:
t−1
a(t) =
1X
E{a(τ )}
t τ =0
pπ (t) =
1X
E{pπ (τ )}
t τ =0
(7)
t−1
(8)
t−1
b (t) =
1X
E{bπ (τ )}
t τ =0
r π (t) =
1X
E{rπ (τ )}
t τ =0
π
(9)
t−1
(10)
Note that 0 ≤ r π (t) ≤ 1 from definition (4).
Optimal scheduling problem formulation: In the considered incentive WiFi offloading network, our objective is to
propose a scheduling policy to maximize the time average expected rewards while meeting energy consumption constraint
pav , 0 < pav < ∞, (12) and queue stability constraint (13). We
formulate this problem as a stochastic optimization problem
[11]:
maximize
π∈P
subject to
lim sup rπ (t)
(11)
t→∞
1) lim sup pπ (t) ≤ pav
(12)
t→∞
2) The queueing process Q(t) is
mean rate stable
3) απ (t) ∈ D, ∀ t ≥ 0
(13)
(14)
Note that the maximum value of lim supt→∞ r π (t) is 1.
If we let
f π (t) = −rπ (t)
y π (t) = pπ (t) − pav
(15)
(16)
and similarly define their time average expectations
π
f (t), y π (t) like (7) (8) (9) (10), then the above stochastic
optimization problem is equivalent to the following one.
π
minimize
lim inf f (t)
(17)
subject to
1) lim sup yπ (t) ≤ 0
(18)
π∈P
t→∞
t→∞
2) The queueing process Q(t) is
mean rate stable
3) απ (t) ∈ D, ∀ t ≥ 0
(19)
(20)
Feasible problem: The above stochastic optimization problem is said to be feasible if there exists at least one policy
π ∈ P that meets all constraints (18)-(20). Throughout this
paper we only consider feasible problem.
V. O PTIMAL S CHEDULING P OLICY
C ONSTRAINT
UNDER
E NERGY
In this section, we propose a scheduling policy that meets
constraints (18)-(20) and achieves the minimum limit value of
π
f (t) (i.e, maximum limit value of rewards r π (t)). Our policy
is named Optimal scheduling Policy under Energy Constraint
(OPEC), OPEC ∈ P.
We need several definitions in describing OPEC. First,
define a virtual queueing process {Z(t); t ≥ 0}, which is used
in OPEC to meet the average energy consumption constraint
(18). The virtual queueing process evolves as follows:
Z(t + 1) = max Z(t) + y OP EC (t), 0 ,
(21)
where y OP EC (t) is defined in (16) and we assume the initial
conditions Z(0) ≥ 0 and E{Z(0)2 } < ∞ 1 .
Then, define the following auxiliary quantities for transmission decision option α(t) ∈ D, which will help OPEC in
making decisions.
for D1,
0
pc
for D2,
p(t) =
(22)
pw
for D3.
n
X
Si (t)αi (t)
(23)
b(t) = α(t)S(t) =
i=1
r(t) =
1
0
for D1 or D3,
for D2.
(24)
The full description of OPEC policy is given in Algorithm
1.
Basic idea of OPEC: To meet queue stability constraint
(19), it is sufficient for one policy to greedily minimize
−Q(t)b(t) every time slot t as proved in [11] (b(t), f (t), y(t)
are defined in Algorithm 1). Similarly, to meet the constraint
on average energy consumption (18), it suffices to greedily
minimize Z(t)y(t) every time slot t [11]. Thus, to meet
both constraints (19) and (18), OPEC greedily minimizes
−Q(t)b(t) + Z(t)y(t) every time slot t. Since our final goal is
to minimize objective function (17) while meeting constraints
(18)-(20), OPEC also needs to consider f (t). To account for
f (t), OPEC minimizes a weighted sum V f (t) − Q(t)b(t) +
Z(t)y(t) every time slot t.
Control parameter V measures the importance of getting
rewards: larger V means we prefer getting more rewards,
which comes at the cost of an increase in queue size Q(t).
Thus, by controlling V we can achieve a tradeoff between
queue size and rewards. It can be proved, by Lyapunov
optimization theory [11], that OPEC policy indeed achieves
π
the minimum limit value of f (t) while meeting constraints
(18)-(20) by greedily minimizing the above weighted sum.
Line 1-3: Initialize the decision αOP EC (t) to be not
transmitting packets. value and tmp are auxiliary variables
to be used in the following steps.
Line 4-13: For each time slot t, OPEC scheduling policy
looks for a transmission decision α(t) ∈ D that minimizes
tmp = V f (t) − Q(t)b(t) + Z(t)y(t).
(25)
Line 14: Update the value of queue Q(t).
Line 15: Update the value of virtual queue Z(t).
1 These two assumptions about Z(0) will be used in proving the optimality
of OPEC. See [11] for more details.
Algorithm 1 OPEC scheduling policy
Input: Q(t) : the number of packets in the queue at slot t
Z(t) : the value of the virtual queue at slot t
S(t) : wireless link states at slot t
a(t) : the number of packets arriving at t
Parameters: n : the maximum number of available links
pw : energy consumption of transmitting packets
by a WiFi link
pc : energy consumption of transmitting packets
by a cellular link
V : control parameter (V ≥ 0)
Output: A transmission decision αOP EC (t) ∈ D
1: value = ∞;
2: αOP EC (t) = (0, 0, · · · , 0);
3: tmp = 0;
4: for all transmission decision options α(t) ∈ D do
5:
f (t) = P
−r(t);
n
6:
b(t) = i=1 Si (t)αi (t);
7:
y(t) = p(t) − pav ;
8:
Calculate tmp = V f (t) − Q(t)b(t) + Z(t)y(t);
9:
if tmp < value then
10:
value = tmp ;
11:
αOP EC (t) = α(t);
12:
end if
13: end for
14: Q(t + 1) = max Q(t) − bOP EC (t), 0 + a(t);
15: Z(t + 1) = max Z(t) + y OP EC (t), 0 ;
16: return αOP EC (t).
t−1
Q=
1X
Q(τ )
t τ =0
(26)
t−1
r OP EC =
1 X OP EC
r
(τ )
t τ =0
pOP EC =
1 X OP EC
p
(τ )
t τ =0
(27)
t−1
(28)
B. Simulation results
After running simulations for 106 slots (i.e., 11.5 days if 1
slot is equal to 1 second) under different settings of V , we
collected data of Q(t), rOP EC (t), pOP EC (t) and calculated
their time averages Q, r OP EC , pOP EC , as summarized in
Figures 2, 3, 4.
VALIDATION
For simulating the considered WiFi offloading network,
we developed a customized simulator, which takes as inputs
packet arrival process {a(t); t ≥ 0}, maximum number of
wireless links n, stochastic process of wireless link states S(t),
energy consumption pc , pw , energy consumption constraint
pav and control parameter V . By configuring these parameters,
we can simulate a wide range of network scenarios.
For illustration purpose, we consider a network scenario
with the following network settings.
• The number a(t) of packets arriving at the queue for
any time slot t ≥ 0 is an Independent and Identically
Distributed (IID) random variable, where P r{a(t) =
0} = 0.2, P r{a(t) = 2} = 0.3, P r{a(t) = 3} = 0.5.
• There are at most two wireless links (n = 2) available
every time slot: one cellular link and one WiFi link.
• Wireless link state vector S(t) is an IID vector random
variable over time slots. Since cellular network is of low
data rates and almost always available, we set cellular
link states and probabilities as P r{S1 (t) = 0} =
0.1, P r{S1 (t) = 1} = 0.2, P r{S1 (t) = 2} = 0.7.
0.7
(J/slot)
A. Simulation settings
0.8
pOPEC
In this section, we verify our OPEC scheduling policy by
simulations.
Time average energy consumption,
VI. S IMULATION
Since WiFi network is of high data rates and intermittently available, we set WiFi link states and probabilities as P r{S2 (t) = 0} = 0.7, P r{S2 (t) = 2} =
0.05, P r{S2(t) = 4} = 0.05, P r{S2 (t) = 10} =
0.1, P r{S2 (t) = 20} = 0.1.
• For energy consumption, we set pc = 1.15 J/slot when
transmitting by cellular link and pc = 1.1 J/slot when
transmitting by WiFi link. Constraint on average energy
consumption is set to be pav = 0.8 J/slot.
We vary control parameter V in our simulations and examine
its impact on performance metrics like time average queue
size Q, time average rewards r OP EC and time average energy
consumption pOP EC defined below
energy constraint,
pav
0.6
0.5
simulation result
0.4
0.3
0
40
80
120
Control parameter,
160
200
V
Fig. 2. Time average energy consumption pOP EC vs. control parameter V .
We first check whether OPEC satisfies the constraint on
average energy consumption. Figure 2 shows the simulated
results on pOP EC as V increases from 1 to 200. From Figure
2, we can see that the time average energy consumption
pOP EC under OPEC scheduling policy is always less than
the constraint pav = 0.8 J/slot. Thus, OPEC meets the given
14
rOPEC
0.9
simulation result
0.8
0.7
0.6
0.5
0
(packets)
12
Q
Time average queue size,
queue size limit
maximum reward
1.0
Time average rewards,
constraint on average energy consumption. Further observation
of Figure 2 shows that as V increases, pOP EC decreases and
approaches a limit value 0.32, indicating that with constraints
on queue stability and average energy consumption, there
exists a minimum value pmin of average energy consumption
for the considered WiFi offloading network. That is to say, the
average energy consumption constraint pav should be larger
than pmin , otherwise no scheduling policy can meet average
energy consumption constraint while keeping queue stable.
40
80
120
Control parameter,
simulation result
160
200
V
10
Fig. 4.
8
6
4
0
40
80
120
Control parameter,
Fig. 3.
V
160
200
Time average queue size Q vs. control parameter V .
We then check whether OPEC satisfies the queue stability
constraint. Figure 3 shows the simulated results on Q as V
increases. From Figure 3, we can see that the time average
queue size Q under OPEC scheduling policy increases as V
increases and tends to reach a limit valuethat is less than
14. As a common practice, we estimate E |Q(t)| by time
average Q. Therefore,
E |Q(t)|
14
Q
= lim
≤ lim
=0
(29)
lim
t→∞ t
t→∞ t
t→∞
t
Thus, OPEC also meets the constraint of queue being mean
rate stable.
Finally, we check whether OPEC can achieve the maximum
average reward 1. Figure 4 shows simulated results on r OP EC
as V increases. We can see from the figure that average
reward r OP EC under OPEC scheduling policy increases as V
increases and reaches the optimal value 1, thus verifying that
OPEC can indeed achieve the maximum average reward 1. In
summary, by simulations we validated that OPEC scheduling
policy can achieve the maximum average rewards while satisfying constraints on queue stability and energy consumption.
VII. C ONCLUSION
In this paper, we focused on maximizing users’ rewards
in incentive WiFi offloading under energy consumption constraint. We proposed a flexible scheduling policy, called Op-
Time average rewards r OP EC vs. control parameter V .
timal scheduling Policy under Energy Constraint (OPEC).
As validated by simulation, OPEC can achieve the maximum rewards while keeping queue stable and meeting energy consumption constraint. One merit of OPEC is that it
automatically adapts to random packet arrivals, intermittent
WiFi connection and time varying wireless link states, not
requiring a priori knowledge of packet arrival and wireless link
probabilities. As a future work, we will consider offloading
cellular traffic by multiple complementary networks, such as
WiFi, WiMax and Femtocell networks.
R EFERENCES
[1] Cisco, “Cisco visual networking index: Global mobile data traffic
forecast update, 2015-2020 white paper,” 2016.
[2] F. Rebecchi, M. D. D. Amorim, V. Conan, A. Passarella, R. Bruno, and
M. Conti, “Data offloading techniques in cellular networks: A survey,”
IEEE Communications Surveys & Tutorials, vol. 17, no. 2, pp. 580–603,
November 2014.
[3] K. Lee, J. Lee, Y. Yi, I. Rhee, and S. Chong, “Mobile data offloading:
How much can wifi deliver?” IEEE/ACM TRANSACTIONS ON NETWORKING, vol. 21, no. 2, pp. 536–551, April 2013.
[4] X. Zhuo, W. Gao, G. Cao, and S. Hua, “An incentive framework for
cellular traffic offloading,” IEEE Transactions on Mobile Computing,
vol. 13, no. 3, pp. 541–554, March 2014.
[5] S. Cai, L. Duan, J. Wang, S. Zhou, and R. Zhang, “Incentive mechanism
design for delayed wifi offloading,” in IEEE International Conference
on Communications (ICC), 2015.
[6] X. Kang and S. Sun, “Incentive mechanism design for mobile data offloading in heterogeneous networks,” in IEEE International Conference
on Communications (ICC), 2015.
[7] G. Iosifidis, L. Gao, J. Huang, and L. Tassiulas, “Incentive mechanisms
for user-provided networks,” IEEE Communications Magazine, vol. 52,
no. 9, pp. 20–27, September 2014.
[8] M.-R. Ra, J. Paek, A. B. Sharma, R. Govindan, M. H. Krieger,
and M. J. Neely, “Energy-delay tradeoffs in smartphone applications,”
in 8th International conference on Mobile systems, applications, and
services(MobiSys), 2010.
[9] L. Al-Kanj and Z. Dawy, “Offloading wireless cellular networks via
energy-constrained local ad hoc networks,” in IEEE Global Telecommunications Conference (GLOBECOM), 2011.
[10] Y. Onoue, M. Tamai, and K. Yasumoto, “Energy-constrained wi-fi offloading method using prefetching,” in IEEE 79th Vehicular Technology
Conference (VTC Spring), 2014.
[11] M. J. Neely, Stochastic Network Optimization with Application to
Communication and Queueing Systems. Morgan & Claypool, 2010.
| 8cs.DS |
Multiplying a Gaussian Matrix by a Gaussian Vector
arXiv:1702.02815v2 [math.PR] 5 Apr 2017
Pierre-Alexandre Mattei
pierre-alexandre.mattei@parisdescartes.fr
Laboratoire MAP5, UMR CNRS 8145
Université Paris Descartes & Sorbonne Paris Cité
45 rue des Saints-Pères, 75006 Paris, France
Abstract
We provide a new and simple characterization of the multivariate generalized Laplace
distribution. In particular, this result implies that the product of a Gaussian matrix with
independent and identically distributed columns by an independent isotropic Gaussian
vector follows a symmetric multivariate generalized Laplace distribution.
Keywords: Gamma distribution; Laplace distribution; Product distribution; Random
matrix; Scale mixture; Variance-gamma distribution
1. Introduction
Wishart and Bartlett (1932) proved that the inner product of two independent bidimensional
standard gaussian vectors follows a Laplace distribution. This result is deeply linked to the
fact that the Laplace distribution can be represented as an infinite scale mixture of Gaussians
with gamma mixing distribution. Specifically, if σ 2 follows a Gamma(1, 1/2) distribution
and x|σ ∼ N (0, σ 2 ), then x follows a standard Laplace distribution1 . This representation –
which was recently named the Gauss-Laplace representation by Ding and Blitzstein (2017)
following a blog post by Christian P. Robert2 – is particularly useful if one wants to simulate
a Laplace random variable: its use constitutes for example the cornerstone of the Gibbs
sampling scheme for the Bayesian lasso of Park and Casella (2008).
In this short paper, we are interested in studying links between multivariate counterparts
of these two characterizations. More specifically, we give a new simple characterization of the
multivariate generalized Laplace distribution of Kotz, Kozubowski, and Podgórski (2001).
As a corollary, we show that the product of a zero-mean Gaussian matrix with in dependent and identically distributed (i.i.d.) columns and a zero-mean isotropic Gaussian vector
follows a symmetric multivariate generalized Laplace distribution, a result that has useful
applications for Bayesian principal component analysis (Bouveyron, Latouche, and Mattei,
2016, 2017).
In the remainder of this paper, p and d are two positive integers and Sp+ denotes the
cone of positive semidefinite matrices of size p × p.
2. The multivariate generalized Laplace distribution
While the definition of the univariate Laplace distribution is widely undisputed, there exist
several different generalizations of this distribution to higher dimensions – a comprehensive
review of such generalizations can be found in the monograph of Kotz, Kozubowski, and Podgórski
1. The shape-rate parametrization of the gamma distribution is used through this paper. Note also that a
standard Laplace distribution is centered with variance two.
2. https://xianblog.wordpress.com/2015/10/14/gauss-to-laplace-transmutation/
1
(2001). In particular, McGraw and Wagner (1968) introduced a zero-mean elliptically contoured bidimensional Laplace distribution with univariate Laplace marginals. This distribution was later generalized to the p-dimensional setting by Anderson (1992), considering
characteristic functions of the form
1
∀u ∈ Rp , φ(u) =
,
1 T
1 + 2 u Σu
where Σ ∈ Sp+ . This distribution was notably promoted by Eltoft, Kim, and Lee (2006)
and is arguably the most popular multivariate generalization of the Laplace distribution
(Kotz, Kozubowski, and Podgórski, 2001, p. 229). Among its advantages, this distribution
can be slightly generalized to model skewness, by building on characteristic functions of the
form
1
,
∀u ∈ Rp , φ(u) =
1 T
1 + 2 u Σu − iµT u
where µ ∈ Rp accounts for asymmetry. Similarly to the univariate Laplace distribution, this
asymmetric multivariate generalization is infinitely divisible (Kotz, Kozubowski, and Podgórski,
2001, p. 256). Therefore, it can be associated with a specific Lévy process (Kyprianou,
2014, p. 5), whose increments will follow yet another generalization of the Laplace distribution, the multivariate generalized asymmetric Laplace distribution. This distribution,
introduced by Kotz, Kozubowski, and Podgórski (2001, p. 257) and further studied by
Kozubowski, Podgórski, and Rychlik (2013), will be the cornerstone of our analysis of multivariate caracterizations of Laplace and Gaussian distributions.
Definition 1 A random variable z ∈ Rp is said to have a multivariate generalized
asymmetric Laplace distribution with parameters s > 0, µ ∈ Rp and Σ ∈ Sp+ if its
characteristic function is
!s
1
p
.
∀u ∈ R , φGALp (Σ,µ,s) (u) =
1 + 12 uT Σu − iµT u
It is denoted by z ∼ GALp (Σ, µ, s).
General properties of the generalized asymmetric Laplace distribution distribution are discussed by Kozubowski, Podgórski, and Rychlik (2013). We list here a few useful ones.
Proposition 2 Let s > 0, µ ∈ Rp and Σ ∈ Sp+ . If z ∼ GALp (Σ, µ, s), we have E(z) = sµ
and Cov(z) = s(Σ + µµT ). Moreover, if Σ is positive definite, the density of z is given by
T −1
2eµ Σ x
QΣ (x) s−p/2
p
√
∀x ∈ R , fz (x) =
Ks−p/2 (QΣ (x)C(Σ, µ)) ,
(2π)p/2 Γ(s) det Σ C(Σ, µ)
√
p
where QΣ (x) = xT Σ−1 x, C(Σ, µ) = 2 + µT Σ−1 µ and Ks−p/2 is the modified Bessel
function of the second kind of order s − p/2.
Note that the GAL1 (2b2 , 0, 1) case corresponds to a centered univariate Laplace distribution
with scale parameter b > 0. In the symmetric case (µ = 0) and when s = 1, we recover the
multivariate generalization of the Laplace distribution of Anderson (1992).
An appealing property of the multivariate generalized Laplace distribution is that it is
also endowed with a multivariate counterpart of the Gauss-Laplace representation.
2
Theorem 3 (Generalized Gauss-Laplace representation) Let s > 0, µ ∈ Rp and Σ ∈
Sp+ . If u ∼ Gamma(s, 1) and x ∼ N (0, Σ) is independent of u, we have
√
uµ + ux ∼ GALp (Σ, µ, s).
(1)
A proof of this result can be found in Kotz, Kozubowski, and Podgórski (2001, chap. 6).
This representation explains why the multivariate generalized Laplace distribution can also
be seen as a multivariate generalization of the variance-gamma distribution which is widely
used in the field of quantitative finance (Madan, Carr, and Chang, 1998). Infinite mixtures
similar to (1) are called variance-mean mixtures (Barndorff-Nielsen, Kent, and Sørensen,
1982) and are discussed for example by Yu (2017).
Another useful property of the multivariate generalized Laplace distribution is that,
under some conditions, it is closed under convolution.
Proposition 4 Let s1 , s2 > 0, µ ∈ Rp and Σ ∈ Sp+ . If z1 ∼ GALp (Σ, µ, s1 ) and z2 ∼
GALp (Σ, µ, s2 ) are independant random variables, then
z1 + z2 ∼ GALp (Σ, µ, s1 + s2 ).
(2)
Proof Since z1 and z2 are independent, we have for all u ∈ Rp ,
φz1 +z2 (u) = φGALp (Σ,µ,s1 ) (u)φGALp (Σ,µ,s2 ) (u) =
1
1 T
1 + 2 u Σu − iµT u
!s1 +s2
which is the characteristic function of the GALp (Σ, µ, s1 + s2 ) distribution.
3. A new characterization involving a product between a Gaussian
matrix and a Gaussian vector
We now state our main theorem, which gives a new characterization of multivariate generalized Laplace distributions with half-integer shape parameters.
Theorem 5 Let W be a p × d random matrix with i.i.d. columns following a N (0, Σ)
distribution, y ∼ N (0, Id ) be a Gaussian vector independent from W and let µ ∈ Rp . We
have
Wy + ||y||22 µ ∼ GALp (2Σ, 2µ, d/2).
(3)
Proof For each k ∈ {1, ..., d} let wk be the k-th column of W, uk = yk2 and ξ k = yk wk +yk2 µ.
To prove the theorem, we will prove that ξ 1 , ..., ξ d follow a GAL distribution and use the
decomposition
d
X
ξk .
Wy + ||y||22 µ =
k=1
Let k ∈ {1, ..., d}. Since y is standard Gaussian, uk = yk2 follows a χ2 (1) distribution, or equivalently a Gamma(1/2, 1/2) distribution. Therefore, uk /2 ∼ Gamma(1/2, 1).
3
Moreover, note that
√
d
uk wk = |yk |wk = yk sign(yk )wk = yk wk since |yk | and sign(yk ) are
d
independent and sign(yk )wk = wk . Therefore, according to the generalized Gauss-Laplace
representation, we have
r
uk √
uk
d
ξk =
2wk + 2µ ∼ GALp (2Σ, 2µ, 1/2).
2
2
Since ξ 1 , ..., ξ d are i.i.d. and following a GALp (2Σ, 2µ, 1/2) distribution, we can use
Proposition 4 to conclude that
Wy + ||y||22 µ =
d
X
k=1
ξ k ∼ GALp (2Σ, 2µ, d/2).
In the symmetric case (µ = 0), this result gives the distribution of the product between
a Gaussian matrix with i.i.d. columns and a isotropic Gaussian vector.
Corollary 6 Let W be a p × d random matrix with i.i.d. columns following a N (0, Σ)
distribution and let y ∼ N (0, αId ) be a Gaussian vector independent from W. Then
Wy ∼ GALp (2αΣ, 0, d/2).
(4)
Moreover, if u is a standard Gamma variable with shape parameter d/2 and if x ∼ N (0, 2αΣ)
is a Gaussian vector independent of u, then
d
Wy =
√
ux.
(5)
Less general versions of Theorem 5 have been proven in the past, dating back to the
derivation of the inner product of two i.i.d. standard Gaussian vectors by Wishart and Bartlett
(1932). In particular, the unidimensional case (p = 1) was recently proven by Gaunt (2014)
in order to obtain bounds for the convergence rate of random sums involving Gaussian
products. The multivariate symmetric isotropic case (µ = 0 and Σ proportional to Ip ) was
proven by Bouveyron, Latouche, and Mattei (2016) in order to derive the marginal likelihood of the noiseless probabilistic principal component analysis model of Roweis (1998).
While the proof of Bouveyron, Latouche, and Mattei (2016) relied on characteristic functions and the properties of Bessel functions, the proof that we presented here is closer in
spirit to the one of Gaunt (2014), based on representations of distributions.
4. Perspectives
The new characterization presented in this paper may notably prove useful in two contexts.
First, it indicates a new way of handling situations involving the product of a Gaussian matrix and a Gaussian vector. An important instance is the Bayesian factor analysis
model (Lopes and West, 2004), of which principal component analysis is a particular case.
4
In this framework, the marginal distribution of the data, which is essential for model selection purposes, can be derived using representation (5) together with the Gauss-Laplace
representation (Bouveyron, Latouche, and Mattei, 2016, 2017).
Moreover, our characterization offers a means to get around problems encountered when
dealing with distributions related to the GAL distribution. For example, representation
(3) might lead to alternative estimation strategies for some problems related to portfolio allocation (Mencía and Sentana, 2009; Breymann and Lüthi, 2013) or cluster analysis
(McNicholas, McNicholas, and Browne, 2013; Franczak, Browne, and McNicholas, 2014).
Acknowledgements
I thank Charles Bouveyron, Pierre Latouche, Brendan Murphy and Christian Robert for
fruitful advices and discussions. Part of this work was made during a visit to University
College Dublin, funded by the Fondation Sciences Mathématiques de Paris (FSMP).
References
D. N. Anderson. A multivariate Linnik distribution. Statistics & Probability Letters, 14(4):
333–336, 1992.
O. Barndorff-Nielsen, J. Kent, and M. Sørensen. Normal variance-mean mixtures and z
distributions. International Statistical Review/Revue Internationale de Statistique, 50(2):
145–159, 1982.
C. Bouveyron, P. Latouche, and P.-A. Mattei. Bayesian variable selection for globally sparse
probabilistic PCA. Technical report, HAL-01310409, Université Paris Descartes, 2016.
C. Bouveyron, P. Latouche, and P.-A. Mattei. Exact dimensionality selection for Bayesian
PCA. Technical report, HAL-01484099, Université Paris Descartes, 2017.
W. Breymann and D. Lüthi. ghyp: A package on generalized hyperbolic distributions.
Manual for R Package ghyp, 2013.
P. Ding and J. K. Blitzstein. On the Gaussian mixture representation of the Laplace distribution. The American Statistician, in press, 2017.
T. Eltoft, T. Kim, and T.-W. Lee. On the multivariate Laplace distribution. IEEE Signal
Processing Letters, 13(5):300–303, 2006.
B. C. Franczak, R. P. Browne, and P. D. McNicholas. Mixtures of shifted asymmetric
Laplace distributions. IEEE Transactions on Pattern Analysis and Machine Intelligence,
36(6):1149–1157, 2014.
R. E. Gaunt. Variance-Gamma approximation via Stein’s method. Electronic Journal of
Probability, 19(38):1–33, 2014.
S. Kotz, T. Kozubowski, and K. Podgórski. The Laplace distribution and generalizations: a
revisit with applications to communications, exonomics, engineering, and finance. Springer
Science & Business Media, 2001.
5
T. Kozubowski, K. Podgórski, and I. Rychlik. Multivariate generalized Laplace distribution
and related random fields. Journal of Multivariate Analysis, 113:59–72, 2013.
A. Kyprianou. Fluctuations of Lévy processes with applications: Introductory Lectures.
Springer Science & Business Media, 2014.
H. F. Lopes and M. West. Bayesian model assessment in factor analysis. Statistica Sinica,
pages 41–67, 2004.
D. B. Madan, P. P. Carr, and E. C. Chang. The variance gamma process and option pricing.
European Finance Review, 2(1):79–105, 1998.
D. McGraw and J. Wagner. Elliptically symmetric distributions. IEEE Transactions on
Information Theory, 14(1):110–120, 1968.
S. M. McNicholas, P. D. McNicholas, and R. P. Browne. Mixtures of variance-gamma
distributions. arXiv preprint arXiv:1309.2695, 2013.
J. Mencía and E. Sentana. Multivariate location–scale mixtures of normals and mean–
variance–skewness portfolio allocation. Journal of Econometrics, 153(2):105–121, 2009.
T. Park and G. Casella. The Bayesian lasso. Journal of the American Statistical Association,
103(482):681–686, 2008.
S. Roweis. EM algorithms for PCA and SPCA. Advances in Neural Information Processing
Systems, 10:626–632, 1998.
J. Wishart and M. S. Bartlett. The distribution of second order moment statistics in a
normal system. Mathematical Proceedings of the Cambridge Philosophical Society, 28(4):
455–459, 1932.
Y. Yu. On normal variance–mean mixtures. Statistics & Probability Letters, 121:45–50,
2017.
6
| 10math.ST |
On Quantitatively Measuring Controllability of
Complex Networks
CAI, Ning1, 2
1
College of Electrical Engineering, Northwest University for Nationalities, Lanzhou, China
2
School of Automation, Hangzhou Dianzi University, Hangzhou, China
E-mail: caining91@tsinghua.org.cn
Abstract: This letter deals with the controllability issue of complex networks. An index
is chosen to quantitatively measure the extent of controllability of given network. The
effect of this index is observed based on empirical studies on various classes of network
topologies, such as random network, small-world network, and scale-free network.
Key Words: Controllability; Complex Network; Quantitative Measure
1. Introduction
The controllability of a dynamic system reflects the ability of influencing from
external input information to the motion of the overall system. Observability and
controllability are dual alternatives. Integrated with stability, they form the theoretical
foundation for most of the systems analysis and synthesis problems. Therefore,
controllability has been one of the most important concepts in modern control theory.
Since about a decade ago, the controllability problems of dynamic networked
large-scale systems have intrigued many scholars from both the control [1-7] and the
physics [8-13] community, and will surely continue to attract the attention from more
and more disciplines.
~1~
Tanner [1] earlier studied the controllability of systems with single leader and
conjectured that excessive connectivity might even be detrimental to controllability,
while giving a definition of graph controllability based on a partition of the associated
Laplacian matrix. Paying attention to the relationship between the extent of symmetry
of graph and controllability, Rahmani and Mesbahi [2] further extended the results in
[1]. Cai et al. addressed the controllability problems of a class of high-order systems,
proposing a scheme of controllability improvement [3-4]. Liu et al. [5] concerned the
controllability of discrete-time systems with switching graph topologies. Ji et al. [6-7]
dealt with the interactive protocols, endeavoring to integrate the influence of three
facets upon controllability, which are the protocol, the vertex dynamics and the
network topology, respectively.
Liu et al. [8] addressed the structural controllability of complex networks. They
selected an index denoted by ND to quantitatively measure the extent of controllability
of complex networks, namely the least amount of independent input signals required.
Along the route of [8], there have emerged plenty of researches from the physics
community, e.g. [9-13]. Particularly, Yan et al. [9] concentrated upon the problem of
minimal energy cost for maneuvering the nodes. Yuan et al. concerned the exact
controllability [10] of undirected networks with identical edge weights and discovered
certain consistency with [8].
The concept of controllability for dynamic systems was originally raised by
Kalman, with a set of algebraic criteria to check whether or not a given system is
exactly controllable. It has formed the foundation of the controllability theory.
However, there are two intrinsic problems limiting the study of the controllability of
complex networks from the viewpoint of exact controllability. The first problem is
that almost any arbitrary system is controllable in the sense of exact controllability.
This fact reduces the significance of being controllable. As the second problem, it is
rather difficult to translate those algebraic criteria into straightforward conditions for
the topologies of networks.
In comparison with exact controllability, the concept of structural controllability
possesses some advantages for coping with the controllability problems of networked
systems. First, it bears no ambiguity like almost exact controllability. Being
structurally controllable or not is essentially distinct for any system. Second, it is
possible to acquire concise and straightforward criteria about the topologies of
networks to check whether or not they are structurally controllable. Nevertheless, the
essence of structural controllability is a minimal requirement for the availability of
~2~
input information across the network, which is only a necessary prerequisite for
controlling, whereas it cannot facilitate to evaluate the efficiency of control. Therefore,
structural controllability is also restrictive.
In the current letter, a third angle on controllability is addressed, other than the
exact controllability and structural controllability. We shall endeavor to explore the
possible way to measure the extent of controllability of any given network,
quantitatively. It is motivated by a wish to overcome some of the limitations of exact
controllability and structural controllability, and to extend the methodology for
controllability analysis of networks, from qualitative to quantitative. An index will be
proposed for evaluating the ability of dynamic network to be easily controllable via
the input information, and simulations will be performed on three distinct types of
complex networks, namely the E-R networks, the WS small-world networks, and the
scale-free networks, to illustrate the diversity of controllability situations.
The remaining part of this letter is organized as follows. Section 2 introduces the
fundamental preliminaries about controllability of complex networks and describes
the model concerned. Section 3 endeavors to analyze the computational controllability
concept theoretically, based on some simple examples. Section 4 presents a series of
experimental results toward three typical complex networks. Finally, Section 5
concludes this letter.
2. Problem Description and Preliminaries
2.1 Exact Controllability
The dynamics of a network concerned with N f followers and N l leaders is
described by the following equation:
Aff Afl
where the vector (t ) R
Nf
(1)
represents the states of the follower vertices; (t ) R Nl
represents the states of the leader vertices, which can be determined externally; and
A R N N is the weighted adjacency matrix of the overall network topology, which is
decomposed as follows according to the leader-follower distribution
Aff
A
Alf
Afl
All
The network is exactly controllable if the state values of the followers can be
~3~
fully controlled via appropriately selected state values of leaders, otherwise, it is
uncontrollable. This
conforms
to
the conventional
definition
of Kalman
controllability.
Definition 1: A dynamic network (1) is exactly controllable if for any initial
values of follower states
1 (0), 2 (0),..., N (0) R
f
there exist and proper leader signals
1 (t ),2 (t ),...,N (t ) ( t [0, ] )
l
such that 1 ( ) 2 ( ) ... Nl ( ) 0 .
Lemma 1 below provides the most fundamental criterion to check controllability,
known as the rank test.
Lemma 1 [14]: A dynamic network (1) is exactly controllable if and only if the
controllability matrix is of full rank, which is
Afl
Aff Afl
A2ff Afl
N 1
Aff f Afl
Remark 1: If the number of leader Nl 1 , then the controllability matrix is
square. In this case, the network is controllable if and only if is nonsingular.
2.2 Almost Exact Controllability
The classic notion of exact controllability of dynamic systems is qualitative, and
could answer only “Yes” or “No”. However, it is well-known by the control theory
community that almost any linear system is controllable, i.e. the probability for any
system (1) to be exactly controllable is 1, so long as the matrix A is continuous-valued;
whilst the uncontrollable cases are just the rare exceptions [4, 15-16]. In this sense,
the exact controllability is less significant, whereas a better way to evaluate the extent
of controllability should be to measure how far a given network is from being
uncontrollable [15-16]. Such a route may be called computational controllability.
The main connotation behind the concept of computational controllability can be
illustrated by Fig. 1.
~4~
Uncontrollable
Set
Overall
Network
Space
0
Measure of
Controllability
Particular
Network
Fig. 1. Network space of R N N .
Solid curves represent the set of networks that cannot be exactly
controlled.
In Fig. 1, any network is represented by a point in the space. Almost all the
networks are exactly controllable. If a particular network happens to locate at the
uncontrollable set, then any infinitesimal perturbation could let it leave this set and
become controllable.
3. Computational Controllability of Complex Networks: Theoretical Analysis
3.1 Introduction of Concept
Almost all the dynamic networks are controllable. Nevertheless, the relative
controllability should be different for various network topologies. Some may possess
relatively higher controllability, whereas some others may be less controllable as
being very close to an uncontrollable network.
According to Lemma 1, controllability is determined by the rank of . If the
rank is nearly reduced, then the controllability of the associated network should be
low. If is square as the single leader case, it is expected to be away from being
singular for higher controllability. Thus, it should be meaningful to measure the extent
of its nonsingularity. In this paper, we choose the conditional number of as this
measure.
Definition 2: The conditional number of a square matrix is
cond () 1
~5~
Conditional number can be regarded as a measure of nonsingularity, which varies
within [1, ] . cond () 1 if is identity matrix, with the maximal possible
nonsingularity. Oppositely, if is nearly singular, then cond () must be large
because 1 is large. The greater the conditional number, the closer the matrix
is to be singular, till cond () as the rank of becomes reduced. A definition
of computational controllability hereby arises, based on conditional number of
controllability matrix.
Definition 3: The computational controllability of a dynamic complex network (1)
can be measured by
1
cond ( )
where is the controllability matrix.
Remark 2: For any network, [0,1] . The greater this value, the higher
controllability a network achieves. A network is uncontrollable if 0 , otherwise, it
is controllable. However, if is very small, the controllability is still quite weak
although the network is thought to be controllable in the conventional sense.
3.2 Basic Analysis
Let us start the theoretical analysis on computational controllability from
investigating several concrete typical instances.
Example 1: Consider the path PN . For simplicity, let N 5 and let the fifth
agent be the only leader. It is always controllable in the sense of exact controllability.
The network topology is illustrated in Fig. 2.
Fig. 2. Fifth order path P5 with default arc weight 1.
If this network is directed, its adjacency matrix is
0 1 0 0 0
0 0 1 0 0
A 0 0 0 1 0
0 0 0 0 1
0 0 0 0 0
~6~
which is already decomposed by dotted lines. The controllability matrix is just
the identity matrix, with 1 , indicating a maximal controllability.
If this network is undirected with the adjacency matrix
0 1 0 0 0
1 0 1 0 0
A 0 1 0 1 0
0 0 1 0 1
0 0 0 1 0
the controllability matrix is
0
0
0
1
0 0 1
0 1 0
1 0 2
0 1 0
The nonsingularity of this matrix is also quite satisfactory with cond () 5.8284
and 0.1716 , indicating a rather high controllability.
Example 2: Consider the complete network C5 . The adjacency matrix is
0 1 1 1 1
1 0 1 1 1
A 1 1 0 1 1
1 1 1 0 1
1 1 1 1 0
with the corresponding controllability matrix being
1
1
1
1
3 9 27
3 9 27
3 9 27
3 9 27
It is singular with the rank being only 1. Thus, is definitely zero.
Example 3: Consider the random weighted directed network, which is known to
be almost controllable in the conventional sense. A random network of 8th order is
generated, with the last agent assigned as the leader. The resulted controllability
matrix is
~7~
0.1897
0.1934
0.6822
0.3028
0.5417
0.1509
0.6979
0.8962 2.7716
8.6052
27.0802
84.4664
1.9397 5.3306 17.1429 53.1950 166.8824
1.0615 3.9149 12.0033 37.6166 117.6977
1.0436 2.9809
9.6179
29.8652
93.5949
1.1522 3.9331 11.9725 37.7541 117.8886
1.5067 3.9574 12.9025 39.9819 125.3981
1.0432 4.1353 12.2815 38.9403 121.4206
264.5718
521.6511
368.1719
292.6930
368.9820
392.0307
380.1693
(2)
In this case, cond () 1.5075 106 , and 6.6335 107 . Such a network
obviously holds a lower controllability.
The controllability of a network is closely relational to its spectral distribution, by
which the reason for lower controllability of Example 3 could partially be explained.
Note that the average values of column elements keep on increasing from left to
right in (2). Such a phenomenon will greatly deteriorate the nonsingularity of the
overall matrix, according to the following lemma.
Lemma 2: The infinite conditional number of square matrix A [aij ] R N N is
generally greater than N , with 1 being the ratio between the maximal and
minimal row/column sum of absolute values of elements.
Proof: If A is singular, cond ( A) is infinite. Consider the case that A is
nonsingular. Without loss of generality, let the rows be sorted according to the sums of
absolute values of elements, with the first row possessing the minimal sum, and
1
N
a11 j 1 a1 j
N
Now consider the matrix A / a11 , which has obviously the same conditional number
with A. Because the sum of absolute values of elements in the first row of A / a11 is
N
j 1
( a1 j / a11 ) N
the sum of absolute values of elements in the last row of A / a11 is N and as a
result
A / a11
N
The inverse of matrix A / a11 can be computed through a series of elementary row
transformations to the following augmented matrix:
~8~
a11 / a11
a21 / a11
aN 1 / a11
a12 / a11
a22 / a11
a1N / a11
a2 N / a11
1 0
0 1
aN 2 / a11
aNN / a11
0 0
0
0
1
Evidently, the absolute value of the element of ( A / a11 )1 with index (1, 1) is 1, and
it follows that
( A / a11 )1
1 □
Remark 3: The bound N given in Lemma 2 is quite conservative. Usually, the
conditional number of a matrix is far greater than this bound. Besides, although there
are different definitions of conditional number, generally their trends are consistent. It
is not possible that the conditional number is very large by one definition but very
small by another.
The spectrum of Aff in Example 3 is
{3.1285, 0.4372 0.445i, 0.3542 0.2556i, 0.7864, 0.7592}
Such a distribution is typical for the class of random networks in Example 3, with one
largest eigenvalue and the remaining being very smaller.
It is known that any random matrix is almost diagnolizable [17]. Suppose that
Aff could be diagnolized as
1
P 1 Aff P
N f
It follows that
1k
Akff P
1
P
k
N f
The magnitude of Akff of Example 3 increases with k because 1 is much larger
than 1.
The controllability can be changed by scaling the spectral radius through zooming
the topology of network, without affecting the structure.
Consider the controllability of A / , where 0 is the zoom factor. Take the
data in Example 3 and let the zoom factor vary across [1, 4] ; the corresponding
controllability variation forms a quite smooth curve, as shown in Fig. 3. One can see
that, at first, the controllability increases as increases, and it reaches a maximum
~9~
when 2.4 ; then the controllability keeps on decreasing because the magnitude of
Akff decreases with k if 1 is less than 1.
Intuitively speaking, the controllability of a network should generally be better if
the eigenvalues locate nearby the unit circle. The spectral radius being either greatly
larger or less than 1 will deteriorate the controllability.
-5.3
-5.4
-5.5
-5.6
-5.7
-5.8
-5.9
-6
-6.1
1
1.5
2
2.5
3
3.5
4
ZoomFactor
Fig. 3. Trend of controllability in Example 3 with zoom factor varying across [1, 4]
4. Computational
Controllability
of
Complex
Networks:
Empirical Analysis
This section will manifest a series of empirical analysis on computational
controllability for the three typical classes of complex networks.
4.1 E-R Network
Consider the E-R networks comprising of 15 followers and 1 leader. Let the
connectivity probability p vary across [0, 0.9]. The value of each data point is the
average of 2000 random trials. The variation of controllability is shown in Fig. 4. One
can see that the logarithm variation trend could be fitted into two straight lines.
It is worth mentioning that throughout this section, the edge weights are mounted
with noises uniformly distributed within [-0.025, 0.025], which are relatively rather
weak.
~ 10 ~
-14
-16
-18
-20
-22
-24
-26
-28
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
p
Fig. 4. Trend of controllability of E-R network with connectivity probability
varying across [0, 0.9]
The controllability achieves a maximum when about p 0.15 . If the value of p
is too low, the network lacks sufficient connectivity to guarantee controllability.
However, the controllability keeps on decreasing as p is large enough. The more a
network approximates a complete graph, the less controllable it would be.
Now fix the connectivity probability p to 0.15 and let the number of leaders N l
vary across 1, 2, …, 15. The variation of controllability is shown in Fig. 5. There is a
positive correlation between the controllability and N l . However, such a correlation
becomes weak as the number of leaders is sufficiently large. For this example,
log( ) seems to approach a limit being less than 10.
-10
-10.5
-11
-11.5
-12
-12.5
-13
-13.5
-14
-14.5
-15
0
5
10
15
Nl
Fig. 6. Trend of controllability of E-R network with number of leaders varying
across 1, 2, …, 15
~ 11 ~
Fix the connectivity probability p to 0.15 and the number of leaders N l to 1. Let
the number of followers vary across 5, 6, …, 104, 105. According to the result, the
controllability is negatively correlated to the number of followers, as shown in Fig. 6.
One can see that the controllability of a network with large-scale is quite weak.
0
-20
-40
-60
-80
-100
-120
-140
-160
0
20
40
60
80
100
120
Nf
Fig. 7. Trend of controllability of E-R network with number of followers
varying across 5, 6, …, 104, 105
4.2 WS Small-World Network
Consider the WS small-world networks containing 16 followers, with K 2 .
First, let the switching probability p vary across [0, 0.9]. However, no direct
relationship can be seen between the controllability and p, according to the simulation
result.
Set the switching probability p to 0.5 and successively increase the quantity of
leaders N l from 1 to 15, with the quantity of followers N f being fixed. A resulted
relationship between the controllability and N l is shown in Fig. 8. It can be seen that
more leaders are generally beneficial to the controllability, although this effect is less
prominent if Nl 1 .
~ 12 ~
-8
-9
-10
-11
-12
-13
-14
-15
0
5
10
Nl
15
Fig. 8. Trend of controllability of WS small-world network with number of
leaders varying across 1, 2, …, 15
Let p 0.5 , Nl 1 , and the quantity of followers be 5, 6, …, 85. The
relationship between the controllability and N f is illustrated in Fig. 9, which may be
approximately fitted into a straight line.
0
-10
-20
-30
-40
-50
-60
0
10
20
30
40
50
60
70
80
90
Nf
Fig. 9. Trend of controllability of WS small-world network with number of
followers varying across 5, 6, …, 85
Let p 0.5 , Nl 1 , and N f 30 ; successively increase the K from 2 to 15. It
should be interesting to see that although being independent of p, the controllability is
negatively correlated to K, as shown in Fig. 10.
~ 13 ~
-20
-25
-30
-35
-40
-45
2
4
6
8
10
12
14
16
K
Fig. 10. Trend of controllability of WS small-world network with K varying
across 2, 3, …, 15
4.3 BA Scale-Free Network
Consider the controllability of BA scale-free networks. First, let the order of the
original complete graph m0 7 , the number of added vertices t 8 , and the number
of new edges per vertex m 3 . The resulted network topology is illustrated in Fig. 11,
with the indices of vertices sorted by degree.
11
1
5
10
3
12
9
2
8
15
14
13
4
7
6
Fig. 11. BA scale-free network with m0 = 7, m = 3, and t = 8
Select each vertex as the single leader, one by one, with the computed controllability
of network successively shown in Fig. 12. A slight tendency of positive correlation
between the controllability and the degree of leader vertex can be observed.
~ 14 ~
-14
7
x 10
6
5
4
3
2
1
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
indexof leader
Fig. 12. Trend of controllability of BA scale-free network with degree of
leader decreasing
Fix m0 7 and t 8 ; meanwhile, let m vary across 1, 2, …, 6. For each case,
the average controllability of 2000 trials is recorded and displayed in Fig. 13.
Evidently, the controllability is positively correlated to m.
-10
-15
-20
-25
1
2
3
m
4
5
6
Fig. 13. Trend of controllability of BA scale-free network with m varying
across 1, 2, …, 6
5. Conclusion
Considering the fact that weighted real-world networks are almost controllable in
the sense of exact controllability, this letter addresses the problem of measuring the
~ 15 ~
controllability of complex networks quantitatively. For this end, an index is selected,
based on the conditional number of controllability matrix. The effect of this index is
observed and discussed, mainly by a series of experiments on various types of
complex networks, e.g. the E-R networks, the WS small-world networks, and the BA
scale-free networks.
Studies along this route could enable the comparisons of controllability between
complex networks with different topologies. The current letter just takes a first step.
There exist abundant directions of potential future extensions, for instance: 1) Other
indices beside the one here can be sought; 2) More experiments can be conducted to a
broader range of networks, for the sake of discovering more empirical regularities; 3)
Certain phenomena observed could be explained analytically. 4) The network model
concerned can take other forms, e.g. being represented by Laplacian matrix.
Acknowledgments
This work is supported by National Natural Science Foundation (NNSF) of China
(Grants 61263002 & 61374054), by Program for Young Talents of State Ethnic
Affairs Commission (SEAC) of China (Grant [2013] 231), and by the Zhejiang Open
Foundation of the Most Important Subjects.
References
[1] H. G. Tanner, “On the controllability of nearest neighbor interconnections”, Proc. 43rd IEEE
Conf. Decision and Control, vol. 3, pp. 2467-2472, 2004.
[2] A. Rahmani, M. Ji, and M. Mesbahi et al., “Controllability of multi-agent systems from a
graph-theoretic perspective”, SIAM J. Control Optim., vol. 48, no.1, pp. 162-186, 2009.
[3] N. Cai, J.-X. Xi, and Y.-S. Zhong et al., “Controllability improvement for linear
time-invariant dynamical multi-agent systems”, Int. J. Innov. Comput. I., vol. 8, no. 5a, pp.
3315-3328, 2012.
[4] N. Cai, J.-W. Cao, and M. J. Khan, “A controllability synthesis problem for dynamic
multi-agent systems with linear high-order protocol”, Int. J. Control Automat. Syst., vol. 12,
no. 6, pp. 1366-1371, 2014.
[5] B. Liu, T.-G. Chu, and L. Wang et al., “Controllability of a leader-follower dynamic network
with switching topology”, IEEE Trans. Autom. Control, vol. 53, no.4, pp. 1009-1013, 2008.
[6] Z.-J. Ji, H. Lin, and H.-S. Yu, “Leaders in multi-agent controllability under consensus
algorithm and tree topology”, Syst. Control Lett., vol. 61, no. 9, pp. 918-925, 2012.
[7] Z.-J. Ji, H. Lin, and H.-S. Yu, “Protocols design and uncontrollable topologies construction
for multi-agent networks”, IEEE Trans. Automat. Control, vol. 60, no. 3, pp. 781-786, 2015.
~ 16 ~
[8] Y.-Y. Liu, J. J. Slotine, and A. L. Barabási, “Controllability of complex networks”, Nature, vol.
473, pp. 167-173, 2011.
[9] G. Yan, J. Ren, and Y.-C. Lai et al., “Controlling complex networks: How much energy is
needed?”, Phys. Rev. Lett., vol. 108, 218703, 2012.
[10] Z.-Z. Yuan, C. Zhao, and Z.-R. Di et al., “Exact controllability of complex networks”, Nat.
Commn., vol. 4, 2013.
[11] T. Jia, Y.-Y. Liu, and E. Csóka et al., “Emergence of bimodality in controlling complex
networks”, Nat. Commun., vol. 4, 2002, 2013.
[12] Y.-Y. Liu, J. J. Slotine, and A. L. Barabási, “Control centrality and hierarchical structure in
complex networks”, PLoS ONE, vol. 7, no. 9, e44459, 2012.
[13] J. Ruths and D. Ruths, “Control profiles of complex networks”, Science, vol. 21, pp.
1373-1376, 2014.
[14] T. Kailath, Linear Systems, Englewood Cliffs: Prentice-Hall, 1980.
[15] C. C. Paige, “Properties of numerical algorithms relating to computing controllability”, IEEE
Trans. Automat. Control, vol. 26, no. 1, pp. 130-138, 1981.
[16] D. L. Boley and W.-S. Lu, “Measuring how far a controllable system is from an
uncontrollable one”, IEEE Trans. Automat. Control, vol. 31, no. 3, pp. 249-251, 1986.
[17] N. Cai, J.-W. Cao, and M. J. Khan, “Almost decouplability of any directed weighted network
topology”, Physica A, vol. 436, pp. 637-645, 2015.
~ 17 ~
| 3cs.SY |
1
Progressive Joint Modeling in Unsupervised
Single-channel Overlapped Speech Recognition
arXiv:1707.07048v2 [cs.CL] 20 Oct 2017
Zhehuai Chen, Student Member, IEEE,, Jasha Droppo, Senior Member, IEEE,, Jinyu Li, Member, IEEE,,
Wayne Xiong, Member, IEEE,,
Abstract—Unsupervised single-channel overlapped speech
recognition is one of the hardest problems in automatic speech
recognition (ASR). Permutation invariant training (PIT) is a state
of the art model-based approach, which applies a single neural
network to solve this single-input, multiple-output modeling
problem. We propose to advance the current state of the
art by imposing a modular structure on the neural network,
applying a progressive pretraining regimen, and improving the
objective function with transfer learning and a discriminative
training criterion. The modular structure splits the problem into
three sub-tasks: frame-wise interpreting, utterance-level speaker
tracing, and speech recognition. The pretraining regimen uses
these modules to solve progressively harder tasks. Transfer
learning leverages parallel clean speech to improve the training
targets for the network. Our discriminative training formulation
is a modification of standard formulations that also penalizes
competing outputs of the system. Experiments are conducted on
the artificial overlapped Switchboard and hub5e-swb dataset. The
proposed framework achieves over 30% relative improvement of
WER over both a strong jointly trained system, PIT for ASR, and
a separately optimized system, PIT for speech separation with
clean speech ASR model. The improvement comes from better
model generalization, training efficiency and the sequence level
linguistic knowledge integration.
Index Terms—unsupervised single channel overlapped speech
recognition, permutation invariant training, progressive joint
training, transfer learning, sequence discriminative training
I. I NTRODUCTION
The cocktail party problem [1], [2], referring to multi-talker
overlapped speech recognition, is critical to enable automatic
speech recognition (ASR) scenarios such as automatic meeting
transcription, automatic captioning for audio/video recordings,
and multi-party human-machine interactions, where overlapping speech is commonly observed and all streams need to be
transcribed. The problem is still one of the hardest problems
in ASR, despite encouraging progresses [3], [4], [5], [6].
In this paper, we address the speech recognition problem
when multiple people speak at the same time and only a single
channel of overlapped speech is available. This is useful when
only a single microphone is present, or when microphone
Zhehuai (Tom) Chen is with the Computer Science and Engineering
Department, Shanghai Jiao Tong University, Shanghai 200240, China. He
conducted this work while working as an intern at Microsoft Research. (email:
chenzhehuai@sjtu.edu.cn)
Jasha Droppo, Jinyu Li and Wayne Xiong are with Microsoft AI and
Research, One Microsoft Way, Redmond, WA 98052, U.S.A. (e-mail:
{jdroppo,jinyli,wayne.xiong}@microsoft.com)
Corresponding authors are Jasha Droppo and Zhehuai Chen (email:
jdroppo@microsoft.com,chenzhehuai@sjtu.edu.cn phone: +1 425 703-7114).
array based algorithms fail to perfectly separate the speech.
Specifically, the paper focuses on an unsupervised inference
method, which does not need any prior knowledge of speakers.
To obtain transcriptions of all speakers from the overlapped
speech, joint inference is conducted based on multiple
knowledge sources: frequency domain voice discrimination,
temporal speaker tracing, linguistic information and speech
recognition.
Prior work in unsupervised single-channel overlapped
speech recognition generally separates the problem into
speech separation and recognition stages. Before the deep
learning era, the most popular speech separation technique
is computational auditory scene analysis (CASA) [3]. There
are two main stages in CASA approaches: segmentation and
grouping. The segmentation stage decomposes mixed speech
into time-frequency segments assumed to be derived from
the corresponding speakers based on perceptual grouping
cues [7]. The grouping stage simultaneously and sequentially
concatenates the segments to generate independent streams for
each speaker. Non-negative matrix factorization (NMF) [8] is
another popular technique which aims to learn a set of nonnegative bases that can be used to estimate mixing factors
during evaluation. Recently, several deep learning based
techniques have been proposed but seldom concentrate on the
unsupervised case, which is more applicable. In [9], [10], [11],
the authors propose deep clustering (DPCL), in which a deep
network is trained to produce spectrogram embeddings that are
discriminative for partition labels given in training data. The
model is optimized so that in the neural network embedding
space the time-frequency bins belonging to the same speaker
are closer and those of different speakers are farther away.
Speech segmentations are therefore implicitly encoded in the
embeddings, and can be obtained by clustering algorithm.
In [12], a DNN-based gender mixture detection system
and three gender-dependent speech separation systems are
constructed. The latter ones directly infer the feature streams
of two speakers respectively. For all these methods, speech
separation and recognition are two separate components and
the latter is applied to the separated feature streams. The
mismatched feature in the speech recognition stage is one of
the limitation in these methods.
In light of permutation invariant training (PIT) proposed
in speech separation [13] originally, the PIT-ASR model [14]
is the first attempt in joint modeling of unsupervised singlechannel mixed speech recognition. Whereas the original PIT
technique jointly models the voice discrimination and speaker
tracing, PIT-ASR further integrates speech recognition into the
2
neural network with a unified cross entropy (CE) criterion.
Although PIT-ASR shows promising results, it suffers from
several disadvantages, which are analyzed in Section II.
In this paper, progressive joint modeling is proposed
to divide the single channel overlapped speech recognition
problem into three sub-problems for initialization: frame-wise
interpreting, speaker tracing and speech recognition (Figure 1).
Each module is initialized by placing it into a series of
networks that solve progressively more difficult problems.
After the initialization, modules are jointly trained with two
novel strategies, namely self-transfer learning and multioutput sequence discriminative training. Transfer learning is
introduced in this problem, which leverages parallel clean
speech to improve the training targets for the network.
Our discriminative training formulation is a modification of
standard formulations, that also penalizes competing outputs
of the system. The proposed framework achieves 30% relative
improvement over both a strong jointly trained system, PITASR, and a separately optimized system, PIT for speech
separation with clean speech ASR. The improvement comes
from better model generalization, training efficiency and the
sequence level linguistic knowledge integration.
The rest of the paper is organized as follows. In Section II,
the unsupervised single-channel overlapped speech recognition
problem is briefly reviewed. In Section III-A, the modular
initialization and progressive joint training is proposed. In
Section III-B, the self-transfer learning is proposed and in
Section III-C, multi-output sequence discriminative training is
proposed. In Section IV, the experimental results are reported
in artificial overlapped Switchboard corpus and Eval2000
hub5e-swb test set, followed by the conclusion in Section V.
II. U NSUPERVISED S INGLE - CHANNEL OVERLAPPED
S PEECH R ECOGNITION
N
X
O(r)
un
(1)
n=1
(r)
P (Lu1 , ..., LuN |O(m)
u )≈
N
Y
(m)
P (L(r)
un |Ou )
(2)
n=1
Another branch of methods assume the overlapped signal can
(c)
(c)
(r)
be separated to Oun : Oun ≈ Oun . Because the original
(r)
streams Oun are conditionally independent with each other,
(c)
the separated signal streams Oun are also assumed to be
conditionally independent. Thus Equation (2) can be derived
to Equation (3),
P (Lu1 , ..., LuN |O(m)
u )≈
N
Y
(c)
P (L(r)
un |Oun )
(3)
n=1
However, neither assumption is precise. For the first
assumption, there is no pre-determined method to obtain the
ideal label arrangements, which is called the speaker tracing
problem. The second assumption is that the speech separation
and recognition are independent processes, which introduces
an artificial information bottleneck.
In [13], the reference streams are treated as an unordered
set. The PIT framework is proposed to address the speech
separation problem by firstly determining the assignment of
the reference stream and inference stream that minimizes the
error at the utterance level based on the forward-pass result.
This is followed by minimizing the error given the utterance
level best assignment. [14] extends this by integrating speech
recognition into the neural network with a unified crossentropy (CE) training criterion.
X 1 X
X
(s0 ) (r)
CE(lutn , lutn )
JCE-PIT =
min
(4)
s0 ∈S
N
t
u
n∈[1,N ]
Unsupervised single-channel overlapped speech recognition
refers to the speech recognition problem when multiple unseen
talkers speak at the same time and only a single channel
of overlapped speech is available. Different from supervised
mode, there’s not any prior knowledge of speakers in the
evaluation stage.
In the problem, only the linearly overlapped single-channel
signal is known, which is defined as,
O(m)
=
u
assumes the inference streams are conditionally independent,
(r)
and tries to assign the correct transcription stream Lun to the
corresponding output stream n,
where Oun is the clean signal stream of speaker n at utterance
(m)
u and Ou is the overlapped speech stream of utterance u.
N is the number of streams. Single channel ASR is always
formulated as a supervised sequence labeling problem given by
P (Lu |Ou ), which is the posterior probability of transcription
sequence Lu given the feature sequence Ou in the utterance
u. Nevertheless, the multi-speaker problem is to model the
joint distribution of N streams of transcriptions given the
(m)
overlapped speech signal, P (Lu1 , ..., LuN |Ou ). Due to the
symmetric labels given the mixture signals, it is no longer
a supervised optimization problem. One branch of methods
Here, S is the permutation set of the 0reference representation
(s )
and the inference representation. lutn is the n-th inference
(r)
label of permutation s0 at frame t in utterance u and lutn is
the corresponding transcription label obtained by clean speech
forced-alignment [15].
The PIT-ASR criterion [14] elegantly integrates speech
separation, speaker tracing and speech recognition together
as Figure 2(a). Its joint modeling approach eliminates the
artificial bottleneck between the speech separation and speech
recognition tasks. But the method suffers from several
disadvantages which deteriorates the performance:
• Previous PIT-ASR work attempts to solve three of the
most difficult problems in speech processing with one
large model. This oversized model can lead to poor
generalization and slow training.
• Speaker tracing and speech recognition are both sequence
modeling problems. PIT-ASR concatenates frame level
CE as the criteria of the sequence level problem, which
limits the modeling effect of the neural network.
• PIT-ASR with a monolithic model is difficult to integrate
with other speech separation and segmentation technologies. For instance, although linguistic information can be
valuable for solving the speaker tracing problem [6], there
is no natural way to incorporate it into PIT-ASR.
3
Fig. 1. The Proposed System Framework. The single monolithic structure (the dashed line box) that predicts independent targets for each speaker, proposed
in [14], is improved through modularization (three solid line boxes) and pretraining. Self-transfer learning and multi-output sequence discriminative training
are conducted on modularly initialized layers.
III. M ETHODS
In this work, we propose three separate enhancements to
improve the performance of PIT-ASR.
First, the structure and accuracy of the model is
improved through modularization and pretraining. Frame-wise
interpreting, speaker tracing, and speech recognition modules
replace the monolithic structures used in previous work. These
modules are progressively pretrained and jointly fine-tuned.
Second, we demonstrate a natural way to incorporate a form
of transfer learning. Clean speech features are used to generate
soft label targets which are interpolated with the reference
label alignments.
Third, multi-output discriminative training is applied to
the system. As with single-stream speech recognition, multistream discriminative training can help with model generalization. Additionally, the objective function is augmented to
reduce cross-speaker word assignment errors.
A. Modularization
In the original formulation, a PIT-ASR model consists of a
single monolithic structure that predicts independent targets for
each speaker. We improve this by replacing the main network
structure with a modular structure, shown in Figure 1.
This modular structure consists of three tasks, namely
interpreting mixed acoustic data, tracing speakers across
time, and predicting acoustic label sequences. First, the
frame-wise module is designed to extract the local timefrequency information necessary to separate the overlapped
speech into individual acoustic representations. It is entirely
local and does not depend on sequence-level information.
Second, the speaker tracing module accepts frame-wise
acoustic representations from the frame-wise module and
traces the speaker information. This process concatenates
adjacent acoustic representations of the same speaker together
to infer the recovered speech features of each speaker. Third,
the speech recognition modules accept the sequences of
recovered acoustic features from each speaker, and produce
sequences of label scores suitable for use in an automatic
speech recognition system. Because each speech recognition
module performs the same task, it is natural to share the
parameters of this module across each instance in the final
model.
Although it is possible to train the modularized network
of Figure 2(e) from random initialization, it is better to use
a progressive training strategy. This strategy is motivated by
the Curriculum learning theory in [16], and integrates both
modular initialization and joint training. We train a simple
model first, and then use it as a pre-trained building block
for a more complicated model and task. Figures 2(b)-(e)
illustrate how the model becomes progressively more complex
while solving more difficult problems, from frame-wise mean
squared error to whole utterance cross entropy.
Our simplest model, shown in Figure 2(b), is trained to solve
a frame-wise speech separation task. For N speakers, given the
(m)
mixed data Ou , the model infers an acoustic representation
outn for each speaker n at frame t of utterance u. The objective
function of the frame-wise training, is given as
X
XX 1
(s0 ) (r)
M SE(outn , outn ) (5)
min
JF-PIT =
0
N s ∈S
u
t
n∈[1,N ]
where, S is the permutation set of the reference
representation
(s0 )
(r)
and the inference representation. outn and outn is the
frame level acoustic representation of permutation s0 and
the reference clean speech, respectively. In each frame t of
the utterance u, the overall minimum square error, M SE,
is obtained by comparing all the reference and inference
representations of each permutation s0 .
The architecture for pre-training the speaker tracing module
is explained in Figure 2(c). The tracing module is combined
with a pre-trained frame-wise module that has had its N
output layers removed. As in [13], the PIT objective function
is applied in utterance level.
X
X 1 X
(s0 ) (r)
JU-PIT =
min
M SE(outn , outn )
(6)
0
s ∈S
N
u
t
n∈[1,N ]
The speech recognition module is separately pretrained in
the same way as a conventional acoustic model, with clean
speech and a cross-entropy objective function, maximizing
p(Lu |Ou ). This is illustrated in Figure 2(d).
The final model, shown in Figure 2(e), is created by stacking
the speech recognition modules onto the outputs of the pretrained speaker tracing and frame-wise modules. It is jointly
trained with an utterance level PIT-CE objective function given
in Equation (4).
Notably, even though the proposed structure has several
copies of the speech recognition module, the numbers of
parameters between Figure 2(a) and Figure 2(e) are similar.
Because the speech recognition modules are solving similar
problems, their parameters can be shared 1 . In preliminary
1 Namely
the “shared” option in CNTK “CloneFunction”.
4
•
UTTwise
PITCE
PIT-ASR
less time compared with the joint training 2 .
Potential to integrate with other technologies. State-ofthe-art technologies in each field can be applied to the
initialization of each module.
B. Transfer Learning Based Joint Training
(a) PIT-ASR Joint Training
Framewise
module
Framewise
PITMSE
CE
Speech
recognition
(d) Speech Recognition
(b) Frame-wise Interpreting
Framewise
module
Speaker
tracing
module
UTTwise
PITMSE
(c) Speaker Tracing
Framewise
module
Speaker
tracing
module
Speech
recognition
Speech
recognition
Transfer learning, or teacher-student training, is a powerful
technique to address domain adaptation problems in speech
recognition. In this section, we show that multi-channel ASR
is a type of domain adaptation, and that transfer learning can
be used to improve model training.
1) Conventional Transfer Learning in Domain Adaptation:
Transfer learning has been proposed to solve the distribution
mismatch problem in feature space [17]. To use this method
in single-output ASR domain adaptation, parallel data must
be available from a source domain and a target domain.
A fully trained model in the source domain (the teacher)
processes data and generates posterior probabilities, which are
sometimes referred to as “soft labels.” These soft labels then
replace or augment the usual “hard labels” when training the
student model with parallel data in the target domain [18].
To train the student, the Kullback-Leibler divergence (KLD)
between the output distributions of the teacher and student
models is minimized as below.
UTTwise
PITCE
(T )
KLD(y (T ) , y (S) ) =
log
X
[
(T )
yi
(T )
log yi
−
yi
(S)
yi
(T )
yi
(S)
log yi
(7)
]
i
(e) Final Joint Training
experiments, the performance gap between sharing and
independent parameters is less than 3%. Thus to make the
comparison fair, without specific explanation, the shared
structure is used.
The advantage of the progressive joint training includes:
• Decreased model complexity leading to better system
generalization and data efficiency. By separating system
into proper modules, the model complexity is less than the
all-in-one system in PIT-ASR. As unsupervised singlechannel overlapped speech recognition contains several of
the hardest components in speech processing, the model
complexity reduction is critical to the empirical training
performance.
• Faster convergence of the training process and better
quality of the local minimum. Curriculum learning
improves both the training speed and the performance of
the model [16]. As shown in Figure 4, the progressive
joint training needs fewer epochs to converge, and it
converges to a better local minimum. An additional
benefit is that the module initializations all take much
(T )
yi
i
=
Fig. 2. PIT-ASR Joint Training versus the Modular Initialization and
Progressive Joint Training. The dash-dot blocks indicate the learnable model
parameters. The dot-dot blocks indicate the learnable and shared model
parameters.
X
=
X
(T )
−yi
(S)
log yi
(8)
i
(T )
(S)
where yi and yi is the teacher and student distributions
respectively. Because the first term is not related to the
student model optimization, only the second term is used for
optimization. Comparing Equation (8) to CE criterion in ASR,
the hard labeling is replaced by the soft distribution inferred
from the source data by the teacher model.
2) Self-transfer Learning: In light of above discussion, selftransfer learning can be extended to the training of any multichannel speech recognition system. The student is, of course
the multi-channel speech recognition system. It operates in
the target domain of mixed speech acoustic data, and must
produce separate outputs for each speaker in the mixture. The
teacher also must produce separate outputs for each speaker,
but has access to the source domain: un-mixed clean speech.
The teacher model is a set of clean speech acoustic models
operating independently on the separate channels of clean
speech.
The self-transfer learning method then minimizes the KLD
between the output distribution of the mixed speech model and
the set of clean speech models. The KL divergence defined for
2 The training curve of PIT speech separation can be referred to [13].
Besides, initializations also can be done in parallel.
5
utterance level PIT training between the clean speech model
distribution and the joint model distribution is as below,
JKLD-PIT =
X
u
X 1
s ∈S
N
t
min
0
X
n∈[1,N ]
(9)
(c)
(s0 )
(m)
KLD(P (lutn |O(r)
un ), P (lutn |Ou ))
where the calculation of each KLD(·) pair is the same to
the adaptation-purpose single-channel case in Equation (8).
Namely,
the joint-trained model distribution, y (S) =
(s0 )
(m)
P (lutn |Ou ), is taken as the student model distribution, and
(c)
(r)
the clean speech model distribution, y (T ) = P (lutn |Oun ), is
taken as the teacher model distribution. It is notable that when
this method is applied to the modular structure proposed in this
work, as in Figure 3, the speech recognition modules can be
initialized with an exact copy of the teacher model.
The training framework for self-transfer learning is shown
in Figure 3. The soft targets generated by the teacher models
are interpolated with the hard labeling as in [19]. The training
procedure is as below:
1) Clone the speaker tracing layers in the bottom. Clone 2N
copies of clean ASR model initialized in Section III-A,
half for stacking upon the speaker tracing layers, half for
model inference given each clean speech stream.
(r)
2) Use simultaneous clean speech streams Oun and the
(m)
overlapped speech stream Ou to do joint training.
a) For each mini-batch, do forward propagation of the
clean ASR model using each clean speech stream
(c)
(r)
to calculate N streams of P (lutn |Oun ) respectively.
Do forward propagation of the joint model using
overlapped speech stream to0 calculate N streams of
(s )
(m)
inference distributions, P (lutn |Ou ).
b) For that mini-batch, calculate the error signal of
Equation (9) and then do back propagation for the joint
model.
c) Update parameters of the joint model and repeat until
convergence.
The proposed method elegantly solves the label mismatch
problem and helps the model convergence. Namely, using hard
labeling obtained from forced-alignment in the clean speech
is not proper, because the feature has been distorted in the
mixed speech. The proposed method replaces it with the soft
distribution. In addition, the proposed method formulates the
joint training of multi-channel ASR by domain adaptation
between clean speech and overlapped speech. Thus the soft
distribution also helps model convergence, because it’s easier
to recreate its performance, compared with training a speech
recognition model from scratch.
The evidence can be observed from the training curve
in Figure 4 that the initial CE of self-transfer learning
based progressive joint training is much better than that of
both joint modeling and progressive joint modeling. Notably,
the different starting points between the progressive joint
modeling and self-transfer learning based progressive joint
modeling is because the CE in the former system is calculated
versus hard labeling, while for the latter system it is versus
the soft distribution inferred from simultaneous clean speech 3 .
Thus with a better starting point and less parameter updating
requirement, finally the model also comes into better minimum
in the figure.
The relationships of the proposed method and previous
works are summarized as below.
• Model space adaptation. The formulation of the proposed
method is similar to KLD-based adaptation [21] and
teacher-student based domain adaptation [18]. In [21], to
conservatively update model parameters using adaptation
data, the target probability distribution is changed from the
ground truth alignment to a linear interpolation with the
distribution estimated from the unadapted model. In [18],
the feature mismatch problem in the target domain is solved
by minimizing the inference distribution divergence between
the target and source domains using parallel-data. The
reader is free to consider the proposed method as analogous
to optimizing student network in the target domain, i.e.
overlapped speech, to behave similarly to the well-trained
teacher network in the source domain, i.e. clean speech,
while bearing in mind the proposed method requires module
stacking because the motivation is to do joint training.
• Stereo piecewise linear compensation for environment
(SPLICE) [22]. The SPLICE algorithm uses stereo data
to do noise reduction and channel distortion compensation.
In [23], the clean feature is used for the teacher model to
provide supervision on the stereo noisy data trained student
model. In [19], the multi-channel enhanced feature is used
for the teacher model. In this work, the teacher-student
framework is also based on stereo data. The student model
is initialized better to cope with the more difficult modeling
problem, and the entire framework is expanded to handle
multiple output streams.
• Progressive stacking transfer learning. [24] proposes to
progressively conduct transfer learning to train speech
enhancement layers. The motivation of the progressive
stacking is only to gradually model a hard task by dividing
into several same but smaller tasks. Thus the criteria of all
tasks are the same. However, the proposed method is to do
joint training of distinct tasks. And each task is fully trained
with specific data and criterion.
• Self-supervised training. [25] proposes to use a teacher
model based on more accurate sensor information as
the supervision of the student model. The motivation of
the proposed method is different, which is to transfer
distribution between two models with feature mismatch.
Besides, the empirical procedure of the proposed method
is to fine-tune the original model in the target feature
distribution from supervision of itself in parallel source
distribution.
3) Learning from Ensemble: Following the transfer learning diagram, the joint model can also benefit from an ensemble
of teachers [26]. Specifically, because the problem includes
several sub-problems discussed in Section III-A, different
3 We didn’t use CE versus hard labeling for all systems, because CE between
inference distribution and the hard labeling is not always a good indicator of
the quality of the acoustic model [20], while the proposed curves better show
the optimization process.
6
(r)
Speech
recognition
+
Framewise
module
Speaker
tracing
module
Speech
recognition
UTTwise
PITCE
Speech
recognition
where Lu is the corresponding reference.
For the overlapped speech recognition problem, the
conditional independence assumption in the output label
streams is still made as in Equation (2). Then the cross-entropy
based PIT can be transformed to sequence discriminative
criterion based PIT as below,
JSEQ-PIT =
X
u
+
Speech
recognition
Fig. 3. Transfer Learning Based Progressive Joint Training. The dash-dot
blocks indicate the learnable model parameters. The dot-dot blocks indicate
the learnable and shared model parameters.
neural network structures can show different superiorities, e.g.,
with different numbers of stream-dependent layers and streamindependent layers. Learning from an ensemble of joint models
with different structures is promising in both performance
improvement and model compression.
p(Ou |Lu )P (Lu )
p(Ou )
(10)
Here, Lu is the word sequence of utterance u. P (Lu ) is the
language model probability. p(Ou |Lu ) is the corresponding
acoustic part. The marginal probability p(Ou ) of the feature
sequence Ou , is modeled by summation of the probability
over all possible hypothesis sequences.
X
X
p(Ou ) =
p(Ou , L) =
P (L)p(Ou |L)
(11)
L
L
Here, L denotes all competing hypotheses. As an example
of the sequence discriminative training criteria, the maximum
mutual information (MMI) [27] of inference distribution
stream Lu in utterance u is defined as below,
(r)
JSEQ (Lu , L(r)
u ) = log P (Lu |Ou )
(12)
0
)
(r)
−JSEQ (L(s
un , Lun )
(13)
n∈[1,N ]
=
X
P u
κ
Lu p(Ou |Lu ) P (Lu )
log P
κ
L p(Ou |L) P (L)
(r)
1) Motivation: Speech recognition is inherently a sequence prediction problem. In single-output ASR, sequence
level criteria such as sequence discriminative training tend
to improve performance. The unsupervised single-channel
overlapped speech recognition problem further includes the
speaker tracing problem, which is also a sequence level
problem. Previous works concatenate frame level CE as the
criteria of the sequence level problem, which limits the
modeling effect of the neural network. In this paper, sequence
discriminative training of multiple output streams is proposed
for the first time.
2) Sequence Discriminative Criterion of Multiple outputs:
In single-output ASR, to form a sequence discriminative
training criterion, it is necessary to calculate the sequence
posterior probability using Bayes’ theorem as below,
P (Lu |Ou ) =
X
Different from
Equation (4), the best permutation is decided
(s0 )
(r)
by JSEQ (Lun , Lun ), which is the sequence discriminative
criterion of taking the s0 -th permutation in n-th output
inference stream at utterance u. Similar to CE-PIT, JSEQ of all
the permutations are calculated and the minimum permutation
is taken to do the optimization.0
(s )
(r)
The definition of JSEQ (Lun , Lun ) is similar to Equation (12) in single output ASR.
X
0
)
(r)
JSEQ (L(s
JMMI =
un , Lun )
u
C. Multi-output Sequence Discriminative Training
1
s ∈S N
min
0
(14)
For simplicity, Lu = Lun and L is all0 the hypothesis
(s )
sequences generated by the output stream Lun . p(Ou |Lu ) and
p(Ou |L) is the conditional likelihood obtained from forward
propagation of the joint model, P (Lu ) and P (L) is the prior
probability obtained from language model.
3) Competing Hypothesis Modeling: The hypothesis sequence L refers to all the competing hypotheses in the
speech recognition. Bad modeling of L, namely ignoring
some modeling errors, results in imprecise estimation of
Equation (11), which hurts ASR performance. Thus competing
hypotheses modeling is key to the discriminative training.
Empirically, it is constrained by linguistic search space. In
the single output ASR, the linguistic search space is further
pruned by the online method, lattice-based discriminative
training [28], or by the offline method, lattice-free (LF)
discriminative training [29], [30], [31].
Compared with single output ASR, the additional error
hypothesis types include: i) Assignment errors: someone said
a word, but it is assigned to the wrong channel. ii) cross talk
errors: one person said a word, but it appears in multiple
channels. They both come from imperfect acoustic modeling
and result in several challenges in discriminative training.
The first problem is linguistic search space modeling.
As discussed in the first paragraph, there are mainly two
branches of methods. When training the joint model, speaker
tracing results can always change, which results in different
permutations of the same utterance between different epochs.
Thus if using lattice-based method, lattices should be updated
after each epoch in case of bias in the search space modeling.
Another choice is to use a pre-pruned senone level language
model as the common search space for all utterances [31].
With this method, the lattice generation problem in multiple
outputs can be solved and the discriminative training can be
conducted efficiently in the shared search space.
7
The second problem is the swapped word modeling
in multiple outputs. Swapped word results in both cross
talk errors and assignment errors. Thus bad swapped word
modeling hurts the ASR performance. Generally, the linguistic
search space is estimated from the transcription of the training
dataset. And then sequence criterion is calculated in this
search space. Because there’s no swapped word phenomenon
in the transcription, thus the search space doesn’t contain
swapped word cases, which results in overestimating the
sequence criterion. Especially for the multiple output streams,
the swapped word errors are critical to the ASR performance.
Three methods are proposed to cope with the problem.
•
•
Artificial swapped words. A very simple method is to
generate several copies of the transcription with artificially
swapped words in each copy. And then the language
model is estimated on the re-generated transcription. With
this method, some of the swapped word cases can still
exist in the search space. Thus the problem is alleviated.
Empirically, the senone level language model is obtained
from the senone level transcription, namely clustered triphone state alignment. In case of significant increasing in
the search space because of the swapped word, we set a rule
that in each frame, the probability of senone swapping is α.
But if the senone is swapped, the senone sequence of the
following β frames won’t be swapped. And γ copies of the
transcriptions are generated.
De-correlated lattice free MMI (LF-DC-MMI). The motivation is that swapped words come from the other
output streams. Thus adding these output streams into the
search space and minimizing them in the denominator of
discriminative training can alleviate the problem.
JLF-DC-MMI =
X
u
P
p(Ou |Lu )κ P (Lu )
·
log[ PLu
( L p(Ou |L)κ P (L) )1−λ
(
•
1
]
κ
λ
Lû p(Ou |Lû ) P (Lû ) )
(15)
P
In Equation (15), the other output streams are denoted as
Lû . An interpolation weight λ is added with the augmented
term in the denominator.
De-correlated lattice free boosted MMI (LF-DC-bMMI).
Analogous to boosted MMI [32] as Equation (16),
JLF-bMMI =
X
u
P
log P
L
Lu
p(Ou |Lu )κ P (Lu )
p(Ou |L)κ P (L)e−b
maxLu A(L,Lu )
(16)
we propose de-correlated lattice free boosted MMI (LF-DCbMMI) as Equation (17). Here, b is the boosting factor.
A(L, Lu ) is the state level accuracy between sequence L and
Lu . By this method, the ASR error hypotheses can be further
minimized in the denominator. In the proposed method,
both the ASR errors between the target inference sequence
and the target reference, and the falsely-recognition of the
interfere streams, are boosted.
X
X
p(Ou |Lu )κ P (Lu )·
JLF-DC-bMMI =
log [
u
Lu
1
P
L
p(Ou |L)κ P (L)e−b
maxLu A(L,Lu )−b̂ maxLû (1−A(L,Lû ))
(17)
where b̂ is the de-correlated boosting factor and A(L, Lû )
measures how many falsely recognitions of the interfere
streams.
Experiments are conducted on all three methods in
Section IV-D and the first method can be further combined
with the other two.
IV. E XPERIMENT
The experimental results are reported in artificial overlapped
Switchboard corpus and Eval2000 hub5e-swb test set.
Although the methods presented here are valid for any number
of overlapped speakers, we focus on the two-talker scenario.
A. Experimental Setup
For training, the Switchboard corpus [33] is used, which
contains about 300 hours of speech. Evaluation is carried
out on the Switchboard (SWB) subset of the NIST 2000
CTS (hub5e-swb) test set. The waveforms were segmented
according to the NIST partitioned evaluation map (PEM) file.
Two-talker overlapped speech is artificially generated by
mixing these waveform segments. To maximize the speech
overlap, we developed a procedure to mix similarly sized
segments at around 0dB. First, we sort the speech segments
by length. Then, we take segments in pairs, zero-padding the
shorter segment so both have the same length. These pairs
are then mixed together to create the overlapped speech data.
The overlapping procedure is similar to [14] except that we
make no modification to the signal levels before mixing 4 .
After overlapping, there’s 150 hours data in the training,
called 150 hours dataset, and 915 utterances in the test set.
After decoding, there are 1830 utterances for evaluation, and
the shortest utterance in the hub5e-swb dataset is discarded.
Additionally, we define a small training set, the 50 hours
dataset, as a random 50 hour subset of the 150 hours dataset.
Results are reported using both datasets.
In the training stage, 80-dimensional log-filterbank features
were extracted every 10 milliseconds, using a 25-millisecond
analysis window. The convolution neural network (CNN)
models use 41 context frames (20 in both left and right) and
the long short term memory networks (LSTM) processed one
frame of input at a time. All neural networks were trained with
the Microsoft Cognitive Toolkit (CNTK) [34]. The detailed
setup of CNN is listed in Section IV-B. The acoustic model
is based on three state left-to-right triphone models with 9000
tied states (senones). The individual senone alignments for the
two-talkers in each mixed speech utterance are from the singlespeaker ASR alignment [31]. For compatibility, the alignment
of the shorter utterance within the mixed speech is padded
4 We
will try to release scripts that create the training data in the future.
]
8
with the silence state at the front and the end. The clean
speech recognition performance in the corpus can be referred
to [31], [30]. Using clean speech model to do decoding in
the overlapped speech isn’t reported as it’s as bad as in [14].
The baseline model of joint training is a PIT-ASR model with
a setup similar to [14]. The PIT-ASR model is composed of
10 bidirectional LSTM layers with 768 memory cells in each
layer 5 , and 80-dimensional feature. The baseline model of
separately optimized system is a PIT for speech separation
(PIT-SS) model combined with a clean speech ASR model. As
PIT-SS model has shown competitive performance compared
with other speech separation systems, only PIT-SS model is
taken as the baseline. The PIT-SS model has a setup similar
to [13] but with 6 bidirectional LSTM layers with 768 memory
cells in each layer, it directly outputs multiple channels of the
80 dimensional log-filterbank features the speech recognition
module expects. The speech recognition module, pretrained as
a clean speech model, is composed of 4 bidirectional LSTM
layers with 768 memory cells in each layer. It is trained from
the corresponding source speech segments used to create the
overlapped corpus. After initialization, the WER performance
of the ASR model in the clean speech test set is 17.0%.
Although it would be easy to incorporate a stronger acoustic
model [31] in conjunction with the proposed method, we chose
a structure that allows for a fair comparison, in terms of
the number of model parameters, among the baselines and
proposed methods. However, as discussed in Section III-A, the
modular system needs fewer parameters and training iterations
to achieve good performance. Notably, the use of enhanced
signals after speech separation as training data of the speech
recognition module tends to degrade the ASR performance and
isn’t included. The reason can be from the sporadic distortions
that signal processing inevitably adds, similar to what has been
observed in [19].
In the evaluation stage, a 30k-vocabulary language model
derived from the most common words in the Switchboard and
Fisher corpora is used. The decoder uses a statically compiled
unigram graph, and dynamically applies the language model
score. The unigram graph has about 300k states and 500k
arcs [31]. Two outputs of the PIT-ASR model are both used
in decoding to obtain the hypotheses for two talkers. For
scoring, we evaluated the hypotheses on the pairwise score
mode against the two references, and used the assignment with
better word error rate (WER) for each utterance [14]. Only the
average WER of two output streams is reported, as the task is
to correctly recognize all words from both speakers.
B. Separate Optimization v.s. Joint Modeling
Table I shows the performance of the naive joint modeling,
PIT-ASR, compared with that of the separately optimized
system.
5 In the updated arxiv version of [14], the number of layers changed from 10
to 4 and a moderate WER improvement is obtained. Since the authors didn’t
provide further information, this work mainly compares with the result of 10
layers. In this work, more layers always bring about better performance. And
we believe that improved performance with fewer layers in [14] is a result of
bad model generalization in the PIT-ASR model framework.
TABLE I
Separate Optimization v.s. Joint Modeling. Fine-tune ST denotes to fine-tune
the speaker tracing module and Fine-tune ASR denotes to fine-tune the
speech recognition module defined in Figure 2.
Layers
10 · 0
Modular
×
×
√
6·4
√
√
Fine-tune ST
n/a
n/a
×
√
√
Fine-tune ASR
n/a
n/a
×
×
√
WER
57.5
52.8
93.4
51.3
50.2
Rel. (%)
0
-8.2
+62.4
-10.7
-12.7
The first row shows the performance of the joint training
baseline model in this corpus, PIT-ASR [14]. Compared with
the 0dB WER result listed in [14], 55.80%, the performance
is reasonable 6 . As discussed in Section III-A, the separately
optimized system has a similar number of parameters but
different model architecture. Thus to make a fair comparison,
the model with 6 stream-independent layers in the bottom
and 4 parameter-shared stream-dependent layers in the top,
denoted as 6·4, is listed in the second row 7 . The learnable
structure is the same to the dash-dot blocks shown in
Figure 2(e), but trained from scratch as 10·0. The performance
of the 6·4 structure is significantly better than that of the
10·0 structure. The reason is that unlike in the pure speech
separation task, the speech recognition stage in this task is
also very hard and needs more nonlinear layers. It also shows
that this task is much harder than the speech separation, so
better joint training method to fulfill the performance of each
module is critical to the success.
Without fine-tuning parameters, the performance of the
separately optimized system is shown in the third row.
The significantly worse performance comes from the feature
mismatch in Equation (3). With fine-tuning parameters, the
performance is restored in the fourth and fifth rows. The
system in the fifth row can be viewed as a strong baseline
with separate optimization and fine-tuning.
The better performance of the progressive joint training is
from better model generalization and training efficiency 8 .
Figure 4 show the effect. Training curves of both joint
modeling, i.e. the second row in the table, and progressive
joint modeling, i.e. the fifth row in the table, are plotted.
From the figure, both better starting point and better converged
minimum can be observed in the joint progressive training.
With better joint training strategy shown in Section IV-C, such
modeling effect can be further fulfilled.
Table II shows the performance of the joint modeling from
different modular initialization setups. All modular systems are
fine-tuned after joint training. The first and second rows show
the naive joint training models with BLSTM and BLSTM
combined with CNN, respectively. 6·4 BLSTM refers to 6
layers BLSTM in the bottom and 4 parameter-shared layers
6 The differences include: the dataset is 50 hours versus 400 hours, while the
clean speech WER is 17.0% versus 26.6% . In Section IV-E, the experiments
are extended to 150 hours corpus.
7 As a comparison, the baseline 10·0 has 10 stream-independent layers in
the bottom and no stream-dependent layers except the output layers.
8 After submitting the paper, we notice that similar progressive joint training
idea is proposed in [35]. Experiments using the AMI corpus [36] shows similar
improvement.
9
7
6.5
6
Cross Entropy Loss
5.5
5
4.5
4
3.5
3
2.5
2
0
10
20
30
40
50
60
Epochs
Joint Model
Pro. Joint Model
Pro. Joint Model+ Transf.
Fig. 4. Validation Curves of Naive Joint Modeling and the Proposed
Methods. Joint modeling, progressive joint modeling and transfer learning
based progressive joint modeling are denoted as Joint Model, Pro. Joint Model
and Pro. Joint Model + Transf. in the figure. Each epoch contains 24 hours
of data.
the acoustic model with modules focusing on different scales
and take advantage of the complementarity of CNN, LSTM
and DNN. The difference is that, to fulfill the respective
advantages in modeling, the proposed method further pretrains
each module with different criteria. ii) As the performance
improvement from modularization and initialization is much
larger in 1 LACE + 5·4 BLSTM, it shows that module-wise
initialization is important to fulfill the modeling power of
neural networks especially with different structures and scales.
To further analyze the frame-wise interpreting ability
of CNN and BLSTM, experiments without frame-wise
interpreting initialization are conducted in the fifth and sixth
rows. This time, the performances are similar both in 6·4
BLSTM and 1 LACE + 5·4 BLSTM. It shows that in
initialization of the speaker tracing module, both BLSTM
and CNN can spontaneously learn the frame-wise interpreting
ability. We notice that in [13], frame-wise PIT training doesn’t
show good performance, which is similar to our observation.
TABLE II
Progressive Joint Modeling Based on Modular Initialization. Initializations
of frame-wise interpreting, speaker tracing and speech recognition are
denoted as FI, ST and ASR respectively.
Modular Init.
for each output in the top, i.e. 6 · 4 in Table I. Layerwise context expansion with attention (LACE) model is used
for CNN [37], which is a TDNN [38] variant where each
higher layer is a weighted sum of nonlinear transformations
of a window of lower layer frame. Each LACE block starts
with a convolution layer with stride 2 which sub-samples the
input and increases the number of channels. This layer is
followed by three RELU-convolution layers with jump links.
The channel size is 48. The slightly different setup compared
with [31] is to make parameter number of one LACE block
comparable with one layer of bidirectional LSTM with 768
memory cells, i.e. 5M parameters. Only one block of LACE
is used to replace one layer of BLSTM as the frame-wise
interpreting module, because the speaker tracing and speech
recognition modules are the hardest parts in the problem. The
other parts of the structure are the same, and the proposed
structure is denoted as 1 LACE + 5·4 BLSTM. From the table,
it can be observed that there’s no improvement by merely
stacking these kinds of neural networks together and jointly
training them.
In the third and fourth rows, the model is firstly initialized
with frame-wise interpreting, Figure 2(b), speaker tracing,
Figure 2(c), and speech recognition, Figure 2(d), tasks
respectively and then jointly trained. Comparing the fourth
row to the third row, 1 LACE + 5·4 BLSTM shows larger
improvement than 6·4 BLSTM. Two conclusions can be
derived from the results: i) CNN structure is more suitable
for the frame-wise interpreting module because it focuses
on the local context and has better modeling power of
frequency variations [39]. Meanwhile, LSTM is good at
temporal modeling, which is more suitable for the speaker
tracing and speech recognition module. The architecture with 1
LACE + 5·4 BLSTM layers combines their strength. Notably,
[39] proposes a similar structure, called CLDNN, to form
Neural network
6·4 BLSTM
1 LACE + 5·4 BLSTM
6·4 BLSTM
1 LACE + 5·4 BLSTM
6·4 BLSTM
1 LACE + 5·4 BLSTM
×
FI+ST+ASR
ST+ASR
WER
52.8
52.9
50.3
47.4
50.2
47.4
Rel. (%)
0
+0.2
-4.9
-10.2
-5.0
-10.2
C. Self-transfer Learning Based Joint Modeling
Table III shows the performance improvement of the transfer
learning applied to joint modeling. For transfer learning, the
interpolation weight between hard and soft labels is 0.5 .
TABLE III
Transfer Learning Based Joint Modeling
Layers
10·0
6·4
Modular
×
×
×
×
×
√
√
teacher
×
9·1 ⊕ 6·4 ⊕ 3·7
clean
×
clean
clean
MMI clean
WER
57.5
55.0
52.5
52.8
47.1
38.9
35.8
Rel. (%)
0
-4.4
-8.7
-8.2
-18.0
-32.4
-37.7
The original PIT-ASR system is in the first row and a better
PIT-ASR baseline with 6·4 structure in Table I is also included
in the fourth row.
The ensemble-based transfer learning proposed in Section III-B3 is tested in the second row. The ensemble contains
3 types of structure, 9·1, 6·4 and 3·7, where the left number
denotes the bottom stream-independent layers and the right
number denotes the top stream-dependent layers. The student
network learns from each teacher one-by-one. Although it’s
not a large gain, it shows improvement after learning from
each teacher.
10
The result of replacing hard labeling with simultaneous
clean speech based transfer learning is listed in the third and
the fifth rows. In both model architectures, transfer learning
brings about a relative 10% improvement over the respective
baseline, which is comparable with the result in [18]. It
shows that soft distribution inferred by the model with similar
architecture is superior to the hard labeling.
The self-transfer learning based progressive joint modeling
is finally listed in the sixth and the seventh row by using CEtrained and MMI-trained clean speech teacher respectively.
Notably, as the model framework discussed in III-B, the
initializations of speech recognition modules are the respective
clean speech teachers. The result shows over 30% relative
improvement. Comparing the third and the fifth rows with
Table II, it can be further observed that, combining progressive
joint training and self-transfer learning brings about even
larger improvement compared with the summation of the
relative improvement from each of the two technologies. The
learning curve of the proposed method is also shown in
Figure 4.
From these results, we conclude: i) The proposed
method brings about faster convergence and better converged
minimum. The reason is discussed in Section III-B. The
better convergence result also comes from the removal of the
inappropriate hard alignment in the joint training. ii) Easier
convergence helps the model fulfill the best performance
in each module. That’s the explanation of the even better
synergy result compared with the summation of the relative
improvements from transfer learning and progressive joint
training. iii) Better teacher generates better student. And the
MMI-trained distribution can also be transferred to the student
model, similar to what has been observed in [40].
Figure 5 further shows the student performance versus
quality of the teacher in transfer learning based joint modeling.
It can be observed that better student can usually be obtained
with better teacher. An important reason is that self-transfer
learning is conducted by minimizing the divergence of its own
distributions in mixed speech and clean speech. Thus better
original distribution, including MMI-trained distribution, can
intrinsically be part of the joint model and brings about better
results. The only inflection point is in epoch=3 of the MMI
teacher, where the student performance is similar to epoch=1
although the teacher model has better WER performance. We
believe the reason is that the distribution of the teacher model
of epoch=3 is hard to transfer to the student model because of
the transition process from CE-trained distribution to MMItrained distribution.
D. Sequence Discriminative Training
Table IV shows the performance improvement of sequence
discriminative training based joint modeling. In this experiment, the baseline is PIT-ASR model without progressive joint
training. All the structures are kept the same as 10·0 and
only criteria are changed. κ = 0.1, which is in accordance
with the decoding setup. The senone level language model for
competing hypothesis modeling is obtained from the clustered
tri-phone state alignment. Tri-gram is used, similar to [31].
44
Student WER (%)
43
epoch=30
42
41
epoch=50
epoch=40
40
39
epoch=3
38
epoch=60
epoch=1
37
epoch=7
epoch=11
36
35
15
16
17
18
19
MMI teacher
CE teacher
20
21
22
Teacher WER (%)
Fig. 5. Student Performance versus Teacher Quality in Transfer Learning
Based Progressive Joint Modeling. All the student models are fully converged
based on the respective teacher model. The epoch number denotes how many
epochs the teacher model is trained in the clean speech. Each epoch contains
24 hours of data. The MMI teachers are initialized from the best CE teacher,
i.e. epoch=60 .
For the artificial swapped word method, the hyper-parameters
in Section III-C3 is decided by the preliminary experiment.
α = 0.4, β = 10 and γ = 2. The swapped word senone level
search graph is 4 times larger than the original tri-gram senone
level language model graph. λ = 0.1, b = 0.1 and b̂ = 0.2.
TABLE IV
Sequence Discriminative Training Based Joint Modeling
corpus
50 hours
150 hours
criterion
PIT-CE
LF-MMI
LF-DC-MMI
LF-bMMI
LF-DC-bMMI
LF-MMI
LF-DC-bMMI
PIT-CE
LF-MMI
LF-DC-bMMI
Den. graph
n/a
senone transcription
+ art. swapped word
n/a
senone transcription
WER
57.5
54.6
53.2
53.5
53.1
53.5
52.7
42.2
40.1
39.1
Rel. (%)
0
-4.9
-7.3
-6.9
-7.4
-6.9
-8.2
0
-4.9
-7.3
The baseline PIT-ASR system is shown in the first row,
denoted as PIT-CE to show the criterion of the system.
Applying the naive sequence discriminative training method
for multiple outputs in the second row, only brings about
4.9% relative improvement. For the recognition result, word
precision is improved but insertion error increases. The reason
is from imperfect search space modeling of swapped words
discussed in III-C.
By applying the proposed LF-DC-MMI method in the third
row, the recognition result is significantly improved compared
with both baseline and the naive LF-MMI. The proposed
method minimizes the swapped words from parallel output
streams in the denominator modeling. Thus the problem can
be alleviated.
The fourth and fifth rows show the effect of using bMMI
instead of MMI in the formulations. The bMMI criterion
boosts the ASR errors in the denominator modeling, implicitly
including possibly swapped words. Although significant
improvement can be observed between LF-MMI and LF-
11
bMMI, the LF-DC-bMMI explicitly includes the swapped
words and achieves an even better result.
The proposed artificial swapped word method is shown in
the sixth and seventh rows. By comparing the sixth row with
the second row, and comparing the seventh row with the fifth
row, it shows slight but consistent improvement in solving the
swapped word problem. And the method can also be combined
with LF-DC-bMMI to achieve 8.2% relative improvement
versus the CE-trained PIT-ASR baseline.
In the eighth to tenth rows, experiments are conducted
on 150 hours corpus. The results are similar, and LF-DCbMMI criterion shows consistent improvement versus PITCE and naive sequence discriminative training criterion. In
Section IV-E, it is shown that sequence discriminative training
can be combined with other technologies and achieves further
consistent and significant improvement.
The discriminative training criterion helps the system
training in two ways. Firstly, sequence level criterion helps
the sequence level speaker tracing problem in PIT modeling.
Specifically, linguistic information is encoded in the senone
level language modeling in discriminative training. Thus the
procedure implicitly integrates linguistic information in the
speaker tracing problem. Secondly, sequence discriminative
training improves the speech recognition module. Notably, all
the sequence discriminative training procedures are applied
after CE initialization as in [31]. With initialization, it also
helps Equation (13) to reach a better minimum.
E. Combination and Extension to Larger Corpus
Table V summarizes the performance improvement of
integrating all the proposed methods.
TABLE V
Performance Summary in SWBD 50 Hours Dataset
Neural network
10·0 BLSTM
6·4 BLSTM
1 LACE + 5·4 BLSTM
Model
PIT-CE
progressive joint training
+ clean teacher
+ MMI clean teacher
+ LF-DC-bMMI
progressive joint training
+ clean teacher
+ MMI clean teacher
+ LF-DC-bMMI
WER
57.5
Rel. (%)
0
50.2
38.9
35.8
35.2
47.4
36.0
34.6
34.0
-13
-32.4
-37.7
-38.8
-17.5
-37.4
-39.8
-40.9
The PIT-ASR model [14], denoted as PIT-CE, is taken as the
baseline of naive joint modeling in the first row. The separately
optimized system, namely PIT-SS+ASR, is not included here.
As shown in Table II, the performance deteriorates because
of feature mismatch. Instead, the proposed progressive joint
training model in the second row can be taken as a stronger
separately optimized system with fine-tuning. The proposed
self-transfer learning based joint training model shows further
significant improvement in the third and fourth rows. Finally,
the multi-output sequence discriminative training is applied
and achieves moderate improvement, although the teacher
model is already MMI-trained, similar to what has been
observed in [41]. Figure 6 shows decoding examples of
the proposed methods versus the PIT baseline. The baseline
contains many errors due to bad model generalization with
limited size of dataset. With the proposed methods, errors
are significantly reduced. Notably, in this example, the selftransfer learning based progressive joint training mainly
reduces errors from similar pronunciations, while sequence
discriminative training mainly reduces explicit syntax or
linguistic errors, which is in line with the expectation.
With similar number of parameters but different neural
networks, namely 1 LACE + 5·4 BLSTM, the system can
be consistently improved in the sixth to ninth rows. We
believe the further improvement comes from the proper
modularization of the problem, which is discussed in
Section IV-B.
Table VI expands the dataset to 150 hours to show the effect
of more training data.
TABLE VI
Performance Summary in SWBD 150 Hours Dataset
Neural network
10·0 BLSTM
6·4 BLSTM
1 LACE + 5·4 BLSTM
Model
PIT-CE
progressive joint training
+ clean teacher
+ LF-DC-bMMI
progressive joint training
+ clean teacher
+ LF-DC-bMMI
WER
Rel. (%)
42.2
41.0
32.8
30.8
39.4
30.4
28.0
0
-2.9
-22.3
-27.0
-6.6
-27.9
-33.6
The naive joint training baseline in the first row significantly
benefits from more data and shrinks the gap to the proposed
progressive joint training model in the second row. However, it
still even significantly worse than the self-transfer learning and
sequence discriminative training based joint model trained in
50 hours data in Table V. It again shows the disadvantages of
large model complexity and insufficient model generalization
discussed in Section II. i.e. compared with merely increasing
data, the better method to solve the problem is to improve
the model generalization. Besides, the convergence speed of
naive joint training model in the larger dataset is even slower,
namely 4 times more epochs versus the proposed method.
Comparing Table VI with Table V, the proposed selftransfer learning based joint training and multi-output
sequence discriminative training show consistent relative
improvement versus the progressive joint training. Compared
with Table V, sequence discriminative training achieves larger
relative improvement on the CE-trained teacher based system.
In both 50 hours and 150 hours corpus, the proposed
method achieves over 30% relative improvement respectively,
versus the PIT-ASR system and the PIT-SS+ASR system.
The improvement comes from better model generalization,
training efficiency and the sequence level linguistic knowledge
integration.
Although this paper addresses the case of simultaneous
speech of two people talking at a relative level of 0dB, we
believe it will be straightforward to extend the system to
handle more realistic conditions. The case where one speaker
is louder than the other has already been observed to be
easier for PIT-style models than the 0dB data explored in
this paper [14]. For more than two speakers, extension of
12
0 reference
1 PIT-CE baseline
2 + self-transf. progressive joint training
3 + MMI teacher
4 + seq. disc. tr.
0 reference
1 PIT-CE baseline
2 + self-transf. progressive joint training
3 + MMI teacher
4 + seq. disc. tr.
Fig. 6. 50 hours dataset decoding examples of the proposed methods versus the PIT baseline. The upper part is from stream A and the lower part is from
stream B. C, S, D, I refer to correct, substitution, deletion and insertion.
the proposed system should follow the same construction
described in [35]. Finally, we expect robustness to background
noise and reverberation to come from standard techniques such
as multi-condition training [43].
ACKNOWLEDGMENT
We thank Chris Basoglu, Frank Seide for their invaluable
assistance with CNTK; Mike Seltzer, Takuya Yoshioka, Hakan
Erdogan and Andreas Stolcke for many helpful conversations.
The first author would like to further thank Jerry and Juncheng
Gu for their supports during the internship.
V. C ONCLUSION
R EFERENCES
In this work, we proposed to divide the single channel
overlapped speech recognition problem into three subproblems: frame-wise interpreting, speaker tracing and speech
recognition. Each module is firstly optimized separately
with specific designed criteria, which significantly improves
the system generalization and training efficiency. After the
initialization, modules are jointly trained with two novel
strategies: self-transfer learning and multi-output sequence
discriminative training. Specifically, in the joint training stage,
the clean speech model fine-tunes its parameters with other
modules in overlapped speech to fit its own distribution in the
simultaneous clean speech. And then sequence discriminative
training designed for multiple outputs is applied to integrate
linguistic and sequence information. The proposed framework
achieves 30% relative improvement over both a strong jointly
trained system, PIT-ASR, and a separately optimized system,
PIT-SS+ASR.
The proposed framework shows promising perspectives of
future improvements, which are: i) Integrating state-of-theart technologies in each module in the initialization stage,
e.g., DPCL [9]. ii) Applying other sequence level criteria to
improve the speaker tracing and speech recognition modules,
i.e. connectionist temporal classification (CTC) [45]. iii)
Explicit integration of language model in the joint modeling,
e.g. joint decoding [6] and end-to-end modeling [46].
[1] E. C. Cherry, “Some experiments on the recognition of speech, with one
and with two ears,” The Journal of the acoustical society of America,
vol. 25, no. 5, pp. 975–979, 1953.
[2] A. S. Bregman, Auditory scene analysis: The perceptual organization of
sound. MIT press, 1994.
[3] D. Wang and G. J. Brown, Computational auditory scene analysis:
Principles, algorithms, and applications. Wiley-IEEE press, 2006.
[4] M. Cooke, J. R. Hershey, and S. J. Rennie, “Monaural speech separation
and recognition challenge,” Computer Speech & Language, vol. 24,
no. 1, pp. 1–15, 2010.
[5] J. Du, Y. Tu, Y. Xu, L. Dai, and C.-H. Lee, “Speech separation of a target
speaker based on deep neural networks,” in Signal Processing (ICSP),
2014 12th International Conference on. IEEE, 2014, pp. 473–477.
[6] C. Weng, D. Yu, M. L. Seltzer, and J. Droppo, “Deep neural
networks for single-channel multi-talker speech recognition,” IEEE/ACM
Transactions on Audio, Speech and Language Processing (TASLP),
vol. 23, no. 10, pp. 1670–1679, 2015.
[7] M. Wertheimer, “Laws of organization in perceptual forms.” 1938.
[8] M. N. Schmidt and R. K. Olsson, “Single-channel speech separation
using sparse non-negative matrix factorization,” in Interspeech 2006,
2006.
[9] J. R. Hershey, Z. Chen, J. Le Roux, and S. Watanabe, “Deep
clustering: Discriminative embeddings for segmentation and separation,”
in Acoustics, Speech and Signal Processing (ICASSP), 2016 IEEE
International Conference on. IEEE, 2016, pp. 31–35.
[10] Z. Chen, Y. Luo, and N. Mesgarani, “Deep attractor network for
single-microphone speaker separation,” in Acoustics, Speech and Signal
Processing (ICASSP), 2017 IEEE International Conference on. IEEE,
2017, pp. 246–250.
[11] Y. Isik, J. L. Roux, Z. Chen, S. Watanabe, and J. R. Hershey, “Singlechannel multi-speaker separation using deep clustering,” arXiv preprint
arXiv:1607.02173, 2016.
13
[12] Y. Wang, J. Du, L.-R. Dai, and C.-H. Lee, “A gender mixture detection
approach to unsupervised single-channel speech separation based on
deep neural networks,” IEEE/ACM Transactions on Audio, Speech, and
Language Processing, vol. 25, no. 7, pp. 1535–1546, 2017.
[13] D. Yu, M. Kolbæk, Z.-H. Tan, and J. Jensen, “Permutation invariant
training of deep models for speaker-independent multi-talker speech
separation,” in Acoustics, Speech and Signal Processing (ICASSP), 2017
IEEE International Conference on. IEEE, 2017, pp. 241–245.
[14] D. Yu, X. Chang, and Y. Qian, “Recognizing multi-talker speech
with permutation invariant training,” CoRR, vol. abs/1704.01985, 2017.
[Online]. Available: http://arxiv.org/abs/1704.01985v3
[15] P. C. Woodland, J. J. Odell, V. Valtchev, and S. J. Young, “Large
vocabulary continuous speech recognition using htk,” in Acoustics,
Speech, and Signal Processing, 1994. ICASSP-94., 1994 IEEE
International Conference on, vol. 2. Ieee, 1994, pp. II–125.
[16] Y. Bengio, J. Louradour, R. Collobert, and J. Weston, “Curriculum
learning,” in Proceedings of the 26th annual international conference
on machine learning. ACM, 2009, pp. 41–48.
[17] S. J. Pan and Q. Yang, “A survey on transfer learning,” IEEE
Transactions on knowledge and data engineering, vol. 22, no. 10, pp.
1345–1359, 2010.
[18] J. Li, M. L. Seltzer, X. Wang, R. Zhao, and Y. Gong, “Large-scale
domain adaptation via teacher-student learning,” in Interspeech 2017,
2017.
[19] S. Watanabe, T. Hori, J. Le Roux, and J. R. Hershey, “Student-teacher
network learning with enhanced features,” in Acoustics, Speech and
Signal Processing (ICASSP), 2017 IEEE International Conference on.
IEEE, 2017, pp. 5275–5279.
[20] A. Graves, N. Jaitly, and A.-r. Mohamed, “Hybrid speech recognition
with deep bidirectional lstm,” in Automatic Speech Recognition and
Understanding (ASRU), 2013 IEEE Workshop on. IEEE, 2013, pp.
273–278.
[21] D. Yu, K. Yao, H. Su, G. Li, and F. Seide, “Kl-divergence regularized
deep neural network adaptation for improved large vocabulary speech
recognition,” in Acoustics, Speech and Signal Processing (ICASSP),
2013 IEEE International Conference on. IEEE, 2013, pp. 7893–7897.
[22] L. Deng, A. Acero, M. Plumpe, and X. Huang, “Large-vocabulary
speech recognition under adverse acoustic environments.” in INTERSPEECH, 2000, pp. 806–809.
[23] K. Markov and T. Matsui, “Robust speech recognition using generalized
distillation framework.” in INTERSPEECH, 2016, pp. 2364–2368.
[24] S. Wang, K. Li, Z. Huang, S. M. Siniscalchi, and C.-H. Lee, “A transfer
learning and progressive stacking approach to reducing deep model sizes
with an application to speech enhancement,” in Acoustics, Speech and
Signal Processing (ICASSP), 2017 IEEE International Conference on.
IEEE, 2017, pp. 5575–5579.
[25] R. Hadsell, P. Sermanet, J. Ben, A. Erkan, M. Scoffier, K. Kavukcuoglu,
U. Muller, and Y. LeCun, “Learning long-range vision for autonomous
off-road driving,” Journal of Field Robotics, vol. 26, no. 2, pp. 120–144,
2009.
[26] J. Ba and R. Caruana, “Do deep nets really need to be deep?” in
Advances in neural information processing systems, 2014, pp. 2654–
2662.
[27] K. Veselỳ, A. Ghoshal, L. Burget, and D. Povey, “Sequencediscriminative training of deep neural networks.” in Interspeech, 2013,
pp. 2345–2349.
[28] D. Povey, “Discriminative training for large vocabulary speech
recognition,” Ph.D. dissertation, University of Cambridge, 2005.
[29] S. F. Chen, B. Kingsbury, L. Mangu, D. Povey, G. Saon, H. Soltau, and
G. Zweig, “Advances in speech transcription at ibm under the darpa
ears program,” IEEE Transactions on Audio, Speech, and Language
Processing, vol. 14, no. 5, pp. 1596–1608, 2006.
[30] D. Povey, V. Peddinti, D. Galvez, P. Ghahremani, V. Manohar, X. Na,
Y. Wang, and S. Khudanpur, “Purely sequence-trained neural networks
for asr based on lattice-free mmi.” in INTERSPEECH, 2016, pp. 2751–
2755.
[31] W. Xiong, J. Droppo, X. Huang, F. Seide, M. Seltzer, A. Stolcke,
D. Yu, and G. Zweig, “Achieving human parity in conversational speech
recognition,” arXiv preprint arXiv:1610.05256, 2016.
[32] D. Povey, D. Kanevsky, B. Kingsbury, B. Ramabhadran, G. Saon,
and K. Visweswariah, “Boosted mmi for model and feature-space
discriminative training,” in Acoustics, Speech and Signal Processing,
2008. ICASSP 2008. IEEE International Conference on. IEEE, 2008,
pp. 4057–4060.
[33] J. J. Godfrey, E. C. Holliman, and J. McDaniel, “Switchboard: Telephone
speech corpus for research and development,” in Acoustics, Speech,
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
[46]
and Signal Processing, 1992. ICASSP-92., 1992 IEEE International
Conference on, vol. 1. IEEE, 1992, pp. 517–520.
F. Seide and A. Agarwal, “Cntk: Microsoft’s open-source deep-learning
toolkit,” in Proceedings of the 22nd ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining. ACM, 2016,
pp. 2135–2135.
Y. Qian, X. Chang, and D. Yu, “Single-channel multi-talker speech
recognition with permutation invariant training,” arXiv preprint
arXiv:1707.06527v1, 2017.
T. Hain, L. Burget, J. Dines, P. N. Garner, F. Grézl, A. El Hannani,
M. Huijbregts, M. Karafiat, M. Lincoln, and V. Wan, “Transcribing
meetings with the amida systems,” IEEE Transactions on Audio, Speech,
and Language Processing, vol. 20, no. 2, pp. 486–498, 2012.
D. Yu, W. Xiong, J. Droppo, A. Stolcke, G. Ye, J. Li, and G. Zweig,
“Deep convolutional neural networks with layer-wise context expansion
and attention.” in INTERSPEECH, 2016, pp. 17–21.
A. Waibel, T. Hanazawa, G. Hinton, K. Shikano, and K. J.
Lang, “Phoneme recognition using time-delay neural networks,” IEEE
transactions on acoustics, speech, and signal processing, vol. 37, no. 3,
pp. 328–339, 1989.
T. N. Sainath, O. Vinyals, A. Senior, and H. Sak, “Convolutional, long
short-term memory, fully connected deep neural networks,” in Acoustics,
Speech and Signal Processing (ICASSP), 2015 IEEE International
Conference on. IEEE, 2015, pp. 4580–4584.
J. Li, R. Zhao, J.-T. Huang, and Y. Gong, “Learning small-size dnn with
output-distribution-based criteria,” in Fifteenth Annual Conference of the
International Speech Communication Association, 2014.
L. Lu and S. Renals, “Small-footprint highway deep neural networks
for speech recognition,” IEEE/ACM Transactions on Audio, Speech, and
Language Processing, vol. 25, no. 7, pp. 1502–1511, July 2017.
X. Zhang and D. Wang, “Binaural reverberant speech separation based
on deep neural networks,” Proc. Interspeech 2017, pp. 2018–2022, 2017.
M. L. Seltzer, D. Yu, and Y. Wang, “An investigation of deep neural
networks for noise robust speech recognition,” in Acoustics, Speech and
Signal Processing (ICASSP), 2013 IEEE International Conference on.
IEEE, 2013, pp. 7398–7402.
L. Drude and R. Haeb-Umbach, “Tight integration of spatial and spectral
features for bss with deep clustering embeddings,” Proc. Interspeech
2017, pp. 2650–2654, 2017.
A. Graves, S. Fernández, F. Gomez, and J. Schmidhuber, “Connectionist
temporal classification: labelling unsegmented sequence data with
recurrent neural networks,” in Proceedings of the 23rd international
conference on Machine learning. ACM, 2006, pp. 369–376.
T. Hori, S. Watanabe, Y. Zhang, and W. Chan, “Advances in joint ctcattention based end-to-end speech recognition with a deep CNN encoder
and RNN-LM,” arXiv preprint arXiv:1706.02737, 2017.
Zhehuai Chen received his B.S. degree in the
Department of Electronic and Information Engineering from Huazhong University of Science and
Technology, China, in 2014. He is currently a Ph.D.
candidate in Shanghai Jiao Tong University working
on speech recognition. His current research interests
include speech recognition, speech synthesis and
deep learning.
14
Jasha Droppo (M’03-SM’07) received the B.S.
degree in electrical engineering (with honors) from
Gonzaga University, Spokane, WA, in 1994, and the
M.S. and Ph.D. degrees in electrical engineering
from the University of Washington, Seattle, in
1996 and 2000, respectively. At the University of
Washington, he helped to develop and promote a
discrete theory for time-frequency representations of
audio signals, with a focus on speech recognition.
He is best known for his research in robust
speech recognition, including algorithms for speech
signal enhancement, model-based speech feature enhancement, robust speech
features, model-based adaptation, and noise tracking. His current interests
include the use of neural networks in acoustic modeling and the application
of large data and general machine learning algorithms to previously handauthored speech recognition components.
Jinyu Li (M’08) received the bachelor and master
degree from University of Science and Technology
of China, in 1997 and 2000, with the highest honor,
and the Ph.D. degree from Georgia Institute of
Technology, Atlanta, in 2008. From 2000 to 2003, he
was a researcher in the Intel China Research Center
and Research Manager in iFlytek Speech, China.
Currently, he is a principal applied scientist and
technical lead in Microsoft Corporation, Redmond,
WA. He leads a team to design and improve speech
modeling algorithms and technologies that ensure
industry state-of-the-art speech recognition accuracy for Microsoft products
such as Cortana and xBox Kinect. His major research interests cover several
topics in speech recognition, including deep learning, noise robustness,
discriminative training, feature extraction, and machine learning methods. He
authored more than 70 refereed publications and around 20 patents. He is the
leading author of the book Robust Automatic Speech Recognition A Bridge to
Practical Applications-Academic Press, Oct, 2015. Currently, he serves as the
associate editor of IEEE/ACM Transactions on Audio, Speech and Language
Processing.
Wayne Xiong (M’07) received the Bachelor degree
in electrical engineering from Tsinghua University,
Beijing, P. R. China, in 2005, and the Master
degree in computer science and engineering from
the Chinese University of Hong Kong, Hong Kong,
in 2007. He joined Microsoft Corporation in 2007.
He is currently working as a Senior Researcher
in the Speech and Dialog Research Group. His
research interests include speech recognition and
next generation of dialog systems.
| 2cs.AI |
arXiv:0912.4360v1 [cs.PL] 22 Dec 2009
Under consideration for publication in Theory and Practice of Logic Programming
1
Polytool: Polynomial Interpretations as a Basis
for Termination Analysis of Logic Programs
MANH THANG NGUYEN
Deceased on June 3, 2009
DANNY DE SCHREYE
Department of Computer Science, K. U. Leuven
Celestijnenlaan 200A, B-3001 Heverlee, Belgium
Danny.DeSchreye@cs.kuleuven.ac.be
JÜRGEN GIESL
LuFG Informatik 2, RWTH Aachen
Ahornstr. 55, D-52074 Aachen, Germany
giesl@informatik.rwth-aachen.de
PETER SCHNEIDER-KAMP
Dept. of Mathematics and Computer Science, U. Southern Denmark
Campusvej 55, DK-5230 Odense M, Denmark
petersk@imada.sdu.dk
submitted 18 February 2008; revised 30 November 2009; accepted 11 December 2009
Abstract
Our goal is to study the feasibility of porting termination analysis techniques developed
for one programming paradigm to another paradigm. In this paper, we show how to adapt
termination analysis techniques based on polynomial interpretations - very well known in
the context of term rewrite systems (TRSs) - to obtain new (non-transformational) termination analysis techniques for definite logic programs (LPs). This leads to an approach
that can be seen as a direct generalization of the traditional techniques in termination
analysis of LPs, where linear norms and level mappings are used. Our extension generalizes these to arbitrary polynomials. We extend a number of standard concepts and results
on termination analysis to the context of polynomial interpretations. We also propose a
constraint-based approach for automatically generating polynomial interpretations that
satisfy the termination conditions. Based on this approach, we implemented a new tool,
called Polytool, for automatic termination analysis of LPs.
KEYWORDS: Termination analysis, acceptability, polynomial interpretations.
1 Introduction
Termination analysis plays an important role in the study of program correctness.
A termination proof is mostly based on a mapping from computational states to
2
Manh Thang Nguyen et al.
some well-founded ordered set. Termination is guaranteed if the mapped values of
the encountered states during a computation, under this mapping, decrease w.r.t.
the order.
For LPs, termination analysis is done by mapping terms and atoms to a wellfounded set of natural numbers by means of norms and level mappings. Proving
termination is based on the search for a suitable norm and level mapping such that
the resulting predicate calls decrease under the mapping.
Until now, most termination techniques for LPs are based on the use of linear
norms and linear level mappings, which measure the size of each term or atom as
a linear combination of the sizes of its sub-terms. For example, the Hasta-La-Vista
system (Serebrenik and De Schreye 2003) infers one specific linear norm and linear
level mapping. In the context of numerical computations, it includes a refinement
on this, based on a case analysis. The tool cTI (Mesnard and Bagnara 2005) uses
a concrete linear norm. The analyzers TermiLog (Lindenstrauss and Sagiv 1997;
Lindenstrauss 2000) and TerminWeb (Codish and Taboch 1999; Taboch et al. 2002)
use a combination of several linear norms to obtain an approximation of the program
and then infer linear level mappings for termination analysis of the approximated
program. However, the restriction to linear norms and level mappings limits the
power of termination analysis considerably. To illustrate this point, consider the following example, der, that formulates rules for computing the repeated derivative of
a function in some variable u. This example from (De Schreye and Serebrenik 2002;
Dershowitz et al. 1997) is inspired by a similar term rewriting example from (Dershowitz 1995).
Example 1 (der )
d(der (u), 1).
(1)
d(der (X + Y ), DX + DY ) :− d(der (X), DX ), d(der (Y ), DY ).
(2)
d(der (X ∗ Y ), X ∗ DY + Y ∗ DX ) :− d(der (X), DX ), d(der (Y ), DY ).
(3)
d(der (der (X)), DDX ) :− d(der (X), DX ), d(der (DX ), DDX ).
(4)
We are interested in proving termination of this program w.r.t. the set of queries
S = { d(t1 , t2 ) | t1 is a ground term and t2 is an arbitrary term}. So the set of
queries is specified by a mode that considers the first argument of d as an input
argument and the second as an output.
As shown in (Dershowitz et al. 1997; Nguyen and De Schreye 2005), the termination proof is impossible when using a linear norm and a linear level mapping.
Indeed, it turns out that all existing non-transformational termination analyzers
for LPs mentioned above fail to prove termination of this example.
In this paper, we propose a general framework for termination proofs of LPs
based on polynomial interpretations. Using polynomial interpretations as a basis
for ordering terms in TRSs was first introduced by Lankford in (Lankford 1979).
It is currently one of the best known and most widely used techniques in TRS
termination analysis.
We develop the approach within an LP context. Classical approaches in LP ter-
Polynomial Interpretations for Termination Analysis of Logic Programs
3
mination use interpretations that map to natural numbers (using linear polynomial functions). In contrast, we will use interpretations that map to polynomials
(using arbitrary polynomial functions). To adapt the classical LP approaches to
polynomial interpretations, we use the concepts of “abstract norm” and “abstract
level mapping” (Verschaetse and De Schreye 1991). We show that with our new
approach, one can also prove termination of programs like Example 1.
We also developed an automated tool (Polytool) for termination analysis based
on our approach (Nguyen and De Schreye 2007). We embedded this within the
constraint-based approach developed in (Decorte et al. 1999) and combined it with
the non-linear Diophantine constraint solver developed by Fuhs et al. (Fuhs et al. 2007)
(implemented in the AProVE system (Giesl et al. 2006)) to provide a completely automated system.
The paper is organized as follows. In the next section, we present some preliminaries. In Section 3, we introduce the notion of polynomial interpretations in logic
programming and show how this approach can be used to prove termination. In
Section 4, we discuss the automation of the approach. In Section 5, we provide and
discuss the results of our experimental evaluation. We end with a conclusion in
Section 6.
2 Preliminaries
After introducing the basic terminology of LPs in Section 2.1, we recapitulate the
concepts of norms and level mappings in Section 2.2 and explain their use for
termination proofs in Section 2.3.
2.1 Notations and Terminology
We assume familiarity with LP concepts and with the main results of logic programming (Apt 1990; Lloyd 1987). In the following, P denotes a definite logic program.
We use Var P , Fun P , and Pred P to denote the sets of variables, function, and predicate symbols of P . Given an atom A, rel (A) denotes the predicate occurring in
A. Let p, q be predicates occurring in the program P . We say that p refers to q if
there is a clause in P such that p is in its head and q is in its body. We say that
p depends on q if (p, q) is in the transitive closure of the relation “refers to”. If
p depends on q and vice versa, p and q are called mutually recursive, denoted by
p ⋍ q. A clause in P with a predicate p in its head and a predicate q in its body,
such that p and q are mutually recursive, is called a (mutually) recursive clause.
Within such a recursive clause, the body-atoms with predicate symbol q are called
(mutually) recursive atoms. Let Term P and Atom P denote, respectively, the sets
of all terms and atoms that can be constructed from P .
In this paper, we focus our attention on definite logic programs and SLD-derivations
where the left-to-right selection rule is used. Such derivations are referred to as LDderivations; the corresponding derivation tree is called LD-tree. We say that a query
Q LD-terminates for a program P , if the LD-tree for (P, Q) is finite (left-termination
4
Manh Thang Nguyen et al.
(Lloyd 1987)). In the following, we usually speak of “termination” instead of “LDtermination” or “left-termination”.
2.2 Norms and Level Mappings
The concepts of norm and level mapping are central in termination analysis of logic
programs.
Definition 1 (norm, level mapping)
A norm is a mapping k.k : Term P → N. A level-mapping is a mapping |.| :
Atom P → N.
Several examples of norms can be found in the literature (Bossi et al. 1991). One
of the most commonly used norms is the list-length norm k.kℓ which maps lists to
their lengths and any other term to 0. Another frequently used norm is the termsize norm k.kτ which counts the number of function symbols in a term. Both of
them belong to a class of norms called linear norms which is defined as follows.
Definition 2 (linear norm and level mapping (Serebrenik 2003))
A norm k.k is a linear norm if it is recursively defined by means of the following
schema:
- kXk = 0 for any variable X,
P
- kf (t1 , . . . , tn )k = f0 + ni=1 fi kti k where fi ∈ N and n ≥ 0.
Similarly, a level mapping |.| is a linear level mapping if it is defined by means of
the following schema:
Pn
- |p(t1 , . . . , tn )| = p0 + i=1 pi kti k where pi ∈ N and n ≥ 0.
2.3 Conditions for Termination w.r.t. General Orders
A quasi-order on a set S is a reflexive and transitive binary relation % defined on
elements of S. We define the associated equivalence relation ≈ as s ≈ t if and only
if s % t and t % s. A well-founded order on S is a transitive relation ≻ where there
is no infinite sequence s0 ≻ s1 ≻ . . . with si ∈ S. A reduction pair (%, ≻) consists of
a quasi-order % and a well-founded order ≻ that are compatible (i.e., t1 % t2 ≻ t3
implies t1 ≻ t3 ). We also need the following notion of a call set.
Definition 3 (call set )
Let P be a program and S be a set of atomic queries. The call set, Call (P, S), is the
set of all atoms A, such that a variant of A is the selected atom in some derivation
for (P , Q ), for some Q ∈ S.
Most often, one regards infinite sets S of queries. For instance, this is the case in
Example 1. As in Example 1, S is then specified in terms of modes or types. As a consequence, in an automated approach, a safe over-approximation of Call (P, S) needs
to be computed, using a mode or a type inference technique (e.g., (Bruynooghe et al. 2005;
Gallagher et al. 2005; Heaton et al. 2000; Janssens and Bruynooghe 1992)).
Polynomial Interpretations for Termination Analysis of Logic Programs
5
In order to obtain a termination criterion that is suitable for automation, one
usually estimates the effect of the atoms in the bodies of clauses by suitable interargument relations. This notion can be defined for arbitrary reduction pairs.
Definition 4 (interargument relation (De Schreye and Serebrenik 2002))
Let P be a program, p be a predicate in P , and (%, ≻) be a reduction pair on Term P .
An interargument relation for p in P w.r.t. (%, ≻) is a relation Rp with the same
arity as p: Rp = {p(t1 , . . . , tn ) | ti ∈ Term P for all 1 ≤ i ≤ n, and ϕp (t1 , . . . , tn )},
where:
- ϕp (t1 , . . . , tn ) is a boolean expression (in terms of disjunction, conjunction,
and negation) of inequalities s % s′ or s ≻ s′ , in which
- s, s′ are constructed from t1 , . . . , tn by applying function symbols from Fun P .
Rp is a valid interargument relation for p in P w.r.t. (%, ≻) if and only if for every
p(t1 , . . . , tn ) ∈ Atom P : P |= p(t1 , . . . , tn ) implies p(t1 , . . . , tn ) ∈ Rp .
Example 2 (interargument relation)
Let P be the standard append program that computes list concatenation. Then
there are a number of valid interargument relations. Consider the reduction pair
(%, ≻) corresponding to the list-length norm k.kℓ , i.e., t1 % t2 if and only if kt1 kℓ ≥
kt2 kℓ and t1 ≻ t2 if and only if kt1 kℓ > kt2 kℓ . For instance, valid interargument
relations for append w.r.t. (%, ≻) are Rappend = {append(t1 , t2 , t3 ) | t1 , t2 , t3 ∈
Term P ∧ ϕappend (t1 , t2 , t3 )}, where ϕappend (t1 , t2 , t3 ) could be:
-
t3 % t2 ∧ t3 % t1 ,
t3 % t2 ,
[t1 , t2 |t3 ] ≻ [t2 |t3 ], or
true
Of course, usually only the first two interargument relations are useful for termination analysis.
Finally, we need the notion of rigidity, in order to deal with bindings that are due
to unification in LD-derivations. These bindings would have to be back-propagated
to the variables in the initial goal. We reformulate rigidity for arbitrary reduction
pairs.
Definition 5 (rigidity - adapted from (De Schreye and Serebrenik 2002))
A term or atom A ∈ Term P ∪ Atom P is called rigid w.r.t. a reduction pair (%, ≻)
if A ≈ Aσ holds for any substitution σ. A set of terms (or atoms) S is called rigid
w.r.t. (%, ≻) if all its elements are rigid w.r.t. (%, ≻).
Example 3 (rigidity)
The list [X|t] (X is a variable, t is a ground term) is rigid w.r.t. the reduction
pair (%, ≻) corresponding to the list-length norm. For any substitution σ, we have
k[X|t]σkℓ = 1 + ktkℓ = k[X|t]kℓ . Therefore, [X|t]σ ≈ [X|t] w.r.t. (%, ≻).
However, the list [X|t] is not rigid w.r.t. the reduction pair (%′ , ≻′ ) corresponding
to the term-size norm k.kτ , i.e., t1 %′ t2 if and only if kt1 kτ ≥ kt2 kτ and t1 ≻′ t2 if
and only if kt1 kτ > kt2 kτ .
6
Manh Thang Nguyen et al.
The following definition introduces the desired termination criterion, i.e., it recalls
the definition of rigid order-acceptability w.r.t. a set of atoms.
Definition 6 (rigid order-acceptability (De Schreye and Serebrenik 2002))
Let S be a set of atomic queries. A program P is rigid order-acceptable w.r.t. S
if there exists a reduction pair (%, ≻) on Atom P where Call (P, S) is rigid w.r.t.
(%, ≻) and where for each predicate p in P , there is a valid interargument relation
Rp in P w.r.t. (%, ≻) such that
- for any clause A :− B1 , B2 , . . . , Bn in P ,
- for any atom Bi ∈ {B1 . . . , Bn } such that rel (Bi ) ⋍ rel(A),
- for any substitution θ such that the atoms B1 θ, . . . , Bi−1 θ are elements of
their associated interargument relations Rrel(B1 ) , . . . , Rrel(Bi−1 ) :
Aθ ≻ Bi θ.
Theorem 1 states that rigid order-acceptability is a sufficient condition for termination. We refer to (Serebrenik 2003), Theorems 3.32 and 3.54, for the proof of
Theorem 1.
Theorem 1 (termination criterion by rigid order-acceptability)
If P is rigid order-acceptable w.r.t. S, then P terminates for any query in S.
Rigid order-acceptability is sufficient for termination, but is not necessary for it
(see (De Schreye and Serebrenik 2002)). With Definition 6 and Theorem 1, proving
termination of a program requires verifying the rigidity of the call set, verifying
the validity of interargument relations for predicates, and verifying the decrease
conditions for the (mutually) recursive clauses.
We will not discuss here the decidability or undecidability results related to various problems concerning: (i) the rigidity of the call set and (ii) the validity of
interargument relations. The interested reader may refer to the relevant literature.
In the remainder of this paper we provide some answers to the question in the
setting of a given set S, an inferred order based on polynomial interpretations,
abstractions of S based on types, type inference to approximate the call set, and
interargument relations based on inequalities between polynomials.
3 Polynomial Interpretation of a Logic Program
The approach presented in the previous section can be considered a theoretical
framework for termination analysis of LPs based on general orders on terms and
atoms. In this section, we specialize it to orders based on polynomial interpretations.
We first introduce polynomial interpretations in Section 3.1. Then in Section 3.2
we reformulate the termination conditions for LPs from Section 2.3 for polynomial
interpretations.
Polynomial Interpretations for Termination Analysis of Logic Programs
7
3.1 Polynomial Interpretations
In this paper, we only consider polynomials with natural numbers as coefficients
(so-called “natural coefficients”). Because natural numbers will occur many times
in this paper, we will simply refer to them as “numbers”.
We say that a variable X occurs in a polynomial p if the polynomial contains
a monomial with a coefficient different from 0 and X occurs in this monomial. If
X1 , . . . , Xn are all the variables occurring in a polynomial p, we often denote p as
p(X1 , . . . , Xn ). For every polynomial p, there is an associated polynomial function
Fp = λX1 , . . . , Xn . p(X1 , . . . , Xn ). For numbers or polynomials x1 , . . . , xn , we often
write “p(x1 , . . . , xn )” instead of “Fp (x1 , . . . , xn )”. Given p(X1 , . . . , Xn ) and m ≥ 1
we also have an associated polynomial function Fp,m = λX1 , . . . , Xn , Y1 , . . . , Ym .
p(X1 , . . . , Xn ). For such an associated function on an extended domain, we often
write “p(x1 , . . . , xn , y1 , . . . , ym )” to denote “Fp,m (x1 , . . . , xn , y1 , . . . , ym )”.
Definition 7 (orders on polynomials)
Let p and q be two polynomials. Let X1 , . . . , Xn be all variables occurring in p or q.
The quasi-order %N is defined as p %N q if and only if p(x1 , . . . , xn ) ≥ q(x1 , . . . , xn )
for all x1 , . . . , xn ∈ N. The strict order ≻N is defined as p ≻N q if and only if
p(x1 , . . . , xn ) > q(x1 , . . . , xn ) for all x1 , . . . , xn ∈ N.
Observe that (%N , ≻N ) is a reduction pair. In other words, ≻N is well-founded
and transitive, %N is reflexive and transitive, and %N and ≻N are compatible. Let
Σ we denote the set of all polynomials with natural coefficients. Note that all
these polynomials p are weakly monotonic, i.e., xi ≥ yi for all 1 ≤ i ≤ n implies
p(x1 , . . . , xn ) ≥ p(y1 , . . . , yn ).
A polynomial interpretation maps each function and each predicate symbol of
the program to a polynomial.
Definition 8 (polynomial interpretation)
A polynomial interpretation I for a logic program P maps each symbol f of arity
n in Fun P ∪ Pred P to a polynomial pf (X1 , . . . , Xn ).
Every polynomial interpretation induces a norm and a level mapping. Although it
is standard in logic programming to distinguish between norms and level mappings,
to simplify the formalization, here we will only introduce a level mapping and define
it on both terms and atoms.
Definition 9 (polynomial level mapping)
The level mapping associated with a polynomial interpretation I, is a mapping
|.|I : Term P ∪ Atom P → Σ, which is defined recursively as:
- |X|I = X if X is a variable,
- |f (t1 , . . . , tn )|I = pf (|t1 |I , . . . , |tn |I ), where pf = I(f ).
Every polynomial interpretation induces corresponding orders.
8
Manh Thang Nguyen et al.
Definition 10 (reduction pair corresponding to polynomial interpretation)
Let I be a polynomial interpretation. We define the relations %I and ≻I on Term P ∪
Atom P as follows:
- s %I t if and only if |s|I %N |t|I for any s, t ∈ Term P ∪ Atom P
- s ≻I t if and only if |s|I ≻N |t|I for any s, t ∈ Term P ∪ Atom P
Again, observe that the orders induced by a polynomial interpretation form a
reduction pair.
Example 4 (polynomial interpretation for “der”)
Let I be a polynomial interpretation with
I(+) = I(∗)
I(u) = I(1)
I(der )
I(d)
=
=
=
=
p+ (X1 , X2 ) = p∗ (X1 , X2 )
pu
=
p1
pder (X)
pd (X1 , X2 )
=
=
=
=
X1 + X2 + 2
1
X 2 + 2X + 2
X1
Then d(der (X + Y ), DX + DY ) ≻I d(der (X), DX), since |d(der (X + Y ), DX +
DY )|I = (X + Y + 2)2 + 2(X + Y + 2) + 2 ≻N |d(der (X), DX)|I = X 2 + 2X + 2.
3.2 Termination of Logic Programs by Polynomial Interpretations
We now re-state Definition 6 and Theorem 1 for the special case of polynomial
interpretations. So instead of interargument relations for arbitrary orders as in
Definition 4, we now use interargument relations w.r.t. polynomial interpretations.
Definition 11 (interargument relation w.r.t. a polynomial interpretation)
Let P be a program, p be a predicate in P , and I be a polynomial interpretation. Rp
is an interargument relation for p in P w.r.t. I iff Rp is an interargument relation
for p in P w.r.t. (%I , ≻I ).
Instead of rigidity w.r.t. general orders as in Definition 5, we define rigidity w.r.t.
polynomial interpretations.
Definition 12 (rigidity w.r.t. a polynomial interpretation)
A term or atom A ∈ Term P ∪ Atom P is called rigid w.r.t. a polynomial interpretation I iff A is rigid w.r.t. (%I , ≻I ), i.e., iff A ≈I Aσ holds for any substitution σ. A
set of terms (or atoms) S is called rigid w.r.t. I if all its elements are rigid w.r.t. I.
For polynomial interpretations, rigidity can also be characterized in an alternative
way using relevant variables.
Definition 13 (relevant variables)
Let I be a polynomial interpretation and A be a term or atom. A variable X in A
is called relevant w.r.t. I if there exists a substitution {X → t} of a term t for X,
such that A{X → t} 6≈I A.
Example 5 (relevant variables)
Let A = [X|Y ] and I be the interpretation corresponding to the list-length norm
k.kℓ , i.e., |[H|T ]|I = 1 + |T |I . Then the only relevant variable of A is Y .
Polynomial Interpretations for Termination Analysis of Logic Programs
9
Proposition 1 (alternative characterization of rigidity)
Let I be a polynomial interpretation and A be a term or atom. Then A is rigid
w.r.t. I iff A has no relevant variables w.r.t. I.
Proof
Obvious from Definitions 12 and 13.
Using the notions of interargument relations and rigidity w.r.t. a polynomial
interpretation, we obtain the following specialization of Theorem 1:
Corollary 1 (termination criterion with polynomial rigid order-acceptability)
Let S be a set of atomic queries and P be a program. Let I be a polynomial
interpretation, where Call (P, S) is rigid w.r.t. I and where for each predicate p in
P , there is a valid interargument relation Rp in P w.r.t. I such that
- for any clause A :− B1 , B2 , . . . , Bn in P ,
- for any atom Bi ∈ {B1 . . . , Bn } such that rel (Bi ) ⋍ rel(A),
- for any substitution θ such that the atoms B1 θ, . . . , Bi−1 θ are elements of
their associated interargument relations Rrel(B1 ) , . . . , Rrel(Bi−1 ) :
Aθ ≻I Bi θ.
Then P terminates for any query in S.
Proof
The corollary immediately follows from Theorem 1.
Corollary 1 can be applied to verify termination of a logic program w.r.t. a set
of queries. More precisely, we have to check that all conditions in the following
termination proof procedure are satisfied by some polynomial interpretation I. In
Section 4 we will discuss how to find such an interpretation automatically.
Procedure 1 (a procedure for automatic termination analysis)
The termination proof procedure derived from Corollary 1 contains the following
three steps:
Step 1: The call set Call (P, S) must be rigid w.r.t. I. In other words, no
query A in the call set may have a relevant variable w.r.t. I.
Step 2: For a clause that has body-atoms between the head and a (mutually)
recursive body-atom, valid interargument relations of those atoms w.r.t. I
need to be inferred.
Step 3: For every clause, the polynomial level mapping of the head w.r.t. I
should be larger than that of any (mutually) recursive body-atom, given that
interargument relations for intermediate body-atoms hold.
For Step 2, we can follow the standard approach for LPs to verify that a relation
R holds for all elements of the Herbrand model (see e.g. (Lloyd 1987)). To this end,
one has to verify TP (R) ⊆ R, where TP is the immediate consequence operator
corresponding to the program P . Thus, we verify the validity of interargument
relations by first checking whether they are correct for the facts in the program.
Then for every clause, if the interargument relations hold for all body-atoms, the
interargument relation for the head should also hold.
10
Manh Thang Nguyen et al.
Example 6 (applying Corollary 1 to the “der”-program)
Consider again the “der ”-program from Example 1 and the set of queries S =
{d(t1 , t2 ) | t1 is a ground term and t2 is an arbitrary term}. Note that here,
Call (P, S) = S. Let I be the polynomial interpretation from Example 4. Then
no A ∈ Call (P, S) has a relevant variable w.r.t. I. This means that Call (P, S) is
rigid w.r.t. I.
Let Rd = {d(t1 , t2 ) | t1 , t2 ∈ T ermP , t1 ≻I t2 } be an interargument relation for
the predicate d. Checking the validity of Rd is equivalent to verifying the correctness
of the following conditions for any substitution θ:
der (u)θ ≻I (1)θ
der (X)θ ≻I DX θ and der (Y )θ ≻I DY θ implies
der (X + Y )θ ≻I (DX + DY )θ
der (X)θ ≻I DX θ and der (Y )θ ≻I DY θ implies
der (X ∗ Y )θ ≻I (X ∗ DY + Y ∗ DX )θ
der (X)θ ≻I DX θ and der (DX )θ ≻I DDX θ implies
der (der (X))θ ≻I DDX θ.
To prove termination, we also need the following decrease conditions for any
substitution θ:
d(der (X + Y ), DX + DY )θ ≻I d(der (X), DX )θ
d(der (X), DX )θ satisfies Rd implies
d(der (X + Y ), DX + DY )θ ≻I d(der (Y ), DY )θ
d(der (X ∗ Y ), X ∗ DY + Y ∗ DX )θ ≻I d(der (X), DX )θ
d(der (X), DX )θ satisfies Rd implies
d(der (X ∗ Y ), X ∗ DY + Y ∗ DX )θ ≻I d(der (Y ), DY )θ
d(der (der (X)), DDX )θ ≻I d(der (X), DX )θ
d(der (X), DX )θ satisfies Rd implies
d(der (der (X)), DDX )θ ≻D d(der (DX ), DDX )θ
The conditions above are equivalent to the following inequalities on the variables
X, Y, DX , DY , DDX . For the conditions on the valid interargument relation, we
obtain:
5
>
1
∀X, Y, DX , DY ∈ N : X 2 + 2X + 2 > DX ∧ Y 2 + 2Y + 2 > DY ⇒
(X + Y + 2)2 + 2(X + Y + 2) + 2 > DX + DY + 2
∀X, Y, DX , DY ∈ N : X 2 + 2X + 2 > DX ∧ Y 2 + 2Y + 2 > DY ⇒
(X + Y + 2)2 + 2(X + Y + 2) + 2 > x + DY + Y + DX + 3
∀X, DX , DDX ∈ N : X 2 + 2X + 2 > DX ∧ DX 2 + 2DX + 2 > DDX ⇒
(X 2 + 2X + 2)2 + 2(X 2 + 2X + 2) + 2 > DDX
Polynomial Interpretations for Termination Analysis of Logic Programs
11
And for the decrease conditions we obtain:
∀X, Y ∈ N :
∀X, Y, DX ∈ N : X 2 + 2X + 2 > DX ⇒
∀X, Y ∈ N :
∀X, Y, DX ∈ N : X 2 + 2X + 2 > DX ⇒
∀X ∈ N : (X 2
∀X, DX ∈ N : X 2 + 2X + 2 > DX ⇒ (X 2
(X + Y + 2)2 + 2(X + Y
(X + Y + 2)2 + 2(X + Y
(X + Y + 2)2 + 2(X + Y
(X + Y + 2)2 + 2(X + Y
+ 2X + 2)2 + 2(X 2 + 2X
+ 2X + 2)2 + 2(X 2 + 2X
+ 2) + 2
+ 2) + 2
+ 2) + 2
+ 2) + 2
+ 2) + 2
+ 2) + 2
>
>
>
>
>
>
X 2 + 2X + 2
Y 2 + 2Y + 2
X 2 + 2X + 2
Y 2 + 2Y + 2
X 2 + 2X + 2
DX 2 + 2DX + 2
The above inequalities are easily verified for all instantiations of the variables by
numbers. Hence, the program terminates w.r.t. the set of queries S.
4 Automating the Termination Proof
A key question is how to automate the search for a polynomial interpretation and
for interargument relations. In other words, to prove termination of a logic program, one has to synthesize the coefficients of the polynomials associated with
the function and predicate symbols as well as the formulas ϕp (t1 , . . . , tn ) defining
the interargument relations. In the philosophy of the constraint-based approach in
(Decorte et al. 1999), we do not choose a particular polynomial interpretation and
particular interargument relations. Instead, we introduce a general symbolic form
for the polynomials associated with the function and predicate symbols and for the
interargument relations. As an example, assume that polynomials of degree 2 are selected for the interpretation. Then instead of assigning the polynomial pq (X1 , X2 ) =
X12 + 2X1 X2 to a predicate symbol q of arity 2, we would, for example, assign the
symbolic polynomial pq (X1 , X2 ) = q00 + q10 X1 + q01 X2 + q11 X1 X2 + q1 X12 + q2 X22 ,
where the qi and qij are unknown coefficients ranging over N. So our approach for
termination analysis works as follows:
• introduce symbolic versions of the polynomials associated with function and
predicate symbols,
• express all conditions resulting from Corollary 1 as constraints on the coefficients (e.g. q00 , q10 , q01 , . . .),
• solve the resulting system of constraints to obtain values for the coefficients.
Each solution for this constraint system gives rise to a concrete polynomial interpretation and to concrete valid interargument relations such that all conditions of
Corollary 1 are satisfied. Therefore, each solution gives a termination proof.
In order to assign symbolic polynomials to the function and predicate symbols,
we make the decision of assigning linear polynomials to predicate symbols and linear
or simple-mixed polynomials to function symbols. These classes of polynomials are
defined as follows:
- The linear class: each monomial of a polynomial in this class contains at most
one variable of at most degree 1:
Pn
p(X1 , . . . , Xn ) = p0 + k=1 pk Xk
- The simple-mixed class: each monomial of a polynomial in this class contains
either a single variable of at most degree 2 or several variables of at most
12
Manh Thang Nguyen et al.
degree 1:
Pn
P
p(X1 , . . . , Xn ) = jk ∈{0,1} pj1 ...jn X1j1 . . . Xnjn + k=1 pk Xk2
The above classes of polynomials have proved to be particularly useful for automated termination proofs of TRSs. For more details on these classes of polynomials
we refer to (Contejean et al. 2005; Steinbach 1992). In our work, these choices resulted from extensive experiments with different kinds of polynomials, where our
goal was to optimize both the efficiency and the power of the termination analyzer.
In Section 4.1, we first reformulate the conditions of our termination criterion in
Corollary 1, using the above symbolic forms of polynomials. Then in Section 4.2, we
transform these symbolic conditions into constraints on the unknown coefficients of
the symbolic polynomials. Afterwards, in Section 4.3 we show how these resulting
Diophantine constraints can be solved automatically. Finally, we conclude with a
comparison of our contributions with related work from term rewriting in Section
4.4.
4.1 Reformulating the Termination Conditions
In this subsection, we reformulate all termination conditions of Corollary 1, i.e., of
Procedure 1. These include the rigidity property (Step 1), the valid interargument
relations (Step 2), and the decrease conditions (Step 3). The reformulation results in
symbolic constraints, based on the symbolic forms of the polynomial interpretations.
4.1.1 Rigidity Conditions (Procedure 1, Step 1)
There are several ways to approximate Call (P, S) (e.g., (Bruynooghe et al. 2005;
Gallagher et al. 2005; Heaton et al. 2000; Janssens and Bruynooghe 1992)). In this
paper, we apply the approximation technique of (Gallagher et al. 2005; Janssens and Bruynooghe 1992).
More precisely, we first specify the set of queries as a set of rigid type graphs. Then
the technique in (Gallagher et al. 2005; Janssens and Bruynooghe 1992) is used to
compute a new, finite set of rigid type graphs which approximate Call (P, S). Each
of these new rigid type graphs represents a so-called call pattern. For further details,
we refer to (Gallagher et al. 2005; Janssens and Bruynooghe 1992).
In the following, we recapitulate the notion of rigid type graphs and show how
rigidity conditions are derived from the set of call patterns. First, we recall and extend some basic definitions from (Janssens and Bruynooghe 1992), which are based
on linear norms and level-mappings, to the case of general polynomial interpretations. Example 7 will illustrate these definitions.
Definition 14 (rigid type graph (Janssens and Bruynooghe 1992))
A rigid type graph T is a 5-tuple, (Nodes, ForArcs, BackArcs, Label , ArgPos), where
1. Nodes is a finite non-empty set of nodes.
2. ForArcs ⊆ Nodes × Nodes such that (Nodes, ForArcs) is a tree.
3. BackArcs ⊆ Nodes × Nodes such that for every arc (m, n) ∈ BackArcs, node
n is an ancestor of node m in the tree (Nodes, ForArcs).
Polynomial Interpretations for Termination Analysis of Logic Programs
13
4. Label is a function Nodes → Fun P ∪ Pred P ∪ {MAX, OR}.
5. If a node n is labelled with f ∈ Fun P ∪ Pred P and f has arity k, then the
node n has exactly k outgoing arcs (counting both ForArcs and BackArcs).
These arcs are labelled with the numbers 1, . . . , k. For every such arc (n, m),
ArgPos(n, m) returns the corresponding label from {1, . . . , k}.
The intuition behind rigid type graphs is related to the tree representation of
terms and atoms in LP. A rigid type graph generalizes the tree representation of
an atom by allowing:
• nodes labeled by MAX, denoting any term,
• nodes labeled by OR, denoting the union of all denotations of the sub-graphs
rooted at this node,
• backarcs, denoting repeated traversals of a sub-graph.
For each rigid type graph representing a set of atoms S, each node MAX in the
graph corresponds to a possible occurrence of a variable in the atoms of S. The
set S is rigid w.r.t. the polynomial interpretation I iff all these variables are not
relevant w.r.t. I. In the following, we formulate this rigidity condition syntactically
based on the rigid type graph.
Definition 15 (critical path (Decorte et al. 1999))
Let T = (Nodes, ForArcs, BackArcs, Label , ArgPos) be a rigid type graph. A critical
path in T is a path of arcs from the tree ForArcs which goes from the root node of
the tree to a node labelled MAX.
The following proposition is extended from (Decorte et al. 1993), where in (Decorte et al. 1993)
each function or predicate symbol is associated with a linear norm or level mapping.
It provides a method to generate constraints for rigidity.
Proposition 2 (checking rigidity by critical paths)
Let P be a program and T = (Nodes, ForArcs, BackArcs, Label , ArgPos) be a
rigid type graph representing a set of atoms S. Let I be a polynomial interpretation, where for any function or predicate symbol f of arity k we have I(f ) =
P
pf (X1 , . . . , Xk ) = 0≤j1 ,...,jk ≤Mf fj1 ...jk X1j1 . . . Xkjk . The set S is rigid w.r.t. I iff on
every critical path of T there exists an arc (n, m) with Label (n) = f , arity(f ) = k,
P
and ArgPos(n, m) = i such that ji >0 fj1 ...jk = 0, where k is the arity of f .
Proof
Since we only regard polynomials with non-negative coefficients fj1 ...jk , the condiP
tion ji >0 fj1 ...jk = 0 is equivalent to the requirement that fj1 ...jk = 0, whenever
ji > 0. This in turn is equivalent to the condition that Xi is not involved in
pf (X1 , . . . , Xk ). Hence, the condition in the above proposition is equivalent to the
requirement that for any MAX node, there is at least one function or predicate
symbol f on the critical path to this MAX node, for which the argument position
corresponding to the path is not involved in pf . So equivalently, the atoms in the set
S have no relevant variables w.r.t. I. According to Proposition 1, this is equivalent
to rigidity w.r.t. I.
14
Manh Thang Nguyen et al.
Fig. 1. Rigid type graph for Example 7
The following corollary shows how to express the above rigidity check as a constraint on the coefficients of the polynomial interpretation. To this end, we express
the existence condition of an appropriate arc (n, m) by a suitable multiplication.
Corollary 2 (symbolic condition for checking rigidity)
Let T be a rigid type graph representing a set of atoms S and let CP be a
critical path of T . Let (n1 , m1 ), . . . , (ne , me ) be all arcs in CP such that for all
d ∈ {1, . . . , e}, Label (nd ) = f d is a function or predicate symbol of some arity k d and
ArgPos(nd ,
md ) = id . If for any such CP we have
e
Y
d=1
(
X
fjd1 ...j(kd ) ) = 0,
(5)
j(id ) > 0
then S is rigid w.r.t. I.
Example 7 (symbolic polynomial interpretation and rigidity constraints for the “der”-program)
For Example 1, we define a symbolic polynomial interpretation I as follows.
I(+)
I(∗)
I(der )
I(u)
I(1)
I(d)
=
=
=
=
=
=
p1 X12 + p2 X22 + p11 X1 X2 + p10 X1 + p01 X2 + p00
m1 X12 + m2 X22 + m11 X1 X2 + m10 X1 + m01 X2 + m00
der 2 X 2 + der 1 X + der 0
cu
c1
d0 + d1 X1 + d2 X2
We will reformulate the termination conditions for this example in symbolic form.
However for reasons of space, we will not give all polynomial constraints. Instead, in
order to illustrate the main ideas, in each sub-section we only present one constraint
for the corresponding type of conditions.
Polynomial Interpretations for Termination Analysis of Logic Programs
15
Instead of checking termination of the “der ”-program w.r.t. the set of queries
S = {d(t1 , t2 ) | t1 is a ground term, t2 is an arbitrary term} as in Example 1, we
now regard the set of queries S1 = {d(t1 , t2 ) | t1 is of the form der(t′1 ), where t′1
is a ground term constructed from the function symbols u, +, ∗, der , and t2 is an
arbitrary term}. S1 is represented by the type graph in Figure 1.
Obviously, termination of the program w.r.t. S1 also implies termination w.r.t.
S. This can be proved easily by showing that for any query Q ∈ S \ S1 , the program
trivially terminates by finite failure.
In our example, type inference (Janssens and Bruynooghe 1992) computes the
call set Call (P, S1 ) = S1 , i.e., the graph in Figure 1 also represents Call (P, S1 ). Its
only critical path consists of just the arc from the root to the node labelled MAX.
Hence from the graph, the following rigidity condition is generated according to
Corollary 2:
d2 = 0
4.1.2 Valid Interargument Relations (Procedure 1, Step 2)
Next we consider the other symbolic constraints, derived for valid interargument
relations and decrease conditions. We will show that they all take the form:
∀X ∈ N :
p1 ≥ q1 ∧ . . . ∧ pn ≥ qn
⇒
pn+1 ≥ qn+1
(6)
where n ≥ 0 and pi , qi are polynomials with natural coefficients. Here, X is the
tuple of all variables occurring in p1 , . . . , pn+1 , q1 , . . . , qn+1 .
There are a number of works on inferring valid interargument relations of predicates. In (Decorte et al. 1999), interargument relations are formulated as inequalities between a linear combination of the “inputs” and a linear combination of the
“outputs”. We will not define input and output arguments formally in this paper,
since we do not use them in our approach, but informally, inputs are the arguments
of a predicate symbol which are only called with ground terms and outputs are the
remaining arguments.
We propose a new form of interargument relation, namely polynomial interargument relations, which are of the following form:
Rp = {p(t1 , . . . , tn ) | ip (|t1 |I , . . . , |tn |I ) %N op (|t1 |I , . . . , |tn |I )}
(7)
where ip and op are polynomials with natural coefficients.
The form of interargument relations in (Decorte et al. 1999) can be considered a
special case of the form (7) above, where ip (|t1 |I , . . . , |tn |I ) is constructed from the
input arguments only and op (|t1 |I , . . . , |tn |I ) is only constructed from the outputs.
Since the approach in (Decorte et al. 1999) only considers relations between the
input and output arguments of the predicates, it has some limitations. In some cases,
the desired relation does not compare inputs with outputs, but the relation holds
among the inputs only or among the outputs only. In particular, if all arguments of
16
Manh Thang Nguyen et al.
a predicate are inputs (or outputs), then the approach in (Decorte et al. 1999) fails
to infer any useful relation among them. The following example shows this point.
It computes the natural division of the first and second arguments of the predicate
div and returns the result in its third argument.
Example 8 (div )
div (X, s(Y ), 0) :− less(X, s(Y )).
div (X, s(Y ), s(Z)) :− sub(X, s(Y ), R), div (R, s(Y ), Z).
(8)
sub(X, 0, X).
sub(s(X), s(Y ), Z) :− sub(X, Y, Z).
less(0, s(Y )).
less(s(X), s(Y )) :− less(X, Y ).
We consider the set of queries S = { div (t1 , t2 , t3 ) | t1 and t2 are ground terms,
and t3 is an arbitrary term}. This program terminates for all these queries. If we look
at Clause (8), the decrease in size between the head and the recursive body-atom can
be established if we can infer a suitable valid interargument relation for sub. This
relation should imply that within Clause (8), the first argument of sub is greater
than its third argument. However, if we apply the approach in (Decorte et al. 1999),
inferring such an interargument relation for sub is impossible. Since the first two
sub-arguments are used as input and the last one is output, the approach can only
infer interargument relations where a linear combination of the sizes of the first
and second arguments is greater than or equal to the size of the third argument.
Then, we cannot conclude that for every successful answer substitution for the call
sub(X, s(Y ), R) in Clause (8), the first sub-argument X is strictly greater than the
third sub-argument R.
In contrast, if we use Form (7), then it is possible to infer the following valid
interargument relation for sub:
Rsub = {sub(t1 , t2 , t3 ) | |t1 |I %N |t2 |I + |t3 |I }
Note that in the right-hand side |t2 |I + |t3 |I of the above inequality, we have both
an input argument t2 and an output argument t3 . This valid polynomial interargument relation guarantees that for any successful answer substitution for the
call sub(X, s(Y ), R) in Clause (8), we have |X|I ≻N |R|I if |s(Y )|I %N 1. Our
implementation in the system Polytool is indeed able to infer this interargument
relation using the constraint solving technique explained below. Therefore, Polytool
can prove termination of “div ”. If we used the form of interargument relations in
(Decorte et al. 1999) instead, Polytool would not be able to solve this problem. .
Similar to the symbolic form of polynomial interpretations, we also use a symbolic form of polynomial interargument relations. To this end, we take symbolic
polynomials ip and op . For the inference of valid interargument relations, we then
apply the technique proposed in (Decorte et al. 1999), cf. Procedure 1, Step 2.
For any sequence of terms t1 , . . . , tn , let Rp (t1 , . . . , tn ) abbreviate the inequality
Polynomial Interpretations for Termination Analysis of Logic Programs
17
ip (|t1 |I , . . . , |tn |I ) ≥ op (|t1 |I , . . . , |tn |I ). The goal is to impose constraints on the
polynomials ip and op which ensure that the corresponding interargument relation
Rp = {p(t1 , . . . , tn ) | ∀X ∈ N : Rp (t1 , . . . , tn )} is valid. To this end, we generate
for every clause of the program:
p(t) :− p1 (t1 ), . . . , pn (tn )
the constraint
∀X ∈ N :
Rp1 (t1 ) ∧ . . . ∧ Rpn (tn ) ⇒ Rp (t).
It is clear that this formula has Form (6).
Example 9 (symbolic interargument relation for the “der”-program)
We continue Example 7 and use linear polynomials for ider and oder , i.e., ider (X, Y ) =
i0 + i1 X + i2 Y and oder = o0 + o1 X + o2 Y . Hence, the the symbolic form of the
polynomial interargument relation for the predicate d is
Rd = {d(t1 , t2 ) | i0 + i1 |t1 |I + i2 |t2 |I %N o0 + o1 |t1 |I + o2 |t2 |I }.
There are four clauses (1) - (4) from which constraints for valid interargument
relations are inferred. We only present the constraint resulting from the last clause
(4):
d(der (der (X)), DDX ) :− d(der (X), DX ), d(der (DX ), DDX )
Here, we obtain the constraint
∀X, DX , DDX ∈ N :
Rd (der (X), DX ) ∧ Rd (der (DX ), DDX ) ⇒ Rd (der (der (X)), DDX ). (9)
4.1.3 Decrease Conditions (Procedure 1, Step 3)
Finally, one has to require the decrease condition between the head and any (mutually) recursive body-atom in any (mutually) recursive clause. So for any clause
p(t) :− p1 (t1 ), . . . , pn (tn )
of the program where p ⋍ pi (i.e., where p and pi are mutually recursive), we require
∀X ∈ N :
Rp1 (t1 ) ∧ . . . ∧ Rpi−1 (ti−1 ) ⇒ |p(t)|I ≥ |pi (ti )|I + 1.
Obviously, the formula is in Form (6).
Example 10 (constraints for the decrease conditions of “der”)
There are three recursive clauses (2) - (4) where decrease conditions can be inferred.
We present the decrease condition for the recursive body-atom d(der (DX ), DDX )
18
Manh Thang Nguyen et al.
of the last clause (4):
∀X, DX , DDX ∈ N :
i0 + i1 (der 2 X 2 + der 1 X + der 0 ) + i2 DX ≥
o0 + o1 (der 2 X 2 + der 1 X + der 0 ) + o2 DX
⇒
(10)
2
2
d0 + d1 (der 2 (der 2 X + der 1 X + der 0 ) +
der 1 (der 2 X 2 + der 1 X + der 0 ) + der 0 ) + d2 DDX ≥
d0 + d1 (der 2 DX 2 + der 1 DX + der 0 ) + d2 DDX + 1.
4.2 From Symbolic Conditions to Constraints on Coefficients
Our goal is to find a polynomial interpretation such that all constraints generated
in the previous section are satisfied. To this end, we transform all these constraints
into Diophantine constraints. In this transformation, we first eliminate implications,
cf. Section 4.2.1. Afterwards, in Section 4.2.2, the universally quantified variables
(e.g., X, DX, DDX, . . .) are removed and the former unknown coefficients (e.g.,
der 0 , der 1 , der 2 , . . .) become the new variables. If the resulting Diophantine constraints can be solved, then the program under consideration is terminating.
As we analyzed in Section 4.1.1, all generated rigidity constraints have the Form
(5). Hence, these are already Diophantine constraints which only contain unknown
coefficients, but no universally quantified variables.
The other constraints, generated for the valid interargument relations and the
decrease conditions, have the following form:
∀X ∈ N :
p1 ≥ q1 ∧ . . . ∧ pn ≥ qn
⇒
pn+1 ≥ qn+1 ,
(6)
where n ≥ 0 and pi , qi are polynomials with natural coefficients.
In the following, we introduce a two-phase method to transform all constraints
of Form (6) into Diophantine constraints on the unknown coefficients.
4.2.1 First Phase: Removing Implications
The constraints of Form (6) are implications. In the first phase, such constraints are
transformed into inequalities without premises, i.e., into constraints of the form
∀X ∈ N :
p ≥ 0.
(11)
However, here p is a polynomial with integer (i.e., possibly negative) coefficients.
The transformation is sound : if the new constraints of Form (11) are satisfied by
some substitution which instantiates the unknown coefficients with numbers, then
this substitution also satisfies the original constraints of Form (6).
The idea for the transformation is the following. Constraints of the form (6) may
Polynomial Interpretations for Termination Analysis of Logic Programs
19
have an arbitrary number n of premises pi ≥ qi . We first transform them into
constraints with at most one premise. Obviously, p1 ≥ q1 ∧ . . . ∧ pn ≥ qn implies
p1 + . . . + pn ≥ q1 + . . . qn . Thus, instead of (6), it would be sufficient to demand
∀X ∈ N :
p1 + . . . + pn ≥ q1 + . . . qn
⇒
pn+1 ≥ qn+1 .
So in order to combine the n polynomials in the premise, we can use the polynomial prem(X1 , . . . , Xn ) = X1 + . . . + Xn . Then instead of (6), we may require
∀X ∈ N :
prem(p1 , . . . , pn ) ≥ prem(q1 , . . . , qn )
⇒
pn+1 ≥ qn+1 .
A similar method was also used for termination analysis of logic programs in
(Decorte et al. 1999) and for termination of term rewriting in (Giesl et al. 2006,
Section 7.2) to transform disjunctions of polynomial inequalities into one single
inequality.
For example, the constraint
∀X1 , X2 , X3 ∈ N :
X1 ≥ X2 ∧ X2 ≥ X3
⇒
X1 ≥ X3
X1 + X2 ≥ X2 + X3
⇒
X1 ≥ X3
can now be transformed into
∀X1 , X2 , X3 ∈ N :
Since the latter constraint is valid, the former one is valid as well.
However, in order to make the approach more powerful, one could also use other
polynomials prem in order to combine the n inequalities in the premise. The reason
is that if prem is restricted to be the addition, then many valid constraints of the
form (6) would be transformed into invalid ones. For example, the valid constraint
∀X1 , X2 , X3 ∈ N :
X1 ≥ X22 ∧ X2 ≥ X32
⇒
X1 ≥ X34
⇒
X1 ≥ X34 .
would be transformed into the invalid constraint
∀X1 , X2 , X3 ∈ N :
X1 + X2 ≥ X22 + X32
For instance, the constraint does not hold for X1 = 4, X2 = 0, and X3 = 2.
To make the transformation more general and more powerful, we therefore permit the use of arbitrary polynomials prem with natural coefficients. In the above
example, now the resulting constraint
∀X1 , X2 , X3 ∈ N :
prem(X1 , X2 ) ≥ prem(X22 , X32 )
⇒
X1 ≥ X34
would indeed be valid for a suitable choice of prem. For instance, one could choose
prem to be the addition of the first argument with the square of the second argument
(i.e., prem(X1 , X2 ) = X1 + X22 ).
By the introduction of the new polynomial prem, every constraint of the form (6)
can now be transformed into an implication with at most one premise. It remains
to transform such implications further into unconditional inequalities. Obviously,
instead of
prem(p1 , . . . , pn ) ≥ prem(q1 , . . . , qn ) ⇒ pn+1 ≥ qn+1 ,
(12)
it is sufficient to demand
pn+1 − qn+1 ≥ prem(p1 , . . . , pn ) − prem(q1 , . . . , qn ).
(13)
20
Manh Thang Nguyen et al.
This observation was already used in the work of (Decorte et al. 1999) and also in
termination techniques for term rewriting to handle such conditional polynomial
inequalities (Brauburger and Giesl 1998; Giesl et al. 2007).
However, the approach can still be improved. Recall that we used an arbitrary
polynomial prem to combine the polynomials in the former premises. In a similar
way, one could also apply an arbitrary polynomial conc to the polynomials pn+1
and qn+1 in the former conclusion. To see why this can be necessary, consider the
valid constraint
∀X ∈ N :
2X ≥ 2
⇒
X ≥ 1.
With the transformation of (12) into (13) above, it would be transformed into the
unconditional constraint
∀X ∈ N :
X − 1 ≥ 2X − 2,
which is invalid. We have encountered several examples of this kind in our experiments, which motivates this further extension. In such examples, it would be better
to apply a suitable polynomial conc to the polynomials X and 1 in the former
conclusion. Then we would obtain
∀X ∈ N :
conc(X) − conc(1) ≥ 2X − 2
instead. By choosing conc(X) = 2X, now the resulting constraint is valid.
So to summarize, in the first phase of our transformation, any constraint of the
form (6) is transformed into the unconditional constraint
∀X ∈ N : conc(pn+1 ) − conc(qn+1 ) ≥ prem(p1 , . . . , pn ) − prem(q1 , . . . , qn ). (14)
Here, prem and conc are two arbitrary new polynomials. The only requirement that
we have to impose is that conc must not be a constant. Indeed, if conc would be
a constant, then (14) no longer implies that (12) holds for all instantiations of the
variables in the polynomials p1 , . . . , pn+1 , q1 , . . . , qn+1 . Note that we do not need a
similar requirement on prem. If a constant prem would satisfy (14), then (6) trivially
holds. The following proposition proves the soundness of this transformation.
Proposition 3 (Soundness of Removing Implications)
Let prem and conc be two polynomials with natural coefficients, where conc is not
a constant. Moreover, let p1 , . . . , pn+1 , q1 , . . . , qn+1 be arbitrary polynomials with
natural coefficients. If
∀X ∈ N :
conc(pn+1 ) − conc(qn+1 ) − prem(p1 , . . . , pn ) + prem(q1 , . . . , qn ) ≥ 0
is valid, then
∀X ∈ N :
p1 ≥ q1 ∧ . . . ∧ pn ≥ qn
⇒
pn+1 ≥ qn+1
is also valid.
Proof
For any tuple of numbers x, let pi (x) and qi (x) denote the numbers that result from
Polynomial Interpretations for Termination Analysis of Logic Programs
21
pi and qi by instantiating the variables X by the numbers x. So if p(X1 , X2 ) is the
polynomial X12 + 2X1 X2 , then p(2, 1) = 8.
Suppose that there is a tuple of numbers x with pi (x) ≥ qi (x) for all i ∈
{1, . . . , n}. We have to show that then pn+1 (x) ≥ qn+1 (x) holds as well.
Since prem only has natural coefficients, it is weakly monotonic. Thus, pi (x)
≥ qi (x) for all i ∈ {1, . . . , n} implies prem(p1 (x), . . . , pn (x)) ≥ prem(q1 (x),
. . . , qn (x)) and thus, prem(p1 (x), . . . , pn (x)) − prem(q1 (x), . . . , qn (x)) ≥ 0. The prerequisites of the proposition ensure
conc(pn+1 ) − conc(qn+1 ) ≥ prem(p1 , . . . , pn ) − prem(q1 , . . . , qn )
for all instantiations of the variables. Hence,
conc(pn+1 (x)) − conc(qn+1 (x)) ≥ 0 or, equivalently,
we
also
obtain
conc(pn+1 (x)) ≥ conc(qn+1 (x)).
(15)
Now suppose that pn+1 (x) 6≥ qn+1 (x). Since pn+1 (x) and qn+1 (x) are numbers
(not polynomials with variables), we would then have pn+1 (x) < qn+1 (x). Since
conc only has non-negative coefficients and since it is not a constant, it is strictly
monotonic. Thus, pn+1 (x) < qn+1 (x) would imply
conc(pn+1 (x)) < conc(qn+1 (x))
in contradiction to (15). Hence, we have pn+1 (x) ≥ qn+1 (x), as desired.
For the symbolic form of prem and conc, we again choose linear or simple-mixed
polynomials. From our experiments, this choice provided good results on the benchmark programs, while remaining reasonably efficient. By applying Proposition 3,
we can now transform all constraints for the termination proof into unconditional
constraints of the form (11). If there exists a substitution of the unknown coefficients by numbers that makes the resulting unconditional constraints valid, then
the same substitution also satisfies the original conditional constraints.
Example 11 (applying Proposition 3 to the “der”-program)
We choose the decrease condition (10) in Example 10 as an example showing how
to transform an implication into an unconditional constraint.
Since the constraint (10) has only one premise, here the polynomial prem has
arity 1. We choose a simple-mixed form for prem and a linear form for conc:
prem(X) = prem 0 + prem 1 X + prem 2 X 2
conc(X) = conc 0 + conc 1 X.
Since conc must not be a constant, one also has to impose the constraint
conc 1 > 0.
Now we can transform (10) into an unconditional constraint. Here, we use the
22
Manh Thang Nguyen et al.
following abbreviations:
p1
q1
p2
q2
= i0 + i1 (der 2 X 2 + der 1 X + der 0 ) + i2 DX
= o0 + o1 (der 2 X 2 + der 1 X + der 0 ) + o2 DX
= d0 + d1 (der 2 (der 2 X 2 + der 1 X + der 0 )2 +
der 1 (der 2 X 2 + der 1 X + der 0 ) + der 0 ) + d2 DDX
= d0 + d1 (der 2 DX 2 + der 1 DX + der 0 ) + d2 DDX + 1
Then (10) is the constraint
∀X, DX , DDX ∈ N :
p1 ≥ q1
⇒ p2 ≥ q2
and its transformation yields
∀X, DX , DDX ∈ N : conc 0 + conc 1 p2 − conc 0 − conc 1 q2
−prem 0 − prem 1 p1 − prem 2 p21
+prem 0 + prem 1 q1 + prem 2 q12
≥ 0.
By applying standard simplifications, the constraint can be rewritten to the following form:
∀X, DX ∈ N : M1 X 4 + M2 X 3 + M3 X 2 + M4 X+
M5 DX 2 + M6 DX + M7 X 2 DX + M8 XDX + M9 ≥ 0
(16)
where M1 , . . . , M9 are polynomials over the unknown coefficients prem j , ij , oj ,
der j , and dj with j ∈ {0, 1, 2} and conc j with j ∈ {0, 1}. For example, we have
M1 =def conc 1 d1 der 32 + prem 2 o21 der 22 − prem 2 i21 der 22 .
4.2.2 Second Phase: Removing Universally Quantified Variables
In this phase, we transform any constraint of the form
∀X ∈ N :
p≥0
(11)
into a set of Diophantine constraints on the unknown coefficients. The transformation is again sound : if there is a solution for the resulting set of Diophantine
constraints, then this solution also satisfies the original constraint (11).
We use a straightforward transformation proposed by (Hong and Jakuš 1998),
which is also used in all related tools for termination of term rewriting. One only
requires that all coefficients of the polynomial p are non-negative integers. Obviously, the criterion is only sufficient, because, for instance, p(X) = (X − 1)2 ≥ 0,
but X 2 − 2X + 1 does not have non-negative coefficients only.
Example 12 (removing universally quantified variables for the “der”-program)
We continue the transformation of Example 11. Here, we obtained the constraint
(16). We derive the following set of Diophantine constraints which contains the
unknown coefficients conc j , prem j , ij , oj , der j , and dj as variables: M1 ≥ 0, M2 ≥
0, . . . , M9 ≥ 0.
Polynomial Interpretations for Termination Analysis of Logic Programs
23
4.3 Solving Diophantine Constraints
The previous sections showed that one can formulate all termination conditions in
symbolic form and that one can transform them automatically into a set of Diophantine constraints. The problem then becomes solving a system of non-linear Diophantine constraints with the unknown coefficients as variables. If the Diophantine
constraints are solvable, then the logic program under consideration is terminating.
Solving such problems has been studied intensively, especially in the context of
constraint logic programming. Moreover, there are approaches from termination
of term rewriting in order to solve such restricted Diophantine constraints automatically e.g., (Borralleras et al. 2009; Contejean et al. 2005; Fuhs et al. 2007).
In (Fuhs et al. 2007), Diophantine constraints are encoded as a SAT-problem, and
then a SAT solver is used to solve the resulting SAT-problem. As shown in (Fuhs et al. 2007),
this approach is significantly more efficient than solving Diophantine constraints by
dedicated solvers like (Contejean et al. 2005) or by standard implementations of
constraint logic programming like in SICStus Prolog.
Example 13 (solving Diophantine constraints for the“der”-program)
We start with the symbolic polynomial interpretation from Example 7 (e.g., with
I(der )
= der 2 X 2 + der 1 X + der 0 ) and obtain the solution der 2 = 1 and der 0 = der 1 = 2,
which corresponds to X 2 + 2X + 2. Similarly, we start with the symbolic form of
the polynomial interargument relation as in Example 9:
Rd = {d(t1 , t2 ) | i0 + i1 |t1 |I + i2 |t2 |I %N o0 + o1 |t1 |I + o2 |t2 |I }.
Then we get the solution i1 = 1, i0 = i2 = 0, o2 = 1, o0 = o1 = 0. This corresponds
to the interargument relation Rd = {d(t1 , t2 ) | |t1 |I %N |t2 |I }. So we obtain the
concrete simple-mixed polynomial interpretation from Example 4 and the concrete
interargument relation from Example 6.
4.4 Relation to Approaches from Term Rewriting
Finally, we briefly discuss the connection between our approach for automated LP
termination proofs from Section 4.1 - 4.3 and related approaches used for termination analysis of TRSs.
Section 4.1 describes how to obtain constraints for a symbolic polynomial order
which guarantee that the requirements of our termination criterion are fulfilled.
This is similar to related approaches used in term rewriting. Here, one also chooses
a symbolic polynomial interpretation and constructs corresponding inequalities. If
one applies polynomial interpretations directly for termination analysis of TRSs,
then these inequalities ensure that every rewrite rule is strictly decreasing. If one
uses more sophisticated termination techniques like the dependency pair method
(Arts and Giesl 2000; Giesl et al. 2006; Hirokawa and Middeldorp 2005), then one
builds inequalities which ensure that dependency pairs are weakly or strictly decreasing and that rules are weakly decreasing. The decrease conditions of dependency pairs correspond to our decrease conditions in Section 4.1.3 and the re-
24
Manh Thang Nguyen et al.
quirement that rules are weakly decreasing roughly corresponds to our symbolic
constraints for valid interargument relations in Section 4.1.2. Still, there are subtle differences. For example, in LPs, a predicate symbol may have several output
arguments which is the reason for the different polynomials ip and op in our polynomial interargument relations. Moreover, while term rewriting uses matching for
evaluation, in logic programming one uses unification. This is the reason for our
additional rigidity conditions in Section 4.1.1.
The approach in Section 4.2 shows how to find suitable values for the symbolic
coefficients. This is the same problem as in the corresponding techniques for term
rewriting. However, the usual techniques in term rewriting can only handle unconditional inequalities. Therefore, we have developed a new method in Section 4.2.1
to remove conditions. This is a new contribution of the present paper. In fact, after having developed this contribution for the current paper, due to its success in
the tool Polytool, two of the authors of the current paper later even adapted this
method to term rewriting (see (Fuhs et al. 2008, Footnote 14)).
The techniques of the short sections 4.2.2 and 4.3 are identical to the corresponding approaches used in term rewriting. We only included them here in order to have
a self-contained presentation of our approach and to finish its illustration with the
“der ”-example.
5 Experimental Evaluation
In this section we discuss the experimental evaluation of our approach. We implemented our technique in a system called Polytool (Nguyen and De Schreye 2007)
written in SICStus Prolog.1 Essentially, the Polytool system consists of four modules: The first module is the type inference engine, where we use the inference system
of (Gallagher et al. 2005). The second module generates all termination conditions
using symbolic polynomials as in Section 4.1. The third module transforms the resulting polynomial constraints into Diophantine constraints, as in Section 4.2. The
final module is a Diophantine constraint solver, cf. Section 4.3. We selected the SATbased Diophantine solver (Fuhs et al. 2007) of the AProVE tool (Giesl et al. 2006).
We tested the performance of Polytool on a collection of 296 examples. The
collection (Table 1) consists of all benchmarks for logic programming from the Termination Problem Data Base (TPDB),2 where all examples that contain arithmetic
or built-in predicates were removed.
Polytool applies the following strategy: first, we search for a linear polynomial
interpretation. If we cannot find such an interpretation satisfying the termination
conditions, then we search for a simple-mixed polynomial interpretation. More precisely, then we still interpret predicate symbols by linear polynomials, but we map
function symbols to simple-mixed polynomials. We use similar symbolic polynomials for conc and prem from Section 4.2.1: if the polynomial interpretation is linear,
then both conc and prem are linear. Otherwise, we use a linear form for conc and
1
2
For the source code, we refer to http://www.cs.kuleuven.be/~ manh/polytool.
http://www.termination-portal.org/wiki/Termination_Competition
Polynomial Interpretations for Termination Analysis of Logic Programs
25
a simple-mixed form for prem. The domain for all unknown coefficients in the generated Diophantine constraints is fixed to the set {0, 1, 2}. The experiments were
performed on an AMD 64 bit, 2GB RAM running Linux.
We performed an experimental comparison with other leading systems for automated termination analysis of logic programs, namely: Polytool-WST07, cTI-1.1
(Mesnard and Bagnara 2005), TerminWeb (Codish and Taboch 1999; Taboch et al. 2002),
TALP (Ohlebusch et al. 2000) and AProVE (Giesl et al. 2006). For TALP, the option
of non-linear polynomial interpretations was chosen. For cTI-1.1, we selected the
“default” option. For AProVE and TerminWeb, the fully automatic modes were chosen. We did not include the tool Hasta-La-Vista (Serebrenik and De Schreye 2003)
in the evaluation because it is a predecessor of Polytool. We used a time limit of
60 seconds for testing each benchmark on each termination tool. This time limit is
also used in the annual termination competition.
In Table 1, we give the numbers of benchmarks which are proved terminating
(”YES”), the number of benchmarks which could not be proved terminating but
where processing ended within the time limit (”FAILURE”), and the number of
benchmarks where the tool did not stop before the timeout (”TIMEOUT”). The
number in square brackets is the average runtime (in seconds) that a particular
tool uses to prove termination of benchmarks (or fails to prove termination of
them within the time limit). The detailed experiments (including also the source
code of the benchmarks and the termination proofs produced by the tools) can be
found at http://www.cs.kuleuven.be/~manh/polytool/POLY/journal07.html.
Note that the two examples der and div presented in this paper do not occur in
the TPDB. For completeness we just mention that Polytool and AProVE succeed on
der , whereas cTI-1.1 and TerminWeb fail, and TALP reaches the timeout. For div ,
all systems except TALP succeed. In the next sub-sections we discuss the results of
the experiments. For a more detailed discussion, we refer to (Nguyen 2009).
TALP
cTI-1.1
TerminWeb
Polytool
AProVE
YES
163 [2.54]
167 [0.06]
177 [0.54]
214 [4.28]
232 [6.34]
FAILURE
112 [1.45]
129 [0.05]
118 [0.6]
62[10.48]
57 [19.08]
TIMEOUT
21
0
1
20
7
Table 1. The results for 296 benchmarks of the TPDB
5.1 Comparison between Polytool and cTI-1.1
Similar to Polytool, cTI-1.1 deploys a global constraint-based approach to termination analysis. However, different from Polytool, in cTI-1.1 termination inference of
the analyzed program relies on its two main abstract approximations: a program in
26
Manh Thang Nguyen et al.
CLP(N), where all terms of the program are mapped to expressions in N according
to a fixed symbolic norm (e.g., the symbolic3 term-size norm by default), and a
program in CLP(B), where B denotes the booleans, which is obtained from the program in CLP(N) by mapping any number to 1, any variable to itself, and addition
to logical conjunction. The purpose of these abstractions is to capture the decrease
conditions (the program in CLP(N)) and the boundedness information (the program
in CLP(B)) of the program.
As shown in Table 1, Polytool outperforms cTI-1.1. The only benchmark where
cTI-1.1 can prove termination and Polytool fails is the example incomplete2.pl in
the directory SGST06 of the TPDB. However, if we reset the range for the values
of the unknown coefficients in the generated Diophantine constraints to {0, . . . , 8},
then Polytool can prove termination for the example as well.
There are several reasons for the less powerful performance of cTI-1.1 in comparison with Polytool. First of all, cTI-1.1 uses a fixed symbolic norm to map the
analyzed program to a program in CLP(N), for which all termination conditions
are formulated. However, in some cases, the selected symbolic norm is not suitable
to capture the decrease in the analyzed program. Then as a result, cTI-1.1 cannot
prove termination. The TPDB contains a number of such benchmarks, e.g., flat.pl,
normal.pl in the talp directory and countstack.pl, factor.pl, flatten.pl in the SGST06
directory.
Secondly, when we use the term-size or list-length norm for the abstract approximation in cTI-1.1, all constant symbols are mapped to the same number in N. As a
result, cTI-1.1 fails for examples where the difference among constant symbols plays
a role for the termination behavior. In Polytool, different constant symbols can be
mapped to different numbers in N. Therefore, termination of examples such as simple.pl in the talp directory, pl2.3.1.pl in the plumer directory, at.pl in the SGST06
directory, etc. can be proved, whereas cTI-1.1 fails.
Thirdly, since termination analysis of cTI-1.1 is based on linear symbolic norms, it
cannot prove termination of programs such as Example 1 or the example hbal tree.pl
in the TPDB. In contrast, Polytool can prove termination of these examples using
simple-mixed polynomial interpretations.
Finally, there are examples like applast.pl, bappend.pl, blist.pl, btappend.pl, btapplast.pl, confdel.pl and btree.pl in the SGST06 directory, whose termination cannot
be proved by cTI-1.1, since cTI-1.1 only uses groundness instead of type analysis.
The termination proof of these examples also fails with TALP for the same reason. In
contrast, Polytool and AProVE succeed for them and TerminWeb succeeds for some
of them (i.e., applast.pl, bappend.pl, blist.pl, confdel.pl). The success of Polytool and
TerminWeb is due to the use of types instead of modes and AProVE succeeds because of so-called argument filterings which remove argument positions of function
and predicate symbols that are irrelevant for termination. But TerminWeb still fails
on some of these examples, since it uses a fixed norm for part of its analysis.
3
The difference between the “term-size norm” and the “symbolic term-size norm” is that the
“term-size norm” maps all variables to 0, whereas the “symbolic term-size norm” maps any
variable to itself (as in polynomial interpretations).
Polynomial Interpretations for Termination Analysis of Logic Programs
27
A strong point of cTI-1.1 is that it is very fast (it is by far the fastest tool in the
experiments). The reason is that cTI-1.1 fixes the norm in advance. Therefore it requires much less unknown coefficients to formulate termination conditions. Another
strong point of cTI-1.1 is its ability of performing termination inference (i.e., it can
try to detect all terminating modes for a program), which is impossible for Polytool at this moment. Finally, recent extensions of cTI-1.1 include non-termination
proofs, which are not supported by the other systems in our experiments.
5.2 Comparison between Polytool and TerminWeb
Similar to cTI-1.1, TerminWeb also uses fixed symbolic norms, e.g., the term-size
norm, the list-length norm, or (as in our experiments) a combination of type-based
norms (Bruynooghe et al. 2007) to approximate the analyzed program. Therefore,
it has similar problems as cTI-1.1. In fact, termination of examples such as flat.pl,
normal.pl, countstack.pl, factor.pl, flatten.pl discussed in Section 5.1 cannot be proved
by TerminWeb either.
Different from Polytool and cTI-1.1, TerminWeb applies a local approach to termination analysis, where different norms and level mappings are used for different
loops in the program (Codish and Taboch 1999). Hence, TerminWeb can prove termination of a class of programs where lexicographic orders are required (e.g., the
benchmarks ackermann.pl and vangelder.pl in the TPDB). In fact, these programs
could already be proven terminating by TermiLog (Lindenstrauss and Sagiv 1997;
Lindenstrauss 2000), the first generally available automatic termination analyzer for
LPs. TermiLog succeeds on these programs due to the query-mapping pairs approach
(Lindenstrauss et al. 2004), which has some similarity to the dependency pair approach (Arts and Giesl 2000; Giesl et al. 2006; Hirokawa and Middeldorp 2005). For
termination of such programs, the global technique based on polynomial interpretations deployed in Polytool is insufficient. We are working on an extension using dependency graphs that is able to deal with such programs as well (Nguyen et al. 2008;
Schneider-Kamp et al. 2009).
Similar to cTI-1.1, TerminWeb is much faster than Polytool. This is again due to
the fact that TerminWeb uses a fixed symbolic norm to approximate the analyzed
program.
5.3 Comparison between Polytool, AProVE, and TALP
A point of similarity between Polytool, TALP, and AProVE is that all these systems
use polynomial interpretations as the basis for the termination analysis. However in
TALP and AProVE, polynomial interpretations are applied indirectly: given a logic
program and a set of queries, these tools first transform them into a TRS whose
termination is sufficient for the termination of the original logic program. Then,
termination analysis is applied to the resulting TRS. Due to this transformational
approach, several other termination techniques developed for TRSs become applicable for the analysis of LPs as well. In particular, AProVE uses many different
methods for proving termination.
28
Manh Thang Nguyen et al.
A limitation of the transformational approach in TALP is that it can only handle
well-moded logic programs. There are many non-well-moded examples in the TPDB
that can be solved by most other tools but not by TALP.
AProVE instead applies a quite strong transformational approach, which can also
deal with non-well-moded logic programs (Schneider-Kamp et al. 2009). Together
with the powerful back-end TRS termination prover, this makes AProVE a very
strong LP termination system. In fact, in both our experiments and in the termination competitions, AProVE was always in the first place. In particular, it can prove
termination of most examples whenever some other tool can. Nevertheless, there
exists one example in the TPDB (i.e., incomplete.pl) where AProVE fails to prove
its termination but Polytool succeeds. In general, the main important observation
when comparing Polytool and AProVE is that although Polytool only uses polynomial interpretations and AProVE uses a large collection of different termination
techniques, Polytool is already almost as powerful as AProVE.
Similar to TerminWeb, cTI-1.1, and TALP, AProVE uses mode analysis and does
not provide the expressivity of types. However, it can express classes like bounded
lists, since it uses argument filterings. Nevertheless, in some cases, the effect of
argument filterings is not “deep” enough to represent redundant argument positions
adequately, cf. (Nguyen 2009). Finally, as shown in Table 1, AProVE is the slowest
tool in the experiments. One reason is that the transformation may generate quite
complex TRSs that require more time for termination analysis. Another reason is
that AProVE contains much more different termination techniques than the other
tools and it tries to apply them all after each other.
6 Conclusions
Since a few years, the LP and the TRS termination analysis communities jointly
organize the “International Workshop on Termination” (WST). As a part of this
workshop, the International Competition of Termination Tools is organized annually, allowing different termination tools from different categories, including term
rewriting and logic programming, to compete. These workshops have raised a considerable interest in gaining a better understanding of each other’s approaches. It
soon became clear that there has to be a close relationship between one of the
most popular techniques for TRSs, polynomial interpretations, and one of the key
techniques for LPs, acceptability with linear norms and level mappings. However,
partly because of the distinction between orders over the numbers (LPs) versus
orders over polynomials (TRSs), the actual relation between the approaches was
unclear.
One main conclusion of the research that led to this paper is that the distinction
is a superficial one. So one outcome of our work is that, indeed, the polynomial
interpretations used for TRSs are a direct generalization of the current practice for
LPs.
On the more technical level, the contribution of this paper is twofold. Firstly, we
provide a complete and revised theoretical framework for polynomial interpretations
in LP termination analysis (cf. Section 3). A first variant of such a framework was
Polynomial Interpretations for Termination Analysis of Logic Programs
29
introduced in a preliminary version of this paper (Nguyen and De Schreye 2005).
Parts of this build on the results in (De Schreye and Serebrenik 2002) on orderacceptability and the results in (Decorte et al. 1999) on the constraint-based approach for termination analysis. Another part extends the results of Bossi et al.
(Bossi et al. 1991) on the syntactic characterization of rigidity. The main revisions
are in the concept of polynomial interpretations and the concept of rigidity. Secondly, we adapt the constraint-based approach in (Decorte et al. 1999) to represent
all termination conditions symbolically, and introduce a new approach to find such
polynomial interpretations automatically (cf. Section 4).
We also developed an automated tool (Polytool (Nguyen and De Schreye 2007))
for termination proofs of LPs based on polynomial interpretations. The main contribution of the implementation is the integration of a number of techniques including the termination framework in Section 3, the call pattern inference tools in
(Bruynooghe et al. 2005; Gallagher et al. 2005; Heaton et al. 2000; Janssens and Bruynooghe 1992),
the constraint-based approach in Section 4, and the Diophantine constraint solver in
(Fuhs et al. 2007), to provide a completely automated termination analyzer. Polytool participated in the annual International Competitions of Termination Tools
since 2007 and reached the second place, just after AProVE.
We have also conducted extensive experimental evaluation for Polytool and compared it empirically with other termination analyzers such as cTI-1.1, TerminWeb,
TALP, and AProVE, cf. Section 5. The evaluation shows that Polytool is powerful
enough to solve a large number of benchmarks. In particular, it can also verify
termination of examples for which non-linear norms are required.
The current paper and the corresponding tool provide a good basis to adapt
further techniques from the area of TRS termination to the LP domain. In this
way, the power of automated termination analysis can be increased substantially.
Moreover, such adaptations will clarify the connections between the numerous termination techniques developed for TRSs and for LPs, respectively. First steps into
this direction are (Nguyen et al. 2008; Schneider-Kamp et al. 2009).
7 Acknowledgements
Manh Thang Nguyen was partly supported by GOA Inductive Knowledge Bases
and partly by FWO Termination Analysis: Crossing Paradigm Borders. Jürgen
Giesl and Peter Schneider-Kamp were supported by the Deutsche Forschungsgemeinschaft (DFG) grant GI 274/5-2. We thank John Gallagher for making his type
inference engine available, Carsten Fuhs for his SAT-based Diophantine constraint
solver within AProVE, Frédéric Mesnard and Roberto Bagnara for providing us the
cTI system and the Parma Polyhedra Library, Michael Codish and Samir Genaim
for their TerminWeb system. We thank the anonymous reviewers for their valuable
comments.
30
Manh Thang Nguyen et al.
References
Apt, K. R. 1990. Logic programming. In Handbook of Theoretical Computer Science.
Vol. B. MIT Press, 493–574.
Arts, T. and Giesl, J. 2000. Termination of term rewriting using dependency pairs.
Theoretical Computer Science 236, 1-2, 133–178.
Borralleras, C., Lucas, S., Navarro-Marset, R., Rodrı́guez-Carbonell, E., and
Rubio, A. 2009. Solving non-linear polynomial arithmetic via SAT modulo linear arithmetic. In Proc. CADE ’09. LNAI 5663. 294–305.
Bossi, A., Cocco, N., and Fabris, M. 1991. Proving termination of logic programs by
exploiting term properties. In Proc. TAPSOFT ’91. LNCS 494. 153–180.
Brauburger, J. and Giesl, J. 1998. Termination analysis by inductive evaluation. In
Proc. CADE ’98. LNAI 1421. 254–269.
Bruynooghe, M., Gallagher, J. P., and Van Humbeeck, W. 2005. Inference of welltypings for logic programs with application to termination analysis. In Proc. SAS ’05.
LNCS 3672. 35–51.
Bruynooghe, M., Codish, M., Gallagher, J. P., Genaim, S., and Vanhoof, W.
2007. Termination analysis of logic programs through combination of type-based norms.
ACM Transactions on Programming Languages and Systems 29, 2.
Codish, M. and Taboch, C. 1999. A semantic basis for the termination analysis of logic
programs. Journal of Logic Programming 41, 1, 103–123.
Contejean, E., Marché, C., Tomás, A. P., and Urbain, X. 2005. Mechanically
proving termination using polynomial interpretations. Journal of Automated Reasoning 34, 4, 325–363.
De Schreye, D. and Serebrenik, A. 2002. Acceptability with general orderings. In
Computational Logic: Logic Programming and Beyond. LNCS 2407. 187–210.
Decorte, S., De Schreye, D., and Fabris, M. 1993. Automatic inference of norms: a
missing link in automatic termination analysis. In Proc. ILPS ’93. MIT Press, 420–436.
Decorte, S., De Schreye, D., and Vandecasteele, H. 1999. Constraint-based automatic termination analysis of logic programs. ACM Transactions on Programming
Languages and Systems 21, 6, 1137–1195.
Dershowitz, N. 1995. 33 examples of termination. In Proc. French Spring School in
Theoretical Computer Science. LNCS 909. 16–26.
Dershowitz, N., Lindenstrauss, N., and Sagiv, Y. 1997.
What norms are
useful for logic programs?
In Proc. WST ’97. Transparencies available from
http://www.cs.huji.ac.il/~ naomil/.
Fuhs, C., Giesl, J., Middeldorp, A., Schneider-Kamp, P., Thiemann, R., and
Zankl, H. 2007. SAT solving for termination analysis with polynomial interpretations.
In Proc. SAT ’07. LNCS 4501. 340–354.
Fuhs, C., Giesl, J., Middeldorp, A., Schneider-Kamp, P., Thiemann, R., and
Zankl, H. 2008. Maximal termination. In Proc. RTA ’08. 110–125.
Gallagher, J. P., Henriksen, K. S., and Banda, G. 2005. Techniques for scaling up
analyses based on pre-interpretations. In Proc. ICLP ’05. LNCS 3668. 280–296.
Giesl, J., Schneider-Kamp, P., and Thiemann, R. 2006. AProVE 1.2: Automatic
termination proofs in the dependency pair framework. In Proc. IJCAR ’06. LNAI 4130.
281–286.
Giesl, J., Thiemann, R., Schneider-Kamp, P., and Falke, S. 2006. Mechanizing and
improving dependency pairs. Journal of Automated Reasoning 37, 3, 155–203.
Giesl, J., Thiemann, R., Swiderski, S., and Schneider-Kamp, P. 2007. Proving
termination by bounded increase. In Proc. CADE ’07. LNAI 4603. 443–459.
Polynomial Interpretations for Termination Analysis of Logic Programs
31
Heaton, A., Abo-Zaed, M., Codish, M., and King, A. 2000. A simple polynomial
groundness analysis for logic programs. Journal of Logic Programming 45, 1-3, 143–
156.
Hirokawa, N. and Middeldorp, A. 2005. Automating the dependency pair method.
Information and Computation 199, 1-2, 172–199.
Hong, H. and Jakuš, D. 1998. Testing positiveness of polynomials. Journal of Automated
Reasoning 21, 1, 23–38.
Janssens, G. and Bruynooghe, M. 1992. Deriving descriptions of possible values of program variables by means of abstract interpretation. Journal of Logic Programming 13, 23, 205–258.
Lankford, D. S. 1979. On proving term rewriting systems are noetherian. Tech. rep.,
Mathematics Dept., Louisiana Tech. University, Ruston, LA.
Lindenstrauss, N. 2000.
TermiLog: Termination analyzer for logic programs.
http://www.cs.huji.ac.il/~ naomil/termilog.php.
Lindenstrauss, N. and Sagiv, Y. 1997. Automatic termination analysis of logic programs. In Proc. ICLP ’97. MIT Press, 63–77.
Lindenstrauss, N., Sagiv, Y., and Serebrenik, A. 2004. Proving termination for
logic programs by the query-mapping pairs approach. In Program Development in
Computational Logic. LNCS 3049. 453–498.
Lloyd, J. W. 1987. Foundations of Logic Programming. Springer Verlag, Berlin.
Mesnard, F. and Bagnara, R. 2005. cTI: A constraint-based termination inference tool
for ISO-Prolog. Theory and Practice of Logic Programming 5, 1-2, 243–257.
Nguyen, M. T. and De Schreye, D. 2005. Polynomial interpretations as a basis for
termination analysis of logic programs. In Proc. ICLP ’05. LNCS 3668. 311–325. Extended version appeared as Technical report, Dept. of Computer Science, K. U. Leuven,
Belgium, http://www.cs.kuleuven.be/publicaties/rapporten/cw/CW412.pdf.
Nguyen, M. T. and De Schreye, D. 2007.
Polytool: Proving termination automatically based on polynomial interpretations.
In Proc.
LOPSTR ’06. LNCS 4407. 210–218.
Extended version appeared as
Technical report, Dept. of Computer Science, K. U. Leuven, Belgium,
http://www.cs.kuleuven.be/~ manh/polytool/publications/tech-reports/reportCW442.pdf.
Nguyen, M. T., Giesl, J., Schneider-Kamp, P., and De Schreye, D. 2008. Termination analysis of logic programs based on dependency graphs. In Proc. LOPSTR ’07.
LNCS 4915. 8–22.
Nguyen, M. T. 2009. Termination analysis: Crossing paradigm borders. Ph.D. thesis,
Dept. of Computer Science, K. U. Leuven, Belgium.
Ohlebusch, E., Claves, C., and Marché, C. 2000. TALP: A tool for the termination
analysis of logic programs. In Proc. RTA ’00. LNCS 1833. 270–273.
Schneider-Kamp, P., Giesl, J., Serebrenik, A., and Thiemann, R. 2009. Automated
termination proofs for logic programs by term rewriting. ACM Transactions on Computational Logic 11, 1. Preliminary version appeared in Proc. LOPSTR ’06, LNCS 4407,
pages 117-193, 2007.
Schneider-Kamp, P., Giesl, J., and Nguyen, M. T. 2009. The dependency triple
framework for termination of logic programs. In Proc. LOPSTR ’09. LNCS. To appear.
Serebrenik,
A.
2003.
Termination
analysis
of
logic
programs.
Ph.D. thesis, Dept. of Computer Science, K. U. Leuven, Belgium.
http://www.cs.kuleuven.ac.be/publicaties/doctoraten/cw/CW2003_02.abs.html.
Serebrenik, A. and De Schreye, D. 2003. Hasta-La-Vista: Termination analyser for
logic programs. In Proc. WLPE ’03. 60–74.
32
Manh Thang Nguyen et al.
Steinbach, J. 1992. Proving polynomials positive. In Proc. FSTTCS ’92. LNCS 652.
18–20.
Taboch, C., Genaim, S., and Codish, M. 2002. TerminWeb: Semantics-based termination analyser for logic programs. http://www.cs.bgu.ac.il/~ mcodish/TerminWeb.
Verschaetse, K. and De Schreye, D. 1991. Deriving termination proofs for logic
programs, using abstract procedures. In Proc. ICLP ’91. MIT Press, 301–315.
| 6cs.PL |
1
Patch-Based Low-Rank Minimization for Image
Denoising
arXiv:1506.08353v2 [cs.CV] 21 Feb 2018
Haijuan Hu, Jacques Froment, Quansheng Liu
Abstract—Patch-based sparse representation and low-rank
approximation for image processing attract much attention in
recent years. The minimization of the matrix rank coupled with
the Frobenius norm data fidelity can be solved by the hard
thresholding filter with principle component analysis (PCA) or
singular value decomposition (SVD). Based on this idea, we
propose a patch-based low-rank minimization method for image
denoising, which learns compact dictionaries from similar patches
with PCA or SVD, and applies simple hard thresholding filters to
shrink the representation coefficients. Compared to recent patchbased sparse representation methods, experiments demonstrate
that the proposed method is not only rather rapid, but also
effective for a variety of natural images, especially for texture
parts in images.
Index Terms—Image denoising, patch-based method, lowrank minimization, principal component analysis, singular value
decomposition, hard thresholding
I. I NTRODUCTION
I
MAGE denoising is a classical image processing problem,
but it still remains very active nowadays with the massive
and easy production of digital images. We mention below
some important works among the vast literature which deals
with image denoising.
One category of denoising methods concerns transformbased methods, for example [1], [2]. The main idea is to
calculate wavelet coefficients of images, shrink the coefficients
and finally reconstruct images by inverse transform. These
methods apply fixed transform dictionaries to whole images.
However, fixed dictionaries do not generally represent whole
images very well due to the complexity of natural images.
Many image details are lost while being denoised.
Another category is related to patch-based methods first
proposed in [3], which explores the non-local self-similarity of
natural images. Inspired by this “patch-based” idea, the authors
of K-SVD [4] and BM3D [5] proposed using dictionaries to
represent small local patches instead of whole images so that
sparsity of coefficients can be increased, where the dictionaries
are fixed or adaptive, and compact or overcomplete. These
methods greatly improve the traditional methods [1], [2],
leading to very good performance. Since these works many
similar methods have been proposed to improve the denoising
process, such as LPG-PCA [6], ASVD [7], PLOW [8], SAIST
Haijuan Hu is with Northeastern University at Qinhuangdao, School of
Mathematics and Statistics, Hebei, 066004, China. She was with Univ
Bretagne-Sud, CNRS UMR 6205, LMBA, Campus de Tohannic, F-56000
Vannes, France (e-mail: huhaijuan61@126.com).
Jacques Froment and Quansheng Liu are with Univ Bretagne-Sud, CNRS
UMR 6205, LMBA, Campus de Tohannic, F-56000 Vannes, France (e-mail:
Jacques.Froment@univ-ubs.fr; Quansheng.Liu@univ-ubs.fr).
[9], NCSR [10], GLIDE [11], and WNNM [12]. However,
many proposed methods are computationally complex. For
example, K-SVD uses overcomplete dictionaries for sparse
representation, which is time-consuming. BM3D and LPGPCA iterate the denoising process twice; SAIST and WNNM
iterate about 10 times. The computational cost is directly
proportional to the number of iterations.
At the same time, the low-rank matrix approximation has
been widely studied and applied to image processing [13],
[14], [15]. Many low-rank models have no explicit solution.
However, the paper [13] proves that the nuclear norm minimization with the Frobenius norm data fidelity can be solved
by a soft thresholding filter. (See also the paper [12] where
an alternative proof is given.) Furthermore, with the help
of Eckart-Young theorem [16], the paper [17] demonstrates
that the solution of the exact low-rank matrix minimization
problem (l0 norm) can be obtained by a hard thresholding
filter.
Inspired by the above theories, in this paper, a patchbased low-rank minimization (PLR) method is proposed for
image denoising. First, similar patches are stacked together
to construct similarity matrices. Then each similarity matrix
is denoised by minimizing the matrix rank coupled with the
Frobenius norm data fidelity. The minimizer can be obtained
by a hard thresholding filter with principle component analysis
(PCA) or singular value decomposion (SVD). The proposed
method is rather rapid, since we use compact dictionaries
which are more computationally efficient than over-completed
dictionaries, and we do not iterate. Moreover, experiments
show that the proposed method is as good as the state-ofthe-art methods, such as K-SVD [4], BM3D [5], LPG-PCA
[6], ASVD [7], PLOW [8], SAIST [9], and WNNM [12].
The rest of the paper is organized as follows. In Section II,
we introduce our method. The experimental results are shown
in Section III. Finally, this paper is concluded in Section IV.
II. PATCH -BASED L OW-R ANK M INIMIZATION
The noise model is:
v = u + η,
where u is the original image, v is the noisy one, and η is
the Gaussian noise with mean 0 and standard deviation σ. The
images u, v, η are with size M ×N . Without loss of generality,
we suppose that M = N .
2
Note that
A. Proposed Algorithm
Divide the noisy image v into overlapped patches of size
d × d. Denote the set of all these patches as S = {xi : i =
1, 2, · · · , (N − d + 1)2 }.
For each patch x ∈ S, called reference patch, consider all
the overlapped patches contained in its n × n neighborhood1
(the total number of such patches is (n − d + 1)2 patches).
Then choose the m (m ≥ d2 ) most similar patches (including
the reference patch itself) to the reference patch among the
(n − d + 1)2 patches. The similarity is determined by the l2 norm distance.
Next, for each reference patch, its similar patches are
reshaped as vectors, and stacked together to form a matrix
of size d2 × m, called similarity matrix. The similarity matrix
is denoted as S = (s1 , s2 , · · · , sm ), where columns of S, i.e.
si , i = 1, 2, · · · , m, are vectored similar patches. Then all the
patches in the matrix S are denoised together using the hard
thresholding method with the principal component (PC) basis,
or equivalently, with the singular value decomposition (SVD)
basis derived from the matrix S; the detailed process will be
given afterward. For convenience, we assume
Pmthat the mean
1
of the patches in S, denoted by sc := m
l=1 sl , is 0. In
practice, we subtract sc from si to form the matrix S, and
add sc to the final estimation s̄l of each patch.
Since the patches are overlapped, every pixel is finally
estimated as the average of repeated estimates.
The process of denoising the matrix S is shown as follows.
Firstly, we derive adaptive basis using PCA. The PC basis
is the set of the eigenvectors of SS T . Write the eigenvalue
decomposition2
SS T = P ΛP −1
(1)
with
P = (g 1 , g 2 , · · · , g d2 ), Λ = diag(mλ21 , mλ22 , · · · , mλ2d2 ),
where g i denotes the i-th column of P and diag(c1 , c2 , · · · )
denotes the diagonal matrix with (c1 , c2 , · · · ) on the diagonal.
The PC basis is the set of the columns of P , that is,
{g 1 , g 2 , · · · , g d2 }.
The original patches si in the similarity matrix S are
estimated as follows:
2
s̄l =
d
X
k=1
ak hsl , g k ig k ,
where
ak =
1
0
l = 1, 2, · · · , m
if λ2k > t2 ,
otherwise,
(2)
(3)
t being the threshold. Or equivalently, the matrix composed
of estimated patches (2) can be written as
S̄ := (s̄1 , s̄2 , · · · , s̄m ) = P h(Λ)P
−1
S,
(4)
with
h(Λ) = diag(a1 , a2 , · · · , ad2 ).
(5)
1 The reference patch is located at the center of the neighborhood, if the
parities of d and n are the same; otherwise, the reference patch is located as
near as possible to the center of the neighborhood.
2 We assume that the matrix SS T has full rank, and it has no identical
eigenvalues, which are generally true in practice.
m
1 X
(hsl , g k i)2 = λ2k
m
(6)
l=1
after a simple calculation. Thus λk can be interpreted as the
standard deviation of the basis coefficients.
We could also consider the singular value decomposition
(SVD) of S:
S = P ΣQT ,
(7)
where P is chosen as the same orthogonal matrix in (1), Σ
is a diagonal matrix, and Q (of size m × d2 ) has orthogonal
columns such that QT Q = I with I the identity matrix. Then
the denoised matrix (4) is equal to
Ŝ := P Ht√m (Σ)QT ,
(8)
where Ht√m (Σ) is a diagonal matrix, with the diagonal of
Ht√m (Σ) obtained by the hard thresholding operator
Ht√m (Σ)kk
=
Σkk
0
√
if Σkk > t m,
otherwise,
k = 1, 2, · · · , d2 .
(9)
In fact, the equality of (4) and (8) can be demonstrated as
follows. By the equations (1) and (7), we have Λ = Σ2 , and
P −1 S = Σ QT . Furthermore, by the equations (5) and (9),
we get h(Λ)Σ = Ht√m (Σ). Thus it follows that S̄ = Ŝ.
B. Low-Rank Minimization
Theorem 2.1 stated below is an unconstrained version of the
Eckart-Young theorem [16], and comes from Theorem 2(ii) in
[17]. According to Theorem 2.1, it easily follows that
Ŝ = arg min kS − Xk2F + mt2 Rank(X),
(10)
X
where the minimum is taken over all the matrices X having
the same size as S, and k · kF is the Frobenius norm. Hence
the denoised matrix Ŝ is the solution of the exact low-rank
minimization problem.
Theorem 2.1: The following low-rank minimization problem
X̂ = arg min kY − Xk2F + µ Rank(X)
(11)
X
has the solution3
X̂ = U H√µ (Σ)V T ,
(12)
where Y = U ΣV T is the SVD of Y , and H√µ is the hard
thresholding operator
√
Σkk if Σkk > µ,
√
H µ (Σ)kk =
0 otherwise.
3 Strictly speaking, if none of the singular values of Y equals with
solution is unique, which is generally true in practice.
√
µ, the
3
C. Choice of the Threshold
The choice of the threshold t in (3) is crucial for the
proposed algorithm. We study it by minimizing the mean
squared error of estimated values of vectored patches sl , (l =
1, 2, · · · , m) in a similarity matrix S. Denote
sl = ul + η l ,
where ul and η l are the vectored patches of the true image
and the noise corresponding to sl respectively.
By (2) or (4), it can be easily obtained that
2
2
ks̄l − ul k =
d
X
k=1
2
2
2
(ak − 1) (hg k , ul i) +
d
X
k=1
a2k (hg k , η l i)2 . (13)
Assume that the PC basis {g 1 , g 2 , · · · , g d2 } only depends
on the true value vectors {u1 , u2 , · · · , um } and hence is
independent of {η 1 , η 2 , · · · , η m }. Then
E(hg k , η l i)2 = σ 2 .
Let
(14)
m
θk2 =
1 X
(hg k , ul i)2 .
m
(15)
l=1
TABLE I
PSNR VALUES FOR REMOVING NOISE FOR OUR PLR AND OTHER
METHODS . C AM IS THE C AMERAMEN IMAGE
Image
Lena Barbara Peppers
σ = 10
K-SVD[4] 35.50 34.82
34.23
LPGPCA[6] 35.72 35.46
34.05
ASVD[7]
35.58 35.58
33.55
PLOW[8]
35.29 34.52
33.56
PLR
35.90 35.50
34.28
BM3D[5]
35.90 35.39
34.68
SAIST[9]
35.87 35.69
34.76
WNNM[12] 36.02 35.92
34.94
σ = 20
K-SVD[4] 32.38 31.12
30.78
LPGPCA[6] 32.61 31.69
30.50
ASVD[7]
33.21 32.96
30.56
PLOW[8]
32.70 31.48
30.52
PLR
33.03 32.12
30.90
BM3D[5]
33.03 32.07
31.28
SAIST[9]
33.07 32.43
31.28
WNNM[12] 33.10 32.49
31.53
Boats Bridge House
Cam
33.62
33.61
33.26
32.94
33.76
33.88
33.87
34.05
30.91 35.96 33.74
30.86 36.16 33.69
27.76 36.46 31.62
29.88 36.22 33.15
30.78 36.57 33.73
31.06 36.71 34.18
31.03 36.52 34.28
31.16 36.94 34.44
30.37
30.26
31.79
30.36
30.64
30.85
30.78
30.98
27.03
26.84
25.51
26.56
27.20
27.14
27.20
27.29
33.07
33.10
33.53
33.56
33.36
33.77
33.80
34.01
30.01
29.77
29.33
29.59
30.12
30.48
30.40
30.75
TABLE II
RUNNING TIME IN SECOND FOR OUR PLR AND OTHER METHODS TO
REMOVE NOISE WITH IMAGES OF SIZE 256 × 256
K-SVD LPG-PCA ASVD PLOW PLR SAIST WNNM
210
138
337
43
2
25
134
Then by (6), we obtain
E(λ2k ) = θk2 + σ 2 .
(16)
Thus, from (13), (14), and (15), it follows that
m
d2
d2
l=1
k=1
k=1
X
X
1 X
ks̄l − ul k2 ≈
(ak − 1)2 θk2 + σ 2
a2k .
m
(17)
After a simple calculation, the optimal value for ak is
1 if θk2 > σ 2 ,
âk =
0 otherwise.
Since λ2k ≈ θk2 +σ 2 by (16), the optimal value of the threshold
in (3) is t2 ≈ 2σ 2 . In practice, we find that t = 1.5σ is a good
choice.
III. E XPERIMENTAL R ESULTS
In this section, we compare the performance of our PLR
method with those of state-of-the-art methods, including the
highly competitive method WNNM [12] proposed very recently. Standard gray images are utilized to test the performance of methods. For the simulation, the level of noise is
supposed to be known, otherwise there are methods to estimate
it; see e.g. [18]. For each image and each level of noise, all
the methods are applied to the same noisy images.
For our algorithm, the patch size is set to d = 7, the
size of neighborhoods for selecting similar patches is set to
n = 35, and the number of similar patches in a similarity
matrix is chosen as m = 5d2 . Image boundaries are handled
by assuming symmetric boundary conditions. For the sake of
computational efficiency, the moving step from one reference
patch to its neighbors both horizontally and vertically is
chosen as the size of patches, that is, 7. For other comparison
algorithms, we utilize the original codes released by theirs
authors.
In Table I, we compare the PSNR (Peak Signal-to-Noise
Ratio) values of our PLR method with other methods. The
PSNR value is defined by
PSNR (v̄) = 20 log10
255N
dB,
kv̄ − ukF
where u is the original image, and v̄ the restored one. As can
be seen in Table I, our method is generally better than K-SVD
[4], LPG-PCA [6] and PLOW [8], and sometimes even better
than BM3D [5]. Furthermore, for the visual comparisons, our
method is also good. For example, as can be seen in Fig.1,
our method preserves the texture parts in Lena and Barbara
the best among all the methods.
To have a clear comparison of complexities of different
methods, we compare the average CPU time to remove noise
with σ = 20 for the testing images of size 256 × 256: Peppers,
House and Cameraman. All the codes are written in M-files
and run in the platform of MATLAB R2011a on a 3.40GHz
Intel Core i7 CPU processor. We do not include BM3D for
comparison since the original code of BM3D contains MEXfiles. The running time is displayed in second in Table II. The
comparisons clearly show that the proposed method is much
faster than the others.
IV. C ONCLUSION
In this paper, a patch-based low-rank minimization method
for image denoising is proposed, which stacks similar patches
into similarity matrices, and denoises each similarity matrix
by seeking the minimizer of the matrix rank coupled with the
Frobenius norm data fidelity. The minimizer can be obtained
by a hard threshoding filter with principle component basis or
4
Original
Noisy
BM3D[5]
WNNM [12]
PLR
Fig. 1. Compare denoised images Lena and Barbara by our method and other methods for σ = 20. From left to right, the images are original images, noisy
images, images denoised by BM3D, WNNM, and our PLR method. To make the differences clearer, the second row and the bottom row display parts of
Lena images and Barbara images extracted from the first row and third row respectively.
left singular vectors. The proposed method is not only rapid,
but also effective compared to recently reported methods.
R EFERENCES
[1] D. Donoho and J. Johnstone, “Ideal spatial adaptation by wavelet
shrinkage,” Biometrika, vol. 81, no. 3, pp. 425–455, 1994.
[2] R. Coifman and D. Donoho, “Translation-invariant denoising,” Wavelets
and statistics, vol. 103, pp. 125–150, 1995.
[3] A. Buades, B. Coll, and J. Morel, “A review of image denoising
algorithms, with a new one,” Multiscale Model. Simul., vol. 4, no. 2,
pp. 490–530, 2005.
[4] M. Elad and M. Aharon, “Image denoising via sparse and redundant
representations over learned dictionaries,” IEEE Trans. Image Process.,
vol. 15, no. 12, pp. 3736–3745, 2006.
[5] K. Dabov, A. Foi, V. Katkovnik, and K. Egiazarian, “Image denoising
by sparse 3-D transform-domain collaborative filtering,” IEEE Trans.
Image Process., vol. 16, no. 8, pp. 2080–2095, 2007.
[6] L. Zhang, W. Dong, D. Zhang, and G. Shi, “Two-stage image denoising
by principal component analysis with local pixel grouping,” Pattern
Recognition, vol. 43, no. 4, pp. 1531–1549, 2010.
[7] Y. He, T. Gan, W. Chen, and H. Wang, “Adaptive denoising by singular
value decomposition,” IEEE Signal Process. Lett., vol. 18, no. 4, pp.
215–218, 2011.
[8] P. Chatterjee and P. Milanfar, “Patch-based near-optimal image denoising,” IEEE Trans. Image Process., vol. 21, no. 4, pp. 1635–1649, 2012.
[9] W. Dong, G. Shi, and X. Li, “Nonlocal image restoration with bilateral
variance estimation: a low-rank approach,” IEEE Trans. Image Process.,
vol. 22, no. 2, pp. 700–711, 2013.
[10] W. Dong, L. Zhang, G. Shi, and X. Li, “Nonlocally centralized sparse
representation for image restoration.” IEEE Trans. Image Process.,
vol. 22, no. 4, pp. 1620–1630, 2013.
[11] H. Talebi and P. Milanfar, “Global image denoising,” IEEE Trans. Image
Process., vol. 23, no. 2, pp. 755–768, 2014.
5
[12] S. Gu, L. Zhang, W. Zuo, and X. Feng, “Weighted nuclear norm
minimization with application to image denoising,” in Proc. IEEE Conf.
Comput. Vis. Pattern Recognit., 2014.
[13] J.-F. Cai, E. J. Candès, and Z. Shen, “A singular value thresholding
algorithm for matrix completion,” SIAM J. Optimiz., vol. 20, no. 4, pp.
1956–1982, 2010.
[14] H. Schaeffer and S. Osher, “A low patch-rank interpretation of texture,”
SIAM J. Imaging Sci., vol. 6, no. 1, pp. 226–262, 2013.
[15] Y. Peng, J. Suo, Q. Dai, and W. Xu, “Reweighted low-rank matrix
recovery and its application in image restoration,” IEEE Trans. Cybern.,
vol. 14, no. 12, pp. 2418–2430, 2014.
[16] C. Eckart and G. Young, “The approximation of one matrix by another
of lower rank,” Psychometrika, vol. 1, no. 3, pp. 211–218, 1936.
[17] J.-B. Hiriart-Urruty and H. Y. Le, “From eckart and young approximation to moreau envelopes and vice versa,” RAIRO-Operations Research,
vol. 47, no. 03, pp. 299–310, 2013.
[18] I. M. Johnstone and B. W. Silverman, “Wavelet threshold estimators
for data with correlated noise,” J. Roy. Stat. Soc. B, vol. 59, no. 2, pp.
319–351, 1997.
| 1cs.CV |
Mean Field Multi-Agent Reinforcement Learning
Yaodong Yang * 1 Rui Luo * 1 Minne Li 1 Ming Zhou 2 Weinan Zhang 2 Jun Wang 1
arXiv:1802.05438v1 [cs.MA] 15 Feb 2018
Abstract
Existing multi-agent reinforcement learning methods are limited typically to a small number of
agents. When the agent number increases largely,
the learning becomes intractable due to the curse
of the dimensionality and the exponential growth
of user interactions. In this paper, we present
Mean Field Reinforcement Learning where the
interactions within the population of agents are
approximated by those between a single agent
and the average effect from the overall population or neighboring agents; the interplay between
the two entities is mutually reinforced: the learning of the individual agent’s optimal policy depends on the dynamics of the population, while
the dynamics of the population change according
to the collective patterns of the individual policies.
We develop practical mean field Q-learning and
mean field Actor-Critic algorithms and analyze
the convergence of the solution. Experiments on
resource allocation, Ising model estimation, and
battle game tasks verify the learning effectiveness
of our mean field approaches in handling manyagent interactions in population.
1. Introduction
Multi-agent reinforcement learning (MARL) is concerned
with a set of autonomous agents that share a common environment (Busoniu & De Schutter). Learning in MARL is
fundamentally difficult since agents not only interact with
the environment but also with each other. Independent Qlearning (Tan, 1993) that considers other agents as a part
of the environment quite often fails as the multi-agent environment breaks the theoretical convergence guarantees and
makes the learning unstable: the changes in strategy of one
agent would affect the strategies of other agents and vice
versa (Matignon et al., 2012).
Instead, accounting for the extra information from conjecturing the policies of other agents would be beneficial to
*
Equal contribution 1 University College London, London,
United Kingdom 2 Shanghai Jiao Tong University, Shanghai, China.
Correspondence to: Jun Wang <j.wang@cs.ucl.ac.uk>.
each single learner (Foerster et al., 2017; Lowe et al., 2017a).
Studies have shown that an agent who learns the effect of
joint actions has better performance than those who does
not, in a number of scenarios including cooperative games
(Panait & Luke, 2005), zero-sum stochastic games (Littman,
1994), and general-sum stochastic games (Littman, 2001;
Hu & Wellman, 2003).
The existing equilibrium based approaches, although principled, are only capable of solving a handful of agents (Hu
& Wellman, 2003; Bowling & Veloso, 2002). The computational complexity of directly solving (Nash) equilibrium
would prevent them from applying for the situations there
are a large body of agents or even a population of agents.
Yet, in practice, many cases do require strategic interactions
among a large number of agents, ranging from gaming bots
in MMORPG (Massively Multiplayer Online Role-Playing
Game) (Jeong et al., 2015) to trading agents in stock markets (Troy, 1997), and to online advertising bidding (Wang
et al., 2017).
In this paper, we tackle MARL when a large number of
agents co-exist. We consider a setting where each agent is
directly interacting with a finite set of other agents; through
a chain of direct interactions, any pair of agents is interconnected globally (Blume, 1993). The scalability is solved by
employing Mean Field Theory (Stanley, 1971): the interactions within the population of agents are approximated by
that of a single agent played with the average effect from the
overall (local) population (Lasry & Lions, 2007). The learning is mutually reinforced between two entities rather than
among many entities: the learning of the individual agent’s
optimal policy is based on the dynamics of the agent population, while at the same time, the dynamics of the population
is updated according to the individual policies. Based on
the formulation, we develop practical mean field Q-learning
and mean field Actor-Critic algorithms, and discuss the convergence of our solution under certain assumptions. Our experiment on a simple multi-agent resource allocation shows
that our mean field MARL is capable of learning over manyagent interactions when others fail. We also demonstrate
that with temporal-difference learning, mean field MARL
manages to learn and solve an Ising model without even
explicitly knowing the energy function. At last, in a mixed
cooperative-competitive battle game, we show that the mean
field MARL achieves high winning rates against other baselines previously reported for many agent systems.
Mean Field Multi-Agent Reinforcement Learning
2. Preliminary
2.2. Nash Q-learning
MARL intersects between reinforcement learning and game
theory. The marriage of the two gives rise to the general
framework of Stochastic Game (van der Wal, 1980).
In MARL, the goal is to learn an optimal strategy for each
agent. In stochastic games, the notion of Nash equilibrium
is thus of great importance (Hu & Wellman, 2003); it is
represented as a set of N policies π∗ = {π∗1, . . . , π∗N } such
that ∀π j ∈ Ω(A j ), it satisfies
2.1. Stochastic Game
An N-agent (or, player) stochastic game Γ is expressed
by a tuple hS, A1, . . . , A N , r 1, . . . , r N , p, γi, where S denotes the state space and A j is the action space of agent
j ( j = 1, . . . , N). At each time step, all agents choose
action simultaneously, which leads to their rewards. r j :
S × A1 × · · · × A N → R is the reward function for agent
j. γ denotes the reward discount factor across the time.
p : S × A1 × · · · × A N → Ω(S) is the transition probability map of the environment, with Ω(S) indicating the
collection of probability distributions over the state space.
Each stage of a stochastic game, depending on the state of
the current game is in, is called a stage game.
The policy (or, strategy) for agent j is π j : S → Ω(A j )
where Ω(A j ) is the collection of probability distribution
over agent j’s action space A j . π = {π 1, . . . , π N } denotes
the joint policy of all N agents. π is usually assumed independent of time, also called stationary strategy. Given initial
state s and certain policy π j , agent j aims to maximize the
value function, namely:
vπj (s) = v j (s; π) =
∞
X
t=0
γ t Eπ, p rtj |s0 = s, π .
(1)
We can define the Q-function in the framework of N-agent
game based on the value function. The Bellman equation
for Q-function of agent j under the joint policy π is Qπj :
S × A1 × · · · × A N → R such that
Qπj (s, a) = r j (s, a) + γEs0 ∼p [vπj (s0 )].
(2)
And the corresponding value function vπj for agent j under
the joint policy π can be expressed in terms of Q-function
vπj (s) = Ea∼π Qπj (s, a) .
(3)
Note that the Q-function for N-agent game extends the conventional formulation for single-agent game by including
actions taken by all N agents a = {a1, . . . , a N }, and by
taking the expectation over the actions of all agents.
We model the MARL problem by the stochastic game in a
discrete-time and non-cooperative setting (no explicit coalitions are considered). We assume the game has incomplete
but perfect information (Littman, 1994), i.e., agents do not
know the game dynamics or the reward functions of others,
but they can observe other agents’ immediate reward and
actions taken previously.
v j (s; π∗ ) = v j (s; π∗j , π∗−j ) ≥ v j (s; π j , π∗−j ),
(4)
where we adopt compact notations for the joint policy of all
players except j as π∗−j = {π∗1, . . . , π∗j−1, π∗j+1, . . . , π∗N }.
In a Nash equilibrium, each agent acts as the best response
π∗j to others, provided that all other agents follow the policy
π∗−j . It has been shown, for every N-agent stochastic game,
there is at least one Nash equilibrium in stationary strategies
(Fink et al., 1964). Given a set of Nash strategy π∗ , the
Nash value function v Nash (s) = (vπ1∗ (s), . . . , vπN∗ (s)) stands
as the expected sum of discounted rewards if all agents
follow these Nash policies from state s onward.
Nash Q-learning (Hu & Wellman, 2003) is an iterative procedure of calculating the Nash policy: 1) solving a Nash
equilibrium through the Lemke-Howson algorithm (Lemke
& Howson, 1964) for the current stage game {Q tj }, 2) improving the estimation of the Nash Q-function using the
new Nash policy. Nash Q keeps iterating the above two
steps until the Q-value converges. It is shown that under certain assumptions, the Nash Q operator HtNash Q =
(HtNash Q1, . . . , HtNash Q N ) defined as
(HtNash Q)(s, a) ≡ Es0 ∼p r(s, a) + γvtNash (s0 )
(5)
is a contraction mapping given that π is an appropriate Nash
equilibrium solver for stage game (Q1 (s0 ), . . . , Q N (s0 )) (s0
is the state at time t + 1); that is, the Q-function will eventually converge to a Nash equilibrium value of the entire
game.
3. Mean Field MARL
As the number of agents increases, the joint action space a
becomes high dimensional. All agents act strategically and
simultaneously learn their values over others actions, which
make the learning of the standard Q function Q j (s, a) in the
previous section infeasible. To handle this, we parametrize
the Q-function by only the pairwise local interactions:
Q j (s, a) ≡
1 X j
Q (s, a j , ak ),
Nj
j
(6)
k∈K
where K j is the set of neighbor agents for agent j. Its size
N j = #K j grows proportionally to the number of agents,
and the ratio depends on applications. It is worth noting that
the pairwise approximation, while significantly reducing the
complexity of the interactions among agents, still preserves
global interactions between any pair of agents implicitly
Mean Field Multi-Agent Reinforcement Learning
(Blume, 1993). Similar approaches in machine learning can
be found such as in factorization machine (Rendle, 2012)
and learning to rank (Cao et al., 2007).
As shown in Eq. (10), we see that under the mean field
approximation, the MARL problem is converted into the
computation of a single agent’s best response with respect
to the action distribution of other agents.
3.1. Mean Field Approximation
To discriminate from the Nash value function v Nash (s), we
refer the mean field value function defined in Eq. (10) as
We consider discrete action space, whereas a continuous
version can be analogically derived. Specially, action a j
is a discrete categorical variable, which is represented as a
one-hot encoding with each component indicating one of
j
the M possible actions: a j ≡ (a1j , . . . , a M
). The pairwise
j
j k
interaction Q (s, a , a ) can be approximated by the mean
field theory (Stanley, 1971). Conveniently the one-hot action of each agent can be represented as a small fluctuation
δak from the mean action within the neighborhood:
ak = ā + δak ,
where ā =
1 X k
a .
Nj
j
(7)
k∈K
Note that the mean action ā ≡ (ā1, . . . , ā M ) is equivalent
to the multinomial distribution of actions from neighboring
agents. With the first-order Taylor expansion, the pairwise
Q-function is approximated as
1 X j
Q (s, a j , ak )
Nj
k∈K j
1 X
≈ j
Q j (s, a j , ā) + ∇ā Q j (s, a j , ā)δak
N
k∈K j
1 X k
= Q j (s, a j , ā) + ∇ā Q j (s, a j , ā) j
δa
N
j
Q j (s, a) =
k∈K
= Q j (s, a j , ā).
(8)
Note that in the second line of Eq. (8) we utilize the firstorder approximation of Q j , where the accuracy improves
when the number of neighbours grows large, since the mean
value of the higher order terms approach zero. The multipleagent interactions have been converted into those of twoagent interactions: a single agent and the mean agent from
the neighbors.
The update rule for the mean field Q-function is defined as
j
Q t+1
(s, a j , ā)
h
i
= (1 − αt )Q tj (s, a j , ā) + αt rtj + γvtj (s0 )
(9)
where αt denotes the learning rate and the mean field value
function is as
X
etj (s, a j ),
vtj (s) =
π j (a j |s)Q
(10)
aj
in which
1 X k
j
j
j
j
e
Q t (s, a ) = Ea− j ∼π− j Q t (s, a , ā = j
a ) . (11)
N
k6= j
v MF (s) = (vπ1 (s), . . . , vπN (s)).
Based on the mean field value function, we finally define the
mean field Q operator as HtMF Q = (HtMF Q1, . . . , HtMF Q N ),
and
(HtMF Q)(s, a) ≡ Es0 ∼p r(s, a) + γvtMF (s0 ) .
(12)
In fact, we can prove that HMF is a contraction mapping; that
is, by updating the Q-function via the mean field operator
HMF , the mean field Q-learning will eventually converge to
the Nash equilibrium under certain assumptions.
3.2. Implementation
In this section we describe the implementation logic of our
proposed algorithm. We parameterize the Q-function and
policy π by φ and θ respectively, and update parameter φ according to Eq.(9). Once Q-function is updated, we improve
the current policy π. For on-policy reinforcement learning,
we use Boltzmann strategy and actor-critic method (Konda
& Tsitsiklis, 2000) to improve policy, presenting MF-Q
(Algorithm 1 in Appendix A) and MF-AC (Algorithm 2 in
Appendix A) respectively.
For agent i, MF-Q can be trained by minimizing the loss
function
L(φ) = (y − Q φj (s, a j , ā))2,
(13)
where y = r + γ Q φj (s0, (a j )0, ā0 ) is the target. Differentiating L(φ) w.r.t the weights produces the following gradient,
∇φ L(φ) = y − Q φj (s, a j , ā) ∇φ Q φj (s, a j , ā).
(14)
The Critic of MF-AC follows the same update rule as above.
The Actor can be trained by using the sampled policy gradient:
∇θ J(θ) ≈ ∇θ log πθj (s)Q φ (s, a j , ā)|a j =π j (s),
θ
(15)
Note that we require the policies of other agents to apply
an update in Eq. 14. For off-policy learning, we can use
DPG (Silver et al., 2014) for deterministic environment or
standard Q-learning (Watkins & Dayan, 1992) for a discrete
case. We iteratively execute parameter update for φ and π
until the Q-function converges.
3.3. Proof of Convergence
We prove the convergence of Q t = (Q1t , . . . , Q tN ) under the
MF-Q iterations to Nash equilibrium Q ∗ = (Q1∗, . . . , Q∗N )
Mean Field Multi-Agent Reinforcement Learning
for learning agents. The proof is conducted through showing
that under the main assumptions below, the mean field Q
operator HMF of Eq. (12) is in fact a contraction mapping,
and its fixed point is Q ∗ .
We start by introducing the main assumptions:
Assumption 1. During training, each action-value pair
will be visited infinitely often, and the reward is bounded by
some constant K.
Assumption 2. Agent’s policy is Greedy in the Limit with
Infinite Exploration (GLIE). In the case of Boltzmann softmax policy, the temperature decays to the zero so that the
policy becomes greedy on the Q-function in the limit.
We make further assumptions on the structure of the game:
Assumption 3. For every stage game (Q1t (s), ..., Q tN (s))
encountered during training, for all t, s and j ∈ {1, . . . , N},
only one of the two scenarios could happen: the Nash
equilibrium π = (π 1, . . . , π N ) is either (I) a global
optimal point,
i.e. Eπ [Q tj (s)] ≥ Eπ̂ [Q tj (s)], ∀π̂ ∈
Q k
j
Ω
k A , or, (II) a saddle point, i.e. Eπ j Eπ − j [Q t (s)] ≥
Eπ̂ j Eπ− j [Q tj (s)], ∀π̂ j ∈ Ω A j ; Eπ j Eπ− j [Q tj (s)] ≤
Q
k
Eπ j Eπ̂− j [Q tj (s)], ∀π̂ −j ∈ Ω
k6= j A .
Note that Assumption 3 asks for a strong requirement on
every individual stage game encountered during training. In
practice, however, we find this is not necessary to hold in
order for the learning algorithm to converge. This is in line
with the empirical findings in Hu & Wellman (2003).
Our proof depends on the following two Lemmas:
Lemma 1. Under the Assumption 3, the Nash Q operator
defined in Eq. (5) is a rigorous contraction mapping on the
complete metric space from Q to Q, its fixed point is the
Nash Q-value of the entire game, i.e., HtNash Q ∗ = Q ∗ .
In Lemma 2, by subtracting Q ∗ (s, a) at both sides in Eq. (9)
and comparing with Eq. (16), we can find in fact
∆t (x) = Q t (s, a) − Q ∗ (s, a),
Ft (x) = rt + γvtMF (st+1 ) − Q∗ (st , at ),
where x = {st , at }. α(t) can be regarded as the learning
rate, and agent will only update the Q-function with the corresponding state st and actions a visited, i.e., αt (s, a) = 0
for (s, a) 6= (st , at ). Note that Lemma 2 suggests the convergence of ∆t (x) to zero, that means, if it holds, the current
Q-function will asymptotically approximate the optimal Q ∗ .
One last piece before the main theorem is the below:
Proposition 1. P
Let the metric space be R N , and the metric
be M(a, b) = j |a j − b j |, ∀ a = {a j }1N , b = {b j }1N . If
the Q function is Lipschitz continuous w.r.t. the input of a j ,
then for sufficiently small temperature β, the operator
exp − βQ(s, a j , ā)
j
B(a ) = P
exp − βQ(s, ak , ā)
a k ∈A j
is a contraction mapping. The proof is in Appendix due to
the space limit.
Theorem 1. In finite-state stochastic game, the Q t values
computed by the MF-Q rule in Eq. (9) converges to the Nash
Q-value Q ∗ = (Q1∗, . . . , Q∗N ), if the Assumptions 1, & 2, 3,
and condition 1 & 2 of Lemma 2 are met.
Proof. We denote the σ-field generated by the random variables in the experience of the stochastic game
{st , αt , at , rt−1, ..., s1, α1, a1, Q 0 } by Ft . Given the fact that
vt , ..., v0 are Ft -measurable, thus both ∆t and Ft−1 are Ft measurable, which satisfies the measurability condition of
Lemma 2.
Proof. See Theorem 17 in Hu & Wellman (2003).
We further need to prove HMF operator satisfies conditions 3
& 4 to employ Lemma 2. We start from condition 3.
Lemma 2. The random process {∆t } in R defined as
Ft (st , at ) = rt + γvtMF (st+1 ) − Q ∗ (st , at )
∆t+1 (x) = (1 − αt (x))∆t (x) + αt (x)Ft (x),
(16)
which converges to zero with probability 1 (w.p.1) under the
following assumptions:
P
P
1. 0 ≤ αt (x) ≤ 1, t αt (x) = ∞, t αt2 (x) < ∞
2. x ∈ X , the set of possible states is finite.
3. kE[Ft (x)|Ft ]kW ≤ γk∆t kW + ct , where γ ∈ [0, 1), ct
converges to zero w.p.1.
2
4. var[Ft (x)|Ft ] ≤ K(1 + k∆t kW
), for constant K > 0.
Here Ft is the filtration of an increasing sequences of σfields that include the history of the process. We assume that
αt , ∆t , Ft ∈ Ft , and k · kW is a weighted maximum norm.
Proof. This Lemma is proved by Jaakkola et al. (1994)
(Theorem 1) and Szepesvári & Littman (1999) (Corollary
5). We include it here to stay self-contained.
= rt + γvtNash (st+1 ) − Q ∗ (st , at )
+ γ[vtMF (st+1 ) − vtNash (st+1 )]
= rt + γvtNash (st+1 ) − Q ∗ (st , at ) + Ct (st , at )
= FtNash (st , at ) + Ct (st , at )
(17)
In fact, we find FtNash (st , at ) is the corresponding Ft in
Lemma 2 as if the convergent proof was about the Nash Q
learning algorithm. Based on Lemma 1, it is easy to know
that FtNash (·, ·) is a contraction mapping where the norm k·k
is the maximum norm over a, thus we have, for all t,
kE[FtNash (·, ·)|Ft ]k∞ ≤ γkQ t − Q ∗ k∞ = γk∆t k∞ .
In meeting the condition 3 of Lemma 2 now we have
kE[Ft (st , at )|Ft ]k ≤ kFtNash (st , at )|Ft k + kCt (st , at )|Ft k
≤ γk∆t k + kCt (st , at )|Ft k.
(18)
Mean Field Multi-Agent Reinforcement Learning
We are left to prove ct = kCt (st , at )|Ft k converges to zero
w.p.1. In fact, under Assumption 3, for each stage game,
all the global optimal equilibrium point(s) share the same
equilibria value, so does any saddle-point equilibrium.
Case I: In each global equilibrium, according to the definition, all players get the same unique maximum value.
Case II: Assume the stage game {Q tj }1N has two saddle-point
equilibrium: π and ρ, then for agent j, we have
Eπ j Eπ− j [Q tj (s)] ≥ Eρ j Eπ− j [Q tj (s)],
Eρ j Eρ− j [Q tj (s)] ≤ Eρ j Eπ− j [Q tj (s)].
Combing them we have
Eπ j Eπ− j [Q tj (s)] ≥ Eρ j Eρ− j [Q tj (s)]
By the definition of saddle point, reversing the order of π
and ρ, the formula still holds; thus the equilibria to agent
i under both saddle-point equilibrium must be the same:
Eπ j Eπ− j [Q tj (s)] = Eρ j Eρ− j [Q tj (s)].
Given the Proposition 1 that the policy of mean field Qfunction is a contraction mapping, and that all global/saddle
points share the same Nash value in each stage game, v MF
value will therefore converge to v Nash value asymptotically.
Condition 3 is thus satisfied.
For condition 4, based on the condition 3 just proved, HMF
is a contraction mapping, i.e., HMF Q∗ = Q ∗ , thus we have:
var[Ft (st , at )|Ft ] = E[(rt + γvtMF (st+1 ) − Q ∗ (st , at ))2 ]
= E[(rt + γvtMF (st+1 ) − HMF (Q ∗ ))2 ]
= var[rt + γvtMF (st+1 )|Ft ]
2
≤ K(1 + k∆t kW
)
(19)
In the last step, we apply the Assumption 1 that the reward
is always bounded by some constant. Finally, by meeting
both condition 3 & 4, according to Lemma 2, ∆t converges
to zero w.p.1, i.e., Q t converges to Q ∗ w.p.1.
Apart from convergence, we also discuss the Rationality
(Bowling & Veloso, 2001; 2002) of the MF-Q algorithm,
due to the space limit, we leave it in the Appendix.
4. Related Work
We continue our discussion on related work started in Section 1 and make comparisons with existing techniques in a
greater scope. Our work follows the same direction as from
(Littman, 1994; Hu & Wellman, 2003; Bowling & Veloso,
2002) in adapting a Stochastic Game (van der Wal, 1980)
in the MARL formulation. Specifically, Littman (1994)
addressed two-player zero-sum stochastic games by introducing a “minimax” operator in Q-learning, whereas Hu
& Wellman (2003) extended it to the general-sum case by
learning a Nash equilibrium in each stage game and considering a mixed strategy. Nash-Q learning is guaranteed to
converge to Nash strategies under the (strong) assumption
that there exists equilibrium for every stage game, defined
by the Q-functions. In the situation where agents can be
identified as either "friends" or "foes", one can simply the
solution by alternating between fully cooperative and zerosum learning (Littman, 2001). Moreover, Littman & Stone
(2005) and de Cote & Littman (2008) draw on the folk
theorem and acquired a polynomial-time Nash equilibrium
algorithm for repeated stochastic games. Bowling & Veloso
(2002) varied the learning rate to improve the convergence.
The recent treatment of MARL was using deep neural networks as a function approximator. For instance, in addressing the non-stationary issue in MARL, various solutions
have been proposed including neural-based opponent modeling (He & Boyd-Graber, 2016), policy parameters sharing
(Gupta et al., 2017) etc. Researchers have also adopted
the paradigm of centralized training with decentralized execution for multi-agent policy-gradient learning: BICNET
(Peng et al., 2017), COMA (Foerster et al., 2016) and MADDPG (Lowe et al., 2017a), which allow the centralized critic
Q-function augmented with actions of other agents, and
the actor needs only local observation history to optimize
agent’s policy.
The above MARL approaches limit their studies mostly to
a few agents. As the number of agents grows, not only the
input space of Q grows linearly, but most critically the accumulated noises by the exploratory actions of other agents
make the Q-function learning no longer feasible. Our work
addresses the issue by making use of mean field approximation (Stanley, 1971) over the joint action space. The
parameters of the Q-function is independent of the number
of agents as it transforms multiple agents interactions into
two entities interactions (single agent vs. a distribution of
other agents). This would also effectively alleviate the exploratory noise (Colby et al., 2015) caused by many other
agents and allow each agent to effectively determine which
actions are beneficial or not, as we have demonstrated from
one of our experiments.
Our work is also closely related to the recent development
of mean field game (MFG) (Lasry & Lions, 2007; Huang
et al., 2006; Weintraub et al., 2006). MFG studies population behaviors resulting from the aggregations of decisions
taken from individuals. Mathematically, the dynamics are
governed by a set of two stochastic differential equations
that model the backward dynamics of individual’s value
function, and the forward dynamics of the aggregate distribution of agent population. Despite that the backward
equation equivalently describes what Bellman equation indicates in MDP, the primarily goal for MFG is rather for a
model-based planning/control and to infer the movements
of the individual density through time. We also employ the
mean field approximation (Stanley, 1971) in physics, but our
1.0
0.8
0.8
0.6
IL
FMQ
Rec-FMQ
MF-Q
MAAC
MF-AC
0.4
0.2
0.0
0
200
400
600
Timestep
(a) N = 100
800
1000
1.0
IL
FMQ
Rec-FMQ
MF-Q
MAAC
MF-AC
0.6
0.4
Performance
1.0
Performance
Performance
Mean Field Multi-Agent Reinforcement Learning
0.2
0.0
0
200
400
600
Timestep
800
1000
(b) N = 500
IL
FMQ
Rec-FMQ
MF-Q
MAAC
MF-AC
0.8
0.6
0.4
0.2
0.0
0
200
400
600
Timestep
800
1000
(c) N = 1000
Figure 1: Learning over N agents in the GS environment with µ = 400 and σ = 200.
work is different in that we focus on a model-free solution of
learning optimal actions when the dynamics of the system
(environment) and the reward function are unknown. Very
recently, Yang et al. (2017) built a connection between MFG
and RL. Their focus is, however, on inverse RL in order to
learn both the reward function and the forward dynamics of
the MFG from the policy data, whereas our goal is to form
a computable Q-function under the framework of temporal
difference learning.
5. Experiments
We analyze and evaluate our algorithms by three different
scenarios, including two stage games (Busoniu & De Schutter), namely Gaussian Squeeze and Ising Model, and a mixed
cooperative-competitive battle game.
5.1. Gaussian Squeeze
5.1.1. E NVIRONMENT
In the Gaussian Squeeze problem (GS) (HolmesParker et al.,
2014), N homogeneous agents determine their individual
PN
j
actions a j to jointly optimize the sum x =
j=1 a for
resource allocation. Each agent has 10 action choices, i.e.,
integer value from 0 to 9 as the quantity of resource utilized
by this agent. This problem is similar to traffic congestion
domains, where each agent is a traffic controller of a single
intersection and wants to send a j vehicles into a shared
main route. Controllers are expected to coordinate with
each other to make the full use of the main route while
avoiding too much traffic. The system objective is defined as
−(x−µ)2
G(x) = xe σ 2 , where µ and σ is the predefined mean
and standard deviation of the system objective’s Gaussian
respectively. The goal of agents is to learn to allocate system
resources efficiently, avoiding over- or under-utilization.
This GS problem is ideal for an ablation study on the impact
of multi-agent exploratory noises toward the learning (Colby
et al., 2015).
5.1.2. M ODEL S ETTINGS
We implement MF-Q and MF-AC following the framework
of centralized training (shared Q) with decentralized execution (independent policy). We compare them with 4
baseline models: (1) Independent Learner (IL) (Watkins &
Dayan, 1992), a decentralized Q-Learning algorithm that
agents only observe local actions and update their Q-values
without regard to the actions performed by other agents; (2)
Frequency Maximum Q-value (FMQ) (Kapetanakis & Kudenko, 2002), a modified IL which increases the Q-values
of actions that frequently produced good rewards in the past;
(3) Recursive Frequency Maximum Q-value (Rec-FMQ)
(Matignon et al., 2012), an improved and robust version
of FMQ which recursively computes the occurrence frequency using a linear interpolation to evaluate and choose
actions; and (4) Multi-agent Actor-Critic (MAAC), a variant of MADDPG architecture (Lowe et al., 2017b) for the
discrete action space in the task. All models use the multilayer perception (MLP) as function approximator. The
full descriptions of implementation details are provided in
Appendix B.1.
5.1.3. R ESULTS
Fig. 1 illustrates our results for different number of agents
in the GS environment with µ = 400 and σ = 200 with
three numbers of agents, i.e., N = 100, 500, 1000, for 3
levels of congestion settings respectively. Fig. 1a shows
that all models perform well in the small GS domain. As
the number of agents becomes larger, Figs. 1b and 1c show
MF-Q and MF-AC are still capable of learning the optimal
allocation effectively after a few iterations, whereas all the
baselines, IL, FMQ, Rec-FMQ and MAAC, fail to learn
at all. This advantage is largely owing to the mean field
methods awareness of other agents’ actions, while still keeping the interactions among agents relatively small, which
reduces the effect of exploratory behaviors of the other
agents. MF-Q converges quicker than MF-AC as it takes
time for the Actor to learn an optimal policy. Both FMQ and
Rec-FMQ fail to reach good performance because agents
cannot distinguish rewards received for the same action due
to various behaviors of other agents and stochastic rewards,
and are thus unable to update their own Q-values w.r.t. the
actual contributions. It is worth noticing that MAAC (all
others actions as independent inputs in the Q function) is
inefficient in learning when the number of agents becomes
large. It fails to handle the non-aggregated noises from the
exploration of many other agents.
Order Parameter
Mean Field Multi-Agent Reinforcement Learning
1.0
0.8
= 1.2
5
5
5
10
10
0.4
15
15
15
0
0.2
0.5
1.0
1.5
2.0
Temperature
2.5
MSE
OP
0.8
0.8
0.6
0.6
0.4
0.4
0.2
0.2
5000
10000
Timestep
15000
200000.0
(a) τ = 0.8
1.0
1.0
MSE
OP
0.8
0.8
0.6
0.6
0.4
0.4
0.2
0.2
0.0 0
5000
10000
Timestep
15000
200000.0
(b) τ = 1.2
Figure 3: Performance against training steps in 20x20 grid
Ising model.
5.2. MARL for Ising Model
5.2.1. E NVIRONMENT
In statistical mechanics, the Ising model is a mathematical model of ferromagnetism (Ising, 1925). It is also
proved to be successful in numerous applications, particularly in sociophysics (Galam & Walliser, 2010). The typical
mean field
model (Stanley, 1971) can iteratively solve
PIsing
j
ha j i =
a a P(a) for every site j on the condition that
the reward mapping function is given. See Eq. (23) in Appendix B.2 for details. Our mean field MARL, however, is
different as the algorithm does not need to know the reward
function, but learns it over time.
To fit our case, we turn the Ising model into a stage game
where
for each site (agent) is as r j = h j a j +
P the reward
λ
j k
j
k∈K j a a , where K is the set of nearest neighbors of
2
j
site j, h ∈ R is the external field affecting site j, and λ ∈ R
is an interaction term determines how much the sites tend
to align in the same direction. Unlike the typical setting
above, each site does not know the reward function exactly,
but aims to maximize its received reward by changing its
individual policy π j to choose the spin state: up or down.
In addition to the observed reward, the order parameter (OP)
(Stanley, 1971) is also a commonly used evaluation metric
|N −N |
for Ising model. OP is defined as ξ = ↑ N ↓ , where N↑
represents the number of sites spinning up, and vice versa.
The closer the OP is to 1, the more orderly the system.
5.2.2. M ODEL S ETTINGS
To prove the correctness of MF-Q value, we implement
MCMC methods (Binder et al., 1993) to simulate the Ising
model and provide the ground truth for our proposed method.
MCMC relies on repeated random sampling to obtain numer-
10
15
< C : = 0.9
0
Mean Squared Error
1.0
Order Parameter
Mean Squared Error
1.0
5
0
5
10
15
0
(a) MF-Q
C:
0
= 1.2
5
5
5
10
10
15
15
15
5
10
15
0
5
10
15
(b) MCMC
5
10
15
> C : = 2.0
0
10
0
> C : = 2.0
0
10
Figure 2: The OP at equilibrium w.r.t. the Ising model
temperature setting in 20x20 grid Ising model.
Order Parameter
C:
0
0.6
0.0
0.0
0.0 0
< C : = 0.9
0
MCMC
MF-Q
0
5
10
15
Figure 4: The spinning pattern of MCMC and MF-Q at equilibrium under different Ising model temperature settings.
ical results and is often used in physical and mathematical
domains. The full descriptions of MCMC and MF-Q for
Ising model are provided in Appendix B.2. Notice that we
do not know exactly the mapping function in Eq. (23), but
rather using Temporal Difference learning to approximate
ha j i during lea rning procedure. One of our learning goals
is to obtain the accurate approximation of ha j i. Once it is
accurately approximated, we should be able to converge to
the simulation result of Ising model provided by MCMC.
5.2.3. R ESULTS
Correctness of MF-Q Value. Fig. 2 illustrates the order
parameter at equilibrium w.r.t. the Ising Model temperature
setting in 20x20 grid Ising model. As shown on the graph,
MF-Q can converge to a similar OP value with MCMC
under different temperature settings, proving the correctness of our algorithm. Also, MF-Q finds a similar Curie
temperature as MCMC, which is around τ = 1.2.
Learning Ability. Fig. 3 illustrates the Mean Squared Error
between learned Q-value and reward target (mean field) in
20x20 grid Ising model w.r.t. training steps. As one can
observe, MF-Q is able to learn the target well under low temperature settings (Fig. 3a). Around the Curie temperature,
the environment will force the site to choose spin state with
too much randomness, resulting in a lower OP and higher
MSE (Fig. 3b).
Convergence Equilibrium. Fig. 4 shows some example
results at equilibrium from our simulation. Most of the
sites choose the same spinning pattern in both MCMC and
MF-Q under a low temperature (τ = 0.9). As the temperature rises, some sites start to change its spinning pattern
and some patches are formed as spontaneous magnetization
with the constraint of neighboring sites and system energy
(τ = 1.2). This phenomenon is mostly observed around
the Curie temperature. After passing the Curie temperature,
the system becomes totally disordered due to large thermal
fluctuations, resulting in a random spinning situation.
Mean Field Multi-Agent Reinforcement Learning
As demonstrated in our previous experiment (Figure 1), a
variant MAAC also fails when the agent number is large.
We thus exclude MADDPG as a baseline, while focusing on
the comparison between mean field methods (MF-Q, MFAC) and their non-mean field counterparts, i.e., independent
Q-learning (IL) and advantageous actor critic (AC).
(a) two groups agents
(b) chasing and cooperation
Figure 5: Battle Game.
1.50
vs IL
vs MF-Q
400
vs MF-AC
vs AC
Total-Reward
Win-Rate
1.25
1.00
0.75
0.50
0.25
0.00
AC
IL
MF-Q
MF-AC
vs IL
vs MF-Q
vs MF-AC
vs AC
300
200
100
0
AC
IL
MF-Q
MF-AC
(a) the rate of wining per round (b) the total reward per round
Figure 6: Performance of algorithms on wining rate, total reward. Each bar cluster shows the index of a set of
competing algorithms, and a higher index is better.
400
MF-AC
250
0
0
Total reward
Total reward
200
200
400
600
250
500
750
1000
MF-Q
1250
800
0
500
1000
Epochs
1500
2000
(a) MF-AC learning curve
0
500
1000
Epochs
1500
2000
(b) MF-Q learning curve
Figure 7: It can be seen from the above two figures that the
MF-AC converges much slower than that of MF-Q.
5.3. Mixed Cooperative-Competitive Battle Game
5.3.1. E NVIRONMENT
The Battle game in the Opensource MAgent system (Zheng
et al., 2017) is a Mixed Cooperative-Competitive scenario
with two armies fighting against each other in a grid world,
shown in Fig. 5. In our settings, each army consists of 64
homogeneous agents. In each timestep, each agent takes
actions to either move to or attack one of the 8 neighbor
grids. The goal of each army is to get more rewards by
collaborating teammates and eliminating all opponents. We
adopt the default setting, i.g., the reward is -0.005 for every
move, 0.2 for attacking an enemy agent, 5 for killing an
enemy agent, -0.1 for attacking a blank grid, and -0.1 for
being attacked or killed.
5.3.2. M ODEL S ETTINGS
Our MF-Q and MF-AC are compared with the baselines
that are successful on the MAgent platform (Zheng et al.,
2017). In the battle game there are a large number of agents,
each of which can only take in observation within its local view, thus the Q-function involves different numbers
of neighbor agents. This makes MADDPG not feasible.
5.3.3. R ESULTS
We get all models with 2000 rounds self-play training, then
use them for cross-comparative experiments. Fig. 6 shows
the result of winning rate and the total reward over 2000
rounds cross-comparative experiments. It is evidence that
on all the metrics mean field methods, i.e., MF-Q and MFAC, largely outperfom the corresponding baselines, i.e., IL
and AC respectively, which shows the effectiveness of the
mean field approximation.
In addition, IL performs much better than AC, and although
worse than its mean field counterpart MF-Q, better than MFAC. This is consistent with the findings in in the Gaussian
Squeeze game (in Figure 1) that IL converges quicker than
AC. This might imply the effectiveness of off-policy learning with shuffled buffer replay in many-agent RL toward a
more stable learning process. The Q-learning family tends
to introduce a positive bias (Hasselt, 2010) by using the maximum action value as an approximation for the maximum
expected action value; we find that such overestimation can
be beneficial for each single agent to find the best response
to others even though the environment itself is still varying.
We further test the situation when the agent number is 8 for
each army, where MF-AC achieves 58% wining rate and get
total rewards of 34.4 while AC achieves 42% wining rate
and get total rewards of 31.6. We see that (i) when the agent
number is small, MF-AC slightly outperforms AC mainly
because of the higher capacity of the Q-network; (ii) when
the agent number gets larger, MF-AC starts to significantly
outperform AC, which verifies the mean field methods work
particularly better for scenarios with a larger number of
agents.
6. Conclusions
In this paper, we have developed mean field reinforcement
learning that tackles the problem of intractable computation
and high dynamics of interactions among a large number
of agents. With mean field theory, we approximated the
Q-function by using the distribution of the actions from
other agents as inputs in order to balance the complexity of
the interactions modeled and the learning ability. It is done
by iteratively learning the best responded actions on the
condition of the distribution of other agents’ actions, which
is subsequently updated each time a new action policy is
obtained. We provided our theoretical analysis on our MFQ learning convergence. The results from three different
Mean Field Multi-Agent Reinforcement Learning
types of environments have shown the effectiveness of our
approaches when the number of agents is large.
There are two interesting future directions. The first is to
investigate the local and global communication mechanisms,
and a more appropriate network architecture for mean field
Q-functions in such a setting. The second is to construct a
Bayesian solution to embed the prior knowledge of agents
optimal policy into the distribution of the agent population.
References
Binder, Kurt, Heermann, Dieter, Roelofs, Lyle, Mallinckrodt, A John, and McKay, Susan. Monte carlo simulation
in statistical physics. Computers in Physics, 7(2):156–
157, 1993.
Foerster, Jakob N., Chen, Richard Y., Al-Shedivat, Maruan,
Whiteson, Shimon, Abbeel, Pieter, and Mordatch, Igor.
Learning with opponent-learning awareness. CoRR,
abs/1709.04326, 2017.
Galam, Serge and Walliser, Bernard. Ising model versus
normal form game. Physica A: Statistical Mechanics and
its Applications, 389(3):481–489, 2010.
Gupta, Jayesh K, Egorov, Maxim, and Kochenderfer, Mykel.
Cooperative multi-agent control using deep reinforcement
learning. In AAMAS, pp. 66–83. Springer, 2017.
Hasselt, Hado V. Double q-learning. In NIPS, pp. 2613–
2621, 2010.
Blume, Lawrence E. The statistical mechanics of strategic
interaction. Games and economic behavior, 5(3):387–
424, 1993.
He, He and Boyd-Graber, Jordan L. Opponent modeling in
deep reinforcement learning. In Balcan, Maria-Florina
and Weinberger, Kilian Q. (eds.), ICML, volume 48, pp.
1804–1813. JMLR.org, 2016.
Bowling, Michael and Veloso, Manuela. Rational and convergent learning in stochastic games. In IJCAI, volume 17,
pp. 1021–1026. Lawrence Erlbaum Associates Ltd, 2001.
HolmesParker, C, Taylor, M, Zhan, Y, and Tumer, K. Exploiting structure and agent-centric rewards to promote
coordination in large multiagent systems. In Adaptive
and Learning Agents Workshop, 2014.
Bowling, Michael and Veloso, Manuela. Multiagent learning using a variable learning rate. Artificial Intelligence,
136(2):215–250, 2002.
Hu, Junling and Wellman, Michael P. Nash q-learning
for general-sum stochastic games. Journal of machine
learning research, 4(Nov):1039–1069, 2003.
Busoniu, Lucian and De Schutter, Bart. A comprehensive
survey of multiagent reinforcement learning.
Huang, Minyi, Malhamé, Roland P, Caines, Peter E, et al.
Large population stochastic dynamic games: closed-loop
mckean-vlasov systems and the nash certainty equivalence principle. Communications in Information & Systems, 6(3):221–252, 2006.
Cao, Zhe, Qin, Tao, Liu, Tie-Yan, Tsai, Ming-Feng, and
Li, Hang. Learning to rank: from pairwise approach
to listwise approach. In Proceedings of the 24th international conference on Machine learning, pp. 129–136.
ACM, 2007.
Colby, Mitchell K., Kharaghani, Sepideh, HolmesParker,
Chris, and Tumer, Kagan. Counterfactual exploration
for improving multiagent learning. AAMAS ’15, pp.
171–179, 2015. ISBN 978-1-4503-3413-6.
de Cote, Enrique Munoz and Littman, Michael L. A
polynomial-time nash equilibrium algorithm for repeated
stochastic games. In McAllester, David A. and Myllymäki, Petri (eds.), UAI 2008, pp. 419–426. AUAI Press,
2008. ISBN 0-9749039-4-9.
Fink, Arlington M et al. Equilibrium in a stochastic n-person
game. Journal of science of the hiroshima university,
series ai (mathematics), 28(1):89–93, 1964.
Foerster, Jakob, Assael, Yannis, de Freitas, Nando, and
Whiteson, Shimon. Learning to communicate with deep
multi-agent reinforcement learning. In NIPS, pp. 2137–
2145, 2016.
Ising, Ernst. Beitrag zur theorie des ferromagnetismus.
Zeitschrift für Physik, 31(1):253–258, 1925.
Jaakkola, Tommi, Jordan, Michael I, and Singh, Satinder P.
Convergence of stochastic iterative dynamic programming algorithms. In NIPS, pp. 703–710, 1994.
Jeong, Seong Hoon, Kang, Ah Reum, and Kim, Huy Kang.
Analysis of game bot’s behavioral characteristics in social
interaction networks of mmorpg. In ACM SIGCOMM,
volume 45, pp. 99–100. ACM, 2015.
Kapetanakis, Spiros and Kudenko, Daniel. Reinforcement
learning of coordination in cooperative multi-agent systems. In NCAI, pp. 326–331, Menlo Park, CA, USA,
2002. ISBN 0-262-51129-0.
Konda, Vijay R. and Tsitsiklis, John N. Actor-critic
algorithms.
In Solla, S. A., Leen, T. K., and
Müller, K. (eds.), Advances in Neural Information
Processing Systems 12, pp. 1008–1014. MIT Press,
2000. URL http://papers.nips.cc/paper/
1786-actor-critic-algorithms.pdf.
Mean Field Multi-Agent Reinforcement Learning
Kreyszig, Erwin. Introductory functional analysis with
applications, volume 1. wiley New York, 1978.
Lasry, Jean-Michel and Lions, Pierre-Louis. Mean field
games. Japanese journal of mathematics, 2(1):229–260,
2007.
Lemke, Carlton E and Howson, Jr, Joseph T. Equilibrium
points of bimatrix games. Journal of the Society for Industrial and Applied Mathematics, 12(2):413–423, 1964.
Littman, Michael L. Markov games as a framework for
multi-agent reinforcement learning. In ICML, volume
157, pp. 157–163, 1994.
Littman, Michael L. Friend-or-foe q-learning in generalsum games. In ICML, volume 1, pp. 322–328, 2001.
Littman, Michael L and Stone, Peter. A polynomial-time
nash equilibrium algorithm for repeated games. Decision
Support Systems, 39(1):55–66, 2005.
Lowe, Ryan, Wu, Yi, Tamar, Aviv, Harb, Jean, Abbeel,
OpenAI Pieter, and Mordatch, Igor. Multi-agent actorcritic for mixed cooperative-competitive environments.
In NIPS, pp. 6382–6393, 2017a.
Lowe, Ryan, Wu, Yi, Tamar, Aviv, Harb, Jean, Abbeel,
Pieter, and Mordatch, Igor. Multi-agent actor-critic for
mixed cooperative-competitive environments. In Guyon,
Isabelle, von Luxburg, Ulrike, Bengio, Samy, Wallach,
Hanna M., Fergus, Rob, Vishwanathan, S. V. N., and
Garnett, Roman (eds.), NIPS, pp. 6382–6393, 2017b.
Matignon, Laetitia, Laurent, Guillaume J, and Le Fort-Piat,
Nadine. Independent reinforcement learners in cooperative markov games: a survey regarding coordination
problems. The Knowledge Engineering Review, 27(1):
1–31, 2012.
Panait, Liviu and Luke, Sean. Cooperative multi-agent
learning: The state of the art. AAMAS, 11(3):387–434,
2005.
Peng, Peng, Yuan, Quan, Wen, Ying, Yang, Yaodong, Tang,
Zhenkun, Long, Haitao, and Wang, Jun. Multiagent
bidirectionally-coordinated nets for learning to play starcraft combat games. arXiv preprint arXiv:1703.10069,
2017.
Rendle, Steffen. Factorization machines with libfm. ACM
Transactions on Intelligent Systems and Technology
(TIST), 3(3):57, 2012.
Silver, David, Lever, Guy, Heess, Nicolas, Degris, Thomas,
Wierstra, Daan, and Riedmiller, Martin. Deterministic
policy gradient algorithms. In ICML, pp. 387–395, 2014.
Stanley, H Eugene. Phase transitions and critical phenomena.
Clarendon, Oxford, 9, 1971.
Szepesvári, Csaba and Littman, Michael L. A unified analysis of value-function-based reinforcement-learning algorithms. Neural computation, 11(8):2017–2060, 1999.
Tan, Ming. Multi-agent reinforcement learning: Independent vs. cooperative agents. In Proceedings of the tenth
international conference on machine learning, pp. 330–
337, 1993.
Troy, Carol Alvarez. Envisioning stock trading where the
brokers are bots. New York Times, 16, 1997.
van der Wal, J. Stochastic dynamic programming: successive approximations and nearly optimal strategies for
markov decision processes and markov games. 1980.
Wang, Jun, Zhang, Weinan, Yuan, Shuai, et al. Display
advertising with real-time bidding (rtb) and behavioural
targeting. Foundations and Trends R in Information Retrieval, 11(4-5):297–435, 2017.
Watkins, Christopher JCH and Dayan, Peter. Q-learning.
Machine learning, 8(3-4):279–292, 1992.
Weintraub, Gabriel Y, Benkard, Lanier, and Van Roy, Benjamin. Oblivious equilibrium: A mean field approximation for large-scale dynamic games. In Advances in
neural information processing systems, pp. 1489–1496,
2006.
Yang, Jiachen, Ye, Xiaojing, Trivedi, Rakshit, Xu, Huan,
and Zha, Hongyuan. Deep mean field games for learning optimal behavior policy of large populations. arXiv
preprint arXiv:1711.03156, 2017.
Zheng, Lianmin, Yang, Jiacheng, Cai, Han, Zhang, Weinan,
Wang, Jun, and Yu, Yong. Magent: A many-agent reinforcement learning platform for artificial collective intelligence. arXiv preprint arXiv:1712.00600, 2017.
Mean Field Multi-Agent Reinforcement Learning
A. Detailed Mean Field Reinforcement Learning Algorithms
Here we provide the pseudo code for Mean Field Q-learning (MF-Q) in Algorithm 1 and Mean Field Actor-Critic (MF-AC)
in Algorithm 2.
Algorithm 1 Mean Field Q-learning (MF-Q)
Initialise Q φ , Q φ− , and ā
while training not finished do
For each agent j, sample action a j w.r.t. Q φ , exploration rate , and ā00 from the last time step
Take the joint action a = (a1, . . . , a N ) and observe the reward r and next state s0
Store (s, a, r, s0, ā) in replay buffer D
for agent j = 1 to N do
Sample a random minibatch of K samples (s, a, r, s0, ā) from D
Sample action (a j )0 w.r.t. Q φ− , exploration rate , and ā
Set y = r j + γ Q φ− ((s j )0, (a j )0, ā0 )
2
P
Update Q-network by minimizing the loss L(φ) = K1 K y − Q φ (s j , a j , ā)
end for
Update target network parameters for each agent j:
j
j
j
φ−
← τφ−
+ (1 − τ)φ−
end while
Algorithm 2 Mean Field Actor-Critic (MF-AC)
Initialize Q φ , Q φ− , πθj , πθj − and ā
while training not finished do
For each agent j, sample action a j = πθj (s j )
Take the joint action a = (a1, . . . , a N ) and observe the reward r and next state s0
Store (s, a, r, s0, ā) in replay buffer D
for agent j = 1 to N do
Sample a random minibatch of K samples (s, a, r, s0, ā) from D
Set y = r j + γ Q φ− ((s j )0, (a j )0, ā0 )|(a j )0 =π j ((s j )0 )
θ−
2
P
Update critic by minimizing the loss L(φ) = K1 K y − Q φ (s j , a j , ā)
Update actor using the sampled policy gradient:
∇θ j J(θ j ) ≈
1 X
∇θ j log πθj ((s j )0 )Q φ− ((s j )0, (a j )0, ā0 )|(a j )0 =π j ((s j )0 )
K K
θ−
end for
Update target network parameters for each agent j:
j
j
j
φ−
← τφ−
+ (1 − τ)φ−
end while
j
j
j
θ−
← τθ −
+ (1 − τ)θ −
B. Experiment Details
B.1. Gaussian Squeeze
IL, FMQ, Rec-FMQ and MF-Q all use a three-layer MLP to approximate Q-value. All agents share the same Q-network
for each experiment. The shared Q-network takes an agent embedding as input and computes Q-value for each candidate
action. For MF-Q, we also feed in the action approximation ā. We use the Adam optimizer with a learning rate of
0.00005 and -greedy exploration unless otherwise specified. For FMQ, we set the exponential decay rate s = 0.006, start
temperature max_temp=1000 and FMQ heuristic c = 5. For Rec-FMQ, we set the frequency learning rate α f = 0.01.
Mean Field Multi-Agent Reinforcement Learning
Algorithm 3 MCMC in Ising Model
initialize spin state a ∈ {−1, 1} N for N sites
while training not finished do
randomly choose site j ∈ I N
j
flip the spin state for site j: (a j )0 ← −aP
compute neighbor energy E(a, h) = − j (h j a j +
randomly choose ∼ U(0, 1)
if exp((E(a j , h) − (E(a j )0, h))/τ) > then
a j ← (a j )0
end if
end while
λ
2
P
k∈K j
a j ak ) for a j and (a j )0
MAAC and MF-AC use the Adam optimizer with a learning rate of 0.001 and 0.0001 for Critics and Actors respectively,
and τ = 0.01 for updating the target networks. We share the Critic among all agents in each experiment and feed in an agent
embedding as extra input. Actors are kept seperate. The discounted factor γ is set to be 0.95 and the mini-batch size is set to
be 200. The size of the replay buffer is 106 and we update the network parameters after every 500 samples added to the
replay buffer.
For all models, we use the performance of the joint-policy learned up to that point if learning and exploration were turned
off (i.e., take the greedy action w.r.t. the learned policy) to compare our method with the above baseline models.
B.2. Ising Model
An Ising model is defined as a stateless system with N homogeneous sites on a finite square lattice. Each site determines
their individual spin a j to interact with each other and aims to minimize the system energy for a more stable environment.
The system energy is defined as
E(a, h) = −
X
j
(h j a j +
λ X j k
a a )
2
j
(20)
k∈K
where K j is the set of nearest neighbors of site j, h j ∈ R is the external field affecting site j, and λ ∈ R is an interaction
term determines how much the sites tend to align in the same direction. The system is said to reach an equilibrium point
when the system energy is minimized, with the probability
exp (−E(a, h)/τ)
P(a) = P
,
a exp(−E(a, h)/τ)
(21)
where τ is the system temperature. When the temperature rises beyond a certain point (the Curie temperature), the system
can no longer keep a stable form and a phase transition happens. As the ground-truth is known, we would be able to evaluate
the correctness of the Q-function learning when there is a large body of agents interacted.
P
The mean field theory provides an approximate solution to ha j i = a a j P(a) through a set of self-consistent mean field
equations
P
exp −[h j a j + λ k∈K j hak i]/τ
j
.
P
ha i =
(22)
1 + exp −[h j a j + λ k∈K j hak i]/τ
which can be solved iteratively by
j (t+1)
ha i
where t represents the number of iterations.
P
exp −[h j a j + λ k∈K j hak i(t) ]/τ
,
P
=
1 + exp −[h j a j + λ k∈K j hak i(t) ]/τ
(23)
To learn an optimal joint policy π ∗ for Ising model, we use the stateless Q-learning with mean field approximation (MF-Q),
defined as
Q j (a j , ā) ← Q j (a j , ā) + α[r j − Q j (a j , ā)],
(24)
Mean Field Multi-Agent Reinforcement Learning
where the mean ā is given as the mean hai i from the last time step, and the individual reward is
λ X j k
r j = hj aj +
a a .
2
j
(25)
k∈K
To balance the trade-off between exploration and exploitation under low temperature settings, we use a policy with Boltzmann
exploration and a decayed exploring temperature. The temperature for Boltzmann exploration of MF-Q is multiplied by a
decay factor exponentially through out the training process.
Without lost of generality, we assume λ > 0, thus neighboring sites with the same action result in lower energy (observe
higher reward) and are more stable. Each site should also align with the sign of external field h j to reduce the system energy.
For simplification, we eliminate the effect of external fields and assume the model to be discrete, i.e., ∀ j ∈ N, h j = 0, a j ∈
{−1, 1}.
We simulate the Ising model using Metropolis Monte Carlo methods (MCMC). After initialization, we randomly change a
site’s spin state and calculate the energy change, select a random number between 0 and 1, and accept the state change only
(E j −(E j )0 )
τ
if the number is less than e
the more probable spin states.
. This is called the Metropolis technique, which saves computation time by selecting
B.3. Battle Game
IL and MF-Q have almost the same hyper-parameters settings. The learning rate is α = 10−4 , and with a dynamic
exploration rate linearly decays from γ = 1.0 to γ = 0.05 during the 2000 rounds training. The discounted factor γ is set to
be 0.95 and the mini-batch size is 128. The size of replay buffer is 5 × 105 .
AC and MF-AC also have almost the same hyper-parameters settings. The learning rate is α = 10−4 , the temperature of
soft-max layer in actor is τ = 0.1. And the coefficient of entropy in the total loss is 0.08, the coefficient of value in the total
loss is 0.1.
C. Details of MF-Q
P
Proposition 1. Let the metric space be R N , and the metric be M(a, b) = j |a j − b j |, ∀ a = {a j }1N , b = {b j }1N . If the
Q function is Lipschitz continuous w.r.t. the input of a j , then for sufficiently small temperature β, the operator
exp − βQ(s, a j , ā)
j
B(a ) = P
exp − βQ(s, ak , ā)
a k ∈A j
is a contraction mapping. The proof is in Appendix due to the space limit.
Proof. Following the contraction mapping theorem (Kreyszig, 1978), in order to be a contraction, the operator has to satisfy:
where 0 ≤ α < 1.
M(B(a), B(b)) ≤ αM(a, b), ∀a, b
Here we start from binary case and then adapt to the general multinomial case. We first rewrite B(a j ) as:
B(a j ) = 1 −
1
.
1 + exp (−βQ(s, a j , ā))
Then we have:
1
1
−
1 + e−βQ(s,a j , ā)
1 + e−βQ(s,b j , ā)
βe−βQ0
=
Q(s, a j , ā) − Q(s, b j , ā)
(1 + e−βQ0 )2
1
· Q(s, a j , ā) − Q(s, b j , ā)
≤
4T
1
≤
· K · a j − bj
4T
|B(a j ) − B(b j )| =
(26)
Mean Field Multi-Agent Reinforcement Learning
In the 2nd equation, we apply the mean value theorem in Calculus: ∃x0 ∈ [x1, x2 ], s.t., f (x1 ) − f (x2 ) = f 0 (x0 )(x1 − x2 ).
In the 3rd equation we use the maximum value for e−βQ0 /(1 + e−βQ0 )2 = 1/4 when Q0 = 0. In the last equation we we
apply the Lipschitz constraint in the assumption where constant K ≥ 0. Finally, we have:
M(B(a), B(b)) ≤
=
X
1
·K·
a j − bj
4T
j
K
M(a, b)
4T
(27)
In order for the contraction to hold, T > K4 . In other words, when the action space is binary for each agent, and the
temperature is sufficiently large, the mean field procedure converges.
This proposition can be easily extended to multinomial case by replacing binary variable a j by a multi-dimensional binary
indicator vector a j , on each dimension, the rest of the derivations would remain essentially the same.
C.1. Discussion on Rationality
In aligned with (Bowling & Veloso, 2001; 2002), we argue that to better evaluate a multi-agent learning algorithm, on top of
the convergence guarantee, discussion on property of Rationality is also needed.
Property 1. (also see (Bowling & Veloso, 2001; 2002)) In an N-agent stochastic game defined in this paper, given all
agents converge to stationary policies, if the learning algorithm converges to a policy that is a best response to the other
agents’ policies, then the algorithm is Rationale.
Our mean field Q-learning is rational in that Eq. (6) converts many agents interactions into two-body interactions between a
single agent and the distribution of other agents actions. When all agents follow stationary policies, their policy distribution
would be stationary too. As such the two-body stochastic game becomes an MDP, and the agent would choose a policy
(based on Assumption 2) which is the best response to the distribution of other stationary policies. As agents are symmetric
in our case, they all the best response to the distributions, therefore are rational.
| 2cs.AI |
Symmetry-Aware Predicate Abstraction
for Shared-Variable Concurrent Programs
(Extended Technical Report)
arXiv:1102.2330v1 [cs.PL] 11 Feb 2011
Alastair Donaldson, Alexander Kaiser, Daniel Kroening, and Thomas Wahl
Oxford University Computing Laboratory, United Kingdom
Abstract. Predicate abstraction is a key enabling technology for applying finitestate model checkers to programs written in mainstream languages. It has been
used very successfully for debugging sequential system-level C code. Although
model checking was originally designed for analyzing concurrent systems, there
is little evidence of fruitful applications of predicate abstraction to shared-variable
concurrent software. The goal of this paper is to close this gap. We have developed a symmetry-aware predicate abstraction strategy: it takes into account
the replicated structure of C programs that consist of many threads executing
the same procedure, and generates a Boolean program template whose multithreaded execution soundly overapproximates the concurrent C program. State
explosion during model checking parallel instantiations of this template can now
be absorbed by exploiting symmetry. We have implemented our method in the
S ATA BS predicate abstraction framework, and demonstrate its superior performance over alternative approaches on a large range of synchronization programs.
1 Introduction
Concurrent software model checking is one of the most challenging problems facing
the verification community today. Not only does software generally suffer from data
state explosion. Concurrent software in particular is susceptible to state explosion due
to the need to track arbitrary thread interleavings, whose number grows exponentially
with the number of executing threads.
Predicate abstraction [1] was introduced as a way of dealing with data state explosion: the program state is approximated via the values of a finite number of predicates over the program variables. Predicate abstraction turns C programs into finite-state
Boolean programs [2], which can be model checked. Since insufficiently many predicates can cause spurious verification results, predicate abstraction is typically embedded into a counterexample-guided abstraction refinement (CEGAR) framework [3]. The
feasibility of the overall approach has been convincingly demonstrated for sequential
software by the success of the S LAM project at Microsoft, which was able to discover
numerous control-dominated errors in low-level operating system code [4].
The majority of concurrent software is written using mainstream APIs such as
POSIX threads (pthreads) in C/C++, or using a combination of language and library
support, such as the Thread class, Runnable interface and synchronized construct in
Java. Typically, multiple threads are spawned — up front or dynamically, in response to
varying system load levels — to execute a given procedure in parallel, communicating
via shared global variables. For such shared-variable concurrent programs, predicate
abstraction success stories similar to that of S LAM are few and far between. The bottleneck is the exponential dependence of the generated state space on the number of
running threads, which, if not addressed, permits exhaustive exploration of such programs only for trivial thread counts. The key to obtaining scalability is to exploit the
symmetry naturally exhibited by these programs, namely their invariance under permutations of the participating threads.
Fortunately, much progress has recently been made on analyzing replicated Boolean
programs, where a non-recursive Boolean template program is executed concurrently
by many threads [5–7]. In this paper, we present an approach to predicate-abstracting
concurrent programs that leverages this recent progress. More precisely:
– We describe a scheme to translate a non-recursive C program P with shared (globalscope) and local (procedure-scope) variables into a Boolean program B such that
the n-thread Boolean program, denoted Bn , soundly overapproximates the n-thread
C program, denoted Pn . We call such an abstraction method symmetry-aware.
– Our method permits predicates over arbitrary C program variables, local or global.
We illustrate below the ramifications of this objective.
We also show in this paper how our approach can be implemented for C-like languages,
complete with pointers and aliasing, and discuss the issues of spurious error detection
and predicate refinement. In the rest of the Introduction, however, we illustrate why
approaching the above two main goals naı̈vely can render abstraction unsound, creating
the danger of missing bugs, which defies the purpose of reliable program verification.
A remark on notation: In program listings, we use == for the comparison operator,
while = denotes assignment (as in C). Concurrent threads are assumed to interleave with
statement-level granularity; see the discussion in the Conclusion on this subject.
1.1 Predicate Abstraction using Mixed Predicates
The Boolean program B to be built from the C program P will consist of Boolean
variables, one per predicate as usual. Since B is to be executed by parallel threads,
its variables have to be partitioned into “shared” and “local”. As these variables track
the values of various predicates over C program variables, the “shared” and “local”
attributes clearly depend on the attributes of the C variables a predicate is formulated
over. We therefore classify predicates as follows.
Definition 1 A local predicate refers solely to local C program variables. Analogously,
a shared predicate refers solely to shared variables. A mixed predicate is neither local
nor shared.
We reasonably assume that each predicate refers to at least one program variable.
A mixed predicate thus refers to both local and shared variables, and the above classification partitions the set of predicates into the three categories.
Given this classification, consider a local predicate φ, which can change only as a
result of a thread changing one of its local C variables; a change that is not visible to any
other thread. This locality is inherited by the Boolean program if predicate φ is tracked
by a local Boolean variable. Similarly, shared predicates are naturally tracked by shared
Boolean variables.
For a mixed predicate, the decision whether it should be tracked in the shared or in
the local space of the Boolean program is non-obvious. Consider first the following program P and the corresponding generated Boolean program B, which tracks the mixed
predicate s != l in a local Boolean variable b:
0: shared int s = 0;
P:
local int l = 1;
1: assert s != l;
2: ++s;
0: local bool b = 1;
B:
1: assert b;
2: b = b ? ⋆ : 1;
Consider the program P2 , a two-thread instantiation of P. It is easy to see that execution of P2 can lead to an assertion violation, while the corresponding concurrent
Boolean program B2 is correct. (In fact, Bn is correct for any n > 0.) As a result,
B2 is an unsound abstraction for P2 . Consider now the following program P′ and its
Boolean abstraction B′ , which tracks the mixed predicate s == l in a shared Boolean
variable b:
s = 0;
bool t = 1;
int l = 0;
B′ :
t ↔ (s == l);
t;
= 0;
0: shared int
shared
local
P′ :
1: assert
2: assume
3: ++l , t
0: shared bool b = 1;
shared bool t = 1;
1: assert t ↔ b;
2: assume t;
3: b , t = ( b ? 0 : ⋆ ) , 0;
Execution of (P′ )2 leads to an assertion violation if the first thread executes P′ without interruption, while (B′ )2 is correct. We conclude that (B′ )2 is unsound for (P′ )2 .
The unsoundness can be eliminated by making b local in B′ ; an analogous reasoning
removes the unsoundness in B as an abstraction for P. It is clear from these examples,
however, that in general a predicate of the form s == l that genuinely depends on s
and l cannot be tracked by a shared or a local variable without further amendments to
the abstraction process.
At this point it may be useful to pause for a moment and consider whether, instead
of designing solutions that deal with mixed predicates, we may not be better off by
banning them, relying solely on shared and local predicates. Such restrictions on the
choice of predicates render very simple bug-free programs unverifiable using predicate
abstraction, including the following program P′′ :
0: shared int r = 0;
P′′ :
shared int s = 0;
local int l = 0;
1: ++r;
2: if (r == 1) then
3:
f ();
f ():
4: ++s, ++l;
5: assert s == l;
6: goto 4;
The assertion in P′′ cannot be violated, no matter how many threads execute P, since
no thread but the first will manage to execute f . It is easy to prove that, over a set of
non-mixed predicates (i.e. no predicate refers to both l and one of {s, r}), no invariant
is computable that is strong enough to prove s == l. We have included such a proof in
Appendix B.
A technically simple solution to all these problems is to instantiate the template P
n times, once for each thread, into programs {P1 , . . . , Pn }, in which indices 1, . . . , n
are attached to the local variables of the template, indicating the variable’s owner. Every
predicate that refers to local variables is similarly instantiated n times. The new program
has two features: (i) all its variables, having unambiguous names, can be declared at the
global scope and are thus shared, including the original global program variables, and
(ii) it is multi-threaded, but the threads no longer execute the same code. Feature (i)
allows the new program to be predicate-abstracted in the conventional fashion: each
predicate is stored in a shared Boolean variable. Feature (ii), however, entails that the
new program is no longer symmetric. Model checking it will therefore have to bear the
brunt of concurrency state explosion. Such an approach, which we refer to as symmetryoblivious, will not scale beyond a very small number of threads.
To summarize our findings: Mixed predicates are necessary to prove properties for
even very simple programs. On the other hand, they cannot be tracked using standard
thread-local or shared variables. Disambiguating local variables avoids mixed predicates, but destroys symmetry. It is the goal of this paper to design a solution without the
need to compromise.
2 Symmetry-Aware Predicate Abstraction
In order to illustrate our method, let P be a program defined over a set of variables V
that is partitioned in the form V = VS ∪VL into shared and local variables. The parallel
execution of P by n threads is a program defined over the shared variables and n copies
of the local variables, one copy for each thread. A thread is nondeterministically chosen
to be active, i.e. to execute a statement of P, potentially modifying the shared variables,
and its own local variables, but nothing else. In this section, we ignore the specific
syntax of statements, and we do not consider language features that introduce aliasing,
such as pointers (these are the subject of Section 3). Therefore, an assignment to a
variable v cannot modify a variable other than v, and an expression φ depends only on
the variables occurring in it, which we refer to as Loc(φ) = {v : v occurs in φ}.
2.1 Mixed Predicates and Notify-All Updates
Our goal is to translate the program P into a Boolean program B such that, for any n,
a suitably defined parallel execution of B by n threads overapproximates the parallel
execution of P by n threads. Let E = {φ1 , . . . , φm } be a set of predicates over P, i.e.
a set of Boolean expressions over variables in V . We say φi is
shared if Loc(φi ) ⊆ VS ,
local if Loc(φi ) ⊆ VL , and
mixed otherwise, i.e. Loc(φi ) ∩ VL 6= ∅ and Loc(φi ) ∩ VS 6= ∅.
We declare, in B, Boolean variables {b1 , . . . , bm }; the intention is that bi tracks the
value of φi during abstract execution of P. We partition these Boolean variables into
shared and local by stipulating that bi is shared if φi is shared; otherwise bi is local.
In particular, mixed predicates are tracked in local variables. Intuitively, the value
of a mixed predicate φi depends on the thread it is evaluated over. Declaring bi shared
would thus necessarily lose information. Declaring it local does not lose information,
but, as the example in the Introduction has shown, is insufficient to guarantee a sound
abstraction. We will see shortly how to solve this problem.
Each statement in P is now translated into a corresponding statement in B. Statements related to flow of control are handled using techniques from standard predicate
abstraction [2]; the distinction between shared, mixed and local predicates does not
matter here. Consider an assignment to a variable v in P and a Boolean variable b
of B with associated predicate φ. We first check whether variable v affects φ, written
affects(v, φ). Given that in this section we assume no aliasing, this is the case exactly
if v ∈ Loc(φ). If affects(v, φ) evaluates to false, b does not change. Otherwise, code
needs to be generated to update b. This code needs to take into account the “flavors” of
v and φ, which give rise to three different flavors of updates of b:
shared update: Suppose v and φ are both shared. An assignment to v is visible to all
threads, so the truth of φ is modified for all threads. This is reflected in B: by our
stipulation above, the shared predicate φ is tracked by the shared variable b. Thus,
we simply generate code to update b according to standard sequential predicate
abstraction rules; the new value of b is shared among all threads.
local update: Suppose v is local and φ is local or mixed. An assignment to v is visible
only by the active (executing) thread, so the truth of φ is modified only for the
active thread. This also is reflected in B: by our stipulation above, the local or
mixed predicate φ is tracked by the local variable b. Again, sequential predicate
abstraction rules suffice; the value of b changes only for the active thread.
notify-all update: Suppose v is shared and φ is mixed. An assignment to v is visible
to all threads, so the truth of φ is modified for all threads. This is not reflected in B:
by our stipulation above, the mixed predicate φ is tracked by the local variable b,
which will be updated only by the active thread. We will solve this problem by (i)
generating code to update b according to standard sequential predicate abstraction
rules, and (ii) notifying all passive threads of the modification of the shared variable
v, so as to allow them to update their local copy of b.
We write must notify(v, φ) if the shared variable v affects the mixed predicate φ:
must notify(v, φ)
=
affects(v, φ) ∧ v ∈ VS ∧ (Loc(φ) ∩ VL 6= ∅) .
This formula evaluates to true exactly when it is necessary to notify passive threads of
an update to v. What remains to be discussed in the rest of this section is how notifications are implemented in B.
2.2 Implementing Notify-All Updates
We pause to recap some terminology and notation from sequential predicate abstraction [2]. Given a set E = {φ1 , . . . , φm } of predicates tracked by variables {b1 , . . . , bm },
an assignment statement st is translated into the following code, in parallel for each
i ∈ {1, . . . , m}:
if
F (WP ( φi , st)) then bi = 1
else if F (WP (¬φi , st)) then bi = 0
(1)
else
bi = ⋆ .
Here, ⋆ is the nondeterministic choice expression, WP the weakest precondition operator, and F the operator that strengthens an arbitrary predicate to a disjunction of
cubes over the bi . For example, with predicate φ :: (l < 10) tracked by variable b,
E = {φ} and statement st :: ++l, we obtain F (WP (φ, st )) = F (l < 9) = false and
F (WP (¬φ, st )) = F (l >= 9) = (l >= 10) = ¬φ, so that (1) reduces to
b = ( b ? ⋆ : 0 ).
In general, (1) is often abbreviated using the assignment
bi = choose(F (WP (φi , st )), F (WP (¬φi , st))) ,
where choose(x, y) returns 1 if x evaluates to true, 0 if (¬x) ∧ y evaluates to true,
and ⋆ otherwise. Abstraction of control flow guards uses the G operator, which is dual
to F : G(φ) = ¬F (¬(φ)).
Returning to symmetry-aware predicate abstraction, if must notify(v, φ) evaluates
to true for φ and v, predicate φ is mixed and thus tracked in B by some local Boolean
variable, say b. Predicate-abstracting an assignment of the form v = χ requires updating the active thread’s copy of b, as well as broadcasting an instruction to all passive
threads to update their copy of b, in view of the new value of v. This is implemented using two assignments, which are executed in parallel. The first assignment is as follows:
b=choose(F (WP (φ, v = χ)), F (WP (¬φ, v = χ))) .
(2)
This assignment has standard predicate abstraction semantics:1 variable b of the active
thread is updated by computing the weakest precondition of predicate φ and its negation
with respect to the statement v = χ, applying the strengthening operator F to make
the precondition expressible over the existing predicates, and by applying the choose
operator, which may introduce nondeterminism. Note that, since expression χ involves
only local variables of the active thread and shared variables, only predicates over those
variables are involved in the defining expression for b.
The second assignment looks similar, but introduces a new symbol:
[b]=choose(F (WP ([φ], v = χ)), F (WP (¬[φ], v = χ))) .
(3)
The notation [b] stands for the copy of local variable b owned by some passive thread.
Similarly, [φ] stands for the expression defining predicate φ, but with every local variable occurring in the expression replaced by the copy owned by the passive thread; this
is the predicate φ in the context of the passive thread. Weakest precondition computation is with respect to [φ], while the assignment v = χ, as an argument to WP , is
1
Our presentation is in terms of the Cartesian abstraction [8], as used in [2], but our method in
general is independent of the abstraction used.
unchanged: v is shared, and local variables appearing in the defining expression χ must
be interpreted as local variables of the active thread. Assignment (3) has the effect of
updating variable b in every passive thread. We call Boolean programs involving assignments of the form [b] = ... Boolean broadcast programs; a formal syntax and
semantics for such programs is given in Appendix A.
Let us illustrate the above technique using a canonical example: consider the assignment s = l, for shared and local variables s and l, and define the mixed predicate
φ :: (s == l). The first part of the above parallel assignment simplifies to b = true.
For the second part, we obtain:
[b] = choose(F (WP (s==[l], s=l)), F (WP (¬(s==[l]), s=l))) .
Computing weakest preconditions, this reduces to:
[b] = choose(F (l==[l]), F (¬(l==[l]))) .
Precision of the Abstraction. To evaluate this expression further, we have to decide on
the set of predicates available to the F operator to express the preconditions. If this set
includes only predicates over the shared variables and the local variables of the passive
thread that owns [b], the predicate l == [l] is not expressible and must be strengthened to false. The above assignment then simplifies to [b] = choose(false, false), i.e.
[b] = ⋆. The mixed predicates owned by passive threads are essentially invalidated
when the active thread modifies a shared variable occurring in such predicates, resulting in a very imprecise abstraction.
We can exploit information stored in predicates local to other threads, to increase the
precision of the abstraction. For maximum precision one could make all other threads’
predicates available to the strengthening operator F . This happens in the symmetryoblivious approach sketched in the Introduction, where local and mixed predicates are
physically replicated and declared at the global scope and can thus be made available
to F . Not surprisingly, in practice, replicating predicates in this way renders the abstraction prohibitively expensive. We analyze this experimentally in Section 5.
A compromise which we have found to work well in practice (again, demonstrated
in Section 5) is to equip operator F with all shared predicates, all predicates of the
passive thread owning [b], and also predicates of the active thread. This arrangement is
intuitive since the update of a passive thread’s local variable [b] is due to an assignment
performed by some active thread. Applying this compromise to our canonical example:
if both s == [l] and s == l evaluate to true before the assignment s=l, we can conclude that [l] == l before the assignment, and hence s == [l] after the assignment.
Using ⊕ to denote exclusive-or, the assignment to [b] becomes:
[b] = choose([b] ∧ b, [b] ⊕ b) .
2.3 The Predicate Abstraction Algorithm
We now show how our technique for soundly handling mixed predicates is used in
an algorithm for predicate abstracting C-like programs. To present the algorithm compactly, we assume a language with three types of statement: assignments, nondeterministic gotos, and assumptions. Control-flow can be modelled via a combination of gotos
and assumes, in the standard way.
Algorithm 1 Predicate abstraction
Input: Program template P, set of predicates {φ1 , . . . , φm }
Output: Boolean program B over variables b1 , . . . , bm
1: for each statement d: stmt of P do
2:
if stmt is v = ψ then
3:
{i1 , . . . , if } ← {i | 1 ≤ i ≤ m ∧ affects (v, φi )}
4:
{j1 , . . . , jg } ← {j | 1 ≤ i ≤ m ∧ must notify(v, φi )}
choose(F(WP(φi1 , v=ψ)), F(WP (¬φi1 , v=ψ)))
bi1
..
..
,
. =
.
choose(F(WP(φif , v=ψ)), F(WP (¬φif , v=ψ)))
bif
5:
output d:
choose(F(WP ([φj1 ], v=ψ)), F(WP(¬[φj1 ], v=ψ)))
[bj1 ]
..
..
. =
.
choose(F(WP ([φjg ], v=ψ)), F(WP(¬[φjg ], v=ψ)))
[bjg ]
6:
else if stmt is goto d1 , . . . , dm then
7:
output d: goto d1 , . . . , dm ;
8:
else if stmt is assume φ then
9:
output d: assume G(φ);
Algorithm 1 processes an input program template of this form and outputs a corresponding Boolean broadcast program template. Statements goto and assume are handled as in standard predicate abstraction: the former are left unchanged, while the latter
are translated directly except that the guard of an assume statement is expressed over
Boolean program variables using the G operator (see Section 2.2).
The interesting part of the algorithm for us is the translation of assignment statements. For each assignment, a corresponding parallel assignment to Boolean program
variables is generated. The affects and must notify predicates are used to decide for
which Boolean variables regular and broadcast assignments are required, respectively.
3 Symmetry-Aware Predicate Abstraction with Aliasing
In Section 2 we outlined our novel predicate abstraction technique, ignoring complications introduced by pointers and aliasing. We now explain how symmetry-aware predicate abstraction is realized in practice, for C programs that manipulate pointers. We
impose one restriction: we do not consider programs where a shared pointer variable, or
a pointer variable local to thread i, can point to a variable local to thread j (with j 6= i).
This scenario arises only when a thread copies the address of a stack or thread-local
variable to the shared state. This unusual programming style allows thread i to directly
modify the local state of thread j at the C program level, breaking the asynchronous
model of computation assumed by our method.
For ease of presentation we consider the scenario where program variables either
have a base type (e.g. int or float), or pointer type (e.g. int* or float**). Our method
can be extended to handle records, arrays and heap-allocated memory in a straightfor-
ward but laborious manner. As in [2], we also assume that input programs have been
processed so that l-values involve at most one pointer dereference.
Alias information is important in deciding, once and for all, whether predicates
should be classed as local, mixed or shared. For example, let p be a local variable
of type int*, and consider predicate φ :: (*p == 1). Clearly φ is not shared since it
depends on local variable p. Whether φ should be regarded as a local or mixed predicate
depends on whether p may point to the shared state: we regard φ as local if p can
never point to a shared variable, otherwise φ is classed as mixed. Alias information also
lets us determine whether a variable update may affect the truth of a given predicate,
and whether it is necessary to notify other threads of this update. We now show how
these intuitions can be formally integrated with our predicate abstraction technique.
This involves suitably refining the notions of local, shared and mixed predicates, and the
definitions of affects and must notify introduced in Section 2 and used by Algorithm 1.
3.1 Aliasing, Locations of Expressions, and Targets of l-values
We assume the existence of a sound pointer alias analysis for concurrent programs,
e.g. [9], which we treat as a black box. This procedure conservatively tells us whether
a shared variable with pointer type may refer to a local variable. As discussed at the
start of Section 3, we reject programs where this is the case.2 Otherwise, for a program
template P over variables V , alias analysis yields a relation 7→d ⊆ V × V for each
program location d. For v, w ∈ V , if v 67→d w then v provably does not point to w at d.
For an expression φ and program point d, we write loc(φ, d) for the set of variables
that it may be necessary to access in order to evaluate φ at d, during an arbitrary program
run. The precision of loc(φ, d) is directly related to the precision of alias analysis.
Definition 2 For a constant value z, v ∈ V and k > 0, we define:
loc(z, d) = ∅
loc(&v, d) = ∅
loc(v, d) = {v}
S
loc(|* .{z
. . *} v, d) = {v} ∪ w∈V {loc(|* .{z
. . *} w, d) | v 7→d w}
k
k−1
For other compound expressions, loc(φ, d) is defined recursively in the obvious way.
Definition 2 captures the fact that evaluating a pointer dereference *v involves reading
both v and the variable to which v points, while evaluating an “address-of” expression,
&v, requires no variable accesses: addresses of variables are fixed at compile time.
Definition 3 For an expression φ, Loc(φ) is the set of variables that may need to be
accessed to evaluate φ at an arbitrary program point during an arbitrary program run:
[
Loc(φ) =
loc(φ, d).
1≤d≤k
Note how this definition of Loc generalizes that used in Section 2.
2
This also eliminates the possibility of thread i pointing to variables in thread j 6= i: the address
of a variable in thread j would have to be communicated to thread i via a shared variable.
Definition 4 We write targets(x, d) for the set of variables that may be modified by
writing to l-value x at program point d:
targets(v, d) = {v}
targets(*v, d) = {w ∈ V | v 7→d w}
Note that we have targets(*v, d) 6= loc(*v, d). This is because writing through *v
modifies only the variable to which v points, while reading the value of *v involves
reading the value of v, to determine which variable w is pointed to by v, and then
reading the value of w.
3.2 Shared, Local and Mixed Predicates in the Presence of Aliasing
In the presence of pointers, we define the notion of a predicate φ being shared, local,
or mixed exactly as in Section 2.1, only with the generalization of Loc presented in
Definition 3. In Section 2.1, without pointers, we could classify φ purely syntactically,
based on whether any shared variables appear in φ. In the presence of pointers, we must
classify φ with respect to alias information; our definition of Loc takes care of this.
Recall from Section 2.1 that we defined affects(v, φ) = (v ∈ Loc(φ)) to indicate
that updating variable v may affect the truth of predicate φ. In the presence of pointers,
this definition no longer suffices. The truth of φ may be affected by assigning to l-value
x if x may alias some variable on which φ depends. Whether this is the case depends
on the program point at which the update occurs. Definitions 2 and 4 of loc and targets
allow us to express this:
affects(x, φ, d) = (targets(x, d) ∩ loc(φ, d) 6= ∅).
We also need to determine whether an update affects the truth of a predicate only
for the thread executing the update, or for all threads. The definition of must notify
presented in Section 2.1 needs to be adapted to take aliasing into account. At first sight,
it seems that we must simply parameterise affects according to program location, and
replace the conjunct v ∈ VS with the condition that x may target some shared variable:
must notify (x, φ, d) = affects(x, φ, d) ∧ (Loc(φ) ∩ VL 6= ∅)
∧ (targets(x, d) ∩ VS 6= ∅) .
However, this is unnecessarily strict. We can refine the above definition to minimise the
extent to which notifications are required, as follows:
must notify(x, φ, d) = (targets(x, d) ∩ Loc(φ) ∩ VS 6= ∅) ∧ (Loc(φ) ∩ VL 6= ∅) .
The refined definition avoids the need for thread notification in the following scenario. Suppose we have shared variables s and t, local variable l, local pointer variable
p, and predicate φ :: (s > l). Consider an assignment to *p at program point d. Suppose that alias analysis tells us exactly p 7→d t and p 7→d l. The only shared variable
that can be modified by assigning through *p at program point d is t, and the truth of
φ does not depend on t. Thus the assignment does not require a “notify-all” with respect to φ. Working through the definitions, we find that our refinement of must notify
correctly determines this, while the direct extension of must notify from Section 2.1
would lead to an unnecessary “notify-all”.
The predicate abstraction algorithm, Algorithm 1, can now be adapted to handle
pointers: parameter d is simply added to the uses of affects and must notify. Handling
of pointers in weakest preconditions works as in standard predicate abstraction [2],
using Morris’s general axiom of assignment [10].
4 Closing the CEGAR Loop
So far we have presented a novel technique for predicate-abstracting symmetric concurrent programs. We have integrated our method with the S ATA BS CEGAR-based
verifier [11], using the Cartesian abstraction method [8] and the maximum cube length
approximation [2]. We now sketch how we have adapted the other phases of the CEGAR loop: model checking, simulation and refinement, to accurately handle concurrency; a detailed description of the entire process is left to an extended version of this
paper.
Model checking Boolean broadcast programs. Our predicate abstraction technique
generates a concurrent Boolean broadcast program. The extended syntax and semantics
for broadcasts mean that we cannot simply use existing concurrent Boolean program
model checkers, such as B OOM [12] or B OPPO [13], for the model checking phase of
the CEGAR loop. We have implemented a prototype extension of B OOM, which we call
B-B OOM. B-B OOM extends the counter abstraction-based symmetry reduction capabilities of B OOM [5] to support broadcast operations. Symbolic image computation for
broadcast assignments is significantly more expensive than image computation for standard assignments. In the context of B OOM it involves 1) converting states from counter
representation to a form where the individual local states of threads are stored using
distinct BDD variables, 2) computing the intersection of n − 1 successor states, one
for each inactive thread paired with the active thread, and 3) transforming the resulting state representation back to counter form using Shannon expansion. The expense of
image computation for broadcasts motivates the careful analysis we have presented in
Sections 2 and 3 for determining tight conditions under which broadcasts are required.
Simulation. To determine the authenticity of abstract error traces reported by B-B OOM
we have extended the S ATA BS simulator. The existing simulator extracts the control
flow from the trace. This is mapped back to the original C program and translated into
a propositional formula (using standard techniques such as single static assignment
conversion and bitvector interpretation of variables). The error is spurious exactly if
this formula is unsatisfiable. In the concurrent case, the control flow information of an
abstract trace includes which thread executes actively in each step. We have extended
the simulator so that each local variable involved in a step is replaced by a fresh indexed
version, indicating the executing thread that owns the variable. The result is a program
trace over the replicated C program Pn , which can be directly checked using a SAT
solver.
As an example, suppose function f from program P′′ (Introduction) is executed by
2 threads (for this example, we ignore the rest of P′′ ). The model checker may return
the error trace shown below on the left, which is converted into the non-threaded form
shown on the right.
T1:
T1:
T2:
T2:
++s,++l;
assert s == l;
++s,++l;
assert s == l;
++s,++l1 ;
assert s == l1 ;
++s,++l2 ;
assert s == l2 ;
The trace on the right is translated into SSA form and finally into the integer arithmetic
formula below, which is shown to be satisfiable, so the error is real.
s0 = 0 ∧ l10 = 0 ∧ l20 = 0 ∧ s1 = s0 + 1 ∧ l11 = l10 + 1
∧ s2 = s1 + 1 ∧ l21 = l20 + 1 ∧ s2 6= l21 ,
Note that broadcast operations do not affect simulation: although, at the Boolean
program level, a broadcast may change the state of multiple threads simultaneously, the
corresponding C program statement is simply an update to a shared variable executed
by a single thread.
Refinement. Our implementation performs refinement by extracting new predicates
from counterexamples via weakest precondition calculations. This standard method requires a small modification in our context: weakest precondition calculations generate
predicates over shared variables, and local variables of specific threads. For example, if
thread 1 branches according to a condition such as l < s, where l and s are local and
shared, respectively, weakest precondition calculations generate the predicate l1 < s,
where l1 is thread 1’s copy of l. Because our predicate abstraction technique works at
the template program level, we cannot add this predicate directly. Instead, we generalize
such predicates by removing thread indices. Hence in the above example, we add the
mixed predicate l < s, for all threads.
An alternative approach is to refine the abstract transition relation associated with
the Cartesian abstraction based on infeasible steps in the abstract counterexample [14].
We do not yet perform such refinement, due to challenge of correctly refining abstract
transitions involving broadcast assignments. This involves some subtle issues which
will require further research to solve.
5 Experimental Results
We evaluate the S ATA BS-based implementation of our techniques using a set of 14
concurrent C programs. We consider benchmarks where threads synchronize via locks
(lock-based), or in a lock-free manner via atomic compare-and-swap (cas) or test-andset (tas) instructions. The benchmarks are as follows:3
– Increment, Inc./Dec. (lock-based and cas-based) A counter, concurrently incremented, or incremented and decremented, by multiple threads [15]
– Prng (lock-based and cas-based) Concurrent pseudorandom number generator [15]
3
All benchmarks and tools are available online: http://www.cprover.org/SAPA
– Stack (lock-based and cas-based) Thread-safe stack implementation, supporting
concurrent pushes and pops, adapted from an Open Source IBM implementation4
of an algorithm described in [15]
– Tas Lock, Ticket Lock (tas-based) Concurrent lock implementations [16]
– FindMax, FindMaxOpt (lock-based and cas-based) Implementations of parallel
reduction operation [17] to find maximum element in array. FindMax is a basic
implementation, and FindMaxOpt and optimized version where threads reduce
communication by computing a partial maximum value locally
Mixed predicates were required for verification to succeed in all but two benchmarks: lock-based Prng, and lock-based Stack. For each benchmark, we consider verification of a simple safety property, specified via an assertion. We have also prepared
a buggy version of each benchmark, where an error is injected into the source code to
make it possible for this assertion to fail. We refer to correct and buggy versions of our
benchmarks as safe and unsafe, respectively.
All experiments are performed on a 3GHz Intel Xeon machine with 40 GB RAM,
running 64-bit Linux, with separate timeouts of 1h for the abstraction and model checking phases of the CEGAR loop. Predicate abstraction uses a maximum cube length of 3
for all examples, and MiniSat 2 (compiled with full optimizations) is used for predicate
abstraction and counterexample simulation.
Symmetry Oblivious
Benchmark n #φ Abs (s) MC (s)
Increment
6 14
13
5
(lock-based) 8 18
29
152
9 20
40
789
10 22
56
T.O.
12 –
–
–
14 –
–
–
16 –
–
–
18 –
–
–
20 –
–
–
Increment
4 26
201
12
(cas-based) 5 32
94
358
6 38
190
T.O.
7 –
–
–
Inc./Dec.
4 26
71
6
(lock-based) 5 31
132
656
6 36
231
T.O.
Inc./Dec.
3 45
372
6
(cas-based) 4 58
872
4043
Tas Lock
4 28
9
114
(tas-based)
5 34
14
T.O.
6 –
–
–
Ticket Lock 2 22
554
1
(tas-based)
3 27 1319
3
4 32
T.O.
–
6 –
–
–
8 –
–
–
Symmetry Aware
#φ Abs (s) MC (s)
4
1
<1
4
1
1
4
1
1
4
1
2
4
1
7
4
1
24
4
1
100
4
1
559
4
1 2882
8
6
1
8
6
13
8
6
116
8
6
997
11
11
2
11
11
50
11
11 1422
20
78
3
20
78
252
8
1
4
8
1
72
8
1
725
19
251
1
19
251
1
19
251
2
19
251
62
19
251 2839
Symmetry Oblivious
Benchmark n #φ Abs (s) MC (s)
Prng
6 57
69
21
(lock-based) 7 68
83
191
8 79
96
T.O.
16 –
–
–
26 –
–
–
Prng
3 32
29
<1
(cas-based)
4 47
40
12
5 62
57
1049
FindMax
6 6
5
30
(lock-based) 7 7
9
244
8 8
14
T.O.
16 –
–
–
25 –
–
–
FindMax
3 18
4
7
(cas-based)
4 24
8
407
FindMaxOpt 4 8
3
40
(lock-based) 5 10
6
1356
6 12
11
T.O.
7 –
–
–
FindMaxOpt 3 21
9
11
(cas-based)
4 28
15
1097
5 35
22
T.O.
Stack
3 16
<1
14
(lock-based) 4 17
<1
945
Stack
3 16
2
29
(cas-based)
4 21
8
3408
Symmetry Aware
#φ Abs (s) MC (s)
13
26
<1
13
26
1
13
26
2
13
26
142
13
26 3023
17
48
1
17
48
38
17
48 1832
1
<1
<1
1
<1
1
1
<1
1
1
<1
125
1
<1
3005
6
1
2
6
1
368
2
<1
3
2
<1
33
2
<1
269
2
<1
1773
7
3
2
7
3
61
7
3 1240
5
<1
8
5
<1
374
6
<1
14
6
<1
813
Table 1. Comparison of symmetry-aware and symmetry-oblivious predicate abstraction over our
benchmarks. For each configuration, the fastest abstraction and model checking times are in bold.
4
http://amino-cbbs.sourceforge.net
Symmetry-aware vs. symmetry-oblivious method. We evaluate the scalability of our
symmetry-aware predicate abstraction technique (SAPA) by comparing it against the
symmetry-oblivious predicate abstraction (SOPA) approach described in Section 1, for
verification of correct versions of our benchmarks. Recall that in SOPA, an n-thread
symmetric concurrent program is expanded so that variables for all threads are explicitly duplicated, and n copies of all non-shared predicate are generated. The expanded
program is then abstracted over the expanded set of predicates, using standard predicate
abstraction. This yields a Boolean program for each thread; the parallel composition of
these n Boolean programs is explored by a model checker. Because symmetry is not
exploited, and no broadcasts are required, any suitable model checker can be used.
We have tried both standard B OOM [5] (without symmetry reduction) and Cadence
SMV [18] to model check expanded Boolean programs. In all cases, we found B OOM
to be faster than SMV, thus we present results only for B OOM.
Table 1 presents the results of the comparison. For each benchmark and each approach we show, for interesting thread counts (including the largest thread count that
could be verified with each approach), the number of predicates required for verification and the elapsed time for predicate abstraction and model checking. For each configuration, the fastest abstraction and model checking times are shown in bold. Model
checking uses standard B OOM, without symmetry reduction (SOPA) and B-B OOM, our
novel extension to B OOM discussed in Section 4 (SAPA), respectively. Entries marked
T.O. indicate that a timeout occurred; succeeding cells are then marked ‘–’.
The results show that our novel SAPA technique significantly outperforms SOPA,
both in terms of abstraction and model checking time. The former can be attributed to
the fact that, with SOPA, the number of predicates grows according to the number of
threads considered, while with SAPA, this is thread count-independent. The latter is due
to the exploitation of template-level symmetry by B-B OOM.
Benchmark
Increment (lock-based)
Incr. (cas-based)
Incr./Dec. (lock-based)
Incr./Dec. (cas-based)
Tas Lock (tas-based)
Ticket Lock (tas-based)
Prng (lock-based)
Prng (cas-based)
FindMax (lock-based)
FindMax (cas-based)
FindMaxOpt (lock-based)
FindMaxOpt (cas-based)
Stack (lock-based)
Stack (cas)
Symmetry-Aware
Safe n Unsafe n
safe >10 unsafe 2
safe
7 unsafe 2
safe
6 unsafe 3
safe
4 unsafe 3
safe
7 unsafe 2
safe
8 unsafe 3
safe >10 unsafe 2
safe
5 unsafe 3
safe >10 unsafe 2
safe
4 unsafe 2
safe
7 unsafe 2
safe
5 unsafe 1
safe
4 unsafe 4
safe
4 unsafe 2
Mixed as local
Safe n Unsafe n
safe >10 error
2
safe
8
safe
5
safe >10
safe >10
safe
6
safe
8
safe
8 error
2
safe >10 unsafe
3
safe >10 unsafe
2
safe
7 unsafe
3
safe >10
safe >10
safe
5
safe
4
safe
7
safe
6
safe
5 unsafe
1
safe
4 unsafe
4
safe
4
safe
6
Mixed as shared
Safe n Unsafe n
safe 10 error 2
error
2 error 2
safe >10 unsafe 3
error
2 error 3
error
2 error 2
safe
5 unsafe 3
safe >10 unsafe 2
safe
6 unsafe 3
safe
2 error 2
safe
2
safe 1
error
2 error 2
error
2 unsafe 1
safe
4 unsafe 4
safe
4 error 2
Table 2. Comparison of sound and unsound approaches; incorrect results in bold.
Comparison with unsound methods. In Section 1, we described two naı̈ve solutions
to the mixed predicate problem: uniformly using local or shared Boolean variables to
represent mixed predicates, and then performing standard predicate abstraction. We de-
note these approaches mixed as local and mixed as shared, respectively. Although we
demonstrated theoretically in Section 1 that both methods are unsound, it is interesting
to see how they perform in practice. Table 2 shows the results of applying CEGARbased model checking to safe and unsafe versions of our benchmarks, using our sound
technique, and the unsound mixed as local and mixed as shared approaches. In all cases,
B-B OOM is used for model checking. For the sound technique, we show the largest
thread count for which we could prove correctness of each safe benchmark, and the
smallest thread count for which a bug was revealed in each unsafe benchmark. The
other columns illustrate how the unsound techniques differ from this, where “error” indicates a refinement failure: it was not possible extract further predicates from spurious
counterexamples. Bold entries indicate cases where the unsound approaches produce
incorrect, or inconclusive results.5 The number of cases where the unsound approaches
produce false negatives, or lead to refinement failure, suggest that little confidence can
be placed in these techniques, even for purposes of falsification. This justifies the more
sophisticated and, crucially, sound techniques developed in this paper.
6 Related Work and Conclusion
There exists a large body of work on the different stages of CEGAR-based program
analysis. We focus here on the abstraction stage, which is at the heart of this paper.
Predicate abstraction goes back to the foundational work by Graf and Saı̈di [1].
It was first presented for sequential programs in a mainstream language (C) by Ball,
Majumdar, Millstein, Rajamani [2] and implemented as part of the S LAM project. We
have found many of the optimizations suggested by [2] to be useful in our implementation as well. Although S LAM has had great success in finding real bugs in system-level
code, we are not aware of any extensions of it to concurrent programs (although this
option is mentioned by the authors). We attribute this to a large part to the infeasibility,
at the time, to handle realistic multi-threaded Boolean programs. We believe our own
work on B OOM [5] has made progress in this direction that has made it attractive again
to address concurrent predicate abstraction.
We are not aware of other work that presents solutions to the problem of “mixed
predicates”. Some approaches avoid it by syntactically disallowing such predicates,
e.g. [19], whose authors don’t discuss, however, the reasons for (or, indeed, the consequences of) doing so. In other work, “algorithmic circumstances” may make the treatment of such predicates unnecessary. The authors of [20], for example, use predicate
abstraction to finitely represent the environment of a thread in multi-threaded programs.
The “environment” consists of assumptions on how threads may manipulate the shared
state of the program, irrespective of their local state. Our case of replicated threads, in
which mixed predicates would constitute a problem, is only briefly mentioned in [20].
In [21], an approach is presented that handles recursive concurrent C programs. The
abstract transition system of a thread (a pushdown system) is formed over predicates
that are projected to the global or the local program variables and thus cannot compare
5
We never expect the unsound techniques to report conclusively that a safe benchmark is unsafe:
this would require demonstrating a concrete error trace in the original, safe, program.
“global against local” directly. As we have discussed, some reachability problems cannot be solved using such restricted predicates. We conjecture this problem is one of the
potential causes of non-termination in the algorithm of [21].
In conclusion, we mention that building a CEGAR-based verification strategy is a
tremendous effort, and our work so far can only be the beginning of such effort. We
have assumed a very strict (and unrealistic) memory model that guarantees atomicity at the statement level. One can work soundly with the former assumption by preprocessing input programs so that the shared state is accessed only via word-length
reads and writes, ensuring that all computation is performed using local variables. Extending our approach to weaker memory models, building on existing work in this
area [22, 23], is future work. Our plans also include a more sophisticated refinement
strategy, and a more detailed comparison with existing approaches that circumvent the
mixed-predicates problem using other means.
References
1. Susanne Graf and Hassen Saı̈di. Construction of abstract state graphs with PVS. In CAV,
1997.
2. Thomas Ball, Rupak Majumdar, Todd D. Millstein, and Sriram K. Rajamani. Automatic
predicate abstraction of C programs. In PLDI, 2001.
3. Edmund Clarke, Orna Grumberg, Somesh Jha, Yuan Lu, and Helmut Veith. Counterexampleguided abstraction refinement for symbolic model checking. J. ACM, 2003.
4. Thomas Ball and Sriram Rajamani. The SLAM project: debugging system software via static
analysis. In POPL, 2002.
5. Gerard Basler, Michele Mazzucchi, Thomas Wahl, and Daniel Kroening. Context-aware
counter abstraction. Formal Methods in System Design (FMSD), 2010.
6. Alexander Kaiser, Daniel Kroening, and Thomas Wahl. Dynamic cutoff detection in parameterized concurrent programs. In CAV, 2010.
7. Salvatore La Torre, Parthasarathy Madhusudan, and Gennaro Parlato. Model-checking parameterized concurrent programs using linear interfaces. In CAV, 2010.
8. Thomas Ball, Andreas Podelski, and Sriram Rajamani. Boolean and Cartesian abstraction
for model checking C programs. STTT, 2003.
9. Radu Rugina and Martin C. Rinard. Pointer analysis for multithreaded programs. In PLDI,
pages 77–90, 1999.
10. J.M. Morris. A general axiom of assignment. In Theoretical Foundations of Programming
Methodology, Lecture Notes of an International Summer School, pages 25–34. D. Reidel
Publishing Company, 1982.
11. Edmund Clarke, Daniel Kroening, Natasha Sharygina, and Karen Yorav. Predicate abstraction of ANSI-C programs using SAT. Formal Methods in System Design (FMSD), 2004.
12. Gerard Basler, Matthew Hague, Daniel Kroening, Luke Ong, Thomas Wahl, and Haoxian
Zhao. Boom: Taking Boolean program model checking one step further. In TACAS, 2010.
13. Byron Cook, Daniel Kroening, and Natasha Sharygina. Verification of Boolean programs
with unbounded thread creation. Theoretical Computer Science (TCS), 2007.
14. Thomas Ball, Byron Cook, Satyaki Das, and Sriram K. Rajamani. Refining approximations
in software predicate abstraction. In TACAS, 2004.
15. Tim Peierls, Brian Goetz, Joshua Bloch, Joseph Bowbeer, Doug Lea, and David Holmes.
Java Concurrency in Practice. Addison-Wesley Professional, 2005.
16. John Mellor-Crummey and Michael Scott. Algorithms for scalable synchronization on
shared-memory multiprocessors. Transactions on Computer Systems (TOCS), 1991.
17. Randy Allen and Ken Kennedy. Optimizing Compilers for Modern Architectures. Morgan
Kaufmann, San Francisco, CA, USA, 2002.
18. Kenneth McMillan. Symbolic Model Checking: An Approach to the State Explosion Problem.
Kluwer Academic, 1993.
19. Nils Timm and Heike Wehrheim. On symmetries and spotlights - verifying parameterised
systems. In ICFEM, 2010.
20. Thomas Henzinger, Ranjit Jhala, Rupak Majumdar, and Shaz Qadeer. Thread-modular abstraction refinement. In CAV, 2003.
21. Sagar Chaki, Edmund Clarke, Nicholas Kidd, Thomas Reps, and Tayssir Touili. Verifying
concurrent message-passing C programs with recursive calls. In TACAS, 2006.
22. J. Lee and D.A. Padua. Hiding relaxed memory consistency with a compiler. IEEE Transactions on Computers, 50:824–833, 2001.
23. J. Alglave. A Shared Memory Poetics. PhD thesis, Université Paris 7 and INRIA, 2010.
A
Boolean Broadcast Programs
Our approach to symmetry-aware predicate abstraction generates a form of concurrent
Boolean program. Although languages for such programs are well-known [13, 5], the
existing ones are not quite suited to our purpose. As discussed in the paper, in order to
handle mixed predicates correctly, we require the facility for a Boolean program thread
to read and update variables of other threads. As this is fundamental to our approach,
we now present syntax and semantics for this variant of concurrent Boolean programs
which we call Boolean broadcast programs.
prog ::= shared name = lit; . . . ; name = lit;
local name = lit; . . . ; name = lit;
1: stmt; . . . ; k: stmt;
lval
stmt ::= lval, . . . , lval = expr, . . . , expr
| goto pc 1 , . . . , pc d (pc j ∈ {1, . . . , k})
| assume expr
lit
::= v | [v]
expr ::= lval | lit | compound expr
::= 0 | 1 | ⋆
name ::= any legal C variable name
Fig. 1. Syntax for Boolean broadcast programs.
Syntax for Boolean broadcast programs is specified by the grammar of Figure 1,
where standard compound expressions are assumed. The language includes standard
features such as shared and local variables, parallel variable assignments, nondeterministic goto statements and assume statements (which together can be used to model
control flow), and the nondeterministic r-value ⋆.
The novel feature of this language is that it supports assignments of the form [v]
= expr, which we call broadcasts. When such a broadcast is executed by thread i, it
causes local variable v to be updated in each passive thread, i.e. in all threads except i.
In this context, expr ranges over shared variables, local variables of the active thread
(thread i), and local variables of the passive thread. The latter are distinguished by the
syntax [v]. We refer to an expression of the form [v] as a passive expression. We refer
to such an expression as a passive l-value or passive r-value depending on which side
of a broadcast it occurs.
We give semantics for parallel assignments in which all passive l-values appear after
non-passive l-values. For parallel assignments where this is not the case, the semantics
are defined to be the same as for any suitable rearrangement that enforces this condition.
In a parallel assignment of the form:
u1 , . . . , ua , [v1 ], . . . , [vb ] = . . .
we require that the ui are mutually distinct, the vi are mutually distinct, and each
vi ∈ VL . We do allow the ui and vj to overlap. We also require that passive r-values
only appear on the right-hand-side of broadcasts (i.e. they cannot be used in standard
assignments).
We now formally define semantics for parallel assignments that may contain broadcasts. Let B be a Boolean broadcast program, and n a positive integer. Let V = VS ∪VL
be the set of shared (VS ) and local (VL ) variable names appearing in B. Define V =
S
VS ∪ 1≤i≤n {li | l ∈ VL }. This is the full set of variables in Bn , an n-thread instantiation of B. A store for Bn is a mapping σ : V → {0, 1}, assigning a Boolean value to
each variable.
For an expression φ, thread indices i and j (1 ≤ i, j ≤ n) and store σ, we write
eval (φ, i, j, σ) ⊆ {0, 1} for the set of possible results which can be obtained by evaluating φ with respect to active thread i and passive thread j in the context of σ. Formally,
for l ∈ VL , s ∈ VS and z ∈ {0, 1}, eval (φ, i, j, σ) is defined for simple expressions as
follows:
eval (z, i, j, σ) = {z}
eval (⋆, i, j, σ) = {0, 1}
eval (l, i, j, σ) = {σ(li )} eval ([l], i, j, σ) = {σ(lj )} eval (s, i, j, σ) = {σ(s)}
and extended to compound expressions in the obvious way.
If φ has no passive subexpressions then it is clear that, for any threads j1 and j2 ,
eval (φ, i, j1 , σ) = eval (φ, i, j2 , σ); the passive thread is irrelevant. In this case we
simply write eval (φ, i, σ).
For a parallel assignment statement assign, store σ and thread id i (1 ≤ i ≤ n),
we write exec(assign, i, σ) for the set of stores that can result from active thread i
executing statement assign in the context of store σ.
Without loss of generality, we can assume that assign has the following form:
s1 , . . . , sa , l1 , . . . , lb , [m1 ], . . . , [mc ] = φ1 , . . . , φa , ψ1 , . . . , ψb , χ1 , . . . , χc ,
where a, b, c ≥ 0, s1 , . . . , sa ∈ VS are mutually distinct shared variables, l1 , . . . , lb ∈
VL are mutually distinct local variables, and m1 , . . . , mc ∈ VL are also mutually distinct local variables (the lj and mj may overlap), φ1 , . . . , φa and ψ1 , . . . , ψb are expressions that do not contain passive subexpressions, and χ1 , . . . , χc are expressions that
may contain passive subexpressions.
With this notation, we define: exec(assign , i, σ) =
σ[ s1 7→ x1 , . . . , sa 7→ xa ,
xf ∈ eval (φf , i, σ),
l1i 7→ y1 , . . . , lbi 7→ yb ,
yg ∈ eval (ψg , i, σ),
1
1
m1 7→ z1,1 , . . . , mc 7→ z1,c , . . .
j
z
∈
eval
(χ
,
i,
j,
σ),
i−1
h
h
m1 7→ zi−1,1 , . . . , mi−1
7→ zi−1,c ,
c
1 ≤ f ≤ a, 1 ≤ g ≤ b,
i+1
i+1
m1 7→ zi+1,1 , . . . , mc 7→ zi+1,c , . . .
1
≤
h
≤
c,
1
≤
j
≤
n,
i
=
6
j
n
n
m1 7→ zn,1 , . . . , mc 7→ zn,c ]
Thus each store in exec(assign, i, σ) is derived from σ by setting each variable sf
and lgi to a value of eval (φf , i, σ) and eval (ψg , i, σ) respectively (1 ≤ f ≤ a, 1 ≤
g ≤ b), and, for each thread j distinct from i, setting variable mjh to a value of
eval (χh , i, j, σ) (1 ≤ h ≤ c).
This definition makes precise the meaning of the [.] notation used in the paper. The
definitions of exec and eval can be used to define the transition system associated with
a Boolean broadcast program in the standard way.
B Proving P′′ Correct Requires Mixed Predicates
Recall program P′′ from Section 1, which, as we show, we cannot prove correct if
executed by 1 thread.
0: shared int r = 0;
P′′ :
shared int s = 0;
local int l = 0;
1: ++r;
2: if (r == 1) then
3:
f ();
f ():
4: ++s, ++l;
5: assert s == l;
6: goto 4;
Let E = E r,s ∪E l for disjoint sets E r,s and E l of predicates over {r, s} and l, resp.;
in particular, no predicate refers to both s and l. Suppose I is an invariant of P′′ expressible over predicates in E such that I ⇒ (s==l) is valid. Since every state satisfying
r==1, s==l is reachable in P′′ , (r==1) ∧ (s==l) ⇒ I is valid. Therefore, for infinitely
many c ∈ N, the assignment r=1, s=l=c satisfies I, written (r, s, l)=(1, c, c) |= I.
Let now {I1 , . . . , Iw } be the cubes in the DNF representation of I. Since this
set is finite, there exist two distinct values a, b and some i ∈ {1, . . . , w} such that
both (r, s, l)=(1, a, a) |= Ii and (r, s, l)=(1, b, b) |= Ii . We split cube Ii into the
sub-cubes Iir,s and Iil that contain the predicates over {r, s} and those over l, resp.:
Ii = Iir,s ∧ Iil . From (r, s, l)=(1, a, a) |= Ii we conclude (r, s)=(1, a) |= Iir,s (l does
not occur in Iir,s ). Similarly, from (r, s, l)=(1, b, b) |= Ii we conclude l=b |= Iil . Hence,
(r, s, l)=(1, a, b) |= Ii , hence (r, s, l)=(1, a, b) |= I, which contradicts the validity of
I ⇒ (s==l) since a 6= b.
2
| 6cs.PL |
1
Bearing Rigidity Theory and its Applications for
Control and Estimation of Network Systems
Life Beyond Distance Rigidity
arXiv:1803.05505v1 [cs.SY] 14 Mar 2018
Shiyu Zhao
Distributed control and estimation of multi-agent systems
has received tremendous research attention in recent years due
to their potential across many application domains [1], [2].
Here, the term “agent” can represent a sensor, an autonomous
vehicle, or any general dynamical system. These multi-agent
systems are becoming increasingly attractive because of their
robustness against system failure, their ability to adapt to
dynamic and uncertain environments, and their economic
advantages compared to the implementation of more expensive
monolithic systems.
Formation control and network localization are two fundamental tasks for multi-agent systems that enable them to
perform complex missions. The goal of formation control is
to control each agent using local information from neighboring
agents so that the entire team forms a desired spatial geometric
pattern (see [2] for a recent survey on formation control).
While the notion of a formation as a geometric pattern has
a natural meaning for robotic systems, it may also correspond
to more abstract configurations for the system state of a team
of agents. The goal of network localization is to estimate the
location of each agent in a network using locally sensed or
communicated information from neighboring agents [3]–[6].
Network localization is usually the first step that must be
completed before a sensor network provides other services like
positioning mobile robots or monitoring areas of interest.
For a formation control or network localization task, the
type of information available to each agent is an important
factor that determines the design of the corresponding control
or estimation algorithms. Most of the existing approaches
for formation control assume that each agent can obtain the
relative positions of their nearest neighbors. In order to obtain
relative positions in practice, each agent can measure their absolute positions using, for example, GPS, and then share their
positions with their neighbors via wireless communications.
This method is, however, not applicable when operating in
GPS-denied environments such as indoors, underwater, or in
deep space. Furthermore, the absolute accuracy of the GPS
may not meet the requirements of high-accuracy formation
control tasks. Rather than relying on external positioning
systems such as GPS, each agent can use onboard sensors
to sense their neighbors.
Optical cameras are widely used onboard sensors for ground
and aerial vehicles to achieve various sensing tasks due to their
characteristics of being low-cost, light-weight, and low-power.
It is notable that optical cameras are inherently bearing-only
sensors. Specifically, once a target has been recognized in an
Daniel Zelazo
image, its bearing relative to the camera can be calculated
immediately from its pixel coordinate based on the pin-hole
camera model [7, Section 3.3]. As a comparison, the range
from the target to the camera is more complicated to obtain
because it requires additional geometric information of the
target and extra estimation algorithms, which may significantly
increase the complexity of the vision sensing system. Although
stereo cameras can be used to estimate the range of a target
by triangulating the bearings of the target [8], the estimation
accuracy degenerates rapidly as the range of the target increases due to the short baseline between the two cameras.
In summary, since it is easy for vision to measure bearings,
but relatively difficult to obtain accurate range information,
vision can be effectively modeled as a bearing-only sensing
approach in multi-agent formation control [9], [10]. In addition
to cameras, other types of sensors such as passive radars,
passive sonars, and sensor arrays are also able to measure
relative bearings [5], [11], [12].
When each agent is only able to access the relative bearings
to their neighbors, two types of strategies can be adopted to
utilize these bearings to achieve formation control or network
localization. The first strategy is to use bearings to estimate
relative positions. This strategy leads to coupled control and
estimation problems whose global stability is difficult to prove
(see, for example, [13]). Moreover, the estimation of relative
positions depends on an observability condition requiring that
the relative motion of each pair of neighboring agents satisfy
certain conditions [14]. Although this observability condition
can be achieved in certain applications, such as bearing-only
circumnavigation [15]–[18], it is difficult to satisfy in general
formation control tasks where all the agents are supposed to
form a target formation with no relative motions among the
agents. This observability condition is not satisfied either in
network localization because all the sensors are stationary. The
second strategy, which is the focus of this article, is to directly
apply bearings in formation control or network localization
without estimating relative positions. This strategy does not
require relative position estimation, but it requires designing
new control and estimation algorithms that only utilize bearing
measurements.
The purpose of this article is to provide a tutorial overview
of recent advances in the area of bearing-based formation
control and network localization. The first problem addressed
in this article is to understand when the formation control
or network localization problems can be solved using only
inter-neighbor bearing measurements. In fact, any distributed
2
control or estimation task requires certain fundamental architectural conditions of the multi-agent system. For example,
in consensus problems, a network must possess a spanning
tree in order to ensure the states of different agents converge
to the same value [19]–[22]. For bearing-based formation
control and network localization, there is also an architectural
requirement to solve these problems - this property is known
as bearing rigidity. The bearing rigidity theory, also called parallel rigidity theory in the literature, was originally introduced
for computer-aided design [23] and has received increasing
attention in recent years due to its important applications in
bearing-based control and estimation problems [24]–[28]. The
bearing rigidity theory studies the fundamental problem of
under what conditions can the geometric pattern of a network
be uniquely determined if the bearing of each edge in the
network is fixed.
The bearing rigidity theory can be interpreted as an analogous theory for the classic rigidity theory based on interneighbor distances, which is referred to as distance rigidity
theory in this article. The classic distance rigidity theory
studies the problem of under what conditions can the geometric pattern of a network be uniquely determined if the
length (distance) of each edge in the network is fixed. It
is a combinatorial theory for characterizing the stiffness or
flexibility of structures formed by rigid bodies connected by
flexible linkages or hinges. The study of distance rigidity has
a long history as a formal mathematical discipline [29]–[36].
In recent years, it has played a fundamental role in distancebased formation control [37]–[45] and distance-based network
localization [4], [5], [46]. One goal of this article is to compare
the distance and bearing rigidity theories by highlighting their
similarities and differences.
This article addresses three important applications of the
bearing rigidity theory in the area of the distributed control and
estimation of multi-agent systems, briefly described below.
(a) Bearing-Based Network Localization: Consider a network
of stationary nodes where only a subset of the nodes
know their own absolute positions - these special nodes
are referred to as the anchors while the other nodes are
called followers. Suppose each follower node is able to
measure the relative bearings of their neighbors and share
the estimates of their own positions with their neighbors
by wireless communication. The aim of bearing-based
network localization is to localize the follower nodes
using the bearing measurements and the anchors’ absolute
positions [6], [47]–[52]. Here the network localization
problem may also be called network self-localization,
which is usually the first step for a sensor network to
provide other services such as positioning or monitoring.
Network localization is essential for sensor networks in
environments where GPS signals are not available, reliable, or sufficiently accurate.
(b) Bearing-Based Formation Control: Consider a group of
mobile agents where each agent is able to obtain the
relative positions of their neighbors. The aim of bearingbased formation control is to steer the agents from some
initial spatial configuration to a target formation with
a desired geometric pattern predefined by inter-neighbor
bearings [24], [53]–[56]. Since the target formation is
invariant to scaling variations, bearing-based formation
control provides a simple solution for formation scale
control, which is a practically useful technique to adjust
the scale of a formation so that the agents can dynamically
respond to the environment to achieve, for example, obstacle avoidance such as passing through narrow passages
[57], [58]. Note that the bearing-based formation control
problem is dual to the bearing-based network localization
problem. When the agent dynamics are modeled as single
integrators and the leaders are stationary, the two problems
are indeed identical. However, this article also considers
a broader range of cases in the formation control problem - namely formation maneuvering using leaders, and
different models for the agent dynamics, including double
integrators and unicycles.
(c) Bearing-Only Formation Control: The aim of bearingonly formation control is to steer a group of mobile
agents to form a desired geometric pattern predefined by
inter-neighbor bearings. Unlike bearing-based formation
control, bearing-only formation control only requires each
agent to measure the relative bearings of their neighbors,
whereas relative positions are not required to be measured
or estimated [10], [25], [59]–[65]. Bearing-only formation
control provides a novel framework for implementing
vision-based formation control tasks where vision may
be modeled as a bearing-only sensing approach. It also
suggests that distance information may be redundant to
achieve certain formation control tasks.
The notations for networks and formations used throughout
this article are given in “Notations for Networks and Formations”.
B EARING R IGIDITY T HEORY
The bearing rigidity theory studies the problem of under
what conditions the geometric pattern of a network can be
uniquely determined if the bearing of each edge in the network
is fixed. Equivalently stated, bearing rigidity studies as under
what conditions do two networks have the same geometric
pattern if they have the same bearings. To illustrate this idea,
the two networks in Figure 1(a) have the same bearings but
different geometric patterns. As a result, they are not bearing
rigid. The two networks in Figure 1(b) have the same bearings
and the same geometric pattern (modulo a scaling and a
translational factor). The two networks can be shown to be
bearing rigid and the rigorous proof of this result relies on the
theory presented in this section.
There are three different notions of bearing rigidity: bearing rigidity, global bearing rigidity, and infinitesimal bearing
rigidity. The first two are not of practical interest because
they cannot ensure unique geometric patterns of networks.
The third, infinitesimal bearing rigidity, is the most important
one whose properties are discussed in detail in this section.
The precise definitions of the three types of bearing rigidity
are given in “Key Definitions in Bearing Rigidity Theory.”
These definitions are analogous to those in the distance rigidity
theory, which are listed in “Key Definitions in Distance
3
(a)
(b)
Fig. 1: Illustration of bearing rigidity. The networks in (a) are not bearing
rigid because the same inter-neighbor bearings may lead to different geometric
patterns of the networks, for example, a square on the left and a rectangle
on the right. The networks in (b) are bearing rigid because the same interneighbor bearings imply the same geometric pattern though the networks may
differ in terms of translation and scale.
Rigidity Theory” for the purpose of comparison. It is worth
noting that an orthogonal projection matrix plays a key role
in the bearing rigidity theory. The properties of the projection
matrix are summarized in “An Orthogonal Projection Matrix.”
Moreover, note that a bearing, which is represented by a unit
vector, must be expressed in a specific reference frame. In
this article, the bearings in a network are all expressed in a
common reference frame.
Properties of Infinitesimal Bearing Rigidity
Infinitesimal bearing rigidity has two key properties. The
first is a geometric property [28, Theorem 6] that the positions
of the nodes in a network can be uniquely determined up
to a translational and scaling factor by the bearings if and
only if the network is infinitesimally bearing rigid. The second
is an algebraic property [28, Theorem 4] that a network is
infinitesimally bearing rigid in d-dimensional space if and only
if the bearing rigidity matrix RB satisfies
Null(RB ) = span{1n ⊗ Id , p},
(1)
or equivalently,
rank(RB ) = dn − d − 1.
(2)
The definition of the bearing rigidity matrix RB is given in
“Key Definitions in Bearing Rigidity Theory.” Due to the
above two properties, infinitesimal bearing rigidity not only
ensures the unique geometric pattern of a network, but also
can be conveniently examined by a mathematical condition.
Examples of infinitesimally bearing rigid networks are given
in Figure 2.
The notion of infinitesimal bearing rigidity is defined based
on the bearing rigidity matrix. The term “infinitesimal” is
due to the fact that the bearing rigidity matrix is the firstorder derivative (the Jacobian) of the bearing vectors with
respect to the positions of the nodes. It must be noted that
infinitesimal bearing rigidity is a global property in the sense
that the bearings can uniquely determine the geometric pattern
of a network. The term “infinitesimal” may be dropped in this
article when the context is clear.
An infinitesimal bearing motion of a network is a motion of
some nodes that preserves all the bearings. All the infinitesimal
bearing motions of a network form the null space of the bearing rigidity matrix. There are two types of trivial infinitesimal
bearing motions: translational and scaling motions of the entire
network. These two types of trivial motions corresponds to the
(a)
(b)
(c)
(d)
Fig. 2: Examples of infinitesimally bearing rigid networks. The networks in
(a) and (b) are two-dimensional and the networks in (c) and (d) are threedimensional. It can be verified that each of these networks satisfies rank(B) =
dn − d − 1. The networks in (a), (b), and (c) also satisfy the Laman condition
and can therefore be generated using a Henneberg construction. Note that the
two networks in (c) and (d) are infinitesimal bearing rigid but not infinitesimal
distance rigid.
(a)
(b)
(c)
(d)
Fig. 3: Examples of non-infinitesimally bearing rigid networks. The red and
solid arrows represent nontrivial infinitesimal bearing motions that preserves
all the inter-neighbor bearings. These networks are not infinitesimally distance
rigid either because they have nontrivial infinitesimal distance motions (see
the blue/dotted arrows). Note that the infinitesimal distance motions are
perpendicular to the infinitesimal bearing motions.
vectors in span{1n ⊗ Id , p}. As a result, the rank condition
in (1) means that a network is infinitesimally bearing rigid
if and only if all infinitesimal bearing motions are trivial.
This provides an intuitive way to examine bearing rigidity.
For example, the networks in Figure 3 are not bearing rigid
because they have non-trivial infinitesimal bearing motions.
An alternative necessary and sufficient condition for infinitesimal bearing rigidity is based on a special matrix termed
the bearing Laplacian [66]. The bearing Laplacian of a
network can be viewed as a weighted graph Laplacian matrix
with weights that are matrices [67]; thus the bearing Laplacian
not only describes the topological structure of the network,
but also the values of the edge bearings. The definition and
properties of bearing Laplacian are summarized in “Bearing
Laplacian of Networks.” For a network with an undirected
graph, the bearing Laplacian has the same rank and null space
as the bearing rigidity matrix [66, Lemma 2]. It then follows
from (1) and (2) that a network is infinitesimally bearing rigid
if and only if
Null(B) = span{1n ⊗ Id , p},
or equivalently,
rank(B) = dn − d − 1.
Compared to the bearing rigidity matrix, the bearing Laplacian
is more convenient to use because it is symmetric and positive
semi-definite for undirected graphs. When the underlying
graph is directed, the bearing Laplacian and the bearing
rigidity matrix may have different ranks and null spaces [68,
Theorem 4].
4
3
3
1
4
1
1
2
Step 1: vertex addition
3
z
5
2
Step 2: edge splitting
z
4
4
2
4
Step 3: edge splitting
3
1
y
2
x
3
1
4
5
3
1
4
5
2
6
Step 4: edge splitting
3
1
2
7
6
Step 5: edge splitting
4
5
8
7
2
6
Step 6: edge splitting
Fig. 4: Illustration of the Henneberg construction procedure. The Henneberg
construction consists of two basic operations: vertex addition and edge
splitting. In this example, the procedure is used to generate an infinitesimally
bearing rigid network in a three-dimensional ambient space. At each step, the
underlying graph of the network is Laman.
Construction of Infinitesimally Bearing Rigid Networks
The previous discussion provided an overview of the properties defining a bearing rigid network. It is also of interest
to explore how to construct a bearing rigid network by
adding well-placed edges and nodes in a network. Although
a network is jointly characterized by its underlying graph and
the configuration of the nodes, the infinitesimal bearing rigidity
of a network is primarily determined by the underlying graph
rather than its configuration [69, Lemma 2]. Given a graph, if
there exists at least one configuration such that the network is
infinitesimally bearing rigid, then for almost all configurations
the corresponding networks are infinitesimally bearing rigid.
Such graphs are called generically bearing rigid [69]. If a
graph is not generically bearing rigid, then the corresponding
network is not infinitesimal bearing rigid for any configurations. As a result, the key to construct infinitesimally bearing
rigid networks is to construct generically bearing rigid graphs.
One of the most well-known methods for rigid graph construction is the Henneberg construction, originally proposed
for the distance rigidity theory [34]. A Henneberg construction
starting from an edge connecting two vertices results in a
Laman graph [70]. For a tutorial on Laman graphs and
Henneberg construction, see “Laman Graphs and Henneberg
Construction.”
In the bearing rigidity theory, the main result about Laman
graphs is that all Laman graphs are generically bearing rigid
in arbitrary dimensions [69, Theorem 1]. That means if the
underlying graph of a network is Laman, then the network
is infinitesimally bearing rigid for almost all configurations
in an arbitrary dimension. Figure 4 illustrates the Henneberg
construction procedure for a three-dimensional infinitesimally
bearing rigid network whose underlying graph is Laman. Note
that the Laman condition is merely sufficient but not necessary
for generic bearing rigidity. A counterexample is given in
Figure 5, where the graph is generically bearing rigid but
not Laman. However, for networks in the plane, a graph
is generically bearing rigid if and only if it is Laman [69,
Theorem 2].
Since a Laman graph has 2n − 3 edges where n is the
3
1
y
2
x
(a)
(b)
Fig. 5: Example of generically bearing rigid graphs that are not Laman. The
configuration (a) is in the x–y plane and the network is not bearing rigid.
The configuration (b) is three-dimensional and the network is bearing rigid.
It can be verified that rank(B) = dn − d − 1 for the configuration in (b).
number of nodes, 2n − 3 edges are sufficient to guarantee
the bearing rigidity of a network in an arbitrary dimension.
For example, every network in Figure 4 is bearing rigid in
the three dimensional space and has 2n − 3 edges. It must
be noted that 2n − 3 is not the minimum number of edges
required to ensure bearing rigidity. The counterexample given
in Figure 5 shows that a graph with less than 2n−3 edges may
be generically bearing rigid in three dimensions. It is still an
open problem to construct all generically bearing rigid graphs
up to now. A comparison between the bearing and distance
rigidity theories is given in “Comparison of Bearing Rigidity
and Distance Rigidity.”
B EARING -BASED N ETWORK L OCALIZATION
This section introduces the theory of bearing-based network
localization that addresses two fundamental problems. The first
problem is localizability, which describes whether or not a
network can possibly be localized. The second problem is how
to localize a network in a distributed manner if it is localizable.
Consider a network of nodes where the first na nodes are
anchors and the remaining nf (nf = n − na ) nodes are
followers. Let Va = {1, . . . , na } and Vf = V \ Va be the sets
of anchors and followers, respectively. The true positions of
the leaders and followers are denoted as pa = [pT1 , . . . , pTna ]T
and pf = [pTn−na , . . . , pTn ]T , respectively. The aim of network
localization is to determine the positions of the followers
{pi }i∈Vf using the edge bearings {gij }(i,j)∈E and the positions of the anchors {pi }i∈Va . All the inter-neighbor bearings
are expressed in a common reference frame.
Bearing-Based Localizability
Localizing the follower nodes is to solve for p̂i , the estimate
of pi , for all i ∈ Vf , obtained from the set of nonlinear
equations,
p̂j − p̂i = g , (i, j) ∈ E,
ij
(3)
kp̂j − p̂i k
p̂i = pi ,
i ∈ Va .
The true location p of the network is a feasible solution to
(3). However, there may exist an infinite number of other
feasible solutions. This leads to the definition of localizability.
A network (G, p) is called bearing localizable if the true
position p is the unique feasible solution to (3). It can be
further shown that p is the unique solution to (3) if and only if
5
anchor
follower
anchor
follower
(a)
(d)
(b)
(e)
(a) (G, p)
(c)
(f)
(g)
Fig. 6: Examples of bearing localizable networks. The networks are localizable
because Bf f of each network is nonsingular. The intuitive interpretation is
that every infinitesimal bearing motion involves at least one anchor. Note
that the networks in (b)-(f) are not infinitesimally bearing rigid but they are
localizable.
anchor
follower
(a)
Fig. 8: Illustration of an augmented network for the localization problem. The
augmented network (Ḡ, p) in (b) is obtained from (G, p) by connecting each
pair of anchors in (G, p). Since deleting or adding the edge between any pair
of anchors only changes Baa but not Bf f , (G, p) and (Ḡ, p) have exactly
the same Bf f , and hence they have the same localizability properties.
to examine network localizability (see, for example, Figure 7).
The following condition indicates how many anchors are
required to guarantee bearing localizability. The number of
anchors in a bearing localizable network in Rd must satisfy
[66, Corollary 1]
na ≥
(b)
(c)
(d)
Fig. 7: Examples of networks that not bearing localizable. The black solid
dots represent the anchors and the white dots for followers. The networks
are not localizable because Bf f of each network is singular. The intuitive
interpretation is that the networks have infinitesimal bearing motions that only
correspond to the followers (see the red arrows).
p is the unique global minimizer of the least-squares problem
[66, Lemma 1]
1X X
min J(p̂) =
kPgij (p̂i − p̂j )k2 = p̂T B p̂, (4)
2
p̂∈Rdn
i∈V j∈Ni
subject to p̂i = pi for i ∈ Va . It has been proven that
p is the unique minimizer of (4) if and only if the matrix
Bf f is nonsingular [66, Theorem 1]. The definition of Bf f
is given in “Bearing Laplacian of Networks.” When Bf f is
nonsingular, the positions of the followers can be solved as
p̂∗f = −Bf−1
f Bf a pa . Examples of bearing localizable and nonlocalizable networks are given in Figures 6 and 7, respectively.
While the nonsingularity of Bf f is an algebraic condition
for bearing localizability, it does not give any intuition on
what a bearing localizable network looks like. The following
conditions can provide more intuition for bearing localizable
networks. First of all, a necessary and sufficient rigidity
condition for bearing localizability is that every infinitesimal
bearing motion of a network must involve at least one anchor
[66, Theorem 2]. More specifically, if there exists a nonzero
infinitesimal bearing motion for a network, there would exist
different networks having exactly the same bearings as the true
network. As a result, infinitesimal bearing motions introduce
ambiguities to the localization of the true network. When
the infinitesimal motion involves at least one anchor, the
ambiguities can be resolved by the anchors whose positions
are known, and hence the network location can be uniquely
determined. This rigidity condition provides an intuitive way
(b) (Ḡ, p)
d+1
dim (Null(B))
≥
.
d
d
(5)
Inequality (5) has two important implications. The first is
that every bearing localizable network must have at least two
anchors because (d + 1)/d > 1. The second is that more
anchors are required when the “degree of bearing rigidity” of
the network is weak. Here, the “degree of bearing rigidity”,
characterized by dim(Null(B)), is strongest if dim(Null(B))
reaches the smallest value d+1 (when the network is infinitesimally bearing rigid) and weak if its value is greater than d+1.
The following two conditions explicitly address the relation
between bearing localizability and bearing rigidity. (i) A
sufficient condition for a network to be bearing localizable
is that it is infinitesimally bearing rigid and has at least two
anchors [66, Corollary 3]. The intuition behind this condition
is as follows. If a network is infinitesimally bearing rigid, then
it can be uniquely determined up to a translation and a scaling
factor. If there are at least two anchors, the translational and
scaling ambiguity can be eliminated by the anchors and thus
the entire network can be fully determined. It must be noted
that infinitesimal bearing rigidity is merely sufficient but not
necessary for bearing localizability. For example, the networks
in Figure 6(b)-(f) are bearing localizable but not infinitesimally
bearing rigid. (ii) Let (Ḡ, p) be the augmented network of
(G, p) which is obtained from (G, p) by connecting each
pair of anchors (see Figure 8 for illustration). Then, another
sufficient condition for bearing localizability is that network
(G, p) is bearing localizable if the augmented network (Ḡ, p) is
infinitesimally bearing rigid and there are at least two anchors
[66, Corollary 2]. This condition is more relaxed in the sense
that it does not require (G, p) to be infinitesimally bearing
rigid. When there are more than two anchors, the infinitesimal
bearing rigidity of (Ḡ, p) is merely sufficient but not necessary
for the bearing localizability of (G, p) (see Figure 6(f) for a
counterexample where the network is bearing localizable but
the augmented network is not infinitesimally bearing rigid).
When there are exactly two anchors, the infinitesimal bearing
rigidity of (Ḡ, p) is both necessary and sufficient for the
bearing localizability of (G, p) [66, Theorem 3].
6
Anchor
Follower: initial estimate
Follower: final estiamte
p̂j (t)
−Pgij (p̂i (t) − p̂j (t))
120
100
gij
z (m)
80
p̂i (t)
Fig. 9: The geometric interpretation of the bearing-based control law in (6).
The term Pgij (p̂j − p̂i ) is perpendicular to gij and it aims to steer agent i
such that ĝij (t) aligns with gij .
60
40
20
120100
80
Distributed Localization Protocols
j∈Ni
T
where Pgij = Id −gij gij
. Protocol (6) is actually the gradientdescent protocol for the objective function in the least-squares
problem (4). The geometric interpretation of this protocol
is illustrated in Figure 9. The expression of protocol (6) is
similar to the well-known linear consensus protocols [19],
[21]. The difference is that the weight for each edge in (6)
is an orthogonal projection matrix, while in the consensus
protocols, the weight for each edge is a scalar. This important
distinction leads to very different properties of the dynamical
system. The unique structure of the projection matrix is the
key feature that enables (6) to solve the bearing-based network
localization problem.
The compact matrix form of (6) is
p̂˙f (t) = −Bf f p̂f (t) − Bf a pa ,
where B is the bearing Laplacian of the true network. This
protocol can globally localize the network if and only if the
network is bearing localizable (that is Bf f is nonsingular)
[66, Theorem 4]. Figure 10 shows a simulation example to
demonstrate protocol (6). The impact of measurement noise
on bearing-based network localization has been discussed in
[66].
B EARING -BASED F ORMATION C ONTROL
This section introduces the theory of bearing-based formation control, which studies how to steer a group of agents to
achieve a bearing-constrained target formation using relative
position measurements. In particular, consider a group of
mobile agents where the first n` agents are leaders and
the remaining nf (nf = n − n` ) agents are followers.
Let V` = {1, . . . , n` } and Vf = V \ V` be the sets of
leaders and followers, respectively. The positions of the leaders and followers are denoted as p` = [pT1 , . . . , pTn` ]T and
pf = [pTn−n` , . . . , pTn ]T , respectively. The target formation is
∗
specified by the constant bearing constraints {gij
}(i,j)∈E and
the leader positions {pi (t)}i∈V` . The control objective is to
120 100 80
20
60
40
20
y (m)
x (m)
(a) Initial and final estimates.
200
Localization error (m)
If a network is bearing localizable, a question that follows is
how to localize it in a distributed manner. Suppose each node
has an initial guess of its own position as p̂i (0). The objective
is to design a distributed protocol to drive p̂i (t) → pi for all
i ∈ Vf as t → ∞. This objective can be achieved by the
protocol [66]
X
p̂˙i (t) = −
Pgij (p̂i (t) − p̂j (t)), i ∈ Vf ,
(6)
60 40
150
100
50
0
0
10
20
30
40
50
Time (sec)
(b) Localization error kp̂i (t) − pi k.
Fig. 10: Simulation example to demonstrate the localization protocol in (6).
The real network is located on a three-dimensional surface. It consists of 210
edges and 64 nodes, four of which are anchors. The network is infinitesimally
bearing rigid because rank(B) = 188 = dn − d − 1. Therefore, the network
is localizable since there are more than two anchors. As can be seen, given a
random initial guess, the localization error of each node converges to zero.
control the positions of the followers {pi (t)}i∈Vf such that
∗
gij (t) → gij
as t → ∞ for all (i, j) ∈ E. All the bearings are
expressed in a common reference frame.
Bearing-Based Formation Control of Single Integrators
First, consider the case where the dynamics of each mobile
agent can be modeled as the single integrator
ṗi (t) = ui (t),
where ui (t) is the velocity input to be designed. If the leaders
are stationary, the bearing-based formation control problem
can be solved by [54]
X
∗ (pi (t) − pj (t)),
ṗi (t) = −
Pgij
i ∈ Vf ,
(7)
j∈Ni
∗
∗ T
∗ = Id − g (g ) . The matrix form of the control
where Pgij
ij ij
law is
ṗf (t) = −Bf f pf (t) − Bf ` p` ,
where B is the bearing Laplacian of the target formation.
Control law (7) can globally stabilize a target formation if and
only if the target formation is bearing localizable (that is Bf f
is nonsingular) [54]. Note that control law (7) has a similar
expression to the network localization protocol in (6). In fact,
the bearing-based formation control problem is mathematically
7
j∈Ni
Z
− kI
t
i
(pi (τ ) − pj (τ ))dτ ,
i ∈ Vf , (8)
Leader
Follower
10
z (m)
equivalent to the bearing-based network localization problem
when the target formation is stationary and each agent is a
single integrator.
If the leaders move at a constant nonzero speed, control law
(7) would yield a constant nonzero tracking error. The tracking
error may be eliminated by using the following proportionalintegral control law proposed in [71],
h
X
∗
kp (pi (t) − pj (t))
Pgij
ṗi (t) = −
0
-10
-10
0
10
x (m)
20
0
80
60
40
y (m)
(a) Generated formation maneuver trajectory (the dark area represents an
obstacle).
0
kp=1, kv=3
where kp and kI are constant positive control gains. The target
formation is globally stable under the action of control law (8)
if and only if it is bearing localizable [71].
If the leader velocities are time-varying, control law (8)
would fail to ensure zero tracking errors. The time-varying
case can be handled by the following control law that requires
velocity feedback:
X
∗ [kp (pi (t) − pj (t)) − ṗj (t)] , i ∈ Vf ,
ṗi (t) = −Ki−1
Pgij
j∈Ni
(9)
P
∗ . The stability of control law (9) can
where Ki = j∈Ni Pgij
be proven as below. First, the nonsingularity of Ki can be
guaranteed by the bearing localizability of the target formation
[55, Lemma 3]. Second, multiplying
P Ki on both sides of (9)
∗ (pi (t) − pj (t))
yields ε̇i = −kp εi where εi = kp j∈Ni Pgij
for i ∈ Vf . It follows that εi → 0 as t → ∞ for all i ∈
∗
Vf , and consequently gij → gij
when the network is bearing
localizable.
Under the action of the control laws (8) and (9), the formation is able to perform translational and scaling formation
maneuvers. A translational maneuver means that all the agents
move at a common velocity such that the formation translates
as a rigid body. A scaling maneuver means that the scale
of the formation, which can be described by the distance
from each agent to the formation centroid, varies while the
geometric pattern of the formation is preserved. In order to
achieve the scaling maneuver, the leaders only needs to adjust
the distances among them. One merit of the bearing-based
control laws is that the desired maneuver is only known to the
leaders and the followers are not required to access or estimate
it.
Bearing-Based Formation Control of Double Integrators
Consider the case where the dynamics of each mobile agent
can be modeled as a double integrator
ṗi (t) = vi (t),
v̇i (t) = ui (t),
where ui (t) is the acceleration input to be designed. If
the velocities of the leaders are constant, the bearing-based
Bearing error
8
6
4
2
0
0
10
20
30
40
50
Time (sec)
(b) Total bearing error of the trajectory,
P
60
(i,j)∈E
70
80
∗ k.
kgij (t) − gij
Fig. 11: Simulation example to demonstrate the bearing formation maneuvering control law in (11). The target formation in the example is a threedimensional cube with two leaders and six followers. The translation and
scale of the formation can continuously vary while the formation pattern is
maintained as desired. This example demonstrates that formation scale control
can be used for obstacle avoidance such as passing through narrow passages.
formation control problem can be solved by [55]
ṗi (t) = vi (t),
h
i
X
∗
v̇i (t) = −
Pgij
kp (pi (t) − pj (t)) + kv (vi (t) − vj (t)) ,
j∈Ni
(10)
where i ∈ Vf and kp , kv are positive constant control gains.
Under control law (10), the target formation is globally stable
if it is bearing localizable.
If the velocities of the leaders are time-varying, the following control law requiring acceleration feedback can be used to
track time-varying target formations [55],
ṗi (t) = vi (t),
v̇i (t) = Ki−1
X
h
∗
Pgij
− kp (pi (t) − pj (t))
j∈Ni
i
− kv (vi (t) − vj (t)) + v̇j (t) ,
(11)
P
∗ . The nonsingularity of
where i ∈ Vf and Ki = j∈Ni Pgij
Ki for any i ∈ Vf is guaranteed by the bearing localizability
of the target formation [55, Lemma 3]. Under control law
(11), the target formation is globally stable if and only if
it is bearing localizable. A simulation example is given in
Figure 11 to demonstrate control law (11). In practice, absolute
acceleration can be measured by each agent using, for example, inertial measurement units, and then transmitted to their
neighbors by wireless communication. Due to measurement
errors and transmission delays, the acceleration measurement
is corrupted by errors. However, since the system is linear,
8
15
pj (t)
2
2
1
1
gij (t)
10
∗
−Pgij (t) gij
y (m)
5
3
4
-5
∗
Pgij (t) gij
4
-10
-15
∗
gij
pi (t)
3
0
-10
-5
0
x (m)
5
10
15
20
Fig. 12: Simulation example to demonstrate the control law in (12). In this
example, there are four unicycle agents whose initial positions and heading
angles are chosen randomly. As can be seen, the formation converges to the
target formation whose square geometric pattern is defined by five bearing
vectors.
bounded acceleration errors would cause bounded tracking errors. Bearing-based formation control in the presence of some
other problems including input disturbance, input saturation,
and collision avoidance have been addressed in [55].
Fig. 13: The geometric interpretation of the bearing-only control law in (13).
∗ is perpendicular to the bearing g , the
Since the control term −Pgij gij
ij
control law aims to reduce the bearing error of gij (t) while maintaining the
distance between agents i and j.
Intial formation
Final formation
1
2
2
1
Bearing-Based Formation Control of Unicycles
Suppose the dynamics of agent i ∈ V can be described by
the unicycle model
ẋi = vi cos θi ,
ẏi = vi sin θi ,
θ̇i = wi ,
where pi = [xi , yi ] ∈ R2 is the coordinate of agent i, θi ∈ S 1
is the heading angle, and vi ∈ R and wi ∈ R are the linear
and angular velocities to be designed. Here, S 1 is the onedimensional manifold on the unit circle. The bearing-based
formation control law for unicycles is [72]
X
∗ (pj (t) − pi (t)),
vi = [cos θi sin θi ]
Pgij
Fig. 14: Simulation example to demonstrate the bearing-only formation control
law in (13). In this example, the formation has two agents and one edge. In the
∗ =
target formation, the bearings are in the horizontal direction; that is g12
∗ = [1, 0]T . The initial formation (the dotted line in the figure) does
−g21
not fulfil the desired bearings. Under the control law in (13), the formation
converges to the desired one (the solid line in the figure). Note that the velocity
of each agent is always perpendicular to the bearing and hence the two agents
move on a circle centered at their midpoint. As a result, the centroid and scale
of the formation are invariant.
T
j∈Ni
wi = [− sin θi cos θi ]
X
∗ (pj (t) − pi (t)).
Pgij
(12)
j∈Ni
When there are no leaders, control law (12) ensures global
∗
∗
stability in the sense that gij (t) converges to either gij
or −gij
as t → ∞ given any initial values of pi (0) and θi (0) if the
target formation is infinitesimally bearing rigid [72]. The final
value of θi is not specified in the control law. A simulation
example is shown in Figure 12.
B EARING -O NLY F ORMATION C ONTROL
This section introduces the theory of bearing-only formation
control, which studies how to steer a group of agents to
achieve a bearing-constrained target formation using bearingonly measurements. Suppose the target formation is specified
∗
by constant bearing constraints {gij
}(i,j)∈E , and there are no
leaders. The control objective is to control the positions of the
∗
agents {pi (t)}i∈V such that gij (t) → gij
for all (i, j) ∈ E as
t → ∞. All the bearings are expressed in a common reference
frame.
The following nonlinear control law, proposed in [28], can
be used to solve the bearing-only formation control problem,
X
∗
ṗi (t) = −
Pgij (t) gij
, i ∈ V,
(13)
j∈Ni
T
where Pgij (t) = Id − gij (t)gij
(t). The geometric interpretation of the control law is illustrated in Figure 13. Some
properties of the control law are highlighted below. First,
the control of each agent only requires bearing measurements
and does not require distance or position estimation. Second,
the control input of (13) is always bounded as kṗi (t)k ≤
P
∗
∗
j∈Ni kPgij (t) kkgij k = |Ni |, since kPgij (t) k = kgij k = 1.
Third, the centroid and scale of the formation are invariant
under the control law [28, Theorem 9]. Here, the centroid is
defined as the average position of the agents and the scale
is defined as the standard deviation of the distances from
the agents to the centroid. Simulation examples are given in
Figures 14 and 15 to demonstrate control law (13).
System (13) is nonlinear and almost globally stable if the
target formation is infinitesimally bearing rigid [28, Theorem 11]. The term “almost” is due to the fact that there are
two isolated equilibriums of the error dynamics, one is desired
9
An optimization-based approach for bearing-only formation
control can be found in [10], [65], where a bearing-only
control law is proposed as
X
∗
(gij (t) − gij
ṗi (t) =
), i ∈ V.
(15)
Initial position
Final position
10
8
j∈Ni
6
This is a gradient-descent control law with the corresponding
objective function as
1 X
1 X
∗ 2
T ∗
keij kkgij − gij
k =
keij k(1 − gij
gij ).
φ2 =
4
2
4
2
0
10
5
0
5
0
10
(a) Initial configuration (grey circle) and final desired
formation (blue circles).
2
(i,j)∈E
(i,j)∈E
∗
Since φ2 contains keij k, φ2 is zero when gij = gij
or eij =
0. As a result, the scale of the formation always decreases
under the action of control law (15). Simulation shows that
this control law may steer all the agents to the same position
given certain initial conditions. To avoid this problem, leaders
must be introduced [65].
Bearing error
1.5
C ONCLUSIONS AND F UTURE D IRECTIONS
1
0.5
0
0
10
20
30
40
50
Time (sec)
(b) Plot of the bearing error,
P
(i,j)∈E
∗ k.
kgij (t) − gij
Fig. 15: Simulation example for the bearing control law in (13) in threedimensional space. In this example, the formation has 27 nodes and 62 edges.
For the target formation, the rank of the bearing rigidity matrix, which equals
the rank of the bearing Laplacian matrix, is 3n − 4 = 77. As a result, the
target formation is infinitesimally bearing rigid and hence the control law (13)
is almost globally stable. As can be seen, given a random initial configuration,
the target formation is achieved and the bearing errors converge to zero.
and the other is undesired. At the desired equilibrium, the
∗
bearings are equal to the desired values; that is gij = gij
for (i, j) ∈ E. At the undesired equilibrium, the bearings
∗
are opposite to the desired values; that is gij = −gij
for
(i, j) ∈ E. The formations at the two equilibriums have the
same centroid and scale but opposite bearings. The almost
global stability means that the formation would converge to
the desired equilibrium unless the initial formation lies exactly
on the undesired equilibrium, which can be shown to be an
unstable equilibrium.
Control law (13) is a modified gradient-descent control law.
In particular, consider the following objective function,
X
1 X
∗ 2
T ∗
kgij − gij
k =
(1 − gij
gij ).
φ1 =
2
(i,j)∈E
(i,j)∈E
∗
The objective function is equal to zero if and only if gij = gij
for all (i, j) ∈ E. The corresponding gradient-descent control
law is
X
1
∗
Pg (t) gij
, i ∈ V.
(14)
ṗi (t) = −
keij (t)k ij
j∈Ni
The two-dimensional version of control law (14) was first
proposed in [24]. This control law requires both bearing and
distance measurements. Removing the distance term keij (t)k
in (14) yields the bearing-only formation control law in (13).
This article presented a review of the bearing rigidity
theory and its applications in distributed formation control
and network localization for multi-agent systems. Motivated
by the fact that many existing approaches rely on measurement
assumptions that may be difficult to realize under certain
circumstances, this article demonstrated how to utilize bearingonly sensors, such as cameras or sensor arrays, to solve the
problems of formation control and network localization. The
article discussed three specific problems including bearingbased network localization, bearing-based formation control,
and bearing-only formation control.
As a newly emerged research area, bearing-based control and estimation is far from being fully explored. Many
important problems in this area remain unsolved. One key
assumption for the results presented in this article is that
the underlying graph is undirected, which means any pair of
neighbors must be able to access each other’s information.
Since this assumption may not be valid in some practical tasks,
it is important to study the case of directed graphs. When the
graph is directed, the control and estimation problem would
become more complicated because undesired equilibriums
may emerge, as observed in [68]. Similar problems also exist
in distance-based formation control [73]–[75]. Despite the
resent progress on bearing-only formation control for some
special directed graphs [76], [77], the problem for general
directed graphs remains an important challenge in this area.
Another key assumption for the results addressed in this
article is that all bearings must be measured in a global
reference frame. Global reference frames, however, may not be
accessible to each agent in some environments such as indoors.
It is important to study how to achieve control or estimation
when bearings are measured in each agent’s local reference
frames. One potential approach is to estimate or synchronize
the orientations of the local reference frames [10], [28]. This
approach has been applied to adapt the bearing-only formation
control law in (13) to use locally measured bearings [28,
Section IV], and a simulation example is given in Figure 16.
This is also a general approach for many types of formation
10
2
2
1.5
1.5
1
1
0.5
0.5
0
0
-0.5
-0.5
-1
-1
2
2
1
-2
-1.5
-1
-0.5
0
0
0.5
1
1.5
2
2.5
(a) Initial formation.
-1
1
-2
-1.5
-1
-0.5
0
0
0.5
1
1.5
2
2.5
-1
(b) Final formation.
Fig. 16: Simulation example for bearing-only formation control without a
global reference frame. The control law is given in [28, Equation (19)]. In
this example, the formation has 8 nodes and 13 edges. The target formation
is a three-dimensional cube that is infinitesimally bearing rigid. The control
is based on inter-neighbor bearings expressed in each agent’s local reference
frames. The orientations of the agents are synchronized in the final formation.
control and network localization tasks in the absence of global
reference frames [78], [79]. However, distributed orientation
estimation or synchronization requires each agent to obtain
their neighbors’ relative orientations, which are usually difficult to measure in practice. Other potential approaches that do
not require an orientation estimation may be based on bearing
rigidity in the special Euclidean group SE(n) [26], [80]–
[83] or complex Laplacian [52], [57]. A brief introduction to
bearing rigidity in SE(2) is given in “Bearing Rigidity Theory
for SE(2).” Nevertheless, the formation control strategies
provided for SE(2) frameworks still require additional sensing
[80], and a complete theory for bearing-only formation control
is still unsolved.
In addition to network localization and formation control,
many other tasks may also be achieved with bearing-only
measurements such as bearing-only rendezvous [84]–[88], and
bearing-only target tracking [16]–[18], [89]–[91] though the
analysis of these tasks may not rely on the bearing rigidity
theory.
The bearing rigidity theory and its application for formation control and network localization is strongly motivated
by the sensing mediums available to distributed and multiagent systems. This work contributed to a broader theory of
cooperative control and estimation for networked systems and
hopes to serve as a starting point for both practitioners and
theoreticians in this community.
11
R EFERENCES
[1] Y. Cao, W. Yu, W. Ren, and G. Chen, “An overview of recent progress
in the study of distributed multi-agent coordination,” IEEE Transactions
on Industrial Informatics, vol. 9, no. 1, pp. 427–438, 2013.
[2] K.-K. Oh, M.-C. Park, and H.-S. Ahn, “A survey of multi-agent
formation control,” Automatica, vol. 53, pp. 424–440, March 2015.
[3] I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci, “A survey
on sensor networks,” IEEE Communications Magazine, vol. 40, pp. 102–
114, Aug 2002.
[4] J. Aspnes, T. Eren, D. K. Goldenberg, A. S. Morse, W. Whiteley, Y. R.
Yang, B. D. O. Anderson, and P. N. Belhumeur, “A theory of network
localization,” IEEE Transactions on Mobile Computing, vol. 12, no. 5,
pp. 1663–1678, 2006.
[5] G. Mao, B. Fidan, and B. D. O. Anderson, “Wireless sensor network
localization techniques,” Comupter Networks, vol. 51, pp. 2529–2553,
2007.
[6] P. Barooah and J. P. Hespanha, “Estimation on graphs from relative
measurements,” IEEE Control Systems Magzine, no. 1, pp. 57–74, 2007.
[7] Y. Ma, S. Soatto, J. Kosecka, and S. Sastry, An Invitation to 3D Vision.
New York: Springer, 2004.
[8] R. I. Hartley and P. Sturm, “Triangulation,” Computer vision and image
understanding, vol. 68, no. 2, pp. 146–157, 1997.
[9] N. Moshtagh, N. Michael, A. Jadbabaie, and K. Daniilidis, “Visionbased, distributed control laws for motion coordination of nonholonomic
robots,” IEEE Transactions on Robotics, vol. 25, pp. 851–860, August
2009.
[10] R. Tron, J. Thomas, G. Loianno, K. Daniilidis, and V. Kumar, “A
distributed optimization framework for localization and formation control: applications to vision-based measurements,” IEEE Control Systems
Magazine, vol. 36, no. 4, pp. 22–44, 2016.
[11] A. Farina, “Target tracking with bearings-only measurements,” Signal
Processing, vol. 78, pp. 61–68, 1999.
[12] Y.-Y. Dong, C.-X. Dong, W. Liu, H. Chen, and G.-Q. Zhao, “2-D
DOA estimation for L-shaped array with array aperture and snapshots
extension techniques,” IEEE Signal Processing letters, vol. 24, no. 4,
pp. 495–499, 2017.
[13] G. Stacey and R. Mahony, “A passivity-based approach to formation
control using partial measurements of relative position,” IEEE Transactions on Automatic Control, vol. 61, no. 2, pp. 538–543, 2016.
[14] K. Becker, “Simple linear theory approach to TMA observability,” IEEE
Transactions on Aerospace and Electronic Systems, vol. 29, no. 2,
pp. 575–578, 1993.
[15] M. Deghat, I. Shames, B. D. O. Anderson, and C. Yu, “Localization
and circumnavigation of a slowly moving target using bearing measurements,” IEEE Transactions on Automatic Control, vol. 59, no. 8,
pp. 2182–2188, 2014.
[16] J. O. Swartling, I. Shames, K. H. Johansson, and D. V. Dimarogonas, “Collective circumnavigation,” Unmanned Systems, vol. 2, no. 3,
pp. 219–229, 2014.
[17] R. Zheng, Y. Liu, and D. Sun, “Enclosing a target by nonholonomic
mobile robots with bearing-only measurements,” Automatica, vol. 53,
pp. 400–407, 2015.
[18] M. Ye, B. D. O. Anderson, and C. Yu, “Bearing-only measurement selflocalization, velocity consensus and formation control,” IEEE Transactions on Aerospace and Electronic Systems, vol. 52, no. 2, pp. 575–586,
2017.
[19] R. Olfati-Saber and R. M. Murray, “Consensus problems in networks
of agents with switching topology and time-delays,” IEEE Transactions
on Automatic Control, vol. 49, no. 9, pp. 1520–1533, 2004.
[20] Z. Lin, B. Francis, and M. Maggiore, “Necessary and sufficient graphical
conditions for formation control,” IEEE Transactions on Automatic
Control, vol. 50, pp. 121–127, January 2005.
[21] W. Ren, R. W. Beard, and E. M. Atkins, “Information consensus
in multivehicle cooperative control,” IEEE Control Systems Magazine,
vol. 27, pp. 71–82, April 2007.
[22] Z. Li, Z. Duan, G. Chen, and L. Huang, “Consensus of multiagent
systems and synchronization of complex networks: A unified viewpoint,”
IEEE Transactions on Circuits and Systems I: Regular Papers, vol. 57,
no. 1, pp. 213–224, 2010.
[23] B. Servatius and W. Whiteley, “Constraining plane configurations in
computer-aided design: Combinatorics of directions and lengths,” SIAM
Journal on Discrete Mathematics, vol. 12, no. 1, pp. 136–153, 1999.
[24] A. N. Bishop, “Stabilization of rigid formations with direction-only
constraints,” in Proceedings of the 50th IEEE Conference on Decision
and Control and European Control Conference, (Orlando, FL, USA),
pp. 746–752, December 2011.
[25] T. Eren, “Formation shape control based on bearing rigidity,” International Journal of Control, vol. 85, no. 9, pp. 1361–1379, 2012.
[26] D. Zelazo, A. Franchi, and P. R. Giordano, “Rigidity theory in SE(2) for
unscaled relative position estimation using only bearing measurements,”
in Proceedings of the 2014 European Control Conference, (Strasbourgh,
France), pp. 2703–2708, June 2014.
[27] R. Tron, L. Carlone, F. Dellaert, and K. Daniilidis, “Rigid components
identification and rigidity control in bearing-only localization using
the graph cycle basis,” in Proceedings of the 2015 American Control
Conference, (Chicago, IL, USA), pp. 3911–3918, 2015.
[28] S. Zhao and D. Zelazo, “Bearing rigidity and almost global bearingonly formation stabilization,” IEEE Transactions on Automatic Control,
vol. 61, no. 5, pp. 1255–1268, 2016.
[29] L. Asimow and B. Roth, “The rigidity of graphs,” Transactions of the
American Mathematical Society, vol. 245, pp. 279–289, November 1978.
[30] R. Connelly and S. D. Guest, Frameworks, Tensegrities and Symmetry: Understanding Stable Structures. 2015. Available online
at http://www.math.cornell.edu/∼web7510/framework.pdf (last accessed
November 2017).
[31] B. Hendrickson, “Conditions for unique graph realizations,” SIAM
Journal on Computing, vol. 21, no. 1, pp. 65–84, 1992.
[32] R. Connelly, “Generic global rigidity,” Discrete & Computational Geometry, vol. 33, pp. 549–563, 2005.
[33] D. Jacobs, “An algorithm for two-dimensional rigidity percolation: The
pebble game,” Journal of Computational Physics, vol. 137, pp. 346–365,
Nov. 1997.
[34] T.-S. Tay and W. Whiteley, “Generating isostatic frameworks,” Structural
Topology, vol. 11, pp. 21–69, 1985.
[35] R. Haas, D. Orde, G. Rote, F. Santos, B. Servatius, H. Servatius,
D. Souvaine, I. Streinu, and W. Whiteley, “Planar minimally rigid graphs
and pseudo-triangulations,” Computational Geometry, vol. 31, no. 1–2,
pp. 31–61, 2005.
[36] B. Jackson, “Notes on the rigidity of graphs,” tech. rep., Queen Mary
University of London, 2007.
[37] B. D. O. Anderson, C. Yu, B. Fidan, and J. Hendrickx, “Rigid graph
control architectures for autonomous formations,” IEEE Control Systems
Magazine, vol. 28, pp. 48–63, December 2008.
[38] L. Krick, M. E. Broucke, and B. A. Francis, “Stabilization of infinitesimally rigid formations of multi-robot networks,” International Journal
of Control, vol. 82, no. 3, pp. 423–439, 2009.
[39] K.-K. Oh and H.-S. Ahn, “Distance-based undirected formations of
single-integrator and double-integrator modeled agents in n-dimensional
space,” International Journal of Robust and Nonlinear Control, vol. 24,
pp. 1809–1820, August 2014.
[40] Y.-P. Tian and Q. Wang, “Global stabilization of rigid formations in the
plane,” Automatica, vol. 49, pp. 1436–1441, May 2013.
[41] D. Zelazo, A. Franchi, and P. R. Giordano, “Distributed rigidity maintenance control with range-only measurements for multi-robot systems,”
The International Journal of Robotics Research, vol. 34, no. 1, pp. 105–
128, 2015.
[42] S. Mou, M.-A. Belabbas, A. S. Morse, Z. Sun, and B. D. O. Anderson,
“Undirected rigid formations are problematic,” IEEE Transactions on
Automatic Control, vol. 61, no. 10, pp. 2821–2836, 2016.
[43] X. Chen, M.-A. Belabbas, and T. Başar, “Global stabilization of triangulated formations,” SIAM Journal on Optimization and Control, vol. 55,
no. 1, pp. 172–199, 2017.
[44] Z. Sun, M.-C. Park, B. D. O. Anderson, and H.-S. Ahn, “Distributed
stabilization control of rigid formations with prescribed orientation,”
Automatica, vol. 78, pp. 250–257, 2017.
[45] H. G. de Marina, B. Jayawardhana, and M. Cao, “Distributed rotational
and translational maneuvering of rigid formations and their applications,” IEEE Transactions on Robotics, vol. 32, no. 3, pp. 684–697,
2016.
[46] T. Eren, O. K. Goldenberg, W. Whiteley, Y. R. Yang, A. S. Morse,
B. D. O. Anderson, and P. Belhumeur, “Rigidity, computation, and randomization in network localization,” in INFOCOM 2004. Twenty-third
Annual Joint Conference of the IEEE Computer and Communications
Societies, (Pasadena, USA), pp. 2673–2684, 2004.
[47] T. Eren, W. Whiteley, A. S. Morse, P. N. Belhumeur, and B. D. O.
Anderson, “Sensor and network topologies of formations with direction,
bearing and angle information between agents,” in Proceedings of
the 42nd IEEE Conference on Decision and Control, (Hawaii, USA),
pp. 3064–3069, December 2003.
[48] A. N. Bishop, B. D. O. Anderson, B. Fidan, P. N. Pathirana, and
G. Mao, “Bearing-only localization using geometrically constrained
optimization,” IEEE Transactions on Aerospace and Electronic Systems,
vol. 45, no. 1, pp. 308–320, 2009.
12
[49] G. Piovan, I. Shames, B. Fidan, F. Bullo, and B. D. O. Anderson,
“On frame and orientation localization for relative sensing networks,”
Automatica, vol. 49, pp. 206–213, January 2013.
[50] I. Shames, A. N. Bishop, and B. D. O. Anderson, “Analysis of noisy
bearing-only network localization,” IEEE Transactions on Automatic
Control, vol. 58, pp. 247–252, January 2013.
[51] G. Zhu and J. Hu, “A distributed continuous-time algorithm for network
localization using angle-of-arrival information,” Automatica, vol. 50,
pp. 53–63, January 2014.
[52] Z. Lin, T. Han, R. Zheng, and M. Fu, “Distributed localization for 2D sensor networks with bearing-only measurements under switching
topologies,” IEEE Transactions on Signal Processing, vol. 64, no. 23,
pp. 6345–6359, 2016.
[53] A. N. Bishop, M. Deghat, B. D. O. Anderson, and Y. Hong, “Distributed
formation control with relaxed motion requirements,” International
Journal of Robust and Nonlinear Control, vol. 25, pp. 3210–3230, 2015.
[54] S. Zhao and D. Zelazo, “Bearing-based distributed control and estimation in multi-agent systems,” in Proceedings of the 2015 European
Control Conference, (Linz, Austria), pp. 2207–2212, July 2015.
[55] S. Zhao and D. Zelazo, “Translational and scaling formation maneuver
control via a bearing-based approach,” IEEE Transactions on Control of
Network Systems, vol. 4, no. 3, pp. 429–438, 2017.
[56] K. Fathian, D. I. Rachinskii, M. W. Spong, and N. R. Gans, “Globally
asymptotically stable distributed control for distance and bearing based
multi-agent formations,” in Proceedings of the 2016 American Control
Conference, (Boston, MA, USA), pp. 4642–4648, 2016.
[57] Z. Han, L. Wang, Z. Lin, and R. Zheng, “Formation control with size
scaling via a complex Laplacian-based approach,” IEEE Transactions
on Cybernetics, vol. 46, no. 10, pp. 2348–2359, 2016.
[58] S. Coogan and M. Arcak, “Scaling the size of a formation using relative
position feedback,” Automatica, vol. 48, pp. 2677–2685, October 2012.
[59] M. Basiri, A. N. Bishop, and P. Jensfelt, “Distributed control of
triangular formations with angle-only constraints,” Systems & Control
Letters, vol. 59, pp. 147–154, 2010.
[60] A. Franchi, C. Masone, V. Grabe, M. Ryll, H. H. Bulthoff, and P. R.
Giordano, “Modeling and control of UAV bearing formations with
bilateral high-level steering,” The International Journal of Robotics
Research, vol. 31, no. 12, pp. 1504–1525, 2012.
[61] A. Cornejo, A. J. Lynch, E. Fudge, S. Bilstein, M. Khabbazian, and
J. McLurkin, “Scale-free coordinates for multi-robot systems with
bearing-only sensors,” The International Journal of Robotics Research,
vol. 32, no. 12, pp. 1459–1474, 2013.
[62] S. Zhao, F. Lin, K. Peng, B. M. Chen, and T. H. Lee, “Distributed
control of angle-constrained cyclic formations using bearing-only measurements,” Systems & Control Letters, vol. 63, no. 1, pp. 12–24, 2014.
[63] S. Zhao, F. Lin, K. Peng, B. M. Chen, and T. H. Lee, “Finite-time
stabilization of cyclic formations using bearing-only measurements,”
International Journal of Control, vol. 87, no. 4, pp. 715–727, 2014.
[64] E. Schoof, A. Chapman, and M. Mesbahi, “Bearing-compass formation
control: a human-swarm interaction perspective,” in Proceedings of the
2014 American Control Conference, (Portland, USA), pp. 3881–3886,
June 2014.
[65] R. Tron, J. Thomas, G. Loianno, K. Daniilidis, and V. Kumar, “Bearingonly formation control with auxiliary distance measurements, leaders,
and collision avoidance,” in Proceedings of the 55th Conference on
Decision and Control, (Las Vegas, USA), pp. 1806–1813, 2016.
[66] S. Zhao and D. Zelazo, “Localizability and distributed protocols for
bearing-based network localization in arbitrary dimensions,” Automatica,
vol. 69, pp. 334–341, 2016.
[67] C. Godsil and G. Royle, Algebraic Graph Theory. New York: Springer,
2001.
[68] S. Zhao and D. Zelazo, “Bearing-based formation stabilization with
directed interaction topologies,” in Proceedings of the 54th IEEE
Conference on Decision and Control, (Osaka, Japan), pp. 6115–6120,
December 2015.
[69] S. Zhao, Z. Sun, D. Zelazo, M. H. Trinh, and H.-S. Ahn, “Laman graphs
are generically bearing rigid in arbitrary dimensions,” in Proceedings
of the 56th IEEE Conference on Decision and Control, (Melbourne,
Australia), December 2017. in press (available online at https://arxiv.
org/abs/1703.04035).
[70] G. Laman, “On graphs and rigidity of plane skeletal structures,” Journal
of Engineering Mathematics, vol. 4, no. 4, pp. 331–340, 1970.
[71] S. Zhao and D. Zelazo, “Bearing-based formation maneuvering,” in Proceedings of the 2015 IEEE Multi-Conference on Systems and Control,
(Sydney, Australia), pp. 658–663, September 2015.
[72] S. Zhao, D. V. Dimarogonas, Z. Sun, and D. Bauso, “A general
approach to coordination control of mobile agents with motion constraints,” IEEE Transactions on Automatic Control. in press (DOI:
10.1109/TAC.2017.2750924).
[73] J. Hendrickx, B. D. O. Anderson, J. Delvenne, and V. Blondel, “Directed
graphs for the analysis of rigidity and persistence in autonomous
agent systems,” International Journal of Robust and Nonlinear Control,
vol. 17, no. 10-11, pp. 960–981, 2007.
[74] C. Yu, B. D. O. Anderson, A. S. Dasgupta, and B. Fidan, “Control of
minimally persistent formations in the plane,” SIAM Journal on Control
and Optimization, vol. 48, pp. 206–233, Februray 2009.
[75] T. H. Summers, C. Yu, S. Dasgupta, and B. D. O. Anderson, “Control
of minimally persistent leader-remote-follower and coleader formations
in the plane,” IEEE Transactions on Automatic Control, vol. 56, no. 12,
pp. 2778–2792, 2011.
[76] D. Mukherjee, M.-H. Trinh, D. Zelazo, and H.-S. Ahn, “Bearing-only
cyclic pursuit in 2-D for capture of moving target,” in 57th Israel Annual
Conference on Aerospace Sciences, (Haifa, Israel), 2017.
[77] M.-H. Trinh, D. Mukherjee, D. Zelazo, and H.-S. Ahn, “Formations on
directed cycles with bearing-only measurements,” International Journal
of Robust and Nonlinear Control, pp. 1–23, 2017. in press (DOI:
10.1002/rnc.3921).
[78] K.-K. Oh and H.-S. Ahn, “Formation control and network localization
via orientation alignment,” IEEE Transactions on Automatic Control,
vol. 59, pp. 540–545, February 2014.
[79] B.-H. Lee and H.-S. Ahn, “Distributed formation control via global
orientation estimation,” Automatica, vol. 73, pp. 125–129, 2016.
[80] D. Zelazo, A. Franchi, and P. R. Giordano, “Formation control using a
SE(2) rigidity theory,” in Proceedings of the 54th IEEE Conference on
Decision and Control, (Osaka, Japan), pp. 6121–6126, 2015.
[81] F. Schiano, A. Franchi, D. Zelazo, and P. R. Giordano, “A RigidityBased Decentralized Bearing Formation Controller for Groups of
Quadrotor UAVs,” in Proceedings of IEEE/RSJ International Conference
on Intelligent Robots and Systems, (Daejeon, Korea), pp. 5099–5106,
2016.
[82] G. Michieletto, A. Cenedese, and A. Franchi, “Bearing rigidity theory
in SE(3),” in Proceedings of the 55th IEEE Conference on Decision and
Control, (5950–5955), pp. 83–92, 2016.
[83] F. Schiano and P. G. Robuffo, “Bearing rigidity maintenance for formations of quadrotor UAVs,” in Proceedings of IEEE International
Conference on Robotics and Automation, (Singapore), 2017.
[84] J. Yu, S. M. LaValle, and D. Liberzon, “Rendezvous without coordinates,” IEEE Transactions on Automatic Control, vol. 57, no. 2, pp. 421–
434, 2012.
[85] R. Zheng and D. Sun, “Rendezvous of unicycles: A bearings-only and
perimeter shortening approach,” Systems & Control Letters, vol. 62,
pp. 401–407, May 2013.
[86] J. Grzymisch and W. Fichter, “Optimal rendezvous guidance with
enhanced bearings-only observability,” Journal of Guidance, Control,
and Dynamics, vol. 38, no. 6, pp. 1131–1139, 2015.
[87] M. Kriegleder, S. T. Digumarti, R. Oung, and R. D’Andrea, “Rendezvous with bearing-only information and limited sensing range,” in
Proceedings of the 2015 IEEE International Conference on Robotics
and Automation, (Seattle, Washington), pp. 5941–5947, 2015.
[88] S. Zhao and R. Zheng, “Flexible bearing-only rendezvous control of
mobile robots,” in Proceedings of the 36th Chinese Control Conference,
(Dalian, China), pp. 8051–8056, July 2017.
[89] S. G. Loizou and V. Kumar, “Biologically inspired bearing-only navigation and tracking,” in Proceedings of the 46th IEEE Conference on
Decision and Control, (New Orleans, LA, USA), pp. 1386–1391, 2007.
[90] M. Deghat, I. Shames, A. N. Bishop, B. D. O. Anderson, and C. Yu,
“Localization and circumnavigation of a slowly moving target using bearing measurements,” IEEE Transactions on Automatic Control,
vol. 59, pp. 2182–2188, August 2014.
[91] M.-H. Trinh, G.-H. Ko, V.-H. Pham, K.-K. Oh, and H.-S. Ahn, “Guidance using bearing-only measurements with three beacons in the plane,”
Control Engineering Practice, vol. 51, pp. 81–91, 2016.
13
N OTATIONS FOR N ETWORKS AND F ORMATIONS
Given a network of n nodes in Rd where n ≥ 2, d ≥ 2, let
the position of node i be pi ∈ Rd and the configuration of the
points be p = [pT1 , . . . , pTn ]T ∈ Rdn . The interaction among
the nodes is described by a graph G = (V, E) which consists
of a vertex set V = {1, . . . , n} and an edge set E ⊆ V × V. If
(i, j) ∈ E, node i receives information from node j, and node
j is called adjacent to i. The set of neighbors of vertex i is
denoted as Ni = {j ∈ V : (i, j) ∈ E}. This article focuses on
undirected graphs where (i, j) ∈ E ⇔ (j, i) ∈ E. Let m be
the number of undirected edges in the graph. An orientation
of an undirected graph is the assignment of a direction to each
edge. An oriented graph is an undirected graph together with
an orientation. The incidence matrix H ∈ Rm×n of an oriented
graph is the {0, ±1}-matrix with rows indexed by edges and
columns by vertices.
A network, denoted as (G, p), is G with its vertex i ∈ V
mapped to pi . Network may be called as formation in the
context of formation control. For a network (G, p), define the
edge and bearing vectors for (i, j) ∈ E as eij = pj − pi and
gij = eij /keij k, respectively. Here gij is the unit vector pointing from pi to pj . It represents the relative bearing of pi with
respect to pj . Note that eij = −eji and gij = −gji . Consider
an orientation of the graph G and suppose (i, j) corresponds to
the kth edge in the oriented graph. Then the edge and bearing
vectors may be reexpressed as ek = pj −pi and gk = ek /kek k
where k ∈ {1, . . . , m}. Denote e = [eT1 , . . . , eTm ]T and
T T
] . Note that e = (H ⊗ Id )p where ⊗ denotes
g = [g1T , . . . , gm
the Kronecker product. In this article, Null(·) and Range(·)
denote the null and range spaces of a matrix, respectively.
Denote 1n , [1, . . . , 1]T ∈ Rn . Let k·k be the Euclidian norm
of a vector or the spectral norm of a matrix, and Id ∈ Rd×d
be the identity matrix.
14
K EY D EFINITIONS IN B EARING R IGIDITY T HEORY
K EY D EFINITIONS IN D ISTANCE R IGIDITY T HEORY
Definition S1 (Bearing Equivalency). Two networks (G, p)
and (G, p0 ) are bearing equivalent if P(pi −pj ) (p0i − p0j ) = 0
for all (i, j) ∈ E.
Definition S1 (Distance Equivalency). Two networks (G, p)
and (G, p0 ) are distance equivalent if kpi − pj k = kp0i − p0j k
for all (i, j) ∈ E.
Definition S2 (Bearing Congruency). Two networks (G, p) and
(G, p0 ) are bearing congruent if P(pi −pj ) (p0i − p0j ) = 0 for all
i, j ∈ V.
Definition S2 (Distance Congruency). Two networks (G, p)
and (G, p0 ) are distance congruent if kpi − pj k = kp0i − p0j k
for all i, j ∈ V.
Definition S3 (Bearing Rigidity). A network (G, p) is bearing
rigid if there exists a constant > 0 such that any network
(G, p0 ) that is bearing equivalent to (G, p) and satisfies kp0 −
pk < is also bearing congruent to (G, p).
Definition S3 (Distance Rigidity). A network (G, p) is distance
rigid if there exists a constant > 0 such that any network
(G, p0 ) that is distance equivalent to (G, p) and satisfies kp0 −
pk < is also distance congruent to (G, p).
Definition S4 (Global Bearing Rigidity). A network (G, p) is
globally bearing rigid if an arbitrary network that is bearing
equivalent to (G, p) is also bearing congruent to (G, p).
Definition S4 (Global Distance Rigidity). A network (G, p) is
globally distance rigid if an arbitrary network that is distance
equivalent to (G, p) is also distance congruent to it.
Consider an oriented graph where the inter-neighbor bearings can be expressed by {gk }m
k=1 . Define the bearing function
FB : Rdn → Rdm as
Consider an oriented graph where the inter-neighbor distances can be expressed by {kek k}m
k=1 . Define the distance
function FD : Rdn → Rdm as
T T
FB (p) = [g1T , . . . , gm
] ∈ Rdm .
FD (p) = [ke1 k2 , . . . , kem k2 ]T /2 ∈ Rm .
The bearing rigidity matrix is defined as the Jacobian of the
bearing function
The distance rigidity matrix is defined as the Jacobian of the
distance function
∂FD (p)
∈ Rm×dn .
(S2)
RD (p) =
∂p
RB (p) =
∂FB (p)
∈ Rdm×dn .
∂p
(S1)
A matrix-vector form RD (p) is
A matrix-vector form RB (p) is
RB (p) = diag(Pg1 /ke1 k, . . . , Pgm /kem k)(H ⊗ Id ).
RD (p) = diag(eT1 , . . . , eTm )(H ⊗ Id ).
Let δp ∈ Rdn be a variation of the configuration p. If
RB (p)δp = 0, then δp is called an infinitesimal bearing motion
of (G, p). An infinitesimal bearing motion is called trivial if
it only corresponds to a translation and a scaling of the entire
network.
Let δp ∈ Rdn be a variation of the configuration p. If
RD (p)δp = 0, then δp is called an infinitesimal distance
motion of (G, p). An infinitesimal distance motion is called
trivial if it only corresponds to a translation and a rotation of
the entire network.
Definition S5 (Infinitesimal Bearing Rigidity). A network is
infinitesimally bearing rigid if all the infinitesimal bearing
motions are trivial.
Definition S5 (Infinitesimal Distance Rigidity). A network is
infinitesimally distance rigid if all the infinitesimal distance
motions are trivial.
The relation between bearing rigidity, global bearing rigidity, and infinitesimal bearing rigidity is illustrated in Figure S1.
Details of these notions can be found in [28].
The relation between distance rigidity, global distance rigidity, and infinitesimal distance rigidity is illustrated in Figure S2. Details of the notions can be found in [29]–[32], [36].
infinitesimal
distance rigidity
infinitesimal
bearing rigidity
bearing rigidity
global
bearing rigidity
Fig. S1: The relation between bearing rigidity, global bearing rigidity, and
infinitesimal bearing rigidity. Infinitesimal bearing rigidity implies both bearing rigidity and global bearing rigidity. Global bearing rigidity and bearing
rigidity imply each other.
distance rigidity
global
distance rigidity
Fig. S2: The relation between distance rigidity, global distance rigidity, and
infinitesimal distance rigidity. Both infinitesimal and global distance rigidity
imply distance rigidity. Infinitesimal and global distance rigidity do not imply
each other.
15
B EARING L APLACIAN OF N ETWORKS
A N O RTHOGONAL P ROJECTION M ATRIX
For any nonzero vector x ∈ R
orthogonal projection matrix as
P (x) = Id −
d
(d ≥ 2), define an
x xT
∈ Rd×d .
kxk kxk
For notational simplicity, denote Px = P (x). The matrix Px is
an orthogonal projection matrix that geometrically projects any
vector onto the orthogonal compliment of x (see Figure S3).
Matrix Px satisfies PxT = Px , Px2 = Px , and Null(Px ) =
span{x}. This matrix is positive semi-definite with one eigenvalue equal to zero and d − 1 eigenvalues equal to one. Other
properties of Px are summarized as below.
(a) Any two nonzero vectors x, y ∈ Rd are parallel if and
only if Px y = 0 [28, Lemma 1].
(b) Any two unit vectors x, y ∈ Rd satisfy xT Py x = y T Px y
[28, Lemma 8].
(c) For any nonzero vectors
∈ Rd where m ≥
Pm x1 , . . . , xmd×d
2, d ≥ 2, the matrix i=1 Pxi ∈ R
is nonsingular if
and only if at least two of x1 , . . . , xm are not collinear
[69, Lemma 3].
(d) For any nonzero vector x ∈ R2 , denote x⊥ ∈ R2 as
a nonzero normal vector that satisfies xT x⊥ = 0. Then
Px = x⊥ (x⊥ )T /kx⊥ k2 . The proof follows from the fact
that the matrix A = [x/kxk, x⊥ /kx⊥ k] ∈ R2×2 satisfies
AT A = AAT = I2 .
(e) For any two nonzero vectors x, y ∈ Rd , if θ ∈ [0, π] is
the angle between them so that xT y = kxkkyk cos θ, then
kPx − Py k = sin θ [66, Lemma 5]. This property has
been used to analyze the perturbation of the orthogonal
projection matrix.
2
(f) If x ∈ R3 is a unit vector, then Px = − [x]× , where
0
−x3 x2
0
−x1 ∈ R3×3
[x]× = x3
−x2 x1
0
is the skew-symmetric matrix associated with x [7, Theorem 2.11]. This property has been used in [72, Equation (6)]
The orthogonal projection matrix plays an important role in
the bearing rigidity theory and its applications.
x
y
0
Px y
Fig. S3: Illustration of the orthogonal projection matrix. Given any nonzero
x, y ∈ Rd , the vector Px y is the orthogonal projection of y onto the
orthogonal compliment of x.
Given network (G, p) with no collocated nodes, define the
bearing Laplacian B ∈ Rdn×dn as [66]
i 6= j, (i, j) ∈
/ E,
0d×d ,
−P
,
i
=
6
j,
(i,
j)
∈
E,
[B]ij =
g
ij
P
P
,
i
=
j,
i
∈
V,
g
ik
k∈Ni
where [B]ij ∈ Rd×d is the ijth block of submatrix of
B. The bearing Laplacian can be viewed a matrix-weighted
Laplacian which describes both the underlying graph and the
inter-neighbor bearings of the network. See Figure S4 for
illustration.
For undirected graphs, the bearing Laplacian has the following properties [66, Lemma 2]:
(a) B is symmetric and positive semi-definite because for any
x = [xT1 , . . . , xTn ]T ∈ Rdn
1X X
(xi − xj )T Pgij (xi − xj ) ≥ 0.
xT Bx =
2
i∈V j∈Ni
(b) rank(B) ≤ dn − d − 1 and span{1 ⊗ Id , p} ⊆ Null(B)
for any network.
(c) rank(B) = dn − d − 1 and Null(B) = span{1 ⊗ Id , p} if
and only if the network is infinitesimally bearing rigid.
In a network with na anchors and nf = n − na followers, the
bearing Laplacian may be partitioned into
Baa Baf
B=
,
Bf a Bf f
where Bf f ∈ Rdnf ×dnf . For any network, Bf f is positive
semi-definite and satisfies Bf f pf = −Bf a pa [66, Lemma 3].
In the context of formation control, the anchors are called
leaders and the subscript a is replaced by `.
g31
g13
1
g12
3
g32
g23
g21
2
B=
Pg12 + Pg13
−Pg12
−Pg13
−Pg21
Pg21 + Pg23
−Pg23
−Pg31
−Pg32
Pg31 + Pg32
Fig. S4: Example to demonstrate bearing Laplacian. The network is the
complete graph on three nodes. The bearing Laplacian has the same structure
as a weighted graph Laplacian matrix [67] with the weights on each edge
corresponding to the projection matrices Pgij .
16
L AMAN G RAPHS AND H ENNEBERG C ONSTRUCTION
An undirected graph G = (V, E) is called Laman if m =
2n − 3 and every subset of k ≥ 2 vertices spans at most
2k − 3 edges [70]. Laman graphs can be characterized by the
Henneberg construction as described below. Given a graph
G = (V, E), a new graph G 0 = (V 0 , E 0 ) is formed by adding
a new vertex v to G and performing one of the following two
operations:
(a) Vertex addition: connect vertex v to any two existing
vertices i, j ∈ V. In this case, V 0 = V ∪ {v} and
E 0 = E ∪ {(v, i), (v, j)}. See Figure S5(a) for illustration.
(b) Edge splitting: consider three vertices i, j, k ∈ V with
(i, j) ∈ E and connect vertex v to i, j, k and delete
(i, j). In this case, V 0 = V ∪ {v} and E 0 = E ∪
{(v, i), (v, j), (v, k)} \ {(i, j)}. See Figure S5(b) for illustration.
A Henneberg construction starting from an edge connecting
two vertices leads to a Laman graph [34]–[36]. The converse is
also true. That is if a graph is Laman, then it can be generated
by a Henneberg construction [35, Lemma 2]. The underlying
graphs of the networks in Figure 2(a)–(c) are Laman. Laman
graphs play critical roles in the construction of distance rigid
and bearing rigid networks.
v
v
i
i
k
j
G
(a) Vertex addition
j
G
(b) Edge splitting
Fig. S5: The two operations of the Henneberg construction. The Henneberg
construction can be used to generate all minimally infinitesimally distance
rigid graphs in the plane. The main idea is to ensure that the vertex addition
and edge splitting operations satisfy the Laman condition at each step.
C OMPARISON OF B EARING R IGIDITY AND D ISTANCE
R IGIDITY
Both of the bearing and distance rigidity theories address
the same problem of when the geometric pattern of a network
can be uniquely determined. The difference is that the bearing
rigidity theory considers inter-neighbor bearings whereas the
distance rigidity theory considers inter-neighbor distances. The
term “unique pattern” in the bearing rigidity theory means the
location of a network can be determined up to a translational
and scaling factor, while in the distance rigidity theory it
means the network can be determined up to a translational
and rotational factor.
One connection between the two rigidity theories is that
infinitesimal bearing rigidity is equivalent to infinitesimal
distance rigidity in two dimensions [28, Theorem 8]. In other
words, a network in the plane is infinitesimally bearing rigid
if and only if it is infinitesimally distance rigid. This equivalence property explains why the distance rigidity theory could
be used to analyze the problems of bearing-based network
localization or formation control in the literature [49], [51],
[64]. It also suggests that the infinitesimal distance rigidity of
a network by be examined by its infinitesimal bearing rigidity.
For example, it may not be straightforward to see that the networks in Figure 3(c)-(d) are not infinitesimally distance rigid.
However, it is intuitive to see these networks are not infinitesimally bearing rigid because there exist nontrivial infinitesimal
bearing motions. It must be noted that the equivalence cannot
be generalized to three or higher dimensions. For example,
the three-dimensional networks shown in Figure 2(c)-(e) are
infinitesimally bearing rigid but not infinitesimally distance
rigid.
Compared to infinitesimal distance rigidity, infinitesimal
bearing rigidity possess some interesting properties. First, infinitesimal bearing rigidity not only ensures the unique pattern
of a network, but also can be examined by a rank condition
easily. As a comparison, infinitesimal distance rigidity may not
be able to ensure a unique pattern though it can be examined
by a rank condition. Second, an infinitesimally bearing rigid
network remains infinitesimally bearing rigid when the dimension is lifted up to a higher dimension [28, Theorem 7]. As a
comparison, a network that is infinitesimally distance rigid in
the plane may be flexible in a higher dimension. Third, in the
bearing rigidity theory, a Laman graph is generically bearing
rigid in arbitrary dimensions and at most 2n−3 edges would be
sufficient to guarantee the bearing rigidity of a network in an
arbitrary dimension. As a comparison, although a Laman graph
embedded in a generic configuration is infinitesimally distance
rigid [30], [34]–[37], this result, known as Laman’s Theorem
[70], is valid merely in two dimensional spaces. In three
or higher dimensions, extra conditions and more edges are
required to guarantee distance rigidity. The above comparison
is summarized in Table S1.
Why bearing rigidity has appealing properties in high dimensions can be explained intuitively from the perspective
of degree of freedom. For example, consider a network of n
nodes in d-dimensional space. The network has dn degrees
of freedoms. In order to ensure the rigidity of the network,
17
there must exist sufficient distance or bearing constraints to
reduce the degrees of freedom of the network to certain desired
values. Given a distance rigid network, when lifted up to
a higher dimension, the degrees of freedom of the network
increases while the number of constraints posed by an interneighbor distance remain the same. As a result, in order to
preserve distance rigidity in higher dimensions, more distance
constraints are required. As a comparison, when lifted to
a higher dimension, the number of independent constraints
posed by an inter-neighbor bearing also increases. For example, a bearing in the plane is equivalent to an azimuth angle
whereas a bearing in the three dimensional space is equivalent
to two bearing angles: azimuth and altitude. As a result, the
same number of bearings are still able to preserve the bearing
rigidity of the network.
B EARING R IGIDITY T HEORY FOR SE(2)
Consider a collection of n nodes in R2 × S 1 . Each point is
described by its position pi ∈ R2 and its orientation ψi ∈ S 1 .
An SE(2) network, denoted as (G, p, ψ), is the directed graph
G = (V, E), and the configuration (p, ψ), where each vertex
i ∈ V in the graph is mapped to the point (pi , ψi ) ∈ SE(2).
Note that SE(2) networks, directed graphs are considered.
Suppose (i, j) ∈ E is the kth directed edge where k =
{1, . . . , m} and m denotes the number of directed edges in E.
Let gk be the relative bearing of pj with respect to pi expressed
in the global frame. Then,
cos ψi
sin ψi
rk =
gk
− sin ψi cos ψi
is the bearing gk expressed in node i’s local reference frame.
Define the directed bearing function associated with the
m
SE(2) network, FSE : SE(2)n → S 2 , as
m
T T
FSE (p, ψ) = [r1T · · · rm
] ∈ S2 .
(S3)
The corresponding directed bearing rigidity matrix is defined
as the Jacobian of the directed bearing function,
RSE (p, ψ) ,
∂FSE (p, ψ)
∈ R2m×3n .
∂(p, ψ)
(S4)
Let δχ ∈ R3n be a variation of the configuration (p, ψ). If
RSE (p, ψ)δχ = 0, then δχ is called an infinitesimal SE(2)
bearing motion of G(p, ψ). There are three types of trivial
infinitesimal SE(2) motions corresponding to translations,
scalings, and coordinated rotations of the entire network. The
coordinated rotation involves an angular rotation of each agent
about its own body axis with a rigid-body rotation of the
network (see Figure S6). An SE(2) network is infinitesimally
bearing rigid if all the infinitesimal bearing motions are trivial.
A necessary and sufficient condition for an SE(2) network to
be infinitesimally bearing rigid is [80], [81]
rank[RSE (p, ψ)] = 3n − 4,
or equivalently
Null[RSE (p, ψ)] = span
1n ⊗ I2
0
⊥
p
p
,
,
,
0
1n
T
⊥ T T
and p⊥
where p⊥ = [(p⊥
1 ) , . . . , (pn ) ]
i = Rπ/2 pi . The
null-space is characterized in this way after a permutation of
the matrix that groups the positions and attitudes of all agents
together. Here Rπ/2 is a rotation matrix that rotates any vector
by π/2.
Detailed definitions in the SE(2) bearing rigidity theory
can be found in [26], [80], [81]. The SE(2) rigidity theory
has been employed for distributed relative position estimation
[26] and formation control [80], [81], [83]. A similar approach
has been extended for SE(3) [82].
18
TABLE S1: Comparison of infinitesimal bearing rigidity and infinitesimal distance rigidity.
Infinitesimal Bearing Rigidity (IBR)
Infinitesimal Distance Rigidity (IDR)
Unique geometric pattern
Yes, IBR ensures the unique pattern of a network.
No, IDR does not ensure the unique pattern of a
network (global distance rigidity does).
Rank condition
Yes, IBR corresponds to a rank condition of the
bearing rigidity matrix.
Yes, IDR corresponds to a rank condition of the
distance rigidity matrix.
Invariance to dimension
Yes, a network that is IBR in a lower dimension
remains IBR in a higher dimension.
No, a network that is IDR in a lower dimension may
be flexible in a higher dimension. (Universal distance
rigidity is invariant to dimensions)
Minimum edge number
In an arbitrary dimension, 2n − 3 edges are sufficient to ensure IBR. Less than 2n − 3 edges may
also be sufficient to ensure IBR in three or higher
dimensions.
In the plane, 2n−3 is the minimum number of edges
to ensure IDR. More than 2n − 3 edges are required
to ensure IDR in three or higher dimensions.
Laman graphs
In an arbitrary dimension, Laman graphs mapped to
almost all configurations result in IBR networks.
In the plane, Laman graphs mapped to almost all
configurations result in IDR networks. A similar
result does not exist in higher dimensions.
3
3
1
1
2
2
Fig. S6: Example of two congruent SE(2) networks. The above two networks
differ in terms of a translation, a scaling, and a coordinated rotation.
19
AUTHOR I NFORMATION
Shiyu Zhao is a Lecturer in the Department of Automatic Control and Systems Engineering at the University of
Sheffield, UK. He received the B.E. and M.E. degrees from
Beijing University of Aeronautics and Astronautics, China, in
2006 and 2009, respectively. He obtained the Ph.D. degree in
Electrical Engineering from the National University of Singapore in 2014. From 2014 to 2016, he served as post-doctoral
researchers at the Technion - Israel Institute of Technology and
University of California at Riverside. He is a corecipient of the
Best Paper Award (Guan Zhao-Zhi Award) in the 33rd Chinese
Control Conference, Nanjing, China, in 2014. His research
interests lie in networked dynamical systems and unmanned
aerial vehicles.
Daniel Zelazo is an Assistant Professor of Aerospace Engineering at the Technion - Israel Institute of Technology, Israel.
He received the B.S. (99) and M.Eng. (01) degrees in Electrical
Engineering from the Massachusetts Institute of Technology.
In 2009, he completed his Ph.D. from the University of
Washington in Aeronautics and Astronautics. From 2010 to
2012, he served as a post-doctoral research associate and
lecturer at the Institute for Systems Theory & Automatic
Control in the University of Stuttgart. His research interests
include topics related to multi-agent systems, optimization,
and graph theory.
| 3cs.SY |
Just-in-Time Static Type Checking for Dynamic Languages
Brianna M. Ren
Jeffrey S. Foster
arXiv:1604.03641v1 [cs.PL] 13 Apr 2016
University of Maryland, College Park
{bren, jfoster}@cs.umd.edu
Abstract
and convoluted; and prior mixed static/dynamic systems are
either cumbersome or make certain limiting assumptions.
(Section 6 discusses prior work in detail.)
In this paper, we introduce Hummingbird,1 a type checking system for Ruby that solves this problem using a new
approach we call just-in-time static type checking. In Hummingbird, user-provided type annotations actually execute at
run-time, adding types to an environment that is maintained
during execution. As metaprogramming code creates new
methods, it, too, executes type annotations to assign types to
dynamically created methods. Then whenever a method m is
called, Hummingbird statically type checks m’s body in the
current dynamic type environment. More precisely, Hummingbird checks that m calls methods at their types as annotated, and that m itself matches its annotated type. Moreover, Hummingbird caches the type check so that it need not
recheck m at the next call unless the dynamic type environment has changed in a way that affects m.
Just-in-time static type checking provides a highly effective tradeoff between purely dynamic and purely static
type checking. On the one hand, metaprogramming code is
very challenging to analyze statically, but in our experience
it is easy to create type annotations at run time for generated code. On the other hand, by statically analyzing whole
method bodies, we catch type errors earlier than a purely
dynamic system, and we can soundly reason about all possible execution paths within type checked methods. (Section 2
shows how several examples of metaprogramming are handled by Hummingbird.)
To ensure our approach to type checking is correct, we
formalize Hummingbird using a core, Ruby-like language in
which method creation and method type annotation can occur at arbitrary points during execution. We provide a flowsensitive type checking system and a dynamic semantics that
invokes the type system at method entry, caching the resulting typing proof. Portions of the cache may be invalidated
as new methods are defined or type annotations are changed.
We prove soundness for our type system. (Section 3 presents
the formalism.)
Our implementation of Hummingbird piggybacks on two
prior systems we developed. We use the Ruby Intermediate
Dynamic languages such as Ruby, Python, and JavaScript
have many compelling benefits, but the lack of static types
means subtle errors can remain latent in code for a long time.
While many researchers have developed various systems to
bring some of the benefits of static types to dynamic languages, prior approaches have trouble dealing with metaprogramming, which generates code as the program executes. In
this paper, we propose Hummingbird, a new system that uses
a novel technique, just-in-time static type checking, to type
check Ruby code even in the presence of metaprogramming.
In Hummingbird, method type signatures are gathered dynamically at run-time, as those methods are created. When
a method is called, Hummingbird statically type checks the
method body against current type signatures. Thus, Hummingbird provides thorough static checks on a per-method
basis, while also allowing arbitrarily complex metaprogramming. For performance, Hummingbird memoizes the static
type checking pass, invalidating cached checks only if necessary. We formalize Hummingbird using a core, Ruby-like
language and prove it sound. To evaluate Hummingbird, we
applied it to six apps, including three that use Ruby on Rails,
a powerful framework that relies heavily on metaprogramming. We found that all apps typecheck successfully using
Hummingbird, and that Hummingbird’s performance overhead is reasonable. We applied Hummingbird to earlier versions of one Rails app and found several type errors that had
been introduced and then fixed. Lastly, we demonstrate using Hummingbird in Rails development mode to typecheck
an app as live updates are applied to it.
Categories and Subject Descriptors F.3.2 [Semantics of
Programming Languages]: Program analysis
Keywords type checking; dynamic languages; Ruby
1. Introduction
Many researchers have explored ways to bring the benefits
of static typing to dynamic languages [1–4, 12, 14, 18, 20–
22, 25, 27, 36, 37]. However, many of these prior systems do
not work well in the presence of metaprogramming, in which
code the program relies on is generated as the program
executes. The challenge is that purely static systems cannot
analyze metaprogramming code, which is often complicated
1A
hummingbird can dynamically flap its wings while statically hovering
in place.
1
2016/4/14
Language [8, 13] to parse input Ruby files and translate them
to simplified control-flow graphs. We use RDL [26, 33], a
Ruby contract system, to intercept method calls and to represent and store method type signatures at run time. Hummingbird supports an extensive set of typing features, including union types, intersection types, code blocks (anonymous
functions), generics, modules, and type casts, among others.
(Section 4 describes our implementation.)
We evaluated Hummingbird by applying it to six Ruby
apps. Three use Ruby on Rails (just “Rails” from now
on), a popular, sophisticated web app framework that uses
metaprogramming heavily both to make Rails code more
compact and expressive and to support “convention over
configuration.” We should emphasize that Rails’s use of
metaprogramming makes static analysis of it very challenging [17]. Two additional apps use other styles of metaprogramming, and the last app does not use metaprogramming,
as a baseline.
We found that all of our subject apps type check successfully using Hummingbird, and that dynamically generated
types are essential for the apps that use metaprogramming.
We also found that Hummingbird’s performance overhead
ranges from 19% to 469%, which is much better than prior
approaches [18, 27], and that caching is essential to achieving this performance. For one Rails app, we ran type checking on many prior versions, and we found a total of six type
errors that had been introduced and then later fixed. We also
ran the app in Rails development mode, which reloads files
as they are edited, to demonstrate how Hummingbird type
check caching behaves in the presence of modified methods.
(Section 5 reports on our results.)
In summary, we believe Hummingbird is an important
step forward in our ability to bring the benefits of static
typing to dynamic languages while still supporting flexible
and powerful metaprogramming features.
1
2
3
4
5
6
7
class Talk < ActiveRecord::Base
belongs to :owner, :class name ⇒ ”User”
....
type :owner?, ”(User) → %bool”
def owner?(user)
return owner == user
end end
8
9
10
11
12
13
14
15
16
17
18
19
module ActiveRecord:: Associations :: ClassMethods
pre (: belongs to ) do |∗ args |
hmi = args[0]
options = args [1]
hm = hmi.to s
cn = options [: class name] if options
hmu = cn ? cn : hm. singularize . camelize
type hm. singularize , ”() → #{hmu}”
type ”#{hm.singularize}=”, ”(#{hmu}) →#{hmu}”
true
end end
Figure 1. Ruby on Rails Metaprogramming.
a keyword, but in fact it is simply a method call. The call
passes the symbol (an interned string) :owner as the first
argument, and the second argument is a hash that maps
symbol :class name to string ”User”.
Now consider the owner? method, defined on line 5. Just
before the method, we introduce a type annotation indicating
the method takes a User and returns a boolean. Given such
an annotation, Hummingbird’s goal is to check whether the
method body has the indicated type.2 This should be quite
simple in this case, as the body of owner? just calls noargument method owner and checks whether the result is
equal to user.
However, if we examine the remaining code of Talk (not
shown), we discover that owner is not defined anywhere in
the class! Instead, this method is created at run-time by belongs to. More specifically, when belongs to is called, it defines several convenience methods that perform appropriate
SQL queries for the relationship [23], in this case to get the
User instance associated with the Talk’s owner. Thus, as we
can see, it is critical for Hummingbird to handle such dynamically created methods even to type check simple Rails
code.
Our solution is to instrument belongs to so that, just
as it creates a method dynamically, it also creates method
type signatures dynamically. The code on lines 9–19 of Figure 1 accomplishes this. Hummingbird is built on RDL,
a Ruby contract system for specifying pre- and postconditions [26, 33]. The precondition is specified via a code
block—an anonymous function (i.e., a lambda) delimited by
do. . .end—passed to pre. Here the code block trivially re-
2. Overview
We begin our presentation by showing some uses of metaprogramming in Ruby and the corresponding Hummingbird
type checking process. The examples below are from the
experiments in Section 5.
Rails Associations. The top of Figure 1 shows an excerpt
from the Talks Rails app. This code defines a class Talk that
is a model in Rails, meaning an instance of Talk represents a
row in the talks database table. The change in case and pluralization here is not an accident—Rails favors “convention
over configuration,” meaning many relationships that would
otherwise be specified via configuration are instead implicitly expressed by using similar or the same name for things.
In this app, every talk is owned by a user, which in
implementation terms means a Talk instance has a foreign
key owner id indicating the owner, which is an instance of
class User (not shown). The existence of that relationship
is defined on line 2. Here it may look like belongs to is
2 In practice
type takes another argument to tell Hummingbird to type check
the body, in contrast to library and framework methods whose types are
trusted. We elide this detail for simplicity.
2
2016/4/14
1
2
3
4
5
6
7
8
module Rolify :: Dynamic
def define dynamic method(role name, resource )
class eval do
define method(” is #{role name}?”.to sym) do
has role ?(”#{role name}”)
end if !method defined?(”is #{role name}?”.to sym)
...
end end
9
10
11
12
13
16
17
18
19
20
2
3
4
5
6
7
8
9
pre :define dynamic method do |role name, resource |
type ” is #{role name}?”, ”() → %bool”
true
end end
14
15
1
10
11
12
13
14
class User; include Rolify :: Dynamic end
user = User. first
user .define dynamic method(”professor”, ...)
user .define dynamic method(”student”, ...)
user . is professor ?
user . is student ?
Transaction = Struct.new(:type, :account name, :amount)
class ApplicationRunner
def process transactions
@transactions .each do | t |
name = t.account name
...
end ... end
field type : @transactions , ”Array<Transaction>”
end
15
16
17
18
19
20
21
class Struct
def self .add types(∗types)
members.zip(types).each {|name, t|
self . class eval do
type name, ”() → #{t}”
type ”#{name}=”, ”(t) →#{t}”
end
}
end
end
Transaction .add types(”String”, ”String”, ”String”)
Figure 2. Methods Dynamically Created by User Code.
Figure 3. Type Signatures for Struct.
turns true so the precondition is always satisfied (last line)
and, as a side effect, creates method type annotations for belongs to.
In more detail, hmi is set to the first argument to belongs to, and options is either nil or the hash argument, if
present. (Here hm is shorthand for “has many,” i.e., since the
Talk belongs to a User, the User has many Talks.) Then hmu
is set to either the class name argument, if present, or hmi
after singularizing and camel-casing it. Then type is called
twice, once to give a type to a getter method created by belongs to, and once for a setter method (whose name ends
with =). Notation #{e} inside a string evaluates the expression e and inserts the result in the string. In this particular
case, these two calls to type evaluate to
code matching the given type signature. Figure 2 shows
an example, extracted from Rolify, in which user code dynamically generates a method. The first part of the figure defines a module (aka mixin) with a two-argument
method define dynamic method. The method body calls define method to create a method named using the first argument, as long as that method does not exist (note the postfix
if on line 6). Similarly to earlier, line 10 adds a precondition to define dynamic method that provides an appropriate
method type. (We do not check for a previous type definition
since adding the same type again is harmless.)
The code starting at line 15 uses the module. This particular code is not from our experiment but is merely for expository purposes. Here we (re)open class User and mix in the
module. Then we create a user; call define dynamic method
twice; and then call the generated methods is professor? and
is student?.
In this case, since the generated methods have type annotations and are in user code, Hummingbird type checks
their bodies when they are called, just like any other userdefined method with a type. For example, consider the call
to is professor?, which is given type () → %bool. At the
call, Hummingbird type checks the code block at line 4 and
determines that it has no arguments and that its body returns
a boolean, i.e., it type checks.
type ”owner”, ”() → User”
type ”owner=”, ”(User) → User”
Now consider executing this code. When Talk is loaded,
belongs to will be invoked, adding those type signatures to
the class. Then when owner? is called, Hummingbird will
perform type checking using currently available type information, and so it will be able to successfully type check the
body. Moreover, notice this approach is very flexible. Rails
does not require that belongs to be used at the beginning
of a class or even during this particular class definition. (In
Ruby, it is possible to “re-open” a class later on and add more
methods to it.) But no matter where the call occurs, it must
be before owner? is called so that owner is defined. Thus in
this case, Hummingbird’s typing strategy matches well with
Ruby’s semantics.
User-provided Type Signatures. In the examples so far,
the types for dynamically created methods could be determined automatically. However, consider Figure 3, which
shows an excerpt from CCT that uses Struct from the Ruby
core library. Line 1 creates a new class, instances of which
are defined to have getters type, account name, and amount,
Type Checking Dynamically Created Methods. In the previous example, we trusted Rails to dynamically generate
3
2016/4/14
values
expressions
v
e
premths
val typs
mth typs
b
τ
τm
::=
::=
|
|
::=
::=
::=
x ∈ var ids
dyn env
dyn cls tab
contexts
E
DT
C
stack
S
type env
Γ, ∆
type tab
TT
cache
X
typ chk deriv DM
subtyp deriv D≤
m ∈ mth ids
:
:
::=
|
::=
:
:
::=
::=
::=
inheritance from our formalism for simplicity, but they are
handled by our implementation.
Expressions e include values, variables x, the special variable self, assignments x = e, and sequencing e; e. Objects
are created with A.new. Conditional if e1 then e2 else e3
evaluates to e2 unless e1 evaluates to nil, in which case it
evaluates to e3 . Method invocation e1 .m(e2 ) is standard, invoking the m method based on the run-time type of e1 .
Expression def A.m = λx.e, defines method m of class
A as taking argument x and returning e. (We refer to λx.e
as a premethod.) This form allows methods to be defined
anywhere during execution, thus it combines the features
of Ruby’s def and define method. As in Ruby, if A.m is
already defined, def overwrites the previous definition. The
def expression itself evaluates to nil.
Finally, expression type A.m : τ → τ ′ asserts that
method m of class A has domain type τ and range type τ ′ .
Types may be either classes A or nil, the type of expression nil. The type expression overwrites the previous type of
A.m, if any. Like Hummingbird, there is no ordering dependency between def and type—the only requirement is that
a method’s type must be declared by the time the method is
called. The type expression itself evaluates to nil.
nil | [A]
v | x | self | x = e | e; e | A.new
if e then e else e | e.m(e)
def A.m = b | type A.m : τm
λx.e
A | nil
τ →τ
A ∈ cls ids
var ids → vals
cls ids → mth ids → premths
✷ | x = C | C.m(e) | v.m(C)
C; e | if C then e else e
· | (E , C) :: S
var ids → val typs
cls ids → mth ids → mth typs
cls ids → mth ids → DM × D≤
TT ⊢ hΓ, ei ⇒ hΓ′ , τ i
τ1 ≤ τ2
Figure 4. Source Language and Auxiliary Definitions.
Type Checking. Figure 5 gives the static type checking
rules. As in Hummingbird, static type checking is performed
at run time at method entry—thus these rules will be invoked
as a subroutine by the dynamic semantics (below). The bottom part of Figure 4 defines the sets and maps used in this
figure and in the dynamic semantics.
In these rules, TT is a type table mapping class and
method ids A.m to their corresponding types, as declared
by type, and Γ is a type environment mapping local variables to their types. These rules prove judgments of the form
TT ⊢ hΓ, ei ⇒ hΓ′ , τ i, meaning with type table TT , in
type environment Γ, expression e has type τ , and after evaluating e, the new type environment is Γ′ . Using an “output”
type environment Γ′ allows us to build a flow-sensitive type
system, in which variables’ types can change at assignments.
Note there is no output TT because the type table does not
change during static type checking—it only changes as the
program is executed by the dynamic semantics.
The type rules are largely standard. (TNil) and (TObject)
give nil and instances the obvious types. (TSelf) and (TVar)
give self and local variables their types according to the type
environment. Since none of these four expressions updates
the state, the output type environment is the same as the input
environment.
(TSeq) types sequencing, threading the type environment
from the output of e1 to the input of e2 . (TAssn) types an
assignment, updating the output type environment to bind
the assigned variable x to the type of the right-hand side.
(TNew) types object creation in the obvious way. (TDef)
trivially type checks method definitions. Notice we do not
type check the method body; that will happen at run time
and setters type=, account name=, and amount=. The
process transactions method iterates through instance field
@transactions (whose type is provided on line 8), and calls
the account name method of each one.
From line 1 we know the account name method exists,
but we do not know its type. Indeed, a “struct field” can hold
any type by default. Thus, to fully type check the body of
process transactions, we need more information from the
programmer to specify the type of account name.
The bottom part of Figure 3 defines a new method,
add types, that the programmer can call to indicate desired
struct field types. The types are given in the same order
as the constructor arguments, and the body of add types
uses zip to pair up the constructor arguments (retrieved via
members) and the types, and then iterates through the pairs,
creating the appropriate type signatures for the getters and
setters. The last line of the figure uses add types to create
type signatures for this example, allowing us to type check
process transactions when it is called.
In this particular case, we could have individually specified type signatures for the methods of Transaction. However, because Hummingbird lets programmers write arbitrary Ruby programs to generate types, we were able to develop this much more elegant solution.
3. Formalism
We formalize Hummingbird using the core, Ruby-like language shown at the top of Figure 4. Values v include nil,
which can be treated as if it has any type, and [A], which
is an instance of class A. Note that we omit both fields and
4
2016/4/14
TT ⊢ hΓ, ei ⇒ hΓ′ , τ i
(TNil)
(TObject)
(TSelf)
(TVar)
TT ⊢ hΓ, nili ⇒ hΓ, nili
TT ⊢ hΓ, [A]i ⇒ hΓ, Ai
TT ⊢ hΓ, selfi ⇒ hΓ, Γ(self)i
TT ⊢ hΓ, xi ⇒ hΓ, Γ(x)i
(TSeq)
TT ⊢ hΓ, e1 i ⇒ hΓ1 , τ1 i
TT ⊢ hΓ1 , e2 i ⇒ hΓ2 , τ2 i
TT ⊢ hΓ, e1 ; e2 i ⇒ hΓ2 , τ2 i
(TAssn)
TT ⊢ hΓ, ei ⇒ hΓ′ , τ i
(TDef)
TT ⊢ hΓ, x = ei ⇒ hΓ′ [x 7→ τ ], τ i
TT ⊢ hΓ, def A.m = λx.ei ⇒ hΓ, nili
(TType)
(TNew)
TT ⊢ hΓ, A.newi ⇒ hΓ, Ai
TT ⊢ hΓ, type A.m : τm i ⇒ hΓ, nili
(TApp)
TT ⊢ hΓ, e0 i ⇒ hΓ0 , Ai
TT ⊢ hΓ0 , e1 i ⇒ hΓ1 , τ i
TT (A.m) = τ1 → τ2
τ ≤ τ1
(TIf)
TT ⊢ hΓ, e0 i ⇒ hΓ′ , τ i
TT ⊢ hΓ′ , e1 i ⇒ hΓ1 , τ1 i
′
TT ⊢ hΓ , e2 i ⇒ hΓ2 , τ2 i
TT ⊢ hΓ, e0 .m(e1 )i ⇒ hΓ1 , τ2 i
TT ⊢ hΓ, if e0 then e1 else e2 i ⇒ hΓ1 ⊔ Γ2 , τ1 ⊔ τ2 i
Figure 5. Type Checking System.
when the method is actually called. (TType) type checks a
type expression, which has no effect during type checking.
Such expressions are only evaluated at run-time, when they
update the type table (see below).
One consequence of (TType) is that our type system forbids typing a method and then immediately calling it in the
same method body. For example, the following method body
would fail to type check:
1
2
3
4
def A.m = λx.
def B.m = ...;
type B.m : ...;
B.new.m
environment of the conditional is the least upper bound of
the output environments of the branches, defined as (Γ1 ⊔
Γ2 )(x) = Γ1 (x) ⊔ Γ2 (x) if x ∈ dom(Γ1 ) ∧ x ∈ dom(Γ2 )
and (Γ1 ⊔ Γ2 )(x) is undefined otherwise.
Dynamic Semantics. Figure 6 gives a small-step dynamic
semantics for our language. The semantics operates on dynamic configurations of the form hX , TT , DT , E , e, S i.
The first two components are the key novelties to support
run-time static type checking. X is a cache mapping A.m to
the type checking proofs for its method body (more details
below). TT is the type table, which is updated at run time
by calls to type. The last four components are standard. DT
is a dynamic class table mapping A.m to its premethod. E
is the dynamic environment mapping local variables to values. e is the expression being reduced. Lastly, S is a stack
of pairs (E , C), where E is the dynamic environment and C
is the evaluation context (defined in the usual way) at a call
site. The semantics pushes onto the stack at calls and pops
off the stack at returns.
The first seven rules in the semantics are standard. (ESelf)
and (EVar) evaluate self and variables by looking them up in
the environment. (EAssn) binds a variable to a value in the
environment. Notice that, like Ruby, variables can be written
without first declaring them, but it is an error to try to read
a variable that has not been written. (ENew) creates a new
instance. Note that since objects do not have fields, we do not
need a separate heap. (ESeq) discards the left-hand side of a
sequence if it has been fully evaluated. (EIfTrue) reduces to
the true branch if the guard is non-nil, and (EIfFalse) reduces
to the false branch otherwise.
The next four rules are the heart of just-in-time static type
checking. Our goal is to statically type check methods once
at the first call, and then avoid rechecking them unless something has changed. To formalize this notion, we define the
# define B.m
# give B.m a type
# type error ! B.m not in type table
Here we type check A.m’s body at the first call to it, so the
type expression has not been run—and hence has not bound
a type to B.m—yet. Thus it is a type error to invoke B.m in
the method body.
While we could potentially solve this problem with a
more complex type system, in our experience (Section 5) we
have not needed such a feature.
Next, (TApp) types method invocation e0 .m(e1 ), where
we look up the method’s type in TT based on the compiletime type of e0 . (Note that since there is no inheritance, we
need not search the inheritance hierarchy to find the type of
A.m.) Here subtyping is defined as nil ≤ A and A ≤ A for
all A. Thus, as is standard in languages with nil, the type
system may accept a program that invokes a non-existent
method of nil even though this is a run-time error. However,
notice that if e0 evaluates to a non-nil value, then (TApp)
guarantees e0 has method m.
Finally, (TIf) types conditionals. Like Ruby, the guard
e0 may have any type. The type of the conditional is the
least upper bound of the types of the two branches, defined
as A ⊔ A = A and nil ⊔ τ = τ ⊔ nil = τ . The output
5
2016/4/14
hX , TT , DT , E , e, S i → hX ′ , TT ′ , DT ′ , E ′ , e′ , S ′ i
hX , TT , DT , E , E (self), S i
hX , TT , DT , E , E (x), S i
hX , TT , DT , E [x 7→ v], v, S i
hX , TT , DT , E , [A], S i
hX , TT , DT , E , e2 , S i
hX , TT , DT , E , e1 , S i
if v 6= nil
hX , TT , DT , E , e2 , S i
hX \A.m, TT , DT [A.m 7→ λx.e], E , nil, S i
h(X \A.m)[TT ′ ], TT ′ , DT , E , nil, S i
TT ′ = TT [A.m 7→ τm ] and A.m 6∈ TApp(S)
(EAppMiss)
hX , TT , DT , E , C[v1 .m(v2 )], S i → hX ′ , TT , DT , [self 7→ v1 , x 7→ v2 ], e, (E , C) :: S i
if A.m 6∈ dom(X ) and v1 = [A] and DT (A.m) = λx.e and TT (A.m) = τ1 → τ2 and type of(v2 ) ≤ τ1 and
DM = (TT ⊢ h[x 7→ τ1 , self 7→ A], ei ⇒ hΓ′ , τ i) holds and D≤ = (τ ≤ τ2 ) holds and
X ′ = X [A.m 7→ (DM , D≤ )]
(EAppHit)
hX , TT , DT , E , C[v1 .m(v2 )], S i → hX , TT , DT , [self 7→ v1 , x 7→ v2 ], e, (E , C) :: S i
if A.m ∈ dom(X ) and v1 = [A] and DT (A.m) = λx.e and TT (A.m) = τ1 → τ2 and type of(v2 ) ≤ τ1
(ERet)
hX , TT , DT , E ′ , v, (E , C) :: S i → hX , TT , DT , E , C[v], S i
(ESelf)
(EVar)
(EAssn)
(ENew)
(ESeq)
(EIfTrue)
(EIfFalse)
(EDef)
(EType)
hX , TT , DT , E , self, S i
hX , TT , DT , E , x, S i
hX , TT , DT , E , x = v, S i
hX , TT , DT , E , A.new, S i
hX , TT , DT , E , (v; e2 ), S i
hX , TT , DT , E , if v then e1 else e2 , S i
hX , TT , DT , E , if nil then e1 else e2 , S i
hX , TT , DT , E , def A.m = λx.e, S i
hX , TT , DT , E , type A.m : τm , S i
→
→
→
→
→
→
→
→
→
(EContext)
hX , TT , DT , E , e, S i → hX ′ , TT ′ , DT ′ , E ′ , e′ , S ′ i
∄v1 , v2 , e′ . e = (v1 .m(v2 )) ∨ e = v1 ∨ e = C[e′ ]
hX , TT , DT , E , C[e], S i → hX ′ , TT ′ , DT ′ , E ′ , C[e′ ], S ′ i
Figure 6. Dynamic Semantics.
cache X as a map from A.m to a pair of typing derivations
(DM , D≤ ). Here DM is a type checking derivation from Figure 5 for the body of A.m, and D≤ is a subtyping judgment
showing that the type of e is a subtype of the declared return type. We need D≤ because our type system is syntaxdirected and hence does not include a standalone subsumption rule.
(EDef) reduces to nil, updating the dynamic class table
to bind A.m to the given premethod along the way. Recall
that we allow a method to be redefined with def. Hence we
need to invalidate anything in the cache relating to A.m so
that A.m will be checked the next time it is called. More
precisely:
derivations that refer to A.m. Hence, cached type derivations
that use TT can safely use TT ′ . Formally, we define:
Definition 2 (Cache upgrading). We write X [TT ′ ] to indicate a new cache that is the same as X , except the type table
in every derivation is replaced by TT ′ .
Thus, in (EType), the output cache is upgraded to the new
type table after invalidation.
The next two rules use the type cache. Both rules evaluate
a method call in a context, written C[v1 .m(v2 )]; we will
discuss the other rule for contexts shortly. In both rules, the
receiver v1 is a run-time object [A]. (EAppMiss) applies
when A.m is not in the cache. In this case, we look up
the type of A.m in TT , yielding some type τ1 → τ2 . We
type check the method body e in an environment in which
formal variable x is bound to τ1 and self is bound to A,
yielding a derivation DM . We check that the resulting type
τ of e is a subtype of the declared type e2 , with subtyping
derivation D≤ . Finally, we check that the run-time type of
v2 —defined as type of(nil) = nil and type of([A]) = A—
is a subtype of τ1 . If all this holds, then it is type-safe to call
the method. Hence we update the cache with the method’s
typing derivations and start evaluating the method body,
pushing the context C and the environment E on the stack.
(EAppHit) is similar but far simpler. This rule applies
when A.m is in the cache. In this case we know its method
body has been successfully type checked, so we need only
check that the run-time type of v2 is a subtype of the declared domain type of v1 . If so, we allow the method call to
proceed.
Definition 1 (Cache invalidation). We write X \A.m to indicate a new cache that is the same as X , except A.m has
been invalidated, meaning:
1. Any entries with A.m as the key are removed.
2. Any entries with a DM that apply (TApp) with A.m are
removed.
Thus, in (EDef), the output cache is the same as the input
cache but with A.m invalidated.
(EType) also reduces to nil, updating the type table to
be TT ′ , which is the same as TT but with new type information for A.m. As with (EDef), we invalidate A.m in
the cache. However, there is a another subtlety. Recall that
cached typing derivations DM include the type table TT .
This is potentially problematic, because we are changing the
type table to TT ′ . However, cache invalidation removes any
6
2016/4/14
However a method is called, the return, handled by
(ERet), is the same. This rule applies when an expression
has been fully evaluated and is at the top level. In this case,
we pop the stack, replacing E ′ with E from the stack and
plugging the value v into the context C from the stack.
Finally, (EContext) takes a step in an subexpression inside a context C. This rule only applies if the subexpression
is not a method call (since that case is handled by (EApp*),
which must push the context on the stack) and not a fully
evaluated value (which is handled by (ERet), which must
pop the context from the stack). We also do not allow the
subexpression to itself be a context, since that could cause
(EApp*) and (ERet) to misbehave.
RDL and Type Checking. Like standard RDL, Hummingbird’s type annotation stores type information in a map and
wraps the associated method to intercept calls to it. We
should emphasize that RDL does not perform any static
checking on its own—rather, it solely enforces contracts
dynamically. In Hummingbird, when a wrapped method is
called, Hummingbird first checks to see if it has already been
type checked. If not, Hummingbird retrieves the method’s
CFG and type and then statically checks that the CFG
matches the given type.
Hummingbird uses RDL’s type language, which includes
nominal types, intersection types, union types, optional
and variable length arguments, block (higher-order method)
types, singleton types, structural types, a self type, generics,
and types for heterogenous arrays and hashes. Hummingbird supports all of these kinds of types except structural
types, self types, heterogeneous collections, and some variable length arguments. In addition, Hummingbird adds support for both instance field types (as seen in Figure 3) and
class field types.
There is one slight subtlety in handling union types: If in
a method call the receiver has a union type, Hummingbird
performs type checking once for each arm of the union and
the unions the possible return types. For example if in call
e.m(. . .) the receiver has type A ∪ B, then Hummingbird
checks the call assuming e has type of A.m, yielding a return
type τA ; checks the call assuming B.m, yielding return type
τB ; and then sets the call’s return type to τA ∪ τB .
Soundness. Our type system forbids invoking non-existent
methods of objects. However, there are three kinds of errors
the type system does not prevent: invoking a method on nil;
calling a method whose body does not type check at run
time; and calling a method that has a type signature but is
itself undefined. (We could prevent the latter error by adding
a side condition to (TApp) that requires the method to be
defined, but we opt not to to keep the formalism slightly
simpler.) To state a soundness theorem, we need to account
for these cases, which we do by extending the dynamic
semantics with rules that reduce to blame in these three
cases. After doing so, we can state soundness:
Theorem 1 (Soundness). If ∅ ⊢ h∅, ei ⇒ hΓ′ , τ i then either
e reduces to a value, e reduces to blame, or e diverges.
We show soundness using a standard progress and preservation approach. The key technical challenge is preservation,
in which we need to show that not only are expression types
preserved, but also the validity of the cache and types of contexts pushed on the stack. The proof can be found in Appendix A.
Hummingbird is implemented using a combination of Ruby
and OCaml. On the OCaml side, we use the Ruby Intermediate Language (RIL) [13] to parse input Ruby files and translate them to control-flow graphs (CFG) on which we perform type checking. On the Ruby side, we extend RDL [26],
a contract system for Ruby, to perform static type checking. We next discuss the major challenges of implementing
Hummingbird.
Eliminating Dynamic Checks. Recall the (EApp*) rules
dynamically check that a method’s actual arguments have
the expected types before calling a statically typed method.
This check ensures that an untrusted caller cannot violate
the assumptions of a method. However, observe that if the
immediate caller is itself statically checked, then we know
the arguments are type-safe. Thus, as a performance optimization, Hummingbird only dynamically checks arguments
of statically typed methods if the caller is itself not statically checked. As a further optimization, Hummingbird also
does not dynamically check calls from Ruby standard library methods or the Rails framework, which are assumed to
be type-safe. The one exception is that Hummingbird does
dynamically check types for the Rails params hash, since
those values come from the user’s browser and hence are
untrusted.
RIL. RIL is essentially the front-end of Diamondback
Ruby (DRuby) [8, 14]. Given an input Ruby program, RIL
produces a CFG that simplifies away many of the tedious
features of Ruby, e.g., multiple forms of conditionals. We
modified DRuby so it emits the RIL CFG as a JSON file and
then exits. When loading each application file at run-time,
we read the corresponding JSON file and store a mapping
from class and method names and positions (file and line
number) to the JSON CFG. At run-time we look up CFGs in
this map to perform static type checking.
Numeric Hierarchy. Ruby has a Numeric tower that provides several related types for numbers. For example, Fixnum
< Integer < Numeric and Bignum < Integer < Numeric.
Adding two Fixnums normally results in another Fixnum,
but adding two large Fixnums could result in a Bignum in
the case of numeric overflow. To keep the type checking system simple, Hummingbird omits the special overflow case
and does not take Bignum into consideration. (This could
be addressed by enriching the type system [32].) Numeric
overflow does not occur in our experiments.
4. Implementation
7
2016/4/14
Code Blocks. As mentioned earlier, Ruby code blocks are
anonymous functions delimited by do. . .end. Hummingbird
allows methods that take code block arguments to be annotated with the block’s type. For example:
Second, by default Hummingbird gives instances of
generic classes their “raw” type with no type parameters.
To add parameters, we use type casts, as in the following
code:
type :m, ”() { (T) → U } → nil ”
a = []
# a has type Array
a. rdl cast (”Array<Fixnum>”) # cast to Array<Fixnum>
a.push(0)
# ok
a.push(”str”)
# type error due to cast
indicates that m takes no regular arguments; one code block
argument where the block takes type T and returns type U;
and m itself returns type nil.
There are two cases involving code blocks that we need
to consider. First, suppose Hummingbird is statically type
checking a call m() do |x| body end, and m has the type
given just above. Then at the call, Hummingbird statically
checks that the code block argument matches the expected
type, i.e., assuming x has type T, then body must produce
a value of type U. Second, when statically type checking m
itself, Hummingbird should check that calls to the block are
type correct. Currently this second case is unimplemented as
it does not arise in our experiments.
Recall from above that Hummingbird sometimes needs
to dynamically check the arguments to a statically typed
method. While this test is easy to do for objects, it is hard
to do for code blocks, which would require higher-order
contracts [11]. Currently Hummingbird does not implement
this higher order check, and simply assumes code block
arguments are type safe. Also, Hummingbird currently assumes the self inside a code block is the same as in the
enclosing method body. This assumption holds in our experiments, but it can be violated using instance eval and instance exec [33]. In the future, we plan to address this limitation by allowing the programmer to annotate the self type
of code blocks.
Here without the type annotation the last line would succeed;
with the annotation it triggers a type error. Note that when
casting an array or hash to a generic type, rdl cast iterates
through the elements to ensure they have the given type.
Modules. Ruby supports mixins via modules, which are
collections of methods that can be added to a class. Recall
that Hummingbird caches which methods have been statically type checked. Because a module can be mixed in to
multiple different classes—and can actually have different
types in those different classes—we need to be careful that
module method type checks are cached via where they are
mixed in rather than via the module name.
For example, consider the following code, where the
method foo defined in module M calls bar, which may vary
depending on where M is mixed in:
1
2
3
module M def foo(x) bar(x) end end
class C; include M; def bar(x) x + 1 end end
class D; include M; def bar(x) x. to s end end
Here method foo returns Fixnum when mixed into C and
String when mixed into D. Thus, rather that track the type
checking status of M#foo, Hummingbird separately tracks
the statuses of C#foo and D#foo.
Type Casts. While Hummingbird’s type system is quite
powerful, it cannot type check every Ruby program, and thus
in some cases we need to insert type casts. Hummingbird
includes a method o.rdl cast(t) that casts o’s type to t. After
such a call, Hummingbird assumes that o has type t. At runtime, the call dynamically checks that o has the given type.
In our experience, type casts have two main uses. First,
sometimes program logic dictates that we can safely downcast an object. For example, consider the following method
from one of our experiments:
Cache Invalidation. Recall from Section 3 that Hummingbird needs to invalidate portions of the cache when certain
typing assumptions change. While Hummingbird currently
does not support cache invalidation in general, it does support one important case. In Rails development mode, Rails
automatically reloads modified files without restarting, thus
redefining the methods in those files but leaving other methods intact [24]. In Rails development mode, Hummingbird
intercepts the Rails reloading process and performs appropriate cache invalidation. More specifically, when a method
is called, if there is a difference between its new and old
method body (which we check using the RIL CFGs), we invalidate the method and any methods that depend on it. We
also maintain a list of methods defined in each class, and
when a class is reloaded we invalidate dependencies of any
method that has been removed. In the next section, we report
on an experiment running a Rails app under Hummingbird
as it is updated.
We plan to add more general support for cache invalidation in future work. There are two main cases to consider.
The first is when a method is redefined or is removed (which
def self . load cache
f = datafile path ([‘‘ cache’ ’ , ‘‘ countries ’ ’ ])
t = Marshal.load( File . binread(f ))
@@cache ||= t. rdl cast (‘‘ Hash<String, %any>’’)
end
Marshal.load returns the result of converting its serialized
data argument into a Ruby object of arbitray type. However,
in our example, the argument passed to Marshal.load is
always an application data file that will be converted to the
annotated Hash.
8
2016/4/14
never happens in our experiments except in Rails development mode). Ruby provides two methods, method added
and method removed, that can be used to register callbacks
when the corresponding actions occur, which could be used
for cache invalidation.
The second case of cache invalidation is method’s type
changes. However, in RDL and Hummingbird, multiple calls
to type for the same method are used to create intersection
types. For example, the core library Array#[] method is
given its type with the following code:
1
2
3
which is an industrial strength web app framework that is
widely deployed; the next two use various metaprogramming styles in different ways than Rails; and the last one
does not use metaprogramming, as a baseline.
Table 1 summarizes the results of applying Hummingbird
to these apps. On the left we list the app name, version number or date (if no version number is available), and lines of
code as measured with sloccount [40]. For the Rails apps,
we ran sloccount on all ruby files in the model, controller,
helper, and mailer directories. We do not include lines of
code for views, as we do not type check views. For Countries and CCT, we ran sloccount on all files in the lib directory. For Rolify, we only statically checked several methods,
comprising 84 lines of code, that use define method in an
interesting way.
type Array, :[], ’ (Fixnum or Float) → t ’
type Array, :[], ’ (Fixnum, Fixnum) →Array<t>’
type Array, :[], ’ (Range<Fixnum>) →Array<t>’
meaning if given a Fixnum or Float, method Array#[] returns the array contents type; and, if given a pair of Fixnums
or a Range<Fixnum>, it returns an array.
In this setting, we cannot easily distinguish adding a
new arm of an intersection type from replacing a method
type. Moreover, adding a new arm to an intersection type
should not invalidate the cache, since the other arms are
still in effect. Thus, full support of cache invalidation will
likely require an explicit mechanism for replacing earlier
type definitions.
Type Annotations. For all apps, we used common type annotations from RDL for the Ruby core and standard libraries.
For several apps, we also added type annotations for thirdparty libraries and for the Rails framework. We trusted the
annotations for all these libraries, i.e., we did not statically
type check the library methods’ bodies.
We also added code to dynamically generate types for
metaprogramming code. For Rails, we added code to dynamically generate types for model getters and setters based
on the database schema; for finder methods such as find by name and find all by password (the method name indicates which field is being searched); and for Rails associations such as belongs to.
In Figure 2, we showed code we added to Rolify to generate types for a method created by calling define dynamic method. Calling define dynamic method also dynamically creates another method, is #{role name} of(arg)?,
which we also provide types for in the pre block.
In CCT, we used the code in Figure 3 to generate types
for Struct getters and setters.
Finally, we wrote type annotations for the app’s own
methods that were included in the lines of code count in Table 1. We marked those methods to indicate Hummingbird
should statically type check their bodies. Developing these
annotations was fairly straightforward, especially since we
could quickly detect incorrect annotations by running Hummingbird.
5. Experiments
We evaluated Hummingbird by applying it to six Ruby apps:
• Talks3 is a Rails app, written by the second author, for
publicizing talk announcements. Talks has been in use in
the UMD CS department since February 2012.
• Boxroom4 is a Rails implementation of a simple file shar-
ing interface.
• Pubs is a Rails app, developed several years ago by the
second author, for managing lists of publications.
• Rolify5 is a role management library for Rails. For this
evaluation, we integrated Rolify with Talks on the User
resource.
• Credit Card Transactions (CCT)6 is a library that per-
forms simple credit card processing tasks.
• Countries7 is an app that provides useful data about each
country.
Type Checking Results. For each program, we performed
type checking while running unit tests that exercised all the
type-annotated app methods. For Talks and Pubs, we wrote
unit tests with the goal of covering all application methods.
For Boxroom, we used its unit tests on models but wrote our
own unit tests on controllers, since it did not have controller
tests. For Rolify, we wrote a small set of unit tests for the
dynamic method definition feature. For CCT and Countries,
we used the unit tests that came with those apps.
In all cases, the app methods type check correctly in
Hummingbird; there were no type errors. The middle group
of columns summarizes more detailed type checking data.
We selected these apps for variety rather than for being representative. We chose these apps because their source code
is publicly available (except Pubs); they work with the latest versions of Ruby and RDL; and they do not rely heavily
on other packages. Moreover, the first three apps use Rails,
3 https://github.com/jeffrey-s-foster/talks
4 http://boxroomapp.com
5 https://github.com/RolifyCommunity/rolify
6 https://github.com/daino3/credit_card_transactions
7 https://github.com/hexorx/countries
9
2016/4/14
App
Talks-1/4/2013
Boxroom-1.7.1
Pubs-1/12/2015
Rolify-4.0.0
CCT-3/23/2014
Countries-1.1.0
LoC
1,055
854
620
84
172
227
Static types
Chk’d App All
111
201 363
127
221 306
47
86 171
14
24
71
23
27
75
33
40 111
Dynamic types
Gen’d
Used
990
45
534
93
445
33
26
2
6
3
0
0
Casts
31
17
13
15
6
22
Phs
1
1
1
12
1
1
Orig
162
263
72.0
5.63
3.06
1.02
Running time (s)
No$ Hum Or. Ratio
1,590
256
1.6×
705
327
1.2×
4,470
217
3.0×
7.79
6.71
1.2×
78.2
17.4
5.7×
18.1
4.62
4.5×
Table 1. Type checking results.
which dynamically defines other methods and adds their
type annotations. The other phases come from the order
in which the type annotation files are required—unlike the
Rails apps, the Rolify type annotation files are loaded piecemeal as the application loads.
The “Static types” columns report data on static type
annotations. The count under “Chk’d” is the number of type
annotations for the app’s methods whose bodies we statically
type checked. The count under “App” is that number plus
the number of types for app-specific methods with (trusted)
static type annotations, e.g., some Rails helper functions
have types that we do not currently dynamically generate.
The count under “All” reports the total number of static
type annotations we used in type checking each app. This
includes the “App” count plus standard, core, and third-party
library type annotations for methods referred to in the app.
The “Dynamic types” columns report the number of types
that were dynamically generated (“Gen’d”) and the number of those that were actually used during type checking
(“Used”). These numbers differ because we tried to make the
dynamic type information general rather than app-specific,
e.g., we generate both the getter and setter for belongs to
even if only one is used by the app.
These results show that having types for methods generated by metaprogramming is critical for successfully typing these programs—every app except Countries requires at
least a few, and sometimes many, such types.
The “Casts” column reports the number of type casts we
needed to make these programs type check; essentially this
measures how often Hummingbird’s type system is overly
conservative. The results show we needed a non-trivial but
relatively small number of casts. All casts were for the reasons discussed in Section 4: downcasting and generics.
The “Phs” column in Table 1 shows the number of type
checking phases under Hummingbird. Here a phase is defined as a sequence of type annotation calls with no intervening static type checks, followed by a sequence of static
type checks with no intervening annotations. We can see that
almost all apps have only a single phase, where the type annotations are executed before any static type checks. Investigating further, we found this is due to the way we added
annotations. For example, we set up our Rails apps so the
first loaded application file in turn loads all type annotation files. In practice the type annotations would likely be
spread throughout the app’s files, thus increasing the number of phases.
Rolify is the only application with multiple phases. Most
of the phases come from calling define dynamic method,
Performance. The last four columns of Table 1 report
the overhead of using Hummingbird. The “Orig” column
shows the running time without Hummingbird. The next
two columns report the running time with Hummingbird,
with caching disabled (“No$”) and enabled (“Hum”). The
last column lists the ratio of the “Hum” and “Orig” column.
For Talks, Boxroom, and Pubs, we measured the running
time of a client script that uses curl to connect to the web
server and exercise a wide range of functionality. For CCT,
we measured the time for running its unit tests 100 times. For
Countries and Rolify, we measured the time for running the
unit tests once (since these take take much more time than
CCT’s tests). For all apps, we performed each measurement
three times and took the arithmetic mean.
These results show that for the Rails apps, where IO
is significant, Hummingbird slows down performance from
24% to 201% (with caching enabled). We think these are
reasonable results for an early prototype that we have not
spent much effort optimizing. Moreover, across all apps, the
ratios are significantly better than prior systems that mix
static and dynamic typing for Ruby [18, 27], which report
orders of magnitude slowdowns.
Investigating further, we found that the main Hummingbird overhead arises from intercepting method calls to statically type checked methods. (Note the interception happens
regardless of the cache state.) The higher slowdowns for
CCT and Countries occur because those applications spend
much of their time in code with intercepted calls, while the
other applications spend most of their time in framework
code, whose methods are not intercepted. We expect performance can be improved with further engineering effort.
We can also see from the results that caching is an important performance optimization: without caching, performance slows down 1.4× to 62×. We investigated pubs, the
app with the highest no-caching slowdown, and found that
while running the application with large array inputs, certain
application methods are called more than 13,000 times while
10
2016/4/14
iterating through the large arrays. This means that each of
these application methods are statically type checked more
than 13,000 times when caching is disabled.
Version
5/14/12
7/24/12
8/24/12-1
8/24/12-2
8/24/12-3
9/14/12
1/4/13
Type Errors in Talks. We downloaded many earlier versions of Talks from its github repository and ran Hummingbird on them using mostly the same type annotations as
for the latest version, changed as necessary due to program
changes. Cumulatively, we found six type errors that were
introduced and later removed as Talks evolved. Below the
number after the date indicates which checkin it was, with 1
for the first checkin of the day, 2 for the second, etc.
∆ Meth
N/A
1
8
1
1
4
Added
N/A
2
1
1
-
Deps
N/A
4
8
-
Chk’d
77
15 / 5
24 / 14
11 / 1
12 / 2
15 / 1
13 / 4
Table 2. Talks Update Results
• 1/8/12-4: This version misspells compute edit fields as
time of Talks in Table 1 ; update the code to the next version;
and repeat.
Table 2 shows the results of our experiment. The “∆
Meth” column lists the number of methods whose bodies or
types were changed compared to the previous version. Note
there are no removed methods in any of these versions. The
“Added” column lists the number of methods added; such
methods will be checked when they are called for the first
time but do not cause any cache invalidations. The “Deps”
column counts the number of dependent methods that call
one or more of the changed methods. These methods plus the
changed methods are those whose previous static type check
are invalidated by the update. The last column, “Chk’d,” reports how many methods are newly or re-type checked after
the update. Currently, Hummingbird always rechecks Rails
helper methods, due to a quirk in the Rails implementation—
the helper methods’ classes get a new name each time the
helper file is reloaded, causing Hummingbird to treat their
methods as new. Thus (except for the first line, since this issue does not arise on the first run), we list two numbers in
the column: the first with all rechecks, including the helper
methods, and the second excluding the helper methods.
These results show that in almost all cases, the second
number in “Chk’d” is equal to the sum of the three previous columns. There is one exception: in 8/24/12/-1, there 14
rechecked methods but 18 changed/added/dependent methods. We investigated and found that the 14 rechecks are composed of six changed methods that are rechecked once; two
changed methods that are rechecked twice because they have
dependencies whose updates are interleaved with calls to
those methods; one added method that is checked; and three
dependent methods that are rechecked. The remaining added
method is not called by the curl scripts, and the remaining
dependent methods are also changed methods (this is the
only version where there is overlap between the changed and
dependent columns).
Finally, as there are no type errors in this sequence of
updates, we confirmed that this streak of updates type checks
under Hummingbird.
copute edit fields. Hummingbird reported this error because the latter was an unbound local variable and was
also not a valid method.
• 1/7/12-5: Instead of calling @list.talks.upcoming.sort{|
a, b | ...}, this version calls @list.talks.upcoming{| a,
b | ...} (leaving off the sort). Hummingbird detects this
error because upcoming’s type indicates it does not take
a block. Interestingly, this error would not be detected
at run-time by Ruby, which simply ignores unused block
arguments.
• 1/26/12-3: This version calls user.subscribed talks(true),
but subscribed talks’s argument is a Symbol.
• 1/28/12: This version calls @job.handler.object, but
@job.handler returns a String, which does not have an
object method.
• 2/6/12-2: This version uses undefined variable old talk.
Thus, Hummingbird assumes old talk is a no-argument
method and attempts to look up its type, which does not
exist.
• 2/6/12-3: This version uses undefined variable new talk
We should emphasize that although we expected there
would be type errors in Talks, we did not know exactly what
they were or what versions they were in. While the second
author did write Talks, the errors were made a long time ago,
and the second author rediscovered them independently by
running Hummingbird.
Updates to Talks Finally, we performed an experiment in
which we launched one version of Talks in Rails development mode and then updated the code to the next six consecutive versions of the app. (We skipped versions in which
none of the Ruby application files changed) Notice that
cache invalidation is particular useful here, since in typical
usage only a small number of methods are changed by each
update.
In more detail, after launching the initial version of the
app, we repeated the following sequence six times: Reset
the database (so that we run all versions with the same initial
data); run a sequence of curl commands that access the same
Talks functionalities as the ones used to measure the running
6. Related Work
There are several threads of related work.
11
2016/4/14
Type Systems for Ruby. We have developed several prior
type systems for Ruby. Diamondback Ruby (DRuby) [14]
is the first comprehensive type system for Ruby that we are
aware of. Because Hummingbird checks types at run-time,
we opted to implement our own type checker rather than
reuse DRuby for type checking, which would have required
some awkward shuffling of the type table between Ruby and
OCaml. Another reason to reimplement type checking was
to keep the type system a little easier to understand—DRuby
performs type inference, which is quite complex for this type
language, in contrast to Hummingbird, which implements
much simpler type checking.
DRuby was effective but did not handle highly dynamic
language constructs well. PRuby [12] solves this problem
using profile-based type inference. To use PRuby, the developer runs the program once to record dynamic behavior, e.g.,
what methods are invoked via send, what strings are passed
to eval, etc. PRuby then applies DRuby to the original program text plus the profiled strings, e.g., any string that was
passed to eval is parsed and analyzed like any other code.
While PRuby can be effective, we think that Hummingbird’s
approach is ultimately more practical because Hummingbird
does not require a separate, potentially cumbersome, profiling phase. We note that Hummingbird does not currently
handle eval, because it was not used in our subject apps’
code, but it could be supported in a straightforward way.
We also developed DRails [17], which type checks Rails
apps by applying DRuby to translated Rails code. For example, if DRails sees a call to belongs to, it outputs Ruby
code that explicitly contains the methods generated from the
call, which DRuby can then analyze. While DRails was applied to a range of programs, its analysis is quite brittle.
Supporting each additional Rails feature in DRails requires
implementing, in OCaml, a source-to-source transformation
that mimics that feature. This is a huge effort and is hard
to sustain as Rails evolves. In contrast, Hummingbird types
are generated in Ruby, which is far easier. DRails is also
complex to use: The program is combined into one file, then
run to gather profile information, then transformed and type
checked. Using Hummingbird is far simpler. Finally, DRails
is Rails-specific, whereas Hummingbird applies readily to
other Ruby frameworks. Due to all these issues, we feel
Hummingbird is much more lightweight, agile, scalable, and
maintainable than DRails.
Finally, RubyDust [18] implements type inference for
Ruby at run time. RubyDust works by wrapping objects
to annotate them with type variables. More precisely, consider a method def m(x) ... end, and let α be the type variable for x. RubyDust’s wrapping is approximately equal to
adding x = Wrap.new(x, α) to the beginning of m. Uses of
the wrapped x generate type constraints on α and then delegate to the underlying object. The Ruby Type Checker [27]
(rtc) is similar but implements type checking rather than type
inference.
Hummingbird has several important advantages over
RubyDust and rtc. First, RubyDust and rtc can only report
errors on program paths they observe. In contrast, Hummingbird type checks all paths through methods it analyzes.
Second, wrapping every object with a type annotation is
extremely expensive. By doing static analysis, Hummingbird avoids this overhead. Finally, RubyDust and rtc have
no special support for metaprogramming. In RubyDust, dynamically created methods could have their types inferred
in a standard way, though RubyDust would likely not infer useful types for Rails-created methods. In rtc, dynamically created methods would lack types, so their use would
not be checked. (Note that it would certainly be possible
to add Hummingbird-style support for metaprogramminggenerated type annotations to either RubyDust or rtc.) In
sum, we think that Hummingbird strikes the right compromise between the purely static DRuby approach and the
purely dynamic RubyDust/rtc approach.
Type Systems for Other Dynamic Languages. Many researchers have proposed type systems for dynamic languages, including Python [3], JavaScript [2, 21, 36], Racket [32,
37, 38], and Lua [22], or developed new dynamic languages
or dialects with special type systems, such as Thorn [5],
TypeScript [4, 25], and Dart [7]. To our knowledge, these
type systems are focused on checking the core language and
can have difficulty in the face of metaprogramming.
One exception is RPython [1], which introduces a notion
of load time, during which highly dynamic features may
be used, and run time, when they may not be. In contrast,
Hummingbird does not need such a separation.
Lerner et al [20] propose a system for type checking
programs that use JQuery, a very sophisticated Javascript
framework. The proposed type system has special support
for JQuery’s abstractions, making it quite effective in that
domain. On the other hand, it does not easily apply to other
frameworks.
Feldthaus et al’s TSCHECK [10] is a tool to check the
correctness of TypeScript interfaces for JavaScript libraries.
TSCHECK discovers a library’s API by taking a snapshot
after executing the library’s top-level code. It then performs
checking using a separate static analysis. This is similar
to Hummingbird’s tracking of type information at run-time
and then performing static checking based on it. However,
Hummingbird allows type information to be generated at any
time and not just in top-level code.
Related Uses of Caching. Several researchers have proposed systems that use caching in a way related to Hummingbird. Koukoutos et al [19] reduce the overhead of
checking data structure contracts (e.g., “this is a binary
search tree”) at run time by modifying nodes to hold key verification properties. This essentially caches those properties.
However, because the properties are complex, the process of
caching them is not automated.
12
2016/4/14
Stulova et al [34] propose memoizing run-time assertion
checking to improve performance. This is similar to Hummingbird’s type check caching, but much more sophisticated
because the cached assertions arise from a rich logic.
Hermenegildo et al [16] proposed a method to incrementally update analysis results at run-time as code is added,
deleted, or changed. Their analysis algorithms are designed
for constraint logic programming languages, and are much
more complicated than Hummingbird’s type checking.
l/F# features and Hummingbird is that Ruby does not have a
separate compile time.
7. Conclusion
We presented Hummingbird, a novel tool that type checks
Ruby apps using an approach we call just-in-time static type
checking. Hummingbird works by tracking type information
dynamically, but then checking method bodies statically at
run time as each method is called. As long as any metaprogramming code is extended to generate types as it creates
methods, Hummingbird will, in a very natural way, be able
to check code that uses the generated methods. Furthermore,
Hummingbird can cache type checking so it need not be unnecessarily repeated at later calls to the same method.
We formalized Hummingbird using a core, Ruby-like language that allows methods and their types to be defined at
arbitrary (and arbitrarily separate) points during execution,
and we proved type soundness. We implemented Hummingbird on top of RIL, for parsing Ruby source code, and RDL,
for intercepting method calls and storing type information.
We applied Hummingbird to six Ruby apps, some of which
use Rails. We found that Hummingbird’s approach is effective, allowing it to successfully type check all the apps even
in the presence of metaprogramming. We ran Hummingbird
on earlier versions of one app and found several type errors. Furthermore, we ran Hummingbird while applying a
sequence of updates to a Rails app in development mode
to demonstrate cache invalidation under Hummingbird. Finally, we measured Hummingbird’s run-time overhead and
found it is reasonable.
In sum, we think that Hummingbird takes a strong step
forward in bringing static typing to dynamic languages.
Staged Analysis. MetaOCaml [35] is a multi-stage extension of OCaml in which code is compiled in one stage and
executed in a later stage. The MetaOCaml compiler performs static type checking on any such delayed code, which
is similar to Hummingbird’s just-in-time type checking. A
key difference between MetaOCaml and Hummingbird is
that Ruby programs do not have clearly delineated stages.
Chugh et al’s staged program analysis [6] performs static
analysis on as much code as is possible at compile time, and
then computes a set of remaining checks to be performed at
run time. Hummingbird uses a related idea in which no static
analysis is performed at compile time, but type checking is
always done when methods are called. Hummingbird is simpler because it need not compute which checks are necessary, as it always does the same kind of checking.
Other. Several researchers have explored other ways to
bring the benefits of static typing to dynamic languages.
Contracts [11] check assertions at function or method entry
and exit. In contrast, Hummingbird performs static analysis
of method bodies, which can find bugs on paths before
they are run. At the same time, contracts can encode richer
properties than types.
Gradual typing [31] lets developers add types gradually
as programs evolve; Vitousek et al recently implemented
gradual typing for Python [39]. Like types [41] bring some
of the flexibility of dynamic typing to statically typed languages. The goal of these systems is to allow mixing of
typed and untyped code. This is orthogonal to Hummingbird, which focuses on checking code with type annotations.
Richards et al [28, 29] have explored how highly dynamic
language features are used in JavaScript. They find such features, including eval, are used extensively in a wide variety
of ways, including supporting metaprogramming.
The GHC Haskell compiler lets developers defer type errors until run-time to suppress type errors on code that is
never actually executed [15]. Hummingbird provides related
behavior in that a method that is never called will never be
type checked by Hummingbird. Template Haskell [30] can
be used for compile-time metaprogramming. Since Haskell
programs contain types, template Haskell is often used to
generate type annotations, analogously to the type annotations generated using Hummingbird. Similarly, F# type
providers [9] allow users to create compile time types, properties and methods. A key difference between these Haskel-
Acknowledgments
Thanks to ThanhVu Nguyen and the anonymous reviewers
for their helpful comments. This research was supported in
part by NSF CCF-1319666 and Subcontract to Northeastern
University, NSF CCF-1518844.
References
[1] D. Ancona, M. Ancona, A. Cuni, and N. D. Matsakis.
RPython: A Step Towards Reconciling Dynamically and Statically Typed OO Languages. In Proceedings of the 2007 Symposium on Dynamic Languages, DLS ’07, pages 53–64, New
York, NY, USA, 2007. ACM. ISBN 978-1-59593-868-8. doi:
10.1145/1297081.1297091.
[2] C. Anderson, P. Giannini, and S. Drossopoulou. Towards Type
Inference for Javascript. In Proceedings of the 19th European
Conference on Object-Oriented Programming, ECOOP’05,
pages 428–452, Berlin, Heidelberg, 2005. Springer-Verlag.
ISBN 3-540-27992-X, 978-3-540-27992-1. doi: 10.1007/
11531142 19.
[3] J. Aycock. Aggressive Type Inference.
Python Conference, 2000.
13
In International
2016/4/14
ence on Automated Software Engineering (ASE), pages 590–
594, Auckland, New Zealand, November 2009. Short paper.
[4] G. Bierman, M. Abadi, and M. Torgersen. Understanding
typescript. In ECOOP 2014–Object-Oriented Programming,
pages 257–281. Springer, 2014.
[18] J. hoon (David) An, A. Chaudhuri, J. S. Foster, and M. Hicks.
Dynamic Inference of Static Types for Ruby. In ACM
SIGPLAN-SIGACT Symposium on Principles of Programming Languages (POPL), pages 459–472, Austin, TX, USA,
January 2011. doi: 10.1145/1926385.1926437.
[5] B. Bloom, J. Field, N. Nystrom, J. Östlund, G. Richards,
R. Strniša, J. Vitek, and T. Wrigstad. Thorn: Robust, Concurrent, Extensible Scripting on the JVM. In Proceedings of
the 24th ACM SIGPLAN Conference on Object Oriented Programming Systems Languages and Applications, OOPSLA
’09, pages 117–136, New York, NY, USA, 2009. ACM. ISBN
978-1-60558-766-0. doi: 10.1145/1640089.1640098.
[19] E. Koukoutos and V. Kuncak. Checking Data Structure Properties Orders of Magnitude Faster. In Runtime Verification 5th International Conference, RV 2014, Toronto, ON, Canada,
September 22-25, 2014. Proceedings, pages 263–268, 2014.
doi: 10.1007/978-3-319-11164-3 22.
[6] R. Chugh, J. A. Meister, R. Jhala, and S. Lerner. Staged Information Flow for Javascript. In Proceedings of the 30th
ACM SIGPLAN Conference on Programming Language Design and Implementation, PLDI ’09, pages 50–62, New York,
NY, USA, 2009. ACM. ISBN 978-1-60558-392-1. doi:
10.1145/1542476.1542483.
[20] B. S. Lerner, L. Elberty, J. Li, and S. Krishnamurthi. Combining Form and Function: Static Types for JQuery Programs.
In Proceedings of the 27th European Conference on ObjectOriented Programming, ECOOP’13, pages 79–103, Berlin,
Heidelberg, 2013. Springer-Verlag. ISBN 978-3-642-390371. doi: 10.1007/978-3-642-39038-8 4.
[7] Dart. Dart, 2015. https://www.dartlang.org.
[8] DRuby.
Diamondback
Ruby,
http://www.cs.umd.edu/projects/PL/druby/.
2009.
[21] B. S. Lerner, J. G. Politz, A. Guha, and S. Krishnamurthi. TeJaS: Retrofitting Type Systems for JavaScript. In Proceedings
[9] F
#.
Type
Provider,
2016.
https://msdn.microsoft.com/en-us/library/hh156509.aspx. of the 9th Symposium on Dynamic Languages, DLS ’13, pages
1–16, New York, NY, USA, 2013. ACM. ISBN 978-1-4503[10] A. Feldthaus and A. Møller. Checking Correctness of Type2433-5. doi: 10.1145/2508168.2508170.
Script Interfaces for JavaScript Libraries. In Proceedings of
[22] A. M. Maidl, F. Mascarenhas, and R. Ierusalimschy. Typed
the 2014 ACM International Conference on Object Oriented
Lua: An Optional Type System for Lua. In Proceedings of the
Programming Systems Languages & Applications, OOPSLA
Workshop on Dynamic Languages and Applications, Dyla’14,
’14, pages 1–16, New York, NY, USA, 2014. ACM. ISBN
pages 3:1–3:10, New York, NY, USA, 2014. ACM. ISBN
978-1-4503-2585-1. doi: 10.1145/2660193.2660215.
978-1-4503-2916-3. doi: 10.1145/2617548.2617553.
[11] R. B. Findler and M. Felleisen. Contracts for Higher-order
[23]
RailsGuides.
Active Record Associations, 2015.
Functions. In Proceedings of the Seventh ACM SIGPLAN
http://guides.rubyonrails.org/association_basics.html.
International Conference on Functional Programming, ICFP
’02, pages 48–59, New York, NY, USA, 2002. ACM. ISBN
[24] RailsGuides. Autoloading and Reloading Constants, 2016.
1-58113-487-8. doi: 10.1145/581478.581484.
http://guides.rubyonrails.org/autoloading_and_reloading_con
[12] M. Furr, J. hoon (David) An, and J. S. Foster. Profile[25] A. Rastogi, N. Swamy, C. Fournet, G. Bierman, and P. Vekris.
Guided Static Typing for Dynamic Scripting Languages. In
Safe & Efficient Gradual Typing for TypeScript. In ProceedACM SIGPLAN International Conference on Object-Oriented
ings of the 42Nd Annual ACM SIGPLAN-SIGACT Symposium
Programming, Systems, Languages and Applications (OOPon Principles of Programming Languages, POPL ’15, pages
SLA), pages 283–300, Orlando, Floria, October 2009. doi:
167–180, New York, NY, USA, 2015. ACM. ISBN 978-110.1145/1639949.1640110. Best student paper award.
4503-3300-9. doi: 10.1145/2676726.2676971.
[13] M. Furr, J. hoon (David) An, J. S. Foster, and M. Hicks.
[26] RDL. RDL, 2015. https://github.com/plum-umd/rdl.
The Ruby Intermediate Langauge. In Dynamic Languages
[27] B. M. Ren, J. Toman, T. S. Strickland, and J. S. Foster.
Symposium (DLS), pages 89–98, Orlando, Florida, October
The Ruby Type Checker. In Object-Oriented Program Lan2009. doi: 10.1145/1837513.1640148.
guages and Systems (OOPS) Track at ACM Symposium on
[14] M. Furr, J. hoon (David) An, J. S. Foster, and M. Hicks.
Applied Computing, pages 1565–1572, Coimbra, Portugal,
Static Type Inference for Ruby. In Object-Oriented Program
March 2013. doi: 10.1145/2480362.2480655.
Languages and Systems (OOPS) Track at ACM Symposium
[28] G. Richards, S. Lebresne, B. Burg, and J. Vitek. An Analysis
on Applied Computing (SAC), pages 1859–1866, Honolulu,
of the Dynamic Behavior of JavaScript Programs. In ProceedHawaii, March 2009. doi: 10.1145/1529282.1529700.
ings of the 31st ACM SIGPLAN Conference on Programming
[15] GHCLanguageFeatures.
DeferLanguage Design and Implementation, PLDI ’10, pages 1–12,
ring
Type
Errors
to
Runtime,
2016.
New York, NY, USA, 2010. ACM. ISBN 978-1-4503-0019-3.
https://downloads.haskell.org/~ ghc/latest/docs/html/users_guide/defer-type-errors.html.
doi: 10.1145/1806596.1806598.
[16] M. Hermenegildo, G. Puebla, K. Marriott, and P. J. Stuckey.
[29] G. Richards, C. Hammer, B. Burg, and J. Vitek. The Eval That
Incremental Analysis of Constraint Logic Programs. ACM
Men Do: A Large-scale Study of the Use of Eval in Javascript
Trans. Program. Lang. Syst., 22(2):187–223, Mar. 2000. ISSN
Applications. In Proceedings of the 25th European Con0164-0925. doi: 10.1145/349214.349216.
ference on Object-oriented Programming, ECOOP’11, pages
[17] J. hoon (David) An, A. Chaudhuri, and J. S. Foster. Static Typing for Ruby on Rails. In IEEE/ACM International Confer-
52–78, Berlin, Heidelberg, 2011. Springer-Verlag. ISBN 9783-642-22654-0.
14
2016/4/14
We show soundness by first showing preservation and
progress. As is typical, the hardest part of the proof is preservation, which shows that an expression’s type is preserved
under a step in the dynamic semantics. To make the theorem
work, we also need to reason about preserving key properties about the typing environment, run-time stack, and cache.
Here is the statement of the theorem, which we explain in
detail next:
[30] T. Sheard and S. P. Jones. Template Meta-programming for
Haskell. SIGPLAN Not., 37(12):60–75, Dec. 2002. ISSN
0362-1340. doi: 10.1145/636517.636528.
[31] J. G. Siek and W. Taha. Gradual Typing for Functional Languages. In Scheme and Functional Programming Workshop,
2006.
[32] V. St-Amour, S. Tobin-Hochstadt, M. Flatt, and M. Felleisen.
Typing the Numeric Tower. In Proceedings of the 14th
International Conference on Practical Aspects of Declarative Languages, PADL’12, pages 289–303, Berlin, Heidelberg, 2012. Springer-Verlag. ISBN 978-3-642-27693-4. doi:
10.1007/978-3-642-27694-1 21.
Theorem 2 (Preservation). If
(1)
(2)
(3)
(4)
(5)
(6)
[33] T. S. Strickland, B. Ren, and J. S. Foster. Contracts for
Domain-Specific Languages in Ruby. In Dynamic Languages
Symposium (DLS), Portland, OR, October 2014. doi: 10.1145/
2661088.2661092.
[34] N. Stulova, J. F. Morales, and M. V. Hermenegildo. Practical
Run-time Checking via Unobtrusive Property Caching. CoRR,
abs/1507.05986, 2015.
Then there exist ∆, ∆′ , TS ′ , τ ′ such that
(a)
(b)
(c)
(d)
(e)
(f)
[35] W.
Taha.
MetaOcaml,
2016.
http://www.cs.rice.edu/~ taha/MetaOCaml.
[36] P. Thiemann. Towards a Type System for Analyzing Javascript
Programs. In Proceedings of the 14th European Conference on Programming Languages and Systems, ESOP’05,
pages 408–422, Berlin, Heidelberg, 2005. Springer-Verlag.
ISBN 3-540-25435-8, 978-3-540-25435-5. doi: 10.1007/
978-3-540-31987-0 28.
TT ′ ⊢ h∆, e′ i ⇒ h∆′ , τ ′ i
τ ′ ≤ TS ′
If S = S ′ then ∆′ ≤ Γ′
∆ ∼ E′
TT ′ ⊢ TS ′ ∼ S ′
X ′ ∼ (TT ′ , DT ′ )
Let’s step through the assumptions and conclusions of the
theorem. (1) and (2) are standard—they assume that e takes
a step and is well-typed, respectively. The corresponding
conclusion (a) states that e′ is also well-typed.
(4) assumes the type and dynamic environments are
consistent—meaning values in E have the corresponding
types in Γ—and conclusion (d) states that they are still consistent after reduction. Formally:
[37] S. Tobin-Hochstadt and M. Felleisen. The Design and Implementation of Typed Scheme. In Proceedings of the 35th
Annual ACM SIGPLAN-SIGACT Symposium on Principles of
Programming Languages, POPL ’08, pages 395–406, New
York, NY, USA, 2008. ACM. ISBN 978-1-59593-689-9. doi:
10.1145/1328438.1328486.
[38] S. Tobin-Hochstadt and M. Felleisen. Logical Types for Untyped Languages. In Proceedings of the 15th ACM SIGPLAN
International Conference on Functional Programming, ICFP
’10, pages 117–128, New York, NY, USA, 2010. ACM. ISBN
978-1-60558-794-3. doi: 10.1145/1863543.1863561.
Definition 3 (Environment consistency). Type environment
Γ is consistent with dynamic environment E , written Γ ∼ E ,
if dom(Γ) ⊆ dom(E ) and for all x ∈ dom(Γ) there exists τ
such that · ⊢ hΓ, E (x)i ⇒ hΓ, τ i and τ ≤ Γ(x).
[39] M. M. Vitousek, A. M. Kent, J. G. Siek, and J. Baker. Design
and Evaluation of Gradual Typing for Python. In Proceedings
of the 10th ACM Symposium on Dynamic Languages, DLS
’14, pages 45–56, New York, NY, USA, 2014. ACM. ISBN
978-1-4503-3211-8. doi: 10.1145/2661088.2661101.
[40] D.
A.
Wheeler.
SLOCCount,
http://www.dwheeler.com/sloccount.
hX , TT , DT , E , e, S i → hX ′ , TT ′ , DT ′ , E ′ , e′ , S ′ i
TT ⊢ hΓ, ei ⇒ hΓ′ , τ i
τ ≤ TS
Γ∼E
TT ⊢ TS ∼ S
X ∼ (TT , DT )
Notice this definition allows E to include some variables
that are not bound in Γ. This is necessary to handle (TIf),
which discards any variables from the type environment that
are bound in one arm of the conditional but not the other.
Next, (3) and (5) concern the type of e and the stack.
The goal of preservation is to show e’s type is preserved, but
consider (EApp*) and (ERet). These rules both push and pop
the stack and change the expression being evaluated—hence
e′ could potentially have an entirely different type than e.
Our solution is to introduce the notion of a type stack
TS to mirror the run-time stack. To understand how the
type stack works, suppose we want to apply preservation to
C[v1 .m(v2 )], i.e., we are about to call a method. The typing judgment is TT ⊢ hΓ, C[v1 .m(v2 )i ⇒ hΓ′ , τ ′ i. In the
dynamic semantics, the (EApp*) rules will push the current
environment E and the context C on the stack. Correspondingly, we will push the current typing judgment onto the type
2015.
[41] T. Wrigstad, F. Z. Nardelli, S. Lebresne, J. Östlund, and
J. Vitek. Integrating Typed and Untyped Code in a Scripting Language. In Proceedings of the 37th Annual ACM
SIGPLAN-SIGACT Symposium on Principles of Programming Languages, POPL ’10, pages 377–388, New York, NY,
USA, 2010. ACM. ISBN 978-1-60558-479-9. doi: 10.1145/
1706299.1706343.
A. Complete Formalism
This section contains the full definitions and proofs for the
formalism.
15
2016/4/14
stack—at least the key pieces of it. More specifically, we
push an element of the form (Γ[τ ], hΓ′ , τ ′ i), where Γ and Γ′
are the initial and final environments of the current typing
judgment; C is the context; and τ is the type of expression
v1 .m(v2 ), i.e., the type that the method must return.
Given this mechanism, the key invariant to maintain is
that the type of the expression is compatible with what the
calling functions expects. We define:
frame is popped when a method returns. Hence in this case
the antecedent of (c) is false, and the conclusion is trivial.
Finally, we need to reason about the cache. As we
saw earlier, the key cache invariant to preserve is that all
the derivations stored in the cache hold and apply to the
premethod stored in DT and the type stored in TT . Formally:
Definition 7 (Cache consistency). We say that cache X is
consistent with type class table TT and dynamic class table
DT , written X ∼ (TT , DT ), if for all A.m ∈ dom(X )
where X (A.m) = (DM , D≤ ), with DM = (TT ⊢ h[x 7→
τ1 , self 7→ A], ei ⇒ hΓ′ , τ i) and D≤ = (τ ≤ τ2 ), it is
the case that DM and D≤ hold and DT (A.m) = λx.e and
TT (A.m) = τ1 → τ2 .
Definition 4 (Stack subtyping). τ0 ≤ (Γ[τ ], hΓ′ , τ ′ i) :: TS
if τ0 ≤ τ .
Then (3) assumes that the type of e is a subtype of the
type expected by the calling function. (At the top-level, we
initialize the type stack with a frame that expects whatever
the top-level type is.) (b) states that the type of e′ is also a
subtype of the type expected by its calling function. Thus, if
the stack does not change, this means that e′ and e have the
same type (up to subtyping). If the stack does change, then
we still maintain the invariant.
Of course, we need this invariant to hold no matter how
many pushes and pops happen. Thus, rather than only talk
about the top element of the type stack, we need to ensure
that all elements of the type stack are consistent with all
elements of the dynamic stack. Formally:
Thus, (6) assumes the cache is consistent, and (f) concludes the new cache is also consistent.
To show preservation, we also need a few lemmas:
Lemma 1. For all Γ1 and Γ2 , it is the case that Γ1 ≤
(Γ1 ⊔ Γ2 ).
Lemma 2 (Contextual Substitution). If
TT ⊢ hΓ, ei ⇒ hΓ, τ ′ i
..
.
Definition 5 (Stack consistency). Type stack element
(Γ[τ ], hΓ′ , τ ′ i) is consistent with dynamic stack element
(E, C), written TT ⊢ (Γ[τ ], hΓ′ , τ ′ i) ∼ (E, C), if Γ ∼ E
and TT ⊢ hΓ[✷ 7→ τ ], Ci ⇒ hΓ′ , τ ′ i. (Here we abuse
notation and treat ✷ as if it’s a variable.)
Type stack TS is consistent with dynamic stack S , written
TT ⊢ TS ∼ S , is defined inductively as
TT ⊢ hΓC , C[e]i ⇒ hΓ′C , τC i
then TT ⊢ hΓC [✷ 7→ τ ′ ], Ci ⇒ hΓ′C , τC i.
Lemma 3 (Substitution). If
1. TT ⊢ h∆[✷ 7→ τC ], Ci ⇒ h∆′ , τC′ i
2. TT ⊢ h·, vi ⇒ h·, τ i
3. τ ≤ τC
1. TT ⊢ · ∼ ·
2. TT ⊢ (Γ[τ ], hΓ′ , τ ′ i) :: TS ∼ (E, C) :: S if
(a) (Γ[τ ], hΓ′ , τ ′ i) ∼ (E, C)
(b) TS ∼ S
(c) τ ′ ≤ TS if TS 6= ·
Then TT ⊢ h∆, C[v]i ⇒ h∆′ , τC′′ i where τC′′ ≤ τC′ .
Finally, we can prove preservation:
Proof. (Preservation) By induction on hX , TT , DT , E , e, S i →
hX ′ , TT ′ , DT ′ , E ′ , e′ , S ′ i.
Thus, (5) assumes the type and dynamic stacks are consistent, and (e) concludes they remain consistent after taking
a step.
Next, (c) relates the output environment of e′ with the
output environment of e.There are two cases. If the stack did
not change (the antecedent of the conclusion is true), then
the output environment of e′ should be compatible with Γ′ .
Again because of (TIf), we need to allow the environment to
shrink:
• Case (EContext). Notice that we cannot have S ′ 6= S ,
since the only cases where that can happen is if (EApp)
or (ERet) apply, and they cannot be used as a hypothesis
of (EContext). Thus the left-hand side of the implication
(c) is true, and we have ∆′ ≤ Γ′ . Using this fact, the
remainder of the proof is routine.
• Case (ESelf). By assumption we have
(1) hX , TT , DT , E , self, S i → hX , TT , DT , E , E (self), S i
by (ESelf)
(2) TT ⊢ hΓ, selfi ⇒ hΓ, Γ(self)i by (TSelf)
(3) Γ(self) ≤ TS
(4) Γ ∼ E
(5) TT ⊢ TS ∼ S
(6) X ∼ (TT , DT )
Definition 6 (Type environment subsumption). We write
Γ1 ≤ Γ2 if dom(Γ2 ) ⊆ dom(Γ1 ) and for all x ∈ dom(Γ2 ),
it is the case that Γ1 (x) ≤ Γ2 (x).
If the stack does change, then the output environment is
irrelevant: It either is captured in the type stack if this is a
push due to a method call. Or it is discarded as the stack
16
2016/4/14
Let ∆ = ∆′ = Γ, and let TS ′ = TS . By (2) and (4)
there exists τ ′ such that · ⊢ h∆, E (self)i ⇒ h∆, τ ′ i and
τ ′ ≤ ∆(self). Then (a) holds, since typing of E (self)
was by (TNil) or (TObject), which do not depend of the
type class table. Also, (b) holds since τ ′ ≤ ∆(self) =
Γ(self) ≤ T S by (3). Also, the right-hand side of the
implication (c) holds trivially. Finally, (d) holds by (4),
(e) holds by (5), and (f) holds by (6).
• Case (EVar). Similar to (ESelf) case.
• Case (EAssn). By assumption we have
(1) hX , TT , DT , E , x = v, S i → hX , TT , DT , E [x 7→
v], v, S i
(2)
TT ⊢ hΓ, vi ⇒ hΓ, τ i
TT ⊢ hΓ, x = vi ⇒ hΓ[x 7→ τ ], τ i
• Case (EIfFalse). Similar to (EIfTrue) case.
• Case (EDef). By assumption we have
(1) hX , TT , DT , E , def A.m = λx.e, S i →
hX \A.m, TT , DT [A.m 7→ λx.e], E , nil, S i by (EDef)
(2) TT ⊢ hΓ, def A.m = λx.ei ⇒ hΓ, nili by (TDef)
(3) nil ≤ TS
(4) Γ ∼ E
(5) TT ⊢ TS ∼ S
(6) X ∼ (TT , DT )
Let ∆ = ∆′ = Γ, let TS ′ = TS , and let τ ′ = nil.
Then (a) holds trivially by (TNil). (c) holds trivially by
definition. (b), (d), and (e) hold trivially by (3), (4), and
(5).
For (f), pick some B.m′ ∈ dom(X \A.m). Observe that
B.m′ 6= A.m, by construction. By (6), we have
1. DM = (TT ⊢ h[y 7→ τy , self 7→ B], ei ⇒ hΓ′y , τy′ i)
2. D≤ = (τy′ ≤ τ2 )
3. DM and D≤ hold
4. DT (B.m′ ) = λy.e′
5. TT (B.m′ ) = τy → τ2
We need to show the above with the same type class table
and with dynamic class table DT [A.m 7→ λx.e]. But
then 1, 2, 3, and 5 are trivial, and since B.m′ 6= A.m we
have (DT [A.m 7→ λx.e])(B.m′ ) = DT (B.m′ ), thus 4
is trivial.
• Case (EType). This case is very similar to (EDef), except
the reduction in the semantics is different. By assumption, we have
(1) hX , TT , DT , E , type A.m : τm , S i →
h(X \A.m)[TT ′ ], TT ′ , DT , E , nil, S i where TT ′ =
TT [A.m 7→ τm ], by (EType)
(2) TT ⊢ hΓ, def A.m = λx.ei ⇒ hΓ, nili by (TDef)
(3) nil ≤ TS
(4) Γ ∼ E
(5) TT ⊢ TS ∼ S
(6) X ∼ (TT , DT )
(a)–(d) hold by the same reasoning above. To see (e),
observe that side condition A.m 6∈ TApp(S) means that
in the typing judgments internal to (5), (TApp) is never
applied with A.m. Hence those same judgments hold
under TT ′ , which only differs from TT in its binding
or A.m.
Let X ′ = (X \A.m)[TT ′ ]. For (f), again pick some
B.m′ ∈ dom(X ′ ). Observe that B.m′ 6= A.m, by construction. By (6), we have
1. DM = (TT ⊢ h[y 7→ τy , self 7→ B], ei ⇒ hΓ′y , τy′ i)
2. D≤ = (τy′ ≤ τ2 )
3. DM and D≤ hold
4. DT (B.m′ ) = λy.e′
5. TT (B.m′ ) = τy → τ2
We need to show the above, but in X ′ and with type
class table TT ′ and the same dynamic class table. By
by (TAssn) and either (TNil) or (TObject)
(3) τ ≤ TS
(4) Γ ∼ E
(5) TT ⊢ TS ∼ S
(6) X ∼ (TT , DT )
Let ∆ = ∆′ = Γ[x 7→ τ ], let TS ′ = TS , and let τ ′ = τ .
Notice that in (2), the hypothesis can only be proven by
either (TNil) or (TObject), both of which are insensitive
to the type environment. Thus, by the hypothesis of (2),
we also have TT ⊢ h∆, vi ⇒ h∆, τ i, which is (a).
Also, (b), (e), and (f) hold trivially by (3), (5), and (6).
Also, the right-hand side of the implication (c) holds
trivially. Finally, from (4) and the hypothesis of (2) we
have ∆ = Γ[x 7→ τ ] ∼ E[x 7→ v], which is (d).
• Case (ENew). Trivial.
• Case (ESeq). Trivial.
• Case (EIfTrue). By assumption we have
(1) hX , TT , DT , E , if v then e1 else e2 , S i →
hX , TT , DT , E , e1 , S i where v 6= nil
(2)
TT ⊢ hΓ, vi ⇒ hΓ, τ i
TT ⊢ hΓ, e1 i ⇒ hΓ1 , τ1 i
TT ⊢ hΓ, e2 i ⇒ hΓ2 , τ2 i
TT ⊢ hΓ, if v then e1 else e2 i ⇒ hΓ1 ⊔ Γ2 , τ1 ⊔ τ2 i
by (TIf) and (TObject), since v 6= nil
(3) τ1 ⊔ τ2 ≤ TS
(4) Γ ∼ E
(5) TT ⊢ TS ∼ S
(6) X ∼ (TT , DT )
Let ∆ = Γ, let ∆′ = Γ1 , let TS ′ = TS , and let
τ ′ = τ1 . From the second hypothesis of (2) we trivially
have (a). Moreover, τ ′ = τ1 ≤ τ1 ⊔ τ2 , so by (3) we
have τ ′ ≤ TS , which is (b). By (4), (5), and (6) we
trivially have (d), (e), and (f). Finally, by Lemma 1 we
have ∆′ = Γ1 ≤ (Γ1 ⊔ Γ2 ), which is the right-hand side
of the implication (c).
17
2016/4/14
′
′
construction, X ′ (B.m′ ) = (DM
, D≤ ) where DM
=
(TT ′ ⊢ h[y 7→ τy , self 7→ B], ei ⇒ hΓ′y , τy′ i), which is
′
1 and 2. Notice by construction that DM and DM
cannot
refer to A.m, thus we have 3. Finally, 4 holds trivially,
and 5 holds since B.m′ 6= A.m by construction.
• Case (EAppMiss). The inductive cases are similar to
(EContext). In the non-inductive case, by assumption we
have
(1) hX , TT , DT , E , C[[A].m(v2 )], S i →
hX [A.m 7→ (DM , D≤ )], TT , DT , [self 7→ [A], x 7→
v2 ], e, (E , C) :: S i where
(1a) DT (A.m) = λx.e
(1b) TT (A.m) = τ1 → τ2
(1c) type of(v2 ) ≤ τ1
(1d) DM = (TT ⊢ h[x 7→ τ1 , self 7→ A], ei ⇒
hΓ′ , τ2′ i) holds
(1e) D≤ = (τ2′ ≤ τ2 ) holds
(1f) A.m 6∈ dom(X )
(2)
3 holds by (1d) and (1e), 4 holds by (1a), and 5 holds by
(1b).
• Case (EAppHit). This case follows mostly the same
reasoning as above. The inductive cases are similar to
(EContext). In the non-inductive case, by assumption we
have
(1) hX , TT , DT , E , C[[A].m(v2 )], S i →
hX , TT , DT , [self 7→ [A], x 7→ v2 ], e, (E , C) :: S i
where
(1a) DT (A.m) = λx.e
(1b) TT (A.m) = τ1 → τ2
(1c) type of(v2 ) ≤ τ1
(1d) A.m ∈ dom(X )
(2)
TT ⊢ hΓ, [A]i ⇒ hΓ, Ai
TT ⊢ hΓ, v2 i ⇒ hΓ, τ i
TT (A.m) = τ1 → τ2
τ ≤ τ1
TT ⊢ hΓ, [A].m(v2 )i ⇒ hΓ, τ2 i
..
.
TT ⊢ hΓ, [A]i ⇒ hΓ, Ai
TT ⊢ hΓ, v2 i ⇒ hΓ, τ i
TT (A.m) = τ1 → τ2
τ ≤ τ1
TT ⊢ hΓ, [A].m(v2 )i ⇒ hΓ, τ2 i
..
.
TT ⊢ hΓC , C[[A].m(v2 )]i ⇒ hΓ′C , τC i
by (TApp) and (TObject) and possible (TNil).
(3) τC ≤ TS
(4) ΓC ∼ EC
(5) TT ⊢ TS ∼ S
(6) X ∼ (TT , DT )
By (6), we have X (A.m) = (DM , D≤ ) where DM =
(TT ⊢ h[x 7→ τ1 , self 7→ A], ei ⇒ hΓ′ , τ2′ i) holds and
D≤ = (τ2′ ≤ τ2 ) holds. Notice that we use properties
4 and 5 of the cache in combination with (1a) and (1b)
to know the assigned types in the cache, and the method
body at run-time, match in the cached derivation.
Let ∆ = [x 7→ τ1 , self 7→ A], let ∆′ = Γ′ , let TS ′ =
(ΓC [τ2 ], hΓ′C , τC i) :: TS , and let τ ′ = τ2′ . Then (a) holds
immediately by DM . (b) holds by D≤ and construction of
TS ′ .
The reasoning for (c)–(e) are the same as the (EAppMiss)
case. Finally, (f) holds trivially by (6), since the cache did
not change.
• Case (ERet). We have
(1) hX , TT , DT , E ′ , v, (E , C) :: S i →
hX , TT , DT , E , C[v], S i
(2) TT ⊢ hΓ′ , vi ⇒ hΓ′ , τ i by either (TObject) or
(TNil).
(3) τ ≤ τC
(4) Γ′ ∼ E ′
(5) TT ⊢ (ΓC [τC ], hΓ′C , τC′ i) :: TS ∼ (E , C) :: S
(6) X ∼ (TT , DT )
Let ∆ = ΓC , let ∆′ = Γ′C , and let TS ′ = TS . By (5),
we have TT ⊢ h∆[✷ 7→ τC ], Ci ⇒ h∆′ , τC′ i. Putting
that together with (2) and (3) via the substitution lemma,
we have TT ⊢ h∆, C[v]i ⇒ h∆′ , τC′′ i where τC′′ ≤ τC′ .
Let τ ′ = τC′′ , and we have (a). By (3) we have τC′ ≤ TS ,
TT ⊢ hΓC , C[[A].m(v2 )]i ⇒ hΓ′C , τC i
by (TApp) and (TObject) and possible (TNil).
(3) τC ≤ TS
(4) ΓC ∼ EC
(5) TT ⊢ TS ∼ S
(6) X ∼ (TT , DT )
Let ∆ = [x 7→ τ1 , self 7→ A], let ∆′ = Γ′ , let TS ′ =
(ΓC [τ2 ], hΓ′C , τC i) :: TS , and let τ ′ = τ2′ . Then (a) holds
immediately by (1d). (b) holds by (1e) and construction
of TS ′ . In this case the stack changes, so the left-hand
side of the implication (c) is false, hence (c) holds trivially. (d) holds because by (TObject) we have [A] has
type A, and by the second hypothesis of (2), which is
either (TObject) or (TNil), we have v2 has type τ , and by
the last hypothesis of (2) we have τ ≤ τ1 .
Next we show (e). By (4) we have ΓC ∼ E, and by
(2) and the Contextual Substitution Lemma we have
TT ⊢ hΓC [✷ 7→ τ2 ], Ci ⇒ hΓ′C , τC i. Thus we
have TT ⊢ (ΓC [τ2 ], hΓ′C , τC i) ∼ (E, C). Further,
by (3) we have τC ≤ TS . Finally, by (5) we have
TT ⊢ TS ∼ S . Putting this all together, we have TT ⊢
(ΓC [τ2 ], hΓ′C , τC i) ∼ (E, C) :: TS ∼ (E, C) :: S ,
which is (e).
Finally, to show (f), pick some element in the domain
of X ′ = X [A.m 7→ (DM , D≤ )]. If we pick some
B.m′ 6= A.m then all the necessary properties hold by
(6). If we pick A.m, then 1 and 2 hold by construction,
18
2016/4/14
and since τC′′ ≤ τC′ we therefore have (b) In this case the
stack changes, so the left-hand side of the implication (c)
is false, hence (c) holds trivially. (d) holds by (5), as does
(e). Finally, (f) holds by (6)
Now there are two cases. If A.m ∈ dom(X ) we can
immediately apply (EAppHit). Otherwise, if A.m 6∈
dom(X ), then we must have
DM = (TT ⊢ h[x 7→ τ1 , self 7→ A], ei ⇒ hΓ′ , τ i)
holds and D≤ = (τ ≤ τ2 ) holds because (EAppNTyp)
did not apply. But then combining this with our previous
assumptions we can apply (EAppMiss).
• Else e = C[e′ ]. Holds by induction and (EContext).
The progress theorem is much simpler:
Theorem 3 (Progress). If
(1)
(2)
(3)
(4)
(5)
TT ⊢ hΓ, ei ⇒ hΓ′ , τ i
τ ≤ TS
Γ∼E
TT ⊢ TS ∼ S
X ∼ (TT , DT )
Finally, we can put these together to prove soundness.
Theorem 4 (Soundness). If ∅ ⊢ h∅, ei ⇒ hΓ′ , τ i then either
e reduces to a value, e reduces to blame, or e does not
terminate.
then one of the following holds
1. e is a value, or
2. There exist X ′ , TT ′ , DT ′ , E ′ , e′ , S ′ such that
hX , TT , DT , E , e, S i → hX ′ , TT ′ , DT ′ , E ′ , e′ , S ′ i, or
3. hX , TT , DT , E , e, S i → blame
Proof. Let X = ∅, let TT = ∅, let Γ = ∅, let E = ∅, let
DT = ∅, let S = (∅, ✷) :: ·, and let TS = (∅[τ ], h∅, τ i).
Then by assumption we have TT ⊢ hΓ, ei ⇒ hΓ′ , τ i. By
construction we have τ ≤ TS and Γ ∼ E and TT ⊢
TS ∼ S and X ∼ (TT , DT ). Thus, these choices of
X , TT , Γ, DT , S , and TS satisfy the preconditions of
progress and preservation. Thus soundness holds by standard
arguments.
Proof. By induction on e.
• Case e = nil or e = [A]. These are values, so the theorem
holds trivially.
• Case self. By assumption (1) we have
TT ⊢ hΓ, selfi ⇒ hΓ, Γ(self)i
•
•
•
•
Thus, self ∈ dom(Γ). But then by (3), self ∈ dom(E ).
Thus (ESelf) can be applied.
(Note that assuming we start executing the program in a
standard environment, self will in fact always be bound
in all type and dynamic environments, unlike variables.)
Case x. Similar to self.
Case x = v, A.new, v; e, def A.m = (λx.e), type A.m :
τm . These cases are trivial, as there is one semantics rule
for each of these forms, and it will always be able to take
a step.
Case if v then e2 else e3 . This case is trivial, since either
(EIfTrue) or (EIfFalse) will apply.
Case v0 .m(v1 ). By assumption (1) we have
TT ⊢ hΓ, v0 i ⇒ hΓ0 , Ai
TT ⊢ hΓ0 , v1 i ⇒ hΓ1 , τ i
TT (A.m) = τ1 → τ2
τ ≤ τ1
TT ⊢ hΓ, v0 .m(v1 )i ⇒ hΓ1 , τ2 i
There are a few cases. If (EAppNil), (EAppNExist), or
(EAppNTyp) apply, then the theorem holds trivially. Otherwise, we must have v1 = [A] and DT (A.m) = λx.e.
More importantly, by (1) we have type of(v1 ) ≤ τ1 ,
since τ = type of(v1 ) by (1), i.e., v1 has the expected argument type. Also by (1) we have TT (A.m) = τ1 → τ2 .
19
2016/4/14
| 6cs.PL |
arXiv:1602.02862v1 [cs.NE] 9 Feb 2016
A Feature-Based Prediction Model of Algorithm
Selection for Constrained Continuous
Optimisation
Frank Neumann
Optimisation and Logistics
School of Computer Science
The University of Adelaide
Adelaide, Australia
Shayan Poursoltan
Optimisation and Logistics
School of Computer Science
The University of Adelaide
Adelaide, Australia
Abstract
With this paper, we contribute to the growing research area of feature-based
analysis of bio-inspired computing. In this research area, problem instances are
classified according to different features of the underlying problem in terms of
their difficulty of being solved by a particular algorithm. We investigate the impact of different sets of evolved instances for building prediction models in the area
of algorithm selection. Building on the work of Poursoltan and Neumann [11, 10],
we consider how evolved instances can be used to predict the best performing
algorithm for constrained continuous optimisation from a set of bio-inspired computing methods, namely high performing variants of differential evolution, particle
swarm optimization, and evolution strategies. Our experimental results show that
instances evolved with a multi-objective approach in combination with random instances of the underlying problem allow to build a model that accurately predicts
the best performing algorithm for a wide range of problem instances.
1
Introduction
Throughout the history of heuristic optimisation, various methods have been proposed
to solve constrained optimisation problems (COPs), specially non-linear ones. The
main idea behind these algorithms is to tackle the constraints. Important approaches in
this area are differential evolution (DE), particle swarm optimisation (PSO) and evolutionary strategies (ES). To handle the constraints, there have been many techniques
applied to these algorithms such as penalty functions, special operators (separating the
constraint and objective function treatment) and decoder based methods. We refer the
reader to [7] for a survey of constraint handling techniques in evolutionary computation. Given a range of different algorithms for constrained continuous optimisation,
we consider algorithm selection problem (ASP) [12] which consists of selection the
best performing algorithm from a suite of algorithms for a given problem instance. In
1
most circumstances, it is difficult to answer the following question: ”Can we estimate
the likelihood that algorithm A will be successful on a given constrained optimisation
problem P?”. Recent works in the field show that it is possible to select the algorithm
most likely to be best suited for a given problem [8, 14, 2]. Based on these studies, it is
possible to find the links between problem characteristics and algorithm performance.
The key to these investigations is problem features which can be used to predict the
most suited algorithm from a set of algorithms.
It is widely assumed that constraints play a vital role in COP’s difficulty. Therefore, in this study we use the meta-learning framework outlined in [12, 16] to build a
prediction model for a given COP. Our model predicts the best algorithm type (DE, ES
and PSO) for a given COP based on their constraint features. The model inputs include
the features of constraints in a given problem. It is shown in [11, 9] that by using an
evolving approach, it is possible to generate problem instances covering a wide range
of problem/algorithm difficulty. Such instances can be used to extract and analyse the
features that make a problem hard or easy to solve for a given algorithm. For a detailed
discussion on these constraints (linear, quadratic and their combination) we refer the
reader to [11].
To build a reliable prediction model, we need to train it with variety of problem
instances that are hard or easy for algorithm(s). Based on the investigations in [10],
a multi-objective evolutionary algorithm can be used to generate constrained problem
instances that are hard/easy for one algorithm but still easy/hard for the others. The
authors show which features of the constraints make the problems hard for certain
algorithm but still easy for the others. Hence, we use the same approach to generate
problem instances to use in our model training phase. This can improve the accuracy of
prediction model since the training instances are used to show the strengths and weaknesses of various algorithm types over constraint features. To illustrate the model’s
efficiency on constraints, we examine our model with generated testing problems such
as hard/easy for one but easy/hard for the others and more general random instances.
To show the model prediction ability over constraints (linear, quadratic and their combination), we also experiment given problems with various objective functions.
The remainder of this paper is organised as follows. In section 2, we discuss constrained continuous optimisation problems. Later, we introduce the evolutionary algorithms that are suggested by our prediction model. Moreover, the background materials
related to multi-objective evolver, algorithm selection problem and meta-learning prediction model are discussed in detail. Section 3 describes and compares all models
trained with different subsets of instances from multi-objective evolver population set.
By choosing the best training data preference in Section 3, the experimental analysis
on various benchmark problems is described in Section 4. We then conclude with some
remarks in Section 5.
2
2
2.1
Preliminaries
Constrained Continuous Optimisation Problems
A constrained optimisation problem (COP) in a continuous space is formulated as follows:
Find x ∈ S ⊆ RD
f (x) = min( f (y); y ∈ S),
subject to gi (x) ≤ 0
h j (x) = 0
∀i ∈ {1, . . . , q}
(1)
∀ j ∈ {q + 1, . . . , p}
In this formulation, f , gi and h j are real-valued functions on the search space S, q is
the number of inequalities and p − q is the number of equalities. The search space S is
defined as a D dimensional rectangle in RD . These equality and inequality constraints
could be linear or nonlinear. The set of all feasible points F ⊆ S which satisfy all
equality and inequality constraints is formulated as:
li ≤ xi ≤ ui ,
1≤i≤D
(2)
where li and ui denote lower and upper bounds for the ith variable respectively. Usually, to simplify COP, the equalities are replaced by the following inequalities [18] as
follows:
|h j (x)| ≤ ε
for j = q + 1 to p
(3)
where ε is a small positive value. In all experiments in this paper, the value of ε is
considered as 1E-4, the same as it was in [5].
2.2
Algorithms
In this section we discuss the basic ideas about algorithms for constrained optimisation problems such as differential evolution, evolutionary strategies and particle swarm
optimisation.
The ε-constrained differential evolution with an archive and gradient-based mutation (εDEag) is the winner of 2010 CEC competition for continuous constrained
optimisation problems [5]. This algorithm uses ε-constrained method technique to
transform algorithms for unconstrained problems to constrained ones. Also, possible
solutions are ordered by ε-level comparison. This means, the lexicographic order is
used in which constraint violation (φ (x)) has more priority and proceeds the function
value ( f (x)). A detailed description of this algorithm can be found in [17].
For evolutionary strategy algorithms, (1+1) CMA-ES for constrained optimisation
[1] is included in our experiment. This algorithm is a variant of (1 + 1)-ES which
adapts the covariance matrix of its offspring distribution in addition to its global step
size. The (1 + 1) CMA-ES for constrained optimisation obtains approximations to the
normal vectors directions in the vicinity of the current solution locations by applying
low-pass filtering steps that violates the constraints and reducing the variance of the
3
offspring distribution in these directions. Adopting this method makes (1 + 1) CMAES as one of the most efficient algorithms for constrained optimisation problems. We
refer the reader to [1] for detailed description and implementation.
The next algorithm that is used in our investigation from particle swarm optimisation algorithms is hybrid multi-swarm particle swarm optimisation (HMPSO). This
algorithm divides the current swarms into sub-swarms and search the solution between
them in parallel. All particles in each sub-swarms locate their fittest local particle
which attracts the particles to fitter positions. Also, having multiple sub-swarms near
different optima increase the diversity of the algorithm. A detailed description and
implementation of HMPSO can be found in [19].
2.3
Multi-objective Investigations
In order to extract information about the strengths and weaknesses of certain algorithms
on constrained optimisation problems, we need problem instances with different kinds
of difficulties for the considered algorithms. The reason behind this idea is that using
instances that are randomly generated are not efficient to cover the full spectrum of difficulty analysis. To do this, we evolve instances to find the ones that are hard/easy for
one algorithm and easy/hard for the others. Analysing the features of these instances
helps us extracting knowledge regarding the strengths and weaknesses of the considered algorithms and give reasons of why an algorithm performs well on one problem
while the others have difficulties. Insights from this analysis can be used to develop
more efficient prediction model for automated algorithm selection.
As mentioned above, we generate problem instances of different difficulties. To do
this, a multi-objective DE algorithm (evolver) [13] is used to evolve constraints that
make problems hard for one algorithm type and easy for the others in the algorithm
suite following the approach in [11, 10]. The feature-based comparison of various
algorithm types has been carried out in these papers. The authors show the constraint
(linear and/or quadratic and their combination) features that are more contributing to
problem difficulty for certain algorithms. We refer the readers to [11, 10] for a detailed
description and implementation.
2.4
Algorithm Selection Problem
There are many algorithms that are proposed for constrained continuous optimisation
problems. These algorithms are categorised as different types such as differential evolution (DE), evolutionary strategy (ES) and particle swarm optimisation (PSO). So, as
a direct consequence of this, it is difficult to understand which algorithms or types of
algorithms are more efficient to solve given COPs. To determine the best algorithm
to solve a problem is referred to ”Algorithm Selection Problem” term in [12] by Rice.
In his work, Rice proposed a model with four main characteristics: a set of problem
instances F, a set of algorithms A, measures for the cost of performing algorithms on
particular problem (Y ) and set of characteristics of problem instances (C). The illustration for Rice general algorithm selection framework is shown in Figure 1 which
predicts the performance y(a( f )) of a given algorithm a on a problem f by extracted
features c. If a solution is found, it is possible to extract features from a given problem
4
Figure 1: The framework of general problem of algorithm selection and performance prediction using problem features
based on [12].
and select the most appropriate algorithm or predict the performance of the algorithm
based on these features. This framework has been extended by [16, 3, 4] in a variety of
computational problem domains using meta-learning framework. So, if the values are
features of problems with algorithm performance measure are known beforehand, then
it is possible to use a learning strategy to predict the algorithm performance based on
the problem features.
2.5
Prediction Model
Our prediction model is based on the [8]. Note that this model is used for unconstrained
continuous optimisation problems using landscapes features. Inputs to the model are
independent problem feature variables (C) and algorithm parameters and output is the
performance measure as an dependant variable which is required function evaluation
number (FEN) of the suggested algorithm. This model can be used to predict the algorithm behaviour on a given problem. To achieve this we use popular basic technique
for model building. A high-level overview of the regression model is shown in Figure
2.
As discussed above, there have been many attempts to train these prediction models
with random generated or benchmark problem instances which could not fully include
all problem instances with difficulty variations. To improve this, we cover the full spectrum of difficulty by evolving two sets of instances with extreme problem difficulties.
These extreme difficulty instances are the ones which are hard for one algorithm and
easy for the others or easy for one and still hard for the rest.
5
Figure 2: Meta-learning prediction model for a constrained continuous optimisation problem
To build our regression model, we implement a multi-layered feed-forward neural
network with 2 hidden layers and 10 neurons in each layer as the regression model.
For training the model, we use a Levenbeg-Marqurd back-propagation algorithm [6]
package using Matlab R2014b. To train this model, we use evolved instances that
are generated from multi-objective evolver in [10]. The prediction model inputs are
given COPs constraint features and algorithms parameter values (the parameters for
the experimented algorithms are identical to [17, 19, 1].)
3
Prediction Model based on Evolved Instances
As mentioned earlier, our goal is to propose a reliable prediction model using constraint features. This reliability can be improved by choosing proper set of learning
data. The accuracy of this prediction model depends on many factors such as the relevance of constraint features, the diversity of instances used to train the model and its
training method. Therefore, to improve this, we train our prediction model obtained
from multi-objective evolver described in [10]. These instances are hard for one algorithm but still easy for the other (or easy for one and hard for the other algorithms).
Analysing these instances shows the effectiveness of constraint features that make a
problem hard or easy for certain algorithms. This set up improves the accuracy of
prediction which is evaluated on the capability to provide realistic ranking of different
algorithm types performances. This can be done by comparing the required function
evaluation numbers (FEN) needed by various algorithms to solve a given COP.
The prediction model uses constraint features (first input) to predict the best algorithm type for a given constrained optimisation problem. These constraint features
are constraint coefficients relationships such as standard deviation, angle between constraint hyperplanes, feasibility ratio in vicinity of optimum and number of constraints.
The details of these features are discussed in [11]. Also, for the second input, since se-
6
lecting various algorithm parameters has different impact on algorithm ability to solve
a given problem, we conduct an experimentally driven meta-learning approach which
has been proposed by Smith-Miles in [15]. For our model, we use the parameters for
DE, CMA-ES and HMPSO suggested in [17, 19, 1].
Our goal of building a prediction model is to identify the best algorithm type for
a given problem. Therefore, the model output predicts the most suited algorithm and
required FEN to solve a given constrained problem. The suggested FENs denote numbers of function evaluation which are needed by different algorithms to solve a given
COP.
In the following we train our prediction model with variety of instance subsets generated by the multi-objective evolver. We choose different combination of subsets of
instances to maximise our prediction model accuracy upon a given constrained problem. These training phase instance subsets are selected from extreme points, Pareto
front line, more random (general) solutions and combination of Pareto front and random points from multi-objective evolver solution population in [10]. We then compare
the prediction accuracy for these prediction model with various training data preferences.
3.1
Extreme Instances
We first train our prediction model with an extreme instance subset which covers the
extreme points of Pareto front in the multi-objective evolver population set. These
extreme solutions are selected from evolved instances that are easiest/hardest for one
and hardest/easiest for the other algorithms at the same time. The reason behind this
selection is to assess the ability of our model to find the most suited algorithm for
a given COP which is fairly hard for one or multiple algorithms. In other words, it
is more beneficial to choose a best algorithm for a given COP in which it cannot be
solved easily by certain other algorithms.
To determine the actual accuracy of our extreme point prediction model (EP-PM),
we select 1500 extreme instances that are hard/easy for one and easy/hard for the other
algorithms for its training phase. To analyse and test the quality of this prediction
model, we use two sets of testing problem instances that we already know their best
algorithm and required FEN. The first one is the set of problem instances that are
hard/easy for one algorithm and easy/hard for the others. This set can improve the
accuracy of EP-PM for given problems that fall into extreme-like evolved problem instances. However, it is very likely that the real world given COP is similar to other
evolved instance subset types. Therefore, as a second set, we use random (general)
testing problem instances to analyse the EP-PM with a potential real world given problem. We need to mention that we already know about their best algorithm and required
FEN.
Our result for EP-PM is summarised in Table 1 for instances that are hard/easy
for one and easy/hard for the other algorithms. Also, the results for testing random
instances are shown in this table. Moreover, DE hard (1 C) denotes testing Sphere
problem instance with 1 linear constraint that is hard for DE algorithm but still easy for
the others. The information about actual and predicted algorithm and required FEN to
solve a problem instance is indicated. The model not only suggests the best algorithm
7
with its required FEN but also predicts the FENs required to solve a given problem
with other algorithms (not the best ones).
Based on the results, EP-PM performs acceptable on extreme-like testing instances.
This performance is acceptable on predicted best algorithm type and required FEN.
Also, the error rates for using other algorithms (not the best ones) are still acceptable.
But, analysing the testing random instances, it is obvious the model is not capable of
predicting algorithm and FEN for these more general form instances. This means, the
proposed model (EP-PM) is not accurate enough for randomly generated real world
instances and the difference between predicted and actual FEN is considerable. Also,
the error rate of other algorithm choices (not the best algorithm) is still high.
To summarize, although the EP-PM model performs fairly accurate on instances
that are grouped into extreme points evolved instances, still needs improvement to handle other subsets (such as random generated instances). The likelihood of given COP
which is more similar to random evolved instances are higher. Thus, this motivates us
to examine other subsets from evolving algorithm population instances for our training
phase. This could be moving along the Pareto front line and choosing more instances
from this category.
3.2
Pareto Front Instances
It is shown that in order to improve the accuracy of our prediction model we need to
include or select different varieties of evolved instance subsets for its training step. The
idea behind this choice is to obtain a model that can predict more general forms of given
constrained problems. Of course it is vital to predict best algorithm for a given problem
which is considerably hard/easy for one and still easy/hard for the other algorithms,
but we also need to include more forms of generality to our prediction model. So, we
need to move along Pareto front line in multi-objective evolver population set for our
learning phase. This could increase the ability of our model to predict algorithms for
more general given COP which is not similar to extreme point instances.
Given a total number of 3000 instances from evolver Pareto front line, we train our
Pareto front prediction model (PF-PM). This preference could increase PF-PM ability
to predict more general forms of given COPs. To compare the quality of our prediction model and other models with different learning phases, we use same extreme and
random generated testing instances used in previous section. Results shown in Table 2
indicate an improving accuracy for random testing instances used for previous model
EP-PM (using extreme instances). Looking at the Table 2, we see that moving towards
Pareto front line in evolver population set for choosing learning instances increases the
accuracy of predicted FEN for predicted algorithm. Also, predicting FEN using other
algorithms (not the best one) represents that the PF-PM is more accurate than EP-PM
for testing random instances. This improvement is acceptable in algorithm type prediction but we still need to improve the predicted required FEN.
As discussed in this section, to improve the accuracy of prediction model, we chose
instances of evolver from its Pareto front line solution population. In other words,
Pareto front prediction model (PF-PM) has some strengths and weaknesses. Although
its error rate for predicting correct algorithms is improved, there is still considerable
difference between the actual required FENs and predicted ones. As testing instances
8
are selected mostly from more general instances (not close to extreme points), we need
to experiment other training instances subsets. In order to address more general form,
based on results so far, our next move is to choose more random instances from evolver
population set for our training set. This could result in increasing the accuracy of our
model for more general forms of given COPs.
3.3
Random Instances
The initial prediction models discussed earlier (EP-PM and PF-PM) has some limitations. In other words, they are not accurate enough to predict more general forms
of given COPs, specially when they are similar to evolver population instances except
extreme and Pareto front points. The results for PF-PM show an increase in accuracy
of prediction for testing COPs which are not similar to extreme points, but the predicted required FEN still needs an improvement. Our goal is to design a prediction
model with an ability to predict all possible given COPs. These COPs are within the
range of extreme to random like instances. Based on previous results, it is shown that
moving from extreme to Pareto front line instances increases the model accuracy (see
Section 3.2). Hence, to decrease the error rate for required FEN for more general testing COPs we choose only random instances for testing phase. These random instances
are selected from evolver population set.
We use 3000 random instances from multi-objective evovler population set for our
random only prediction model (RO-PM). To assess the accuracy of our random only
prediction model we use the same testing instances applied to EP-PM and PF-PM.
Table ?? indicates the actual and predicted FEN and algorithm of hard/easy and random testing instances for RO-PM. As it is observed, the random only prediction model
(RO-PM) fails to predict suitable algorithm for a given COP. This failure include both
predicted algorithm and required FEN. Results show that moving through random instances in evolver population and choose only random instances increase the number
of incorrect predictions. Also, comparing to previous models, RO-PM accuracy is decreased for testing instances similar to extreme points (easy/hard instances).
It is shown that the accuracy of resulting model with Pareto front instances (PFPM) is improved by selecting different subsets (Pareto front line) than extreme points.
This improvement is analysed by experimenting more general form of testing COPs.
Therefore, this motivated us to experiment only random solutions in order to build more
accurate prediction model for given COPs. By selecting only random instances to train
the new model, it is observed the predicted algorithm and required FEN is not accurate
as Pareto front (PF-PM) and extreme points (EP-PM) models. It can be translated as
excluding instances from Pareto front line for training step decreases the accuracy of
prediction model. Also, the RO-PM model failed to predict testing instances which are
similar to evolver extreme points (easy/hard instances). So, other possibility is to use
a combination of both Pareto front and random instances from evolver population for
model training.
9
3.4
Pareto Front with Random Instances
In previous sections we experimented three types of prediction models (EP-PM, PF-PM
and RO-PM). These models have different prediction accuracy upon choosing different
varies of evolver subsets for their training phases. Our goal of building a prediction
model is to minimise its error rate in both algorithm and required FEN. So far, we
understand that moving from extreme points (EP-PM) to Pareto front (PF-PM) for
training step increases the accuracy of prediction model. However, moving further and
choosing only random points from evolver (RO-PM) is not the solution for covering
all possible given COPs (extreme and random like instances). In other words, there
should be a trade-off relation between moving towards random points from extreme
and random instances in multi-objective evolver population. Therefore, our preference
for training phase is a combination set of Pareto front and random instances from multiobjective evolver. Not only it covers instances that are hard/easy for one and easy/hard
for the other algorithms, but also it can predict general given COPs more accurately.
To train our Pareto front with random instances prediction model (PFR-PM), we
use 3000 points from evolver population set (1500 each). To compare and assess the
prediction model accuracy we experiment our PFR-PM with the same testing COPs for
the former models. The results for Pareto front and random prediction model (PFRPM) are shown in Table 4. It is observed that including both Pareto front line and
random only instances can be effective in accuracy improvement. Analysing the results, it is obvious the error rate for both predicted algorithm and required FEN are
decreased. Also, the model is able to predict required FEN using other algorithms (not
the best one) more accurate. This can be concluded by having lower error rates for
FENs of other algorithms (not the best one) for PFR-PM. The reason behind this is that
to cover all possible given COPs, we use both Pareto front and random points from
evolver population for out training phase. In other words, our model is trained with
constraint characteristics and features of both types of COPs (Pareto front and random
instances).
It is shown that choosing the proper subsets for training phase is effective in prediction model accuracy. The results for four prediction models suggest that moving
from extreme points to random instances can improve the quality of prediction model.
It is found that there is a trade-off relationship in choosing instances that are close to
random or extreme points (from evolver) for training phase. Results analysis shows
by selecting a combination subsets from both random and Pareto front instances (PFRPM) for training step, we can improve our model prediction quality. This improvement
is in both selected algorithm and also its required FEN. By selecting the best model
(PFR-PM), in the following, we examine it in a more detailed approach.
4
Experiments on Benchmarks
Our goal is to design highly accurate prediction model for a given COP based on its
constraint features. As mentioned earlier, in order to improve the model accuracy, it
needs to be trained with COP instances that are generated using multi-objective evolver.
So far, we analysed the results for all four types of prediction models trained with ex-
10
treme points (EP-PM), Pareto front (PF-PM), random only (RO-PM) and combination
of Pareto front with random (PFR-PM) instances. We have experimented our prediction
model with various subsets of training data to analyse the best preference. As results
indicate, the most accurate prediction model is the one which is trained with combination of Pareto front and random subset of evolver population (PFR-PM). This model is
capable of predicting algorithms for almost all possible given testing COPs such as the
ones similar to extreme (hard/easy) or ordinary instances (random) in a multi-objective
evolver population. Also, the prediction ability for required function evaluation number (FEN) has been significantly increased. Therefore, in order to assess our optimised
prediction model (see Section 3.4), we decide to experiment it with our newly designed
benchmark. In order to analyse the capability of the prediction model (PFR-PM) on
constraints, we use fixed objective function with various numbers of linear, quadratic
(and their combination) constraints. Then, we test other well-known objective function
to see the relationship of constraints and our prediction model.
For this experiment, we train our PFR-PM with 3000 instances from both Pareto
front and random instances of evolving algorithm population set (1500 each). We also
use optimised algorithm parameter settings for each algorithm suggested in [17, 19, 1].
In order to show the accuracy of model on prediction over constraints, we examine
the model on various well-known objective functions such as Sphere (bowl-shaped),
Ackley (many local optima) and Rosenbrock (valley-shaped). Also, to evaluate the
effectiveness of our model on constrained problems, we use various numbers and types
of constraints. Tables 5, 6 and 7 show the prediction results for Sphere, Ackley and
Rosenbrock objective functions respectively. The results show the number of correct
algorithm types prediction (success rate) from 30 different tests. Also, one step further,
the average deviations of required FEN (the correct and predicted one) for the predicted
algorithms are calculated.
Table 5 compares the prediction results for our proposed model (PFR-PM) and random only model (RO-PM) for Sphere COPs. The results indicate the effectiveness of
choosing the proper subset training instances. It is observed that the prediction algorithm success rate for our proposed model (PFR-PM) is significantly better than ROPM for all Sphere COPs using various combinations of constraints. The success rate
(out of 30 tests) for newly testing given COP is significantly higher for PFR-PM comparing to RO-PM. Also, the low value average deviation of predicted FEN and actual
one for PFR-PM represents its higher accuracy in predicting the algorithm performance
in terms of function evaluation number.
By observing the Tables 6 and 7, we realise that our prediction model (PFR-PM)
is reliable in predicting with only constraints. In other words, experimenting other
types of objective functions (Bowl-shaped, many local-optima and valley shaped) with
accurate results shows the ability of the model to predict based on constraints. Based
on the Table 6, the lower value of FEN average deviation indicates the higher accuracy
of PFR-PM for Ackley COPs. Also, the results for Rosenbrock COPs with 1 to 5
linear, quadratic constraints (and their combination) shows the accuracy of PFR-PM
comparing to RO-PM. The average deviation of FEN for Rosenbrock problems denotes
the significantly close predicted FEN with PFR-PM (see Table 7).
As mentioned before, the output of our proposed prediction model (PFR-PM) includes predicted algorithm with its required FEN. It is observed that the prediction
11
model is capable of suggesting the best algorithm and required FEN based on constraint features of given COP. Due to the stochastic nature of evolutionary optimisation,
the above benchmark tests are repeated 30 times and the two-tail t-test significance is
performed for average deviations of FEN. The significant level α is considered as 0.05.
The p-values for significance of a difference between FEN average deviation of Pareto
front with random (PFR-PR) and random only (RO-PR) models for each Sphere, Ackley and Rosenbrock are shown in Tables 5, 6 and 7 respectively. The results show that
the difference in FEN Average deviation are significant and less than 0.05.
As discussed earlier, the idea of designing a prediction model based on instance
features is rather a novel approach in algorithm selection problem. Training a model
with COP instances from multi-objective evolver improves the prediction accuracy.
The performance prediction (FEN) and suggested algorithm can be used to produce
the final output of our prediction model. As we know selecting a suitable algorithm
for a given problem requires substantial amount of time. In contrast, in our approach,
we only need to extract features of a problem once and the model produces the final
output. It is observed that selecting different sets of training instances improves the prediction model success rate. We designed and examined various prediction model using
different subsets of problem instances from evolver population set. In order to show
the ability of the prediction model only based on constraints features we use various
objective functions. Results for these COPs with different combinations of objective
functions and constraints indicate that the model is highly accurate in algorithm and
required FEN prediction.
5
Conclusion
In this study, we examined the impact of different types of problem instances that can
be used in prediction models for constrained continuous optimisation.
Our resulting prediction model captures the links between constraint features, algorithm type performance and the required function evaluation number. The model
inputs are considered as constraint features and selected parameter settings. The outputs includes the required function evaluation number and most suited algorithm type
to solve the given COP.
The model was trained (using NN learning strategy) with evolved COP instances.
To improve the accuracy of the model we used evolved instances that are hard/easy
for one and easy/hard for the other algorithms. These training instances are generated
with multi-objective evolver. We first, chose various subsets of instances from multiobjective evolver population set. It is observed that the a model using combination of
Pareto front line and random points from population set has the highest accuracy in predicting best algorithm types for a given COP. We then, tested our prediction model with
different objective functions and constraints types. The results indicate our prediction
model is reliable to suggest and predict most suited algorithm and required FEN using
problem’s constraint features. Our approach shows the relationship between constraint
features and various algorithm performances. The results clearly demonstrate the ability of prediction model to predict the algorithm and required FEN using only constraint
features.
12
References
[1] D. V. Arnold and N. Hansen. A (1+ 1)-cma-es for constrained optimisation. In Proceedings of the 14th annual conference on Genetic and evolutionary computation,
pages 297–304. ACM, 2012.
[2] B. Bischl, O. Mersmann, H. Trautmann, and M. Preuß. Algorithm selection based
on exploratory landscape analysis and cost-sensitive learning. In Proceedings of the
14th annual conference on Genetic and evolutionary computation, pages 313–320.
ACM, 2012.
[3] F. Hutter, Y. Hamadi, H. H. Hoos, and K. Leyton-Brown. Performance prediction
and automated tuning of randomized and parametric algorithms. In Principles and
Practice of Constraint Programming-CP 2006, pages 213–228. Springer, 2006.
[4] K. Leyton-Brown, E. Nudelman, and Y. Shoham. Empirical hardness models:
Methodology and a case study on combinatorial auctions. Journal of the ACM
(JACM), 56(4):22, 2009.
[5] R. Mallipeddi and P. N. Suganthan. Problem definitions and evaluation criteria
for the cec 2010 competition on constrained real-parameter optimization. Nanyang
Technological University, Singapore, 2010.
[6] D. W. Marquardt. An algorithm for least-squares estimation of nonlinear parameters. Journal of the Society for Industrial & Applied Mathematics, 11(2):431–441,
1963.
[7] E. Mezura-Montes and C. A. Coello Coello. Constraint-handling in nature-inspired
numerical optimization: past, present and future. Swarm and Evolutionary Computation, 1(4):173–194, 2011.
[8] M. A. Muñoz, M. Kirley, and S. K. Halgamuge. A meta-learning prediction model
of algorithm performance for continuous optimization problems. In Parallel Problem Solving from Nature-PPSN XII, pages 226–235. Springer, 2012.
[9] S. Poursoltan and F. Neumann. A feature-based analysis on the impact of linear
constraints for ε-constrained differential evolution. In Evolutionary Computation
(CEC), 2014 IEEE Congress on, pages 3088–3095. IEEE, 2014.
[10] S. Poursoltan and F. Neumann. A feature-based analysis on the impact of set of
constraints for ε-constrained differential evolution. In S. Arik, T. Huang, W. K. Lai,
and Q. Liu, editors, Neural Information Processing, volume 9491 of Lecture Notes
in Computer Science, pages 344–355. Springer International Publishing, 2015.
[11] S. Poursoltan and F. Neumann. A feature-based comparison of evolutionary computing techniques for constrained continuous optimisation. In S. Arik, T. Huang,
W. K. Lai, and Q. Liu, editors, Neural Information Processing, volume 9491 of
Lecture Notes in Computer Science, pages 332–343. Springer International Publishing, 2015.
13
[12] J. R. Rice. The algorithm selection problem. 1975.
[13] T. Robič and B. Filipič. Demo: Differential evolution for multiobjective optimization. In Evolutionary Multi-Criterion Optimization, pages 520–533. Springer,
2005.
[14] K. Smith-Miles. Towards insightful algorithm selection for optimisation using
meta-learning concepts. In WCCI 2008: IEEE World Congress on Computational
Intelligence, pages 4118–4124. IEEE, 2008.
[15] K. A. Smith-Miles. Cross-disciplinary perspectives on meta-learning for algorithm selection. ACM Computing Surveys (CSUR), 41(1):6, 2008.
[16] K. A. Smith-Miles, R. J. James, J. W. Giffin, and Y. Tu. A knowledge discovery approach to understanding relationships between scheduling problem structure
and heuristic performance. In Learning and intelligent optimization, pages 89–103.
Springer, 2009.
[17] T. Takahama and S. Sakai. Constrained optimization by the ε constrained differential evolution with an archive and gradient-based mutation. In Evolutionary
Computation (CEC), 2010 IEEE Congress on, pages 1–9. IEEE, 2010.
[18] T. Takahama and S. Sakai. Efficient constrained optimization by the ε constrained
adaptive differential evolution. In Evolutionary Computation (CEC), 2010 IEEE
Congress on, pages 1–8. IEEE, 2010.
[19] Y. Wang and Z. Cai. A hybrid multi-swarm particle swarm optimization to
solve constrained optimization problems. Frontiers of Computer Science in China,
3(1):38–52, 2009.
14
Table 1: Predicted and actual most suited algorithm type/required FEN for Sphere function with linear constraint(s). The
prediction model is trained with extreme points (EP-PM) from multi objective evolver population. DE hard/easy (1 C) is a
problem instance that is hard/easy for DE algorithm but easy/hard for the others.
Instances
name
DE hard (1 c)
ES hard (1 c)
PSO hard (1 c)
DE hard (2 c)
ES hard (2 c)
PSO hard (2 c)
DE hard (3 c)
ES hard (3 c)
PSO hard (3 c)
DE hard (4 c)
ES hard (4 c)
PSO hard (4 c)
DE hard (5 c)
ES hard (5 c)
PSO hard (5 c)
DE easy (1 c)
ES easy (1 c)
PSO easy (1 c)
DE easy (2 c)
ES easy (2 c)
PSO easy (2 c)
DE easy (3 c)
ES easy (3 c)
PSO easy (3 c)
DE easy (4 c)
ES easy (4 c)
PSO easy (4 c)
DE easy (5 c)
ES easy (5 c)
PSO easy (5 c)
Rnd. 1 (1 c)
Rnd. 2 (1 c)
Rnd. 3 (2 c)
Rnd. 4 (2 c)
Rnd. 5 (3 c)
Rnd. 6 (3 c)
Rnd. 7 (4 c)
Rnd. 8 (4 c)
Rnd. 9 (5 c)
Rnd. 10 (5 c)
Rnd. 11 (1 c)
Rnd. 12 (1 c)
Rnd. 13 (2 c)
Rnd. 14 (2 c)
Rnd. 15 (3 c)
Rnd. 16 (3 c)
Rnd. 17 (4 c)
Rnd. 18 (4 c)
Rnd. 19 (5 c)
Rnd. 20 (5 c)
Predicted Actual Error Predicted
alg.
alg.
FEN for
DE
ES
ES
NO
83.2K
PSO
DE
YES 43.4K
DE
DE
NO
38.9K
ES
ES
NO
83.5K
DE
DE
NO
47.2K
DE
DE
NO
41.5K
ES
ES
NO
87.3K
PSO
PSO
NO
45.6K
DE
DE
NO
47.2K
PSO
ES
YES 95.2K
PSO
PSO
NO
51.3K
DE
ES
YES 49.2K
ES
ES
NO
94.7K
DE
DE
NO
51.4K
DE
DE
NO
49.1K
DE
DE
NO
45.3K
ES
ES
NO
87.3K
PSO
PSO
NO
92.5K
DE
DE
NO
49.4K
ES
ES
NO
78.4K
PSO
PSO
NO
84.2K
DE
DE
NO
41.8K
ES
ES
NO
91.4K
PSO
PSO
NO
88.4K
DE
DE
NO
51.6K
ES
ES
NO
85.3K
PSO
DE
YES 57.2K
DE
DE
NO
59.2K
ES
ES
NO
88.3K
PSO
PSO
NO
89.3K
DE
PSO
YES 53.2K
ES
DE
YES 77.2K
ES
DE
YES 71.4K
PSO
ES
YES 61.2K
DE
DE
NO
45.2K
ES
PSO
YES 71.9K
DE
ES
YES 71.6K
PSO
DE
YES 84.7K
ES
DE
YES 91.8K
DE
PSO
YES 78.4K
ES
DE
YES 56.3K
ES
DE
YES 54.2K
DE
PSO
YES 48.2K
ES
PSO
YES 70.2K
DE
PSO
YES 63.4K
PSO
DE
YES 77.5K
ES
ES
NO
69.3K
DE
PSO
YES 71.4K
DE
ES
YES 81.2K
PSO
PSO
NO
81.2K
Actual
FEN
for DE
86.3K
45.7K
37.2K
87.4K
46.4K
43.6K
92.4K
51.2K
49.6K
93.7K
49.2K
52.7K
96.3K
53.8K
51.3K
48.9K
81.4K
87.1K
44.5K
83.6K
81.4K
48.4K
87.8K
89.5K
53.2K
89.4K
55.5K
57.8K
91.9K
91.3K
65.7K
61.9K
59.8K
72.2K
65.4K
82.4K
83.2K
71.1K
82.7K
93.2K
44.6K
45.2K
61.2K
79.0K
71.9K
69.2K
73.9K
85.3K
93.9K
90.9K
Error
for DE
-3.1K
-2.3K
+1.7K
-3.9K
+0.8K
-2.1K
-5.1K
-5.6K
-2.4K
-1.5K
+2.1K
-3.5K
-1.6K
-2.4K
-2.2K
-3.6K
+5.9K
+5.4K
+4.9K
-5.2K
+2.8K
-6.6K
+3.6K
-1.1K
-1.6K
-4.1K
+1.7K
+1.4K
-3.6K
-2.0K
-12.5K
15.3K
+11.6K
-11.0K
-20.2K
-10.5K
-11.6K
+13.6K
+9.1K
-14.8K
+11.7K
+9.0K
-13.0K
-8.8K
-8.5K
+8.3K
-4.6K
15
-13.9K
-12.7K
-9.7K
Predicted
FEN for
ES
43.7K
80.9K
43.2K
42.5K
84.3K
46.3K
43.8K
95.5K
51.2K
52.4K
85.3K
53.4K
49.3K
94.2K
57.9K
81.5K
48.3K
81.4K
89.3K
51.4K
78.3K
85.8K
48.2K
93.2K
83.5K
51.2K
84.1K
93.2K
53.2K
84.3K
65.3K
43.5K
48.1K
68.8K
77.2K
45.6K
80.3K
68.8K
83.9K
79.2K
49.3K
48.2K
78.5K
61.2K
71.3K
89.5K
68.5K
81.2K
83.2K
81.3K
Actual
FEN
for ES
41.5K
84.2K
41.8K
45.2K
88.3K
45.1K
45.2K
91.2K
53.5K
47.2K
89.1K
50.8K
53.3K
92.6K
55.2K
85.2K
54.7K
84.2K
83.1K
55.3K
83.9K
89.4K
46.2K
87.3K
85.1K
48.9K
92.4K
95.3K
55.7K
82.9K
56.9K
64.2K
65.3K
59.4K
66.9K
78.3K
65.8K
79.2K
93.8K
89.5K
59.2K
58.9K
64.3K
75.3K
79.3K
79.3K
75.9K
78.3K
91.2K
93.2K
Error
for ES
+2.2K
-3.3K
+1.4K
-2.7K
-4.0K
+1.2K
-1.4K
+4.3K
-2.3K
+5.2K
-3.8K
+2.6K
-4.0K
+1.6K
+2.7K
-3.7K
-6.4K
-2.8K
+6.2K
-3.9K
-5.6K
-3.6K
+2.0K
+5.9K
-1.6K
+2.3K
-8.3K
-2.1K
-2.5K
+1.4K
+8.4K
-20.7
-17.2K
+9.4K
+10.3K
-32.7K
+14.5K
-10.4K
-9.9K
-10.3K
-9.9K
-10.7K
+14.2K
-14.1K
-8.0K
+10.2K
-7.4K
+2.9K
-8.0K
-11.9K
Predicted
FEN for
PSO
46.8K
41.6K
76.2K
44.2K
49.4K
85.5K
45.6K
43.5K
92.1K
49.2K
46.3K
89.3K
57.1K
55.2K
89.5K
75.4K
79.4K
48.2K
78.9K
79.3K
51.6K
86.3K
51.3K
46.8K
91.4K
89.4K
54.2K
81.3K
86.3K
55.8K
62.4K
51.9K
60.2K
55.3K
64.5K
50.3K
89.2K
62.7K
89.0K
89.3K
65.2K
61.2K
50.9K
71.2K
73.1K
74.3K
69.K
75.3K
87.2K
78.2K
Actual
FEN for
PSO
43.2K
48.3K
80.1K
43.6K
46.8K
83.2K
47.2K
48.2K
89.5K
50.2K
48.9K
87.5K
55.3K
54.7K
93.2K
79.1K
81.7K
41.9K
81.6K
78.9K
49.4K
84.5K
48.0K
49.1K
93.8K
87.2K
68.9K
81.5K
81.1K
57.6K
53.2K
65.2K
69.6K
71.4K
74.6K
61.4K
78.9K
73.2K
97.3K
68.4K
64.2K
69.4K
59.1K
63.4K
65.3K
70.1K
77.4K
68.3K
95.9K
87.3K
Error
for
PSO
+3.6K
-6.7K
-3.9K
+1.4K
+2.6K
+2.2K
-1.6K
-4.7K
+2.6K
-1.0K
-2.6K
+1.8K
+1.8K
-2.5K
-3.7K
-3.7K
-2.3K
+6.3K
-2.7K
+0.4K
2.2K
+1.8K
+3.3K
-2.3K
-2.4K
+2.2K
-14.7K
-0.2K
5.2K
-1.8K
+9.2K
-13.3K
-9.4K
-16.1K
-10.1K
-11.1K
+10.3K
-10.5K
-8.3K
+20.9K
-8.0K
-8.2K
-8.2K
+7.8K
+7.8K
+6.2K
-7.7K
+7.0K
-8.7K
-9.1K
Table 2: Predicted and actual most suited algorithm type/required FEN for Sphere function with linear constraint(s). The
prediction model is trained with pareto front line (PF-PM) from multi objective evolver population. DE hard/easy (1 C) is a
problem instance that is hard/easy for DE algorithm but easy/hard for the others.
Instances
name
DE hard (1 c)
ES hard (1 c)
PSO hard (1 c)
DE hard (2 c)
ES hard (2 c)
PSO hard (2 c)
DE hard (3 c)
ES hard (3 c)
PSO hard (3 c)
DE hard (4 c)
ES hard (4 c)
PSO hard (4 c)
DE hard (5 c)
ES hard (5 c)
PSO hard (5 c)
DE easy (1 c)
ES easy (1 c)
PSO easy (1 c)
DE easy (2 c)
ES easy (2 c)
PSO easy (2 c)
DE easy (3 c)
ES easy (3 c)
PSO easy (3 c)
DE easy (4 c)
ES easy (4 c)
PSO easy (4 c)
DE easy (5 c)
ES easy (5 c)
PSO easy (5 c)
Rnd. 1 (1 c)
Rnd. 2 (1 c)
Rnd. 3 (2 c)
Rnd. 4 (2 c)
Rnd. 5 (3 c)
Rnd. 6 (3 c)
Rnd. 7 (4 c)
Rnd. 8 (4 c)
Rnd. 9 (5 c)
Rnd. 10 (5 c)
Rnd. 11 (1 c)
Rnd. 12 (1 c)
Rnd. 13 (2 c)
Rnd. 14 (2 c)
Rnd. 15 (3 c)
Rnd. 16 (3 c)
Rnd. 17 (4 c)
Rnd. 18 (4 c)
Rnd. 19 (5 c)
Rnd. 20 (5 c)
Predicted Actual Error Predicted
alg.
alg.
FEN for
DE
ES
ES
NO
82.7K
PSO
DE
YES 42.3K
DE
DE
NO
35.1K
ES
ES
NO
89.4K
DE
DE
NO
48.5K
DE
DE
NO
42.3K
ES
ES
NO
87.6K
DE
PSO
YES 49.3K
DE
DE
NO
47.5K
ES
ES
NO
94.9K
PSO
PSO
NO
52.4K
ES
ES
NO
55.1K
ES
ES
NO
94.1K
DE
DE
NO
51.9K
ES
DE
YES 54.7K
PSO
DE
YES 72.3K
ES
ES
NO
78.4K
PSO
PSO
NO
91.4K
DE
DE
NO
48.1K
ES
ES
NO
79.4K
PSO
PSO
NO
83.5K
DE
DE
NO
42.4K
ES
ES
NO
90.4K
PSO
PSO
NO
86.4K
DE
DE
NO
50.8K
ES
ES
NO
86.3K
PSO
DE
YES 60.4K
DE
DE
NO
56.7K
ES
ES
NO
87.3K
PSO
PSO
NO
88.2K
PSO
PSO
NO
60.8K
DE
DE
NO
67.3K
DE
DE
NO
51.5K
ES
ES
NO
65.9K
ES
DE
YES 59.2K
PSO
PSO
NO
74.9K
ES
ES
NO
87.6K
ES
DE
YES 78.3K
DE
DE
NO
87.9K
PSO
PSO
NO
79.1K
DE
DE
NO
51.2K
DE
DE
NO
49.2K
PSO
PSO
NO
53.2K
PSO
PSO
NO
72.9K
ES
PSO
YES 65.1K
DE
DE
NO
63.1K
ES
ES
NO
69.9K
PSO
PSO
NO
74.8K
DE
ES
YES 85.1K
PSO
PSO
NO
85.1K
Actual
FEN
for DE
86.3K
45.7K
37.2K
87.4K
46.4K
43.6K
92.4K
51.2K
49.6K
93.7K
49.2K
52.7K
96.3K
53.8K
51.3K
48.9K
81.4K
87.1K
44.5K
83.6K
81.4K
48.4K
87.8K
89.5K
53.2K
89.4K
55.5K
57.8K
91.9K
91.3K
65.7K
61.9K
59.8K
72.2K
65.4K
82.4K
83.2K
71.1K
82.7K
93.2K
44.6K
45.2K
61.2K
79.0K
71.9K
69.2K
73.9K
85.3K
93.9K
90.9K
Error
for DE
-3.6
-3.4K
-2.1K
2.0K
2.1K
-1.3K
-4.8K
-1.9K
-2.1K
1.2K
3.2K
2.4K
-2.2
-1.9K
3.4K
23.4K
-3.0K
4.3K
3.6K
-4.2K
2.1K
-6.0K
2.6K
-3.1K
-2.4K
-3.1K
4.9K
-1.1K
-4.6K
-3.1K
-4.9K
+5.4K
-8.3K
-6.3K
-6.2K
-7.5K
+4.4K
+7.2K
+5.2K
-14.1K
+6.6K
+4.0K
-8.0K
-6.1K
-6.8K
-7.9K
-4.0K
16
-10.5K
-8.1K
-5.8K
Predicted
FEN for
ES
43.5K
86.4K
43.8K
43.1K
86.9K
47.2K
42.9K
94.5K
55.7K
49.2K
86.9K
51.9K
55.9K
95.1K
53.1K
80.4K
47.2K
80.7K
85.9K
52.9K
79.3K
86.2K
47.9K
94.1K
84.2K
46.2K
88.2K
92.5K
52.8K
85.2K
60.2K
68.1K
59.9K
64.1K
57.1K
69.2K
71.3K
70.5K
88.2K
85.6K
52.1K
52.1K
60.5K
65.9K
59.9K
72.5K
68.4K
84.0K
85.8K
87.9K
Actual
FEN
for ES
41.5K
84.2K
41.8K
45.2K
88.3K
45.1K
45.2K
91.2K
53.5K
47.2K
89.1K
50.8K
53.3K
92.6K
55.2K
85.2K
54.7K
84.2K
83.1K
55.3K
83.9K
89.4K
46.2K
87.3K
85.1K
48.9K
92.4K
95.3K
55.7K
82.9K
56.9K
64.2K
65.3K
59.4K
66.9K
78.3K
65.8K
79.2K
93.8K
89.5K
59.2K
58.9K
64.3K
75.3K
79.3K
79.3K
75.9K
78.3K
91.2K
93.2K
Error
for ES
2.0K
2.2K
2.0K
-2.1K
-1.4K
2.1K
-2.3K
3.3K
2.2K
2.0K
-2.2K
1.1K
2.6K
2.5K
-2.1K
-4.8K
-7.5K
-3.5K
2.8K
-2.4K
-4.6K
-3.2K
1.7K
6.8K
-0.9K
-2.7K
-4.2K
-2.7K
-2.9K
2.3K
+3.3K
+3.9K
-5.2K
+4.7K
-9.8K
-9.1K
+5.5K
-8.7K
-5.6K
-3.9K
-7.1K
-6.8K
-3.8K
-9.4K
-19.4K
-6.8K
-7.5K
+5.7K
-6.1K
-5.3K
Predicted
FEN for
PSO
47.3K
40.3K
77.4K
45.7K
48.2K
84.6K
46.2K
50.7K
93.1K
53.6K
45.1K
85.2K
58.3K
52.9K
89.4K
71.2K
78.9K
45.8K
85.1K
80.7K
47.1K
87.5K
50.5K
47.9K
95.3K
90.3K
58.8K
83.2K
85.4K
56.2K
48.8K
71.3K
63.8K
64.7K
67.3K
55.1K
72.1K
77.9K
91.9K
75.6K
57.1K
62.4K
52.7K
59.6K
60.9K
74.2K
70.3K
72.6K
98.9K
82.5K
Actual
FEN for
PSO
43.2K
48.3K
80.1K
43.6K
46.8K
83.2K
47.2K
48.2K
89.5K
50.2K
48.9K
87.5K
55.3K
54.7K
93.2K
79.1K
81.7K
41.9K
81.6K
78.9K
49.4K
84.5K
48.0K
49.1K
93.8K
87.2K
68.9K
81.5K
81.1K
57.6K
53.2K
65.2K
69.6K
71.4K
74.6K
61.4K
78.9K
73.2K
97.3K
68.4K
64.2K
69.4K
59.1K
63.4K
65.3K
70.1K
77.4K
68.3K
95.9K
87.3K
Error
for
PSO
4.1K
-8.0K
-2.7K
2.1K
1.4K
1.4K
-1.0
2.5K
3.6K
3.4K
-3.8K
-2.3K
3.0K
-1.8K
-3.8K
-7.9K
-2.8K
3.9K
3.5K
1.8K
-2.3K
3.0K
2.5K
-1.2K
1.5K
3.1K
-10.1K
1.7K
4.3K
-1.4K
-4.4K
+6.1K
-5.8K
-6.7K
-7.3K
-6.3K
-6.8K
+4.7K
-5.4K
+7.2K
-7.1K
-7.0K
-6.4K
-3.8K
-4.4K
6.1K
-7.1K
+4.3K
+3.0K
-4.8K
Table 3: Predicted and actual most suited algorithm type/required FEN for Sphere function with linear constraint(s). The
prediction model is trained with random points (RO-PM) from multi objective evolver population. DE hard/easy (1 C) is a
problem instance that is hard/easy for DE algorithm but easy/hard for the others.
Instances
name
DE hard (1 c)
ES hard (1 c)
PSO hard (1 c)
DE hard (2 c)
ES hard (2 c)
PSO hard (2 c)
DE hard (3 c)
ES hard (3 c)
PSO hard (3 c)
DE hard (4 c)
ES hard (4 c)
PSO hard (4 c)
DE hard (5 c)
ES hard (5 c)
PSO hard (5 c)
DE easy (1 c)
ES easy (1 c)
PSO easy (1 c)
DE easy (2 c)
ES easy (2 c)
PSO easy (2 c)
DE easy (3 c)
ES easy (3 c)
PSO easy (3 c)
DE easy (4 c)
ES easy (4 c)
PSO easy (4 c)
DE easy (5 c)
ES easy (5 c)
PSO easy (5 c)
Rnd. 1 (1 c)
Rnd. 2 (1 c)
Rnd. 3 (2 c)
Rnd. 4 (2 c)
Rnd. 5 (3 c)
Rnd. 6 (3 c)
Rnd. 7 (4 c)
Rnd. 8 (4 c)
Rnd. 9 (5 c)
Rnd. 10 (5 c)
Rnd. 11 (1 c)
Rnd. 12 (1 c)
Rnd. 13 (2 c)
Rnd. 14 (2 c)
Rnd. 15 (3 c)
Rnd. 16 (3 c)
Rnd. 17 (4 c)
Rnd. 18 (4 c)
Rnd. 19 (5 c)
Rnd. 20 (5 c)
Predicted Actual Error Predicted
alg.
alg.
FEN for
DE
PSO
ES
YES 93.6K
DE
DE
NO
51.8K
PSO
DE
YES 58.8K
ES
ES
NO
66.8K
PSO
DE
YES 67.2K
DE
DE
NO
55.7K
PSO
ES
YES 73.4K
DE
PSO
YES 62.7K
ES
DE
YES 59.2K
PSO
ES
YES 88.2K
ES
PSO
YES 39.8K
DE
ES
YES 63.9K
PSO
ES
YES 102.6K
PSO
DE
YES 47.9K
ES
DE
YES 59.9K
DE
DE
NO
41.2K
ES
ES
NO
68.2K
ES
PSO
YES 81.2K
DE
DE
NO
54.2K
PSO
ES
YES 75.3K
PSO
PSO
YES 68.3K
DE
DE
NO
59.1K
PSO
ES
YES 73.9K
PSO
PSO
NO
69.2K
DE
DE
NO
64.2K
ES
ES
NO
78.3K
PSO
DE
YES 61.3K
PSO
DE
YES 71.4K
ES
ES
NO
86.9K
PSO
PSO
NO
84.1K
PSO
PSO
NO
59.1K
DE
DE
NO
55.2K
DE
DE
NO
49.2K
ES
ES
NO
64.1K
ES
DE
YES 68.3K
PSO
PSO
NO
73.9K
PSO
ES
YES 78.3K
ES
DE
YES 76.9K
DE
DE
NO
88.3K
PSO
PSO
NO
80.8K
ES
DE
YES 52.9K
DE
DE
NO
50.2K
PSO
PSO
NO
67.3K
PSO
PSO
NO
70.3K
ES
PSO
YES 66.9K
DE
DE
NO
61.2K
ES
ES
NO
67.9K
PSO
PSO
NO
72.7K
DE
ES
YES 83.1K
PSO
PSO
NO
84.1K
Actual
FEN
for DE
86.3K
45.7K
37.2K
87.4K
46.4K
43.6K
92.4K
51.2K
49.6K
93.7K
49.2K
52.7K
96.3K
53.8K
51.3K
48.9K
81.4K
87.1K
44.5K
83.6K
81.4K
48.4K
87.8K
89.5K
53.2K
89.4K
55.5K
57.8K
91.9K
91.3K
65.7K
61.9K
59.8K
72.2K
65.4K
82.4K
83.2K
71.1K
82.7K
93.2K
44.6K
45.2K
61.2K
79.0K
71.9K
69.2K
73.9K
85.3K
93.9K
90.9K
Error
for DE
+7.3K
+6.1K
+21.6K
-20.6K
+20.8K
+12.1K
-19.0K
+11.5K
+9.6K
-5.5K
-9.4K
+11.2K
+6.3K
-5.9K
+8.6K
-7.7K
-13.2K
4.1K
9.7K
-8.3K
-13.1K
10.7K
-13.9K
-20.3K
11K
-11.1K
5.8K
13.6K
-5.0K
-7.2K
-6.6K
-6.7K
-10.6K
-8.1K
+2.9K
-8.5K
-4.9K
+5.8K
+5.6K
-12.4K
+8.3K
+5K
6.1K
-8.7K
-5K
-8K
-6K
17
-12.6K
-10.8K
-6.8K
Predicted
FEN for
ES
53.5K
78.1K
73.5K
57.9K
73.2K
73.6K
59.2K
73.6K
43.2K
51.2K
96.3K
64.6K
62.3K
84.8K
45.2K
78.2K
64.2K
62.1K
68.4K
69.2K
94.1K
78.2K
55.2K
76.4K
78.3K
59.1K
85.9K
99.1K
62.3K
76.9K
51.3K
67.4K
57.9K
63.2K
63.5K
66.2K
73.5K
68.3K
90.9K
83.5K
52.4K
51.2K
60.8K
68.3K
55.2K
70.5K
65.2K
80.3K
83.9K
85.9K
Actual
FEN
for ES
41.5K
84.2K
41.8K
45.2K
88.3K
45.1K
45.2K
91.2K
53.5K
47.2K
89.1K
50.8K
53.3K
92.6K
55.2K
85.2K
54.7K
84.2K
83.1K
55.3K
83.9K
89.4K
46.2K
87.3K
85.1K
48.9K
92.4K
95.3K
55.7K
82.9K
56.9K
64.2K
65.3K
59.4K
66.9K
78.3K
65.8K
79.2K
93.8K
89.5K
59.2K
58.9K
64.3K
75.3K
79.3K
79.3K
75.9K
78.3K
91.2K
93.2K
Error
for ES
+12K
-6.1K
+31.7K
+12.7K
-15.2K
+28.5K
+14.0K
-17.6K
-10.3K
+4.0K
+7.2K
+13.8K
+9.0K
-7.8K
-10.0K
-7K
9.6K
-22.1K
-14.7K
13.9K
10.2K
-11.2K
9K
-10.9K
-6.8K
10.2K
-6.5K
3.8K
6.6K
-6.0K
-5.6K
+3.2K
-7.4K
+3.8
-3.4K
-12.1K
+7.7K
-10.9K
-2.9K
-6K
-6.8K
-7.7K
-3.5K
-7K
-24.1K
-8.8K
-10.7K
+2K
-7.3K
-7.3K
Predicted
FEN for
PSO
38.1K
53.7K
57.3K
72.8K
53.7K
77.8K
57.3K
64.3K
103.6K
49.1K
41.2K
83.8K
49.2K
47.1K
82.7K
67.2K
72.9K
63.4K
71.4K
65.1K
58.1K
80.4K
42.4K
58.2K
87.3K
82.5K
60.4K
70.9K
78.9K
63.4K
50.9K
69.8K
73.4K
65.3K
69.9K
56.8K
71.2K
76.3K
90.2K
75.3K
54.3K
62.5K
49.2K
57.9K
58.3K
73.9K
69.2K
70.2K
99.1K
83.5K
Actual
FEN for
PSO
43.2K
48.3K
80.1K
43.6K
46.8K
83.2K
47.2K
48.2K
89.5K
50.2K
48.9K
87.5K
55.3K
54.7K
93.2K
79.1K
81.7K
41.9K
81.6K
78.9K
49.4K
84.5K
48.0K
49.1K
93.8K
87.2K
68.9K
81.5K
81.1K
57.6K
53.2K
65.2K
69.6K
71.4K
74.6K
61.4K
78.9K
73.2K
97.3K
68.4K
64.2K
69.4K
59.1K
63.4K
65.3K
70.1K
77.4K
68.3K
95.9K
87.3K
Error
for
PSO
-5.1K
+5.4K
-22.8K
+29.2K
+6.9K
-5.4K
+10.1K
+16.1K
+14.1K
-1.1K
-7.7K
-3.7K
-6.1K
-7.6K
-10.5K
-11.9K
-8.8K
21.5K
-10.2K
-13.8K
8.7K
-4.1K
-5.6K
9.1K
-6.5K
-4.7K
-8.5K
-10.6K
-2.2K
5.8K
-2.3K
+4.6K
+3.8K
-6.1K
-4.7K
-4.6K
-7.7K
+3.1K
-7.1K
+6.9K
-9.9K
-6.9K
-9.9K
-5.5K
-7K
3.8K
-8.2K
+1.9K
+3.2K
-3.8K
Table 4: Predicted and actual most suited algorithm type/required FEN for Sphere function with linear constraint(s). The
prediction model is trained with combination of pareto front and random points (PFR-PM) from multi objective evolver
population. DE hard/easy (1 C) is a problem instance that is hard/easy for DE algorithm but easy/hard for the others.
Instances
name
DE hard (1 c)
ES hard (1 c)
PSO hard (1 c)
DE hard (2 c)
ES hard (2 c)
PSO hard (2 c)
DE hard (3 c)
ES hard (3 c)
PSO hard (3 c)
DE hard (4 c)
ES hard (4 c)
PSO hard (4 c)
DE hard (5 c)
ES hard (5 c)
PSO hard (5 c)
DE easy (1 c)
ES easy (1 c)
PSO easy (1 c)
DE easy (2 c)
ES easy (2 c)
PSO easy (2 c)
DE easy (3 c)
ES easy (3 c)
PSO easy (3 c)
DE easy (4 c)
ES easy (4 c)
PSO easy (4 c)
DE easy (5 c)
ES easy (5 c)
PSO easy (5 c)
Rnd. 1 (1 c)
Rnd. 2 (1 c)
Rnd. 3 (2 c)
Rnd. 4 (2 c)
Rnd. 5 (3 c)
Rnd. 6 (3 c)
Rnd. 7 (4 c)
Rnd. 8 (4 c)
Rnd. 9 (5 c)
Rnd. 10 (5 c)
Rnd. 11 (1 c)
Rnd. 12 (1 c)
Rnd. 13 (2 c)
Rnd. 14 (2 c)
Rnd. 15 (3 c)
Rnd. 16 (3 c)
Rnd. 17 (4 c)
Rnd. 18 (4 c)
Rnd. 19 (5 c)
Rnd. 20 (5 c)
Predicted Actual Error Predicted
alg.
alg.
FEN for
DE
ES
ES
NO
83.8K
DE
DE
NO
43.6K
DE
DE
NO
39.1K
PSO
ES
YES 85.2K
DE
DE
NO
45.8K
DE
DE
NO
41.8K
ES
ES
NO
89.5K
PSO
PSO
NO
47.8K
DE
DE
NO
46.8K
ES
ES
NO
96.2K
PSO
PSO
NO
50.4K
ES
ES
NO
50.3K
PSO
ES
YES 97.2K
DE
DE
NO
52.5K
DE
DE
NO
50.2K
DE
DE
NO
46.3K
ES
ES
NO
86.6K
PSO
PSO
NO
86.3K
DE
DE
NO
48.1K
ES
ES
NO
77.2K
PSO
PSO
NO
83.2K
DE
DE
NO
42.5K
PSO
ES
YES 90.9K
PSO
PSO
NO
87.8K
DE
DE
NO
50.2K
ES
ES
NO
86.7K
DE
DE
NO
56.7K
DE
DE
NO
55.9K
ES
ES
NO
89.4K
PSO
PSO
NO
92.4K
PSO
PSO
NO
63.4K
DE
DE
NO
64.6K
DE
DE
NO
55.0K
ES
ES
NO
68.6K
DE
DE
NO
61.9K
PSO
PSO
NO
78.3K
ES
ES
NO
86.4K
ES
DE
YES 74.6K
DE
DE
NO
84.2K
PSO
PSO
NO
83.7K
DE
DE
NO
48.5K
DE
DE
NO
48.7K
PSO
PSO
NO
55.7K
PSO
PSO
NO
76.0K
ES
PSO
YES 68.8K
DE
DE
NO
66.5K
ES
ES
NO
70.4K
PSO
PSO
NO
75.4K
DE
ES
YES 86.9K
PSO
PSO
NO
87.4K
Actual
FEN
for DE
86.3K
45.7K
37.2K
87.4K
46.4K
43.6K
92.4K
51.2K
49.6K
93.7K
49.2K
52.7K
96.3K
53.8K
51.3K
48.9K
81.4K
87.1K
44.5K
83.6K
81.4K
48.4K
87.8K
89.5K
53.2K
89.4K
55.5K
57.8K
91.9K
91.3K
65.7K
61.9K
59.8K
72.2K
65.4K
82.4K
83.2K
71.1K
82.7K
93.2K
44.6K
45.2K
61.2K
79.1K
71.9K
69.2K
73.9K
85.3K
93.9K
90.9K
Error
for DE
-2.5K
-2.1K
1.9K
-2.2K
-0.6K
-1.8K
-2.9K
-3.4K
-2.8K
2.5K
1.2K
-2.4K
0.9K
-1.3K
-1.1K
-2.6K
5.2K
-0.8K
3.6K
-6.4K
1.8K
-5.9K
3.1K
-1.7K
-3.0K
-2.7K
1.2K
-1.9K
-2.5K
1.1K
-2.3K
+2.7K
-4.8K
-3.6K
-3.5K
-4.1K
+3.2K
+3.5K
+1.5K
-9.5K
+3.9K
+3.5K
-5.5K
-3.1K
-3.1K
-2.7K
-3.5K
18
-9.9K
-7.0K
-3.5K
Predicted
FEN for
ES
40.2K
80.2K
42.4K
43.2K
85.3K
47.9K
43.1K
94.7K
50.2K
49.2K
84.7K
53.6K
55.9K
94.2K
56.3K
82.7K
48.2K
82.4K
87.4K
52.3K
78.4K
86.8K
48.9K
92.4K
83.4K
50.8K
87.9K
93.1K
56.3K
84.1K
55.4K
64.7K
62.3K
61.4K
62.5K
73.9K
69.4K
73.7K
87.6K
84.3K
55.8K
53.6K
65.9K
66.6K
62.7K
72.2K
70.3K
83.5K
88.5K
90.3K
Actual
FEN
for ES
41.5K
84.2K
41.8K
45.2K
88.3K
45.1K
45.2K
91.2K
53.5K
47.2K
89.1K
50.8K
53.3K
92.6K
55.2K
85.2K
54.7K
84.2K
83.1K
55.3K
83.9K
89.4K
46.2K
87.3K
85.1K
48.9K
92.4K
95.3K
55.7K
82.9K
56.9K
64.2K
65.3K
59.4K
66.9K
78.3K
65.8K
79.2K
93.8K
89.5K
59.2K
58.9K
64.3K
75.3K
79.3K
79.3K
75.9K
78.3K
91.2K
93.2K
Error
for ES
-1.3K
-4.0K
0.6K
-2.0K
-3.0K
2.8K
-2.1K
3.5K
-3.3K
2.0K
-4.4K
2.8K
2.6K
1.6K
1.1K
-2.5K
-6.5K
-1.8K
4.3K
-3.0K
-5.5K
-2.6K
2.7K
5.1K
-1.7K
1.9K
-4.5K
-2.2K
0.6K
1.2K
-1.5K
+0.5K
-3.0K
+2.0K
-4.4K
-4.4K
+3.6K
-5.5K
-6.2K
-5.2K
-0.4K
-5.3K
+1.6K
-8.7K
-16.6K
-7.1K
-5.6K
+5.2K
-2.7L
-2.9K
Predicted
FEN for
PSO
45.8K
45.1K
78.1K
41.9K
48.9K
85.3K
44.2K
42.9K
92.5K
51.2K
46.8K
90.2K
52.4K
53.8K
90.5K
76.2K
80.2K
45.7K
79.4K
77.4K
48.1K
87.4K
46.9K
47.0K
91.8K
89.8K
55.2K
82.4K
85.9K
56.1K
51.7K
65.7K
67.1K
68.2K
70.8K
57.8K
74.7K
75.8K
94.2K
65.8K
60.9K
63.6K
54.9K
61.7K
63.1K
71.5K
71.6K
71.2K
97.4K
85.9K
Actual
FEN for
PSO
43.2K
48.3K
80.1K
43.6K
46.8K
83.2K
47.2K
48.2K
89.5K
50.2K
48.9K
87.5K
55.3K
54.7K
93.2K
79.1K
81.7K
41.9K
81.6K
78.9K
49.4K
84.5K
48.0K
49.1K
93.8K
87.2K
68.9K
81.5K
81.1K
57.6K
53.2K
65.2K
69.6K
71.4K
74.6K
61.4K
78.9K
73.2K
97.3K
68.4K
64.2K
69.4K
59.1K
63.4K
65.3K
70.1K
77.4K
68.3K
95.9K
87.3K
Error
for
PSO
2.6K
-3.2K
-2.0K
-1.7K
2.1K
2.1K
-3.0K
-5.3K
3.0K
1.0K
-2.1K
2.7K
-2.9K
-0.9K
-2.7K
-2.9K
-1.5K
3.8K
-2.2K
-1.5K
-1.3K
2.9K
-1.1K
-2.1K
-2.0K
2.6K
-13.7K
0.9K
4.8K
-1.5K
-1.5K
+0.5K
-2.5K
-3.2K
-3.8K
-3.6K
-4.2K
+2.6K
-3.1K
-2.6K
-3.3K
-5.8K
-4.2K
-1.7K
-2.2K
+3.4K
-5.8K
+2.9K
-1.5K
-1.4K
Table 5: Comparison of PFR-PM with RO-PM models for Sphere function with various types of constraints (linear,
quadratic and their combination). Average deviation of FEN denotes the average of differences between actual and
predicted required FEN for PFR-PM and RO-PM.
Problem
Success Rate
RO-PM
Success Rate
FR-PM
Sphere, 1lin
Sphere, 2lin
Sphere, 3lin
Sphere, 4lin
Sphere, 5lin
Sphere, 1Quad
Sphere, 2Quad
Sphere, 3Quad
Sphere, 4Quad
Sphere, 5Quad
Sphere, 4Lin, 1Quad
Sphere, 3Lin, 2Quad
Sphere, 2Lin, 3Quad
Sphere, 1Lin, 4Quad
2
1
1
5
1
1
1
3
1
5
2
2
3
5
26
27
26
28
28
27
26
28
29
28
24
26
28
27
Average deviation of FEN
for RO-PM
7.8K
8.6K
12.6K
13.6K
17.4K
11.9K
13.4K
15.8K
19.3K
21.6K
13.4K
13.8K
16.1K
18.9K
Average deviation of FEN
for PFR-PM
2.4K
1.8K
3.2K
3.5K
2.7K
2.1K
2.6K
3.7K
3.1K
4.3K
2.5K
2.8K
3.8K
3.7K
P
value
0.004
0.013
0.018
0.006
0.028
0.035
0.038
0.043
0.026
0.035
0.007
0.004
0.031
0.016
Table 6: Comparison of PFR-PM with RO-PM models for Ackley function with various types of constraints (linear,
quadratic and their combination). Average deviation of FEN denotes the average of differences between actual and
predicted required FEN for PFR-PM and RO-PM.
Problem
Success Rate
RO-PM
Success Rate
FR-PM
Ackley, 1lin
Ackley, 2lin
Ackley, 3lin
Ackley, 4lin
Ackley, 5lin
Ackley, 1Quad
Ackley, 2Quad
Ackley, 3Quad
Ackley, 4Quad
Ackley, 5Quad
Ackley, 4Lin, 1Quad
Ackley, 3Lin, 2Quad
Ackley, 2Lin, 3Quad
Ackley, 1Lin, 4Quad
0
1
2
1
6
2
1
0
3
2
4
2
3
0
27
27
25
28
29
29
27
25
27
29
24
26
24
28
19
Average deviation of FEN
for RO-PM
9.3K
11.5K
10.3K
14.7K
13.8K
16.3K
17.7K
18.3K
16.9K
21.9K
15.8K
16.7K
16.8K
19.8K
Average deviation of FEN
for PFR-PM
2.5K
3.2K
2.7K
3.6K
4.5K
3.5K
4.1K
3.7K
5.1K
5.8K
4.2K
4.6K
4.9K
4.3K
P
value
0.043
0.016
0.004
0.008
0.025
0.046
0.026
0.043
0.048
0.034
0.032
0.012
0.006
0.021
Table 7: Comparison of PFR-PM with RO-PM models for Rosenbrock function with various types of constraints (linear,
quadratic and their combination). Average deviation of FEN denotes the average of differences between actual and
predicted required FEN for PFR-PM and RO-PM.
Problem
Success Rate
RO-PM
Success Rate
FR-PM
Rosenbrock, 1lin
Rosenbrock, 2lin
Rosenbrock, 3lin
Rosenbrock, 4lin
Rosenbrock, 5lin
Rosenbrock, 1Quad
Rosenbrock, 2Quad
Rosenbrock, 3Quad
Rosenbrock, 4Quad
Rosenbrock, 5Quad
Rosenbrock, 4Lin, 1Quad
Rosenbrock, 3Lin, 2Quad
Rosenbrock, 2Lin, 3Quad
Rosenbrock, 1Lin, 4Quad
2
0
3
4
5
2
1
4
3
2
1
0
4
2
26
26
25
27
28
27
29
26
28
28
28
24
25
27
20
Average deviation of FEN
for RO-PM
10.3K
11.5K
12.7K
15.8K
19.4K
17.4K
21.5K
21.3K
18.5K
24.6K
14.7K
17.4K
19.5K
21.3K
Average deviation of FEN
for PFR-PM
3.3K
4.6K
3.6K
5.2K
5.1K
4.1K
4.7K
5.7K
5.2K
6.9K
3.6K
4.7K
3.6K
5.2K
P
value
0.038
0.035
0.002
0.035
0.028
0.017
0.043
0.037
0.043
0.004
0.004
0.024
0.029
0.006
| 9cs.NE |
Information-geometrical characterization of
statistical models which are
statistically equivalent to probability simplexes
arXiv:1701.07736v2 [cs.IT] 8 Feb 2017
Hiroshi Nagaoka
Graduate School of Informatics and Engineering
The University of Electro-Communications
Chofu, Tokyo 182-8585, Japan
Email: nagaoka@is.uec.ac.jp
Abstract—The probability simplex is the set of all probability distributions on a finite set and is the most fundamental
object in the finite probability theory. In this paper we give
a characterization of statistical models on finite sets which are
statistically equivalent to probability simplexes in terms of αfamilies including exponential families and mixture families. The
subject has a close relation to some fundamental aspects of
information geometry such as α-connections and autoparallelity.
II. S TATEMENT OF
THE MAIN RESULT
We begin with giving some basic definitions which are
necessary to state our problem.
For an arbitrary finite set X , let P(X ) and P(X ) be
the sets of probability distributions and of strictly positive
probability distributions on X ;
X
P(X ) := {p | p : X → [0, 1],
p(x) = 1}
x
I. A N
INTRODUCTORY EXAMPLE
Let X = {0, 1, 2} and let M = {pλ | 0 < λ < 1} be the
set of probability distributions on X of the form
pλ = (pλ (0), pλ (1), pλ (2)) = (λ, (1 − λ)/2, (1 − λ)/2) .
The statistical model M has the following three properties.
Firstly, it is a mixture family since
pλ = λ (1, 0, 0) + (1 − λ) (0, 1/2, 1/2).
Secondly, it is an exponential family since
log pλ = θF − ψ(θ),
where θ = log(2λ/(1 − λ)), (F (0), F (1), F (2)) = (1, 0, 0)
and ψ(θ) = − log(1 − λ)/2 = log(2 + eθ ). Lastly, M is
statistically equivalent to the 1-dimensional open probability
simplex P1 = {(λ, 1 − λ) | 0 < λ < 1} in the sense that there
exist a channel V from {0, 1} to X and a channel W from
X to {0, 1} such that M is the set of output distributions of
V for input distributions in P1 and that V is invertible by W .
The matrix representations of these channels are given by
1 0
1 0 0
V = 0 1/2 , W =
.
0 1 1
0 1/2
Note that the invertibility W V = I holds.
Our aim is to show the equivalence between the first two
properties and the last one.
P(X ) := {p | p : X → (0, 1),
X
p(x) = 1}.
x
In particular, let for an arbitrary positive integer d
P d := P({0, 1, . . . , d})
Pd := P({0, 1, . . . , d}),
which we call the d-dimensional (closed and open) probability
simplexes.
A mapping Γ : P(X ) → P(Y ), where X and Y are
finite sets, is called a Markov map when there exists a channel
W (y|x) from X to Y such that, for any p ∈ P(X ),
X
Γ(p) =
W ( · |x)p(x).
x
i.e., Γ(p) is the output distribution of the channel W corresponding to the input distribution p. Note that a Markov
map is affine; Γ(λp + (1 − λ)q) = λΓ(p) + (1 − λ)Γ(q) for
∀p, q ∈ P(X ) and 0 ≤ ∀λ ≤ 1.
Let M and N be smooth submanifolds (statistical models)
of P(X ) and P(Y ), respectively. When there exist a pair
of Markov maps Γ : P(X ) → P(Y ) and ∆ : P(Y ) →
P(X ) such that their restrictions Γ|M and ∆|N are bijections
between M and N and are the inverse mappings of each other,
we say that M and N are Markov equivalent or statistically
equivalent and wite as M ≃ N .
The aim of this paper is to give a characterization of statistical models which are statistically equivalent to probability
simplexes. The main result is as follows.
Theorem 1 For an arbitrary smooth submanifold M of
P(X ), the following conditions are mutually equivalent.
For an arbitrary α ∈ R, define a function L(α) : R+ (=
(0, ∞)) → R by1
1−α
(α 6= 1)
u 2
(α)
(5)
L (u) =
log u (α = 1).
(i) M ≃ Pd , where d = dim M .
(ii) M is an exponential family and is a mixture
family.
(iii) ∃α 6= ∃β, M is an α-family and is an β-family.
(iv) ∀α, M is an α-family.
Explanation of exponential family, mixture family and αfamily for arbitrary α ∈ R as well as the proof of the theorem
will be presented in subsequent sections. Here we only give
a few remarks on condition (i). Firstly, (i) is equivalent to the
condition that ∃d′ , M ≃ Pd′ , since if M ≃ Pd′ then M and
Pd′ must be diffeomorphic, so that dim M = dim Pd′ = d′ .
Secondly, (i) is equivalent to the condition M ≃ P d , where
M denotes the topological closure of M , and means that M
is the set of output distributions of an invertible (erro-free)
channel.
III. S OME
FACTS ABOUT CONDITION
(i)
From the definition of the relation ≃, condition (i) implies
that there exist Γ : P(X ) → P d and ∆ : P d →
P(X ) satisfying Γ ◦ ∆ = id (the identity on P d ). Let
{q0 , q1 , . . . , qd } ⊂ P(X ) be defined by
∆(δi ) = qi , ∀i ∈ {0, 1, . . . , d},
(1)
where δi is the delta distributions on {0, 1, . . . , d} concentrated
on i. Then it is easy to see, as is shown in Lemma 9.5 and its
“Supplement” of [1] where our ∆ is called a congruent embedding (of P d into P(X )), that the supports Ai := supp (qi )
constitute a partition of X in the sense that
Ai ∩ Aj = φ if i 6= j, and
d
[
Ai = X ,
(2)
i=0
The function L(α) is naturally extended to a mapping
(R+ )X → RX (f 7→ L(α) (f )) by
L(α) (f ) (x) = L(α) (f (x)).
(6)
For a submanifold M of P(X ), its denormalization M̃ is
defined by
M̃ := τ p | p ∈ M and τ ∈ R+ ,
(7)
where τ p denotes the function X ∋ x 7→ τ p(x) ∈ R+ . The
denormalization is an extended
Pmanifold obtained by relaxing
the normalization constraint x p(x) = 1. Obviously, M̃ is
f ), and P(X
f ) = (R+ )X is an open
a submanifold of P(X
X
subset of R . When the image
o
n
L(α) (M̃ ) = L(α) (τ p) p ∈ M and τ ∈ R+
forms an open subset of an affine subspace, say Z, of RX , M
is called an α-family. In this paper, it is assumed for simplicity
that M is maximal in the sense that
L(α) (M̃ ) = Z ∩ L(α) (R+ )X .
(8)
Since it follows from the definition (5) of L(α) that
(R+ )X (α 6= 1)
L(α) (R+ )X =
RX
(α = 1),
(8) is written as
and the left inverse Γ of ∆ is represented as
Γ(p) =
d
X
L
p(Ai ) δi , ∀p ∈ P(X ),
where p(Ai ) := x∈Ai p(x). In addition, condition (i) implies
M = ∆(Pd ) := {∆(λ) | λ ∈ Pd }, so that from (1) we have
)
( d
X
(4)
λi qi (λ0 , . . . , λd ) ∈ Pd .
M=
i=0
Conversely, if a statistical model M ⊂ P(X ) is represented
in the form (4) by a collection of d + 1 distributions {qi } on
X whose supports {Ai } constitute a partition of X , then we
see that M satisfies condition (i) by defining ∆ and Γ by (1)
and (3). Thus a necessary and sufficient condition for (i) is
obtained, which will be used in later arguments to prove the
theorem.
IV. α- FAMILY, e- FAMILY
AND
(M̃ ) =
(3)
i=0
P
(α)
m- FAMILY
Following the way developed in [5] (see also [3], [4]),
we give the definition of α-family, which includes that of
exponential family and mixture family as special cases.
Z ∩ (R+ )X
Z
(α 6= 1)
(α = 1).
(9)
Note that, as is pointed out in section 2.6 of [4], an affine
subspace Z satisfying (9) must be a linear subspace when
α 6= 1. Note also that P(X ) is an α-family for ∀α ∈ R,
corresponding to the case when Z = RX .
When α = 1, the notion of α-family is equivalent to that
of exponential family, whose general form is M = {pθ | θ =
(θ1 , . . . , θd ) ∈ Rd } such that
#
"
d
X
i
θ Fi (x) − ψ(θ) ,
(10)
pθ (x) = exp C(x) +
i=1
where C, F1 , . . . , Fd are functions on X and ψ is a function
on Rd defined by
#
"
d
X
X
i
θ Fi (x) .
(11)
ψ(θ) = log
exp C(x) +
x
i=1
1 L(α) (u) can be replaced with aL(α) (u) + b by arbitrary constants a 6= 0
and b, possibly depending on α. In [3], [4], [5], these constants are properly
chosen so that the ±α-duality and the limit of α → 1 can be treated in a
convenient way.
When α = −1, on the other hand, the notion of α-family
is equivalent to that of mixture family, whose general form is
M = {pθ | θ = (θ1 , . . . , θd ) ∈ Θ} such that
pθ (x) = C(x) +
d
X
i
θ Fi (x),
(12)
i=1
pθ (x) =
j=0
2
o 1−α
.
ξ (θ)Fj (x)
j
(13)
See §2.6 of [4] for further details.
V. P ROOF
OF
Let α be an arbitrary real number such that α 6= 1. Since
L(α) (0) = 0 in this case, it follows from the disjointness of
the supports of {qi } that
!
X 1−α
X
(α)
λi 2 L(α) (qi )
λi qi =
L
i
for any (λ0 , . . . , λd ) ∈ (R+ )d+1 . From this we have
L(α) (M̃ )
)
( d
X 1−α
(α)
+ d+1
2
λi L (qi ) (λ0 , . . . , λd ) ∈ (R )
=
i=0
=
ξi L(α) (qi ) (ξ0 , . . . , ξd ) ∈ (R+ )d+1
i=0
)
= Z ∩ (R+ )X ,
where Z is the (d + 1)-dimensional linear subspace of RX
spanned by L(α) (qi ), i ∈ {0, 1, . . . , d}. This proves that M is
an α-family for any α 6= 1.
Let α = 1. For any x ∈ X , we have
!
!
X
X
(1)
λi qi (x)
λi qi (x) = log
L
i
i
= log(λj qj (x))
= log λj + log qj (x)
X
(log λi + log qi (x)) 1Ai (x),
=
i
(
C+
i=0
d
X
+ d+1
(log λi )1Ai (λ0 , . . . , λd ) ∈ (R )
ξi 1Ai (ξ0 , . . . , ξd ) ∈ R
d+1
i=0
)
)
,
which is an affine subspace of RX . This proves that M is a
1-family (an exponential family).
The implication (i) ⇒ (iv) has thus been proved.
OF
(ii), (iii) AND (iv)
The implications (iv) ⇒ (ii) ⇒ (iii) are obvious. To see (iii)
⇒ (iv), some results of information geometry are invoked.
(i) ⇒ (iv)
i=0
( d
X
=
C+
d
X
VI. E QUIVALENCE
Assume (i), which implies that there exists a collection of
d + 1 probability distributions {qi } ⊂ P(X ) whose supports
{Ai } constitute a partition of X and that M is represented
as (4). Then the denormalization M̃ is represented as
)
( d
X
(14)
λi qi (λ0 , . . . , λd ) ∈ (R+ )d+1 .
M̃ =
i
L(1) (M̃ )
(
=
where F1 , . . . , Fd are functions on X satisfying
P
d
x Fi (x) = 0 and Θ := {θ ∈ R | ∀x, pθ (x) > 0}.
When α 6= 1, the general form of α-family M = {pθ | θ =
(θ1 , . . . , θd ) ∈ Θ} is
d
nX
where j denotes the element of {0, 1, . . . P
, d} such that x ∈ Aj .
Letting C ∈ RX be defined by C(x) = i (log qi (x))1Ai (x),
we have
Remark 1: The notion of affine connections appears only in
this section. Since the implication (ii) ⇒ (i) will be proved
in the next section without using affine connections (at least
explicitly), we do not need them in proving the equivalence
of the conditions of Theorem 1 except for (iii).
We first introduce some concepts from general differential
geometry. Let S be a smooth manifold and denote by T (S)
the set of smooth vector fields on S. Here, by a vector field
on S we mean a mapping, say X, such that X : S ∋ p 7→
Xp ∈ Tp (S), where Tp (S) denotes the tangent space of S
at p. An affine connection on S is represented by a mapping
∇ : T (S)×T (S) ∋ (X, Y ) 7→ ∇X Y ∈ T (S), which is called
a covariant derivative, satisfying certain conditions. Let M be
a smooth submanifold of S. Then ∇ is naturally defined on
T (M ) × T (M ), so that ∇X Y is defined for any vector fields
on M . However, the value ∇X Y in this case is a mapping
M ∋ p 7→ (∇X Y )p ∈ Tp (S) in general and is not a vector
field on M (i.e., ∇X Y 6∈ T (M )) unless
(∇X Y )p ∈ Tp (M ), ∀p ∈ M.
(15)
When (15) holds for ∀X, Y ∈ T (M ), M is said to be
autoparallel w.r.t. ∇ or ∇-autoparallel in S.
Let ∇, ∇′ and ∇′′ be affine connection on S for which there
exists a real number a satisfying2
∇′′ = a∇ + (1 − a)∇′ .
(16)
If a submanifold M is ∇-autoparallel and ∇′ -autoparallel,
then it is also ∇′′ -autoparallel. This implication is obvious
from (∇′′X Y )p = a(∇X Y )p + (1 − a)(∇′X Y )p and the
autoparallelity condition (15), which will be invoked later.
As was independently introduced by Čencov [1] and Amari
[2], a one-parameter family of affince connections, which are
called the α-connections (α ∈ R), are defined on a manifold
2 For arbitrary affine connections ∇ and ∇′ , their affine combination a∇ +
(1 − a)∇′ always becomes an affine connection.
of probability distributions. After Amari’s notation, the αconnection is written in the form of affine combination
1 + α (1) 1 − α (−1)
∇ +
∇
,
(17)
∇(α) =
2
2
which implies that
∇(γ) =
γ − β (α) α − γ (β)
∇ +
∇
α−β
α−β
(18)
for any α, β, γ ∈ R such that α 6= β.
When a submanifold M of S is autoparallel w.r.t. the αconnection in S, we say that M is α-autoparallel in S.
Since (18) is of the form (16), it follows that if M is αautoparallel and β-autoparallel in S for some α 6= β, then it
is γ-autoparallel in S for all γ ∈ R. On the other hand, it
was shown in [5] (see also section 2.6 of [4]) that, for any
submanifold M of P(X ) and for any real number α, M
is an α-family if and only if M is α-autoparallel in P(X ).
Combination of these two results proves (iii) ⇒ (iv).
Remark 2: Since the e-connection and the m-connection are
dual w.r.t. the Fisher information metric [3], [4], [5], condition
(ii) is a special case of doubly autoparallelity introduced by
Ohara; see [6], [7] and the reference cited there. It is pointed
out in [7] that the α-autoparallelity for all α follows from that
for α = ±1.
VII. P ROOF
OF
(ii) ⇒ (i)
Assume (ii), which means that there exist two affine subspaces Z (e) and Z (m) of RX such that
L(e) (M̃ ) = {log µ | µ ∈ M̃ } = Z (e)
(19)
L(m) (M̃ ) = M̃ = Z (m) ∩ (R+ )X ,
(20)
(e)
(1)
(m)
(−1)
(e)
where L := L and L
:= L
. Let V
and V (m)
(e)
be the linear spaces of translation vectors of Z and Z (m) ,
respectively, so that we have Z (e) = f +V (e) for any f ∈ Z (e)
and Z (m) = g + V (m) for any g ∈ Z (m)3 .
Lemma 1 V (e) is closed w.r.t. multiplication of functions;
i.e., a, b ∈ V (e) ⇒ ab ∈ V (e) , where the product ab
is defined by (ab)(x) = a(x)b(x).
Proof. The map
Φ := L(e) |M̃ : M̃ ∋ µ 7→ log µ ∈ Z (e)
(m)
(e)
(dΨa,ν )µ ◦ (dΦ)−1
∋ b 7→ ab ∈ V (e) .
ν :V
This proves that a, b ∈ V (e) ⇒ ab ∈ V (e) .
Lemma 2 V (e) contains the constant functions on X .
Proof. From the definition (7) of M̃ , for any µ ∈ M̃ and
any positive constant τ = ec , we have τ µ ∈ M̃ . This implies
that both log µ and log(τ µ) belong to Z (e) , and hence the
translation log(τ µ) − log µ = log τ = c belongs to V (e) .
These two lemmas state that V (e) is a subalgebra of the
commutative algebra RX with the unit element 1 (: the
constant function x 7→ 1) of RX contained in V (e) . From
a well known result on such subalgebras4 , it is concluded
that there exists a partition {Ai }di=0 of X such that
)
( d
X
d+1
(e)
.
(21)
ci 1Ai (c0 , . . . , cd ) ∈ R
V
=
i=0
Let an element p0 of M (⊂ M̃ ) be arbitrarily fixed. Then we
have
Z (e) = log p0 + V (e) .
(22)
From (19), (21) and (22) and the disjointness of {Ai }, we
have
M̃ = {µ | log µ ∈ Z (e) }
= {µ | log µ − log p0 ∈ V (e) }
n
= µ ∃(c0 , . . . , cd ) ∈ Rd+1 ,
log µ = log p0 +
+ X
is a diffeomorphism from M̃ = Z
∩ (R ) , which is an
open subset of Z (m) , onto Z (e) . The differential map of Φ at
a point µ ∈ M̃ is defined by
dµ(t)
d
(dΦ)µ
= Φ(µ(t))
dt t=0
dt
t=0
for any smooth curve µ(t) in M̃ and is represented as
(dΦ)µ : V (m) ∋ f 7→
This gives a linear isomorphism from V (m) onto V (e) . Therefore, for any two points µ, ν ∈ M̃ , we can define
µa
(e)
∈ V (e) .
(dΦ)ν ◦ (dΦ)−1
∋ a 7→
µ :V
ν
This means that, for any a ∈ V (e) and any µ, ν ∈ M̃ , we have
µa
(e)
. For arbitrary a ∈ V (e) and ν ∈ M̃ , let us define a
ν ∈ V
map Ψa,ν by
µa
Ψa,ν : M̃ ∋ µ 7→
∈ V (e) .
ν
Then its differential at a point µ ∈ M̃ is given by
ga
(dΨa,ν )µ : V (m) ∋ g 7→
∈ V (e) .
ν
Composing this map with the inverse of
g
(dΦ)ν : V (m) ∋ g 7→ ∈ V (e) ,
ν
we have
f
∈ V (e) .
µ
3 Actually, Z (m) is a linear space as mentioned in section IV, and therefore
Z (m) = V (m) .
=
=
(
(
d
X
i=0
p0
d
X
o
c i 1 Ai , ,
ci
e 1Ai (c0 , . . . , cd ) ∈ R
d+1
i=0
d
X
i=0
+ d+1
λi qi (λ0 , . . . , λd ) ∈ (R )
)
)
,
4 Although various mathematical extensions of this result including infinitedimensional and/or noncommutative versions are known, the author of the
present paper could find no appropriate reference describing the result for the
finite-dimensional commutative case with an elementary proof. So, we give a
proof in the appendix for the readers’ sake.
where
1
p 0 1 Ai ,
qi :=
p0 (Ai )
Proof. Let
i ∈ {0, . . . , d}.
Then {qi } are probability distributions on X whose supports
are supp (qi ) = Ai , and
M = M̃ ∩ P(X )
)
( d
X
λi qi (λ0 , . . . , λd ) ∈ Pd .
=
i=0
Since this is the same form as (4), condition (i) has been
derived.
VIII. C ONCLUSION
We have shown Theorem 1 which gives an informationgeometrical characterization of statistical models on finite
sample spaces which are statistically equivalent to open probability simplexes Pd . The statistical equivalence (also called
the Markov equivalence) to probability simplexes played a
crucial role in Čencov’s pioneering work [1] on information
geometry, where the notions of Fisher information metric
and the α-connections were characterized in terms of the
statistical equivalence. The present work shed another light on
the relation between the statistical equivalence and information
geometry.
ACKNOWLEDGMENT
This work was partially supported by JSPS KAKENHI
Grant Number 16K00012.
B := f −1 (λ) | λ ∈ R, f ∈ V
⊂ 2X ,
(24)
which is the totality of the level sets f −1 (λ) = {x | f (x) =
λ} ⊂ X of functions in V . We first show that, for any B ⊂
X,
B ∈ B ⇔ 1B ∈ V.
(25)
Since ⇐ is obvious, it suffices to show ⇒. Assume B ∈ B,
so that B = f −1 (λ) for some f ∈ V and λ ∈ R. When B
is the empty set φ, we have 1B = 0 ∈ V . So we assume
B 6= φ, which means that λ ∈ f (X ). Let the elements of
f (X ) be λ0 , λ1 , . . . λk , where λ0 = λ and λi 6= λj if i 6= j,
Pk
and let Bi := f −1 (λi ). Then we have f = i=0 λj 1Bi with
B0 = B. Let a(t) = a0 tk + a1 tk−1 + · · · + ak be a polynomial
satisfying a(λ0 ) = 1 and a(λi ) = 0 for any i 6= 0. Explicitly,
a(t) is expressed as
a(t) =
It follows that
a(f ) =
k
X
k
Y
t − λi
.
λ − λi
i=1 0
a(λi )1Bi = 1B0 = 1B .
i=0
In addition, a(f ) = a0 f k + a1 f k−1 + · · · + ak belongs to V
since V is a subalgebra of RX with 1 ∈ V . Hence we have
1B ∈ V .
Using (25), we see that
R EFERENCES
X ∈ B,
(26)
B ∈ B ⇒ B c ∈ B,
B1 , B2 ∈ B ⇒ B1 ∩ B2 ∈ B
(27)
(28)
[1] N. N. Čencov (Chentsov), Statistical Decision Rules and Optimal Inference, AMS, 1982 (original Russian edition: Nauka, Moscow, 1972).
[2] S. Amari, “Differential geometry of curved exponential families—
as
curvature and information loss”, The Annals of Statistics, 10, 357–385,
1982.
1X = 1 ∈ V ⇒X ∈ B,
(29)
[3] S. Amari, Differential-Geometrical Methods in Statistics, Springer, LecB ∈ B ⇒ 1B ∈ V ⇒ 1B c =1 − 1B ∈ V
ture Notes in Statistics 28, 1985.
[4] S. Amari and H. Nagaoka, Methods of information geometry, AMS &
⇒ B c ∈ B,
(30)
OUP, 2000.
[5] H. Nagaoka and S. Amari, “Differential geometry of smooth
B1 , B2 ∈ B ⇒ 1B1 , 1B2 ∈ V ⇒ 1B1 ∩B2 = 1B1 1B2 ∈ V
families of probability distributions”, Technical Report METR 82⇒ B1 ∩ B2 ∈ B.
(31)
7, Dept. of Math. Eng. and Instr. Phys, Univ. of Tokyo, 1982.
(http://www.keisu.t.u-tokyo.ac.jp/research/techrep/data/1982/METR82-07.pdf)
Properties (26)-(28) implies that B is an additive class of
[6] A. Ohara, “Information geometric analysis of an interior point method
for semidefinite program- ming”, Geometry in Present Day Science (eds.
sets (σ-algebra) on the finite entire set X . Therefore, B is
O. E. Barndorff-Nielsen and E. B. V. Jensen), pp.49-74, World Scientific,
generated by a partition {A1 , · · · , An } of X in the sense that
1999.
every element of B is the union of some (or no) elements
[7] A. Ohara, “Geodesics for dual connections and means on symmetric
cones”, Integr. equ. oper. theory, 50, 537–548, 2004.
of {A1 , · · · , An }. Recalling the definition (24) of B, we
A PPENDIX
Proposition Let X be a finite set and V be a subalgebra
of RX containing the constant functions. Then there exists a
partition {Ai }ni=1 of X such that
)
( n
X
n
.
(23)
ci 1Ai (c1 , . . . , cn ) ∈ R
V =
i=1
conclude (23).
| 10math.ST |
Scalable photonic reinforcement learning by time-division
multiplexing of laser chaos
Makoto Naruse1*, Takatomo Mihana2, Hirokazu Hori3, Hayato Saigo4,
Kazuya Okamura5, Mikio Hasegawa6 and Atsushi Uchida2
1
Network System Research Institute, National Institute of Information and Communications
Technology, 4-2-1 Nukui-kita, Koganei, Tokyo 184-8795, Japan
2
Department of Information and Computer Sciences, Saitama University, 255 Shimo-Okubo,
Sakura-ku, Saitama, Saitama 338-8570, Japan
3
Interdisciplinary Graduate School, University of Yamanashi, Takeda, Kofu, Yamanashi 400-8510,
Japan
4
Nagahama Institute of Bio-Science and Technology, 1266 Tamura, Nagahama, Shiga 526-0829,
Japan
5
Graduate School of Informatics, Nagoya University, Furo, Chikusa, Nagoya, Aichi 464-8601, Japan
6
Department of Electrical Engineering, Tokyo University of Science, 6-3-1 Niijuku, Katsushika, Tokyo
125-8585, Japan
* Corresponding author. Email: naruse@nict.go.jp
1
Abstract
Reinforcement learning involves decision making in dynamic and uncertain environments and
constitutes a crucial element of artificial intelligence. In our previous work, we experimentally
demonstrated that the ultrafast chaotic oscillatory dynamics of lasers can be used to solve the
two-armed bandit problem efficiently, which requires decision making concerning a class of
difficult trade-offs called the explorationexploitation dilemma. However, only two selections
were employed in that research; thus, the scalability of the laser-chaos-based reinforcement
learning should be clarified. In this study, we demonstrated a scalable, pipelined principle of
resolving the multi-armed bandit problem by introducing time-division multiplexing of
chaotically oscillated ultrafast time-series. The experimental demonstrations in which bandit
problems with up to 64 arms were successfully solved are presented in this report. Detailed
analyses are also provided that include performance comparisons among laser chaos signals
generated in different physical conditions, which coincide with the diffusivity inherent in the
time series. This study paves the way for ultrafast reinforcement learning by taking advantage
of the ultrahigh bandwidths of light wave and practical enabling technologies.
2
Recently, the use of photonics for information processing and artificial intelligence has been
intensively studied by exploiting the unique physical attributes of photons. The latest examples include
a coherent Ising machine for combinatorial optimization1, photonic reservoir computing to perform
complex time-series predictions2,3, and ultrafast random number generation using chaotic dynamics in
lasers4,5 in which the ultrahigh bandwidth attributes of light bring novel advantages. Reinforcement
learning, also called decision making, is another important branch of research, which involves making
decisions promptly and accurately in uncertain, dynamically changing environments6 and constitutes
the foundation of a variety of applications ranging from communication infrastructures7,8 and robotics9
to computer gaming10.
The multi-armed bandit problem (MAB) is known to be a fundamental reinforcement learning
problem where the goal is to maximize the total reward from multiple slot machines whose reward
probabilities are unknown and could dynamically change6. To solve the MAB, it is necessary to
explore higher-reward slot machines. However, too much exploration may result in excessive loss,
whereas too quick of a decision or insufficient exploration may lead to missing the best machine; thus,
there is a trade-off referred to as the explorationexploitation dilemma11.
In our previous study, we experimentally demonstrated that the ultrafast chaotic oscillatory
dynamics of lasers2–5 can be used to solve the MAB efficiently12. With a chaotic time series generated
by a semiconductor laser with a delayed feedback sampled at a maximum rate of 100 GSample/s
followed by a digitization mechanism with a variable threshold, ultrafast, adaptive, and accurate
3
decision making was demonstrated. Such ultrafast decision making is unachievable using conventional
algorithms on digital computers11,13,14, which rely on pseudorandom numbers. It was also demonstrated
that the decision-making performance is maximized by utilizing an optimal sampling interval that
exactly coincides with the negative autocorrelation inherent in the chaotic time series12. Moreover,
even when assuming that pseudorandom numbers and coloured noise were available in such a highspeed domain, the laser chaos method outperformed these alternatives; that is, chaotic dynamics yields
superior decision-making abilities12.
However, only two options, or slot machines, were employed in the MAB investigated therein;
that is, the two-armed bandit problem was studied. A scalable principle and technologies toward an
N-armed bandit with N being a natural number are strongly demanded for practical applications. In
addition, detailed insights into the relations between the resulting decision-making abilities and
properties of chaotic signal trains should be pursued to achieve deeper physical understanding as well
as performance optimization at the physical or photonic device level.
In this study, we experimentally demonstrated a scalable photonic reinforcement learning
principle based on ultrafast chaotic oscillatory dynamics in semiconductor lasers. Taking advantage of
the high-bandwidth attributes of chaotic lasers, we incorporated the concept of time-division
multiplexing into the decision-making strategy; specifically, consecutively sampled chaotic signals are
used in the proposed method to determine the identity of the slot machine in a binary digit form.
4
In the recent literature on photonic decision making, near-field-mediated optical excitation
transfer15,16 and single photon17,18 methods have been discussed; the former technique involves
pursuing the diffraction-limit-free spatial resolution19, whereas the latter reveals the benefits of the
wave–particle duality of single light quanta20. A promising approach for achieving scalability by
means of near-field-coupled excitation transfer or single photons is spatial parallelism; indeed, a
hierarchical principle has been successfully demonstrated experimentally in solving the four-armed
bandit problem using single photons18. In contrast, the high-bandwidth attributes of chaotic lasers
accommodate time-division multiplexing and have been successfully used in optical communications21.
In this study, we transformed the hierarchical decision-making strategy18 into the time domain,
transcending the barrier toward scalability. We also successfully resolved the bandit problem with up
to 64 arms.
Meanwhile, four kinds of chaotic signals experimentally generated in different conditions, as
well as quasiperiodic sequences, were subjected to performance comparisons and characterizations,
including diffusivity analysis. In addition, computer-generated pseudorandom signals and coloured
noise were used to clarify the similarities and differences with respect to chaotically fluctuating
random signals. Detailed dependency analysis with regard to the precision of parameter adjustments,
sampling interval of chaotic time series, and difficulties of given decision-making problems as well as
diffusivity analyses of time series were also performed. The experimental findings will facilitate
5
understanding of the characteristics of laser-chaos-based decision making and the future design of
integrated systems.
Principle
We considered an MAB problem in which a player selects one of N slot machines, where N = 2M with
M being a natural number. The N slot machines are distinguished by the identity given by natural
numbers ranging from 0 to N1, which are also represented in an M-bit binary code given by
S1S2 S M with Si (i = 1, …, M) being 0 or 1. For example, when N = 8 (or M = 3), the slot machines
are numbered by S1S2S3 = {000, 001, 010, …, 111} (Fig. 1a). The reward probability of slot machine
i is represented by Pi (i = 0, …, N1), and the problem addressed herein is the selection of the machine
with the highest reward probability. The reward amount dispensed by each slot machine per play is
assumed to be the same in this study. That is, the probability of winning by playing slot machine i is
Pi, and the probability of losing by playing slot machine i is 1Pi.
The principle consists of the following three steps: [STEP 1] decision making for each bit of
the slot machine in a pipelined manner, [STEP 2] playing the selected slot machine, and [STEP 3]
updating the threshold values. The exact details and general formula are given in the Methods section.
[STEP 1] Decision for each bit of the slot machine
6
The identity of the slot machine to be chosen is determined bit by bit from the most significant bit
(MSB) to the least significant bit in a pipelined manner. For each of the bits, the decision is made based
on a comparison between the measured chaotic signal level and the designated threshold value.
First, the chaotic signal s(t1) measured at t = t1 is compared to a threshold value denoted as TH1
(Fig. 1b). The output of the comparison is immediately the decision of the MSB concerning the slot
machine to choose. If s(t1) is less than or equal to the threshold value TH1, the decision is that the MSB
of the slot machine to be chosen is 0, which we denote as D1 = 0. Otherwise, the MSB is determined
to be 1 (D1 = 1). Here we suppose that s(t1) < TH1; then, the MSB of the slot machine to be selected is
0.
Based upon the determination of the MSB, the chaotic signal s(t2) measured at t = t2 is subjected
to another threshold value denoted by TH2,0. The first number in the suffix, 2, means that this threshold
is related to the second-most significant bit of the slot machine, while the second number of the suffix,
0, indicates that the previous decision, related to the MSB, was 0 (D0 = 0). If s(t2) is less than or equal
to the threshold value TH2,0, the decision is that the second-most significant bit of the select slot
machine to be chosen is 0 (D2 = 0) (Fig. 1b). Otherwise, the second-most significant bit is determined
to be 1 (D2 = 1). Note that the second-most significant bit is determined by the other threshold value
TH2,1 if the MSB is 1 (D0 = 1).
All of the bits are determined in this manner. In general, there are 2k-1 kinds of threshold values
related to the k-th bit; hence, there are 2M 1 = N 1 kinds of threshold values in total. What is
7
important is that the incoming signal sequence is a chaotic time series, which enables efficient
exploration of the searching space, as discussed later.
[STEP 2] Slot machine play
Play the selected slot machine.
[STEP 3] Threshold values adjustment
Suppose that the selected slot machine yields a reward (i.e. the player wins the slot machine play).
Then, the threshold values are adjusted so that that same decision will be highly likely to be selected
in the subsequent play. Therefore, for example, if the MSB of the selected machine is 0, TH1 should
be increased because doing so increases the likelihood of obtaining the same decision regarding MSB
being 0. All of the other threshold values involved in determining the decision are updated in the same
manner.
It should be noted that due to the irregular nature of the incoming chaotic signal, the possibility
of choosing the opposite values of bits is not 0 if the above-described threshold adjustments have been
made. This feature is critical in exploration in reinforcement learning. For example, even when the
value of TH1 is sufficiently small (indicating that slot machines whose MSBs are 1 are highly likely to
be better machines), the probability of the decision to choose machines whose MSBs are 0 is not 0.
This mechanism is of particular importance when the given decision-making problem is difficult (i.e.
the differences among the reward probabilities are minute); this situation will be discussed in detail
later.
8
If the selected slot machine does not yield a reward (i.e. the player loses the slot machine play),
then the threshold values are adjusted so that that same decision will not be highly likely to be selected
in the subsequent play. Therefore, for example, if the MSB of the selected machine is 0, TH1 should
be decreased because doing so decreases the likelihood of obtaining the same decision regarding MSB
being 0. All of the other threshold values involved determining the decision are revised.
As described above, the threshold adjustment involves increasing or decreasing the threshold
values based on the betting results, which seems to be symmetric between the cases of winning and
losing. However, the adjustment must be made asymmetrically except in special cases for the following
reason.
Suppose that the reward probabilities of Machines 0 and 1 are given by 0.9 and 0.7, respectively,
where the probability of receiving a reward is rather high. Indeed, the probability of receiving a reward
regardless of the decision is (0.9 + 0.7)/2 = 0.8 while that of no reward is (0.1 + 0.3)/2 = 0.2. Thus,
the event of losing is rare and should occur four times (0.8/0.2 = 4) less than the event of winning.
Hence, if the amount of threshold adjustment in the case of winning is set to 1, that in the case of losing
should be 4. On the contrary, if the reward probabilities of Machines 0 and 1 are given by 0.1 and 0.3,
respectively, the tendency becomes the opposite since most of the betting results in losing; hence, the
amount of threshold adjustment in the case of losing must be attenuated by four times compared to
that in the case of winning.
9
In the present study, the amount of threshold adjustment in the case of wining is given by 1
while that of losing is given by the parameter Ω. The detailed definition is provided in the Methods
section. Ω is also updated during the course of play based on the betting history concerning the
numbers of wins and selections. Notably, Ω must be configured differently based on the designated bit.
For the MSB, for example, the win/lose events should be related to the two groups of slot machines
whose MSBs are 0 and 1, while for the second-most significant bit when the MSB is 0, the win/lose
events are related to the two groups of slot machines whose second-most significant bits are 0 and 1
and have MSBs of 0.
Results
A schematic diagram of the laser-chaos-based scalable decision-making system is shown in Fig. 1c. A
semiconductor laser operated at a centre wavelength of 1547.785 nm is coupled with a polarizationmaintaining (PM) coupler. The light is connected to a variable fibre reflector, which provides delayed
optical feedback to the laser, generating laser chaos22–24. The output light at the other end of the PM
coupler is detected by a high-speed, AC-coupled photodetector through an optical isolator (ISO) and
optical attenuator. The signal is sampled by a high-speed digital oscilloscope at a rate of 100
GSample/s (a 10 ps sampling interval) with an eight-bit resolution; the signal level takes integer values
ranging from 127 to 128. The details of the experimental setup are described in the Methods section.
10
Figure 2a shows examples of the chaotic signal trains. Four kinds of chaotic signal trains were
generated, which are referred to as (i) Chaos 1, (ii) Chaos 2, (iii) Chaos 3, and (iv) Chaos 4 in Fig. 2a,
by varying the reflection by the variable reflector by letting 210, 120, 80, and 45 μW of optical power
be fed back to the laser, respectively. A quasiperiodic signal train was also generated, as depicted in
Fig. 2a(v), by the variable reflector by providing a feedback optical power of 15 μW. Figure 2b
summarizes the experimentally observed radio-frequency (RF) power spectra obtained using Chaos 1,
2, 3, and 4 and quasiperiodic signals. It can be seen that the chaotic time series contain wide bands of
signals24 and that there are clear differences among the shapes of the RF spectra corresponding to
Chaos 1–4, even though the time-domain waveforms shown in Fig. 2a(i–iv) look similar. The
experimental details of the RF spectrum evaluation are provided in the Methods section.
In addition, Fig. 2a(vii) shows an example of a coloured noise signal train containing negative
autocorrelation calculated using a computer based on the Ornstein–Uhlenbeck process using white
Gaussian noise and a low-pass filter25 with a cut-off frequency of 10 GHz12. Also, the black curve in
Fig. 2a(vi) marked with RAND depicts a sequence generated by a pseudorandom generator based on
the Mersenne Twister. For RAND, the horizontal axis of Fig. 2a should be read as ‘cycles’ instead of
physical time, but we dealt with RAND as if it were available at the same sampling rate as the laser
chaos signals to investigate the performance differences between the laser chaos sequences and
pseudorandom numbers both in qualitative and quantitative ways.
Two-armed bandit
11
We began with the two-armed bandit problem, which is the simplest case12. The slot machine was
played 250 times consecutively, and such play was repeated 10,000 times. The reward probabilities of
the two slot machines, referred to as Machines 0 and 1, were 0.9 and 0.7, respectively; hence the
correct decision was to choose Machine 0 because it was the machine with the higher reward
probability. The red, green, blue, and cyan curves in Fig. 2c show the evolution of the correct decision
ratio (CDR), defined as the ratio of the number of times when the selected machine has the highest
reward probability at cycle t based on the time series of Chaos 1, 2, 3, and 4, respectively. The chaotic
signal was sampled every 50 ps; that is, a single cycle corresponds to 50 ps in physical time. The
magenta, yellow, and black curves in Fig. 2c represent the CDRs obtained based on quasiperiodic,
coloured noise, and RAND sequences. Clearly, the chaotic sequences approach a CDR of unity more
quickly than the other signals. Although the difference is subtle, Chaos 3 exhibits the best adaptation
among the four chaotic time series; a CDR of 0.95 was achieved at cycle 122, corresponding to 6.1 ns.
In the previous study, an exact coincidence between the autocorrelation of the laser chaos signal
trains and the resulting decision-making performance was obtained12; specifically, it was found that
the sampling interval yielding the negative maximum of the autocorrelation provided the fastest
decision-making abilities. To solve a two-armed bandit problem, a single threshold (TH1) and single
chaotic signal sample are needed to derive a decision (D1 = 0 or 1). The sampling interval, or more
precisely the inter-decision sampling interval, of chaotic signals to configure the threshold (TH1) is
defined by ΔS, which is shown in Fig. 1b. Figure 2d compares the autocorrelations of Chaos 1–4 as
12
well as the quasiperiodic and coloured noise. Chaos 1–4 exhibit negative maxima at time lags of around
5 and 6 (and 5 and 6), whereas the quasiperiodic and coloured noise yield negative maxima at time
lags of around 7 (and 7). The amount of time lag corresponds to the physical time difference
multiplied times 10 ps, which is the sampling interval; hence, for example, a time lag of 5 means that
the time difference is 50 ps.
Correspondingly, Fig. 3 characterizes the CDRs as a function of the inter-decision sampling
interval ΔS by setting the reward probabilities of the two slot machines to 0.1 and 0.5. In Fig. 3a, the
CDRs at cycle 10 are compared among Chaos 1–4, while Fig. 3b shows the CDRs at cycle 100 for the
quasiperiodic, coloured noise, RAND, and Chaos 3 series. In Fig. 3a, the CDRs obtained using the
chaotic time series show maxima around the sampling intervals of 50 ps and 60 ps, which nicely
coincide with the negative maxima of the autocorrelations, as we reported previously12. At the same
time, the negative maxima of the chaotic time series follow the order Chaos 4, 3, 2, and 1 in Fig. 2d,
whereas the greater decision-making performances follow the order Chaos 3, 2, 4, and 1 in Fig. 3a
with a sampling interval of 50 ps. That is, the order of the absolute values of the autocorrelation does
not explain the resulting decision-making performances. We will discuss the relation between the
decision-making performance and the characteristics of chaotic time series via other metrics at the end
of the paper. Meanwhile, the black curve in Fig. 3b, which corresponds to RAND, does not show
dependency on the inter-decision sampling interval, whereas the magenta and yellow curves
corresponding to quasiperiodic and coloured noise exhibit peaky characteristics with respect to the
13
sampling interval, clearly indicating the qualitative differences between correlated times series and
conventional pseudorandom signals.
Multi-armed bandit
We applied the proposed time-division multiplexing decision-making strategy to bandit problems with
more than four arms. Here, we first describe the problem to be solved and the assignment of reward
probabilities (Fig. 4a).
(1) Two-armed: The reward probabilities of Machines 0 and 1 are given by 0.9 and 0.7, respectively
(Fig. 4a(i)). Note that the difference is 0.2, which is retained in the subsequent settings.
(2) Four-armed: In addition to the threshold used to determine the MSB (TH1), two more thresholds
are necessary to determine the second bit ( TH 2,D1 (D1 = {0, 1})). The reward probabilities of
Machines 0, 1, 2, and 3 are defined as 0.7, 0.5, 0.9, and 0.1, respectively, where the correct decision
is to select Machine 3 (Fig. 4a(ii)). Note that the difference between the highest and second-highest
reward probabilities is 0.2, as in the two-armed bandit problem. In addition, the sum of the reward
probabilities of the first two machines (Machines 0 and 1: 0.7 + 0.5 = 1.2) is larger than that of the
second two machines (Machines 2 and 3: 0.9 + 0.1 = 1.0). This situation is called contradictory18
since the maximum-reward-probability machine (Machine 3) belongs to the latter group whose
reward-probability sum is greater than that of the former group.
(3) Eight-armed: In addition to the thresholds used to determine the MSB (TH1) and the second bit
TH 2,D1 (D1 = {0, 1}), four more thresholds are needed to decide the third bit ( TH 3, D1 , D2 (D1 = {0,
14
1}, D2 = {0, 1})). The reward probabilities of Machines 0, 1, 2, 3, 4, 5, 6, and 7 are given by 0.7,
0.5, 0.9, 0.1, 0.7, 0.5, 0.7, and 0.5, respectively. First, the difference between the highest and
second-highest reward probabilities is 0.2, as in the two- and four-armed bandit problems described
above. Second, the sum of the reward probabilities of the slot machines whose MSBs are 0 and 1
are 2.2 and 2.4, respectively, whereas the maximum-reward-probability machine (Machine 2) has
an MSB of 0, which is a contradictory situation. Similarly, the sums of the reward probabilities of
the slot machines whose second MSBs are 0 and 1 (as well as whose MSBs are 1) are 1.2 and 1,
respectively, while the best machine belongs to the latter group, which is also a contradiction (Fig.
4a(iii)). In the following bandit problem definitions, all of these contradictory conditions are
satisfied for the sake of coherent comparison with the increased arm numbers.
(4) 16-armed: In addition to the thresholds used to determine the MSB (TH1), the second bit TH 2,D1
(D1 = {0, 1}), and the third bit ( TH 3, D1 , D2 ( TH 3, D1 , D2 (D1 = {0, 1}, D2 = {0, 1})), eight more
thresholds are required for the fourth bit ( TH 4, D1 , D2 , D3 ( TH 3, D1 , D2 (D1 = {0, 1}, D2 = {0, 1}, D3 = {0,
1})). The reward probabilities of Machines 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, and 15 are
given by 0.7, 0.5, 0.9, 0.1, 0.7, 0.5, 0.7, 0.5, 0.7, 0.5, 0.7, 0.5, 0.7, 0.5, 0.7, and 0.5, respectively.
The best machine is Machine 2. The contradiction conditions are satisfied, as in the four- and eightarmed problems (Fig. 4a(iv)).
15
(5) 32-armed: A 32-armed bandit requires thresholds to determine five bits. The best machine is
Machine 2. The contradiction rules apply, as in the previous cases. The details are described in the
Methods section (Fig. 4a(v)).
(6) 64-armed: A 64-armed bandit requires thresholds to determine six bits. The best machine is
Machine 2. The contradiction rules apply, as in the previous cases. The details are described in the
Methods section (Fig. aA(vi)).
Figures 4c(i), (ii), (iii), (iv), (v), and (vi) summarize the results of the two-, four-, eight-, 16-,
32-, and 64-armed bandit problems, respectively. The red, green, blue, and cyan curves show the CDR
evolution obtained using Chaos 1, 2, 3, and 4, respectively, while the magenta, black, and yellow
curves depict the evolution obtained using quasiperiodic, RAND, and coloured noise, respectively.
The threshold values take integer values ranging from 128 to 128. The sampling interval of the
chaotic signal trains for the MSB (ΔS) is 50 ps, whereas that of the subsequent bits, called the inter-bit
sampling interval (ΔL), is 100 ps. (The impacts due to the choice of ΔL are discussed later.) From Fig.
4c, it can be seen that Chaos 3 provides the promptest adaptation to the unity value of the CDR, whereas
the nonchaotic signals (quasiperiodic, RAND, and coloured noise) yield substantially deteriorated
performances, especially in bandit problems with more than 16 arms. The number of cycles necessary
to obtain the correct decision increases as the number of bandits increases. The square marks in Fig.
4d indicate the numbers of cycles required to reach a CDR of 0.95 as a function of the number of slot
machines, where the required number of cycles grows in the form of the power-law relation aNb, where
16
a and b are approximately 55 and 1.16, respectively. These results support the successful operation of
the proposed scalable decision-making principle using laser-generated chaotic time series.
Discussion
Inter-bit sampling interval dependencies
In resolving MAB problems in which the number of bandits is greater than four and is given by 2M, M
samples are needed with the interval being specified by ΔL, as schematically shown in Fig. 1b. In this
study, we investigated the ΔL dependency by analysing the four kinds of four-armed bandit problems
shown in Fig. 4b and labelled as Types 1–4. The reward probabilities of Type 1 are equal to those in
the case shown in Fig. 4a(ii); P0, P1, P2, and P3 are given by 0.7, 0.5, 0.9, and 0.1, respectively. The
correct decision is to select Machine 2; that is, the machine identity is given by S1S2 = 10. In deriving
the correct decision, the first sample should be greater than the threshold (TH0) to obtain the decision
S1 = 1, whereas the second sample should be smaller than the threshold (TH2,1) to obtain the decision
S2 = 0. Consequently, if ΔL is 0, the search for the best selections does not work well since the same
sampling provides the same searching traces that do not satisfy the conditions for both bits. Indeed,
the cyan circular marks in Fig. 5a characterize the CDR at cycle 100 as a function of ΔL, where ΔL = 0
ps yields a CDR of 0. Chaos 3 was used for the evaluation. The CDR exhibits the maximum value
when ΔL = 50 ps, which is reasonable because 50 ps is the interval that provides the negative
autocorrelation that easily allows oppositely arranged bits to be found (S1S2 = 10).
17
Types 2, 3, and 4 contain the same reward probability values but in different arrangements. In
Type 3, the correct decision is to select Machine 1, or S1S2 = 01, which is similar to the correct decision
in Type 1 in the sense that the two bits have opposite values. Consequently, the inter-bit-interval
dependence, shown by the yellow circular marks in Fig. 5a, exhibits traces similar to those of Type 1,
where ΔL = 0 ps gives a CDR of 0, whereas ΔL values that yield negative autocorrelations provide
greater CDRs. In Types 2 and 4, on the other hand, the correct decisions are given by Machines 0 and
3, or S1S2 = 00 and S1S2 = 11. For such problems, ΔL = 0 ps gives a greater CDR due to the eventually
identical values of the first and second bits, whereas ΔL values corresponding to negative
autocorrelations yield poorer performance, unlike for Types 1 and 3, as clearly represented by the
magenta and black circular marks in Fig. 5a.
It is also noteworthy that pseudorandom numbers provide no characteristic responses with
respect to the inter-bit intervals, as shown by the square marks in Fig. 5a, which is another clear
indication that the temporal structure inherent in chaotic signal trains affects the decision-making
performance.
The decision-making system must deal with all of these types of problems, namely, all kinds
of bit combinations; thus, temporal structures, such as positive and negative autocorrelations, may lead
to inappropriate consequences. To derive a moderate setting, the circular marks in Fig. 5b show the
coefficient of variation (CV), which is defined as the ratio of the standard deviation to the mean value,
for the four types of problems shown in Fig. 5a. A smaller CV is preferred. The inter-bit sampling
18
interval of 30 ps eventually provides the minimum CV, although slight changes could lead to larger
CVs. Indeed, the autocorrelation is about 0 with this inter-bit sampling interval (Fig. 2d). An inter-bit
sampling interval of approximately 100 ps constantly offers smaller CV values. The square marks in
Fig. 5b correspond to RAND, and no evident inter-bit interval dependency related to the CV is
observable, in clear contrast to the chaotic time series cases.
Decision-making difficulties
The adjustment precision of the thresholds is important when searching for the maximum-rewardprobability machine, especially in many-armed bandit problems that include the contradictory
arrangement discussed earlier18. Here, we discuss the dependency of the decision-making difficulty by
focussing on the two-armed bandit problem; even in simple two-armed cases, the threshold precision
clearly affects the resulting decision-making performance.
Figures 6a and 6b present two-armed bandit problems in which the reward probability of
Machine 0 is 0.9 whereas that of Machine 1 is 0.5 and 0.7, respectively. Since the probability difference
is larger in the former case, it is easier to derive the maximum-reward-probability machine in that case.
Indeed, the curves in the former case shown in Fig. 6a provide steeper adaptation than in the latter case
depicted in Fig. 6b. The eight curves shown therein depict the CDRs corresponding to numbers of
threshold levels given by 2K+1, where K takes integer values from 1 to 8. From Figs. 6a and 6b, it can
be seen that the CDR is saturated before approaching unity when the number of threshold levels is
19
smaller than it is in the K = 5 case; in particular, the CDR is limited to significantly lower values in
difficult problems (Fig. 6a).
One of the reasons behind such phenomena is insufficient exploration due to the smaller
threshold levels. In the case of nine-level thresholding (K = 3), the estimated reward probabilities of
Machines 0 and 1 based on Eq. (6) at cycle 200 are 0.742 and 0.307, respectively, which significantly
differ from the true values (0.9 and 0.7, respectively). Eventually, Ω1 became approximately 1.1, which
is far from the value of Ω1 = 4 based on Eq. (7) assuming true reward probabilities, indicating that the
threshold could not be biased toward the positive or negative maximum value (here it should be a
positive maximum because the correct decision is to choose Machine 0); indeed, the threshold TH1 is
limited to around 45 at cycle 200, which is far from the negative maximum of 128. Based on Eqs.
(1) and (5), the absolute value of the threshold is decreased due to the forgetting parameter, which is
0.45; this value is larger than the average of the decrement and increment caused by the first terms of
Eqs. (1) and (5), leading to the saturation of TH1 and resulting in a limited CSR. Figure 6c summarizes
the CDR at cycle 200 as a function of decision difficulty, where precise threshold control is necessary
to obtain a higher CDR, especially in difficult problems.
Diffusivity and decision-making performance
In the results shown for bandit problems with up to 64 arms in Fig. 4, Chaos 3 provides the best
performance among the four kinds of chaotic time series. The negative autocorrelation indeed affects
the decision-making ability, as discussed in Fig. 5; however, the value of the negative maximum of
20
the autocorrelation shown in Fig. 2b does not coincide with the order of performance superiority,
indicating the necessity of further insights into the underlying mechanisms.
In this respect, we analysed the diffusivity of the temporal sequences based on the ensemble
averages of the time-averaged mean square displacements (ETMSDs)26,27 in the following manner. We
first generated a random walker via comparison between the chaotic time series. If the value of the
random number, which was generated based on the Mersenne Twister, is smaller than s(t), the walker
moves to the right. X(t) = +1; otherwise, X(t) = 1. Hence, the position of the walker at time t is given
by x(t) = X(1) + X(2) + … + X(t). We then calculate the ETMSD using
ETMSD( )
1
T
T
x (t ) x ( t )
2
,
(1)
t 1
where x(t) is the time series, T is the last sample to be evaluated, and denotes the ensemble
average over different sequences. The ETMSDs corresponding to Chaos 1, 2, 3, and 4 and
quasiperiodic, RAND, and coloured noise are shown in the inset of Fig. 7, all of which monotonically
increase as a function of the time difference τ. It should be noted that at τ = 1000, Chaos 3 exhibits the
maximum ETMSD, followed by Chaos 2, 1, and 4, as shown by the circular marks in Fig. 7a. This
order agrees with the superiority order of the decision-making performance in the 64-armed bandit
problem shown in Fig. 4c. At the same time, RAND derives an ETMSD of 1000 at τ = 1000, which is
a natural consequence considering the fact that the mean square displacement of a random walk is
given by
x (t )
x (t )
2
4 pqt , where p and q are the probabilities of flight to the right and left,
21
respectively. If p = q = 1/2 (via RAND), then the mean square displacement is t. From Fig. 7a, RAND
and coloured noise actually exhibit larger ETMSD values than Chaos 1–4, although the decisionmaking abilities are considerably poorer for RAND and coloured noise, implying that the ETMSD
alone cannot perfectly explain the performances.
Figure 7b explains diffusivity in another way, where the average displacement x(t ) and
x ( t D ) are plotted for each time series superimposed in the XY plane with D = 10,000. Although
the quasiperiodic and coloured noise, shown by the magenta and yellow curves, respectively, move
toward positions far from the Cartesian origin, their trajectories are biased toward limited coverage in
the plane. Meanwhile, the trajectories of the chaotic time series cover wider areas, as shown by the red,
green, blue, and cyan curves. The trajectories generated via RAND, shown by the black curve, remain
near the origin.
To quantify such differences, we evaluated the covariance matrix Θ = cov(X1, X2) by
substituting x(t) and x(t+D) for X1 and X2, where the ij-element of Θ is defined by
1 N
X1 X1
N 1 i 1
X
2
X 2 , with N denoting the number of samples and X i denoting the average
of Xi. The condition number of Θ, which is the ratio of the maximum singular value to the minimum
singular value28,29, indicates the uniformity of the sample distribution. A larger condition number
means that the trajectories are skewed toward a particular orientation, whereas a condition number
closer to unity indicates uniformly distributed data. The square marks in Fig. 7a show the calculated
22
condition numbers, where Chaos 1–4 achieve smaller values, which are even smaller than that
achieved by RAND, and the quasiperiodic and coloured noise yield larger scores.
Through these analyses using the ETMDSs and condition numbers related to the diffusivity of
the time series, a clear correlation between the greater diffusion properties inherent in laser-generated
chaotic time series and the superiority in the resulting decision-making ability is observable.
Simultaneously, however, we consider further insight to be necessary to draw general
conclusions regarding the origin of the superiority of chaotic time series in the proposed decisionmaking principle. For example, as can be seen in Fig. 2b, the RF spectrum differs among the chaotic
time series, which suggests a potential relation to the resulting decision-making performances.
Ultimately, an artificially constructed optimal chaotic time series that provides the best decisionmaking ability could be derived, which is an important and interesting topic requiring future study.
Pipelined processing
The design, implementation, and performance analysis were performed offline in this study since the
primary objectives were to verify the scalability of the principle and seek better chaotic sequences by
physically tuning the original laser chaos. Although we demonstrated scalability up to 64 arms, no
greater numbers of arms were employed due to the technological limitations regarding the computing
power required to emulate all of the slot machines and the external environment. Simultaneously,
however, it should be noted that the proposed decision-making principle and architecture have simple
structures with a particular emphasis on pipelined processing30. (1) The decision of the first bit (S1) of
23
the slot machine depends only on the first threshold (TH1) and first sampled data. No other information
is required. (2) The decision of the second bit (S2) depends on the decision of the first bit (D1) obtained
in the previous step, the second threshold ( TH 2, D1 ), and the second sampled data. Simultaneously, the
decision of the first bit can proceed to the next decision by sampling the next signal. (3) The same
architecture continues until the M-th bit. The earlier stages do not depend on the results obtained in the
latter stages. Such a structure is particularly preferred due to the benefits of the ultrahigh-speed chaotic
time series signals and greater throughput of the total system.
Conclusion
We proposed a scalable principle of ultrafast reinforcement learning or decision making using chaotic
time series generated by lasers. We experimentally demonstrated that multi-armed bandit problems
with N = 2M arms can be successfully solved using M points of signal sampling from the laser chaos
and comparison to multiple thresholds. Bandit problems with up to 64 arms were successfully solved
even though the presence of difficulties that we call contradictions can potentially lead to trapping in
local minima. Based on the experimental results, the required latency scales as N1.16, with N being the
number of slot machines or bandits. Furthermore, by physically changing the laser chaos operation
conditions, four kinds of chaotic time series were subjected to the decision-making analysis; a
particular chaos sequence provided superiority over the other chaotic time series. Diffusivity analyses
through the ETMSDs and covariance matrix condition numbers related to the time sequences well
24
accounted for the underlying mechanisms for quasiperiodic sequences and computer-generated
pseudorandom numbers and coloured noise. This study is the first demonstration of photonic
reinforcement learning with scalability to larger decision problems and paves the way for new
applications of chaotic lasers in the realm of artificial intelligence.
Methods
Optical system
The laser was a distributed feedback semiconductor laser mounted on a butterfly package with optical
fibre pigtails (NTT Electronics, KELD1C5GAAA). The injection current of the semiconductor laser
was set to 58.5 mA (5.37Ith), where the lasing threshold Ith was 10.9 mA. The relaxation oscillation
frequency of the laser was 6.5 GHz, and its temperature was maintained at 294.83 K. The optical output
power was 13.2 mW. The laser was connected to a variable fibre reflector through a fibre coupler,
where a fraction of light was reflected back to the laser, generating high-frequency chaotic oscillations
of optical intensity22–24. The length of the fibre between the laser and reflector was 4.55 m,
corresponding to a feedback delay time (round trip) of 43.8 ns. PM fibres were used for all of the
optical fibre components. The optical signal was detected by a photodetector (New Focus, 1474-A, 38
GHz bandwidth) and sampled using a digital oscilloscope (Tektronics, DPO73304D, 33 GHz
bandwidth, 100 GSample/s, eight-bit vertical resolution). The RF spectrum of the laser was measured
by an RF spectrum analyser (Agilent, N9010A-544) up to 44 GHz every 44 MHz. The observed raw
25
data were subjected to moving averaging over a 20-point window, yielding the RF spectrum curves
shown in Fig. 2b.
Details of the principle
1. Decision of the most significant bit
The chaotic signal s(t1) measured at t = t1 is compared to a threshold value denoted as TH1 (Fig. 1b).
The output of the comparison is immediately the decision of the most significant bit (MSB) concerning
the slot machine to choose. If s(t1) is less than or equal to the threshold value TH1, the decision is that
the MSB of the select slot machine to be chosen is 0, which we denote as D1 = 0. Otherwise, the MSB
is determined to be 1 (D1 = 1).
2. Decision of the second-most significant bit
Suppose that s(t1) < TH1; then, the MSB of the slot machine to be selected is 0. The chaotic signal s(t2)
measured at t = t2 is subjected to another threshold value denoted by TH2,0. The first number in the
suffix, 2, means that this threshold is related to the second-most significant bit of the slot machine,
while the second number of the suffix, 0, indicates that the previous decision, related to the MSB, was
0 (D0 = 0). If s(t2) is less than or equal to the threshold value TH2,0, the decision is that the second-most
significant bit of the select slot machine to be chosen is 0 (D2 = 0). Otherwise, the second-most
significant bit is determined to be 1 (D2 = 1).
3. Decision of the least significant bit
26
Suppose that s(t2) > TH2,0; then, the second-most significant bit of the slot machine to be selected is 1.
In such a case, the third comparison with regard to the chaotic signal s(t3) measured at t = t3 is
performed using another threshold adjuster value denoted by TH3,0,1. The 3 in the subscript 3,0,1
indicates that the threshold is related to the third-most significant bit, and the second and third numbers,
0 and 1, indicate that the most and second-most significant bits were determined to be 0 and 1,
respectively. Such threshold comparisons continue until all M bits of information ( D1 , D2 ,, DM ) that
specify the slot machine have been determined. If M = 3, the result of the third comparison corresponds
to the least significant bit of the slot machine to be chosen. Suppose that the result of the comparison
is s(t3) < TH3,0,1; then, the third bit is 0 (D3 = 0). Finally, the decision is to select the slot machine with
D1D2D3 = 010; that is, the slot machine to be chosen is 2.
In general, there are 2k-1 kinds of threshold values related to the k-th bit; hence, there are 2M
1 = N 1 kinds of threshold values in total.
4. Threshold values adjustment
I. When winning the slot machine play
If the identity of the selected slot machine is D1D2 DM and it yields a reward (i.e. the player wins
the slot machine play), then the threshold values are updated in the following manner.
(1) MSB
The threshold value related to the MSB of the slot machine is updated according to
27
TH1(t 1) Δ TH1(t ) if D1 0
,
TH1(t 1) Δ TH1(t ) if D1 1
(2)
where α is referred to as the forgetting (memory) parameter12 and Δ is the constant increment (in this
experiment, Δ = 1 and α = 0.99). The intuitive meaning of the update given by Eq. (2) is that the
threshold value is revised so that the likelihood of choosing the same machine in the next cycle
increases.
(2) Second-most significant bit
The threshold adjuster values related to the second-most significant bit are revised based on the
following rules:
TH2,0 (t 1) Δ TH2,0 (t ) if D1 0, D2 0
TH2,0 (t 1) Δ TH2,0 (t ) if D1 0, D2 1
(3)
when the MSB has been determined to be 0 (D1 = 0) and
TH2,1 (t 1) Δ TH2,1 (t ) if D1 1, D2 0
TH 2,1 (t 1) Δ TH2,1 (t ) if D1 1, D2 1
(4)
when the MSB has been determined to be 1 (D1 = 1).
(3) General form
As was done with the most and second-most significant bits in Eqs. (2)–(4), all of the threshold values
are updated. In a general form, the threshold value for the K-th bit is given by
28
TH K ,S1 ,S2 ,,SK 1 (t 1) Δ TH K ,S1 ,S2 ,,SK 1 (t ) if SK 0, D1 S1, D2 S2 ,, DK 1 SK 1
TH K ,S1 ,S2 ,,SK 1 (t 1) Δ TH K ,S1 ,S2 ,,SK 1 (t ) if SK 1, D1 S1 , D2 S2 ,, DK -1 SK -1
(5)
when the decisions from the MSB to the (K-1)-th bit have been determined by D1 = S1, D2 = S2, …, and
DK1 = SK1.
II. When losing the slot machine play
If the selected machine does not yield a reward (i.e. the player loses in the slot machine play), the
threshold values are updated as follows.
(1) MSB
The threshold value of the MSB is updated according to
TH1 (t 1) Ω1 TH1 (t ) if D1 0
,
TH1 (t 1) Ω1 TH1 (t ) if D1 1
(6)
where the parameter Ω1 is determined based on the history of betting results.
Let the number of times that slot machines for which the MSB is 0 (S1 = 0) and 1 (S1 = 1) are
selected be given by CS1 0 and CS1 1 , respectively. Let the number of wins by selecting slot machines
for which the MSB is 0 and 1 be given by LS1 0 and LS1 1 , respectively. Then, the estimated reward
probability, or winning probability, by choosing slot machines for which the MSB is k is given by
PˆS1 k
LS1 k
CS1 k
,
(7)
where k is 0 or 1. Ω1 is then given by
29
Ω1
PˆS1 0 PˆS1 1
.
2 ( PˆS1 0 PˆS1 1 )
(8)
The initial Ω1 is assumed to be unity, and a constant value is assumed when the denominator
of Eq. (8) is 0. Ω1 is the figure that designates the degrees of winning and losing. Indeed, the numerator
of Eq. (8) indicates the degree of winning, whereas the denominator shows that of losing.
(2) Second-most significant bit
The threshold adjuster values related to the second-most significant bit are updated when the MSB of
the decision is 0 (D1 = 0) by using the following formula:
TH 2,0 (t 1) Ω2,0 TH 2,0 (t ) if D1 0, D2 0
.
TH 2,0 (t 1) Ω2,0 TH 2,0 (t ) if D1 0, D2 1
(9)
Let the number of times that slot machines for which the MSB is 0 (S1 = 0) and the secondmost significant bit is 0 (S2 = 0) are selected be given by CS1 0,S2 0 . Let the number of times that slot
machines for which the MSB is 0 (S1 = 0) and the second-most significant bit is 1 (S2 = 1) are selected
be given by CS1 0,S2 1 . Let the numbers of wins by selecting slot machines for which the MSB is 0 (S1
= 0) and the second-most significant bit is 0 (S2 = 0) or 1 (S2 = 1) be given by LS1 0,S2 0 and LS1 0,S2 1 ,
respectively. Then, the estimated reward probability, or winning probability, by choosing slot
machines for which the MSB is 0 and the second-most significant bit is k is given by
PˆS1 0,S2 k
LS1 0,S2 k
CS1 0,S2 k
.
(10)
Ω2,0 is then given by
30
Ω 2,0
PˆS1 0,S2 0 PˆS1 0,S2 1
.
2 ( PˆS1 0, S2 0 PˆS1 0, S2 1 )
(11)
Ω2,0 concerns the ratios of winning and losing within the slot machine groups whose MSBs are given
by 0.
(3) General form
All of the threshold values are updated. In a general form,
THK ,S1 ,S2 ,,SK 1 (t 1) ΩK ,S1 ,S2 ,,SK 1 THK ,S1 ,S2 ,,SK 1 (t ) if SK 0, D1 S1, D2 S2 ,, DK 1 SK 1
THK ,S1 ,S2 ,,SK 1 (t 1) ΩK ,S1 ,S2 ,,SK 1 THK ,S1 ,S2 ,,SK 1 (t ) if SK 0, D1 S1, D2 S2 , L , DK -1 SK-1.
(12)
ΩK ,S1 ,S2 ,,SK 1 is the increment defined as follows.
Let the number of times that slot machines whose upper K-1 bits are specified by S1, S2 ,, SK 1
and whose K-th bits are given by k (Sk = k) are selected be denoted by CS1S2SK 1 ,SK k . Let the number of
wins by selecting such machines be given by LS1S2SK 1 ,SK k . Then, the estimated reward probability, or
winning probability, by choosing such machines is given by
PˆS1S2SK 1 ,SK k
LS1S2SK 1 ,SK k
LS1S2SK 1 ,SK k
.
(13)
ΩK ,S1 ,S2 ,,SK 1 is then given by
Ω K , S1 , S2 ,,S K 1
PˆS1S2 S K 1 , S K 0 PˆS1S2 S K 1 ,S K 1
.
2 ( PˆS1S2 S K 1 , S K 0 PˆS1S2 S K 1 , S K 1 )
31
(14)
Initially, all of the threshold values are 0; hence, for example, the probability of determining
the MSB of the slot machine to be 1 or 0 is 0.5 since TH1 = 0. As time elapses, the threshold values
are biased towards the slot machine with the higher reward probability based on the updates described
by Eqs. (2)–(10). It should be noted that due to the irregular nature of the incoming chaotic signal, the
possibility of choosing the opposite values of bits is not 0, and this feature is critical in exploration in
reinforcement learning. For example, even when the value of TH1 is sufficiently small (indicating that
slot machines whose MSBs are 1 are highly likely to be better machines), the probability of the decision
to choose machines whose MSBs are 0 is not 0.
The number of threshold levels was limited to a finite value in the experimental implementation.
Furthermore, the threshold resolution affects the decision-making performance, as discussed below. In
this study, we assumed that the actual threshold level takes the values Z, …, 1, 0, 1, …, Z, where Z
is a natural number; thus, the number of the threshold levels is 2Z+1 (Fig. 1c). More precisely, the
actual threshold value is defined by
T (t ) a TH (t ) ,
(15)
where TH ( t ) is the nearest integer to TH(t) rounded to 0, and a is a constant for scaling to limit the
range of the resulting T(t). The value of T(t) ranges from –aZ to aZ by assigning the limits T(t) = aZ
when TH ( t ) is greater than Z and T(t) = aZ when TH ( t ) is smaller than Z. In the experimental
results shown below, the chaotic signals s(t) take integer values from 127 to 128; hence, a was given
by a = 128/Z in the present study.
32
Data analysis
(1) Chaotic and quasiperiodic time series: Four kinds of chaotically oscillating signal trains
(Chaos 1, 2, 3, and 4) and a quasiperiodically oscillated sequence were sampled at a rate of 100
GSample/s using 10,000,000 points, which took approximately 10 s. Such 10 M points were
stored 120 times for each signal train; hence, there were 120 kinds of 10-M-long sequences. The
processing for decision making was performed offline using a personal computer (Hewlett-Packard,
Z-800, Intel Xeon CPU, 3.33 GHZ, 48 GB RAM, Windows 7, MATLAB R2016b). Pseudorandom
sequences were generated using the random number generator by Mersenne Twister implemented
in MATLAB.
(2) Coloured noise: Coloured noise was calculated based on the Ornstein–Uhlenbeck process using
white Gaussian noise and a low-pass filter in numerical simulations25. We assumed that the
coloured noise was generated at a sampling rate of 100 GHz, and the cut-off frequency of the lowpass filter was set to 10 GHz. Forty sequences of 10,000,000 points were generated. The reduction
in the number of sequences was due to the excessive computational cost.
(3) Two-armed bandit problem shown in Fig. 2c: Five hundred consecutive plays were repeated
10,000 times; hence, the total number of slot machine plays for a single chaotic sequence was
5,000,000. Such plays were performed for all 120 sets of sequences. (In the case of coloured noise,
40 sets of sequences were used.) The CDR described in the main text was derived as the average
of these sets. In Fig. 2c, the first 250 plays are shown to highlight the initial adaptation.
33
(4) Two-armed bandit problem shown in Fig. 3: The number of consecutive plays was 100. All
of the other conditions were the same as in (3).
(5) Multi-armed bandit demonstrations in Figs. 4–6: The numbers of plays in the multi-armed
bandit problems whose results are depicted in Figs. 4–6 are summarized in Table 1. The
suppression of the number of repetitions is due to the limitations of our computing environment.
Table 1 Setting of the slot machine playing
Number of bandits
Number of consecutive plays
Number of repetitions
2
500
10,000
4
1000
1000
8
5000 (shown only until 2000 in Fig. 4)
100
16
5000 (shown only until 3000 in Fig. 4)
100
32
5000
100
64
10,000
100
(6) 32-armed bandit problem setting (Fig. 4a(v)): The reward probabilities of Machines 0–31
were 0.7, 0.5, 0.9, 0.1, 0.7, 0.5, 0.7, 0.5, 0.7, 0.5, 0.7, 0.5, 0.7, 0.5, 0.7, 0.5, 0.7, 0.5, 0.7, 0.5, 0.7,
0.5, 0.7, 0.5, 0.7, 0.5, 0.7, 0.5, 0.7, 0.5, 0.7, and 0.5, where the best machine was Machine 2. The
contradiction conditions were satisfied. For example, the MSB of the best machine (Machine 2)
was 0, but the sum of the reward probabilities of the machines whose MSBs were 0 was smaller
than that for the machines whose MSBs were 1 (9.4 versus 9.6); the same structure held from the
second to fourth bits as well.
34
(7) 64-armed bandit problem setting (Fig. 4a(vi)): The reward probabilities of Machines 0–63
were 0.7, 0.5, 0.9, 0.1, 0.7, 0.5, 0.7, 0.5, 0.7, 0.5, 0.7, 0.5, 0.7, 0.5, 0.7, 0.5, 0.7, 0.5, 0.7, 0.5, 0.7,
0.5, 0.7, 0.5, 0.7, 0.5, 0.7, 0.5, 0.7, 0.5, 0.7, 0.5, 0.7, 0.5, 0.7, 0.5, 0.7, 0.5, 0.7, 0.5, 0.7, 0.5, 0.7,
0.5, 0.7, 0.5, 0.7, 0.5, 0.7, 0.5, 0.7, 0.5, 0.7, 0.5, 0.7, 0.5, 0.7, 0.5, 0.7, 0.5, 0.7, 0.5, 0.7, and 0.5,
where the best machine was Machine 2. The contradiction conditions were satisfied similarly to in
the 32-armed bandit problem described in (6).
(8) Autocorrelation of chaotic signals (Fig. 2b): The autocorrelation was computed based on all
10,000,000 sampling points of each sequence. The autocorrelation in Fig. 2b is the average of 120
kinds of autocorrelations. (For the coloured noise, the number of sequences was 40.)
References
1. Inagaki, T., et al. A coherent Ising machine for 2000-node optimization problems. Science
10.1126/science.aah4243 (2016).
2. Larger, L., et al. Photonic information processing beyond Turing: an optoelectronic 3
implementation of reservoir computing. Opt. Express 20, 3241–3249 (2012).
3. Brunner, D., Soriano, M. C., Mirasso, C. R. & Fischer, I. Parallel photonic information processing
at gigabyte per second data rates using transient states. Nat. Commun. 4, 1364 (2013).
4. Uchida, A., et al. Fast physical random bit generation with chaotic semiconductor lasers. Nat.
Photon. 2, 728–732 (2008).
5. Argyris, A., et al. Implementation of 140 Gb/s true random bit generator based on a chaotic
photonic integrated circuit. Opt. Exp. 18, 18763-18768 (2010).
35
6. Sutton, R. S. & Barto, A. G. Reinforcement Learning: An Introduction (The MIT Press,
Massachusetts, 1998).
7. Awerbuch, B. & Kleinberg, R. Online linear optimization and adaptive routing. J. Comput. Syst.
Sci. 74, 97–114 (2008).
8. Kuroda, K., Kato, H., Kim, S.-J., Naruse, M. & Hasegawa, M. Improving throughput using multiarmed bandit algorithm for wireless LANs. NOLTA, IEICE 9, 74-81 (2018)
9. Kroemer, O. B., Detry, R., Piater, J. & Peters, J. Combining active learning and reactive control
for robot grasping. Robot. Auton. Syst. 58, 1105–1116 (2010).
10. Silver, D., et al. Mastering the game of go without human knowledge. Nature 550, 354 (2017).
11. Daw, N., O’Doherty, J., Dayan, P., Seymour, B. & Dolan, R. Cortical substrates for exploratory
decisions in humans. Nature 441, 876–879 (2006).
12. Naruse, M., Terashima, Y., Uchida, A. & Kim, S. -J. Ultrafast photonic reinforcement learning
based on laser chaos. Sci. Rep. 7, 8772 (2017).
13. Robbins, H. Some aspects of the sequential design of experiments. B. Am. Math. Soc. 58, 527–535
(1952).
14. Auer, P., Cesa-Bianchi, N. & Fischer, P. Finite-time analysis of the multi-armed bandit problem.
Machine Learning 47, 235–256 (2002).
15. Kim, S. -J., Naruse, M., Aono, M., Ohtsu, M. & Hara, M. Decision Maker Based on Nanoscale
Photo-Excitation Transfer. Sci Rep. 3, 2370 (2013).
36
16. Naruse, M., et al. Decision making based on optical excitation transfer via near-field interactions
between quantum dots. J. Appl. Phys. 116, 154303 (2014).
17. Naruse, M., et al. Single-photon decision maker. Sci. Rep. 5, 13253 (2015).
18. Naruse, M., Berthel, M., Drezet, A., Huant, S., Hori, H. & Kim S.-J. Single Photon in Hierarchical
Architecture for Physical Decision Making: Photon Intelligence. ACS Photonics 3, 2505–2514
(2016).
19. Naruse, M., Tate, N., Aono, M. & Ohtsu, M. Information physics fundamentals of nanophotonics.
Rep. Prog. Phys. 76, 056401 (2013).
20. Eisaman, M. D., Fan, J., Migdall, A. & Polyakov, S. V. Single-photon sources and detectors. Rev.
Sci. Instrum. 82, 071101 (2011).
21. Tucker, R. S., Eisenstein, G. & Korotky, S. K. Optical time-division multiplexing for very high
bit-rate transmission. J. Lightwave Technol. 6, 1737–1749 (1988).
22. Soriano, M. C., García-Ojalvo, J., Mirasso, C. R. & Fischer, I. Complex photonics: Dynamics and
applications of delay-coupled semiconductors lasers. Rev. Mod. Phys. 85, 421–470 (2013).
23. Ohtsubo, J. Semiconductor lasers: stability, instability and chaos (Springer, Berlin, 2012).
24. Uchida, A. Optical communication with chaotic lasers: applications of nonlinear dynamics and
synchronization (Wiley-VCH, Weinheim, 2012).
25. Fox, R. F., Gatland, I. R., Roy, R. & Vemuri, G. Fast, accurate algorithm for numerical simulation
of exponentially correlated colored noise. Phys. Rev. A 38, 5938–5940 (1988).
37
26. Miyaguchi, T. & Akimoto, T. Anomalous diffusion in a quenched-trap model on fractal lattices.
Phys. Rev. E 91, 010102 (2015).
27. Kim, S. -J., Naruse, M., Aono, M., Hori, H. & Akimoto, T. Random walk with chaotically driven
bias. Sci. Rep. 6, 38634 (2016).
28. Gentle, J. E. Computational statistics (Springer Science & Business Media, New York, 2009).
29. Naruse, M. & Ishikawa, M. Analysis and Characterization of Alignment for Free-Space Optical
Interconnects Based on Singular-Value Decomposition. Appl. Opt. 39, 293–301 (2000).
30. Ishikawa, M. & McArdle, N. Optically interconnected parallel computing systems. Computer 31,
61–68 (1998).
Acknowledgements
This work was supported in part by CREST project (JPMJCR17N2) funded by the Japan Science and
Technology Agency, the Core-to-Core Program A. Advanced Research Networks and the Grants-inAid for Scientific Research (A) (JP17H01277) and (B) (JP16H03878) funded by the Japan Society for
the Promotion of Science.
Author contributions
M.N. and A.U. directed the project. M.N. designed the system architecture and the principle. T.M. and
A.U designed and implemented the laser chaos. T.M., A.U., and M.N. conducted the optical
38
experiments and data processing. M.N., T.M., A.U., H.H., H.S., and K.O. analysed the data. M.N. and
A.U. wrote the paper.
Competing interests
The authors declare no competing interests.
Correspondence and requests for materials should be addressed to M.N.
39
Fig. 1 | Architecture for scalable reinforcement learning using laser chaos. a Solving the
multi-armed bandit problem with N = 2M arms using a pipelined arrangement of comparisons between
thresholds and a series of chaotic signal sequences. b Chaotic time series with the definitions of the
40
inter-decision sampling interval (ΔS) and inter-bit sampling interval (ΔL) to arrive at a single decision.
The 2Z+1 threshold levels are also depicted, where Z is a natural number. c Schematic diagram of the
decision-making system architecture based on laser chaos and pipelined threshold processing.
41
Fig. 2 | Chaotic time series, inherent time-correlated structures, RF spectra, and
decision-making performance for the two-armed bandit problem. a Snapshots of the time
series used for solving the N-armed bandit problem. Four kinds of chaotic signals (Chaos 1–4) as well
42
as quasiperiodic sequences, pseudorandom numbers (RAND), and coloured noise are used. b Radiofrequency (RF) power spectra in Chaos 1–4 and quasiperiodic signal cases. c Evolution of the correct
decision ratio (CDR) indicating the likelihood of choosing the highest-reward-probability slot machine.
d Autocorrelation inherent in Chaos 1–4, quasiperiodic, and coloured noise cases.
Fig. 3 | Decision-making performance as a function of inter-decision sampling interval.
CDR comparison at cycles a 10 and b 100.
43
Fig. 4 | Scalable decision making of N-armed bandit problems. a Definitions of the two-,
four-, eight-, 16-, 32, and 64-armed problems. The conditions are equally arranged to ensure fair
44
comparisons. b Four kinds of reward probability arrangements in the four-armed bandit problem. c
CDR evolution. d Number of cycles necessary to reach a CDR of 0.95 as a function of N, which varies
approximately as N1.16.
Fig. 5 | Inter-bit sampling interval dependency. a Effects of the temporal structure inherent in
laser chaos on the decision-making performance, which can be negative or positive depending on the
given problem. b Evaluation of a moderate inter-bit sampling interval choice based on the coefficient
of variation (CV) of a. ΔL = 30 ps or 100 ps yields a lower CV for the Chaos 3 case. Meanwhile, the
unstructured time-domain properties of pseudorandom numbers are clearly shown by the flat
dependency on ΔL in the RAND case.
45
Fig. 6 | Difficulty of the given decision-making problem and threshold precision control.
While retaining the higher reward probability in the two-armed bandit problem (P0 = 0.9), the lower
reward probability P1 was set to a 0.5 and b 0.7 to examine the decision difficulty. The CDR increases
more rapidly in the easier decision-making problem (a) than in the harder one (b). In addition, a
decrease in the number of threshold levels prevents the system from reaching the correct decision,
especially for harder problems, due to insufficient exploration. c CDR at cycle 200 as a function of
decision difficulty.
46
Fig. 7 | Diffusivity analysis of the time series to investigate the underlying structure. a
Comparison of the ensemble averages of the time-averaged mean square displacements (ETMSDs),
where the values for Chaos 1–4 agree with the superiority of the resulting decision-making
performance. b Trajectories configured in a two-dimensional plane by x(t ) and x (t D ) with D
47
= 10,000 to examine the coverage by the time series. The trajectories spanned by the chaotic time
series are uniformly distributed and are also quantitatively analysed using the smaller condition
numbers shown in a.
48
| 2cs.AI |
УДК 004.9:66.013.512
ИНФОРМАТИКА
МОДУЛЬНАЯ ТЕХНОЛОГИЯ РАЗРАБОТКИ РАСШИРЕНИЙ САПР:
МОЛНИЕЗАЩИТА ЗДАНИЙ И СООРУЖЕНИЙ
В.В.Мигунов, Р.Р.Кафиятуллов, И.Т.Сафин
ЦЭСИ РТ при КМ РТ, г. Казань
Аннотация
Модульная технология разработки проблемно-ориентированных расширений САПР применена к задаче
проектирования молниезащиты зданий и сооружений с реализацией в программной системе TechnoCAD GlassX.
Разработана системная модель проектов молниезащиты, включающая структурированное параметрическое
представление (свойства объектов и их связи, общие установки и установки по умолчанию) и операции над ним,
автоматизирующие проектирование.
Библ.2
Abstract
V.V. Migunov, R.R. Kafiyatullov, I.T. Safin. The modular technology of development of the CAD expansions:
protection of the buildings from the lightning // Izvestiya of the Tula State University/ Ser. Mathematics. Mechanics.
Informatics. Tula: TSU, 2003. V._. N _. P. __–__.
The modular technology of development of the problem-oriented CAD expansions is applied to a task of designing of
protection of the buildings from the lightning with realization in program system TechnoCAD GlassX. The system model
of the drawings of lightning protection is developed including the structured parametric representation (properties of
objects and their interdependence, general settings and default settings) and operations with it, which efficiently automate
designing.
Bibl.2
Настоящая работа посвящена применению модульной технологии разработки проблемноориентированных расширений систем автоматизированного проектирования (САПР) реконструкции
предприятия, общие положения которой изложены в [1]. Объект приложения технологии автоматизация подготовки чертежей проектов молниезащиты зданий и сооружений (МЗС), выполненная
в САПР TechnoCAD GlassX. Кроме обычных требований единой системы конструкторской
документации и системы проектной документации для строительства, чертежи МЗС должны отвечать
правилам [2]. В [2] задаются методы расчета горизонтальных и вертикальных сечений зон защиты
молниеприемников различных типов. Изображение и характеристика этих сечений на чертеже
защищаемых зданий и сооружений и является проектом их молниезащиты.
Учитываются и рассчитываются комплексы из стержневых (одиночных, двойных и многократных),
тросовых (одиночных и двойных) и сеточных молниеприемников (МП). Основой автоматизации
является быстрый (мгновенный) расчет зон защиты по [2] для подбираемых проектировщиком
вариантов молниезащитных сооружений с демонстрацией на чертеже нескольких сечений зон защиты в
нескольких проекциях и таблицы основных данных ("Таблица расчета молниезащиты").
Сечения зон защиты всегда простые, без изломов секущих плоскостей. Все вертикальные сечения –
вынесенные (располагаются вне изображения предмета). На виде сверху располагаются одно или
несколько горизонтальных сечений зон защиты для разных высот, на каждом вертикальном сечении –
только одно сечение зон защиты. Каждое из сечений зон защиты может учитывать свой набор
молниеприемников.
При разработке МЗС используется заранее подготовленный чертеж с изображениями зданий и
сооружений (например, строительная подоснова зданий), для которых разрабатывается МЗС. Чертеж
МЗС представляется в виде одного обязательного вида сверху (плана), нескольких необязательных
вертикальных сечений и содержит следующие специфичные элементы:
• изображения зданий и сооружений в нескольких видах, опор молниеотводов и токоотводов,
подземные инженерные сооружения: подземный трубопровод, резервуары, фундамент и пр. Эти
элементы вычерчиваются обычными средствами TechnoCAD GlassX, без автоматизации в рамках
данного проблемно-ориентированного расширения;
• изображения стержневых молниеприемников. Отдельно стоящий на земле МП на виде сверху
изображается квадратом с диагональными линиями; стоящий на каком-либо объекте МП на виде сверху
изображается жирной точкой; на других проекциях и видах МП изображаются равнобедренным
треугольником, вытянутым вверх по его высоте;
• изображения сеточных МП в виде ограничивающих многоугольников. При разработке МЗС
1
многоугольники для наглядности штрихуются как неметаллы (2х45°), при дооформлении чертежа они
штрихуются проектировщиком более точно, с учетом шага сетки на различных участках;
• изображения одиночных и двойных тросовых МП;
• позиционные обозначения МП;
• обозначение вида сверху;
• обозначения вертикальных сечений на виде сверху (включая изображение секущего отрезка) и
на самом вертикальном сечении;
• размеры, характеризующие расстояния между МП;
• линии горизонтальных и вертикальных сечений зон защиты, порождаемых различными
наборами МП;
• текстовые обозначения горизонтальных сечений зон защиты на виде сверху с указанием высот;
• размеры радиусов зон защиты одиночных МП на нулевой высоте;
• размеры минимальной ширины зон защиты двойных МП на нулевой высоте;
• таблица расчета молниезащиты;
• изображения заземлителей, обеспечивающих растекание тока молнии в земле;
• другая текстовая информация, более общего характера, относящаяся к зонам молниезащиты,
конструкции МП, обозначениям имеющихся на чертеже видов.
Используется внутренняя трехмерная прямоугольная декартова правая система координат с
единицами измерения мм натуры, начало которой считается находящимся на нулевой отметке высоты и
называется далее базовой точкой. На виде сверху оси X и Y внутренней системы координат идут вдоль
одноименных осей чертежа, а ось Z идет от плоскости чертежа к пользователю.
Параметрическое представление содержит списки объектов МЗС, установки построения по
умолчанию и общие установки МЗС, вводимые, выбираемые или указываемые проектировщиком. Для
краткости ниже словами "установки шрифта" будут называться размер, угол наклона и степень сжатия
чертежного шрифта системы TechnoCAD GlassX. Все координаты и размеры задаются во внутренней
системе координат, если не оговаривается иное. Если речь идет о размерах в отпечатанном чертеже,
говорится о системе отсчета Бумага.
Молниеприемники
Список молниеприемников для каждого МП содержит: позиционное обозначение (однострочный
текст); тип (однострочный текст); вариант конструкции (стержневой, сеточный, тросовый и двойной
тросовый); признак того, что стержневой МП – отдельно стоящий, а не установлен на каком-либо
объекте; высота конструкции (кроме сеточного); список трехмерных точек, цвет и тип линии. Для
стержневого варианта задается одна точка – верхняя точка стержня, для сеточного – не менее трех
вершин ломаной без самопересечений с обходом против часовой стрелки, ограничивающей плоский
горизонтальный многоугольник сетки (координаты Z одинаковы для всех углов многоугольника), для
тросового – две точки опор на их высоте (для обеих точек координаты Z одинаковы), для двойных
тросовых – то же плюс смещение второго троса от первого и его высоту. Позиционное обозначение и
высота МП размещаются в таблице расчета молниезащиты соответственно в графах "N°N°
молниеприемников" и "Высота молниеприемника (h)";
Вертикальные сечения чертежа
Список (вертикальных) сечений чертежа – проекций для каждого сечения содержит: буквенное
обозначение сечения (возможно, включая слово "Повернуто"); масштаб Бумага/Натура; проекцию
базовой точки внутренней системы координат на плоскость сечения в системе отсчета Бумага всего
чертежа; секущий отрезок на плоскости XY внутренней системы координат (след секущей плоскости);
параметры для вычерчивания буквенного обозначения вертикального сечения на виде сверху: длина
штрихов на краях; смещение стрелки от края штриха; смещение центра габаритов буквенного
обозначения от ближайшего конца секущего отрезка (все размеры в мм Бумаги); параметры для вывода
буквенного обозначения сечения чертежа на своем виде: смещение точки середины полки буквенного
обозначения относительно проекции базовой точки в системе отсчета Бумага (текст обозначения
центрируется относительно нее); смещения от полки до текстов над и под ней; вариант включения
масштаба сечения в его буквенное обозначение (вывести в одной строке; вывести отдельной строкой
под полкой); установку шрифта. Положение буквенного обозначения относительно секущего отрезка
задает направление взгляда. На каждом вертикальном сечении чертежа располагается одно сечение зоны
защиты.
Сечения зон защиты
Список сечений зон защиты (горизонтальных и вертикальных) для каждого сечения содержит:
2
ссылку на вертикальное сечение чертежа из их списка (в том числе 0 для вида сверху); список
молниеприемников, учитываемых при построении сечения (не указанные в этом списке МП не
изображаются на сечении чертежа и их влияние на сечение зоны защиты игнорируется); отметку высоты
сечения (для горизонтальных сечений зон защиты, размещаемых на виде сверху); цвет и тип линии
границы сечения зоны защиты.
Тексты молниеприемников
Список текстов молниеприемников для каждого текста содержит: ссылку на МП из их списка;
ссылку на сечение чертежа из их списка (для текстов, располагающихся на вертикальном сечении);
смещение точки начала текста относительно проекции базовой точки в системе отсчета Бумага;
смещение точки указания сноски от первой точки МП в системе отсчета Бумага; признак направления
сноски на конец или на начало полки.
Тексты сечений зон защиты на виде сверху
Список текстов (отметок высот) сечений зон защиты на виде сверху для каждого текста содержит:
смещение точки начала текста от проекции базовой точки в системе отсчета Бумага; ссылку на элемент
списка сечений зон защиты; угол направления сноски из точки начала текста; признак направления
сноски на конец или на начало полки; признак расположения текста в двух строках, а не в одной.
Размеры расстояний между осями стержневых молниеприемников
Список (линейных) размеров – расстояний между осями стержневых молниеприемников в метрах
для каждого из них содержит: ссылки на два МП из их списка, расстояние между которыми указывается
в размере; ссылку на вертикальное сечение чертежа из их списка (для размеров, расположенных на
вертикальных сечениях); смещение размерной от первой точки начала выносных и смещение текстов от
размерной линии в направлении допустимого расположения текстов (система отсчета Бумага); сведения
о засечках или стрелках (стрелка внутрь, стрелка наружу, засечка).
Размеры радиусов зон защиты одиночных стержневых молниеприемников на виде сверху
Список размеров радиусов (в метрах) зон защиты одиночных стержневых молниеприемников на
виде сверху (представляют собой линейные размеры типа "Радиус" с текстом типа "Rx1 = 17.00") для
каждого такого размера содержит: текст обозначения параметра, например "Rx1"; ссылка на МП из их
списка (начало размерной); ссылка на сечение зоны защиты из их списка; угол направления размерной
против часовой от положительного направления оси X чертежа; признак автоматического
позиционирования текста размера посередине размерной линии; смещение текстов от размерной линии
в направлении допустимого расположения текстов в системе отсчета Бумага (учитывается при
автоматическом позиционировании текста размера); положение точки начала текста относительно
базовой точки в системе отсчета Бумага (учитывается при ручном позиционировании); признак наличия
сноски (отсутствует (по умолчанию), от середины размерной, от начала или конца размерной); признак
проведения сноски к началу (по умолчанию) или концу полки; сведения о засечках или стрелках
(стрелка внутрь, стрелка наружу или засечка).
Размеры радиусов зон защиты одиночных молниеприемников на вертикальных сечениях
Список размеров (в метрах) радиусов зон защиты одиночных молниеприемников на вертикальных
сечениях на нулевой отметке высоты (представляют собой линейные размеры с текстом типа "R1 =
25.00") для каждого такого размера содержит: текст обозначения размера – радиуса, например, "R1";
ссылку на молниеприемник из их списка (начало размерной); ссылку на вертикальное сечение из их
списка; смещение размерной от первой точки начала выносных и смещение текстов от размерной линии
в направлении допустимого расположения текстов (оба параметра в системе отсчета Бумага); сведения о
засечках или стрелках (стрелка внутрь, стрелка наружу или засечка); направления размерной от МП
(влево или вправо).
Размеры минимальных ширин зон защиты двойных молниеприемников на виде сверху
Список размеров (в метрах) минимальных ширин зон защиты двойных молниеприемников на виде
сверху (представляют собой линейные размеры типа "Радиус" с текстом типа "Rсx = 13.50") для каждого
размера содержит: текст обозначения параметра, например "Rcx"; ссылки на два МП из их списка (их
порядок важен: размерная идет налево перпендикулярно направлению от первого МП ко второму);
ссылка на сечение зоны защиты из их списка; признак автоматического позиционирования текста
размера посередине размерной линии, а не вручную пользователем; смещение текстов от размерной
линии в направлении допустимого расположения текстов в системе отсчета Бумага (учитывается при
автоматическом позиционировании текста размера); положение точки начала текста относительно
базовой точки в системе отсчета Бумага (учитывается при ручном позиционировании); признак наличия
сноски (отсутствует (по умолчанию), от середины размерной, от начала или конца размерной); признак
проведения сноски к началу (по умолчанию) или концу полки; сведения о засечках или стрелках
(стрелка внутрь, стрелка наружу или засечка).
3
Молниеприемники, вносимые в таблицу расчета молниезащиты
Список молниеприемников, вносимых в таблицу расчета молниезащиты, для каждого элемента
списка содержит информацию для одного одиночного или двух связанных друг с другом (двойной) МП:
ссылку на молниеприемник из их списка (это либо одиночный молниеприемник, либо первый из пары
молниеприемников, составляющих двойной МП); ссылка на второй молниеприемник (для двойного
МП); значение отметки высоты защищаемого уровня относительно нулевой отметки.
Заземлители на виде сверху
Список заземлителей для каждого элемента списка содержит: смещение центра заземлителя от
проекции базовой точки в системе отсчета Натура; тип линии отрезков (окружности берут из него
толщину, штриховыми не бывают); число стержней в заземлителе (от 1 до 32); угол направления
отрезков заземлителя против часовой от положительного направления оси X чертежа; расстояние между
стержнями (от 3000 до 25000 мм Натуры); диаметр стержней в мм Натуры.
В приведенных списках возникают связи принадлежности у объектов, отвечающих за оформление
чертежа МЗС (тексты и размеры – к молниеприемникам и сечениям зон защиты, сечения зон защиты – к
проекциям). Эти связи позволяют автоматизировать различные операции по модификации МЗС.
Например, при удалении МП удаляются и связанные с ним тексты, радиусы и выносные линии размеров
(сами размеры автоматически перегенерируются или удалятся). При переносе МП автоматически
перечерчиваются учитывающие его сечения зон защиты, размеры радиусов зон защиты и др. При
изменении высоты горизонтальных сечений зон защиты на виде сверху тексты этих сечений и радиусы
зон защиты молниеприемников автоматически перегенерируются.
Установки МЗС
Установки МЗС делятся на установки, действующие по умолчанию, и общие установки. Общие
установки при изменении приводят к соответствующим изменениям во всех элементах. Установки,
действующие по умолчанию, применяются только при добавлении новых объектов в МЗС.
Общие установки:
• положение базовой точки внутренней системы координат (вида сверху) в чертеже, система
отсчета Бумага чертежа;
• тип зоны защиты (А или Б), определяющий уровень жесткости требований к зоне защиты;
• установки вида сверху: смещение точки середины полки буквенного обозначения относительно
базовой точки в системе отсчета Бумага; обозначение вида сверху; масштаб Бумага/Натура вида сверху;
смещения от полки до текстов над и под ней в системе отсчета Бумага; признак добавления масштаба
вида сверху к его обозначению отдельной строкой под полкой, а не в одной строке с ним; установка
шрифта обозначения вида сверху.
• установки таблицы расчета молниезащиты: вектор смещения верхнего левого угла таблицы от
базовой точки в системе отсчета Бумага; единица измерения выводимых в таблице числовых значений
(миллиметр, сантиметр или метр); установки точности представления числовых значений в таблице
(число знаков после десятичной точки, 0-7); высота строк в области данных, в мм Бумаги (3-50 мм);
типы линий шапки, краев таблицы и линий-разделителей данных таблицы; установка шрифта для всех
текстов таблицы в области данных (кроме текстов шапки); способ упорядочивания строк (не
сортировать – без упорядочивания; сортировать – сквозное упорядочивание строк в алфавитном порядке
по позиционным обозначениям МП; сортировать по группам – упорядочивание строк в алфавитном
порядке по позиционным обозначениям МП, причем сначала идут все одиночные, а затем все двойные
молниеприемники); признак объединения строк для одиночных молниеприемников при совпадении у
них всех числовых значений;
• молниеприемников: сторона квадрата условного графического обозначения (УГО) МП на плане;
диаметр жирной точки УГО МП на плане; диаметр жирной точки – сечения троса на вертикальном
сечении; ширина основания треугольника УГО МП на вертикальном сечении (все в мм Бумаги);
• вертикальных сечений: установка шрифта букв "А А" буквенного обозначения на виде сверху;
установка шрифта для "А – А повернуто М 1: 100" на самом вертикальном сечении; длина хвоста
стрелки; длина самой стрелки; цвета обозначения сечения на виде сверху и на своем виде (все в мм
Бумаги);
• текстов молниеприемников: установка шрифта и цвет;
• текстов (отметок высот) сечений зон защиты: установка шрифта; установки точности
представления значения отметки высоты (число знаков после десятичной точки, 0-7) и цвет;
• общие для всех размеров: установка шрифта; длина перехода выносных за размерную (0-7.5мм);
размер стрелки или засечки (0-12.6мм);
• каждого вида размера: число знаков после десятичной точки и цвет;
4
• заземлителей: цвет;
• для штрихования сетки для сеточного варианта МП во время работы: установка штрихования –
базовая, не линейчатая.
Установки, действующие по умолчанию:
• молниеприемников: вариант конструкции (стержневой, сеточный, тросовый и двойной
тросовый); отдельно стоящий МП, или устанавливается на объекте (для стержневого МП); цвет и тип
линии;
• обозначений вертикальных сечений: длина штрихов на краях (0-25.5 мм); смещение стрелки от
края штриха (0-25.5 мм); смещения от полки до текстов над и под ней (0-25.5 мм), все в мм Бумаги;
признак добавления масштаба сечения к его буквенному обозначению отдельной строкой под полкой, а
не в одной строке с ним;
• сечений зон защиты: цвет и тип линии;
• общие для всех текстов: признак соединения сноски с полкой текста (сноска на конец полки или
на начало);
• текстов (отметок высот) сечений зон защиты: признак расположения текста в двух строках или в
одной;
• общие для всех размеров: смещение текстов всех размеров от размерной линии в направлении
допустимого расположения текстов в мм Бумаги; признак автоматического позиционирования текста
размера посередине размерной линии, а не задается вручную пользователем; признак наличия сноски и
указания на размерную (нет, сноска указывает на середину, начало или конец размерной линии);
признак проведения сноски к началу или к концу полки текста; сведения о засечках или стрелках
(стрелка внутрь, стрелка наружу или засечка);
• размеров – радиусов зон защиты одиночных стержневых МП на виде сверху: признак включения
в текст значения высоты сечения зоны защиты (включать или не включать);
• заземлителей: тип линии отрезков (окружности берут из него толщину, штриховыми не бывают);
число стержней в заземлителе (от 1 до 32); угол направления отрезков заземлителя против часовой
стрелки от положительного направления оси X чертежа; расстояние между стержнями (от 3000 до 5000
мм Натуры); диаметр стержней в мм Натуры.
Описанное параметрическое представление МЗС может сохраняться на диске как комплект
параметров, без геометрической части. Однако при выборе комплекта параметров на диске для работы с
ним пользователь просматривает изображение МЗС, которое генерируется при перемещении по меню в
режиме on-line.
Операции по подготовке чертежа проекта МЗС
Над параметрическим представлением МЗС проектировщик последовательно выполняет различные
операции, выбирая их из основного меню.
"С диска" – чтение из дискового файла ранее сохраненного в нем параметрического представления
МЗС.
"Из чертежа" – взятие параметрического представления из проекта МЗС в чертеже.
"Добавление" – группа операций по добавлению в модуль МЗС: молниеприемников, сечений
чертежа, сечений зон защиты, текстов молниеприемников, текстов сечений зон защиты на виде сверху,
размеров расстояний между осями стержневых молниеприемников, размеров радиусов зон защиты
одиночных стержневых молниеприемников на виде сверху, размеров радиусов зон защиты одиночных
молниеприемников на вертикальных сечениях, размеров минимальных ширин зон защиты двойных
молниеприемников на виде сверху, заземлителей на виде сверху, молниеприемников, вносимых в
таблицу расчета молниезащиты. Добавление объектов выбранного типа повторяется циклически до
отказа от очередного добавления.
"Удаление" – запускается режим повторяющегося удаления объектов МЗС: молниеприемников,
сечений чертежа, сечений зон защиты, текстов молниеприемников, текстов сечений зон защиты на виде
сверху, размеров расстояний между осями стержневых молниеприемников, размеров радиусов зон
защиты одиночных стержневых молниеприемников на виде сверху, размеров радиусов зон защиты
одиночных молниеприемников на вертикальных сечениях, размеров минимальных ширин зон защиты
двойных молниеприемников на виде сверху, заземлителей на виде сверху, молниеприемников,
вносимых в таблицу расчета молниезащиты, выбираемых курсором по одному.
При удалении молниеприемника на вертикальном сечении он удаляется лишь из него (это
равнозначно удалению молниеприемника из сечения зоны защиты на вертикальном сечении чертежа),
на виде сверху он не удаляется.
"Перенос" – запускается циклический режим переноса объектов: молниеприемников, сечений
5
чертежа, текстов молниеприемников, текстов сечений зон защиты на виде сверху, размерной линии
размеров расстояний между осями стержневых молниеприемников, размерной линии размеров радиусов
зон защиты одиночных стержневых молниеприемников на виде сверху, размерной линии размеров
радиусов зон защиты одиночных молниеприемников на вертикальных сечениях, заземлителей на виде
сверху, выбираемых курсором по одному.
Перенос молниеприемника на вертикальном сечении чертежа возможен только по вертикали, это
равнозначно изменению высоты молниеприемника.
"Редактирование" – группа операций по редактированию объектов МЗС:
•
"Добавление МП в сечение зоны защиты";
•
"Удаление МП из сечения зоны защиты";
•
"Добавление вершины сеточного МП";
•
"Перенос вершины (опоры) МП";
•
"Удаление вершины сеточного МП";
•
"Перенос обозначения сечения чертежа";
•
"Изменение секущего отрезка";
•
"Совмещение видов";
•
"Перенос точки указания сноски";
•
"Перенос текста";
•
"Правка данных в таблице".
"Копирование" – запускается циклический режим молниеприемников и заземлителей, выбираемых
курсором по одному. Цикл работы с одним объектом выглядит так:
•
в чертеже выбирается один из перечисленных выше объектов;
•
создается копия объекта, изображение которой начинает перемещаться по полю чертежа вместе
с курсором;
•
в нужном месте чертежа нажимается клавиша Enter (левая кнопка мыши).
Цикл повторяется до нажатия Esc (правой кнопки мыши) в ситуации выбора очередного объекта
или указания положения копии объекта.
"Свойства" – запускается циклический режим изменения свойств объектов МЗС:
молниеприемников, сечений чертежа, сечений зон защиты, текстов молниеприемников, текстов сечений
зон защиты на виде сверху, размеров расстояний между осями стержневых молниеприемников,
размеров радиусов зон защиты одиночных стержневых молниеприемников на виде сверху, размеров
радиусов зон защиты одиночных молниеприемников на вертикальных сечениях, размеров минимальных
ширин зон защиты двойных молниеприемников на виде сверху, заземлителей на виде сверху,
выбираемых курсором по одному.
"Перенос МЗС" – перенос всего проекта МЗС в другое место на поле чертежа.
"Установки" – изменение установок, действующих при подготовке МЗС:
•
"Тип зоны, установки плана";
•
"Молниеприемники";
•
"Сечения чертежа";
•
"Сечения зон защиты";
•
"Тексты";
•
"Тексты молниеприемников";
•
"Тексты сечений зон защиты";
•
"Размеры";
•
"Размеры расстояний";
•
"Радиусы на виде сверху";
•
"Радиусы на верт.сечениях";
•
"Размеры мин.ширин";
•
"Заземлители";
•
"Таблица расчета".
"Высота в точке" – подсветка высоты полной зоны защиты (от всех молниеприемников). Эта
операция позволяет в любой точке, указываемой курсором на плане, отображать значение высоты
полной зоны защиты.
"Сечение в точке" – операция идентична предыдущей, но только вместе с высотой полной зоны
защиты отображается и ее горизонтальное сечение на этой высоте. Текст значения высоты и
изображение сечения зоны защиты меняются вместе с движением курсора.
6
"Рельеф" – позволяет просмотреть "послойный" набор
горизонтальных сечений полной зоны защиты, аналогичный
изобазам в географических картах. Горизонтальные сечения
строятся, начиная с нулевой отметки высоты и до
максимальной из габаритных высот молниеприемников с
шагом в 5% этого диапазона. Далее включается режим
просмотра с подсветкой высоты полной зоны защиты
(рис.1).
"На диск" – запись на диск текущего параметрического
представления МЗС. Его можно будет повторно
использовать в дальнейшем, копировать на другие рабочие
места.
"В чертеж" – помещение проекта МЗС в чертеж.
Таблица расчета молниезащиты
Изображение
таблицы
расчета
молниезащиты
вычерчивается автоматически на основе текущего
состояния МЗС, выбора одиночных и двойных
молниеприемников и соответствующих им значений высот
защищаемых уровней (рис.2).
Рис.1. Рельеф полной зоны защиты
Рис.2. Таблица расчета молниезащиты
Выводы
Изложенные в настоящей работе состав и особенности реализации операций подготовки проекта
МЗС вместе с их структурированным параметрическим представлением являются системной моделью
чертежей молниезащиты зданий и сооружений, используемой при проектировании реконструкции
предприятий согласно системе проектной документации для строительства. Модель разработана по
модульной технологии [1] и прошла практическую апробацию в условиях проектно-конструкторского
отдела крупного предприятия химической промышленности..
7
Литература
[1] Мигунов В.В. Модульная технология разработки проблемно-ориентированных расширений САПР
реконструкции предприятия / Материалы Второй международной электронной научно-технической
конференции "Технологическая системотехника" (ТСТ'2003), г.Тула, 01.09.2003-30.10.2003
[Электронный ресурс] / Тульский государственный университет. – Режим доступа:
http://www.tsu.tula.ru/aim/, свободный. – Загл. с экрана. – Яз. рус., англ.
[2] РД 34.21.122-87 "Инструкция по устройству молниезащиты зданий и сооружений". – М.:
Министерство энергетики и электрификации СССР, 1988.
8
| 5cs.CE |
Compact Environment-Invariant Codes for Robust Visual Place Recognition
arXiv:1709.08103v1 [cs.CV] 23 Sep 2017
Unnat Jain
Dept. of Computer Science
University of Illinois Urbana-Champaign
Urbana, USA
uj2@illinois.edu
Vinay P. Namboodiri
Dept. of Computer Science & Engg.
Indian Institute of Technology
Kanpur, India
vinaypn@iitk.ac.in
Abstract—Robust visual place recognition (VPR) requires
scene representations that are invariant to various environmental challenges such as seasonal changes and variations due
to ambient lighting conditions during day and night. Moreover,
a practical VPR system necessitate compact representations of
environmental features. To satisfy these requirements, in this
paper we suggest a modification to the existing pipeline of
VPR systems to incorporate supervised hashing. The modified
system learns (in a supervised setting) compact binary codes
from image feature descriptors. These binary codes imbibe
robustness to the visual variations exposed to it during the
training phase, thereby, making the system adaptive to severe
environmental changes. Also, incorporating supervised hashing makes VPR computationally more efficient and easy to
implement on simple hardware. This is because binary embeddings can be learnt over simple-to-compute features and the
distance computation is also in the low dimensional hamming
space of binary codes. We have performed experiments on
several challenging data sets covering seasonal, illumination
and viewpoint variations. We also compare two widely used
supervised hashing methods of CCAITQ [1] and MLH [2] and
show that this new pipeline out-performs or closely matches
the state-of-the-art deep learning VPR methods that are based
on high-dimensional features extracted from pre-trained deep
convolutional neural networks.
Gaurav Pandey
Dept. of Electrical Engg.
Indian Institute of Technology
Kanpur, India
gpandey@iitk.ac.in
(a) Nordland spring-summer & summer-winter data sets: Mild to
severe appearance change & no viewpoint change.
(b) Alderley Day-Night data set: Severe appearance & mild viewpoint change.
(c) St. Lucia data set: Mild appearance & severe viewpoint change.
Keywords-visual place recognition; similarity learning; hashing; convolutional neural network; dynamic time warping
Figure 1: Appearance and viewpoint variations in data sets
on which we test our visual place recognition pipeline.
I. I NTRODUCTION
Robots are often required to operate in environments
for a long period of time ranging from days, months to
even years. For autonomous operation of robots in such
cases, the robot has to recognize places that it has visited
before. This ability of the robot to recognize places has
a wide range of applications in its autonomous navigation
capabilities that include global localization and loop-closure
detection. The task of VPR for a long-term autonomous
visual navigation system becomes extremely challenging
because over a long period of time the appearance of a place
can drastically change. Traditional visual place recognition
approaches like [3], [4] focused on situations where robot
has to recognize places that it has recently visited, where
the difference between query and database images was
mainly due to different view-point of sensors. However,
for a long-term autonomous visual navigation system the
VPR method should be robust to seasonal, illumination
and viewpoint variations (Fig. 1). It is also desired that
the VPR system could imbibe robustness to an unfamiliar
variation from some learning examples. It should also be
real-time & storage efficient for it to have utility for robotic
applications. Additionally, it is advantageous if a visual place
recognition system is implementable on a simple hardware
configuration so that it can have much wider application.
Several methods that are robust to certain visual variations
have been proposed in the past. Sünderhauf et al. [5] used
concatenated BRIEF-gist descriptor to incorporate some robustness to viewpoint variations. Milford et al. [4] proposed
to use the video sequence instead of independent images,
thereby utilising the continuity constraint of consecutive
images in-order to remove outliers. Lowry et al. [6] employed linear regression techniques to predict the temporal
appearance of a place based on the time of the day. Neubert
et al. [7] uses a vocabulary of superpixels to predict the
change in appearance of the scene.
Many visual place recognition methods based on deep
convolutional neural network (CNN)s which are pre-trained
on the task of either object recognition or scene classification, have recently been proposed [8]–[11]. The deep CNN
based methods have shown to outperform previous state-ofthe-art visual place recognition techniques. However, CNN
feature based methods like [8] with no dimensionality reduction are slow because of computations on high dimensional
feature vectors. Several dimensionality reduction techniques
have been proposed to speed-up the algorithm. Milford et
al. [12] uses principal component analysis (PCA) as an
extra step to reduce the dimensionality of feature vectors.
Sünderhauf et al. [11] utilized Gaussian random projections
to obtain shorter features than the raw conv3 CNN features.
Sünderhauf et al. [10] borrows from research in unsupervised hashing methods to obtain binary codes of 8192 bits
to describe the images. It is important to note that all the
above dimensionality reduction methods are unsupervised
and have lower accuracies than a visual place recognition
method which uses the ‘non-reduced’ raw features.
Unlike the existing methods, the proposed VPR pipeline
uses supervised dimensionality reduction to reduce high
dimensional real features to compact and more semantic
binary codes. All other VPR methods [10]–[12] demonstrate
lower performance due to the dimensionality reduction step.
On the contrary, a supervised VPR system based on the
learning of compact binary codes from image features significantly improves the image retrieval performance when
compared to VPR methods based on the corresponding raw
features. We improve on accuracy while still using a binary
code representation. This representation keeps our VPR
pipeline real-time and space efficient. Also, the proposed
method is feature agnostic. Therefore, our VPR pipeline is
capable of processing both simple gist [13] features (whose
computation requires only simple convolutions with Gabor
filters) as well as complex deep learning based CNN features.
With the boost of accuracy, our VPR method is capable of
bootstrapping the accuracies of simple-to-compute features
like gist to performance comparable to (usually better than)
existing VPR methods based on pre-trained CNN features.
The rest of the paper is organized in the following
sections:
Section II describes how gist & CNN features have been
utilized for VPR research. It also motivates the need of
a system capable of adaptively learning robustness rather
than relying only on the pre-learnt robustness of popular
image descriptors. Section III describes CCAITQ [1], the
supervised hashing method that we utilize in our VPR
pipeline. Section IV gives details of the experimentation data sets, training & testing set division, assigning similarity
labels, running hashing methods and testing. We explain
important inferences of our experiments in section IV-C and
conclude the paper with section V.
Feature Dimension
512 or
Gist
2048
fc6
4096
conv3
43264
(VGG-f)
Advantage
Compact global
representation
Robust to viewpoint
variations
Robust to severe
appearance
variations
Disadvantage
Low performance in
severe changes
Susceptible to appearance
variations
Very high dimensional
and susceptible to
viewpoint variations
Table I: Details about popular image features for VPR
II. BACKGROUND
A. Gist based Visual Place Recognition methods
Initial feature based VPR methods (like FAB-MAP [3])
built a visual vocabulary from local SIFT or SURF descriptors and then used a bag-of-words (BoW) image descriptor to
find the best matching image frame corresponding to a query
frame. Later, success in scene classification based on gist
features [14] led to gist features being applied to the place
recognition tasks. Gist features were adapted to panoramic
views for VPR [15]. Visual Loop Closing methods which
earlier utilized SIFT or SURF BoW, demonstrated much
superior performance after adapting gist descriptors [5], [16],
[17].
B. Deep CNN based VPR
AlexNet [18] made deep CNNs popular in Computer
Vision research in 2012. Studies in [19], [20] demonstrated
that features extracted from deep CNN (which are pretrained on object recognition task) can be used for generic
visual recognition tasks. CNN features performed better than
features which were handcrafted specific for the tasks like
domain adaptation, fine grained recognition and scene recognition. Thereafter, research in VPR has almost suddenly
turned to explore the power of CNN based features. Work
of Sünderhauf et al. [10], [11] has extensively compared
features extracted from different CNN layers, on challenging
VPR data sets. Instead of extracting global CNN image
descriptors like [10], [11] extracts pooled local conv3 CNN
features of fifty landmark regions and then use similarity
matching of these local features to obtain the image which is
the best match to a query image. Inferring from the detailed
empirical study of [10], we focus our experiments on features extracted from two layers - lower level convolutional
layer (third layer - conv3) and higher level fully connected
layer (sixth layer - fc6).
Drawbacks of pre-trained CNN based VPR approaches:
• Choice of layer: [10] empirically validates the appearance invariance of conv3 and and viewpoint invariance
of fc6 layers. However, for a new environment displaying a unknown weighted mixture of multiple visual
variations, one has no insight of which layer features
to utilize.
•
•
•
Dimensionality: We validated [10]’s claim that lower
depth features from conv3 layer are invariant to severe
appearance and mild viewpoint variations. Despite their
good pre-learnt robustness, raw conv3 features are
not useful for VPR due to their huge dimensionality,
64896 for [18] and 43264 for [21], which significantly
increases the computation time.
Storage size: Each dimension for raw real-valued features is 256 bits leading to 1.3Mb size of one conv3
feature vector. Table II gives details about the storage
capacity required for storing different features.
Hardware requirement: Deep CNNs have complex architecture and require much advanced hardware devices
for training and feature extraction. The model of a CNN
has over millions of parameters and merely loading
a CNN model requires huge amount of RAM space,
which is often impractical for use in many robotic
vision applications.
C. Learning invariance vs. pre-learnt invariance
Pre-trained CNN features acquire invariance to the variations that the CNN has been exposed to while training.
Since the popular pre-trained CNNs are trained for object
recognition, the variations in images of the same object
might not cover environmental variations prevalent in places.
For example, glaring lights in the night traversal of Alderley
data set, change of texture from grassy (in summer) to snowy
(in winter) are variations which are rarely seen in object
recognition data sets. Such variations are not as generic as
simple illumination, rotation and scale variations which pretrained CNNs have already been exposed to, while training
on data sets like ImageNet [22]. Thus, any feature based
VPR system needs some bootstrapping to learn robustness
against unfamiliar variations. We incorporate this bootstrapping by supervised hashing methods discussed later.
Is obtaining supervision easy: Obtaining GPS tagged
images is cheap and easy, and can be done by any vehicle
which travels with a camera and GPS. Multiple traversals of
the same tracks helps create cluster of images of the same
places at different times of the day and different seasons of
the year. This forms the ground truth for a supervised VPR
system. There are many publicly available data sets which
have multiple images of the same place (under different
conditions). These include Nordland [23], Alderley [4], St.
Lucia [24], KITTI [25], Pittsburgh 250k [26] and Tokyo 24/7
[27] data sets. Thus, supervision is simple to incorporate in
VPR and is especially important when a VPR system is put
in a new environment.
III. H ASHING METHODS
In efficient retrieval systems, high dimensional vectors are
often compressed using similarity-preserving hash functions
which map long features to compact binary codes. Hashing
methods can be used to map similar images to nearby binary
Feature
Dimension
Gist512 (binary)
Gist2048 (binary)
fc6 (binary)
Gist512 (raw)
Gist2048 (raw)
fc6 (raw)
conv3 (raw)
512
2048
4096
512
2048
4096
43264
Size
(bits)
512
2048
2048
131072
524288
1048576
11075584
Size of entire Alderley data (MB)
2 Mb
8 Mb
8 Mb
493 Mb
1973 Mb
3946 Mb
41678 Mb
Table II: Comparing storage size of binary codes and their
corresponding raw/precursor features. To illustrate the impact on VPR, we also calculate how much space is required
to store a data set (here, Alderley Night/Day data set)
codes (with small hamming distance between them) and
map dissimilar images to far-away binary codes (with large
hamming distance between them). Two main advantages of
using hashing methods for visual place recognition are:
• Storage space: Table II compares the storage size of
real-valued features with that of binary code representations. Even though [10] applies hashing to obtain
compact 8192 bit binary codes, it can preserve 95% of
the accuracy obtained by using raw real-valued features.
Our supervised approach obtains even shorter binary
codes and simultaneously shows better performance
than that obtained by using raw real-valued features.
• Speed: The hamming distance computation required to
compare similarity between compact binary codes is
much faster than euclidean or cosine distance computations required to compare very high dimensional
feature vectors. Hamming distances can be efficiently
calculated using low level (hardware) bit-wise operations.
Hashing methods can be categorized into - unsupervised
and supervised, based on the notion of similarity that these
methods preserve. Unsupervised hashing methods preserve
distance based similarity whereas supervised methods preserve label based similarity.
A. Unsupervised hashing
Distance based similarity means that images are considered similar if their feature vectors have small distance
(euclidean, cosine etc.) between them. Unsupervised hashing
methods output binary codes which aim to preserve the
original distances in real-valued feature space. They leverage
no other information except the image features, therefore,
their performance is lower than original real-valued features.
Despite slightly lower performance, they help overcome two
of the issues with CNN features - dimensionality and storage
size. Locality sensitive hashing (LSH) [28], [29] is a widely
used unsupervised hashing method which preserves cosine
distance between original data points. Random hyperplane
adaptation of LSH in [30] is most commonly used. We too
80
learn the matrix W such that V = XW transforms feature
vectors (rows of X) to a more semantic real space. After
obtaining transformed representations vi ∈ Rk (ith row of
V = XW ) from the CCA step, we ought to quantize this
representation to obtain binary codes. This can be directly
done by using indicator function on each of the dimensions:
f (v) = 1R+ (v). However, a better binary embedding is
obtained by rotating the features obtained after the CCA step
in such a manner that the quantization loss is minimized.
Gong et al. [1] describe a Procustean approach to solve
this quantization problem by minimizing the following loss
function:
Gist 512 CCAITQ
Gist 512 MLH
Gist 512 LSH
Fc6 CCAITQ
Fc6 MLH
Fc6 LSH
70
60
Recall
50
40
30
20
10
arg minkf (V R) − V Rk s.t. R0 R = RR0 = I
R
32
64
128
256
Bits
Figure 2: Comparing hashing methods - CCAITQ [1], MLH
[2] and LSH [28].
use it in this work to replicate results of previous LSH-based
VPR approaches [10].
B. Supervised hashing
Supervised hashing leverage supervision of labels. Using
the knowledge from labels, their performance is higher
than original real-valued features. For the task of longterm visual place recognition, feature vectors of spatially
proximal places might not be nearby in the feature space. For
example, feature vector corresponding to winter image of a
place X might be closer to the feature vector corresponding
to summer image of place Y (rather than that of place X). In
such cases externally provided labels can support supervised
hashing, which defines a better notion of similarity. We
use the well known technique of Canonical Correlation
Analysis combined with Iterative Quantization (CCAITQ)
[1] to perform supervised hashing for robust VPR. Another
supervised hashing technique that was developed at the
same time is Minimal Loss Hashing (MLH) [2]. In our
comparison in section IV-A we observe that for the task
of VPR, CCAITQ performs better than MLH and is also
more computationally efficient. Hence, we use CCAITQ in
the proposed VPR method.
CCAITQ [1] is a linear hashing method which combines
Canonical Correlation Analysis (CCA) [31] and Iterative
Quantization (ITQ) techniques to eventually obtain binary
codes. CCA is the supervised analog of the PCA, which
learns a matrix W ∈ Rd×k where d is the dimensionality
of original features and k is the desired dimensionality
of output binary codes. This matrix helps in finding the
directions in which the feature vectors and the label vectors
have maximum correlation. The input feature matrix is
X = [x1 ; x2 ; ...xn ] ∈ Rn×d where n is the number of data
points and xi ’s are rows of features. The aim of CCA is to
By solving this minimization problem we obtain the
orthogonal matrix R which minimizes the quantization loss.
IV. E XPERIMENTATION
We perform experiments on four data sets chosen for the
variety of variation in them (details of data sets is given
in table III). Each data set has two traversals of the same
route - database and query traversal, with appropriate ground
truth match. In Nordland data set the frames of the database
and query traversal are synchronized i.e. ith winter frame
matches the ith summer frame, whereas the ground truth in
Alderley and St. Lucia data sets is provided externally using
a frame matching matrix (fm). More generally, fm(i) = j
stores that the ith training frame in query traversal corresponds to the same locations as j th training frame in
the database traversal. We use some portion of both the
traversals for training CCAITQ to learn the transformation
matrices W and R as described in section III-B.
For extracting gist feature descriptors we use the original
implementation1 of gist, made available by Oliva & Torralba.
For extracting deep CNN features we use the matconvnet
toolbox made available by VGG group.
We consider each training image of both traversals as a
label. Label vector (for CCAITQ) of a particular training
image i has 10 s for i± margin frames of the same traversal
and also fm(i) ± margin frames of the other traversal.
Thus, the obtained binary code of a given frame learns
similarity to neighbouring frames (variation in viewpoint)
and also learn similarity to corresponding frames in the other
traversal (variation in appearance). We employ the original
implementation of CCA-ITQ2 .
A. Comparing supervised hashing methods
The works of MLH [2] and CCAITQ [1] have performed well in comparison to other methods of supervised
hashing such as spectral hashing and binary reconstructive
embedding. We therefore restrict ourselves in this paper to
1 http://people.csail.mit.edu/torralba/code/spatialenvelope
2 http://slazebni.cs.illinois.edu/research/smallcode.zip
100
90
80
GIST 2048
GIST 512
CNN - Fc6 (pre ReLU)
Conv3 benchmark (43264D)
Supervised hashing (CCA-ITQ)
Euclidean baseline (2048D)
Unsupervised hashing (LSH)
Conv3 benchmark (43264D)
Supervised hashing (CCA-ITQ)
Euclidean baseline (512D)
Unsupervised hashing (LSH)
Conv3 benchmark (43264D)
Supervised hashing (CCA-ITQ)
Euclidean baseline (4096D)
Unsupervised hashing (LSH)
70
Recall
60
50
40
30
20
10
32
64
128
256
Bits
512
1024
2048
(a) CNN-Fc6 features
32
64
128
256
Bits
512
1024
(b) Gist 512D features
2048
32
64
128
256
Bits
512
1024
2048
(c) Gist 2048D features
Figure 3: Comparing effect of code length (in bits) on recall (in %). Our model can convert raw real valued features (green)
to binary codes (red) which boosts the performance. Binary codes obtained by popular unsupervised hashing method (LSH)
(blue) have lower performance than the corresponding raw real-valued features.
these two popularly used supervised hashing methods. Fig.
2 shows a comparison between the two hashing methods.
We also compare it to the baseline of unsupervised hashing
method LSH (used in recent VPR methods [10], [11]). We
find MLH is intractable for learning more than 256 bit
binary codes, requiring more than a day of training time
on our data sets on a 16Gb i7 computer. On the contrary,
CCAITQ requires only few minutes to train over 4096D
fc6 features for data sets of considerable size. Moreover,
it always outperforms MLH’s performance. Hence, we conduct experiments and report results using CCAITQ in our
supervised VPR pipeline.
B. Bit-wise study of supervised VPR
We compare the recall performance of binary codes for
different code lengths varying from 32 to 2048 bits for the
most challenging Nordland winter-summer data set. While
testing our VPR pipeline, we extract binary codes for each
test set image of query and database traversal (using learnt
W and R matrices). For every query image code we find the
closest database image code (in hamming space) and count
it a true positive if it is within the margin around ground
truth match. CCAITQ algorithm outputs binary codes with
lengths lesser than the dimensionality of the raw features.
Hence, 3b does not go beyond 512 bits (learnt from 512D
gist features). The black benchmark is calculated using
conv3 raw features, which is the best performing feature
as suggested in [10]. We compare conv3 features with
our VPR pipeline’s results, obtained from much smaller
features - simple gist and fc6 CNN features. We observed
that the learnt binary codes (red) perform significantly
better than the corresponding raw features (green). Fig. 3c
shows the VPR system with the proposed modification helps
outperforms conv3 raw features [10] by using only 2048
bit binary codes. Hence we are able to bootstrap simpleto-compute & low dimensional gist features to match the
performance of pre-trained CNN based VPR systems and
simultaneously reduce dimensionality & storage space.
C. Comparing precision-recall performance
Precision-Recall (PR) curves are used to compare image
retrieval techniques. VPR is similar to image retrieval and
research in VPR [4], [8], [10]–[12] plot PR curves by
altering a parameter. Achieving high precision with high
recall is desired. Hence, farther a PR plot is from the axes,
better the performance. The procedure for plotting PR curve
for a VPR system is described ahead. Each query image has
n = 2 ∗ margin + 1 ground truth positives. We retrieve top
k matches for a query image out of which m (≤ k) are true
positives. We use P recision = m/k and Recall = m/n,
which is averaged over all query images to make the PR
plot. The value of k is varied from 1 to the total number
of test images to obtain multiple points on the curve. We
extract different top matches for each VPR method and
compare the performance in fig. 4. We observe that the
recognition performance of the binary codes (coloured solid
lines) learnt over all three features - 512D gist, 2048D
gist and 4096D fc6, improve over their corresponding raw
feature versions (coloured dashed lines). Moreover, 2048 bit
binary codes learnt over 2048D gist features shows better
1
0.7
0.6
0.5
0.8
0.7
0.4
0.5
0.4
0.3
0.2
0.2
0.1
0.1
0
0.1
0.2
0.3
0.4
0.5
Recall
0.6
0.7
0.8
0.9
0
1
(a) Nordland summer-spring data set (A: Mild, V: No)
0.5
0.3
0.25
0.2
0.3
0.4
0.5
Recall
0.6
0.7
0.8
0.9
1
St. Lucia 01-10 data set
Conv3 benchmark (43264D)
Gist 2048 CCAITQ (2048 bit)
Gist 2048D raw feature
Gist 512 CCAITQ (512 bit)
Gist 512D raw feature
Fc6 CCAITQ (2048 bit)
Fc6 4096 raw feature
0.7
0.6
0.5
0.2
0.15
0.4
0.3
0.2
0.1
0.1
0.05
0
0.1
(b) Nordland summer-winter data set (A: Severe, V: None)
Precision
0.4
0.35
0
0.8
Alderley night-day data set
Gist 512 CCAITQ (512 bit)
Gist 2048 CCAITQ (2048 bit)
Fc6 CCAITQ (2048 bit)
Gist 512D raw feature
Gist 2048D raw feature
Fc6 raw feature (4096D)
Conv3 benchmark (43264D)
0.45
Precision
0.6
0.3
0
Nordland Winter-Summer data set
Conv3 benchmark (43264D)
Gist 2048 CCAITQ (2048 bit)
Gist 2048 raw feature
Gist 512 CCAITQ (512 bit)
Gist 512 raw feature
Fc6 CCAITQ (2048 bit)
Fc6 raw feature
0.9
Precision
0.8
Precision
1
Nordland Summer-Spring data set
Conv3 benchmark (43264D)
Gist 2048 CCAITQ (2048 bit)
Gist 2048D raw feature
Gist 512 CCAITQ (512 bit)
Gist 512D raw feature
Fc6 CCAITQ (2048 bit)
Fc6 4096D raw feature
0.9
0
0.1
0.2
0.3
0.4
0.5
Recall
0.6
0.7
0.8
0.9
1
(c) Alderley Night-Day data set (A: Severe, V: Mild)
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Recall
(d) St. Lucia 01-10 data set (A: Mild, V: Severe)
Figure 4: Precision-Recall curves of data sets exhibiting difference mixture of appearance (A) and viewpoint (V) variations
(or marginally less) performance than the benchmarking
raw conv3 features (solid black).
D. Leveraging continuity information using contrast enhanced Dynamic Time Warping
While most VPR methods are similar to an image retrieval
for each query image, some methods like [4], [32] leverage
the fact that we always have a sequence of images as
opposed to a single query image. The common assumption
being that the two traversals have no negative velocities
(reverse/backward travel). Authors of [4] suggest using
Dynamic Time Warping (DTW) [33] in future work to tackle
variations in velocities. We apply DTW over our binary
codes (red dashed) to show an improvement in results. The
results are compiled in fig. 5 where the individual imagelevel VPR method (black circles) explained till now gives
slightly incongruous (non continuous) retrievals. Cost of a
path has two factors - number of elements (length of path)
and cost of each element (divergence from ground truth).
We contrast enhanced the DTW distance matrix, thus, giving
more weight to the factor of divergence from ground truth.
Hence, we found such a cost function (blue dots) to give
better performance on non-diagonal trajectories (zoomed in
plots in fig. 5).
E. Implementation details
Additional details of implementation needing explanation
are as follows:
• CCAITQ is applied over pre-ReLU layer features instead of post-ReLU layer features: Rectified linear unit
(ReLU) layer: f (x) = max(x, 0) add the non-linearity
quotient to deep CNNs. Since [10], [11] directly utilize
pre-trained CNN features without doing any learning,
they may choose to use features after ReLU layers
(post-ReLU) which are much sparser than pre-ReLU.
Studies in [20] have shown that SVM learning over
CNN features works better when we use pre-ReLU
features. Since our task of supervised hashing method
is equivalent to learning multiple (equal to the length of
binary codes) hyperplanes, we too use pre-ReLU CNN
features in our experiments. For features requiring no
learning, we report results using post-ReLU (same as
[10], [11]).
• Use of VGG-f: VGG-f [21] has five convolutional, three
fully connected layers and takes a 224 × 224 image (all
same as AlexNet [18]). While the old AlexNet architecture is configured rigidly to accommodate training
distributed over two GPUs, VGG-f gets rid of such
dependency. With more minor variations, it is shown
Data set
Database
traversal
Query
traversal
Nordland
Summer
Spring
Nordland
Summer
Winter
Alderley
Night
Day
1st
10th
traversal
traversal
St. Lucia
Variation
present
A: Mild,
V: None
A: Severe,
V: None
A: Severe,
V: Mild
A: Mild,
V: Severe
Train frames
Test frames
1 − 10k
(both traversals)
1 − 10k
(both traversals)
1 − 9k (day)
1 − 11301 (night)
1 − 10k (10)
1 − 11145 (01)
11k − 16k
(both traversals)
11k − 16k
(both traversals)
10k − 14607 (day)
12051 − 16960 (night)
11001 − 16k (10)
12145 − 17512 (01)
Sample
rate (fps)
Margin
(± frames)
2
5
2
5
25
10
15
10
Table III: Experimental details and variations present in data sets used for evaluation
18000
hashing to the Visual Place Recognition research. The
application is a perfect fit as there is a need for both
learning and compact representation for any VPR to be
robust and real-time, respectively. There is the alternative to
learn these compact embeddings by training a CNN for this
task. Training of CNNs require a long time and also require
much complex hardware capable systems. We conclude that
for a widespread application of VPR, supervised hashing is
ideal as it is both quick to train and outputs compact binary
representation of images.
Ground truth
CCAITQ
16000
CCAITQ + DTW
CCAITQ +
enchanced DTW
Alderley night traversal frame number
14000
12000
10000
8000
ACKNOWLEDGMENT
6000
The authors would like to thank Ford Motor Company for supporting this research through the project
FMT/EE/2015241
4000
R EFERENCES
2000
0
0
5000
10000
15000
Alderley day travesal frame number
Figure 5: Impact of leveraging sequencing or continuity
information for Alderley data set.
[1] Y. Gong, S. Lazebnik, A. Gordo, and F. Perronnin, “Iterative
quantization: A procrustean approach to learning binary codes
for large-scale image retrieval,” IEEE Transactions on Pattern
Analysis and Machine Intelligence, vol. 35, no. 12, pp. 2916–
2929, Dec 2013.
[2] M. Norouzi and D. J. Fleet, “Minimal loss hashing for
compact binary codes,” in ICML, 2011, pp. 353–360.
•
to give better performance than AlexNet on recognition
tasks. Hence, we report results using this variant and
achieve better performance for both raw features as well
as binary codes.
Supervised hashing not applied on conv3 layer:
Whether we choose VGG-f or AlexNet variant of the
deep CNN, the dimensionality of conv3 layer is too
high to tractably run CCAITQ (MLH is even slower).
Both aims of our VPR pipeline - Speed and storage
efficiency (Sec. III), are incongruent to the use of
conv3 layer features. Supervised hashed codes that we
obtain from gist or fc6 CNN features are compared to
the raw conv3 layer (black plots in fig. 3 and 4).
V. C ONCLUSION
To the best of our knowledge, our work is the first to
introduce the successful learning methods of supervised
[3] M. Cummins and P. Newman, “Fab-map: Probabilistic localization and mapping in the space of appearance,” The
International Journal of Robotics Research, vol. 27, no. 6,
pp. 647–665, 2008.
[4] M. J. Milford and G. F. Wyeth, “Seqslam: Visual route-based
navigation for sunny summer days and stormy winter nights,”
in Robotics and Automation (ICRA), 2012 IEEE International
Conference on, May 2012, pp. 1643–1649.
[5] N. Sünderhauf and P. Protzel, “Brief-gist-closing the loop
by simple means,” in Intelligent Robots and Systems (IROS),
2011 IEEE/RSJ International Conference on. IEEE, 2011,
pp. 1234–1241.
[6] S. M. Lowry, M. J. Milford, and G. F. Wyeth, “Transforming
morning to afternoon using linear regression techniques,” in
Robotics and Automation (ICRA), 2014 IEEE International
Conference on. IEEE, 2014, pp. 3950–3955.
[7] P. Neubert, N. Sunderhauf, and P. Protzel, “Appearance
change prediction for long-term navigation across seasons,”
in Mobile Robots (ECMR), 2013 European Conference on.
IEEE, 2013, pp. 198–203.
[8] Z. Chen, O. Lam, A. Jacobson, and M. Milford, “Convolutional neural network-based place recognition,” arXiv preprint
arXiv:1411.1509, 2014.
[9] P. Neubert and P. Protzel, “Local region detector + cnn
based landmarks for practical place recognition in changing
environments,” in Mobile Robots (ECMR), 2015 European
Conference on, Sept 2015, pp. 1–6.
[10] N. Sünderhauf, F. Dayoub, S. Shirazi, B. Upcroft, and M. Milford, “On the performance of convnet features for place
recognition,” CoRR, vol. abs/1501.04158, 2015.
[11] N. Suenderhauf, S. Shirazi, A. Jacobson, F. Dayoub, E. Pepperell, B. Upcroft, and M. Milford, “Place recognition
with convnet landmarks: Viewpoint-robust, condition-robust,
training-free,” in Proceedings of Robotics: Science and Systems, Rome, Italy, July 2015.
[12] Z. Chen, S. Lowry, A. Jacobson, Z. Ge, and M. Milford, “Distance metric learning for feature-agnostic place recognition,”
in Intelligent Robots and Systems (IROS), 2015 IEEE/RSJ
International Conference on, Sept 2015, pp. 2556–2563.
[13] A. Oliva and A. Torralba, “Modeling the shape of the scene: A
holistic representation of the spatial envelope,” International
journal of computer vision, vol. 42, no. 3, pp. 145–175, 2001.
[14] C. Siagian and L. Itti, “Rapid biologically-inspired scene classification using features shared with visual attention,” Pattern
Analysis and Machine Intelligence, IEEE Transactions on,
vol. 29, no. 2, pp. 300–312, 2007.
[15] A. C. Murillo and J. Kosecka, “Experiments in place recognition using gist panoramas,” in Computer Vision Workshops
(ICCV Workshops), 2009 IEEE 12th International Conference
on. IEEE, 2009, pp. 2196–2203.
[16] Y. Liu and H. Zhang, “Visual loop closure detection with a
compact image descriptor,” in Intelligent Robots and Systems
(IROS), 2012 IEEE/RSJ International Conference on. IEEE,
2012, pp. 1051–1056.
[17] G. Singh and J. Kosecka, “Visual loop closing using gist
descriptors in manhattan world,” in ICRA Omnidirectional
Vision Workshop. Citeseer, 2010.
[18] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet
classification with deep convolutional neural networks,” in
Advances in Neural Information Processing Systems 25,
F. Pereira, C. J. C. Burges, L. Bottou, and K. Q. Weinberger,
Eds. Curran Associates, Inc., 2012, pp. 1097–1105.
[19] J. Donahue, Y. Jia, O. Vinyals, J. Hoffman, N. Zhang,
E. Tzeng, and T. Darrell, “Decaf: A deep convolutional activation feature for generic visual recognition,” arXiv preprint
arXiv:1310.1531, 2013.
[20] A. Sharif Razavian, H. Azizpour, J. Sullivan, and S. Carlsson, “Cnn features off-the-shelf: An astounding baseline for
recognition,” in The IEEE Conference on Computer Vision
and Pattern Recognition (CVPR) Workshops, June 2014.
[21] K. Chatfield, K. Simonyan, A. Vedaldi, and A. Zisserman,
“Return of the devil in the details: Delving deep into convolutional nets,” in British Machine Vision Conference, 2014.
[22] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh,
S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein,
A. C. Berg, and L. Fei-Fei, “ImageNet Large Scale Visual
Recognition Challenge,” International Journal of Computer
Vision (IJCV), vol. 115, no. 3, pp. 211–252, 2015.
[23] N. B. Corporation, “Nordlandsbanen: minute by
minute, season by season,” https://nrkbeta.no/2013/01/
15/nordlandsbanen-minute-by-minute-season-by-season/.
[24] M. Warren, D. McKinnon, H. He, and B. Upcroft, “Unaided
stereo vision based pose estimation,” in Australasian Conference on Robotics and Automation, G. Wyeth and B. Upcroft,
Eds. Brisbane: Australian Robotics and Automation Association, 2010.
[25] A. Geiger, P. Lenz, C. Stiller, and R. Urtasun, “Vision meets
robotics: The kitti dataset,” International Journal of Robotics
Research (IJRR), 2013.
[26] A. Torii, J. Sivic, T. Pajdla, and M. Okutomi, “Visual place
recognition with repetitive structures,” in CVPR, 2013.
[27] A. Torii, R. Arandjelović, J. Sivic, M. Okutomi, and T. Pajdla,
“24/7 place recognition by view synthesis,” in CVPR, 2015.
[28] P. Indyk and R. Motwani, “Approximate nearest neighbors:
towards removing the curse of dimensionality,” in Proceedings of the thirtieth annual ACM symposium on Theory of
computing. ACM, 1998, pp. 604–613.
[29] A. Gionis, P. Indyk, R. Motwani et al., “Similarity search in
high dimensions via hashing,” in VLDB, vol. 99, no. 6, 1999,
pp. 518–529.
[30] M. S. Charikar, “Similarity estimation techniques from rounding algorithms,” in Proceedings of the thiry-fourth annual
ACM symposium on Theory of computing. ACM, 2002, pp.
380–388.
[31] H. Hotelling, “Relations between two sets of variates,”
Biometrika, vol. 28, no. 3/4, pp. 321–377, 1936.
[32] E. Pepperell, P. I. Corke, and M. J. Milford, “All-environment
visual place recognition with smart,” in 2014 IEEE International Conference on Robotics and Automation (ICRA).
IEEE, 2014, pp. 1612–1618.
[33] H. Sakoe and S. Chiba, “Dynamic programming algorithm
optimization for spoken word recognition,” IEEE transactions
on acoustics, speech, and signal processing, vol. 26, no. 1,
pp. 43–49, 1978.
| 1cs.CV |
arXiv:1706.01294v1 [math.CO] 5 Jun 2017
THE SMITH GROUP AND THE CRITICAL GROUP
OF THE GRASSMANN GRAPH OF LINES IN FINITE
PROJECTIVE SPACE AND OF ITS COMPLEMENT
JOSHUA E. DUCEY AND PETER SIN
Abstract. We compute the elementary divisors of the adjacency
and Laplacian matrices of the Grassmann graph on 2-dimensional
subspaces in a finite vector space. We also compute the corresponding invariants of the complementary graphs.
1. Introduction
The Smith group of a graph is the abelian group defined by using
an adjacency matrix of the graph as a relation matrix. The name
comes from the fact that a cyclic decomposition of this group is given
by the Smith normal form of the adjacency matrix. If we use the
Laplacian matrix of the graph as a relation matrix then the finite part
of the abelian group defined is called the the critical group. By the
well-known Matrix Tree Theorem of Kirchhoff, the order of the critical
group of a connected graph is equal to the number of spanning trees in
the graph. This group also arises in chip-firing games on graphs and
in the closely-related sandpile model in physics, where it is also known
as the sandpile group. It is therefore of some interest to compute the
Smith groups and critical groups for individual graphs or for families of
graphs. A particular class that has proved amenable to computations
is the class of strongly regular graphs (e.g. [4]). In this paper we treat
these questions for the Grassmann graph defined by the incidence of
lines in a finite projective space PG(n − 1, q), as well as its complementary graph, the skew lines graph. As well as being strongly regular
2010 Mathematics Subject Classification. 05C50.
Key words and phrases. invariant factors, elementary divisors, Smith normal
form, critical group, Jacobian group, sandpile group, adjacency matrix, Laplacian,
Grassmann.
This work was partially supported by a grant from the Simons Foundation
(#204181 to Peter Sin).
Research was (partially) completed while the second author was visiting the
Institute for Mathematical Sciences, National University of Singapore in 2016.
1
2
DUCEY AND SIN
these graphs admit the action of the general linear group of the projective space, which allows the application of extensive machinery from
representation theory. The problem of computing the Smith normal
forms or, equivalently, the elementary divisors of the adjacency and
Laplacian matrices, falls naturally into two parts, one for the characteristic p of the projective space and the second for all other primes.
In §3, we establish some general lemmas needed in both parts. The
charactersistic p case was handled in projective dimension 3 in an earlier paper [1], by making use of the computations of the p-elementary
divisors of point-subspace incidence matrices in [3]. Fortunately, it is
possible to extend the method to higher dimensions without great difficulty, and this is done in §4. We then turn to the cross-characteristic
case. Here we rely heavily on James’s theory [6] of unipotent representations of GL(n, q), over a field of characteristic ℓ 6= p. From general
considerations we know that the multiplicity of a given power ℓa as
an elementary divisor is the dimension of some GL(n, q)-subquotient
of the ℓ-modular permutation module on lines, and detailed information about the structure of the submodule lattice of this permutation
module allows us to determine all such multiplicities. In §5 we collect
together the required results from the ℓ-modular representation theory
of GL(n, q). The detailed computations are then made in the following
four sections, treating in turn the critical group and the Smith group
of the Grassmann graph, followed by the critical group and the Smith
group of the skew lines graph. Finally, we include an appendix displaying all the possible submodule structures of the ℓ-modular permutation
module, corresponding to various arithmetic conditions satisfied by ℓ
and n. Although we never use these informal pictures in our proofs,
we have found them very helpful in navigating through the many cases
into which the computations are divided.
2. Parameters of the graphs
Let V be an n-dimensional vector space (n ≥ 4) over a finite field
Fq of q = pt elements. Consider the graph with vertex set the 2dimensional subspaces of V (or 2-spaces for short), where a pair of
2-spaces are adjacent when their intersection is trivial. This is the
skew-lines graph of the associated projective space, which we denote by
Γ. Under some ordering of the vertices, we let A denote the adjacency
matrix.
The complement of this graph is the Grassmann graph Γ′ on 2-spaces,
and we denote by A′ its adjacency matrix.
GRASSMANN GRAPHS
3
These graphs are strongly regular [2, Chapter 8]. The parameters
v ′ , k ′ , λ′ , µ′ of Γ′ are given below. We use the q-binomial coefficients
[m
k ]q
(q m − 1)(q m−1 − 1) · · · (q m−k+1 − 1)
=
(q k − 1)(q k−1 − 1) · · · (q − 1)
which count the number of k-dimensional subspaces in an m-dimensional
vector space over Fq .
v ′ = [ n2 ]q
k ′ = q(q + 1)[ n−2
1 ]q
2
λ′ = [ n−1
1 ]q + q − 2
µ′ = (q + 1)2 .
Let v, k, λ, µ denote the parameters of the skew-lines graph Γ. Since
this graph is complementary to Γ′ we have
v = v′
k = v − k ′ − 1 = q 4 [ n−2
2 ]q
λ = v − 2k ′ + µ′ − 2 = 2k + µ − v
µ = v − 2k ′ + λ′ .
We also record that A has spectrum
2 n−3
q 4 [ n−2
2 ]q , −q [ 1 ]q , q
with respective multiplicities 1, [ n1 ]q − 1, [ n2 ]q − [ n1 ]q .
Let L = kI − A be the Laplacian matrix of Γ and L′ = k ′ I − A′ that
of Γ′ . We note that A and A′ are invertible matrices, while L and L′
have corank 1. We shall make use of relations satisfied by A, L, A′ and
L′ that follow from the strong regularity of Γ and Γ′ .
These are well known [2, Theorem 8.1.2].
Lemma 2.1. Let à be the adjacency matrix of a strongly regular graph
with parameters (ṽ, k̃, λ̃, µ̃) and restricted eigenvalues r̃ and s̃, and let
L̃ = k̃I − Ã be the Laplacian matrix. Then
(i) Ã2 + (µ̃ − λ̃)Ã + (µ̃ − k̃)I = µ̃J, where J is the matrix all of
whose entries are 1.
(ii) (Ã − r̃I)(Ã − s̃I) = µ̃J, and (L̃ − (k̃ − r̃)I)(L̃ − (k̃ − s̃)I) = µ̃J.
3. General results on modules
Let (R, (π)) be a local PID with residue field R/(π) = F , field of
quotients K, and let M a finitely generated free R-module. Let M =
4
DUCEY AND SIN
M/πM ∼
= F ⊗R M, and for any submodule U of M, we denote by U
its image in M . Let φ ∈ EndR (M), and for i ≥ 0 let
(3.1)
Mi = Mi (φ) = {n ∈ M|φ(n) ∈ π i M}.
Then we have
(3.2)
M = M0 ⊇ M1 ⊇ · · · ⊇ Ker(φ),
and
(3.3)
M = M 0 ⊇ M 1 ⊇ · · · ⊇ Ker(φ).
Then ei = ei (φ) := dimF (M i /M i+1 ) is the multiplicity of π i as an
elementary divisor of φ. (e0 is the rank of φ.) We will often use the
formula
X
(3.4)
dimF M i = dimF Ker(φ) +
ej .
j≥i
Note that Ker(φ) is not, in general, equal to the kernel of the induced
map φ ∈ EndF (M ). Indeed, Ker(φ) = M 1 .
One can also define an ascending chain of submodules of M:
Ni = Ni (φ) = {π −i φ(n)|n ∈ Mi }.
A formula similar to the above holds, namely,
(3.5)
dimF N i =
i
X
ej .
j=0
Lemma 3.1. Let M, and φ be as above. Let d be the π-adic valuation
of the product of the nonzero elementary divisors of φ, counted with
multiplicities. Suppose that in (3.3) we have an increasing sequence of
indices 0 < a1 < a2 < · · · < ah and a corresponding sequence of lower
bounds b1 > b2 > · · · > bh satisfying the following conditions.
(1) dimF M aj ≥ bj for j = 1,. . . , h.
P
(2) hj=1 (bj − bj+1 )aj = d, where we set bh+1 = dimF Ker(φ).
Then the following hold.
(1) eaj (φ) = bj − bj+1 for j = 1,. . . , h.
(2) e0 (φ) = dimF M − b1 .
(3) ei (φ) = 0 for i ∈
/ {0, a1 , . . . , ah }.
GRASSMANN GRAPHS
Proof. We have
(3.6) X
d=
iei
5
i≥0
≥
h−1
X
j=1
≥
h−1
X
j=1
=
h−1
X
X
aj ≤i<aj+1
aj
iei +
X
aj ≤i<aj+1
X
iei
i≥ah
ei + ah
X
ei
i≥ah
aj (dimF M aj − dimF M aj+1 ) + ah (dimF M ah − dimF Ker(φ))
j=1
= a1 dimF M a1 +
h
X
(aj − aj−1 ) dimF M aj − ah dimF Ker(φ)
j=2
≥ a1 b1 +
h
X
(aj − aj−1 )bj − ah bh+1
j=2
=
h
X
(bj − bj+1 )aj = d.
j=1
Therefore we must have equality throughout. The equality of the first
three lines implies that ei = 0 for i ∈
/ {0, a1 , . . . , ah }, proving (3). It
follows that for j = 1,. . . , h − 1 we have eaj = dimF M aj − dimF M aj+1 ,
and eah = dimF M ah − dimF Ker(φ). The equality of the second and
third last rows then shows that inequalities dimF M aj ≥ bj are all
equalities, for j = 1,. . . , h, so eaj = (bj − bj+1 ) and (1) is proved.
Finally, e0 (φ) is the rank of the induced endomorphism φ ∈ EndF (M),
whose kernel is M 1 by definition, and our discussion shows that M 1 =
M a1 , proving(2).
Let MK = K ⊗R M. The following lemma is well known (cf. [2,
Proposition 12.8.3]).
Lemma 3.2. Let φ ∈ EndR (M). Suppose the induced endomorphism
φK of MK has an eigenvalue u ∈ R. Let d be the dimension of the
eigenspace
(MK )u = {m ∈ MK | φK (m) = um}.
Then d ≤ dimF M b , for any b with π b | u.
6
DUCEY AND SIN
Proof. The submodule M ∩ (MK )u is pure ([5, p.84]), has rank equal
to d and clearly lies in Mb . It follows that d = dimF M ∩ (MK )u ≤
dimF M b .
Assume that M is an RG-module for a group G and φ an RG-module
homomorphism. Then we can strengthen the conclusion of Lemma 3.2
to a statement about the F G-composition factors of Mb .
Lemma 3.3. With the notation above and Lemma 3.2, let E be a KGsubmodule of (MK )u . Then the following hold.
(i) M ∩ E is a pure submodule of M such that K ⊗R (M ∩ E) ∼
= E.
∼
(ii) If X be any R-free RG-module with K ⊗R X = E, then X/πX
and M ∩ E have the same F G-composition factors, counting
multiplicities.
(iii) For any b with π b | u, the F G-composition factors of Mb include
those of X/πX, counting multiplicities.
Proof. Part (i) is immediate from the definitions of pure submodule
([5, p.84]) and M ∩ E. By (i), the image M ∩ E of M ∩ E in M is
isomorphic to (M ∩ E)/π(M ∩ E), and this has the same composition
factors as X/πX since X and M ∩E are both R-forms of E, by a general
principle [5, Proposition 16.16] in modular representation theory. Thus
(ii) holds. To see (iii), we note that from the hypothesis on b we have
M ∩ E ⊆ M b , and the statement about composition factors holds by
(ii).
4. The p-elementary divisors
4.1. Skew lines, A.
By Lemma 2.1, we have
(4.1)
A2 = kI + λA + µ(J − A − I).
Rewriting, we get
(4.2)
A[A − (λ − µ)I] = (k − µ)I + µJ,
where one checks that
λ − µ = −q n−2 − q n−3 − · · · − q 2 + q
= −q(q n−3 + q n−4 + · · · + q − 1)
and
k − µ = q n−1 + q n−2 + · · · + q 3
= q 3 (q n−4 + q n−5 + · · · + 1).
GRASSMANN GRAPHS
7
We can use Equation 4.2 to learn much about the p-elementary divisors of A. We may work over Zp , the ring of p-adic integers. Let L2
denote our vertex set of 2-spaces. The matrix A defines a Zp -module
homomorphism:
A : M → M,
where M = ZLp 2 consists of all formal Zp -linear combinations of the
elements of L2 , and the map A sends a 2-space to the formal sum of
the 2-spaces that intersect it trivially. Since |L2 | is a unit in Zp , we
have
M = Zp 1 ⊕ Y,
where Y is the kernel of the augmentation map ǫ : M → Zp sending all
2-spaces to 1, and 1 is the formal sum of all of the 2-spaces. The map
A respects this decomposition. Note that
(4.3)
A(1) = k1 = q 4 [ n−2
2 ]q ,
so p4t occurs at least once as an elementary divisor of A. We will see
shortly that it occurs exactly once. Now we focus our attention on Y
and the restriction map A|Y : Y → Y .
The Equation 4.2 now becomes
(4.4)
A|Y [A|Y − qz1 I] = q 3 z2 I,
with z1 , z2 being units in Zp . It follows immediately that A|Y has no
elementary divisors pi for i > 3t (and so p4t does indeed have multiplicity 1 as an elementary divisor of A). It also follows from Equation 4.4
that the multiplicity of pi as an elementary divisor of A|Y is equal to
the multiplicity of p3t−i as an elementary divisor of A|Y − qz1 I. Since
A|Y and A|Y − qz1 I are congruent modulo pt , they have the same multiplicities for elementary divisors pi with 0 ≤ i < t. Hence, pi and
p3t−i occur as elementary divisors for A with the same multiplicity, for
0 ≤ i < t.
Now we use the eigenspaces of A to determine exactly which pelementary divisor multiplicities are nonzero. From our argument above
we again only need consider A|Y and those pi where i ≤ 3t. It will turn
out that only those elementary divisors pi satisfying 0 ≤ i ≤ t or
2t ≤ i ≤ 3t will have a nonzero multiplicity. Write Qp for the field
of quotients of Zp . Denote by V−q2 the Qp -eigenspace for A for the
eigenvalue −q 2 [ n−3
1 ]q , and denote by Vq the eigenspace for q. We have
the inclusions
V−q2 ∩ M ⊆ M2t (A)
Vq ∩ M ⊆ Nt (A).
8
DUCEY AND SIN
Recall that ei denotes the multiplicity of pi as an elementary divisor of
A. We can apply Lemma 3.2 to get the inequalities
(4.5)
[ n1 ]q − 1 ≤ e2t + · · · + e3t
(4.6)
[ n2 ]q − [ n1 ]q ≤ e0 + · · · + et .
Since the left sides of the inequalities 4.5 sum to the dimension of Y ,
both must in fact be equalities. Thus we see that, aside from one
occurrence of the elementary divisor p4t , the only other p-elementary
divisors of A that occur are those pi satisfying 0 ≤ i ≤ t or 2t ≤ i ≤ 3t.
Since we have equality throughout in 4.5, we see that we only need to
determine the multiplicities of the elementary divisors pi where 0 ≤
i < t in order to know them all.
In general, if Ar,s denotes the zero-intersection incidence matrix between r-spaces and s-spaces, then we have that
−Ar,s ≡ Ar,1 A1,s
(mod pt )
and so the multiplicities of pi as elementary divisors agree for Ar,s and
Ar,1 A1,s when 0 ≤ i < t. In [1] the p-elementary divisors of the matrix
Ar,1 A1,s were computed. Our matrix A is just A2,2 , so we want to look
at A2,1 A1,2 . Before we can state the theorem, we need some notation.
Let [n − 1]t denote the set of t-tuples with entries from {1, 2, . . . , n −
1}. For nonnegative integers α, β, define the following subsets of [n−1]t :
)
(
t−1
X
max{0, 2 − si } = α
Hα = (s0 , . . . , st−1 ) ∈ [n − 1]t |
i=0
and
= {(n − s0 , . . . , n − st−1 )|(s0 , . . . , st−1 ) ∈ Hβ } .
To each tuple ~s ∈ [n − 1]t we will associate a number d(~s) as follows.
First define the integer tuple ~λ = (λ1 , . . . , λt−1 ) by
βH
λi = psi+1 − si
where the subscripts are read modulo t. For an integer k, set dk to
be the coefficient of xk in the expansion of (1 + x + · · · + xp−1 )n , or,
explicitly,
⌊k/p⌋
X
n + k − jp − 1
j n
.
dk =
(−1)
n−1
j
j=0
Q
Finally, we set d(~s) = t−1
i=0 dλi .
Theorem 4.1 ([1], Theorem 5.1 and Remark following its proof).
Let ei denote the multiplicity of pi as a p-adic elementary divisor of
A2,1 A1,2 .
GRASSMANN GRAPHS
9
(1) e4t = 1.
(2) For i 6= 4t,
ei =
X
d(~s),
~s∈Γ(i)
where
Γ(i) =
[
βH
∩ Hα .
α+β=i
0≤α≤t
0≤β≤t
Summation over an empty set is interpreted to result in 0.
4.2. Skew lines, L.
The Laplacian matrix L of the skew-lines graph is defined as
L = kI − A = q 4 [ n−2
2 ]q I − A.
Thus L ≡ −A (mod p4t ), and so the p-elementary divisor multiplicities
are the same for A and L, for prime powers pi with 0 ≤ i < 4t. This
actually accounts for all of the p-elementary divisors of L; this can be
read off from the order of the critical group which we now compute.
Since the skew-lines graph is regular, the spectrum of L easily follows
from that of A. For L we have eigenvalues
2 n−3
4 n−2
0, q 4 [ n−2
2 ]q + q [ 1 ]q , q [ 2 ]q − q
with respective multiplicities 1, [ n1 ]q −1, [ n2 ]q −[ n1 ]q . By the Matrix Tree
Theorem, the order of the critical group is the product of the nonzero
eigenvalues of L divided by the number of vertices. We have
|K(Γ)| =
[ n1 ] −1
[ n2 ] −[ n1 ]
q
q
q
4 n−2
2 n−3
4 n−2
q [ 2 ]q + q [ 1 ]q
· q [ 2 ]q − q
[ n2 ]q
,
where K(Γ) denotes the critical group of Γ. Let S(Γ) denote the Smith
group of Γ, and compare the above order with
[ n1 ] −1
[ n2 ] −[ n1 ]
q
q
q
4 n−2
2 n−3
·
q
[
]
−
q
.
|S(Γ)| = q 4 [ n−2
]
·
−q
[
]
2 q
2 q
1 q
Looking carefully, one sees that the p-parts of the two orders differ
only by an extra factor of p4t in the order of the Smith group of Γ.
This is accounted for by the image of 1 under A, whereas L(1) = 0.
In other words, the p-elementary divisors of L and A agree on Y and
these account for all of the p-elementary divisors of L.
10
DUCEY AND SIN
4.3. Grassmann, A′ .
The spectrum of A′ is
2 n−3
q(q + 1)[ n−2
1 ]q , q [ 1 ]q − 1, −q − 1
with respective multiplicities 1, [ n1 ]q − 1, [ n2 ]q − [ n1 ]q . Thus we see that
there is only one p-elementary divisor of A′ , and that is determined by
the image of 1 under A′ . This image is exactly divisible by pt , and so
the p-Sylow subgroup of S(A′ ) is cyclic of order pt .
4.4. Grassmann, L′ .
The spectrum of L′ is
n−2
n−2
0, −q 2 [ n−3
1 ]q + 1 + q(q + 1)[ 1 ]q , q + 1 + q(q + 1)[ 1 ]q
with respective multiplicities 1, [ n1 ]q − 1, [ n2 ]q − [ n1 ]q . By the Matrix
Tree Theorem the order of the critical group of Γ′ depends only on the
nonzero eigenvalues. Since none of these are divisible by p, we see that
L′ has no p-elementary divisors.
5. The ℓ-elementary divisors for ℓ 6= p
Let ℓ be a prime different from p. Let R be the integral extension
of the ℓ-adic integers obtained by adding a primitive p-th root of unity
and let K be its field of quotients. The prime ℓ is unramified in R and
we set F = R/ℓR. Let Li be the set of i-dimensional subspaces of V
and let M = RL2 be the free R-module with basis L2 , which has the
structure of an R GL(n, q)-permutation module. Let Y be the kernel
of the homomorphism M → R sending each 2-space to 1, namely the
map given by the matrix J in the standard basis of M. As before, we
shall identify the matrices A, L, A′ and L′ with the endomorphisms of
M that they define with respect to the standard basis.
We have an A-invariant decomposition
M = R1 ⊕ Y
if and only if ℓ ∤ [ n2 ]q .
On extending scalars to K, we have
(5.1)
MK = K ⊕ S̃1 ⊕ S̃2 ,
where S̃1 and S̃2 are simple K GL(n, q)-modules (Specht modules over
K). The matrices A, L, A′ and L′ induce endomorphisms of MK , and
the decomposition above is the decomposition in to eigenspaces of these
endomorphisms. The submodule S̃1 affords the “r-eigenvalues” rA , rL ,
rA′ and rL′ , while S̃2 affords the “s-eigenvalues” sA , sL , sA′ and sL′ . We
fix the notation f = dimK S̃1 = [ n1 ]q −1 and g = dimK S̃2 = [ n2 ]q −[ n1 ]q .
GRASSMANN GRAPHS
11
In this section we shall apply the results of §3 to the local P ID
(R, (ℓ)) and the R GL(n, q)-modules M and Y , to to determine the
ℓ-elementary divisors of the matrices A, L, A′ and L′ . We shall also
need to consider the R GL(n, q)-permutation modules RLi on the set
of i-dimensional subspaces of V , for i = 0, 1, 2 and 3, as well as the
corresponding permutation modules over F and K. These modules
have been studied in depth by James [6], to which we shall refer for the
definition and basic properties of the important class of related modules
called Specht modules, which are analogous to the better known Specht
modules for the symmetric group. Lemma 3.3 will be applied to M or
Y , with G = GL(n, q), E = S̃i , for i = 1 or 2, and X equal to the
corresponding Specht module over R. In this case, X/πX ∼
= Si , the
Specht module over F . By the theory of Specht modules, we also have
f = dimF S1 and g = dimF S2 .
James [6, Theorem 13.3] has shown that F Li has a a descending
filtration whose subquotients are isomorphic to the Specht modules Sj ,
for 0 ≤ j ≤ i, each with multiplicity 1. Moreover, the only simple
modules that can be composition factors of the Sj are certain simple
modules Dk , k ≤ j. We assume n ≥ 4 to avoid trivialities. Then the
modules Dk , k = 0, 1, 2 are nonisomorphic. We have D0 = S0 ∼
= F,
the trivial module. The structure of S1 is also easy to describe; if
ℓ ∤ [ n1 ]q then S1 = D1 , while if ℓ | [ n1 ]q , then S1 has a one-dimensional
trivial radical. In the latter case, one easily sees by self-duality of F L1
that F L1 is uniserial with socle series F , D1 , F . The structure of S2
varies according to arithmetic properties; its composition factors can
be read off from James’s results [6]. It is also possible to describe all the
possible submodule lattices, but we shall only give details in the cases
we need for computing the elementary divisors. A general fact is that
S2 has a unique maximal submodule with quotient D2 . A consequence
of this fact that we shall use later on is that unless S2 = D2 , F L2 has
no submodule isomorphic to D2 .
The composition multiplicities [Si : Dj ] are given by [6, Theorem
20.7]; they are no greater than one, with [Si : Di ] = 1. (In James’s
notation the module F Li is denoted M(n−i,i) , the Specht module Sj is
denoted S(n−j,j) and the simple module Dj is denoted D(n−j,j) .) The
module D0 is the trivial module and the module D1 is the unique
nontrivial composition factor of F L1 (and of S1 ). The module D2 is
the unique composition factor of F L2 (and of S2 ) that is not isomorphic
to D0 or D1 .
We summarize the consequences of [6, Theorem 20.7] that we need,
using our notation.
12
DUCEY AND SIN
Lemma 5.1. We have the following composition multiplicities.
(a) [S1 : D0 ] = 1 if ℓ | [ n1 ]q and is zero otherwise.
(b) [S2 : D1 ] = 1 if ℓ | [ n−2
1 ]q and is zero otherwise.
(c) [S2 : D0 ] = 1 if one of the following conditions holds and is zero
otherwise.
(i) ℓ = 2 and n ≡ 1 or 2 (mod 4).
⌋.
(ii) ℓ 6= 2, ℓ | q + 1 and ℓ | ⌊ n−1
2
(iii) ℓ 6= 2, ℓ | q − 1 and ℓ | n − 1.
(iv) ℓ 6= 2, ℓ ∤ q − 1 and ℓ | [ n−1
1 ]q .
Here ⌊x⌋ denotes the greatest integer that is no greater than x.
We end this section with some elementary properties of q-binomial
coefficients that will be used frequently without further mention.
Lemma 5.2. Suppose ℓ | q + 1. Then
(i) ℓ | [ m1 ]q if and only if m is even.
(ii) Assume that m is even. Then
[ m1 ]q = (q + 1) m/2
:= (q + 1)(q m−2 + q m−4 + · · · + q 2 + 1).
1
q2
6. Grassmann, L′
6.0.1. Parameters.
k ′ = q(q + 1)[ n−2
1 ]q ,
(6.1)
µ′ = (q + 1)2 ,
2
λ′ = [ n−1
1 ]q + q − 2,
2 n−3
n
rL′ = q(q + 1)[ n−2
1 ]q − q [ 1 ]q + 1 = [ 1 ]q ,
n−1
sL′ = q(q + 1)[ n−2
1 ]q + (q + 1) = (q + 1)[ 1 ]q .
In this section we shall write r for rL′ and s for sL′ . The multiplicity
of r is f = dimK S̃1 = [ n1 ]q −1 and that of s is g = dimK S̃2 = [ n2 ]q −[ n1 ]q .
′
(6.2)
|K(Γ )| =
g
[ n1 ]q f ((q + 1)[ n−1
1 ]q )
[ n2 ]q
= [ n1 ]q f −1 (q + 1)g+1 [ n−1
1 ]q
g−1
.
′
6.0.2. ℓ ∤ q + 1. If ℓ ∤ [ n1 ]q and ℓ ∤ [ n−1
1 ]q . then, ℓ ∤ |K(Γ )|, so we
can assume that ℓ divides exactly one of [ n1 ]q and [ n−1
1 ]q , since these
numbers are coprime.
GRASSMANN GRAPHS
13
n
′
Case (i). ℓ | [ n1 ]q and ℓ ∤ [ n−1
1 ]q . Let a = vℓ ([ 1 ]q ). Then vℓ (|K(Γ )|) =
a(f − 1). By Lemma 3.2 and the definitions of Mi and ei , we have
(6.3)
f ≤ dimF M a .
Therefore, by Lemma 3.1 with d = a(f − 1), h = 1, b2 = 1, b1 = f and
a1 = a, we obtain
(6.4)
ea = f − 1,
e0 = g + 1,
and ei = 0 otherwise.
(Moreover, from Lemma 5.1, dimF D1 = f − 1 and in fact it is not
hard to deduce that M a /M a+1 ∼
= D1 .)
n
n−1
Case (ii). ℓ ∤ [ 1 ]q and ℓ | [ n−1
1 ]q . Let a = vℓ ([ 1 ]q ). Then
vℓ (|K(Γ′ )|) = a(g − 1). By Lemma 3.2
(6.5)
g ≤ dimF M a .
Therefore, by Lemma 3.1 with d = a(g − 1), h = 1, b2 = 1, b1 = g and
a1 = a, we obtain
(6.6)
ea = g − 1,
e0 = f + 1,
and ei = 0 otherwise.
(Moreover, from Lemma 5.1, dimF D2 = g − 1 and in fact it is not hard
to deduce that M a /M a+1 ∼
= D2 .)
(n−1)/2
. Let a =
6.0.3. ℓ | q + 1. Case (i) ℓ ∤ [ n1 ]q , ℓ | [ n−1
1 ]q , ℓ ∤
1
q2
(n−1)/2
vℓ (q + 1). Then since [ n−1
, we have vℓ (|K(Γ′ )|) =
1 ]q = (q + 1)
1
q2
a(g +1)+a(g −1) = 2ag and vℓ (s) = 2a. Also, ℓ ∤ [ n2 ]q , so M = R1⊕Y ,
where Y is the kernel of the map J sending each 2-subspace to 1, and L′
maps each summand to itself. We have R1 = Ker(L′ ) and YK contains
the s-eigenspace of dimension g. Hence by Lemma 3.2, applied to Y ,
we have g ≤ dimF M2a ∩ Y , so
(6.7)
g + 1 ≤ dimF M 2a .
Therefore, by Lemma 3.1 with d = 2ag, h = 1, b2 = 1, b1 = g + 1
and a1 = 2a, we obtain
(6.8)
e2a = g,
e0 = f,
and ei = 0 otherwise.
(Moreover, from Lemma 5.1, dimF D2 = g, dimF D1 = f and in fact it
∼
is not hard to deduce that M ∼
= F 1⊕D
1 ⊕D2 , with M 2a /M 2a+1 = D2 .)
(n−1)/2
Case (ii) ℓ ∤ [ n1 ]q , ℓ | [ n−1
. Let a = vℓ (q + 1) and
1 ]q , ℓ |
1
q2
(n−1)/2
′
). Then vℓ (|K(Γ )|) = a(g + 1) + (a + b)(g − 1) =
b = vℓ (
1
q2
(2a + b)(g − 1) + 2a and vℓ (s) = 2a + b.
We will extract some information from the equation (Lemma 2.1)
(6.9)
(L′ − sI)(L′ − rI) = µ′ J.
14
DUCEY AND SIN
Modulo ℓ we have
(L′ )(L′ − rI) = 0.
(6.10)
Since r 6= 0 (mod ℓ) this equation shows that M has a direct sum
decomposition into eigenspaces Ker L′ and Ker(L′ − rI). Since s ≡ 0
(mod ℓ), the (algebraic) multiplicity of 0 as an eigenvalue of L′ is g + 1,
whence dimF Ker L′ = g + 1 and dimF Ker(L′ − rI) = f . It also follows
that Ker L′ = Im(L′ − rI), so we have shown that dimF Im(L′ − rI) =
g + 1. Now (6.9) implies that for any x ∈ Im(L′ − rI) we have L′ x ∈
sx + µ′ R1, so, as ℓ2a | s and ℓ2a | µ′ we have Im(L′ − rI) ⊆ M2a . We
have proved that
(6.11)
g + 1 ≤ dimF M 2a .
By Lemma 3.2 we also have
(6.12)
g ≤ dimF M 2a+b .
Therefore, by Lemma 3.1 with d = (2a + b)(g − 1) + 2a, h = 2, b3 = 1,
b2 = g, b1 = g + 1, a2 = 2a + b and a1 = 2a, we obtain
(6.13)
e2a+b = g − 1,
e2a = 1 e0 = f,
and ei = 0 otherwise.
For the next two cases, we need the following lemma.
Lemma 6.1. Assume that ℓ | q + 1 and that ℓ | [ n1 ]q . Let a = vℓ (q + 1).
Let WP be the set of 2 − spaces containing a fixed 1-space P , and let
[WP ] ∈ M be its characteristic vector. Then [WP ] ∈ Ma .
Proof. Since a = vℓ (q + 1) and P
q + 1 | k ′ , it suffices to show that
′
′
[WP ] ∈ Ma (A ). Let A ([WP ]) = m∈L2 am m. Then the coefficient am
is the number of lines through P that are distinct from m and meet m.
n−2
This number is [ n−1
1 ]q − 1 = q[ 1 ]q if m passes through P and q + 1
if m does not. Since q + 1 | [ n1 ]q under our assumptions, we also have
q + 1 | [ n−2
1 ]q , so the Lemma holds.
n/2
Case (iii) ℓ | [ n1 ]q , ℓ ∤ [ n−1
, ℓ ∤ ⌊ n−1
⌋. Then [ n1 ]q =
1 ]q , ℓ ∤
2
1 q2
n/2
(q + 1) 1 q2 . Let a = vℓ (q + 1). Then vℓ (|K(Γ′ )|) = a(f + g). We
have ℓ ∤ [ n2 ]q , so shall always apply Lemma 3.2 to the complementary
summand Y to R1 in M. Write
(6.14)
(L′ − rI)(L′ − sI) = µ′J
as
(6.15)
L′ (L′ − (r + s)I) = −rsI + µ′ J
to get Im(L′ − (r + s)I) ⊆ M2a , hence M 2a ⊇ Im(L′ − (r + s)I) =
Im(L′ ) = Im(L′ ). It is easy to see that Im(L′ ) is not in the span of 1,
GRASSMANN GRAPHS
15
and that 1 ∈ M2a . Since S2 is not simple in this case, by Lemma 5.1,
the only simple module that can appear in the socle of Y is D1 . It
follows that dimF M 2a ≥ 1 + dimF D1 = f . Next, we will show that
dimF (M a ) ≥ g + 2. Over K, the submodule S̃2 is the s-eigenspace
of MK , so by Lemma 3.3 and Lemma 5.1 the composition factors of
M ∩ S̃2 are D1 and D2 .
The map η from RL1 → M sending P to [WP ] is an R GL(n, q)homomorphism. From the structure of F L1 , it is not hard to see
that Im(η) has composition factors F and D1 , and simple socle D1 .
In particular, Im(η) does not contain 1. Using the preceding paragraph and Lemma 6.1 we now see that M a has among its composition
factors F (occurring twice), D1 and D2 . This proves our claim that
dimF (M a ) ≥ g + 2.
Therefore, by Lemma 3.1 with d = a(f + g), h = 2, b3 = 1, b2 = f ,
b1 = g + 2, a2 = 2a and a1 = a, we obtain
(6.16)
e2a = f − 1, ea = g − f + 2, e0 = f − 1, and ei = 0 otherwise.
n/2
. In this case ℓ | n2 so
Case(iv) ℓ | [ n1 ]q , ℓ ∤ [ n−1
1 ]q , ℓ |
1 q 2
⌋. Let a = vℓ (q + 1) and b = vℓ ( n/2
). Then vℓ (|K(Γ′ )|) =
ℓ ∤ ⌊ n−1
2
1 q2
(a + b)f + ag − b. We claim that dimF (M a ) ≥ g + 2. Consider the
map η : RL1 → M sending a 1-subspace P to [WP ]. Since ℓ | q + 1,
the all-one vector of F L1 is in the kernel of η, and from the uniserial
structure of F L1 it is not hard to see that its span equals the kernel.
It follows that Im(η) is of dimension f and has no fixed points, and, in
particular, does not contain 1. However, by Lemma 6.1 Im(η) ⊆ Ma ,
so we know that M a has composition factors F (twice), D1 . Also,
since M ∩ S̃2 ⊆ Ma , we know by Lemma 3.3 that M a contains the
composition factors D1 and D2 of S2 . It follows that M a must have
at least composition factors F (twice), D1 and D2 , so is of dimension
≥ g + 2.
Let Y = Ker(J). Here Y is not a direct summand of M as an
R GL(n, q)-module but it is a pure R GL(n, q)-submodule.
Lemma 6.2. Under the current hypotheses on ℓ, we have F 1 ⊆ Y and
Y /F 1 is uniserial with socle series D1 , D2 , D1 .
Proof. The composition factors are given by Lemma 5.1. Since ℓ | [ n2 ]q
we have F 1 ⊆ Y = (F 1)⊥ , and Y /F 1 is self-dual with composition
factors D1 , D2 , D1 . The image of the Specht S2 in Y /F 1 has composition factors D2 and D1 , so is uniserial. The lemma now follows from
the self-duality of Y /F 1.
16
DUCEY AND SIN
We claim that dimF (M 2a+b ) ≥ f . To see this consider the image of
Y under L′ − (r + s)I. Since Y = Ker(J), the equation
(6.17)
L′ (L′ − (r + s)I) = −rsI + µ′ J
implies that (L′ −(r+s)I)(Y ) ⊆ M2a+b . It follows that L′ (Y ) ⊆ M 2a+b ,
and it is easy to check that L′ (Y ) is not in the span of 1. Since also
1 ∈ M2a+b , it follows that M 2a+b has at least one trivial composition
factor and one composition factor isomorphic to D1 , so its dimension
is at least f .
Therefore, by Lemma 3.1 with d = (a + b)f + ag − b, h = 2, b3 = 1,
b2 = f , b1 = g + 2, a2 = 2a + b and a1 = a, we obtain
(6.18)
e2a+b = f − 1, ea = g − f + 2, e0 = f − 1, and ei = 0 otherwise.
n/2
Case(v) ℓ | [ n1 ]q , ℓ ∤ [ n−1
, ℓ | ⌊ n−1
⌋. Let a = vℓ (q + 1).
1 ]q , ℓ ∤
2
1 q2
a
Here, r and s are exactly divisible by ℓ and ℓ ∤ [ n2 ]q , so vℓ (|K(Γ′ )|) =
a(f + g). Also, since ℓ ∤ [ n2 ]q , M = F 1 ⊕ Y , where Y = Ker(J). This
is an orthogonal decomposition with respect to the dot product on M.
Lemma 6.3. Y is a uniserial F GL(n, q)-module with socle series D1 ,
F , D2 , F , D1 .
Proof. The composition factors of Y and the Specht modules can be
read off from Lemma 5.1. Since the space of GL(n, q)-fixed points on
M is F 1, there are no fixed points on Y , hence none on its submodule
S2 . Since D2 is the unique simple quotient of S2 , it now follows that
S2 is uniserial with socle series D1 , F , D2 . The lemma is now implied
by the self-duality of Y .
Lemma 6.4. Let Wπ be the set of 2 − spaces contained in a fixed 3space π, and let [Wπ ] ∈ M be its characteristic vector. Then 1−[Wπ ] ∈
Ma (L′ ) and 1 − [Wπ ] ∈ Ma (A′ ).
′
′
a
Proof. Since
P q + 1 | k , it suffices to prove that A [Wπ ] ∈ ℓ M. Write
′
A [Wπ ] = L cL L, where L runs over the 2-subspaces. Then it is easy
to see that cL = 0 if L has zero intersection with π, that cL = q + 1 if L
meets π in a 1-dimensional space, and that cL = q 2 + q if L is contained
in π.
Lemma 6.5. Let h−, −i denote the natural dot product on M . Let π
be a 3-subspace and P a 1-subspace of π. Under the current hypotheses
on ℓ, the following hold, where we denote images in M of elements of
M by the same symbols.
(a) h1, 1 − [WP ]i = 0.
GRASSMANN GRAPHS
17
(b) h1, 1 − [Wπ ]i = 0.
(c) h1 − [WP ], 1 − [Wπ ]i =
6 0.
Proof. Under the hypothesis on ℓ, we know that n is even, that [ n2 ]q ≡
n
≡ 1 (mod ℓ), as ℓ | ( n2 − 1). The number of 2-spaces containing
2
P is [ n−1
1 ]q ≡ 1 (mod ℓ). The number of 2-spaces contained in π is
2
q + q + 1 ≡ 1 (mod ℓ), and of these there are q + 1 which contain P .
All parts of the lemma now follow.
We claim that dimF (M a ) ≥ g + 2. To see this we make use of
the uniserial structure of Y given in Lemma 6.3. Since ℓ | n, the
permutation module on points F L1 is uniserial with series F , D1 , F .
Let U1 be the submodule of M generated by the elements 1 − [WQ ]
as Q ranges over L1 . Then U1 ⊆ Y by Lemma 6.5(a). Also, U1 is
the image of F L1 under the F GL(n, q)-map sending Q to 1 − [WQ ], so
U1 is uniserial, and from the structures of Y and F L1 , we see that U1
has socle series D1 , F , so is equal to the second socle, soc2 (Y ). Thus,
from the structure of Y , we see that U1⊥ ∩ Y is equal to soc3 (Y ). Let
U3 be the submodule of M generated by the elements 1 − [Wπ ] as π
ranges over L3 . Then U3 ⊆ Y by Lemma 6.5(b), but U3 * U1⊥ ∩ Y by
Lemma 6.5(c). Therefore, from the uniserial structure of Y , it follows
that U3 ⊇ soc4 (Y ), so contains at least the composition factors D1 , F
(twice) and D2 . Hence dimF U3 ≥ g + 1. Our claim now follows from
Lemma 6.4 and the fact that 1 ∈ Ker(L′ ).
We next claim that dimF (M 2a ) ≥ f . From the equation
L′ (L′ − (r + s)I) = −rsI + µ′ J
(6.19)
we have Im(L′ − (r + s)I) ⊆ M2a , so Im(L′ ) ⊆ M 2a . This image is
easily seen not to be in the span of 1, hence from the structure of M,
must have a composition factor isomorphic to D1 . Also 1 ∈ M2a , so
M 2a has dimension at least 1 + dimF D1 = f .
Therefore, by Lemma 3.1 with d = a(f + g), h = 2, b3 = 1, b2 = f ,
b1 = g + 2, a2 = 2a and a1 = a, we obtain
(6.20)
e2a = f − 1, ea = g − f + 2, e0 = f − 1, and ei = 0 otherwise.
7. Grassmann, A′
7.0.1. Parameters.
(7.1)
k ′ = q(q + 1)[ n−2
1 ]q ,
µ′ = (q + 1)2 ,
n−2
rA′ = q 2 [ n−3
1 ]q − 1 = q[ 1 ]q − (q + 1),
2
λ′ = [ n−1
1 ]q + q − 2,
sA′ = −(q + 1).
18
DUCEY AND SIN
In this section we shall write r for rA′ and s for sA′ . The multiplicity
of r is f = dimK S̃1 = [ n1 ]q −1 and that of s is g = dimK S̃2 = [ n2 ]q −[ n1 ]q .
(7.2)
f n−2
|S(Γ′ )| = q(q + 1)g+1 (q 2 [ n−3
1 ]q − 1) [ 1 ]q .
Lemma 7.1.
(a) If n is odd then gcd(r, s) = 1, gcd(k ′ , s) = q + 1
′
and gcd(k , r) = 1.
(b) If n is even, then gcd(r, s) = q + 1, gcd(s, k ′ ) = q + 1, and any
common prime divisor of r and k ′ must divide q + 1.
7.0.2. ℓ ∤ q + 1. In this case, ℓ ∤ s. In the nontrivial case where ℓ | k ′ r,
we know by Lemma 7.1 that ℓ divides exactly one of k ′ and r.
Case(i). Suppose vℓ (k ′ ) = a, a > 0, ℓ ∤ r, ℓ ∤ s. Here, vℓ (|S(Γ′)|) = a
and we have 1 ∈ Ma , so dimF M a ≥ 1. Therefore, by Lemma 3.1 with
d = a, h = 1, b2 = 0, b1 = 1 and a1 = a, we obtain
(7.3)
ea = 1,
e0 = f + g,
and ei = 0 otherwise.
Case(ii). Suppose vℓ (r) = a, a > 0, ℓ ∤ k ′ , ℓ ∤ s. Here, vℓ (|S(Γ′ )|) =
af and by Lemma 3.2 we have dimF M a ≥ f . Therefore, by Lemma 3.1
with d = af , h = 1, b2 = 0, b1 = f and a1 = a, we obtain
(7.4)
ea = f,
e0 = g + 1,
and ei = 0 otherwise.
7.0.3. ℓ | q + 1. Since ℓ | q + 1, we have for any m that ℓ | [ m1 ]q if and
only of m is even, so we divide into two cases according to the parity
of n.
i) n even. Then
k ′ = q(q + 1)2 h, r = (q + 1)(qh − 1), s = −(q + 1),
h n−2 i
2
where h =
and that
. Note that ℓ | h if and only if ℓ | n−2
2
1
(7.5)
q2
gcd(h, qh − 1) = 1.
Let a = vℓ (q + 1), b = vℓ (qh − 1) and c = vℓ (h). Then either c = 0
or b = 0.
Note that since ℓ | q + 1 and n is even, we have ℓ | [ n2 ]q iff ℓ | n2 =
n−2
+ 1. But modulo ℓ, we have −( n−2
+ 1) ≡ −h − 1 ≡ qh − 1, so we
2
2
see that b = 0 if and only if ℓ ∤ [ n2 ]q .
a) Suppose c = 0 and b = 0. In this case ℓ ∤ [ n2 ]q , so we have
a decomposition M = R1 ⊕ Y . By Lemma 5.1 the module Y has
composition factors D1 (twice), D2 , F . From the self-duality of Y and
the fact that GL(n, q) has no fixed points in Y , one sees that Y has
socle series D1 , F ⊕ D2 , D1 . By Lemma 2.1, on Y = Ker(J) we have
(A′ − rI)(A′ − sI) = 0, which we can write as
(7.6)
A′ (A′ − (r + s)I) = −rsI.
GRASSMANN GRAPHS
19
This shows that (A′ − (r + s)I)(Y ) ⊆ M2a ∩ Y , so as ℓ | (r + s), we have
A′ (Y ) ⊆ M2a ∩ Y . It is easy to see that A′ (Y ) is nonzero, so it has D1
as a composition factor. In particular, dimF (M2a ∩ Y ) ≥ f − 1 and so
dimF M 2a ≥ f , as 1 ∈ M2a . By Lemma 3.3 applied to Y we know from
the fact that YK = S̃1 ⊕ S̃2 that Ma ∩ Y must contain the composition
factors of S1 and of S2 , so has at least the composition factors D1 , F
and D2 . This shows dimF (Ma ∩ Y ) ≥ (f − 1) + 1 + (g − f + 1) = g + 1,
so dimF M a ≥ g + 2. Now vℓ (|S(Γ′ )|) = a(f + g) + 2a. Therefore, by
Lemma 3.1 with d = a(f + g) + 2a, h = 2, b3 = 0, b2 = f , b1 = g + 2,
a2 = 2a and a1 = a, we obtain
(7.7) e2a = f,
ea = g − f + 2,
e0 = f − 1,
and ei = 0 otherwise.
b) Suppose b > 0, which implies c = 0. Here, vℓ (|S(Γ′)|) = (a+ b)f +
ag + 2a.
Lemma 7.2. Let Y = Ker(J). Then F 1 ⊆ Y and Y /F 1 is uniserial
with socle series D1 , D2 , D1 .
We will prove the following bounds.
Lemma 7.3.
(1) dimF M 2a+b ≥ f .
(2) dimF M a ≥ g + 2.
Proof. The restriction of A′ to Y satisfies the equation
(7.8)
A′ (A′ − (r + s)I) = −rsI.
This shows that (A′ − (r + s)I)(Y ) ⊆ M2a+b , so (A′ − (r + s)I)(Y ) ⊆
M 2a+b It is easy to see that (A′ − (r + s)I)(Y ) is not in the span of 1
so it has a composition factor D1 . By explicit computation we have
(7.9)
A′ ([Wp ]) = (q[ n−2
1 ]q −(q+1))[Wp ]+(q+1)1 = (q+1)(qh−1)[Wp ]+(q+1)1
and
(7.10)
A′ (1) = k ′ 1 = q(q + 1)2 h1.
Therefore, as h is a unit of R, we have (q + 1)[Wp ] − (qh)−1 1 ∈ M2a+b ,
whence F 1 ⊆ M 2a+b . This proves (1).
By Lemma 3.3 we know that M a has at least the composition factors
D2 and D1 of S2 , and we have already seen that F 1 ⊆ M a . We shall
show that M a contains a submodule with no GL(n, q)-fixed points but
with a trivial composition factor. This will show that M a at least two
trivial composition factors and (2) will be proved. The submodule in
question is the F -submodule W generated by the [Wp ], for 1-spaces p.
It is the image of the homomorphism η : F L1 → M sending a 1-space
to the sum of 2-spaces containing it. A direct computation shows that
20
DUCEY AND SIN
the image is not in the span of 1 and that the kernel contains the allone vector of F L1 . Since F L1 is uniserial with socle series F , D1 , F ,
it follows that W is uniserial with socle series D1 , F . Equation (7.9)
shows that A′ [Wp ] ∈ Ma , so we are done.
In view of the above lemma can apply Lemma 3.1 with d = (a +
b)f + ag + 2a, h = 2, b3 = 0, b2 = f , b1 = g + 2, a2 = 2a + b and a1 = a,
we obtain
(7.11)
e2a+b = f, ea = g − f + 2, e0 = f − 1, and ei = 0 otherwise.
c) Suppose c > 0, which implies b = 0. Then ℓ ∤ [ n2 ]q . We have a
decomposition M = R1 ⊕ Y . Here, vℓ (|S(Γ′ )|) = af + ag + 2a + c. We
will prove the following bounds.
(1) dimF M 2a+c ≥ 1.
(2) dimF M 2a ≥ f .
(3) dimF M a ≥ g + 2.
(1) is by simply computing A′ (1). For (2) and (3) we note that the
current hypotheses on ℓ imply those of case (v) for L′ in §6, and that
as ℓ2a | k ′ we have M2a (A′ ) = M2a (L′ ) and Ma (A′ ) = Ma (L′ ).
Therefore, by Lemma 3.1 with d = af + ag + 2a + c, h = 3, b4 = 0,
b3 = 1, b2 = f , b1 = g + 2, a3 = 2a + c, a2 = 2a and a1 = a, we obtain
(7.12)
e2a+c = 1, e2a = f −1, ea = g −f +2, e0 = f −1, and ei = 0 otherwise.
ii) n odd. Then ℓ ∤ [ n1 ]q and ℓ ∤ [ n−2
1 ]q , Let a = vℓ (q + 1). Then, by
′
Lemma 7.1, a = vℓ (k ) = vℓ (s) and ℓ ∤ r. Here, vℓ (|S(Γ′)|) = a + ag.
a) ℓ ∤ ⌊ n−1
⌋. In this case we have a decomposition M = R1 ⊕ Y .
2
Applying Lemma 3.2 to Y , we have dimF (Ma ∩ Y ) ≥ g, and so since
1 ∈ Ma , we have dimF M a ≥ g + 1.
Therefore, by Lemma 3.1 with d = a(g + 1), h = 1, b2 = 0, b1 = g + 1
and a1 = a, we obtain
(7.13)
ea = g + 1,
e0 = f,
and ei = 0 otherwise.
b) ℓ | ⌊ n−1
⌋. Let a = vℓ (q + 1). We claim that dimF M a ≥ g + 1. By
2
Lemma 2.1, we have (A′ −sI)(A′ −rI) = µ′ J, where µ′ = (q+1)2 . Thus,
A′ (A′ − rI) = s(A′ − rI) + µ′ J, which shows that Im(A′ − rI) ⊆ Ma
and hence Im(A′ − rI) ⊆ M a . Now on M, A′ has a quadratic minimal
polynomial with distinct roots 0 and r, so we have an F GL(n, q)decomposition
(7.14)
M = Ker(A′ − rI) ⊕ Ker A′ = Ker(A′ − rI) ⊕ Im(A′ − rI).
GRASSMANN GRAPHS
21
The dimension of Ker A′ is equal to the algebraic multiplicity of 0 as
an eigenvalue of A′ , which equals g + 1 since ℓ divides both s and k ′ .
Therefore dimF M a ≥ g + 1. It now follows from Lemma 3.1 with
d = a(g + 1), h = 1, b2 = 0, b1 = g + 1 and a1 = a, that
(7.15)
ea = g + 1,
e0 = f,
and ei = 0 otherwise.
8. Skew lines, L
8.0.1. Parameters.
µ=
(8.1)
rL =
n−1
[
]
−
(q
+
2)
q 3 [ n−3
]
1 q
1 q
q+1
q 4 [ n−2
2 ]q
−
q 2 [ n−3
1 ]q
sL = q 4 [ n−2
2 ]q − q =
=
n
q 2 [ n−3
1 ]q [ 1 ]q
q+1
,
n−1
[
]
−
(q
+
2)
q[ n−1
]
1 q
1 q
.
q+1
In this section we shall write r for rL and s for sL . The multiplicity of
r is f = dimK S̃1 = [ n1 ]q − 1 and that of s is g = dimK S̃2 = [ n2 ]q − [ n1 ]q .
(8.2)
|K(Γ)| = q
2f +g
[ n1 ]q [ n−3
1 ]q
q+1
f
!g
n−1 −(q+2)
[ n−1
1 ]q [ 1 ]q
q+1
[ n2 ]q
Lemma 8.1.
(a) If n is odd then no prime ℓ 6= p can divide both r
and s.
(b) If n is even, then any common prime divisor of r and s must
divide q(q + 1).
(c) [ n−1
1 ]q divides s(q + 1).
Proof. (a) Suppose ℓ 6= p divides both r and s. Then ℓ divides both
n
3 n−2
3 n−2
n−3
([ n−3
1 ]q /(q + 1))[ 1 ]q and q [ 2 ]q − 1 = q [ 1 ]q ([ 1 ]q /(q + 1)) − 1.
n
Obviously, ℓ cannot divide [ n−3
1 ]q /(q + 1), so ℓ must divide [ 1 ]q =
3 n−2
q 2 [ n−2
1 ]q + (q + 1). Hence q [ 1 ]q ≡ −q(q + 1) (mod ℓ) and so
n−3
n−3
n−2
0 ≡ q 3 [ n−2
1 ]q ([ 1 ]q /(q + 1)) − 1 ≡ −q[ 1 ]q − 1 ≡ −[ 1 ]q
(mod ℓ),
n−3
a contradiction since ℓ divides q 3 [ n−2
1 ]q ([ 1 ]q /(q + 1)) − 1.
(b) Suppose that ℓ divides r and s, but ℓ does not divide q(q + 1).
3 n−2
n−3
Then ℓ divides ([ n1 ]q /(q + 1))[ n−3
1 ]q and q ([ 1 ]q /(q + 1))[ 1 ]q − 1 so,
n
2 n−2
obviously, ℓ does not divide [ n−3
1 ]q . Hence ℓ divides [ 1 ]q = q [ 1 ]q +
22
DUCEY AND SIN
q + 1. Thus q 3 [ n−2
1 ]q ≡ −q(q + 1) (mod ℓ), and it follows that ℓ divides
n−3
n−2
−q[ 1 ]q − 1 = −[ n−2
1 ]q . But then ℓ divides [ 1 ]q /(q + 1), which
n−3
contradicts the fact that ℓ divides q 3 ([ n−2
1 ]q /(q + 1))[ 1 ]q − 1.
(c) This is clear from the factorization of s given at the beginning of
this section.
8.0.2. ℓ ∤ q + 1. By Lemma 8.1, ℓ can divide at most one of r and s, so
we consider the two nontrivial cases in turn.
Case(i) ℓ | r, ℓ ∤ s.
n
n
a) ℓ | [ n1 ]q and ℓ ∤ [ n−3
1 ]q . Let a = vℓ ([ 1 ]q ). Then a = vℓ ([ 2 ]q ),
and vℓ (|K(Γ)|) = a(f − 1). By Lemma 3.2, we have dimF (M a ) ≥ f .
Therefore, by Lemma 3.1 with d = a(f − 1), h = 1, b2 = 1, b1 = f and
a1 = a, we obtain
(8.3)
ea = f − 1,
e0 = g + 1,
and ei = 0 otherwise.
2
b) ℓ | [ n1 ]q and ℓ | [ n−3
1 ]q . Note that we must have ℓ | q + q + 1. Let
a = vℓ ([ n1 ]q ), b = vℓ ([ n−3
1 ]q ),
n
Here, vℓ ([ 2 ]q ) = a, so vℓ (|K(Γ)|) = a(f − 1) + bf . By Lemma 3.2 we
have dimF M a+b ≥ f . We claim dimF M b ≥ f + 1.
From Lemma 2.1 we have L(L − (s + r)I) = −rsI + µJ, which shows
that Im(L − (s + r)I) ⊆ Mb , hence Im(L − sI) ⊆ M b . Now on M,
L has quadratic minimal polynomial with eigenvalues s and 0, which
gives the decomposition of F GL(n, q)-modules
(8.4)
M = Ker(L) ⊕ Ker(L − sI) = Im(L − sI) ⊕ Ker(L − sI).
The dimension of Ker L is equal to the algebraic multiplicity of 0 as an
eigenvalue of L, which equals f + 1 since ℓ divides both r and k. We
have shown that dimF M b ≥ f + 1.
Therefore, by Lemma 3.1 with d = a(f − 1) + bf , h = 2, b3 = 1,
b2 = f , b1 = f + 1, a2 = a + b and a1 = b, we obtain
(8.5)
ea+b = f − 1,
eb = 1,
e0 = g,
and ei = 0 otherwise.
c) ℓ ∤ [ n1 ]q and ℓ | [ n−3
1 ]q . Under the present assumptions, we have
n
ℓ ∤ [ 2 ]q , so we have a decomposition M = R1 ⊕ Y as R GL(n, q)modules. Let a = vℓ ([ n−3
1 ]q ). Then vℓ (|K(Γ)|) = af . Using Lemma 3.2
applied to Y , we see that dimF (Ma ∩ Y ) ≥ f , so dimF (M a ) ≥ f + 1.
Therefore, by Lemma 3.1 with d = af , h = 1, b2 = 1, b1 = f + 1 and
a1 = a, we obtain
(8.6)
ea = f,
Case (ii) ℓ | s, ℓ ∤ r.
e0 = g,
and ei = 0 otherwise.
GRASSMANN GRAPHS
23
n
a) ℓ ∤ [ n−1
1 ]q . Let a = vℓ (s). Then since ℓ ∤ [ 2 ]q , we have vℓ (|K(Γ)|) =
ag. Also M = R1 ⊕ Y . Using Lemma 3.2 applied to Y , we see that
dimF (Ma ∩ Y ) ≥ g, so dimF (M a ) ≥ g + 1. Therefore, by Lemma 3.1
with d = ag, h = 1, b2 = 1, b1 = g + 1 and a1 = a, we obtain
(8.7)
ea = g,
e0 = f,
and ei = 0 otherwise.
n−1
b) ℓ | [ n−1
1 ]q . Let a = vℓ ([ 1 ]q ). Then vℓ (s) = a+b, where b = vℓ (µ).
Then vℓ (|K(Γ)|) = (a + b)g − a.
If b = 0, then by Lemma 3.2, dimF (M a ) ≥ g. Therefore, by
Lemma 3.1 with d = a(g − 1), h = 1, b2 = 1, b1 = g and a1 = a,
we obtain
(8.8)
ea = g − 1,
e0 = f + 1,
and ei = 0 otherwise.
Suppose b > 0. Here we have vℓ (|K(Γ)|) = (a + b)g − a. By
Lemma 3.2 we have dimF M a+b ≥ g, and we claim that dimF M b ≥
g +1. By Lemma 2.1, we have L(L−rI) = s(L−rI)+µJ, which shows
that Im(L − rI) ⊆ Mb , hence Im(L − rI) ⊆ M b . The same equation
also shows that on M , L has a quadratic minimal polynomial with
distinct roots r and 0, so M = Ker(L − rI) ⊕ Ker(L). The algebraic
multiplicity of 0 is g + 1, the sum of the multiplicities of s and 0 as
eigenvalues of L, so dimF Ker(L) = g + 1. Then, from the decomposition above, we have dimF Im(L − rI) = g + 1, and our claim follows.
Therefore, by Lemma 3.1 with d = (a + b)g − a, h = 2, b3 = 1, b2 = g,
b1 = g + 1, a2 = a + b and a1 = b, we obtain
(8.9)
ea+b = g − 1,
eb = 1,
e0 = f,
and ei = 0 otherwise.
8.0.3. ℓ | q + 1. As ℓ | q + 1, we have for any m, ℓ | [ m1 ]q if and only if
m is even, so we consider two cases according to the parity of n.
Case (i) n even.
Lemma 8.2. ℓ | r iff ℓ |
n
2
iff ℓ | s.
Proof. As n is even, ℓ divides [ n1 ]q and [ n−2
neither [ n−1
1 ]q but divides
1 ]q
n/2
2 n−3
n−3
nor [ 1 ]q . Since ℓ divides q + 1 and r = q [ 1 ]q 1 q2 , it follows
that ℓ divides r if and only if ℓ divides n/2. Since
3 (n−4)/2
n−1
n−1
s = q[ n−1
+ (q − 1)),
1 ]q ([ 1 ]q − (q + 2))/(q + 1) = q[ 1 ]q (q
1
q2
and q ≡ −1 (mod ℓ), we see that ℓ divides s if and only if ℓ divides
−(n − 4)/2 − 2 = −n/2.
Let a = vℓ (r), b = vℓ (s). We can assume a, b > 0, or else vℓ (|K(Γ)|) =
0. Then vℓ ([ n2 ]q ) = a and we have vℓ (|K(Γ)|) = af + bg − a.
24
DUCEY AND SIN
We claim (1) dimF M a+b ≥ f and (2) dimF M b ≥ g +1. To prove (1),
we observe that the equation L(L − (r + s)I) = −rsI + µJ implies that
(L − (r + s)I)(Y ) ⊆ Ma+b . It is easy to see that L(Y ) is not in the span
of 1, so contains a nontrivial composition factor. Since [S2 : D1 ] 6= 0
we see that L(Y ) must have a composition factor D1 , whose dimension
is f − 1. Also F 1 ⊆ M a+b so (1) is proved. For (2), we note that S2
has composition factors D2 and D1 , so by Lemma 3.3, M ∩ S̃2 also has
these composition factors, so M b has at least these composition factors.
In addition we have F 1 ⊆ M b , so dimF M b ≥ g + 1. Therefore, by
Lemma 3.1 with d = af + bg − a, h = 2, b3 = 1, b2 = f , b1 = g + 1,
a2 = a + b and a1 = b, we obtain
(8.10) ea+b = f −1,
eb = g−f +1,
e0 = f,
and ei = 0 otherwise.
Case (ii) n odd. Let a = vℓ (r), b = vℓ (s). We can assume vℓ (|K(Γ)|) >
0. By Lemma 8.1, exactly one of a and b is nonzero.
. In this case ℓ ∤ [ n2 ]q , so vℓ (|K(Γ)|) = af + bg and we have
a) ℓ ∤ n−1
2
a decomposition M = R1 ⊕ Y .
If a > 0, b = 0, then from Lemma 3.2, we have dimF Ma ∩ Y ≥ f , so
dimF Ma ≥ f + 1. Therefore, by Lemma 3.1 with d = af + bg, h = 1,
b2 = 1, b1 = f + 1 and a1 = a, we obtain
(8.11)
ea = f,
e0 = g,
and ei = 0 otherwise.
If a = 0, b > 0, then from Lemma 3.2, we have dimF Mb ∩ Y ≥ g, so
dimF Mb ≥ g + 1. Therefore, by Lemma 3.1 with d = af + bg, h = 1,
b2 = 1, b1 = g + 1 and a1 = b, we obtain
(8.12)
eb = g,
n−1
.
2
e0 = f,
and ei = 0 otherwise.
[ n−1
1 ]q
= vℓ ([ n2 ]q ). We
h n−1 i
2
,
so
since
ℓ
|
q
+
1,
we
have
≡
are assuming that ℓ | n−1
2
1
q2
h n−1 i
[ n−1
1 ]q
n−1
2
(mod
ℓ),
so
ℓ
|
and we have b > 0. Then by
=
2
q+1
1
b) ℓ |
Here we have b = vℓ (s) = vℓ
q+1
q2
Lemma 8.1 a = 0. Then vℓ (|K(Γ)|) = bg − b. By Lemma 3.2, we have
dimF M b ≥ g. Therefore, by Lemma 3.1 with d = bg − b, h = 1, b2 = 1,
b1 = g and a1 = b, we obtain
(8.13)
eb = g − 1,
e0 = f + 1,
and ei = 0 otherwise.
GRASSMANN GRAPHS
25
9. Skew lines, A
9.0.1. Parameters.
(9.1)
k = q 4 [ n−2
2 ]q ,
n−1
q
− q2 − q + 1
3 n−3
µ = q [ 1 ]q
q2 − 1
!
n−1
[
]
−
(q
+
2)
1
q
= q 3 [ n−3
,
1 ]q
q+1
rA = −q 2 [ n−3
1 ]q ,
sA = q.
In this section we shall write r for rA and s for sA . The multiplicity of
r is f = dimK S̃1 = [ n1 ]q − 1 and that of s is g = dimK S̃2 = [ n2 ]q − [ n1 ]q .
f
n−3
|S(Γ)| = q 4+2f +g [ n−2
2 ]q ([ 1 ]q )
(9.2)
= q 4+2f +g
f +1
n−3
[ n−2
1 ]q ([ 1 ]q )
q+1
!
.
n−3
9.0.2. ℓ ∤ q + 1. If ℓ ∤ [ n−2
1 ]q and ℓ ∤ [ 1 ]q the ℓ-part of S(Γ) is trivial,
so we consider the nontrivial cases.
n−3
n−2
Case (i) ℓ | [ n−2
1 ]q , ℓ ∤ [ 1 ]q . Let a = vℓ ([ 2 ]q ). Then as we have
1 ∈ Ma , we have dimF (M a ) ≥ 1 and by Lemma 3.1 with d = a, h = 1,
b2 = 0, b1 = 1 and a1 = a, we obtain
(9.3)
ea = 1,
e0 = f + g,
and ei = 0 otherwise.
n−3
n−3
Case (ii) ℓ ∤ [ n−2
1 ]q , ℓ | [ 1 ]q . Let a = vℓ ([ 1 ]q ). We claim
dimF M a ≥ f + 1. Suppose first that ℓ ∤ [ n1 ]q . Then ℓ ∤ [ n2 ]q , so we have
a decomposition M = R1 ⊕ Y . The r-eigenspace of A is contained in
YK , so f ≤ dimF Ma ∩ Y . Also 1 ∈ Ma , so the claim is true in this
case. Now assume that ℓ | [ n1 ]q . In this case, F L1 is uniserial with its
unique simple submodule spanned by the all-one vector 11 of F L1 . The
map η : F L1 → M maps 11 to (q + 1)1 6= 0, so it is an injective map.
Its (f + 1)-dimensional image is the span of the elements [Wp ]. Direct
computation shows that A([Wp ]) = q 2 [ n−3
1 ]q (1 − [Wp ]), so Im(η) ⊆ Ma .
Therefore, by Lemma 3.1 with d = a(f + 1), h = 1, b2 = 0, b1 = f + 1
and a1 = a, we obtain
(9.4)
ea = f + 1,
e0 = g,
and ei = 0 otherwise.
26
DUCEY AND SIN
9.0.3. ℓ | q + 1. Case (i) ℓ | [ n−2
1 ]q , but ℓ ∤
h n−2 i
2
1
q2
. In this case
n−3
ℓ ∤ [ n−2
of S(Γ) is trivial.
2 ]q and ℓ ∤ [ 1 ]q , so the ℓ-part
h n−2 i
2
. In this case, ℓ ∤ [ n−3
Case (ii) ℓ | [ n−2
1 ]q . Let
1 ]q , and ℓ |
1
q2
a = vℓ ([ n−2
2 ]q ). Then 1 ∈ Ma , so by Lemma 3.1 with d = a, h = 1,
b2 = 0, b1 = 1 and a1 = a, we obtain
(9.5)
ea = 1,
e0 = f + g, and ei = 0 otherwise.
h n−3 i
2
Case (iii) ℓ | [ n−3
]
,
but
ℓ
∤
. Let a = vℓ ([ n−3
1 q
1 ]q ). Then
1
q2
n−3
[ n−2
1 ]q [ 1 ]q
. We have f ≤ dimF M a
a = vℓ (q + 1) and so ℓ ∤ [ n−2
2 ]q =
q+1
by Lemma 3.2. Therefore, by Lemma 3.1 with d = af , h = 1, b2 = 0,
b1 = f and a1 = a, we obtain
(9.6)
ea = f,
e0 = g + 1, and ei = 0 otherwise.
h n−3 i
2
Case (iv) ℓ | [ n−3
]
,
and
ℓ
|
. Let a = vℓ (q + 1) and b =
1 q
1
q2
[ n−3
1 ]q
vℓ ( q+1 ). Then b = vℓ ([ n−2
2 ]q ), so 1 ∈ Mb , and F 1 ⊆ M b . Now
n−3
n−1
ℓ | 2 , so ℓ ∤ 2 , and it follows that ℓ ∤ [ n2 ]q . Thus, M = R1 ⊕ Y ,
where YK contains the whole r-eigenspace in MK . As vℓ (r) = a + b it
follows from Lemma 3.2 that f ≤ dimF Ma+b ∩ Y . As M = F 1 ⊕ Y ,
and 1 ∈ Mb , we also have dimF M b ≥ 1 + f . Therefore, by Lemma 3.1
with d = f (a + b) + b, h = 2, b3 = 0, b2 = f , b1 = f + 1, a2 = a + b and
a1 = b, we obtain
(9.7)
ea+b = f,
eb = 1,
e0 = g,
and ei = 0 otherwise.
References
1. Andries E. Brouwer, Joshua E. Ducey, and Peter Sin, The elementary divisors
of the incidence matrix of skew lines in PG(3, q), Proc. Amer. Math. Soc. 140
(2012), no. 8, 2561–2573.
2. Andries E. Brouwer and Willem H. Haemers, Spectra of graphs, Universitext,
Springer, New York, 2012. MR 2882891
3. David B. Chandler, Peter Sin, and Qing Xiang, The invariant factors of the
incidence matrices of points and subspaces in PG(n, q) and AG(n, q), Trans.
Amer. Math. Soc. 358 (2006), no. 11, 4935–4957. MR 2231879
4.
, The Smith and critical groups of Paley graphs, J. Algebraic Combin. 41
(2015), no. 4, 1013–1022. MR 3342710
5. Charles W. Curtis and Irving Reiner, Methods of representation theory. Vol. I,
John Wiley & Sons, Inc., New York, 1981, With applications to finite groups
and orders, Pure and Applied Mathematics, A Wiley-Interscience Publication.
MR 632548
GRASSMANN GRAPHS
27
6. G. D. James, Representations of general linear groups, London Mathematical
Society Lecture Note Series, vol. 94, Cambridge University Press, Cambridge,
1984.
10. Appendix
The two tables in this appendix show the F GL(n, q)-submodule
structure of the permutation module M , according to the relation of ℓ,
q and n.
ℓ ∤ [ n−2
1 ]q
ℓ∤q+1
ℓ ∤ [ n−1
1 ]q
S 2 = D2
M = F ⊕ D1 ⊕ D2
S 2 = D2 + F
F
ℓ | [ n−1
1 ]q
M = D1 ⊕ D2
F
S 2 = D2
⌋
ℓ ∤ ⌊ n−1
2
ℓ|q+1
M = F ⊕ D1 ⊕ D2
S 2 = D2 + F
F
⌋
ℓ | ⌊ n−1
2
M = D1 ⊕ D2
F
Table A. ℓ ∤ [ n1 ]q .
ℓ | [ n−2
1 ]q
S 2 = D2 + D1
M =F⊕
N/A
N/A
N/A
D1
D2
D1
28
ℓ∤q+1
ℓ|q+1
DUCEY AND SIN
ℓ | [ n−1
1 ]q
ℓ ∤ [ n−2
1 ]q
S 2 = D2
F
M = D1 ⊕ D2
F
N/A
⌋
ℓ ∤ ⌊ n−1
2
N/A
ℓ ∤ [ n−1
1 ]q
ℓ | [ n−2
1 ]q
N/A
N/A
S 2 = D2 + D1
ℓ ∤ [ n2 ]q
M=
F⊕
④
④④
D2 ❈
❈❈
ℓ | ⌊ n−1
⌋
2
N/A
D1 ❅
ℓ | [ n2 ]q
D1
❅❅
F
M=
⑦
⑦⑦⑦
D1
S 2 = D2 + D1 + F
D1
F
M = F ⊕ D2
F
D1
Table B. ℓ | [ n1 ]q .
Dept. of Mathematics and Statistics, James Madison University,
Harrisonburg, VA 22807, USA
E-mail address: duceyje@jmu.edu
Dept. of Mathematics, University of Florida, P.O. Box 118105, Gainesville,
FL 32611, USA
E-mail address: sin@ufl.edu
F
⑦
⑦⑦
D2
D1
F
⑦
⑦⑦⑦
| 4math.GR |
TRAINING PROBABILISTIC SPIKING NEURAL NETWORKS WITH FIRST-TO-SPIKE
DECODING
Alireza Bagheri† , Osvaldo Simeone‡ , and Bipin Rajendran†
arXiv:1710.10704v3 [stat.ML] 22 Feb 2018
†
Department of Electrical and Computer Engineering (ECE),
New Jersey Institute of Technology, Newark, NJ 07102, USA.
Email: {ab745; bipin}@njit.edu
‡
Centre for Telecommunications Research, Department of Informatics,
King’s College London, London, WC2R 2LS, UK.
Email: osvaldo.simeone@kcl.ac.uk
ABSTRACT
Third-generation neural networks, or Spiking Neural Networks
(SNNs), aim at harnessing the energy efficiency of spike-domain
processing by building on computing elements that operate on,
and exchange, spikes. In this paper, the problem of training a
two-layer SNN is studied for the purpose of classification, under
a Generalized Linear Model (GLM) probabilistic neural model
that was previously considered within the computational neuroscience literature. Conventional classification rules for SNNs
operate offline based on the number of output spikes at each
output neuron. In contrast, a novel training method is proposed
here for a first-to-spike decoding rule, whereby the SNN can
perform an early classification decision once spike firing is
detected at an output neuron. Numerical results bring insights
into the optimal parameter selection for the GLM neuron and on
the accuracy-complexity trade-off performance of conventional
and first-to-spike decoding.
Index Terms— Spiking Neural Network (SNN), Generalized Linear Model (GLM), first-to-spike decoding, neuromorphic computing
1. INTRODUCTION
Most current machine learning methods rely on secondgeneration neural networks, which consist of simple static
non-linear neurons. In contrast, neurons in the human brain
are known to communicate by means of sparse spiking processes. As a result, they are mostly inactive, and energy is
consumed sporadically. Third-generation neural networks, or
Spiking Neural Networks (SNNs), aim at harnessing the energy
efficiency of spike-domain processing by building on computing
elements that operate on, and exchange, spikes [1]. SNNs can be
natively implemented on neuromorphic chips that are currently
being developed within academic projects and by major chip
manufacturers. Proof-of-concept implementations have shown
remarkable energy savings by multiple orders of magnitude with
respect to second-generation neural networks (see, e.g., [2], [3]).
Notwithstanding the potential of SNNs, a significant stumbling block to their adoption is the dearth of flexible and ef-
1
1 2
T
1 2
T
1
1 2
T΄
NY 1 2
T΄
Output layer
1 2
T
NX
Input layer
Fig. 1. Two-layer SNN for supervised learning.
fective learning algorithms. Most existing algorithms are based
on variations of the unsupervised mechanism of Spike-Timing
Dependent Plasticity (STDP), which updates synaptic weights
based on local input and output spikes, and supervised variations that leverage global feedback [4], [5]. Another common
approach is to convert trained second-generation networks to
SNNs [6], [7]. Among the learning methods that attempt to
directly maximize a spike-domain performance criterion, most
techniques assume deterministic Spike Response Model (SRM)
neurons, and propose various approximations to cope with the
non-differentiability of the neurons’ outputs (see [8], [9] and
references therein).
While the use of probabilistic models for spiking neurons is
standard in the context of computational neuroscience (see, e.g.,
[10]), probabilistic modeling has been sparsely considered in the
machine learning literature on SNNs. This is despite the known
increased flexibility and expressive power of probabilistic models [11], [12]. In the context of SNNs, as an example, probabilistic models have the capability of learning firing thresholds
using standard gradient based methods, while in deterministic
models these are instead treated as hyperparameters and set by
using heuristic mechanisms such as homeostasis [13]. The state
of the art on supervised learning with probabilistic models is
set by [14] that considers Stochastic Gradient Descent (SGD)
for Maximum Likelihood (ML) training, under the assumption
that there exist given desired output spike trains for all output
neurons.
In this paper, we study the problem of training the two-layer
SNN illustrated in Fig. 1 under a probabilistic neuron model, for
the purpose of classification. Conventional decoding in SNNs
operates offline by selecting the output neuron, and hence the
corresponding class, with the largest number of output spikes
[14]. In contrast, we study here a first-to-spike decoding rule,
whereby the SNN can perform an early classification decision
once a spike firing is detected at an output neuron. This generally
reduces decision latency and complexity during the inference
phase. The first-to-spike decision method has been investigated
with temporal, rather than rate, coding and deterministic neurons
in [15]–[18], but no learning algorithm exists under probabilistic
neural models.
To fill this gap, we first propose the use of the flexible
and computationally tractable Generalized Linear Model (GLM)
that was introduced in [19] in the context of computational
neuroscience (Section 3). Under this model, we then derive
a novel SGD-based learning algorithm that maximizes the
likelihood that the first spike is observed at the correct output
neuron (Section 4). Finally, we present numerical results that
bring insights into the optimal parameter selection for the GLM
neuron and on the accuracy-complexity trade-off performance
of conventional and first-to-spike decoding rules.
2. SPIKING NEURAL NETWORK WITH GLM
NEURONS
In this section, we describe the architecture of the two-layer
SNN under study and then we present the proposed GLM neuron
model.
Architecture. We consider the problem of classification
using a two-layer SNN. As shown in Fig. 1, the SNN is fully
connected and has NX presynaptic neurons in the input, or
sensory layer, and NY neurons in the output layer. Each output
neuron is associated with a class. In order to feed the SNN,
an input example, e.g., a gray scale image, is converted to a
set of NX discrete-time spike trains, each with T samples,
through rate encoding. The input spike trains are fed to the NY
postsynaptic GLM neurons, which output discrete-time spike
trains. A decoder then selects the image class on the basis of the
spike trains emitted by the output neurons.
Rate encoding. With the conventional rate encoding
method, each entry of the input signal, e.g., each pixel for
images, is converted into a discrete-time spike train by generating an independent and identically distributed (i.i.d.) Bernoulli
vectors. The probability of generating a “1”, i.e., a spike, is
proportional to the value of the entry. In the experiments in Sec.
5, we use gray scale images with pixel intensities between 0 and
255 that yield a spike probability between 0 and 1/2.
GLM neuron model. The relationship between the input
spike trains from the NX presynaptic neurons and the output
spike train of any postsynaptic neuron i follows a GLM, as
illustrated in Fig. 2. To elaborate, we denote as xj,t and yi,t
x1 τ
y
tT
1
α1,i
γi
ui,t activation g(ui,t) Bernoulli
+
function
r.v. gen.
xNX
1
τy
tT
yi τ'y
1
t T΄
βi
αNX,i
Fig. 2. GLM neuron model.
the binary signal emitted by the j -th presynaptic and the ith postsynaptic neurons, respectively, at time t. Also, we let
xbj,a = (xj,a , ..., xj,b ) be the vector of samples from spiking
process of the presynaptic neuron j in the time interval [a, b].
b
Similarly, the vector yi,a
= (yi,a , ..., yi,b ) contains samples
from the spiking process of the neuron i in the interval [a, b].
As seen in Fig. 2, the output yi,t of postsynaptic neuron i at time
t is Bernoulli distributed, with firing probability that depends on
t−1
the past spiking behaviors {xj,t−τ
} of the presynaptic neurons
y
j = 1, ..., NX in a window of duration τy samples, as well as
t−1
on the past spike timings yi,t−τ
0 of neuron i in a window of
y
0
duration τy samples. Mathematically, the membrane potential of
postsynaptic neuron i at time t is given by
NX
X
T t−1
(1)
αTj,i xt−1
ui,t =
j,t−τy + β i yi,t−τ 0 + γi ,
y
j=1
τy
where αj,i ∈ R is a vector that defines the synaptic kernel
(SK) applied on the {j, i} synapse between presynaptic neuron
0
j and postsynaptic neuron i; β i ∈ Rτy is the feedback kernel
(FK); and γi is a bias parameter. The vector of variable parameters θ i includes the bias γi and the parameters that define the
SK and FK filters, which are discussed below. Accordingly, the
T
log-probability of the entire spike train yi = [yi,1 , ..., yi,T ]
NX
conditioned on the input spike trains x = {xj }j=1 can be
written as
T
X
log pθi (yi |x ) =
[yi,t log g (ui,t ) + ȳi,t log ḡ (ui,t )],
t=1
(2)
where g (·) is an activation function, such as the sigmoid
function g (x) = σ (x) = 1/ (1 + exp (−x)), and we defined
ȳi,t = 1 − yi,t and ḡ (ui,t ) = 1 − g (ui,t ).
Unlike prior work on SNNs with GLM neurons, we adopt
here the parameterized model introduced in [19] in the field
of computational neuroscience. Accordingly, the SK and FK
filters are parameterized as the sum of fixed basis functions with
learnable weights. To elaborate, we write the SK αj,i and the
FK β i as
αj,i = Awj,i , and β i = Bvi ,
(3)
respectively, where we have defined
the matrices A =
[a1 , ..., aKα ] and B = b1 , ..., bKβ and the vectors wj,i =
T
T
[wj,i,1 , ..., wj,i,Kα ] and vi = vi,1 , ..., vi,Kβ ; Kα and
Kβ denote the respective number of basis functions; ak =
1
2
1
3
4
6
5
7
4. TRAINING WITH FIRST-TO-SPIKE DECODING
8
0
1
2
3
4
5
6
7
8
Time index t
Fig. 3. Basis functions used in Sec. 5 (a = 7500 and c = 1 in
[19, Sec. Methods]).
iT
h
T
are the basis
and bk = bk,1 , ..., bk,τy0
ak,1 , ..., ak,τy
vectors; and {wj,i,k } and {vi,k } are the learnable weights for
the kernels αj,i and β i , respectively. This parameterization generalizes previously studied models for machine learning application. For instance, as a special case, if we set Kα = Kβ = 1,
set a1 and b1 as in [14, eqs. (4) and (5)], and fix the weights
vi,1 = 1, equation (2) yields a discrete-time approximation of
the model considered in [14]. As another example, if we set
Kα = τy , Kβ = τy0 , ak = 1k , bk = 1k , where 1k is the
all-zero vector except for a one in position k , (1) yields the unstructured GLM model considered in [20]. For the experiments
discussed in Sec. 5, we adopt the time-localized raised cosine
basis functions introduced in [19], which are illustrated in Fig.
3. Note that this model is flexible enough to include the learning
of synaptic delays [21], [22].
In this section, we introduce the proposed learning approach
based on GLM neurons and first-to-spike decoding.
During the inference phase, with first-to-spike decoding, a
decision is made once a first spike is observed at an output
neuron. In order to train the SNN for this classification rule, we
propose to follow the ML criterion by maximizing the probability to have the first spike at the output neuron corresponding to
the correct label c. The logarithm of this probability for a given
example x can be written as
!
T
X
L (θ) = log
pt (θ) ,
(5)
t=1
where
N
Y
Y
pt (θ) =
t
Y
ḡ (ui,t0 )g (uc,t )
i=1,i6=c t0 =1
t−1
Y
ḡ (uc,t0 ),
(6)
t0 =1
is the probability of having the first spike at the correct neuron
c at time t. In (6), the potential ui,t for all i is obtained from (1)
by setting yi,t = 0 for all i and t. The log-likelihood function
L (θ) in (5) is not concave, and we tackle its maximization via
SGD.
To this end, the gradients of the log-likelihood function for
a given input x can be computed after some algebra as (see
Appendix B for details)
∇wj,i L (θ)
T
P
i 6= c
ρi,t ht g (ui,t ) AT xt−1
−
j,t−τy
(7)
t=1
,
=
T
P
t−1
−
ρc,t (ht g (uc,t ) − qt ) AT xj,t−τy i = c
t=1
3. TRAINING WITH CONVENTIONAL DECODING
In this section, we briefly review ML training based on conventional rate decoding for the two-layer SNN. During the inference
phase, decoding is conventionally carried out in post-processing
by selecting the output neuron with the largest number of spikes.
In order to facilitate the success of this decoding rule, in the
training phase, the postsynaptic neuron corresponding to the
correct label c ∈ {1, ..., NY } is typically assigned a desired
output spike train yc with a number of spikes, while a zero
output is assigned to the other postsynaptic neurons yi with
i 6= c.
Using the ML criterion, one hence maximizes the sum of
the log-probabilities (2) of the desired output spikes y (c) =
{y1 (c) , ..., yNY (c)} for the correct label c given the NX input
spike trains x = {x1 , ..., xNX }, i.e.,
NY
X
L (θ) =
log pθi ( yi (c)| x).
(4)
i=1
The sum is further extended to all examples in the training set.
The parameter vector θ = {W, V, γ} includes the parameters
NY
NY
NY
W = {Wi }i=1
, V = {vi }i=1
and γ = {γi }i=1
. The
negative log-likelihood −L (θ) is convex with respect to θ and
can be minimized via SGD. For completeness, we report the
gradients of L (θ) in Appendix A.
for the weights and
T
P
ρi,t ht g (ui,t )
−
t=1
∇γi L (θ) =
T
P
−
ρc,t (ht g (uc,t ) − qt )
i 6= c
, (8)
i=c
t=1
for the biases, where we have defined
g 0 (ui,t )
,
ρi,t =
g (ui,t ) ḡ (ui,t )
and
T
t−1
X
X
ht =
qt0 = 1 −
qt0 ,
t0 =t
(9)
(10)
t0 =1
with
qt =
pt (θ)
T
P
.
(11)
pt0 (θ)
t0 =1
Note that we have ρi,t = 1 when g is the sigmoid function.
Based on (7), the resulting SGD update can be considered as
a neo-Hebbian rule [23], since it multiplies the contributions of
the presynaptic neurons and of the postsynaptic activity, where
the former depends on x and the latter on the potential ui,t .
Furthermore, in (7)−(8), the probabilities g (ui,t ) and g (uc,t )
of firing at time t are weighted by the probability ht in (10). By
(11), this is the probability that the correct neuron is the first to
spike and that it fires at some time t0 ≥ t, given that it is the
first to spike at some time in the interval [1, 2, ..., T ].
99
100
2 digits
90
ANN
98.5
T = 32
T = 16
T=8
80
Test Accuracy [%]
Test Accuracy [%]
98
4 digits
70
10 digits
60
T = 32
T = 16
97.5
T=4
97
T=8
96.5
96
50
40
Conventional (rate) dec.
First-to-spike dec.
1
2
3
Conventional (rate) dec.
First-to-spike dec.
95.5
4
K
Fig. 4. Test accuracy versus the number K of basis functions for
both conventional (rate) and first-to-spike decoding rules when
T = 4.
As a practical note, in order to avoid vanishing values
in calculating the weights (11), we compute each probability term pt (θ) in the log-domain, and normalize all the
resulting terms with
P respect to the minimum probability as
qt = exp (at ) / Tt0 =1 exp (at0 ), where at = ln (pt ) −
mint (ln (pt )).
5. NUMERICAL RESULTS
In this section, we numerically study the performance of the
probabilistic SNN in Fig. 1 under conventional and first-to-spike
decoding rules. We use the standard MNIST dataset [24] as the
input data. As a result, we have NX = 784, with one input
neuron per pixel of the x images. Following [3], we consider
different number of classes, or digits, namely, the two digits
{5, 7}, the four digits {5, 7, 1, 9} and all 10 digits {0, ..., 9},
and we use 1000 samples of each class for training and the same
number for test set. We use a desired spike train with one spike
after every three zeros for training the conventional decoding.
SGD with minibatch size of one with 200 training epochs is
used for both schemes. Ten-fold cross-validation is applied for
selecting between 10−3 or 10−4 for the constant learning rates.
The model parameters θ are randomly initialized with uniform
distribution between -1 and 1.
We evaluate the performance of the schemes in terms of
classification accuracy in the test set and of inference complexity. The inference complexity is measured by the total number
of elementary operations, namely additions and multiplications,
for input image that are required by the SNN during inference.
The number of arithmetic operations needed to calculate the
membrane potential (1) of neuron i at time instant t is of
the order of O NX τy + τy0 . As a result, in the conventional
decoding method, the inference complexity per output neuron,
or per class, is of the order O (T (NX sx + sy )), where sx and
sy are the fraction of spikes in x and y, respectively. In contrast,
with the first-to-spike decoding rule, the SNN can perform an
95
T=4
0
100
200
300
400
500
600
700
800
Inference Complexity [# elem. op.]
Fig. 5. Test accuracy versus per-class inference complexity for
both conventional (rate) and first-to-spike decoding rules.
early decision once a single spike is detected, and hence its
complexity order is O (t (NX sx + sy )), where 1 ≤ t ≤ T is
the (random) decision time.
We first consider the test classification accuracy as a function
of the number K of basis functions in the GLM neural model.
The basis functions are numbered as in Fig. 3, and we set T = 4.
From Fig. 4, we observe that conventional decoding requires a
large number K in order to obtain its best accuracy. This is
due to the need to ensure that the correct output neuron fires
consistently more than the other neurons in response to the input
spikes. This, in turn, requires a larger temporal reception field,
i.e., a larger K , to be sensitive to the randomly located input
spikes. We note that for small values of T , such as T = 4, firstto-spike decoding obtains better accuracies than conventional
decoding.
Fig. 5 depicts the test classification accuracy versus the
inference complexity for both conventional and first-to-spike
decoding rules for two digits when K = T . The classification
accuracy of a conventional two-layer artificial neural network
(ANN) with logistic neurons is added for comparison. From
the figure, first-to-spike decoding is seen to offer a significantly
lower inference complexity, thanks to its capability for early
decisions, without compromising the accuracy. For instance,
when the classification accuracy equals to 98.4%, the complexity of the conventional decoding method is five times larger than
the first-to-spike method. Note also that conventional decoding
generally requires large values of T to perform satisfactorily.
6. CONCLUSIONS
In this paper, we have proposed a novel learning method for
probabilistic two-layer SNN that operates according to the firstto-spike learning rule. We have demonstrated that the proposed
method improves the accuracy-inference complexity trade-off
with respect to conventional decoding. Additional work is
needed in order to generalize the results to multi-layer networks.
7. ACKNOWLEDGMENT
This work was supported by the U.S. NSF under grant ECCS
#1710009. O. Simeone has also received funding from the
European Research Council (ERC) under the European Unions
Horizon 2020 research and innovation program (grant agreement #725731).
APPENDIX
A. GRADIENTS FOR CONVENTIONAL DECODING
For a given input x, the gradients of the log-likelihood function
L (θ) in (4) for conventional decoding are given as
T
X
∇wj,i L (θ) =
ei,t ρi,t AT xt−1
(12)
j,t−τy ,
t=1
∇vi L (θ) =
T
X
T
ei,t ρi,t B
t−1
yi,t−τ
0,
y
(13)
t=1
and
∇γi L (θ) =
T
X
ei,t ρi,t ,
(14)
t=1
we have
t
t
t
Y
Y
X
g 0 (ui,t0 )
∇wj,i
ḡ (ui,t0 ) = −
∇wj,i ui,t0
ḡ (ui,t0 )
ḡ (ui,t0 )
t0 =1
t0 =1
t0 =1
=−
t
Y
t0 =1
ei,t = yi,t − g (ui,t ) ,
=−
t=1
∇wj,i pt (θ)
t=1
T
P
(17)
pt (θ)
t0 =1
(21)
ht ρi,t g (ui,t ) ∇wj,i ui,t ,
t=1
where we have defined qt and ht as in (11) and (10), respectively. Given that we have the equality ∇wj,i ui,t = AT xt−1
j,t−τy ,
we have (7) for ∇wj,i L (θ) when i 6= c.
The gradient of L (θ) with respect to wj,i for i = c can be
calculated as:
!
T
t−1
X
Y
∇wj,c L (θ) =
kc,t ∇wj,c g (uc,t )
ḡ (uc,t0 ) , (22)
t0 =1
t=1
N
QY
(16)
The gradient of L (θ) with respect to wj,i for i 6= c can be
calculated as:
!
T
X
∇wj,i L (θ) = ∇wj,i log
pt (θ)
=
t=1
T
X
kc,t =
B. CALCULATION OF GRADIENTS FOR FIRST-TOSPIKE DECODING
ρi,t0 g (ui,t0 ) ∇wj,i ui,t0 ,
t0 =1
(15)
is the error signal, and ρi,t is given as
g 0 (ui,t )
ρi,t =
,
g (ui,t ) ḡ (ui,t )
∆ dg(ui,t )
where g 0 (ui,t ) = dui,t
.
T
P
t
X
(20)
where we have used the equality ḡ 0 (u) = −g 0 (u) and ρi,t is
defined as in (9). After substituting (20) into (17), we have
T
t
X
X
∇wj,i L (θ) = −
qt
ρi,t0 g (ui,t0 ) ∇wj,i ui,t0
where
where
ḡ (ui,t0 )
t
Q
ḡ (ui,t0 )
i=1,i6=c t0 =1
T
P
.
(23)
pt0 (θ)
t0 =1
Using (19), we have
∇wj,c L (θ)
"
#
T
t−1 0
X
X
g 0 (uc,t )
g (uc,t0 )
=
qt
∇wj,c uc,t −
∇wj,c uc,t0
g (uc,t )
ḡ (uc,t0 )
t=1
t0 =1
"
#
T
t
X
X
=
qt ρc,t ∇wj,c uc,t −
ρc,t0 g (uc,t0 ) ∇wj,c uc,t0 .
t=1
t0 =1
(24)
into
(24)
and
Thus, by substituting ∇wj,c uc,t = AT xt−1
j,t−τy
after simplification, equation (7) is obtained for i = c, which
completes the proof. Note that, the same procedure is done for
∇γi L (θ) by considering the equality ∇γi ui,t = 1 for all i.
t=1
=
where
ki,t =
T
X
ki,t ∇wj,i
t
Q
ḡ (ui0 ,t0 )g (uc,t )
i0 =1,i0 6=i,c t0 =1
t−1
Q
t0 =1
T
P
t0 =1
8. REFERENCES
ḡ (ui,t0 ),
t0 =1
t=1
N
QY
t
Y
ḡ (uc,t0 )
.
(18)
pt0 (θ)
Using the generalized product rule for derivative of k factors
[25] as
k
k
X
Y
d Y
d
fi (x)
fi (x) =
fj (x)
dx i=1
dx
i=1
j6=i
(19)
! k
!
k
Y
X f 0 (x)
i
=
fi (x)
,
f
(x)
i
i=1
i=1
[1] H. Paugam-Moisy and S. Bohte, “Computing with spiking
neuron networks,” in Handbook of natural computing,
2012, pp. 335–376.
[2] J.
Vincent,
“Intel
investigates
chips
designed like your brain to turn the AI tide,”
https://www.theverge.com/2017/9/26/16365390/intelinvestigates-chips-designed-like-your-brain-to-turn-the
-ai-tide, accessed: Sept. 26, 2017.
[3] A. Diamond, T. Nowotny, and M. Schmuker, “Comparing neuromorphic solutions in action: implementing a
bio-inspired solution to a benchmark classification task
on three parallel-computing platforms,” Front. Neurosci.,
vol. 9, p. 491, 2016.
[4] F. Ponulak and A. Kasiński, “Supervised learning in spiking neural networks with ReSuMe: sequence learning,
classification, and spike shifting,” Neural Comput., vol. 22,
no. 2, pp. 467–510, 2010.
[5] R. V. Florian, “Reinforcement learning through modulation of spike-timing-dependent synaptic plasticity,” Neural
Comput., vol. 19, no. 6, pp. 1468–1502, 2007.
[6] P. O’Connor and M. Welling, “Deep spiking networks,”
arXiv preprint. arXiv:1602.08323, 2016.
[7] E. Hunsberger and C. Eliasmith, “Spiking deep networks
with LIF neurons,” arXiv preprint. arXiv:1510.08829,
2015.
[8] N. Anwani and B. Rajendran, “NormAD-normalized approximate descent based supervised learning rule for spiking neurons,” in Proc. IEEE Int. Joint Conf. on Neural
Netw. (IJCNN), 2015.
[9] J. H. Lee, T. Delbruck, and M. Pfeiffer, “Training deep
spiking neural networks using backpropagation,” Front.
Neurosci., vol. 10, p. 508, 2016.
[10] J. W. Pillow, L. Paninski, V. J. Uzzell, E. P. Simoncelli, and
E. J. Chichilnisky, “Prediction and decoding of retinal ganglion cell responses with a probabilistic spiking model,” J.
Neurosci., vol. 25, no. 47, pp. 11 003–11 013, 2005.
[11] D. Koller and N. Friedman, Probabilistic graphical models: principles and techniques. MIT press, 2009.
[12] O. Simeone, “A brief introduction to machine learning for
engineers,” arXiv preprint arXiv:1709.02840, 2017.
[13] R. Jolivet, A. Rauch, H. Lüscher, and W. Gerstner, “Predicting spike timing of neocortical pyramidal neurons by
simple threshold models,” J. Comput. Neurosci., vol. 21,
no. 1, pp. 35–49, 2006.
[14] B. Gardner and A. Grüning, “Supervised learning in spiking neural networks for precise temporal encoding,” PloS
ONE, vol. 11, no. 8, p. e0161335, 2016.
[15] T. Masquelier and S. J. Thorpe, “Unsupervised learning of
visual features through spike timing dependent plasticity,”
PLoS Comput. Biol., vol. 3, no. 2, p. e31, 2007.
[16] M. Mozafari, S. Kheradpisheh, T. Masquelier, A. NowzariDalini, and M. Ganjtabesh, “First-spike based visual categorization using reward-modulated STDP,” arXiv preprint
arXiv:1705.09132, 2017.
[17] J. Wang, A. Belatreche, L. P. Maguire, and T. M. McGinnity, “SpikeTemp: An enhanced rank-order-based learning
approach for spiking neural networks with adaptive structure,” IEEE Trans. Neural Netw. Learn. Syst., vol. 28, no. 1,
pp. 30–43, 2017.
[18] Z. Lin, D. Ma, J. Meng, and L. Chen, “Relative ordering
learning in spiking neural network for pattern recognition,”
Neurocomputing, vol. 275, pp. 94–106, 2018.
[19] J. W. Pillow, J. Shlens, L. Paninski, A. Sher, A. M. Litke,
E. J. Chichilnisky, and E. P. Simoncelli, “Spatio-temporal
correlations and visual signalling in a complete neuronal
population,” Nature, vol. 454, no. 7207, pp. 995–999,
2008.
[20] J. Shlens, “Notes on generalized linear models of neurons,”
arXiv preprint arXiv:1404.1999, 2014.
[21] P. Baldi and A. F. Atiya, “How delays affect neural dynamics and learning,” IEEE Transactions on Neural Networks,
vol. 5, no. 4, pp. 612–621, 1994.
[22] A. Taherkhani, A. Belatreche, Y. Li, and L. P. Maguire,
“DL-ReSuMe: a delay learning-based remote supervised
method for spiking neurons,” IEEE Trans. Neural Netw.
Learn. Syst., vol. 26, no. 12, pp. 3137–3149, 2015.
[23] N. Frémaux and W. Gerstner, “Neuromodulated spiketiming-dependent plasticity, and theory of three-factor
learning rules,” Front. Neural Circuits, vol. 9, 2015.
[24] Y. LeCun, “The MNIST database of handwritten digits,”
http://yann.lecun.com/exdb/mnist/, 1998.
[25] Wikipedia,
“Product
rule,”
https://en.wikipedia.org/wiki/Product rule,
accessed:
Oct. 27, 2017.
| 9cs.NE |
Object-Oriented Modeling of Programming Paradigms
arXiv:cs/0603016v2 [cs.SE] 9 Nov 2006
M.H. van Emden
University of Victoria
Victoria, B.C., Canada
Abstract
For the right application, the use of programming paradigms such as functional or logic programming can enormously increase productivity in software development. But
these powerful paradigms are tied to exotic programming
languages, while the management of software development
dictates standardization on a single language.
This dilemma can be resolved by using object-oriented
programming in a new way. It is conventional to analyze
an application by object-oriented modeling. In the new approach, the analysis identifies the paradigm that is ideal
for the application; development starts with object-oriented
modeling of the paradigm. In this paper we illustrate the
new approach by giving examples of object-oriented modeling of dataflow and constraint programming. These examples suggest that it is no longer necessary to embody a
programming paradigm in a language dedicated to it.
1. Introduction
What programming language should we use? The answer to this question has changed over the decades. In the
1970s the answer was: “The Right One”. Since then it has
become: “What Everyone Else Is Using”. For example, in
the 1970s one company embarked on the design and implementation of a language that was to be ideal for developing
telephone switch software. Although they were successful,
in the 1990s they judged it more important to use a standardized language with multiple and competing vendors. Accordingly, the ideal language was replaced by C and C++,
much to the detriment of their subsequent software development.
Not only in this company, but in almost every other organization, a similar shift has occurred. In this paper, we want
to re-examine the now discarded answer, “The Right One.”
Why was this ever considered the right answer?
We believe it was based on the observation that with the
languages such as Prolog, Scheme or ML, some problems
become miraculously easy to program. But that depends
S.C. Somosan
NewHeights Software Corporation
Victoria, B.C., Canada
on the problem: if it is easy in Scheme or ML, it may not
be so in Prolog, and vice versa. Thus, the effect depends
on the programming paradigm on which the language is
based: functional programming in the case of Scheme and
ML; logic programming in the case of Prolog.
The fact that, with the right choice of language some
problems become miraculously easy to program we call the
Whitehead effect, inspired by the following quote from Alfred North Whitehead (1861-1947):
“By relieving the brain of unnecessary work, a
good notation sets it free to concentrate on more
advanced problems, . . . ”
Whitehead goes on to claim that the effect is to increase the
mental capacity of those who use the notation.
In this connection one should also note the Sapir-Whorf
hypothesis, stating that what one can think is determined
by one’s language. In this form the hypothesis is vague in
the extreme. Though various attempts at making it concrete
have been discredited, the hypothesis may have something
to do with the fact that, with the right language, some things
become miraculously easy to program.
1.1. Radical Software Development
The Whitehead effect suggests that one first determines
the programming paradigm that makes the application
miraculously easy to program and that one then uses a language that embodies the paradigm. This we call Radical
Software Development.
It is easy to see why this approach is not practical: Radical Software Development tends to lead to different paradigms, as required by the different applications. As it is,
paradigms are locked up in programming languages. As a
result Radical Software Development requires writing modules in several programming languages.
1.2. A One-Language World
Meanwhile, in the real world, companies are increasingly under pressure to “stick to their knitting”. For example, this means that a company making telephone switches
gets out of developing and maintaining its own ideal programming language. Managers want to cut costs of training new hires and insist on a widely used, standardized language. For further economies, such a language should come
with tools and multiple, competing vendors. This explains
why the world of programming has become a one-language
world, or at least has been trying to.
It is not just managers who are opposed to multiple programming languages. The Right Language tends to be an
exotic one, and does not sit well on the resume of the programmer who may soon need to look for another job. Another problem with The Right Language is what we call the
Ninety-Percent Phenomenon, which we will now explain by
an example.
Suppose one finds that logic programming is the
paradigm that makes the application miraculously easy. At
present, this means using Prolog. A browse of the manual
shows that ninety-percent of it is taken up by matters that
have nothing to do with logic programming. This ninetypercent is taken up with the mundane infrastructure that all
languages seem to require, regardless of paradigm: what
characters are allowed in names, what kind of numbers
there are, how you write them, strings, characters, I/O, and
so on. To their credit, the designers of C++ and Java have
been careful to do this as much as possible as it has been
done in C. In Prolog it has been done independently of C, or
of any other language. As a result, Prolog does it sometimes
the way you guess it will do it, and sometimes not. Rather
time consuming, and frustrating.
Thus it was apparently common to regard functions as
objects before 1977. It was only a matter of time to work out
the consequences of this view. In fact, it happened twenty
years later in Friedman and Felleisen’s A Little Java, A Few
Patterns [5]. However, it took the form of a cryptic onepage appendix.
From this page, it is apparent that the absence of generics in Java makes the modelling of higher-order functional
programming a tiresome exercise. To overcome this limitation, one could have followed C++ and added templates.
However, Wadler and Odersky judged the Hindley-Milner
type system, used in ML, as a superior alternative. They
extended Java to include this type system. This resulted
languages such as Pizza and GJava [15, 2] that directly allow higher-order functional programming, without the need
for object-oriented modelling in the sense of this paper. The
Java extended in this way became the standard in 2004 under the name “Java 2 Platform, Standard Edition 5”.
Pizza and GJava are examples of Multi-Paradigm Object
Oriented Programming Languages. Here the object-orientation of Java is used to extend it to include another programming paradigm. Several other projects in this direction were
reported at MPOOL 2001 [3], a workshop in conjunction
with the ECOOP conference.
As we explained, there is a strong tendency for one
language to dominate practical programming. Although a
multi-paradigm OO language can be more elegant and powerful, it is unlikely to be accepted in practice in the foreseeable future. It is therefore of interest to see how far one can
go in the direction of multiparadigm programming within
a programming language that is widely used in practice.
In this paper we show by worked examples in C++ that
multi-paradigm programming is not only possible, but can
be quite simple and elegant. To fit the paper’s format, we
need to restrict ourselves to small paradigms. As we will
see, dataflow and constraints are small enough to exhibit as
worked examples here.
2. OO modelling of the paradigm
In object-oriented programming (OOP) it is routine to
model an application from the ground up in terms of objects. In one widely used approach [24], one notes the nouns
and the verbs of a description of the application in English.
The nouns are then considered as candidates for the classes,
the verbs as candidates for the methods. If one can do this
for transactions, invoices, customers, . . ., why not for the
key concepts that make a particular programming paradigm
miraculously effective for the application? In this paper we
argue that OOP can be used to do this for programming paradigms.
Let us consider functional programming. In procedureoriented languages, numbers are privileged in that one can
(1) give them names, (2) assign them to variables, and (3)
return them as function results. In these languages, functions are underprivileged in that they share with numbers
property (1), but not (2) and (3). The motivation for functional programming languages is to accord to them all the
privileges of numbers, and thus make them “first-class objects”, as was the going terminology in functional programming [20].
3. Dataflow programming
In the dataflow paradigm all computation happens in a
network consisting of nodes connected by unidirectional
datapipes. Thus each node has zero or more input pipes
(when the output end of the pipe is connected to the node),
and zero or more output pipes (when the input end of the
pipe is connected to the node). A pipe is used by repeatedly placing data items at its input end. These items can
be retrieved from its output end in the same order in which
they entered at the opposite end. Abstractly viewed, the
pipes behave like the abstract data structure referred to as a
queue.
The dataflow paradigm is justified by the class of problems that it makes easy to solve. For example, such prob2
lems arise in business process re-engineering. Such processes are naturally analyzed and re-designed in terms of
workflow [9]. The automation of workflow then naturally
translates to dataflow.
An older method is Structured Systems Analysis (also
called SADT, for Structured Analysis and Design Technique), which was at one time widely applied in commercial dataprocessing [8]. However, Structured Systems Analysis was a world unto itself, apparently not aware of the
larger context of dataflow programming. However, Structured Systems Analysis was a world unto itself, apparently
not aware of the larger context of dataflow programming.
Structured Systems Analysis discovered dataflow by inspired thinking about commercial dataprocessing applications. Ashcroft and Wadge [22] arrived at the idea starting
from studies in semantics of programming languages. Dennis, Arvind and others at MIT have arrived at the dataflow
paradigm from computer architecture [21]. Dataflow has
been identified as a Software Architecture [19].
As early as 1977 the paradigm was already sufficiently
compelling that a programming language was designed to
make dataflow programming as natural as possible [13].
The paper just mentioned also contains some simple and
widely appealing examples showing the paradigm at its
best. Note that the paradigm was independently arrived at
from disparate areas: business applications, semantics, and
architecture (hardware and software). This suggests that it’s
“real” in some sense.
• merge is the result of merging its two sorted input sequences into a sorted output sequence, suppressing duplicates
• t2 is result of multiplying by 2 its input sequence, element by element; similarly for t3 and t5
The question whether the solution to Hamming’s problem is the only solution to the equation is addressed by the
methods developed in Kahn [12].
To turn this observation into a dataflow network, we take
the above equation with a complex expression and turn it
into a system of simple equations by introducing auxiliary
variables. We do this in two steps. In the first step, we get
rid of nested expressions:
a = t2 (x2 ) b = t3 (x2 ) d = t5 (x2 )
c = merge(a, b) x1 = merge(c, d)
x2 = 1 ◦ x1
The resulting equations can, if considered in isolation,
each be translated directly into a node of a dataflow network. Each of a, b, c, d, and x1 correspond to a datapipe
because, in the above set of equations, they have exactly
one occurrence in a left-hand side and exactly one occurrence in a right-hand side. An occurrence on the left-hand
(right-hand) side corresponds to the output (input) side of
the datapipe.
However, x2 has too many occurrences in right-hand
sides. We can avoid this problem by making these occurrences into different variables, say f , g, and h. But how
do we tell that these are the same sequence? To do that,
we introduce a node type, with one input and two output
pipes, that outputs two identical copies of each item that it
removes from the input pipe. Let us call this node split.
3.1. Hamming’s problem solved in dataflow
A good introduction to the dataflow paradigm is Hamming’s problem [13]:
to print out in increasing order all positive integers that have no prime factors other than 2, 3, or
5.
i = x2 h = x2 f = i g = i
a = t2 (f ) b = t3 (g) d = t5 (h)
Thus, the sequence starts with 1, 2, 3, 4, 5, 6, 8, 9, 10,
12, 15, . . . This problem is attributed to R.W. Hamming
by Dijkstra [4], who provided an ingenious solution. It is
more efficient than, but is not as easy to understand as, the
dataflow version given by Kahn and McQueen [13], which
we follow here.
One approach to a solution starts with the observation
that the infinite sequence x of numbers required by Hamming’s problem satisfies the following equation:
c = merge(a, b)
x2 = 1 ◦ x1
x1 = merge(c, d)
The entire top two lines each correspond each to a split
node; the remaining six equations correspond to a node
each. This makes eight nodes in all. See the network diagram in Figure 1.
3.2. An object-oriented implementation of
Hamming’s problem
x = 1 ◦ merge(merge(t2 (x), t3 (x)), t5 (x))
where
The most widely known principle of object-oriented
modelling is to consider the nouns of the specification as
candidates for classes.
• the meaning of ◦ is defined by u ◦ v being the result of
prefixing the sequence v by the element u
3
a
t2
f
m1
sp2
g
c
b
m2
t3
d
i
h
sp1
t5
x1
x2
1
p
01: int main() {
02:
const int MaxTimes = 50;
03:
queue a(10), b(10), c(10), d(10),
04:
x1(10), x2(10),
05:
f(10), g(10), h(10), i(10);
06:
merge m1(&a,&b,&c), m2(&c,&d,&x1);
07:
times t2(2,&f,&a), t3(3,&g,&b),
08:
t5(5,&h,&d);
09:
split sp1(&x2,&h,&i), sp2(&i,&f,&g);
10:
print p(&x1,&x2);
11:
12:
node* ar[] = { &m1, &m2, &t2, &t3,
13:
&t5, &sp1, &sp2, &p };
14:
int arSz = sizeof(ar) / sizeof(node*);
15:
16:
x1.add(1);
17:
for ( int i=0; i < MaxTimes; i++ )
18:
for ( int j=0; j < arSz; j++ )
19:
ar[j]->run();
20:
return 0;
21: }
Figure 1. Dataflow network for Hamming’s
problem in the initial state, where all pipes
are empty except a 1 in x1.
In any dataflow network, particularly conspicuous nouns
are node and datapipe. As observed before, datapipes behave like queues, a commonly used class.
According to the principle just mentioned, we should
consider a class suitable for creating all required nodes as
instances. The principle, though a good first approximation, needs some refinement. This is because objects of the
same class should have states of the same form, though not
necessarily of the same content. The state of a node object
includes the states of the abutting datapipes. And of course,
instances of the same class should have the same behaviour.
These considerations suggest that the merge nodes are
instances of the same class (merge; the individual instances are m1 and m2), as are the split nodes (of the class
split, with instances sp1 and sp2), as are the nodes t2 ,
t3 , and t5 (of the class times, with instances t1, t2, and
t3).
Lines 3–5 create instances of class queue to act as datapipes, each with a maximum size arbitrarily set at 10.
Given these pipes, lines 6–10 create the nodes, with the correct pipe connections. These lines seem the most succinct
possible textual representation of the diagram in Figure 1.
In so far as this is true, C++ comes close to the best possible
dataflow programming language.
Lines 3–10 create the dataflow network; they do not
cause it to execute its computation. The attraction of the
dataflow paradigm is to avoid the difficulty of conventional
programming, namely to ensure that events happen in the
right sequence. To execute a dataflow network, each node
executes, independently of the others, the following simple
computation:
Figure 2. C++ code for the dataflow network
for Hamming’s problem. The node print
has been added to allow the solution to be
printed.
4
};
If any of the input datapipes is empty, or if any
of the output datapipes is full, do nothing. Otherwise, remove the next item from each of the input
pipes, perform on them the specialized computation characteristic for the type of node, and place
the results, if any, on the output pipes.
An attractive characteristic of dataflow is that the nodes can
run concurrently subject to mutual exclusion on the datapipes. We have considered doing this example in Java to
make it easy to have every instance of node run in its own
thread. But although threads are simple to use, the result is
still not as simple as doing without. To find Hamming numbers, threads are not essential. So we make our point better
by showing a simpler program that just gets the numbers.
The computation just described is invoked by a method
called run, which is defined for each of the classes merge,
times, and split.
To execute the entire network, one invokes the run
method for each node, as in lines 18–19 of Figure 2. Typically several of these invocation have no effect because of
full output or empty input pipes. But if the network can do
anything at all, then at least one node will do something.
In large networks it is worth optimizing the invocations of
the run method. One can keep track of which nodes are
blocked. A blocked node connected to a pipe of which the
content changed may no longer be blocked and becomes
a candidate for being run. Such an optimization is reminiscent of the constraint propagation algorithm of D. Waltz
[23].
Note that in lines 12–13 the nodes of the dataflow are
placed into an array. The order of the nodes in this array
gives the order in which the run methods are called (line
19). However, this order does not matter, since in dataflow
the order in which things are done matters less than it does
in conventional programming.
Space limitations prevent us from listing the entire program, which is about a hundred lines, including the queue
implementation. We just add some representative code:
4. Constraint programming
Many problems in resource planning and numerical
computation can be solved in a declarative way: one states
the relations that are to hold between the unknowns; a suitable solver then finds values such that all relations are satisfied. The relations are referred to as constraints; the method
is known as constraint programming [14, 11].
Constraint programming is useful because systems can
be solved that contain thousands of constraints. Here we
will of course illustrate with a very small example.
4.1. Complex constraints
A particular constraint programming method is interval
constraints. Specific for interval constraints is that the unknowns are real numbers and that their domains are intervals. As an example consider the problem of finding the intersection points of the circle x2 + y 2 = 1 and the parabola
y = x2 . This means finding values for x and y such that
both relations are satisfied.
√Of course a student in secondary school will identify y as
( 5 − 1)/2, and x accordingly. The point here is to develop
a system that determines solutions directly from the set of
constraints as given, for a wide class of constraints.
Suppose that initially all we know is that x and y are
in [−∞, +∞]. Considering each of the two given relations
separately would already remove some values for x or y
from consideration. For example, it is clear from the constraint x2 + y 2 = 1 that x and y have to be in [−1, +1].
Whatever it is that allows us to reduce the original intervals
[−∞, +∞] to [−1, +1] we call a contraction operator. It
is an operator associated with the constraint that allows one
to make such an inference. In constraint programming, one
applies in turn the contraction operators associated with the
constraints until no more contraction results. The remaining
intervals for the variables then give all information about the
solutions of the problem that this method can give.
To make this method practical, contraction operators
have to be widely applicable and efficiently implementable.
Such operators have only been discovered for a relatively
small repertoire of primitive constraints. These include constraints with binary relations such as x ≤ y, x = y, and
class node {
public:
virtual void run() = 0;
};
class times : node {
private:
int mult; queue *in, *out;
public:
times(int Mult, queue *In,
queue *Out) {
mult = Mult;
in = In;
out = Out;
}
void run() {
if (in->empty()
|| out->full())
return;
out->add(mult*(in->next()));
in->remove();
}
5
y = x2 . There are also constraints with ternary relations
such as x + y = z and x × y = z. The primitive constraints
do not include the constraint x2 + y 2 = 1, as it does not
occur sufficiently often to justify its presence in a generalpurpose solver. This complex constraint is therefore decomposed into primitive constraints with the aid of auxiliary
variables. In this way the circle-and-parabola problem is
expressed by the system
x2 = x2 , y 2 = y2 , x2 + y2 = 1, x2 = y
concept, in this case a real number. It hides the representation, which is the description of a set of real numbers to
which the real number belongs. The set is restricted to the
form of an interval [lb,ub], which has the floating-point
numbers lb and ub as lower and upper bounds.
In the case of real numbers this distinction between the
abstract concept, which is public, and its hidden representation is especially valuable because most real numbers do
not even have a finite representation. So far, this perfectly
ordinary piece of object-oriented modelling has been, as far
as we know, the only practical way to directly compute with
mathematical models involving continuously varying quantities.
The zero-argument constructor creates a real of which
nothing is known. Accordingly it is represented by the interval [−∞, +∞], which is the set of all real numbers. If
more is known about the real number, then it is represented
by a smaller interval. In our example we need the real numbers 1 and 0.5, which are represented by the intervals only
containing these numbers and are created by the constructor
calls real _1(1.0,1.0) and real _0(0.5,0.5).
To model constraints, we note that their indentity is determined both by the relation and by the variables connected
by the relation. Each instance of a constraint needs to record
the variables of that instance. It is essential that different
constraints be able to share reals. This sharing is modelled
by members that are pointers to instances of class real.
The relation is represented by ensuring that constraints are
instances of the same class if, and only if, they have the
same relation. In this way the contraction operator for that
relation can be a method of that class. This method is
named shrinc1 . As all constraints need to go into the
same container, and as they all have a shrinc() method,
all constraint classes derive from an abstract class named
constraint. These considerations are embodied in the
code in Figure 3.
Now that we have the interface with the interval constraint machinery, let us use it to express the system (1) for
the circle-and-parabola problem; see Figure 4.
In line 3 the variables of system (1) are created. Lines
7 through 11 create the constraints of this system and place
them in a container. At the same time, this specifies the
interconnections of the constraints via the shared variables.
After the constraint system has been built, contraction
operators are applied. This is done in a simplistic way in
lines 13 through 18. In actual practice, it is done by means
of a constraint propagation algorithm. These algorithms
have a long and interesting history; see [1] for recent contributions and a history.
The result of the propagation is printed on lines 18 and
19 with the result: x ∈ [−1, 1] and y ∈ [0, 1]. This result
(1)
4.2. Contraction operators
To introduce contraction operators, let us look at an example of one in action: the one for primitive constraints of
the form x + y = z. Such a constraint expresses that the
variables are in a ternary relation that we call sum. If it is
known initially that x and y are in [0, 2] and that z is in
[3, 5], then it is clear that neither x nor y can be near 0 and
that z cannot be near 5. In fact the contraction operator reduces the intervals for x and y to [1, 2] and the one for z to
[3, 4].
This is optimal and can be computed in a few floatingpoint operations. One can solve complex systems of nonlinear equations and inequalities by first reducing them to
primitive constraints, and by then applying contraction operators until nothing changes. Typically, the remaining intervals are too large to be useful, in which case one interval
is split and the constraint propagation is repeated.
4.3. OO modelling of interval constraints
Let us now do a perfectly straightforward exercise in
object-oriented modelling in the spirit of [24]. Clearly,
“constraint” is an important noun, so justifies a class of
that name. The same holds for “variable”. However, this
is a dangerously ambiguous concept: in programming languages, in logic, and in calculus it means different things.
The role of x in the example can be stated precisely: it is
the name of a real number that we do not know. Thus x is
an instance of a class of which the instances represent real
numbers. This suggests the following class definition:
class real {
private:
FLPT lb, ub;
public:
real();
real(FLPT lb, FLPT ub);
};
FLPT is a generic floating-point number type; it could be of
single or double length. The class real is a classic example of object-oriented modelling. It stands for an abstract
1 It is intended to shrink intervals. Moreover, the class is part of a
system called “Sound High-Resolution Interval Numeric Calculator”.
6
is large enough to contain both solutions to (1). If one is
interested in one solution, say, with positive x, then one can
add a constraint as in line 10, using a constant created as
in line 4. In that case there is one solution, enclosed in a
narrow interval:
patterns found wide acceptance. However, to quote Erich
Gamma [6]
When we wrote “Design Patterns” we were excited about frameworks and forecast that there
would be lots of them in the near future. Time has
shown, however, that frameworks aren’t for everyone. Developing a good framework is hard and
requires a large up-front investment that doesn’t
always pay off.
x ∈ [0.7861513777574229, 0.7861513777574236]
y ∈ [0.6180339887498944, 0.6180339887498954].
We include this example because we believe that it shows
that a special-purpose constraint programming language
would not be able to specify the system (1) in a clearer or
more succinct way. At the same time, a special-purpose
language would probably not have any original insights on
how to do containers, nor on how to do control primitives,
so that these aspects of the program would be the same or
gratuitously different. Hence, this straightforward exercise
in object-oriented modelling of constraint programming is
probably as well as one can do in the form of text.
Originally, frameworks were intended to help specific applications, such as graphical user interfaces or document processing. Like design patterns, object-oriented frameworks
are defined as systems of customizable co-operating classes.
What one obtains as a result of modelling a programming
paradigm also answers to this description, but is not a design pattern. Instead, it is useful as a new way of discovering useful systems of customizable co-operating classes.
One of the most important advantages of design patterns
is as an aid to documentation: just because the pattern has
a name, the use of this name in the documentation speeds
up understanding of the part of the code concerned. The
object-oriented frameworks arising from programming paradigms have this same advantage.
5. Object-oriented frameworks
The foregoing examples suggest that a distinctive programming paradigm need not have a language of its own.
It is adequately supported by a suitable class library. Both
our examples include an abstract class and a generic algorithm expressed in terms of that class. Programming in the
paradigm requires one to define derived class suitable to the
application. In our dataflow example, each type of node
and its specific activity becomes a class derived from node.
Similarly, in the constraint example, each type of constraint
and its specific contraction operator becomes a class derived
from constraint.
But the library, with its abstract classes and generic algorithms is not enough. If it indeed embodies a distinctive
programming paradigm, it comes with a view of how to do
things, a mindset. This can be expressed by a set of tutorial
examples. The collection of these things is what is called a
framework by Gamma, Helm, Johnson, and Vlissides [7].
The trend towards frameworks rather than dedicated programming languages may have started with Puget’s approach to constraint programming. At first, constraint programming followed the conventional route with a dedicated programming language. It grew out of logic programming, which had Prolog as dedicated programming
language. Accordingly, constraint programming was implemented in various dialects of Prolog: Prolog II and III,
CHIP and its descendants, BNR Prolog, and Prolog IV.
However, Puget adopted Saraswat’s comprehensive view of
constraint programming [18] called Concurrent Constraint
Programming, dropped the concurrency, and based a C++
class library on it [16, 17].
Design Patterns and object-oriented frameworks seemd
equally promising when these ideas first arose. Design
6. Conclusions
So far there has been little scope in practice for the
Whitehead effect to ease software development. Applying
the paradigm that’s right for the application seemed to require switching to a different programming language. Practical concerns necessitate sticking to a single language.
This situation has changed since the single language is
object-oriented. Programming paradigms such as dataflow
or constraints, which once were thought to need a dedicated language, are easily modelled in an object-oriented
language. In fact, as easily as the textbook example of
object-oriented modelling of simple applications. Thus it
is not overly ambitious to model the ideal programming
paradigm, rather than the application. This modelling typically takes the form of an OO framework.
In this we can distinguish between minor and major paradigms. The minor ones are those that can be modelled in
one of the few languages most widely accepted in industry. We saw that dataflow and constraint programming fit in
this category. Järvi and Powell [10] implement partial function application in this way in C++. The fact that functional
programming can be characterized by functions being firstclass objects, suggests that functional programming is no
harder to do. It has turned out that it is and that it seems
to require a new programming language. Apparently, functional programming is a major paradigm. However, even
in this case OO programming has made a difference. The
7
new language turned out to be an extension of Java that was
needed anyway, independently of functional programming.
The question remains whether logic programming, so far
sequestered in Prolog and other dedicated languages, is a
major paradigm requiring a language extension or a minor
one that can easily be modelled in an existing OO language.
The work of combining Smalltalk with logic programming
by symbiotic reflection [25] seems to transcend the simplistic criterion used to distinguish between minor and major
paradigms.
It used to be that not only paradigms, but certain applications got their own programming languages. We have argued that this is something of the past. It was even the case
that one application, namely simulation, had several languages devoted to it. One of these was Simula, and the rest
is history.
Programming with Object-Oriented Languages Workshop, pages 43–60, Budapest, Hungary, June 2001.
[11] Luc Jaulin, Michel Kieffer, Olivier Didrit, and Éric
Walter. Applied Interval Analysis. Springer-Verlag,
2001.
[12] Gilles Kahn. The semantics of a simple language
for parallel processing. In Information Processing
74. North Holland, 1974. Proceeding of the IFIP
Congress.
[13] Gilles Kahn and David B. MacQueen. Coroutines and
networks of parallel processes. In B. Gilchrist, editor, Information Processing 77. North Holland, 1977.
Proceeding of the IFIP Congress.
[14] Kim Marriott and Peter J. Stuckey. Programming With
Constraints : An Introduction. The MIT Press, 1998.
References
[15] Martin Odersky and Philip Wadler. Pizza into Java:
Translating theory into practice. In Proc. 24th ACM
Symposium on Principles of Programming Languages
(POPL), 1997.
[1] K.R. Apt. The essence of constraint propagation. Theoretical Computer Science, 221(1-2):179–210, 1999.
[2] Gilad Bracha, Martin Odersky, David Stoutamire, and
Philip Wadler. Making the future safe for the past:
Adding genericity to the Java programming language.
In OOPSLA 98, 1998.
[16] Jean-François Puget. A C++ implementation of CLP.
In SPICIS94, 1994.
[17] Jean-François Puget and Michel Leconte. Beyond the
glass box: Constraints as objects. In Proc. Int. Logic
Programming Symposium. MIT Press, 1995.
[3] Kei Davis, Yannis Smaragdakis, and Jörg Striegnitz,
editors. Multiparadigm Programming with ObjectOriented Languages. John van Neumann Institute for
Computing, 2001. ISBN 3-00-007968-8.
[18] Vijay A. Saraswat. Concurrent Constraint Programming. MIT Press, 1993.
[4] E.W. Dijkstra. A Discipline of Programming. Prentice
Hall, 1976.
[5] Matthias Felleisen and Daniel P. Friedman. A Little
Java, a Few Patterns. MIT Press, 1997.
[19] Mary Shaw and David Garlan. Software Architecture:
Perspectives on an Emerging Discipline. PrenticeHall, 1996.
[6] Erich Gamma. Design patterns — ten years later. In
Manfred Broy and Ernest Denert, editors, Software Pioneers, pages 688–700. Springer-Verlag, 2002.
[20] Joseph E. Stoy. Denotational Semantics: The ScottStrachey approach to Programming Language Theory.
MIT Press, 1977.
[7] Erich Gamma, Richard Helm, Ralph Johnson, and
John Vlissides. Design Patterns. Addison-Wesley,
1994.
[21] A.H. Veen. Dataflow machine architecture. Computing Surveys, pages 365–396, 1986.
Structured Systems
[22] William W. Wadge and Edward A. Ashcroft. Lucid,
the Dataflow Language. Academic Press, 1985.
[9] Michael Jackson and Graham Twaddle. Business
Process Implementation: Building Workflow Systems.
Addison-Wesley, 1997.
[23] D. Waltz. Understanding line drawings in scenes with
shadows. In Patrick Henry Winston, editor, The Psychology of Computer Vision, pages 19–91. McGrawHill, 1975.
[10] J. Järvi and G. Powell. Side effects and partial function application in C++. In Kei Davis, Yannis Smaragdakis, and Jörg Striegnitz, editors, Multiparadigm
[24] Rebecca Wirfs-Brock, Brian Wilkerson, and Lauren Wiener. Designing Object-Oriented Software.
Prentice-Hall, 1990.
[8] Chris Gane and Trish Sarson.
Analysis. Prentice-Hall, 1979.
8
[25] Roel Wuyts and Stéphane Ducasse. Symbiotic reflection between and object-oriented and a logic programming language. In Kei Davis, Yannis Smaragdakis, and Jörg Striegnitz, editors, Multiparadigm
Programming with Object-Oriented Languages Workshop, pages 81–96, Budapest, Hungary, June 2001.
class constraint {
public:
virtual bool shrinc() = 0;
virtual ˜constraint {}
// Applies contraction operator.
// Returns false iff
// an empty interval results.
};
// constraint is x <= y
class leq: public constraint {
real *x, *y;
public:
leq(real *x, real *y);
bool shrinc();
};
// constraint is x == y
class eq: public constraint {
real *x, *y;
public:
eq(real *x, real *y);
bool shrinc();
};
// constraint is x+y = z
class sum: public constraint {
real *x, *y, *z;
public:
sum(real *x, real *y, real *z);
bool shrinc();
};
// constraint xˆ2=y
class square: public constraint {
real *x, *y;
public:
square(real *x, real *y);
bool shrinc();
};
Figure 3. Definitions for the constraint
classes.
9
00: int main () {
01: //solve {xˆ2=x2, yˆ2=y2, x2+y2=1, y=x2}
02: //create variables:
03: real x, y, x2, y2, _1(1.0,1.0);
04: //real _0(0.5,0.5);
05:
06: //create constraint system:
07: constraint* array[] = {
08:
new square(&x,&x2), new square(&y,&y2),
09:
new sum(&x2,&y2,&_1), new eq(&y,&x2)
10:
//, new leq(&_0,&x)
11:
};
12: //propagate:
13: const int MAX = 1000;
14: int size =
sizeof(array)/sizeof(constraint*);
15: for (int i=0; i < MAX; i++)
16:
for (int j=0; j < size; j++)
17:
array[j].shrinc();
18: cout << "x: " << x << endl;
19: cout << "y: " << y << endl;
20: for (int j=0; j < size; j++)
21:
delete array[j];
22: }
Figure 4. C++ code for finding the intersection of a circle with a parabola using the interval constraints method.
10
| 6cs.PL |
DuctTeip: An efficient programming model for distributed
task based parallel computingI
Afshin Zafaria , Elisabeth Larssona,∗, Martin Tilleniusa
arXiv:1801.03578v1 [cs.DC] 10 Jan 2018
a
Uppsala University, Department of Information Technology, Box 337, SE-751 05
Uppsala, Sweden
Abstract
Current high-performance computer systems used for scientific computing
typically combine shared memory computational nodes in a distributed
memory environment. Extracting high performance from these complex
systems requires tailored approaches. Task based parallel programming has
been successful both in simplifying the programming and in exploiting the
available hardware parallelism for shared memory systems. In this paper we
focus on how to extend task parallel programming to distributed memory
systems. We use a hierarchical decomposition of tasks and data in order
to accommodate the different levels of hardware. We test the proposed
programming model on two different applications, a Cholesky factorization,
and a solver for the Shallow Water Equations. We also compare the performance of our implementation with that of other frameworks for distributed
task parallel programming, and show that it is competitive.
Keywords: Task-based parallel programming, distributed memory system,
high performance computing, scientific computing, hierarchical
decomposition, data versioning
2000 MSC: 65Y05, 65Y10, 68Q10
I
This work was supported in part by the Swedish Research Council and carried out
within the Linnaeus centre of excellence UPMARC, Uppsala Programming for Multicore
Architectures Research Center.
∗
Corresponding author
Email addresses: afshinzafari@gmail.com (Afshin Zafari),
elisabeth.larsson@it.uu.se (Elisabeth Larsson), martin.tillenius@gmail.com
(Martin Tillenius)
Preprint submitted to Elsevier
January 14, 2018
1. Introduction
Task based parallel programming models [1] have, over the last decade,
been developed to meet the contemporary challenge of programming for
rapidly evolving multicore hardware. In this programming paradigm, the
work for the programmer is reduced to expressing an algorithm in terms
of tasks and their accesses to shared data, while scheduling and execution
of tasks in parallel are managed by an application independent run-time
system. The run-time system is also responsible for ensuring that data
dependencies are respected. The dependencies can, e.g., be deduced by the
run-time system from user annotations of data accesses.
While moving away from optimized software for a specific architecture,
the role of the run-time system is to extract as high performance as possible
from a wide range of multicore CPUs. Task based programming models
provide some of the key components of multicore performance. Task sizes
can be selected (when the application allows it) such that their data fits into
the lowest level cache for efficient use of the memory hierarchy. Fine grained
tasks also provide a high level of parallelism that can supply high numbers
of threads with work. Local synchronization points between tasks (as opposed to global barriers), allow for asynchronous task execution, which can
compensate for performance variations between threads due to, e.g., variable task sizes, resource contention, non-uniform memory access (NUMA)
effects, or interference from simultaneously running processes in the kernel,
run-time or other applications.
A large number of frameworks for shared memory task parallel programming that address the common issues of portability of software, ease
of programming, and performance on multicore and possibly heterogeneous
computer systems have been developed. The type of parallelism that is provided varies from purely fork-join as in Cilk [2] and Wool [3], fork-join that
allows data dependencies between sibling tasks, as implemented in Swan [4],
OpenMP [5] version 4.0 or higher, and XKaapi [6], to directed acyclic graph
(DAG) based, where tasks are organized as nodes in a graph, where edges
represent must-execute-before dependencies, as in Intel Threading Building
Blocks [7], OmpSs [8], and StarPU [9].
In this paper, we develop a distributed task programming model, and
implement it in the distributed task parallel framework DuctTeip . We use
a hybrid parallel implementation with MPI for the distributed layer and
Pthreads for the shared memory layer. We build on the shared memory
2
library SuperGlue1 [10], developed by one of the authors. The reason for
this choice is based on both the performance of SuperGlue and the suitability
of the programming model for an extension to distributed memory.
In [10], SuperGlue was shown to perform better than ten of the other
shared memory frameworks available at that time. A key component for the
performance is the dependency management through data versioning. The
resulting programming model is very similar to the DAG based models, but
has some additional flexibility regarding dependencies.
DuctTeip has been developed, not primarily to compete with other frameworks, but to perform research on different ways to promote both flexibility
and performance in the distributed setting. Understanding which trade-offs
can be made and what their consequences are is relevant for the community that develops task parallel run-times, and accordingly, in the long term
perspective, also for the users of the frameworks.
When we evaluate the performance of DuctTeip, we compare with the
three major task frameworks/run-times with support for distributed memory architectures that implement a similar programming model: StarPU [11],
Cluster OmpSS [12], and PaRSEC [13]. We introduce each in some detail
below.
The MPI version of StarPU [11] maps the task DAG to the computational nodes. The edges in the DAG (dependencies) that cross the boundary
between two computational nodes are mapped to MPI-based data transfer
routines. Any MPI process can create and submit tasks, and the global task
graph is known to all. The communication calls are initiated at task submission time, which may lead to high memory consumption if too many receive
buffers are allocated in advance. This issue is investigated and handled
through pausing task submission in [14].
Cluster OmpSs [12] employs centralized task creation and submission
with one master process that submits tasks to all other processes. A task
can generate child tasks at execution time. These child tasks can have
dependencies to their sibling tasks, but not to other tasks.
PaRSEC [13, 15] provides a highly efficient platform for distributed task
based parallel execution. An annotated user code is here first (automatically) translated into a parametrized DAG representation. In particular,
this representation can generate information about the predecessors and
successors of each task. There is no need to store or expand the full graph,
instead task specific information can be examined locally through queries on
1
https://github.com/tillenius/superglue
3
the DAG representation. Scheduling and task queues are fully NUMA aware
and exploit the memory hierarchy to promote data locality. Communication
is asynchronous and managed by a dedicated thread on each node.
In the DuctTeip programming model, we do not assume any pre-knowledge
regarding the task graph. Tasks can be generated dynamically, and tasks
can be submitted by other tasks. Dependencies are of the DAG type, but
with an extension to commutative dependencies through the data versioning
model. Dependencies between tasks that are children of other tasks are not
limited to siblings, but can be to other tasks on that level, which will be
described in more detail later. Furthermore, the data that tasks work on
can in principle have any format as protection of shared data is managed
indirectly through data handles. An early description of the DuctTeip programming model is available in [16]. In this paper, we develop and evaluate
two main ideas within DuctTeip for distributed task parallel programming
• a hierarchical task and data model based on data versioning, which
has several benefits regarding flexibility, utilization of the memory
hierarchy, and network loading,
• an efficient communication model, which avoids redundant messaging,
and enables task and data migration (for future work on dynamic load
balancing).
Distributed task frameworks that are based on, to some extent, other
programming models are, e.g., Charm++ [17, 18], which implements messagedriven execution, where the arrival of a data object prompts the execution
of a task, and Chunks and Tasks [19], where chunks are immutable data
objects that can be split hierarchically into smaller chunks, and tasks are
objects that work on chunks, and can generate new tasks.
The paper is organized in the following way: In Section 2 we briefly introduce dependency management through data versioning. Sections 3 and 4
describe the task and data model, and the communication model. In Section 5 we discuss how to support programmers in using the DuctTeip model
efficiently. In Section 6 the performance and scalability of DuctTeip are
investigated and compared to that of StarPU, Cluster OmpSs, PaRSEC,
and also ScaLAPACK. Finally, conclusions based on the results are made in
Section 7.
2. Tracking dependencies through data versioning
A detailed description of the SuperGlue programming model including
data versioning can be found in [10]. Here we give brief account to serve as
4
a basis for the discussion in the following sections of the distributed case.
In this model, shared data is protected by a construct that we call a
handle. Since the handle is not physically connected to a memory location,
it provides flexibility to the programmer in the sense that we can protect
any type of shared data or shared resource [10, 20].
Each data handle is equipped with a version counter, which allows the
run-time system to order data accesses (at execution time) such that data
dependencies are respected. The versioning system is initialized at task
submission time based on the assumption that tasks are submitted in a
consistent sequential order.
At task submission time, the run-time system counts the (future) accesses to each data handle, and for each task, records the required version
number of each of the accessed data handles. How the required version is
incremented depends on the type of access. Tasks that require the same
version number can be reordered.
During execution, the run-time version number of each handle is incremented after each access. Technically, this operation is performed at the
completion of the execution of a task. In this way, the run-time system can
compare the required version numbers for a certain task with the current
run-time version numbers of the involved handles in order to determine if
the task is ready to run.
Figure 1 provides a small example of a task graph and the corresponding
versions. The version of each handle starts at 0. The two read tasks can
run immediately and in any order, but the first task that modifies the data
must wait until both reads have finished. Hence, the modify task requires
version 2. The three adds can run in any order, and when all of them are
finished, the final modify task requiring version 6 is allowed run.
When the arguments of a task are examined to determine if the task is
ready to run, and an argument that is not ready is encountered, the task is
placed in a queue for that particular data version. This means that as soon
as the data is ready, the run-time system knows which tasks to wake up
either for execution or for examining of the next data in the argument list.
In this way, the need for traversing long lists of waiting tasks is reduced,
and the run-time can deploy the ready tasks at the thread where the data
is locally cached. This improves the data locality of the execution, and also
reduces the need for a particular scheduling strategy. In SuperGlue load
balancing is achieved dynamically through task stealing.
Scheduling to improve data locality and minimizing NUMA effects in
shared memory task parallel execution is an active research area [6, 21, 22,
23, 24, 25, 26, 27] and can also be coupled to energy considerations [28, 29].
5
read
read
modify
add
add
add
modify
Figure 1: Left: A small task graph where all accesses (the type is indicated for each
task) are assumed to be to the same shared data. Dashed lines indicate a commutative
dependency (any order, but one at a time). Right: The version v required by the task to
run, and the run-time version r produced by the task given in the format v : r. A range
of numbers indicates that any number in the range can be realized in an execution.
3. The hierarchical DuctTeip task and data model
We exploit a hierarchical task and data model to extract performance
from the hardware in the distributed task parallel case. The corresponding
hierarchical target architecture is a distributed memory system (cluster) of
shared memory computational nodes. A key idea in this approach is to use
large data partitions for communication and small data partitions for computations. Communicating a single large data block allows for reducing the
overhead of the communication compared with multiple smaller messages.
Then locally, further decomposing the large data into smaller parts increases
the cache locality of the data in the local computations.
To make the description of the hierarchical model general, we introduce
the concept of levels, where level 0 is the top level (root), and level L is the
finest level (leaves). The typical data at level 0 consists (symbolically) of
the unpartitioned data structure and tasks submitted at level ` works on
level ` + 1 data. We also allow for data to be allocated and created at levels
` > 0.
We exemplify with a basic three level configuration, where the distributed memory framework DuctTeip is used at level 0 (working on level
1 data), and the shared memory library SuperGlue is employed at level 1
(working on level 2 data). For an illustration, see Figure 2. In the discussion
of data and tasks below, we describe the general case with L data levels.
In the following subsections, we explain the hierarchical approach for
defining data and tasks and how it affects the dependency tracking and the
6
Data
Algorithm
level 0
Task submission
Tasks
level 1
DuctTeip tasks
level 2
SuperGlue tasks
Figure 2: Schematic illustration of hierarchical data and tasks in the currently implemented three level model.
task execution in the context of distributed parallel programming.
3.1. The hierarchical data structures
Conceptually, we consider general data types such as dense matrices,
sparse matrices, and tree structures. Examples are given in terms of dense
matrices below to provide clear illustrations. For each type of data that
can be represented, a parametrized splitting scheme needs to be provided.
The parametrization allows for tuning partition sizes with respect to the
numbers of nodes and cores as well as the local cache sizes. For a square
matrix, a standard splitting is to divide a matrix block A`−1 at level ` − 1
into n` × n` blocks A`−1 (i, j), i, j = 1, . . . , n` at level `.
As indicated, we also need an indexing method associated with the splitting. The algorithm at level ` − 1 submits tasks formulated in terms of the
data partitions A`−1 (i1 , . . . , id ), where d is the number of dimensions of the
splitting and the indexing.
When a shared data structure A0 within the parallel framework is initialized at level 0, we do not want to allocate the whole structure anywhere. The
distributed level is in our case level ` = 1, and this is where the data is associated with a node and allocated locally. The data ownership is determined
from a process grid topology and a distribution scheme. An example of how
a block-cyclic distribution of a dense matrix interacts with the process grid
is shown in Figure 3.
For efficient memory access during the computations as well as efficient
communication of data, we want partitions at each level, if possible, to be
stored in contiguous memory locations. For more exotic data structures,
this means that packing and unpacking routines may be required at the
distributed level(s) in addition to the splitting scheme. For (dense) matrices,
defining an efficient storage scheme is relatively straightforward. Figure 4
7
P0 P1 P2
P0 P1 P2 P0
P0 P1 P2
P3 P4 P5
P3 P4 P5 P3
P0 P1 P2 P0
P0 P1 P2 P0
P0 P1 P2 P0
P3 P4 P5 P3
P0 P1 P2 P0
(a) Block-cyclic ownership.
P0 P1 P2 P0
(b) Column-cyclic ownership.
Figure 3: Two different process grids (p, q) = (2, 3) and (p, q) = (1, 3) for the processes
Pi , and the resulting block cyclic distributions for a 4 × 4 block matrix.
shows an example of partitioning a matrix into two hierarchical levels. The
matrix elements are ordered such that blocks (tiles) at each level are stored
in contiguous memory blocks.
As described in Section 2, accesses to shared data structures are managed
through data handles. When a DuctTeip data structure is initiated, handles
are created for the data partitions at each level. In the same way as in the
shared memory case, the handles track the versions of the corresponding
data and host request queues for particular versions of the data.
3.2. Hierarchical task generation and task execution
At level 0, the code is traversed by all participating computational nodes.
Each task submission statement is inspected with respect to task ownership.
The ownership of a task is decided based on the location of its output data.
If the task is locally owned, it is generated and submitted to the level 0
run-time system. In the case of several outputs located in different computational nodes, a rule (e.g., first output) is implemented. Due to the
hierarchical structure, the number of tasks submitted at the highest level
is smaller than in a corresponding one-level approach, thereby reducing the
overhead from the global inspection of the tasks in the initial task generation
phase.
A task at level ` contains task submission statements for tasks at level
` + 1. When the level ` task is ready to run, the corresponding level ` + 1
tasks are generated and submitted to the level ` + 1 run-time system. Note
that even though tasks at levels ` > 0 are submitted by different nodes,
the order is globally preserved because these are submitted at the execution
time of the parent task, which ensures that the dependencies at the parent
level are respected and tasks are submitted in a consistent sequential order
at each level. This is the mechanism that allows dependencies between
any tasks scheduled by the same run-time, not only sibling tasks. (At the
8
A0 (i, j)
a00
A0 (0, 0)
a20
Single message in
communication
a22
A0 (i, j)
...
...
a00 a10
A0 (3, 0)
a20
a22
... Memory
alignment
A0 (3, 3)
Small blocks in
computations
Figure 4: An example of hierarchical data decomposition. A matrix is partitioned into 4×4
large blocks at level 1, each of which contains 3×3 small blocks at level 2. The elements
of each block at each level are linearly indexed in contiguous memory, as indicated by the
blue line traversing one sub block at a time. The notation amn is used for partitions of
the level 1 data.
shared memory level, there is one SuperGlue run-time deployed in each
computational node.)
The level ` and level `+1 tasks have a parent-child relationship. The task
completion of the parent task depends on the completion of all of its children
tasks. However, the run-time system handling the parent tasks can start the
next ready parent task in the queue without waiting for completion of the
previous one, as all ready tasks already have their dependencies fulfilled.
The tasks working on the last level data, contain the actual computational kernels. A benefit that follows from the hierarchical task submission
is that the submission of lower level tasks is paced in relation to the speed of
execution due to the synchronization with the parent tasks. This prevents
the number of tasks in the work queues from growing in an uncontrolled
way.
When tasks are placed according to the location of their output data as
described above, the resulting load balance at the distributed level is static.
If the data distribution scheme is not well chosen, this can result in severe
performance degradation in the overall execution time of an algorithm. Distributed dynamic load balancing/scheduling [30, 31, 32] is challenging as
latency is much higher than for the shared memory case, and data movements both before and after the execution of a task must be considered.
Another issue that becomes increasingly relevant as the scale of the dis9
tributed system increases, but which we will not discuss further here, is
fault tolerant execution, see [33, 34, 35, 36, 37].
3.3. The data allocation scheme
DuctTeip creates a memory pool at each computational node at startup time before the task submission begins. By using a memory pool, the
overall time spent for memory allocation and deallocation is reduced as well
as shifted out of the task execution phase. Within the memory pool, space
is allocated for all the locally owned data and for data that will be received
from other nodes during the task execution phase. The memory pool is
dynamically extended when needed until the node runs out of memory. The
pooling mechanism can be switched off for any individual data, and then
allocation is implemented by the application code instead.
Each data item in the memory pool resides in a contiguous memory block
M that contains a header part (of size H) and a content part (of size C). The
header part holds meta-data including the data handle with current version
and memory address of the data, data size, and data ownership. A local task
uses the content part directly in its computations. When a task at another
node needs the data, the DuctTeip run-time initiates a communication call
for the header address and size H + C. The data is received in a buffer
within the memory pool of the remote node. The same memory area is used
for further partitioning the data into smaller blocks, hence the received data
can immediately be used by tasks at deeper levels (see Figure 4).
When using data versioning in the distributed case, it is possible for two
or more versions of the same data to exist at the same time. A new version of
a data may arrive while an older one is still involved in a computation. Data
that is received by a node gets a new place in the memory pool. Therefore,
memory must also be released at run-time when a data version is no longer
needed. When a newer version is accessed by a task, it means that all older
versions are obsolete and the memory slots of these can be released to the
pool.
Precisely how memory is allocated has implications on the NUMA effects
during the execution. When the memory pool is allocated, the allocation
is by default distributed evenly (interleaved in a round-robin mode by the
NUMA controller) among the NUMA nodes (sockets/cores) of the computational node. If we instead allocate space for each data individually, the
same NUMA allocation policy is applied per block. This is especially bad
for the smaller data blocks used at the shared memory level. In the experiments, we can see a significant performance improvement when the memory
pool is used, and finer partitions are derived in place as offsets from the
10
parent data starting address compared with individual allocations. There
are further possibilities to optimize the NUMA behavior that we have not
explored so far. Using the hwloc library [38], fine grained control over the
association of memory with NUMA locations can be implemented.
4. The distributed communication model
In a distributed task based programming model, it is expected that the
necessary communication should be handled transparently by the run-time
system. One drawback of this model can be that efficient implementations
of global communication patterns like all-to-all or all-reduce are not immediately available. In the task based view, communication is pairwise and per
task. A mixed programming model could be allowed, where user MPI code
is interleaved with task-based regions. Here, we describe the constructs used
in DuctTeip for the point wise communication.
4.1. The listener concept
The object that we call listener is used for managing communication at
the distributed level(s). At task generation time, when data accesses are
registered, the location of the data is also examined. In cases where a task
requires a specific version of a remote data, a listener is generated by (or
sent to) the data owner. This means that the remote node knows that the
data will be needed later and can send it as soon as the required version
becomes available. That is, the communication is initiated at the time the
data should be sent, not earlier. In this way, we do not need to have a long
list of outstanding communication requests.
At execution time, when a data handle is upgraded to a new version,
tasks that are waiting for that version are examined for execution. Additionally, the listeners waiting for that data version are processed, and the
new data is sent to all nodes that have one or more listeners in the queue.
Duplicate listeners can occur if several tasks hosted in the same remote
node need the data, but then only one message is sent to avoid unnecessary
communication traffic.
Sending data corresponds to remote data accesses, and the data version
needs to be updated in a consistent way. If the remote accesses are of the
read type, the version can be upgraded immediately, while if the access involves modifying the data, the version cannot be changed until the modified
data has been returned from the modifying node and received by the data
owner. In the basic DuctTeip model, the output data is always located in
11
the node that is running the task, and then it is enough to count the number of listeners that are processed and upgrade the version accordingly. An
example of the events associated with one listener is given below
• During task submission, task t1 owned by process P1 , that needs a
remote data d2 at process P2 is encountered.
• At P2 , a listener is created for the version v of d2 that is required by
t1 .
• During task execution, version v of d2 becomes ready at P2 .
• The listeners waiting for version v are examined, and a message containing d2 is sent from P2 to P1 .
• The version of d2 is incremented, and the listener(s) are deallocated.
• The message with d2 is received by P1 . Task t1 waiting for version v
of d2 is notified.
• If all other data dependencies are fulfilled, task t1 is ready to run.
A benefit derived from having listeners is that it enables task and data
migration/stealing. When a data is moved, listeners waiting for that data are
moved along. Moving tasks can be done in different ways. In our ongoing
work on dynamic load balancing, only ready tasks are moved, and then
together with copies of their input data. It is assumed that a task that has
been moved still returns the output data to the original data location.
4.2. Asynchronous communication and threading models
For the implementation of the communication model, the main objective is to overlap communication and computations. We also want to avoid
stalling of the communication due to receive requests that are not fulfilled.
An attractive solution would be to use multithreaded MPI, and let one
thread manage send requests and one thread manage the receive state. Using more than one thread for MPI calls requires that the MPI library used
is fully compliant with the MPI-2 specification of threading models. In particular MPI THREAD MULTIPLE, in which many threads can call MPI routines
independently.
If the MPI implementation is not fully compliant, the request for using MPI THREAD MULTIPLE may instead result in one of two other threading
models. In the serialized threading model, MPI THREAD SERIALIZED, multiple threads can call MPI routines, but only one thread at a time is allowed to
make progress. The others are blocked until the working thread completes.
12
In the funnel threading model, MPI THREAD FUNNEL, instead only one thread
is allowed to call MPI routines.
When the current version of DuctTeip was implemented, the latest versions of the most common MPI implementations were not fully compliant
with respect to multi-threading. This means that in practice, only one
thread could be used for communication. We could have chosen to split
task management and communication over two threads, but since the task
management and the communication are tightly connected, we decided to
use only one thread per MPI process. To achieve overlapping of the task handling and the communication, all send and receive MPI calls are therefore in
asynchronous non-blocking mode, using MPI routines that return immediately. The state of the requests is also checked in non-blocking mode, which
means that a periodic polling is needed.
The MPI process works in a loop until termination, where the send and
receive states, the task states, and the listeners are checked. The MPI thread
is co-located with the SuperGlue administration thread, that handles task
submission, at the same core. In this way, computations and administration are interleaved at this core, while the threads at the other cores are
left without interference. How well the asynchronous communication works
also depends on how well the MPI threading model is implemented. In the
experiments reported later, the communication is working well in this way.
The listener based communication, where data is sent at the time it becomes available helps to reduce the risk for message flooding as requests are
separated over time.
4.3. Hybrid approaches versus pure MPI approaches
We have chosen to use a hybrid MPI–Pthreads approach in DuctTeip. In
this section, we describe the main reasons for this choice from a performance
perspective. Whether using a hybrid or a pure MPI approach, the parallel
program runs on p distributed computational nodes, each with n cores. In
the pure MPI case, pn instances of the program run concurrently, with n
instances within each node. In the hybrid MPI–Pthreads case instead p
instances of the program run on the computational nodes, and within every
node the work is distributed over n threads.
Using only MPI instances of a program to exploit the parallelism provided by the underlying hardware degrades the performance in comparison
with a hybrid approach for the following reasons: First, the memory allocated for each instance of an MPI process within a node is private to that
process and can only be accessed by other processes via the main memory of the node, using techniques like inter process communication (IPC).
13
Therefore, communicating even cached data between two processes running
on the same CPU, and hence potentially sharing the same cache, requires
a round-trip for the data to the main memory. Secondly, even if the communication within one node can be performed efficiently, the traffic between
nodes is likely to increase as each of the processes within one node may need
to communicate with each of the processes running at a remote node. In the
hierarchical hybrid case, communication requests are instead aggregated at
the node level.
Additionally, in the pure MPI case, the amount of memory that needs to
be allocated by one node to receive messages becomes unnecessarily large.
For example, if processes P1 and P2 running on node i own data A1 and
A2 respectively, and communicate these data between each other, then we
need memory space for each of the two items at node i. When instead using
a hybrid approach, no extra space is needed for communication within the
node. This means that in a pure MPI mode, we run out of memory faster
when increasing the node local work size.
5. Creating and tuning applications in the DuctTeip framework
Even if using a programming framework relieves the programmer from
many of the more technical aspects of parallel programming, there are still
things that the programmer can influence both regarding the algorithm and
the program settings that affect the performance of the parallel software.
In this section, we start by discussing in general terms how to express an
algorithm in a form that makes it suitable for task parallel execution. Then
we take a closer look at the two algorithms that are used in the numerical
experiments in this paper, Cholesky factorization and a solver for the shallow
water equations. Finally, we discuss how a simple simulation tool can give
insights concerning the expected performance.
5.1. Taskifying algorithms
As explained in Section 3, an algorithm in the DuctTeip or SuperGlue frameworks is working on (hierarchically) partitioned data. The first
step in preparing an algorithm for task-based parallelization is to decide
how to partition the shared data. For matrices, blocks or slices are natural
partitions. The algorithm then needs to be expressed as working on these
data partitions. This could for example correspond to a transition from an
element wise Cholesky decomposition algorithm to a block Cholesky factorization.
14
A scalar control variable or a global vector that local results are gathered
into can also be shared variables. Common for all shared data is that it is
protected by a handle, which is used for managing accesses to the data.
Operations that can become tasks have a clearly defined interface to the
shared data in terms of inputs and outputs, and do not modify any other
shared data (no global variables). In the taskified code, such operations
are replaced by task submission statements. The operations provide the
computational kernel of the tasks. APIs for defining and submitting tasks
are provided by the framework.
In the current implementation, the APIs for task submission and task
definition with respect to the SuperGlue and DuctTeip run-time systems
have some small differences, and task definitions need to be provided for each
participating run-time. Ideally, they should be the same in a general L-level
framework. A unified interface that allows different run-times to cooperate
transparently using a single application code is described, implemented and
shown to perform well in [39].
Since the DuctTeip framework uses the same logic, concepts and abstracts as the SuperGlue framework does, any algorithm that can be implemented using SuperGlue can similarly be implemented using DuctTeip. For
some examples of algorithms implemented with SuperGlue, see [10, 40, 41,
42].
To learn more about the details of the framework implementations, as
well as the application examples, visit the SuperGlue framework source
repository at https://github.com/tillenius/superglue/ and the DuctTeip framework source repository at https://github.com/afshin-zafari/
DuctTeip.
5.2. Task parallel hierarchical Cholesky factorization
The data in the Cholesky factorization algorithm is a dense symmetric
matrix. Due to the symmetry, only the upper/lower part of the matrix
is stored. The factorization is performed in place, such that the original
matrix is overwritten by the factorization. The level 0 data is the whole
matrix A0 . At level 1, the matrix is subdivided into n × n blocks. A natural
formulation of an algorithm that works on the level 1 data is a block Cholesky
factorization. The algorithm is given below in terms of calls to LAPACK
subroutines.
for ( int i = 0; i < n ; i ++) {
for ( int j = 0; j < i ; j ++) {
syrk ( A (i , j ) ,A (i , i ) ) ;
15
// Aii = Aij ATij
for ( int k = i +1; k < n ; k ++) {
gemm ( A (k , j ) ,A (i , j ) ,A (k , i ) ) ; // Aki = Akj Aij
}
}
potrf ( A (i , i ) ) ;
for ( int j = i +1; j < n ; j ++) {
trsm ( A (i , i ) ,A (j , i ) ) ;
}
// Aii = LLT
// Aji = A−1
ii Aji
}
The block Cholesky factorization contains four types of tasks syrk, gemm,
potrf, and trsm. To continue to a hierarchical algorithm, each of the level
1 blocks are in turn subdivided into m × m blocks, and the algorithm of
each task is expressed in terms of the level 2 data. The potrf operation is
a Cholesky factorization, and is hence subdivided in the same way as the
original algorithm, while the gemm operation is a matrix-matrix multiplication subdivided into gemm operations on the level 2 blocks. Figure 5 shows
an example of a level 0 task graph for a 4 × 4 block Cholesky factorization
and some of the level 1 subgraphs for a division into 3 × 3 level 2 blocks.
The complexity of the dependencies between tasks and data in this factorization illuminates the flexibility of the framework in handling algorithms
that are not straightforward to parallelize. In our implementation, we have
not given priority to the potrf tasks which constitute the critical path.
Priority can be included in the run-time scheduling approach to improve
performance for algorithms where the critical path is known.
5.3. A task parallel solver for the shallow water equations
Dense linear algebra operations as the Cholesky factorization are common test cases for parallel implementation. They have certain advantages
such as a large proportion of BLAS level 3 operations, which means that
they are not so sensitive to the bandwidth limitations of modern hardware,
they have regular data structures that allow for memory to be traversed linearly, and with a proper organization, the blocks are well suited for efficient
communication.
Therefore, we wanted to include also a different type of application example. We have chosen a PDE solver for the shallow water equations on
the surface of the Earth. The numerical method, based on radial basis
function scattered node stencil approximations, is introduced in [43], and
was parallelized using an MPI-enabled version of the SuperGlue framework
16
A00
A30
A10
A31
A11
Aij
A20
a00
a10
a20
a01 a02
a11 a12
a21 a22
a01
A33
A32
A21
A22
a11
a21
a02
a12
a22
a02
a12
a22
A11
A31
Ajj
A21
a00
A33
A32
A22
a20
a22
a10
a21
a11
A22
a11
a21
A32
a22
a22
A33
A33
SYRK
POTRF
TRSM
GEMM
Figure 5: To the left, a level 0 task graph of a 4 × 4 block Cholesky factorization of a
matrix A is shown. The labels of the nodes, Aij , are the output data of the tasks. An
arrow between tasks indicates a must execute before dependency, while the dashed lines
indicate that the tasks can be reordered, but cannot execute concurrently. To the right,
two examples of level 1 task graphs (a block TRSM and a block POTRF) for a 3 × 3
splitting of the level 1 data are shown.
(not using hierarchical tasks and data) in [41]. Here we instead use the
hierarchical programming model of the DuctTeip framework.
The solver algorithm is data parallel in the sense that there is an explicit time stepping scheme, and each new value in the vector of unknowns
can be computed based on data from the previous time step. However, the
time stepping scheme is inherently sequential, which means that the potential parallelism has to be found in the spatial operator. The time-stepping
scheme that is used is the fourth order Runge-Kutta scheme, and one time
step has the following structure:
f(F1,H);
add(H1, H , 0.5*dt, F1);
f(F2,H1);
add(H2, H , 0.5*dt, F2);
f(F3,H2);
add(H3, H ,
dt, F3);
f(F4,H3);
step(H,F1,F2,F3,F4);
//
//
//
//
//
//
//
//
F1
H1
F2
H2
F3
H3
F4
H
17
=
=
=
=
=
=
=
=
f(H)
H + 0.5*dt*F1
f(H1)
H + 0.5*dt*F2
f(H2)
H +
dt*F3
f(H3)
H + dt/6*(F1+2*F2+2*F3+F4)
The different sub steps within each time step also has a sequential dependency structure. The DuctTeip tasks for the time steps are submitted in
two layers. There is one task for each time step, that in turn submits the
tasks within that particular time step. If all time-steps were submitted in
the beginning of the execution, there would be a problem with task overflow
since the number of time steps in these simulations typically is very large.
Therefore, the submission of time step tasks is controlled by a handle to a
dummy data object that is owned by all processes. Each time step task has
a modify access to that handle. A few time step tasks are submitted initially
to fill the task pipelines, and after that, a new time step task is submitted
whenever a previous time step task finishes.
The computationally heavy operations within a time step are the calls
to the right hand side function f(F,H), that evaluates the spatial part of the
PDE operator. From the programming perspective, it consists of a number
of sparse matrix-vector multiplications between differentiation matrices and
instances of the vector of unknowns H. The data distribution of the matrix
and vectors is shown in Figure 6. The performance of the parallel impleP0
P1
P2
←
P3
Figure 6: Left: One block row of each data structure is assigned to each process Pi . The
shaded blocks illustrate input and output in the block matrix multiplication. Right: An
actual matrix structure together with a partitioning into level 1 blocks is shown. Tasks
are only generated for blocks with non-zero elements. Tasks associated with the shaded
blocks require communication between different processes.
mentation will depend on how well the sparse matrix-vector multiplication
scales. The main challenges are the low computational density, which can
lead to bandwidth contention and suboptimal scaling within a shared memory node, and the unstructured layout of the non-zero elements that leads
to scattered memory accesses to the input vector, which is less efficient than
linear accesses due to cache effects. For the sparse matrices, compressed row
storage (CRS) is used. As shown in [41] the computational density can be
increased by combining operations. There are four sparse matrices with the
same structure that are multiplied with the four unknown vectors, and by
18
storing the matrices and the unknowns together, the sixteen matrix-vector
operations can be performed together. When increasing the number of computational nodes, we instead expect near optimal scaling since the amount
of inter node communication is small, see Figure 6.
A critical aspect for this type of application is the synchronization points,
because the number of tasks that can run in parallel is relatively small. The
problem sizes that have been used for this particular application range from
6 000 to 600 000 data points in space. Since the hierarchical model employs
synchronization between child tasks and their parent task, the potential mixing of tasks from different time steps, which can be beneficial for utilization,
can be prevented if the number of tasks at the parent level are too few.
5.4. Simulation mode
The actual performance of a task parallel application code will depend
on how data is partitioned and distributed over the processes. This determines the amount of communication that has to be performed, how evenly
the work is distributed over the processes, and how well the data sizes fit
the memory hierarchy. In some cases, the properties of an algorithm are
easy to understand theoretically and the best distribution can be derived
analytically. In other cases, this may be less clear and some experiments
are needed. To perform full simulations for many different parameter combinations to find the optimal ones is a waste of computational resources. A
cheaper alternative is to use what is called simulation mode in DuctTeip.
The application code can be executed in simulation mode as it is, but the
run-time replaces all computational kernels with an empty kernel and all
messages with one byte messages. Statistics from the execution regarding
numbers of tasks and messages are recorded to provide information to the
user for deciding on partitioning and distribution schemes.
To illustrate how this can work, we use the Cholesky algorithm as an
example to see if we can recover the expected advice about how to best
run it. Execution in simulation mode can not tell us what block size is
appropriate for the caches. However, results for the level 2 block size can be
obtained by running a small problem on one node. In all experiments shown
here, this size is fixed to 462 × 462 elements. Statistics for the total amount
of communication and the maximum work on one node are given for three
different process grids combined with three different blockings in Figure 7.
Starting with the communication, we can observe that the total amount
is the same for the 9×1 and 1×9 process grids, while it is significantly lower
for the 3 × 3 process grid. In all cases, the amount grows with the number
of level 1 blocks, B, but only mildly for the 3 × 3 process grid. Hence, from
19
a communication perspective, a square process grid is desirable. Proceeding
to the computational work performed by one node, the 1×9 process grid has
the smallest variance between nodes, and therefore also the lowest maximum
work size. The 9 × 1 process grid leads to smaller maximum work size than
the 3 × 3 process grid if B is large enough, but suffers from significant load
imbalance in the case B = 18. In all cases, the work becomes more evenly
distributed if the number of level 1 tasks B is larger.
The statistics for communication and computation point in different directions, but when comparing with actual run times, we have found that the
communication size is the most informative measure. Having a large total
communication size is likely to be detrimental to performance as the risk of
tasks left waiting for remote data increases as well as the risk of congestion
of messages. A square process grid is the factor that has the largest impact.
Regarding the block sizes, having a large B improves the load balance, but
increases the amount of communication as well as the number of messages
(another indicator that is not shown in the graphics).
9x1 process grid
1x9 process grid
3x3 process grid
250
230
Total communication size (GB)
Comparison of total computations
Comparison of total communication
210
190
170
150
130
B=36, b=6
B=27, b=8
10
Maximum total computations (TFLOP)
270
9
9x1 process grid
1x9 process grid
3x3 process grid
8
7
6
5
B=36, b=6
B=18, b=12
B=27, b=8
B=18, b=12
Figure 7: Simulation statistics for Cholesky factorization of a matrix of size N = B × b ×
462 = 99 792, where B is the level 1 block size and b is the level 2 block size. The simulation
is executed on 9 nodes using three different process grids. The total communication is the
sum of the number of GBytes communicated by each node, while the work is measured
as the maximum number of level 2 tasks executed by one node multiplied by the typical
work size, which is proportional to n3 for the Cholesky factorization, where n = 462 is the
level 2 block size. The error bars show the variance over the nodes.
6. Experimental results
The experiments presented in this section were performed during the
time period when the main HPC cluster at the Uppsala Multidisciplinary
20
Center for Advanced Computational Science (UPPMAX) was replaced. Therefore, the experiments where we compare the performance of different task
parallel frameworks are performed at the Tintin cluster, while the experiments on the performance of the individual applications are performed on
the Rackham cluster that replaced Tintin. We discuss how the properties
of the different systems impact the results in connection with the specific
experiments.
The Tintin cluster had 180 dual socket computational nodes with 64
GB/128 GB memory each. Each socket was equipped with an eight core
AMD Opteron 6220 (Bulldozer) processor running at 3.0 GHz. The theoretical maximum performance for one Bulldozer processor is 192 GFlops and
the theoretical memory bandwidth is 51.2 GB/s corresponding to 6.4 double precision loads per second. This means that 192/6.4 = 30 floating point
operations per load are needed to not suffer from bandwidth limitations.
The new Rackham cluster taken into production in April 2017 currently has 334 dual socket nodes with 128 GB/256 GB memory each. Each
socket is equipped with a 10 core Intel Xeon E5 2630 v4 (Broadwell) processor running at 2.2 GHz. For a dual socket Broadwell node, the corresponding performance numbers are 704 GFlops and 136.6 GB/s leading
to 704/(136.6/8) ≈ 41 floating point operations per double precision load.
That is, to achieve a similar utilization as on Tintin, the computational
intensity needs to be higher for Rackham.
6.1. Weak scalability for the Cholesky factorization
These experiments are run on the new Rackham cluster, and the application code is compiled with the Intel Parallel Studio XE 18.0.1 C++
compiler. When performing the parallel experiments, the problem size is
increased with the number of nodes such that the amount of data that one
node needs to communicate is fixed. The reason for fixing the communication size rather than the work size is that our experiment indicate that
communication has a larger impact on the actual performance than the computations. The communication size is approximately proportional to N 2 /p,
where N is the matrix size and p is the number of processes (computational
nodes). Hence, if we start with a matrix size N1 for one node, then the
√
matrix size for p nodes becomes Np = pN1 . A square process grid is used
in all cases, which is beneficial for reducing the total amount of communication, see Section 5.4. The matrix is split into B × B level one blocks, and
then each level 1 block is split into b × b level 2 blocks.
We start by investigating the performance on one node. The size n of
the level 2 blocks has a large impact on performance. It influences both
21
Actual performance achieved for factorization of
different matrices
0.8
0.7
0.6
Thread
TFLOPS
0.5
0.4
0.3
0.2
Theoretical Peak
Various N
N=40K
N=10K
0.1
0.0
0
500
1000 1500 2000 2500 3000 3500 4000
Block Size
19
18
17
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1
0
0
500
1000
Time (million cycles)
1500
2000
Figure 8: Left: The computational speed as a function of the level 2 block size when the
task parallel Cholesky application is executed on one computational node. Right: The
beginning of the execution trace on one node for N = 40K, B = 5, and b = 30. Each
triangle represents a task, and the color indicates the type, with red for potrf, blue for
trsm, green for gemm, and gray for DuctTeip parent tasks.
how the blocks align with the cache sizes and the amount of computations
per memory load. The left panel in Figure 8 shows how the computational
speed varies with the block size. For the experiments marked with stars, the
problem size is allowed to grow, while the number of blocks B × b = 10 × 3 is
fixed. Then the computational speed increases with block size until it reaches
a steady level. When we instead fix N, we see that the performance follows
the same curve for a while, but then performance goes down again. For
N = 10 000, the performance starts going down for n = 500 this corresponds
to a product B × b = 20. The parallelism of the Cholesky task graph
varies over the execution, see Figure 5, but for example, the first instance
of the trsm operation has width B × b for a two-level graph. That is, the
performance goes down when we no longer supply enough parallelism to
keep the 20 threads busy. We are left with a trade-off situation, where the
block size should be large enough for performance, but small enough for
parallelism.
The best performance achieved on one node for block sizes n & 2000 is
about 85% of the theoretical peak performance, which is computed by assuming that the software manages to use both vectorization (AVX instructions) and fused multiply-add (FMA) instructions fully. Then the Broadwell
processors in Rackham can manage 16 floating point operations per cycle
and core at 2.2 GHz, which gives a theoretical peak performance of 704
GFlops for one node. Studying execution traces such as the right panel
in Figure 8, some of the causes for the distance from the theoretical peak
22
can be understood. For the investigated problem sizes, thread 0 does not
perform any computational work. Thread 0 handles task submission at the
shared memory level, and if task execution is fast in relation to task submission (fine task granularity), it does not have time to do computational
work. This stands for a 5% performance loss in relation to peak. Another
issue is the lack of parallelism in the beginning and the end of the Cholesky
execution. The first panel factorization needs to be completed before more
tasks become ready. When a relatively small b is used, such as in the best
performing results, there is not enough parallelism in this step to occupy
more than a fraction of the threads. The rest are idle until the panel factorization completes. Similarly at the end of the execution the amount of
parallelism decreases and there is some idle time among the threads. The
first tasks also exhibit longer execution times than later tasks. This is likely
due to memory effects, since data needs to be fetched from the main memory
until the caches are warmed up. In problems where block sizes and block
numbers are well chosen, the idle time for the actively working threads is
small, around 0.5%.
For the multi-node weak scaling experiments, we have opted for having
enough parallelism, which means that the block sizes are too small to give
optimal performance. For the largest problem sequence, the typical block
size is n = 300, which means that we cannot expect more than around 0.4
TFlops/computational node. The parameter values used in the experiments
are given in Table 1. The right panel of Figure 9 shows the weak scaling
Table 1: The blocking parameters used in the Cholesky weak scaling experiments.
p
B
b
1
5
30
4
8
30
9
12
30
16
16
30
25
20
30
36
18
20
49
28
30
results for the multinode experiments for three different initial matrix sizes.
We can observe that the absolute performance improves with matrix size.
The efficiency Ep of the scaling, computed as Sp , the actual speed for p
processes, divided by the theoretical speed for linear scaling
Ep =
Sp
,
pS1
is given in Table 2. For the largest matrix size we achieve about 50% of
the expected performance when increasing p, while the smaller problem
sizes have a reduction also in efficiency. The traces again reveal the reason
23
Weak scalability for DuctTeip implementation of
Cholesky factorization
35
Theoretical Peak
Actual Peak
30
N1 =4 £104 ; N49 =2:8 £105
N1 =2 £104 ; N49 =1:4 £105
N1 =1 £104 ; N49 =7 £104
Thread
TFLOPS
25
20
15
10
5
0
1
4
9
16
25
Nodes
36
49
19
18
17
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1
0
0
50000
100000
150000
200000
Time (million cycles)
250000
Figure 9: Left: Weak scaling for three different initial matrix sizes. The theoretical peak
corresponds to 704 GFlops/node while the actual peak is the speed on one node for the
largest problem multiplied with the number of nodes. Right: A simplified representation
of the trace at node 0 when running the Cholesky application at 4 nodes. Busy time is
indicated by filled blocks, while the idle time has just the 0 line.
for the results. In the right panel of Figure 9, we see the trace for the
smallest matrix size running on four nodes. There are a number of gaps
in the execution. We believe that these are connected with the inter-node
communication, and that the computational work is not enough to hide the
latency of the communication operations. There could potentially be some
improvement through introducing priority for the critical path to ensure
that data is communicated as early as possible.
A positive aspect of the results is that the code scales up to 49 nodes
for the tested problem sizes. There is some reduction in efficiency with the
number of nodes, but the main loss is between running on one and multiple
nodes.
Table 2: The efficiency in the weak scaling experiments for the Cholesky application for
different initial matrix sizes N1 . The speed on one node, S1 , is measured in TFlops.
N1
10 000
20 000
40 000
S1
0.33
0.44
0.46
E4
0.50
0.60
0.71
E9
0.38
0.38
0.55
E16
0.40
0.42
0.55
E25
0.34
0.36
0.54
E36
0.29
0.31
0.50
E49
0.27
0.29
0.43
6.2. Strong scalability for the shallow water application
For the shallow water solver, we look at the performance for fixed problem sizes. These problems are benchmarks for global climate simulations
24
and are typically run for a certain resolution of the Earth. The largest matrix size we are using is N = 612 416, which is not very large for a sparse
matrix. The main reason that the problems still take a long time to run is
the large number of time steps that are needed. For example, for a simulation over 6 days with a time step of 30 s, 17 280 time steps are needed. In
the numerical experiments, we perform 20 time steps, which is long enough
to capture the long term behavior of the solver. The block sizes used in the
experiments are given in Table 3.
Table 3: The blocking parameters used for the shallow water experiments.
p
B
b
1
5
18
2
6
18
3
15
5
4
40
5
5
50
5
6
60
5
7
35
5
8
32
8
9
36
5
10
50
4
First we look at the speedup Sp = T1 /Tp , where Tp is the execution time
on p nodes. The left part of Figure 10 shows the results when compiling
with gcc 7.2.0 and linking with OpenMPI. The speedup is close to linear for
small numbers of nodes, and gradually moves away from linear when the
node number increases. The speedup improves with problem size. However,
the picture is different in the right part of Figure 10, where the Intel C++
compiler has been used together with Intel MPI. Execution is unexpectedly
fast at two nodes, but in all other cases, it is slower than or similar to
execution on one node. Furthermore, the speedup at two nodes is highest
for the smallest problem size.
10
9
8
Speedup for different problem sizes of
the Shallow Water application (gcc)
Perfect speedup
Actual speedup, N=600K
Actual speedup, N=300K
Actual speedup, N=150K
10
8
Speedup for different problem sizes of
the Shallow Water application (intel)
Perfect speedup
Actual speedup, N=600K
Actual speedup, N=300K
Actual speedup, N=150K
7
6
6
5
4
4
3
2
2
1
1
2
3
4
5
Nodes
6
7
8
9
10
0
1
2
3
4
5
Nodes
6
7
8
9
10
Figure 10: Speedup for the shallow water application using the gcc (left) and Intel compilers (right).
To understand the differences between the two versions, we continue with
25
comparing the absolute execution times for the largest problem size. Table 4
shows that the smallest run time is actually the same for both compilers, but
it is achieved at 10 nodes for gcc, and for two nodes by the Intel compiler.
Also, the run time on one node is three times smaller for the Intel compiler.
One expected effect of this is that it will be harder for the code compiled
with the Intel compiler to scale since the perceived work size is smaller.
Table 4: Run times in seconds for the shallow water application with N = 612 416 when
the code is compiled with the gcc (g) and Intel (I) compiler, respectively.
p
(g)
Tp
(I)
Tp
1
60.4
20.3
2
32.4
7.3
3
22.4
31.5
4
15.6
257.9
5
12.6
269.4
6
10.8
273.7
7
10.5
30.8
8
9.0
34.1
9
8.0
35.5
10
7.4
33.8
Finally, we take a look at the execution traces. In the gcc case, all
the traces have the same appearance, but with a different time scale. An
example is shown in the left part of Figure 11. Thread 0 is occupied with task
submission during the whole execution. The computations are performed on
the remaining 19 threads. The different time steps are easy to identify, and
the desired mixing that allows the work to be balanced over the threads is
visible. This works as long as the number of tasks in each time step at both
the DuctTeip and SuperGlue levels are large enough to provide parallelism.
In the right part of Figure 11, we show a trace for the code compiled
with the Intel compiler. It is clear that there is significant idle time between
the tasks. There is also a clear variation in how long time one time step
takes. To further quantify what we see in the traces, we compare the time
spent in tasks for the two versions running on four nodes. The tasks are on
average computed 13 times faster in the Intel version than in the gcc version,
in spite of this, the overall execution time for the gcc version is more than
16 times faster.
The drop in performance experienced with the Intel compiler can be related to observations made for SuperGlue regarding small task sizes, see [10,
Fig. 6]. When the task size is too small, and the number of threads is comparatively large, the work stealing algorithm used for load balancing results
in contention between the threads, preventing productive work.
6.3. Comparison between hybrid and pure MPI parallel implementations
In this section, we try to quantify the communication disadvantage of
having a pure MPI implementation instead of the hybrid implementation
using Pthreads within the shared memory nodes with an example. We run
26
0
Thread
Thread
19
18
17
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1
0
5000
10000
15000
20000
Time (million cycles)
25000
30000
19
18
17
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1
0
0
100000
200000
300000
Time (million cycles)
400000
500000
Figure 11: The execution traces for node zero for the code compiled with gcc (left) and
for the code compiled with Intel (right) running on four nodes. The matrix and vectors
are here sliced into B = 40 and b = 4 and b = 5 blocks respectively.
the Cholesky factorization code on two computational nodes with a 6 × 6
process grid in the pure MPI case (a square process grid is optimal), and a
2 × 1 process grid for the hybrid implementation.
Table 5 shows the effect on the number of messages and the resulting
execution times for the two cases. The total number of messages increases
significantly with the increase in the number of processes for a pure MPI
implementation. Furtermore, in the Cholesky case, the number of messages
between the nodes also increases. The number of messages per process is
however relatively stable. A negative effect that will become even worse
when the number of nodes is increased is the length of the message queues.
Here, the increase in queue length is 3.5, which has an effect on the time it
will take to process messages and hence the communication overhead. The
execution time in the pure MPI case is increased by 50%. As we are actually
using only 36 cores compared with 40 for the hybrid case, we should expect
an 11% increase, but not as much as 50%.
Table 5: Execution statistics for a Cholesky factorization of a matrix of size N = 90 000
comparing the hybrid DuctTeip implementation and a pure MPI version.
Total messages
Max messages/process
Max pending messages
Execution time (s)
Hybrid
16
12
2
18
27
Pure MPI
455
20
7
27
Ratio MPI/Hybrid
28.4
1.7
3.5
1.5
6.4. Benchmarking against other frameworks
When comparing DuctTeip against other frameworks, we use the Cholesky
factorization as a benchmark because this is a standard test case, which has
been implemented also by the other framework developers. We compare
against PaRSEC, Cluster OmpSs and StarPU, but we also include ScaLAPACK to represent an application specific implementation with an optimized
static parallel schedule for the work. To be compatible with ScaLAPCK that
uses a square process grid, we only use square numbers of nodes in the experiments. As mentioned in the introduction to the experimental section,
these experiments are performed on the older cluster Tintin.
Compilers and libraries have been selected to provide the best possible
performance for the target hardware at the time when the experiments were
made. Since the software is continuously upgraded, improved versions may
be available. Below, we specify the choices for the frameworks and other
components that were used.
BLAS: The ACML 5.3.1 implementation of BLAS is used, except for StarPU,
which uses the ATLAS library 3.10.2, in both cases the libraries are
optimized for the target system.
MPI: All applications are linked with OpenMPI 1.6.5.
hwloc: Version 1.9.1 is used by PaRSEC and StarPU and built by gcc 4.4.7.
DuctTeip: Compiled with the Intel C/C++ 13.1 compiler.
Cluster OmpSS: The customized compilers Mercury 1.99.7 and Nanos+
0.9a from the Barcelona Supercomputing Center are built on top of
gcc 4.4.7. The Cholesky application is then compiled and built using
these.
ScaLAPACK: Version 2.0.2 is compiled and built with the PGI 13.8 FORTRAN compiler and optimized for the target system. The Cholesky
factorization is performed by calling the DPOTRF subroutine.
StarPU: Version 1.1.5 is built using gcc 4.8 and calibrated for the Cholesky
factorization program using the dmdar (deque model data aware ready)
task scheduling policy.
PaRSEC: DPLASMA 1.2.1 is compiled and built by the Intel C/C++ 13.1
compiler.
The most frequent operation in the block Cholesky factorization is a
GEMM operation (matrix–matrix multiplication). To have a reference to
what could be considered the maximal achievable performance we include
28
what we call the GEMM peak in our comparisons. The maximum throughput of the dgemm BLAS subroutine on one node (16 cores) that is achieved
on the Tintin cluster is 162 GFlops for square matrices of size N = 24 000.
The GEMM peak is then computed as the one-node-result scaled by the
number of nodes.
6.4.1. Strong scaling experiments
For the strong scaling experiments, we have identified the largest problem that can be solved on a single node without performance (throughput)
degradation. This corresponds to matrix size N = 43 200. Then, with the
problem size fixed, we increase the number of nodes. Theoretically, linear
scalability with the number of nodes is expected. The experimental results
in terms of TFlops are shown in the left part of Figure 12.
Two frameworks, PaRSEC and DuctTeip are faster than the highly optimized ScaLAPACK Cholesky factorization. Potential reasons for this could
be that the ScaLAPACK implementation is a pure MPI implementation,
and that the matrix storage is not tiled. Furthermore, a dynamical schedule
without global barriers can be more efficient in exploiting the variabilities
of modern hardware than a static schedule with global synchronization.
PaRSEC is the fastest of all frameworks. One reason for this can be
that with the embedded parametrized task graph, decisions about which
task to run next can be taken quickly leading to a low overhead. DuctTeip
comes out well in the comparison. A difference compared with PaRSEC
is that no pre-knowledge about the data dependencies is used. We draw
the conclusion that the programming model with hierarchical partitioning
of data and tasks coupled with decentralized task submission and scheduling
is appropriate for the architecture we are using.
Both StarPU and Cluster OmpSs perform significantly worse, especially
for larger numbers of nodes. A contributing reason for StarPU can be the
single level data and tasks, which increases the amount of communication,
whereas in Cluster OmpSs, the centralized task submission can be an explanation for the lack of scalability.
The right part of Figure 12 shows the speedup for the same problem
(dashed lines) for up to 49 nodes (784 cores) as well as for a problem with a
larger matrix size, optimized to give maximum performance on four nodes
(64 cores) (solid lines). The speedup numbers are computed as Sp = T1 /Tp
for the first experiment, and Sp = 4T4 /Tp for the second experiment. The
relative performance of the frameworks is the same in both experiments, but
the second experiment also shows that increasing the workload results in a
reduction of the relative size of the overhead, and the speedup and scaling
29
101
Strong Scaling (total work = 27 TF)
N=43200,86400
49
36
TFLOPS
SpeedUp
100
25
16
10-1
10-2 1
4
GEMM Peak
PaRSEC
9
Number of Nodes
DuctTeip
ScaLAPACK
16
25
36
49
StarPU
ClusterSs
9
4
1
1 4
9
Theoretical
PaRSEC
16
25
36
Number of Nodes
DuctTeip
ScaLAPACK
49
StarPU
ClusterSs
Figure 12: The left subfigure shows the scaling of the different frameworks for Cholesky
factorization of a matrix of size N = 43 200. In the right subfigure, the fixed size speedup
for problem sizes N = 43 200 (dashed lines) and N = 86 400 (solid lines) is shown.
are thereby improved. Both StarPU and Cluster OmpSs had problems with
handling large matrices and/or larger number of nodes, and therefore no
results are shown for these frameworks in those cases.
6.4.2. Weak scaling experiments
For the weak scaling experiments, we increase the problem size such
that the communication size is constant as we also did in Section 6.1. The
√
problem sizes are given by N = 10 800 p leading to to N = 75 600 on
49 nodes (784 cores). Figure 13 shows the experimental result. In absolute numbers, PaRSEC shows the best performance, followed by DuctTeip,
and then ScaLAPACK. However, looking at the slope of the scaling results,
PaRSEC scales almost perfectly, while DuctTeip scales slightly worse than
ScaLAPACK.
7. Conclusions
We have introduced a dependency-aware hierarchical task-based programming model, where the execution order of the tasks is controlled by a
versioning system. The data versioning approach has previously been shown
to be very effective in a shared memory setting [10], and here we have shown
that it is also very useful in a distributed context. In particular, the consistency of the hierarchical task submission is automatically handled by the
framework, and tasks at all levels can be executed in parallel. The hierarchical task decomposition helps to pace the submission of tasks such that task
overflow is avoided. It also gives a lot of flexibility in choosing appropriate
30
Weak Scaling
101
TFLOPS
100
10-1
10-2 1
GEMM Peak
PaRSEC
4
9
Number of Nodes
DuctTeip
16
25
36 49
ScaLAPACK
Figure 13: Weak scaling scaling results for the Cholesky factorization when the problem
size is scaled up from N = 10 800 one node such that N 2 /p, where p is the number of
processes, is held constant.
task sizes for the different levels. There are however a number of trade-offs
that influence the performance that make it hard to select the blocking parameters. The simulation mode of DuctTeip gives some guidance regarding
load balancing over nodes and communication volumes. Furthermore, we
have proposed an abstract software construct, the listener, which enables
the run-time system to control the communication of data. Without this
abstraction level, the run-time system would not be able to detect multiple
requests of the same data from a single requester or resolve the arrival of
multiple versions of the same data by a single receiver.
In the current DuctTeip implementation, the programmer needs to supply the task definitions and algorithms at each level. This is an unnecessary
obstacle to the ease of programming in the framework. Therefore, in an ongoing project, we are developing a unified interface, such that the algorithm
and the task kernels only need to be defined once and then can be reused
for all levels [39].
A potential obstacle to performance is load imbalance between the computational nodes. The task ownership is an implicit result of the data distribution scheme, the block sizes, the process grid, and the algorithm. Clearly,
these factors together may result in load imbalance. In the DuctTeip framework both tasks and data are designed to be movable objects. Hence, the
framework is prepared for distributed dynamical load balancing. In another
31
ongoing project, we are investigating whether distributed load balancing
can be implemented within the framework in such a way that overall performance is improved.
We have, to our knowledge, performed the first comparison involving
several of the established distributed task parallel frameworks. The experiments show that DuctTeip compares very well with similar frameworks and
libraries regarding both absolute performance and relative speedup. Only
PaRSEC, which is extremely scalable, but also encodes more information
about the task graph in the compiled program, performs better.
Acknowledgments
The computations were performed on resources provided by SNIC through
Uppsala Multidisciplinary Center for Advanced Computational Science (UPPMAX) under Project p2009014. Lennart Karlsson at UPPMAX is acknowledged for assistance concerning technical and implementational aspects in
making the code run on the UPPMAX resources.
Javier Bueno Hedo, Barcelona Supercomputing Centre, Barcelona, Spain,
is acknowledged for assistance with the installation of the Cluster OmpSs
compilers and for providing the Cholesky factorization code for Cluster
OmpSs.
Finally, we thank Igor Tominec, Uppsala University for constructive feedback on the text.
References
[1] J. Dongarra, M. Abalenkovs, A. Abdelfattah, M. Gates, A. Haidar,
J. Kurzak, P. Luszczek, S. Tomov, I. Yamazaki, A. YarKhan, Parallel
programming models for dense linear algebra on heterogeneous systems,
Supercomputing Frontiers and Innovations 2 (4).
URL http://superfri.org/superfri/article/view/90
[2] R. D. Blumofe, C. F. Joerg, B. C. Kuszmaul, C. E. Leiserson, K. H.
Randall, Y. Zhou, Cilk: An efficient multithreaded runtime system, in:
Proceedings of the Fifth ACM SIGPLAN Symposium on Principles &
Practice of Parallel Programming (PPOPP), Santa Barbara, California,
USA, July 19-21, 1995, pp. 207–216. doi:10.1145/209936.209958.
[3] K.-F. Faxén, Wool-a work stealing library, SIGARCH Comput. Archit.
News 36 (5) (2009) 93–100. doi:10.1145/1556444.1556457.
32
[4] H. Vandierendonck, G. Tzenakis, D. S. Nikolopoulos, A unified scheduler for recursive and task dataflow parallelism, in: Proceedings of the
2011 International Conference on Parallel Architectures and Compilation Techniques (PACT ’11), 2011, pp. 1–11.
[5] OpenMP Architecture Review Board, OpenMP 4.0 Complete specifications, http://openmp.org/wp/openmp-specifications/ (July 2013).
[6] P. Virouleau, F. Broquedis, T. Gautier, F. Rastello, Using Data Dependencies to Improve Task-Based Scheduling Strategies on NUMA Architectures, Springer International Publishing, Cham, 2016, pp. 531–544.
doi:10.1007/978-3-319-43659-3_39.
[7] Intel® TBB, Intel® threading building blocks documentation,
https://www.threadingbuildingblocks.org/ (June 2016).
[8] A. Duran, E. Ayguadé, R. M. Badia, J. Labarta, L. Martinell, X. Martorell, J. Planas, Ompss: a proposal for programming heterogeneous
multi-core architectures, Parallel Processing Letters 21 (2) (2011) 173–
193. doi:10.1142/S0129626411000151.
[9] C. Augonnet, S. Thibault, R. Namyst, P. Wacrenier, StarPU: a unified
platform for task scheduling on heterogeneous multicore architectures,
Concurrency Computat.: Pract. Exper. 23 (2) (2011) 187–198. doi:
10.1002/cpe.1631.
[10] M. Tillenius, SuperGlue: A shared memory framework using data versioning for dependency-aware task-based parallelization, SIAM J. Sci.
Comput. 37 (6) (2015) C617–C642. doi:10.1137/140989716.
[11] C. Augonnet, O. Aumage, N. Furmento, R. Namyst, S. Thibault,
StarPU-MPI: Task programming over clusters of machines enhanced
with accelerators, in: Recent Advances in the Message Passing Interface - 19th European MPI Users’ Group Meeting, EuroMPI 2012,
Vienna, Austria, September 23–26, 2012, pp. 298–299. doi:10.1007/
978-3-642-33518-1_40.
[12] E. Tejedor, M. Farreras, D. Grove, R. M. Badia, G. Almasi, J. Labarta,
ClusterSs: a task-based programming model for clusters, in: Proceedings of the 20th ACM International Symposium on High Performance
Distributed Computing, HPDC 2011, San Jose, CA, USA, June 8–11,
2011, pp. 267–268. doi:10.1145/1996130.1996168.
33
[13] G. Bosilca, A. Bouteiller, A. Danalis, M. Faverge, T. Herault, J. J.
Dongarra, PaRSEC: Exploiting heterogeneity to enhance scalability,
Comput. Sci. Eng. 15 (6) (2013) 36–45. doi:10.1109/MCSE.2013.98.
[14] M. Sergent, D. Goudin, S. Thibault, O. Aumage, Controlling the memory subscription of distributed applications with a task-based runtime
system, in: Parallel and Distributed Processing Symposium Workshops,
2016 IEEE International, IEEE, 2016, pp. 318–327.
[15] G. Bosilca, A. Bouteiller, A. Danalis, T. Hérault, P. Lemarinier, J. Dongarra, DAGuE: A generic distributed DAG engine for high performance computing, Parallel Comput. 38 (1–2) (2012) 37–51. doi:
10.1016/j.parco.2011.10.003.
[16] A. Zafari, M. Tillenius, E. Larsson, Programming models based on data
versioning for dependency-aware task-based parallelisation, in: Proceedings of the 15th IEEE International Conference on Computational
Science and Engineering, CSE 2012, Paphos, Cyprus, December 5–7,
2012, pp. 275–280. doi:10.1109/ICCSE.2012.45.
[17] L. V. Kalé, Charm++, in: Encyclopedia of Parallel Computing, Springer US, Boston, MA, 2011, pp. 256–264. doi:10.1007/
978-0-387-09766-4_242.
[18] L. V. Kalé, S. Krishnan, CHARM++: A portable concurrent object
oriented system based on C++, in: Proceedings of the Eighth Annual
Conference on Object-Oriented Programming Systems, Languages, and
Applications (OOPSLA), Washington, DC, USA, September 26 - October 1, 1993, pp. 91–108. doi:10.1145/165854.165874.
[19] E. H. Rubensson, E. Rudberg, Chunks and tasks: A programming
model for parallelization of dynamic algorithms, Parallel Comput.
40 (7) (2014) 328–343. doi:10.1016/j.parco.2013.09.006.
[20] M. Tillenius, E. Larsson, R. M. Badia, X. Martorell, Resource-aware
task scheduling, ACM Trans. Embedded Comput. Syst. 14 (1) (2015)
5:1–5:25. doi:10.1145/2638554.
[21] G. Ceballos, T. Grass, A. Hugo, D. Black-Schaffer, Taskinsight: Understanding task schedules effects on memory and performance, in: Proceedings of the 8th International Workshop on Programming Models
and Applications for Multicores and Manycores, PMAM’17, ACM, New
York, NY, USA, 2017, pp. 11–20. doi:10.1145/3026937.3026943.
34
[22] G. Ceballos, E. Hagersten, D. Black-Schaffer, Stattask: Reuse distance
analysis for task-based applications, in: Proceedings of the 2015 Workshop on Rapid Simulation and Performance Evaluation: Methods and
Tools, RAPIDO ’15, ACM, New York, NY, USA, 2015, pp. 1:1–1:7.
doi:10.1145/2693433.2693434.
[23] A. Drebes, K. Heydemann, N. Drach, A. Pop, A. Cohen, Topologyaware and dependence-aware scheduling and memory allocation for
task-parallel languages, ACM Trans. Archit. Code Optim. 11 (3) (2014)
30:1–30:25. doi:10.1145/2641764.
[24] A. Muddukrishna, P. A. Jonsson, M. Brorsson, Locality-aware task
scheduling and data distribution for openmp programs on numa systems
and manycore processors, Sci. Program. 2015 (2016) 5:5–5:5. doi:
10.1155/2015/981759.
[25] A. Pan, V. S. Pai, Runtime-driven shared last-level cache management
for task-parallel programs, in: SC15: International Conference for High
Performance Computing, Networking, Storage and Analysis, 2015, pp.
1–12. doi:10.1145/2807591.2807625.
[26] M. Pericàs, A. Amer, K. Taura, S. Matsuoka, Analysis of Data Reuse
in Task-Parallel Runtimes, Springer International Publishing, Cham,
2014, pp. 73–87. doi:10.1007/978-3-319-10214-6_4.
[27] R. M. Yoo, C. J. Hughes, C. Kim, Y.-K. Chen, C. Kozyrakis, Localityaware task management for unstructured parallelism: A quantitative
limit study, in: Proceedings of the Twenty-fifth Annual ACM Symposium on Parallelism in Algorithms and Architectures, SPAA ’13, ACM,
New York, NY, USA, 2013, pp. 315–325. doi:10.1145/2486159.
2486175.
[28] A. Jimborean, K. Koukos, V. Spiliopoulos, D. Black-Schaffer, S. Kaxiras, Fix the code. don’t tweak the hardware: A new compiler approach
to voltage-frequency scaling, in: Proceedings of Annual IEEE/ACM International Symposium on Code Generation and Optimization, ACM,
2014, p. 262.
[29] K. Koukos, D. Black-Schaffer, V. Spiliopoulos, S. Kaxiras, Towards
more efficient execution: a decoupled access-execute approach, in: Proceedings of the International Conference on Supercomputing, ICS’13,
Eugene, OR, USA, June 10–14, 2013, pp. 253–262. doi:10.1145/
2464996.2465012.
35
[30] M. Garcia Gasulla, Dynamic load balancing for hybrid applications,
Doctor of philosophy in computer architecture, Department of Computer Architecture, Universitat Politècnica de Catalunya (2017).
[31] M. Moniruzzaman, K. Idrees, M. Rossbory, J. Gracia, An adaptive
load-balancer for task-scheduling in fastflow, INFOCOMP 2015 (2015)
18.
[32] G. Zheng, Achieving high performance on extremely large parallel
machines: performance prediction and load balancing, Ph.D. thesis,
Department of Computer Science, University of Illinois at UrbanaChampaign (2005).
[33] M. C. Kurt, S. Krishnamoorthy, K. Agrawal, G. Agrawal, Fault-tolerant
dynamic task graph scheduling, in: Proceedings of the International
Conference for High Performance Computing, Networking, Storage and
Analysis, SC ’14, IEEE Press, Piscataway, NJ, USA, 2014, pp. 719–730.
doi:10.1109/SC.2014.64.
URL https://doi.org/10.1109/SC.2014.64
[34] J. Lifflander, E. Meneses, H. Menon, P. Miller, S. Krishnamoorthy, L. Kale, Scalable Replay with Partial-Order Dependencies for
Message-Logging Fault Tolerance, in: Proceedings of IEEE Cluster
2014, Madrid, Spain, 2014.
[35] A. Bouteiller, T. Herault, G. Bosilca, P. Du, J. Dongarra, Algorithmbased fault tolerance for dense matrix factorizations, multiple failures
and accuracy, ACM Trans. Parallel Comput. 1 (2) (2015) 10:1–10:28.
doi:10.1145/2686892.
[36] C. Cao, T. Herault, G. Bosilca, J. Dongarra, Design for a soft error resilient dynamic task-based runtime, in: 2015 IEEE International
Parallel and Distributed Processing Symposium, 2015, pp. 765–774.
doi:10.1109/IPDPS.2015.81.
[37] X. Ni, Mitigation of failures in high performance computing via runtime techniques, Ph.D. thesis, Dept. of Computer Science, University
of Illinois (2016).
[38] F. Broquedis, J. Clet-Ortega, S. Moreaud, N. Furmento, B. Goglin,
G. Mercier, S. Thibault, R. Namyst, hwloc: A generic framework for
managing hardware affinities in HPC applications, in: Proceedings of
36
the 18th Euromicro Conference on Parallel, Distributed and Networkbased Processing, PDP 2010, Pisa, Italy, February 17–19, 2010, pp.
180–186. doi:10.1109/PDP.2010.67.
[39] A. Zafari, TaskUniVerse: A task-based unified interface for versatile
parallel execution, in: R. Wyrzykowski, E. Deelman, et al. (Eds.), Parallel Processing and Applied Mathematics—12th International Conference, PPAM 2017, Krakow, Poland, September 10–13, 2017, Lecture
Notes in Computer Science, Springer, 2018, 14 pp., to appear.
[40] M. Holm, S. Engblom, A. Goude, S. Holmgren, Dynamic autotuning
of adaptive fast multipole methods on hybrid multicore CPU and GPU
systems, SIAM J. Sci. Comput. 36 (4) (2014) C376–C399. doi:10.
1137/130943595.
[41] M. Tillenius, E. Larsson, E. Lehto, N. Flyer, A scalable RBF-FD
method for atmospheric flow, J. Comput. Phys. 298 (2015) 406–422.
doi:10.1016/j.jcp.2015.06.003.
[42] P. Bauer, S. Engblom, S. Widgren, Fast event-based epidemiological simulations on national scales, Int. J. High Perform. Comput. Appl.Electronic publication ahead of print.
doi:10.1177/
1094342016635723.
[43] N. Flyer, G. B. Wright, A radial basis function method for the shallow
water equations on a sphere, Proc. R. Soc. Lond. Ser. A Math. Phys.
Eng. Sci. 465 (2106) (2009) 1949–1976. doi:10.1098/rspa.2009.0033.
37
| 5cs.CE |
EVALUATING A SUBLINEAR-TIME ALGORITHM
FOR THE MINIMUM SPANNING TREE WEIGHT PROBLEM
arXiv:1701.08128v1 [cs.DS] 27 Jan 2017
GABRIELE SANTI∗ AND LEONARDO DE LAURENTIIS†
Abstract. We present an implementation and an experimental evaluation of an algorithm that,
given a connected graph G (represented by adjacency lists), estimates in sublinear time, with a
relative error e, the Minimum Spanning Tree Weight of G (see [1] for a theoretical exposure of
the algorithm). Since the theoretical performances have already been shown and demonstrated in
the above-mentioned paper of Chazelle et al. our goal is, exclusively, to experimental evaluate the
algorithm and at last to present the results. Some technical insights are given on the implementation
of the algorithm and on the dataset used in the test phase, hence to show how the experiment has
been carried out even for reproducibility purposes; the results are then evaluated empirically and
widely discussed, comparing these with the performances of the Prim algorithm and the Kruskal
algorithm, launching several runs on a heterogeneous set of graphs and different theoretical models
for them.
We assume hereafter that the reader has knowledge about the cited paper as we will just recap
the theoretical results.
Key words. minimum spanning tree, sublinear time algorithms, randomized algorithm, approximation algorithm, minimum spanning tree weight, experimental evaluation
AMS subject classifications. 68W20, 68W25, 68R10
1. Introduction. We will discuss here some preliminary observations and assumptions. First of all, we observe that we need a set of graphs that satisfies the
following points:
• they should be finite and should not be multigraphs;
• they should be undirected;
• they should have weighted edges;
• the weights on the edges should be integers; since the graph is finite, it is
enough to show that there exist W such that it is the maximum weight on
the edges of the graph;1
• they might contain self-loops;
• they have to be connected (the graph has only one connected component);
• they should be represented with adjacency lists;
• they should be represented in the same manner (we need an unvarying file
format).
Unfortunately, the graphs and their representations, which can easily be found on the
Internet, don’t accomplish all of this requirements at the same time, although many
standards for the file format are available.
Given this observation, our choice was to use randomly generated graphs, hence
to implement our own graphs generator. This gives us the opportunity to generate
a wide set of connected graphs, with tunable parameters, carefully chosen looking
forward to the tests; these parameters include the number of nodes, the number of
edges and the edges weight, nonetheless the distribution law for the edges. The edges
between the nodes are step-by-step randomly constructed, respecting the connection
∗ Master’s
Degree in Computer Science,
University of Rome “Tor Vergata”
(gabriele.santi@acm.org).
† Master’s
Degree in Computer Science,
University of Rome “Tor Vergata”
(ldelaurentiis@acm.org).
1 more generally, it is enough to have a numerable set of values for the weights which is always
true when the graph is finite
1
2
GABRIELE SANTI AND LEONARDO DE LAURENTIIS
requirement. The different types of graphs that we use in our experimental evaluation
are presented afterwards.
After studying the paper we made some assumptions. One of the problem we
encountered is that the theoretical algorithm assumes to have as input only graph
G and to have direct access to the family of graphs Gi 2 ; with “direct” we intend
that no computation is required to extract Gi , which is not true. In fact, we can
show easily that a lower bound for the extraction of all the family is, at least, O(m)
(i.e. is linear on the number of edges). A naı̈ve approach would cost O(m) for the
extraction of Gi for a given i, hence O(wm) for the whole family; a better approach
could order the edges in O(m log m) and build the family in a single pass on the
edges, achieving O(m + m log m). Having as input only G, it would seem that the
algorithm is responsible for the extraction of the family, but this is not desirable due
to this lower bound that would sabotage the overall performance. Finally we decided
to consider this cost as a part of the construction of the data structure of the graph,
to be done prior to the call of the algorithm.
2. Design studies and choices.
2.1. Random Graphs Generator. As mentioned before, our choice is to implement our own graphs generator. The aim is to generate connected random graphs
with a specific set of parameters, like the number of nodes, the number of edges, the
maximum edges weight and the average degree. Moreover, we want to test how our
algorithm behaves in different environments, so we want to control the distribution
law of the edges. Keep in mind that the graph has to satisfy all the points of the
previous section, among which the connectivity requirement. Given a desired number
of nodes n and e ≥ n − 1 edges, the key-concepts under the implementation of our
connected graphs generator are the following:
• we begin by generating a random permutation of the vertices v0 , v1 , . . . vn
• we generate a random spanning tree by iteratively adding edges in this vector
in a uniform random manner; suppose we have added τ vertices to the tree
T , we randomly select a vertex s in the set {T0 , . . . , Tτ } and add a new edge
< Ts , vτ +1 >. At the end we will have an acyclic graph with n − 1 edges.
• following a certain probability law, we add the remaining e − (n − 1) edges
note that every time we add an edge, a weight is associated to it, with a value uniformly
choosen in [1, w].
We even decided to use a custom file format to save the data, which is a .ssv file,
standing for space separated values: every line of the file corresponds to two edges,
reporting the values < vs , vt , w > that means source and target nodes, edge weight.
Being the graph undirected, it is implicit that the edge < vt , vs , w > also exists.
2.2. Graph Data Structures. We implemented two versions of the algorithm:
the first time using the well-known Boost ’s BGL3 for the graphs data structures
and efficient implementations of Kruskal’s and Prim’s algorithms; the latter instead
embodies our own implementation of both the structures and all the side algorithms.
We decided to do so in order to obtain more control over the code for testing purposes
and because, at the moment (version 1.61.0), the BGL subgraphs framework presents
some still unfixed bug.
2 we recall that G is an induced subgraph of G such that the maximum weight on his edges is i
i
(e.g. Gw is exactly G)
3 Boost Graph Libray, [3]
EVALUATING A SUBLINEAR-TIME ALGORITHM FOR THE MST WEIGHT PROBLEM
3
Unfortunately, our Kruskal algorithm, although using union-find structures with
path compression4 , is not as fine-tuned as that proposed by the Boost’s libraries;
we didn’t take further time on this because on the other side, our version of Prim’s
algorithm shows the same exact performances of the Boost version and it counts as a
theoretical lower bound, being an optimal algorithm and way better than the Kruskal’s
solution. Our data structures have been called FastGraph for their optimization over
the operation required for the test. In any way our data structures can nor do “help”
the CRT5 algorithm in improving his time complexity.
2.3. Tuning the algorithm. In the implementation of the algorithm, the graph
stored in the text file is read into a FastGraph structure, which is of course a representation by adjacency lists. We want to compare the performances of CRT algorithm
with a standard MST algorithm that in addition computes the total weight.
We want to emphasize now that the CRT algorithm is based on probabilistic assumptions; some parameters, that depend asymptotically on ε should be selected carefully
in order to provide a good approximation very fast. These includes:
• r, the number of vertices uniformly choosen in the “approx-number-connectedcomponents”. This number is critical for the performances, as it directly
determines the running time of the entire algorithm.
• C, the large costant that we use to pick the number of vertices determinants
for the application of the original paper’s Lemma 4.
Both these values largely affect the overall performances because they undirectly
decide how many BFS will be carried out by the CRT algorithm; the BFSes represent
the fundamental cost driver of the entire run.
Initially in our opinion, the choice of these parameters must depend on the number
of vertices in a way that they are dynamic, keeping their dependency on ε. We tried
to untie this bond to n but having static values for this parameters showed poor
performances and a relative error exploding too fast as we grow the input instance.
As the paper says, the only requirement for r is that it is O( ε12 ) and C ∈ O( 1ε ).
The demonstration reported on the original paper bound these values in a way
that helped us on choosing the right function to compute them; in fact, we have that6
r
r
C
√ < ε < 1/2
n
w
< ε < 1/2,
n
hence we choose
r=
$ p
C=
n
wε −
ε2
1
%
∈O
1
ε2
√
1
⌊ nε − 1⌋
∈O
ε
ε
4 it is known that the Kruskal algorithm time complexity is O(m log m), but can be improved to
O(m log n) using union find structures with some euristic conveniently applied
5 Short for B. Chazelle, R. Rubinfeld, and L. Trevisan.[1]
6 see Theorem 7 and Theorem 8 of [1]
4
GABRIELE SANTI AND LEONARDO DE LAURENTIIS
2.4. Random Sequences. Another problem we encountered was the random
selection of k distinct vertices out of n with k < n; after some research, we found that
this kind of problem cannot be solved in sublinear time on the number of vertices,
which can’t be done for the same reasons exposed in the introduction about the
subgraph family. The problem here is that we can’t extract with no reinsertion k
values without using an external data structure; this structure has a linear cost for
the maintainance that depends on n. A special case is that in which n is a prime that
fulfills certain properties; the key idea is to make use of the properties of the quadratic
residues of n7 . That permits us to extract a non-repeating random value in constant
time in the range 0 . . . n − 1; sadly, this solution is not affordable here because we
need a dynamic range for each call, so that we cannot fulfill such constraints for n.
The solution we found is to use Fisher-Yates sequences8 which permits us to
prepare in advance the sequences and then get different, distinct values at each call
in constant time and with dynamic bounds. The cost of the preparation of those
sequences is linear and is not considered in the total cost.
3. Implementation choices. We choose to implement the algorithm in C++
because we wanted a language that offers the advantages of an OOP language, not
least a faster development, that had a standard library and sufficiently “near” the
machine to have a tighter control over the performances and the memory occupation
(e.g. performances high variance due to overhead of a VM or an interpreter). We
considered that absolute performances are not important here, whereas relative performances of the CRT algorithm and other MST algorithms are the focus, so finally
the choice of the language is something not of the greatest importance. The IDE tool
used for the implementation is JetBrains Clion. We also used GitHub, as our code
versioning control system.
Also, as already mentioned, we did use the Boost library to extend the STL that
C++ already offers; we did implement FastGraph in place of BGL also to address
memory issues in relation to the method Boost uses to store subgraphs of a graph; we
in fact used an advanced method to store the family of subgraphs Gi , i = 0, 1, . . . , w
that simply store the difference of vertices between them, ∆Gi := V (Gi ) − V (Gi−1 ).
It is always possibile to reconstruct every Gi because this family only has induced
subgraphs, and this cost is not taken into account in the computation of the total
time.
The main function of the implementation is in the AlgoWEB.cpp file. Launching
the program from this file allows us to run either the CRT algorithm or the Prim algorithm,or the Kruskal algorithm, and to view either the running time or the computed
weight of the Minimum Spanning Tree. It takes some argument in input, namely the
file containing the graph, a suitable value for ε and the path where to save the results
of the single run.
Besides, it is possible to run the random graph generator by the grandom utility
we developed apart. It takes many arguments in input, that you can find just by
calling
$> grandom -h|--help
3.1. Random Graphs Models. In order to give exhaustive results, we designed
grandom to produce 4 classes of random graphs:
7 this
is widely used in cryptography for “format preserving encryption”; see [5] and [6]
decribed in [4] and then redesigned for computer use in [7]
8 originally
EVALUATING A SUBLINEAR-TIME ALGORITHM FOR THE MST WEIGHT PROBLEM
5
• Erdős-Rényi model, that builds random graphs using a uniform distribution
of the edges over the vertices; his average degree d ≈ 2m
n ;
• Gaussian model, that builds random graphs with a “cluster zone” where the
edges are more frequent, hence the gaussian shape of the degree distribution;
we have still d ≈ 2m
n ;
• Barabási-Albert model, that builds random scale-free graphs. The average
degree is d ≈ m
n;
• Watts-Strogatz model, that builds random graphs with small-world properties and d ≈ 2m
n .
We want to emphasize here the fact that for the Barabási-Albert model the average degree results to be different respect to the other models; this is due to the
algorithm the scientific literature offers to build such graphs. For the other models is
always possible to add an arbitrary number of edges keeping the theoretical properties
valid; having, on the contrary, for the Barabási-Albert model a certain probability pk
at each step to successfully add an edge, it is not possible to build a graph with an
arbitrary number of edges; the user can solely give the number of vertices. But then
again, the theory states that if the algorithm starts with a complete graph of m0
vertices (hence m0 − 1 edges), it will produce a Barabási-Albert graph whose average
degree is scaled by this quantity. Our initial complete graph has m0 = 2 vertices, so
=m
we will have d ≈ m2m
n . A little insight on the proof is the following: the distribu0n
tion of the degree in the Barabási-Albert model is a power law with a cubic exponent.
Fortunately, in that case this distribution has a well defined mean. Applying a little
of arithmetic we can easily see the truthfulness of what stated previously.
This difference in the models has to be bore in mind when reading the test results, because when comparing the performances over a Barabási-Albert graph with
n vertices and m edges and any other one of a different model with same number of
vertices and edges, we will have different average degrees. Since the CRT algorithm
is designed to only depend on the latter and not on m nor n, a multiplicative factor
of 2 is to be taken into account.
Our implementation provide a memory-aware 9 subroutine that calculates the
exact value of d at each run; the values obtained so far agree with the statements
above.
4. Tests. The following section reports the results we had launching the algorithm over a variegate dataset of graphs, as described in the next paragraph.
4.1. Dataset. For each random graph model listed in the previous section we
decided to produce a dataset; each dataset has a “family” of graphs that differs one
from each other following a pattern for the parameters. Precisely, we composed the
dataset grouping sets of graphs based on the value of the parameters, following the
rules:
• build one set of graphs for each model (Uniform, Gaussian, Small-World, ScaleFree);
• every set contains in turn other sets, one for each selected value of n, i.e. the
number of nodes (5000, 30000, 55000, 80000, 105000, 130000, 155000, 180000,
205000, 230000);
• then again, for each selected value of d, i.e. the average degree (20, 100, 200);
this way we also determine the desired value for m, since we have, for our
models and for the fact that m ∝ n, a proportion between the two. It’s easy
9 it
calculates the average using the progressive mean technique, avoiding arithmetic overflow
6
GABRIELE SANTI AND LEONARDO DE LAURENTIIS
n⇀+∞
to see that d −−−−−→ 2m
n , so we have the selected values for m (10, 50, 100
times n)10 ;
• finally, a set for each selected value of w, i.e. the weight (20, 40, 60, 80).
In conclusion we have the number of models × the number of selected values for
n × the number of selected values for d (hence m) × the number of selected values for
w for a total of 4 · 10 · 3 · 4 = 480 graphs. This dataset, made of plain text files as
already described, is quite heavy: 42.1 GiB of data.
4.2. Runs. Using the dataset just described, we used a pattern for the runs,
in order to have a complete view of the behavior of the algorithm in the domain of
the various parameters; every single result consists of three plots, and we obtained,
inspired by the structure of the dataset:
• a set of results for each model;
• this set containing in turn a set for each selected value of ε (0.2, 0.3, 0.4, 0.49999);
• then a set for each selected value of d/2 ≃ m
n (10, 50, 100);
• finally a set for each selected value of w.
The first two plots report the absolute and relative error of the CRT result compared to the correct one, calculated with Prim’s algorithm; the third report the trend
of the used time. As we mentioned, we indeed used Kruskal too but it’s not reported
here for it was not faster respect to Prim’s times and unuseful for the computation of
the MST weight, since already done with Prim’s method.
We had this way a total of 3 · 4 · 3 · 4 = 144 plots, or better 48 different cases
of study. A comparison between those cases we consider meaningful finally concludes
our work, as reported below in section 5.
As mentioned before, the parameters should be selected carefully in order to
provide a good approximation very fast. In this sense ε plays a crucial role here: since
the CRT algorithm is a probabilistic algorithm and ε is indeed the driver parameter
either for performances and for the accuracy of the estimate MST weight, it does not
make sense to choose values too small for this parameter, as the performances could
dramatically degrades (as, indeed, is expected). So, although it could be of interest
to study the right limit of 1/2, we empirically noted that values of ε below 0.2 shows
undesired behaviour of the CRT algorithm, either for the computed MST weights and
for the running times. This is not unusual dealing with theoretical algorithms that
shows asymptotical properties; the class this algorithm belongs to is known as property
testing algorithms, whose requirement is to have a query complexity much smaller than
the instance size of the problem. Given that and the fact that the algorithm is not
required to compute an exact value, but to estimate a value probabilistically near to
the correct one11 , we are not surprised that if the instance of the problem is small, the
algorithm shows worse performances respect to a “deterministic” method. Because of
this, results for ε < 0.2 were not comprehensive nor interesting and are not reported.
Given the intrinsically probabilistic nature of the CRT algorithm, we had to
launch several runs on the same graph to have a good estimate of the CRT running
time. For this purpose we decided to launch 10 runs for each graph of the dataset,
and to take the average running time as the estimate of the running time; the amount
of runs has been decided as a compromise between having a low variance between
execution time’s mode and mean values, and keeping a restrained amount of tests to
do over the whole dataset. For the output value of the approximated MST weight we
10 operatively
11 iff
we fixed the desired number of edges m, so d is a random variable with mean
the property we are looking for is a probabilistically checkable proof, see [2]
2m
n
EVALUATING A SUBLINEAR-TIME ALGORITHM FOR THE MST WEIGHT PROBLEM
7
instead took one over the ones computed, randomly, to preserve the information on
the tolerance of the estimate.
5. Results. Following there are considerations about the meaningful line charts
regarding the results.
dw
We know that the CRT time complexity is O(dwε−2 log
), where d is the
ε
average degree of a graph; on the other hand, we know that the accuracy depends
on ε also, so we expect an inversely proportional relation with the running time.
Therefore what we expect is that:
• by increasing one or more of the parameters d, w, there should be a worsening
of the average running time;
• keeping all the other parameters unchanged, if we consider an increase in the
value of ε there must be an improvement of the running time (as well as a
worsening of the result, though);
• viceversa we expect the opposite behaviour if we decrease d and/or w or
decrease ε with the other parameters unchanged.
Let us call the above crucial parameters.
What we still need to know is what happens to the error; we should expect a
direct proportion with the running times, so the above considerations could be also
valid for the error. On the contrary, we’ll see that this is not exactly respected.
First thing to show is that the CRT algorithm is sublinear in the number of
edges, hence is better than any other exact algorithm. This can be easily seen in
figures from 5.1a to 5.1c. For the rest of the plots listed from now on it is possibile
to see, above each graph, the values of the parameters for the presented run.
It is interesting to note that the correct value, computed with Prim’s algorithm,
it’s linear in the number of edges (figure 5.1b); we want to stress the fact that this
is not the general case, and that this trend is due to the uniform distribution of the
weights over the edges. We will dissect this point more in section 8.
Given that the CRT Algorithm respects the sub-linearity constraint, let’s now see
the variations that occur when selectively changing other parameters. For the sake
of completeness, in figures 5.1 informations about Kruskal’s runs are reported, yet we
won’t report them in the charts that follow.
5.1. Variations of crucial parameters.
5.1.1. Average degree d. Let us now see the behaviour for the variation of d;
we will initially concentrate on the running time and solely for the uniform model.
The selected values of ε and w are respectivey 0.3 and 40.
As we can see in figures 5.2a to 5.2c, there is a worsening in the performance for
small instances: an increase of the average degree d is somewhat correlated to a loss
of performances to the point that our property testing algorithm needs more time
that the deterministic one; still, that seems to be true under a certain dimension of
the instance of the graph, so that we don’t lose the truthfulness of the theoretical
time complexity because for a fixed d∗ it will always be possibile to find empirically a
certain number of edges m∗ ∝ d∗ beyond which the running time function is always
dw
below C · dwε−2 log
for a certain C.12 We want here to highlight the crucial fact
ε
that the algorithm behaves better on big instances, where with “big” we refer to the
parameters the performances depend on, namely d, w. Almost all the trends reported
12 that
comes directly from the definition of asymptotic complexity
8
GABRIELE SANTI AND LEONARDO DE LAURENTIIS
Eps 0.3, m/navg 100, Weight 40
12
10
time (s)
8
Kruskal
Prim
CRT
6
4
2
2x106
4x106
6x106
# edges
8x106
1x107
(a) Running time; for completeness, Kruskal’s time is also reported.
Eps 0.3, m/navg 50, Weight 40
Eps 0.3, m/navg 50, Weight 40
140000
300000
120000
250000
MST Weight (Prim)
MST Weight (CRT)
Upper Bound
Lower Bound
150000
|mstW - mstWCRT|
|mstW - mstWCRT|
100000
200000
80000
Relative Error
Maximum Allowed Error
60000
100000
40000
50000
20000
2x106
4x106
6x106
# edges
8x106
1x107
(b) Relative error.
2x106
4x106
6x106
# edges
8x106
1x107
(c) Absolute error.
Fig. 5.1: Sublinearity of the CRT algorithm and related error trends.
in this paper, in fact, show this initial “bad” curve and, from a certain value onward,
a sublinear behaviour. We will discuss further about this in section 6 on page 20.
We would speculate that the error could be related to this. Indeed in figure 5.3
we can see that to an increase of d corresponds a dramatic annihilation of the error;
to explain this point we use a simplified example. Let us consider a minimal complete
graph13 ; the algorithm launches some BFSes on a strict subset of the graph, in more
steps that could be interrupted according to certain rules, based in part on a stochastic
process. It is easily provable that in this kind of graphs we have a worst case of n − 1
hops between to vertices u and v; if we add a random edge to this graph, the length
of this path decrease at least of one hop. By induction we can hence prove that the
diameter of the graph decreases as |E(G)| grows. In other words having a stronger
connection within the graph (i.e. a greater probability to visit a vertex v from u in
k hops of a random walk) increases the probability to have a complete view of the
13 a
connected graph with the least number of edges, i.e. n − 1
EVALUATING A SUBLINEAR-TIME ALGORITHM FOR THE MST WEIGHT PROBLEM
Eps 0.3, m/navg 10, Weight 40
0.9
0.8
0.7
time (s)
0.6
0.5
Prim
CRT
0.4
0.3
0.2
0.1
1x106
500000
1.5x106
2x106
# edges
(a) d ≃ 20
Eps 0.3, m/navg 50, Weight 40
2.5
2
time (s)
1.5
Prim
CRT
1
0.5
2x106
4x106
6x106
# edges
8x106
1x107
(b) d ≃ 100
Eps 0.3, m/navg 100, Weight 40
6
5
time (s)
4
Prim
CRT
3
2
1
5x106
1x107
1.5x107
2x107
# edges
(c) d ≃ 200
Fig. 5.2: Behaviour for the increase of d. (Uniform model)
9
10
GABRIELE SANTI AND LEONARDO DE LAURENTIIS
Eps 0.3, m/navg 10, Weight 40
400000
350000
300000
|mstW - mstWCRT|
250000
Relative Error
Maximum Allowed Error
200000
150000
100000
50000
1x106
500000
1.5x106
2x106
# edges
(a) d ≃ 20
Eps 0.3, m/navg 50, Weight 40
140000
120000
|mstW - mstWCRT|
100000
80000
Relative Error
Maximum Allowed Error
60000
40000
20000
2x106
4x106
6x106
# edges
8x106
1x107
(b) d ≃ 100
Eps 0.3, m/navg 100, Weight 40
120000
|mstW - mstWCRT|
100000
80000
Relative Error
Maximum Allowed Error
60000
40000
20000
0
5x106
1x107
1.5x107
2x107
# edges
(c) d ≃ 200
Fig. 5.3: Error behaviour for the increase of d. (Uniform model)
EVALUATING A SUBLINEAR-TIME ALGORITHM FOR THE MST WEIGHT PROBLEM
11
graph, that is more information about the MST weight.
Moreover, we saw in our study of all the results showed in this paper, that the
performances of the CRT are completely untied from the number of vertices n and
from the number of edges m of the input graph; this suggests us also that the error
is in turn driven solely by the parameters responsible of the algorithm’s complexity,
as the results that follow are in fact going to prove.
In figure 5.4 we summarize the results for the Gaussian and Small-World models,
noticing that they equate the one we showed about the uniform model. This suggests that the algorithm complexity does not depend on the dimension and clustering
coefficient of the graphs, being those the main differences from one model to another.
In figure 5.5 we summarize instead the trend of the relative error; we see here a
slightly different evolution. We cannot conclude, as we did for the time complexity,
that the error doesn’t suffer from the different graph model. The error in fact depends
on the clustering coefficient, because it is going to grow dependently on the number
of not accomplished BFSes: each one of them cause in fact a loss of information. The
algorithm, as well explained in [1], during the BFSes phase avoid to explore nodes
that shows a high degree and even stops when encounters hubs.
In other words, having equals values of d in two different runs of the algorithm,
we see that its time complexity trend remain the same; assuming that δ is the sample
mean of the vertices degree, this tells us that the time complexity is bound to the
average of δ and, since we have a growth of the error on graphs that contains hubs,
we also conclude that the relative error is bound to the variance of δ.
5.2. Maximum weight w. Here we will manipulate the value of w, similarly to
what we have already done with the average degree. Figures 5.6 and 5.7 are hereafter
proposed; this time the other fixed values are ε = 0.4, d = 50. Still this graphs has
been build using a uniform model.
This time we see the error growing as w increase. So we see here a direct proportion of the execution time with the maximum weight, unlike the inverse proportion it had with d. This is due to the fact that every iteration of the subroutine
approx-number-connected-components that the reader can find in the original paper and remembered in pseudocode 1, adds a further approximation to the final result,
because approximates the addend ĉ that contributes to the total approximation, that
is, the final error.
We see also that the dimension of the initial curve described here so far, grows
proportionally to the worsening of the excution time’s trend. This evidence also is
observable in all the trends studied.
5.3. Error tolerance ε. We will test our algorithm for the values of ε =
0.2, 0.3, 0.4, 0.49999 over uniform generated graphs. As already explained, no values
below 0.2 are investigated. We see in figures 5.8 and 5.9 the trends.
As expected, we do note that the time trends tend to decrease as ε increases, since
we tolerate a higher error for the computed value, so the algorithm is less aggressive
on computation and takes less time.
12
GABRIELE SANTI AND LEONARDO DE LAURENTIIS
Eps 0.3, m/navg 10, Weight 40
Eps 0.3, m/navg 10, Weight 40
1
0.9
0.9
0.8
0.8
0.7
0.7
0.6
Prim
CRT
0.5
time (s)
time (s)
0.6
0.5
Prim
CRT
0.4
0.4
0.3
0.3
0.2
0.2
0.1
0.1
1x106
500000
1.5x106
2x106
1x106
500000
1.5x106
# edges
2x106
# edges
(a) gaussian, d ≃ 20
(b) small-world, d ≃ 20
Eps 0.3, m/navg 50, Weight 40
Eps 0.3, m/navg 50, Weight 40
2.5
2.5
2
2
Prim
CRT
time (s)
time (s)
1.5
1.5
Prim
CRT
1
1
0.5
0.5
2x106
4x106
6x106
# edges
8x106
1x107
2x106
(c) gaussian, d ≃ 100
4x106
6x106
# edges
8x106
1x107
(d) small-world, d ≃ 100
Eps 0.3, m/navg 100, Weight 40
Eps 0.3, m/navg 100, Weight 40
6
5
5
4
Prim
CRT
3
time (s)
time (s)
4
3
Prim
CRT
2
2
1
1
5x106
1x107
1.5x107
# edges
(e) gaussian, d ≃ 200
2x107
5x106
1x107
1.5x107
# edges
(f) small-world, d ≃ 200
Fig. 5.4: Execution times behaviour for the increase of d, different models.
2x107
EVALUATING A SUBLINEAR-TIME ALGORITHM FOR THE MST WEIGHT PROBLEM
Eps 0.3, m/navg 10, Weight 40
13
Eps 0.3, m/navg 10, Weight 40
400000
400000
350000
350000
300000
300000
Relative Error
Maximum Allowed Error
200000
|mstW - mstWCRT|
|mstW - mstWCRT|
250000
250000
Relative Error
Maximum Allowed Error
200000
150000
150000
100000
100000
50000
50000
1x106
500000
1.5x106
2x106
1x106
500000
# edges
1.5x106
2x106
# edges
(a) gaussian, d ≃ 20
(b) small-world, d ≃ 20
Eps 0.3, m/navg 50, Weight 40
Eps 0.3, m/navg 50, Weight 40
160000
140000
140000
120000
120000
100000
Relative Error
Maximum Allowed Error
80000
|mstW - mstWCRT|
|mstW - mstWCRT|
100000
80000
Relative Error
Maximum Allowed Error
60000
60000
40000
40000
20000
20000
2x106
4x106
6x106
# edges
8x106
0
1x107
2x106
(c) gaussian, d ≃ 100
4x106
6x106
# edges
8x106
1x107
(d) small-world, d ≃ 100
Eps 0.3, m/navg 100, Weight 40
Eps 0.3, m/navg 100, Weight 40
140000
120000
120000
100000
80000
Relative Error
Maximum Allowed Error
60000
|mstW - mstWCRT|
|mstW - mstWCRT|
100000
80000
Relative Error
Maximum Allowed Error
60000
40000
40000
20000
20000
0
5x106
1x107
1.5x107
2x107
0
5x106
1x107
# edges
(e) gaussian, d ≃ 200
1.5x107
2x107
# edges
(f) small-world, d ≃ 200
Fig. 5.5: Error behaviour for the increase of d, different models.
14
GABRIELE SANTI AND LEONARDO DE LAURENTIIS
Eps 0.4, m/navg 50, Weight 20
Eps 0.4, m/navg 50, Weight 40
2
2.5
2
Prim
CRT
1
time (s)
time (s)
1.5
1.5
Prim
CRT
1
0.5
0.5
2x106
4x106
6x106
# edges
8x106
1x107
2x106
4x106
(a) w = 20
6x106
# edges
8x106
1x107
(b) w = 40
Eps 0.4, m/navg 50, Weight 60
Eps 0.4, m/navg 50, Weight 80
2.5
2.5
2
Prim
CRT
1.5
time (s)
time (s)
2
1
1
0.5
0.5
2x106
4x106
6x106
# edges
8x106
1x107
Prim
CRT
1.5
2x106
(c) w = 60
Fig. 5.6: Behaviour for the increase of w. (Uniform model)
4x106
6x106
# edges
(d) w = 80
8x106
1x107
EVALUATING A SUBLINEAR-TIME ALGORITHM FOR THE MST WEIGHT PROBLEM
Eps 0.4, m/navg 50, Weight 20
15
Eps 0.4, m/navg 50, Weight 40
200000
180000
160000
140000
150000
100000
Relative Error
Maximum Allowed Error
80000
|mstW - mstWCRT|
|mstW - mstWCRT|
120000
Relative Error
Maximum Allowed Error
100000
60000
50000
40000
20000
2x106
4x106
6x106
# edges
8x106
1x107
2x106
4x106
(a) w = 20
6x106
# edges
8x106
1x107
(b) w = 40
Eps 0.4, m/navg 50, Weight 60
Eps 0.4, m/navg 50, Weight 80
250000
200000
200000
Relative Error
Maximum Allowed Error
100000
|mstW - mstWCRT|
|mstW - mstWCRT|
150000
150000
Relative Error
Maximum Allowed Error
100000
50000
50000
2x106
4x106
6x106
# edges
8x106
1x107
2x106
4x106
(c) w = 60
Fig. 5.7: Error behaviour for the increase of w. (Uniform model)
6x106
# edges
8x106
(d) w = 80
1x107
16
GABRIELE SANTI AND LEONARDO DE LAURENTIIS
Eps 0.2, m/navg 50, Weight 40
Eps 0.3, m/navg 50, Weight 40
2.5
2.5
2
2
1.5
Prim
CRT
time (s)
time (s)
1.5
Prim
CRT
1
1
0.5
0.5
2x106
4x106
6x106
# edges
8x106
1x107
2x106
4x106
(a) ε = 0.2
6x106
# edges
8x106
1x107
(b) ε = 0.3
Eps 0.4, m/navg 50, Weight 40
Eps 0.49999, m/navg 50, Weight 40
2.5
2.5
2
1.5
Prim
CRT
time (s)
time (s)
2
1.5
1
1
0.5
0.5
2x106
4x106
6x106
# edges
8x106
1x107
Prim
CRT
2x106
(c) ε = 0.4
Fig. 5.8: Behaviour for the increase of ε. (Uniform model)
4x106
6x106
# edges
8x106
(d) ε = 0.49999
1x107
Eps 0.2, m/navg 50, Weight 40
Eps 0.3, m/navg 50, Weight 40
Eps 0.4, m/navg 50, Weight 40
100000
Eps 0.49999, m/navg 50, Weight 40
200000
250000
200000
120000
80000
150000
80000
Relative Error
Maximum Allowed Error
Relative Error
Maximum Allowed Error
100000
|mstW - mstWCRT|
Relative Error
Maximum Allowed Error
|mstW - mstWCRT|
60000
|mstW - mstWCRT|
|mstW - mstWCRT|
100000
Relative Error
Maximum Allowed Error
100000
60000
40000
150000
40000
50000
50000
20000
20000
2x106
4x106
6x106
# edges
8x106
1x107
2x106
4x106
(a) ε = 0.2
6x106
# edges
8x106
1x107
2x106
4x106
(b) ε = 0.3
Eps 0.2, m/navg 50, Weight 40
6x106
# edges
8x106
1x107
2x106
(c) ε = 0.4
Eps 0.3, m/navg 50, Weight 40
6x106
# edges
8x106
1x107
(d) ε = 0.49999
Eps 0.4, m/navg 50, Weight 40
300000
4x106
Eps 0.49999, m/navg 50, Weight 40
350000
350000
300000
300000
250000
300000
250000
250000
250000
MST Weight (Prim)
MST Weight (CRT)
Upper Bound
Lower Bound
150000
200000
MST Weight (Prim)
MST Weight (CRT)
Upper Bound
Lower Bound
150000
|mstW - mstWCRT|
150000
200000
|mstW - mstWCRT|
MST Weight (Prim)
MST Weight (CRT)
Upper Bound
Lower Bound
|mstW - mstWCRT|
|mstW - mstWCRT|
200000
MST Weight (Prim)
MST Weight (CRT)
Upper Bound
Lower Bound
200000
150000
100000
100000
50000
100000
50000
2x106
4x106
6x106
# edges
8x106
(e) ε = 0.2
1x107
100000
50000
2x106
4x106
6x106
# edges
8x106
(f) ε = 0.3
1x107
50000
2x106
4x106
6x106
# edges
8x106
(g) ε = 0.4
1x107
2x106
4x106
6x106
# edges
8x106
1x107
(h) ε = 0.49999
Fig. 5.9: Absolute and relative value of the error in a Uniform dataset, as ε changes; fixed values are d¯ = 50, wmax = 40.
EVALUATING A SUBLINEAR-TIME ALGORITHM FOR THE MST WEIGHT PROBLEM
140000
17
18
GABRIELE SANTI AND LEONARDO DE LAURENTIIS
For the error trend instead we note an increase, still as expected; figures from 5.9a
to 5.9d has to be read with attention because every function graph has a different
scale. The reader must not confuse the apparent lowering of the error since it is not
the case. Looking at the other graphs about the absolute error, from 5.9e to 5.9h, we
see an expansion of the tolerance cone (light blue and ochre yellow lines) as increasing
ε means admitting an higher deviation from the correct MST weight value. Here too
the different scaling must not confuse about the increasing trend of the error.
We see that the result is coherent with the theoretical model as the error increases
with ε, but his variation is, after all, contained.
5.4. Variations of graph model. As a last comparison, instead of varying one
by one the crucial parameters and see the time and error trends, we fix all of them and
try to change the model the graph belongs to. Figures 5.10 on the next page show
those results. We see here that both uniform and small-world models keep the same
trend for the error, but the small-world one behaves slightly better on small instances.
On the contrary, the gaussian show the same time trend respect to the uniform case,
but his error has a higher growth curve. The scale-free model seems to be the worse
case both regarding time and error trend, but we might remember, as observed earlier
at 3.1 on page 5, that a real term of comparison requires to considerate a scale factor
of 2 as done in figure 5.11 on page 20. We see in fact that, compared to the uniform
case, the scale-free model has even a better behaviour, looking carefully at the function
graph scale: but after all, both of them show sublinear complexity for instances of
more than 107 edges, so have both the same transients as the same steady state trend.
Another thing we note is that a bad trend in execution times are always bond to
an explosion of the error, as we can see in the charts so far. This means that using
more time to compute the value doesn’t mean it will be nearer to the correct one.
Eps 0.4, m/navg 50, Weight 60
Eps 0.4, m/navg 50, Weight 60
Eps 0.4, m/navg 50, Weight 60
Eps 0.4, m/navg 50, Weight 60
3.5
2.5
2.5
3
2.5
2
2.5
2
Prim
CRT
1.5
time (s)
Prim
CRT
1.5
time (s)
time (s)
time (s)
2
Prim
CRT
1.5
2
Prim
CRT
1.5
1
1
0.5
0.5
1
1
0.5
0.5
2x106
4x106
6x106
# edges
8x106
1x107
2x106
(a) uniform, time trend
4x106
6x106
# edges
8x106
1x107
2x106
(b) gaussian, time trend
Eps 0.4, m/navg 50, Weight 60
4x106
6x106
# edges
8x106
1x107
2x106
(c) small world, time trend
Eps 0.4, m/navg 50, Weight 60
4x106
6x106
# edges
8x106
1x107
(d) scale free, time trend
Eps 0.4, m/navg 50, Weight 60
Eps 0.4, m/navg 50, Weight 60
250000
250000
200000
200000
200000
200000
150000
Relative Error
Maximum Allowed Error
|mstW - mstWCRT|
100000
|mstW - mstWCRT|
|mstW - mstWCRT|
Relative Error
Maximum Allowed Error
Relative Error
Maximum Allowed Error
100000
100000
50000
4x106
6x106
# edges
8x106
1x107
(e) uniform, relative error
150000
Relative Error
Maximum Allowed Error
100000
50000
50000
2x106
|mstW - mstWCRT|
150000
150000
50000
2x106
4x106
6x106
# edges
8x106
1x107
(f) gaussian, relative error
0
2x106
4x106
6x106
# edges
8x106
1x107
(g) small world, relative error
2x106
4x106
6x106
# edges
8x106
1x107
(h) scale free, relative error
Fig. 5.10: Different behaviours for different graph models; fixed values are ε = 0.4, d¯ = 50, wmax = 60.
EVALUATING A SUBLINEAR-TIME ALGORITHM FOR THE MST WEIGHT PROBLEM
3
19
20
GABRIELE SANTI AND LEONARDO DE LAURENTIIS
Eps 0.4, m/navg 100, Weight 60
Eps 0.4, m/navg 50, Weight 60
5
3.5
4.5
3
4
2.5
3.5
Prim
CRT
2.5
time (s)
time (s)
3
2
Prim
CRT
1.5
2
1.5
1
1
0.5
0.5
5x106
1x107
1.5x107
# edges
(a) uniform, d¯ = 100
2x107
2x106
4x106
6x106
# edges
8x106
(b) scale free, d¯ = 50
Fig. 5.11: A more correct comparison between the general case and scale-free graphs
(ε = 0.4, wmax = 60.)
6. A specific case of study. At this point, all of our graphs show a bad initial curve every time we burden the input instance; in a specific case of study this
anomalous curve was so persistent that all the function graph was more than linear;
we decided to deepen this special case of study, performing a longer run to see the tail
of this particular trend. The results are reported on figure 6.1 on the facing page.
We can see that the original case of study hinted a sublinear trend beyond 20
millions edges instances, but we considered to investigate further, and figure 6.1b
confirms the sublinearity. We have a good behaviour on the error trend (figure 6.1c).
7. Conclusions and final thoughts. As expected, a probabilistic algorithm
like the CRT allows us to compute an approximation of the Minimum Spanning Tree
weight in sublinear time on the number of edges, under certain conditions. Tunable
parameters, that depends on ε, allows us to perform either a better or a worse approximation, implying respectively a very slow and a very fast computation. The choice of
a small value of ε can lead to terrible running times, and for these values it does not
make sense to compare the CRT algorithm with any other deterministic algorithm.
For other ε values, instead, we prove the good performances of the CRT. The reader
can easily view the better performances of CRT algorithm versus Prim algorithm or
Kruskal algorithm watching the line charts in the previous section of this paper.
More in general, we see that execution time and error depend on the number of
BFSes successfully completed during the computation The more BFSes are completed,
the more information the algorithm has to predict a correct value for the MST weight,
but on the other hand, completing the BFSes takes time. If instead we have a lot of
interrupted BFSes, we waste a large amount of time without gathering information,
hence resulting in both high execution time and error.
We considered so far different theoretical graph models, and we conclude that a
high clustering coefficient tends to increase the probability to have interrupted BFSes.
This because one of the reasons the algorithm has to interrupt the search is having
encountered a hub, i.e. a vertex with a high degree. We saw in fact that when
1x107
EVALUATING A SUBLINEAR-TIME ALGORITHM FOR THE MST WEIGHT PROBLEM
21
Eps 0.3, d 100, Weight 80
18
16
14
Eps 0.3, m/navg 100, Weight 80
16
12
time (s)
14
12
10
Prim
CRT
8
time (s)
10
Prim
CRT
8
6
6
4
4
2
2
5x106
1x107
1.5x107
1x107
2x107
2x107
3x107
4x107
# edges
# edges
(a) Heavy case of uniform, ε = 0.3, d ≃ 200, w =
80
(b) Long run, up to 50 millions edges
Eps = 0.3, d 100, Weight 80
300000
|mstW - mstWhat|
250000
200000
Relative Error
Maximum Allowed Error
150000
100000
50000
1x107
2x107
3x107
4x107
5x107
# edges
(c) Long run, error trend
Fig. 6.1: Long run results; figure (a) shows the original case of study, while (b) his
extension.
changing the model there is a slight perturbation of the trend, although it remains
sublinear. The key concept is the “distance” from the hubs of the graph of the root
vertex from which our BFS starts.
Generally we also concluded that increasing the average degree let our algorithm
gather more information, because there is a growing of the probability to visit the
generic node u of our graph. On the other side, increasing the maximum weight correspond to an increase in the number of iterations our algorithm does, that leads to
summing more intermediate approximated results that imply a higher final approxi-
5x107
22
GABRIELE SANTI AND LEONARDO DE LAURENTIIS
mation.
7.1. Parallel implementation. We observe that the CRT algorithm lends itself
very well to a parallel implementation. Indeed the majority of the algorithm’s code
is organized into independent sections, and in most cases they don’t need to comunicate to each other. We also observe that three levels of paralellism can be achieved
within the code. In the first level we parallelize each of the w independent calls to
approx-number-connected-components, as depicted in pseudocode 1, below; every
of this calls internally performs r independent BFSes from r different roots, that could
in turn run in parallel, achieving a second level of parallelism. Moreover, considering
that in the academic world there already exist different parallel implementations of
the BFS algorithm, we can use one of them to perform an additional third level of
paralellism.
Algorithm 1 First level parallelism for the CRT algorithm
function approx-MST-weight(G, ε)
⊲ G - input graph, ε - error tolerance
d∗ ← approx-avg-degree(ε) ⊲ sequential, runs in O(d/ε) as shown in [1]
ĉ ← 0
for i = 1, . . . , i = w do
⊲ this w − 1 calls can be run in w − 1 parallel threads
ĉ += approx-number-connected-components(G(i) , ε, d∗ )
end for
return v̂ ← n − w + ĉ
end function
To make it even more simpler, the number of different flows is known a priori, so
a static pre-instantiation and a smart scheduling of the threads can be performed. At
the first and second levels a master-slave model can be used in a fork-join structure,
while in the third level a shared variable is needed between the different BFSes. As
a final remark, parallelizing the BFSes could have too much overhead given that the
algorithm is optimized to run them very fast and to stop the ones that seem to cost
too much.
8. Future prospects. During an e-mail exchange with one of the original authors of [1], Dr. Ronitt Rubinfeld, another topic of discussion and study has emerged,
about the distribution of the weight on the edges.
In our code, the generation of random graphs only assume a uniform distribution
of the weights on the edges, i.e. a given edge can have an integer weight k ∈ [1, w] with
probability w1 . That implies a linear growth of the dimension of E(Gi ), ∀i ∈ [1, w],
namely the set of Gi ’s edges; this is well depicted in figure 8.1a, where, as i grows, the
size of E(Gi ) increases at each step of a quantity “near” |E(G)|
w , and it is more true
as |E(G)| is big for the law of large numbers. On the other side, having a generic law
of distribution for the edges weight implies having a different behavior as depicted
on 8.1b.
This difference could be of interest because it means that the input of different subsequent iterations of approx-number-connected-components will have a non
regularly increasing size, or even the same size for some calls; it can be easily shown
indeed that the function in 8.1 is nondecreasing. Since the cost of those calls finally
determines the overall cost, we might argue that this could lead to a minor difference,
but at the same time think that only with another set of tests we could conclude
something relevant about this observation.
23
16
16
14
14
12
12
10
10
# of edges of Gi
# of edges of Gi
EVALUATING A SUBLINEAR-TIME ALGORITHM FOR THE MST WEIGHT PROBLEM
8
6
more or less same gap
4
8
6
different gaps
4
2
2
0
0
0
1
2
3
4
5
6
7
8
0
i
1
2
3
4
5
6
7
i
(a) Assuming a uniform distribution
(b) Assuming a generic distribution
Fig. 8.1: Expected behaviour with different laws for the distribution of weights.
REFERENCES
[1] B. Chazelle, R. Rubinfeld, and L. Trevisan, Approximating the minimum spanning tree
weight in sublinear time. SIAM J Computing, 34, 2005.
[2] R. Rubinfeld, A. Shapira, Sublinear Time Algorithms. SIAM Journal on Discrete Mathematics, 2011, Vol. 25, No. 4 : pp. 1562-1588.
[3] Beman Dawes,
David Abrahams,
Rene
Rivera,
Boost C++ Libraries.
http://www.boost.org/.
[4] Fisher, Ronald A.; Yates, Frank (1948) [1938], Statistical tables for biological, agricultural
and medical research (3rd ed.). London: Oliver & Boyd. pp. 2627.
[5] Lenore Blum, Manuel Blum, Mike Shub; Comparison of Two Pseudo-Random Number
Generators. Advances in Cryptology: Proceedings of CRYPTO ’82 pp. 61-78, Plenum
1982.
[6] John Black, Phillip Rogaway; Ciphers with Arbitrary Finite Domains. Topics in Cryptology
- CT-RSA 2002, The Cryptographer’s Track at the RSA Conference, 2002, San Jose, CA,
USA, February 18-22, 2002, Proceedings, pp. 114-130
[7] Durstenfeld, R. (July 1964), “Algorithm 235: Random permutation”. Communications of
the ACM.
8
| 8cs.DS |
Quantum Information Processing manuscript No.
(will be inserted by the editor)
Two Gilbert-Varshamov Type Existential Bounds for
Asymmetric Quantum Error-Correcting Codes
arXiv:1705.04087v2 [quant-ph] 19 Oct 2017
Ryutaroh Matsumoto
19 October 2017
Abstract In this note we report two versions of Gilbert-Varshamov type existential bounds
for asymmetric quantum error-correcting codes.
Keywords asymmetric error · quantum error correction
PACS 03.67.Pp
Mathematics Subject Classification (2010) 94B65
1 Introduction
Quantum error-correcting codes (QECC) are important for construction of quantum computers, as the fault-tolerant quantum computation is based on QECC [13]. There are two
kinds of errors in quantum information, one is called a bit error and the other is called a
phase error. Steane [16] first studied the asymmetry between probabilities of the bit and
the phase errors, and he also considered QECC for asymmetric quantum errors, which are
called asymmetric quantum error-correcting codes (AQECC). Research on AQECC has become very active recently, see [7, 9, 16] and the references therein.
On the other hand, in the study of error-correcting codes, it is important to know the
optimal performance of codes. For classical error-correcting codes, the Gilbert-Varshamov
(GV) bound [11] is a sufficient condition for existence of codes whose parameters satisfies
the GV bound. By the GV bound, one can know that the optimal performance of classical
codes is at least as good as the GV bound.
For QECC, Ekert and Macchiavello obtained a GV type existential bound for general
QECCs. An important subclass of general QECCs is the stabilizer codes [2, 3, 8], as they
enable efficient encoding and decoding. Calderbank et al. [2] obtained a GV type existential
bound for the stabilizer QECCs. After that, Feng and Ma [6] and Jin and Xing [10] obtained
improved versions of GV type bounds for the stabilizer QECCs.
This note will be published in Quantum Information Processing with DOI:10.1007/s11128-017-1748-y
Ryutaroh Matsumoto
Department of Communication and Computer Engineering, Nagoya University, 464-8603 Japan.
http://www.rmatsumoto.org/research.html
2
Ryutaroh Matsumoto
The Calderbank-Shor-Steane (CSS) QECCs [4, 15] are an important subclass of the stabilizer QECCs, as the CSS codes enable more efficient implementation of the fault-tolerant
quantum computation than the stabilizer codes.
Those existential bounds [2, 4, 5, 6, 10] did not consider the asymmetric quantum errors,
while the asymmetry in quantum errors is important in practice [14]. As far as the author
know, nobody has reported existential bounds for the stabilizer or the CSS QECC for asymmetric quantum errors. In this note we report such ones. Our proof arguments are similar to
ones in [2, 4].
2 A GV type existential bound for the CSS codes
An [[n, k, d x , dz ]]q QECC encodes k q-ary qudits into n q-ary qudits and detects up to d x bit
errors and up to dz phase errors. It is known [1, 3] that a nested classical code C 2 ⊂ C 1 ⊂ Fnq
with dimensions k2 and k1 can construct an [[n, dim C 1 − dim C 2 ]]q CSS code, where Fq is
a finite field with q elements. A quantum error can be expressed as a pair (e x , ez ), where
e x ∈ Fnq corresponds to the bit error component of a quantum error and e x ∈ Fnq does to the
phase error component.
Let GLn (Fq ) be the group of n × n invertible matrices over Fq . Let Bn = {(C 1 , C 2 ) |
C 2 ⊂ C 1 ⊂ Fnq , dim C 1 = k1 , dim C 2 = k2 }. For a nonzero vector e ∈ Fnq , let Bn,x (e) (resp.
Bn,z (e)) be the set of nested code pairs that cannot detect e as a bit error (resp. a phase error),
that is, Bn,x (e) = {(C 1 , C 2 ) ∈ Bn | e ∈ C 1 \ C 2 } (resp. Bn,z (e) = {(C 1 , C 2 ) ∈ Bn | e ∈ C 2⊥ \ C 1⊥ }),
where C 1⊥ is the dual code of C 1 with respect to the standard inner product.
Lemma 1 For nonzero e, we have
qk1 − qk2
♯Bn ,
qn − 1
qn−k2 − qn−k1
♯Bn .
♯Bn,z (e) =
qn − 1
♯Bn,x (e) =
Proof As each pair C 2 ⊂ C 1 has ♯C 1 \ C 2 = qk1 − qk2 undetectable errors, we have
P
0,e∈Fnq ♯Bn,x(e)
= qk1 − qk2 .
♯Bn
For nonzero e1 , e2 ∈ Fnq , we claim ♯Bn,x (e1 ) = ♯Bn,x(e)2 . Assuming the claim, we have
X
♯Bn,x (e) = (qn − 1)♯Bn,x (e).
0,e∈Fnq
Combining these two equalities, we have
♯Bn,x (e) =
qk1 − qk2
♯Bn .
qn − 1
We finish the proof by proving the claim. Let e1 , e2 be nonzero vectors. We have
♯Bn,x (e1 ) = ♯{(C 1 , C 2 ) ∈ Bn | e1 ∈ C 1 \ C 2 }
= ♯{(τC 1 , τC 2 ) | τ ∈ GLn (Fq ), e1 ∈ C 1 \ C 2 }
= ♯{(τC 1 , τC 2 ) | τ ∈ GLn (Fq ), τ′ e1 ∈ C 1 \ C 2 }
= ♯{(C 1 , C 2 ) ∈ Bn | τ′ e1 ∈ C 1 \ C 2 }
= ♯Bn,x(τ′ e1 ),
Gilbert-Varshamov Bounds for Asymmetric Quantum Codes
where τ′ ∈ GLn (Fq ) such that τ′ e1 = e2 .
For phase errors, we can make a similar argument with C 2⊥ ⊃ C 1⊥ .
3
⊔
⊓
Theorem 2 Let n, k1 , k2 , d x and dz be positive integers such that
dz −1 !
d x −1 !
qn−k2 − qn−k1 X n
n
qk1 − qk2 X
i
(q − 1)i < 1,
(q
−
1)
+
i
qn − 1 i=1 i
qn − 1
i=1
(1)
then an [[n, k1 − k2 , d x , dz ]]q CSS QECC exists.
Proof Recall that each quantum error can be expressed by its bit error component e x ∈ Fnq
and its phase error component ez ∈ Fnq . The bit error component e x cannot be detected by
codes in Bn,x (e x ) and the phase error component ez cannot be detected by codes in Bn,z (ez ).
The detectabilities of the bit errors and the phase errors are independent of each other. Therefore, if Eq. (1) holds then there exists at least one (C 1 , C 2 ) ∈ Bn that can detect all the bit
errors with weight up to d x − 1 and all the phase errors with weight up to dz − 1, which
implies it is an [[n, k1 − k2 , d x , dz ]]q quantum code. ⊓
⊔
Classical coding theorists often have interest in asymptotic versions of GV type existential bounds [11]. They are stated in terms of information rate and relative distance of
classical error-correcting codes. In the classical error correction, information rate is the ratio
of the number of information symbols to the code length, and relative distance is the ratio
of the minimum distance to the code length.
We can also derive an asymptotic version of Theorem 2. For an [[n, k, d x , dz ]]q AQECC,
we may define the relative distance δ x for bit errors as d x /n, and the relative distance δz for
bit errors as dz /n. The information rate of an [[n, k]]q QECC is defined as k/n [13].
Recall [11] that for 0 ≤ δ ≤ 1 − 1/q we have
!
⌊nδ⌋
X
n
(q − 1)i ≤ qnhq (δ) ,
i
i=1
(2)
where hq (δ) = δ logq (q − 1) − δ logq δ − (1 − δ) logq (1 − δ).
Corollary 3 Let δ x and δz be real numbers such that 0 ≤ δ x ≤ 1 − 1/q and 0 ≤ δz ≤ 1 − 1/q.
If
hq (δ x ) < 1 − R1 ,
(3)
hq (δz ) < R2 , and
(4)
0 ≤ R1 − R2 ,
then, for sufficiently large n, there exists an [[n, ⌊nR1 ⌋ − ⌈nR2 ⌉, ⌊nδ x ⌋, ⌊nδz ⌋]]q CSS QECC
exists.
In Corollary 3, R1 is the information rate of classical ECC C 1 , and R2 is the information rate
of classical ECC C 2 . The corresponding quantum CSS code has information rate R1 − R2 ,
relative distance δ x for bit errors, and relative distance δz for phase errors.
4
Ryutaroh Matsumoto
Proof Assume that Eq. (3) holds. Then for sufficiently large n we have
nhq (δ x ) < n − nR1
qn
⇒ qnhq (δx ) < (1/2) nR1
q
nR1
q
⇒ n qnhq (δx ) < 1/2
q
!
⌊nδ x ⌋−1
n
q⌊nR1 ⌋ − q⌈nR2 ⌉ X
(q − 1)i < 1/2.
⇒
i
qn − 1
i=1
(5)
Similarly, for sufficiently large n Eq. (4) implies
nhq (δz ) < nR2
⇒ qnhq (δz ) < (1/2)
n(1−R2 )
⇒
⇒
q
qn
qn
qn(1−R2 )
qnhq (δz ) < 1/2
!
⌊nδz ⌋−1
qn−⌈nR2 ⌉ − qn−⌊nR1 ⌋ X n
(q − 1)i < 1/2.
i
qn − 1
i=1
(6)
Equations (5) and (6) imply that the assumption of Theorem 2 becomes true for sufficiently
large n, which shows Corollary 3. ⊓
⊔
3 A GV type existential bound for the stabilizer codes
Let C ⊂ F2n
q be a Fq -linear space of dimension n − k self-orthogonal with respect to the
standard symplectic inner product in F2n
q . C can be viewed as an [[n, k]]q stabilizer QECC.
Let An be the set of all such C’s. A nonzero e ∈ F2n
q can be viewed as a quantum error
on n qudits. Let An (e) be the set of stabilizer codes that cannot detect e as an error, that is,
An (e) = {C ∈ An | e ∈ C ⊥s \ C}, where C ⊥s is the dual of C with respect to the symplectic
1−q−2k
1
inner product. Then ♯An (e) ≤ 1−q
−2n · qn−k ♯An [12, Lemma 9].
Recall that, for C to be [[n, k, d x , dz ]]q , C must be able to detect all d x or less bit errors
and all dz or less phase errors. The number of such errors is
dX
x −1
i=1
!
!
dX
z −1
n
n
(q − 1)i ×
(q − 1)i .
i
i
i=1
By the same argument as [12][Remark 10] (or as the last section), we have the following
theorem:
Theorem 4 Let n, k1 , k2 , d x and dz be positive integers such that
!
dX
d x −1 !
z −1
1 − q−2k
1 X
n
n
i
·
(q − 1)i < 1
(q
−
1)
×
i
1 − q−2n qn−k i=1 i
i=1
then there exists an [[n, k, d x , dz ]]q stabilizer QECC. ⊓
⊔
Gilbert-Varshamov Bounds for Asymmetric Quantum Codes
5
By almost the same argument as Corollary 3 we can derive the following asymptotic
version of Theorem 4.
Corollary 5 Let δ x and δz be real numbers such that 0 ≤ δ x ≤ 1 − 1/q and 0 ≤ δz ≤ 1 − 1/q.
If
hq (δ x ) + hq (δz ) < 1 − R ≤ 1,
(7)
then, for sufficiently large n, there exists an [[n, ⌊nR⌋, ⌊nδ x ⌋, ⌊nδz ⌋]]q stabilizer QECC. ⊓
⊔
The quantum statilizer code in Corollary 5 has information rate R, relative distance δ x for
bit errors, and relative distance δz for phase errors.
By the relation between the CSS and the stabilizer QECCs [3], we see that the assumption in Corollary 3 is less demanding than that in Corollary 5 for the same n, R = R1 − R2 ,
δ x and δz , which means that Corollary 5 is a stronger existential bound than Corollary 3.
Remark 6 Theorems 2 and 4, and Corollaries 3 and 5 do not admit direct comparisons
against previously known GV type bounds even when d x = dz . The reason is as follows:
For a binary QECC to be [[n, k, 2, 2]]2 , it must detect at least n2 different errors. On the
other hand, for a binary [[n, k]]2 QECC to detect all single symmetric errors, it only has to
detect 3n errors, which is generally much fewer than n2 . The above example shows that the
number of asymmetric quantum errors is much different from that of corresponding symmetric quantum errors, even if we assume the same number of bit errors and phase errors in
asymmetric quantum errors.
In addition, the famous [[5, 1, 3]]2 binary stabilizer code in [3, 8] can detect up to four
bit errors if there is no phase error, and can detect up to four phase errors if there is no bit
error. Thus it is simultaneously both [[5, 1, 1, 5]]2 AQECC and [[5, 1, 5, 1]]2 AQECC. This
phenomenon makes the direct comparison even more difficult.
Acknowledgements The author would like to thank an anonymous reviewer for careful reading and the helpful report that improved this note, and Prof. Fernando Hernando for drawing his attention to the asymmetric
quantum error correction.
References
1. Ashikhmin, A., Knill, E.: Nonbinary quantum stabilizer codes. IEEE Trans. Inform. Theory 47(7),
3065–3072 (2001)
2. Calderbank, A.R., Rains, E.M., Shor, P.W., Sloane, N.J.A.: Quantum error correction and orthogonal
geometry. Phys. Rev. Lett. 78(3), 405–408 (1997)
3. Calderbank, A.R., Rains, E.M., Shor, P.W., Sloane, N.J.A.: Quantum error correction via codes over
GF(4). IEEE Trans. Inform. Theory 44(4), 1369–1387 (1998)
4. Calderbank, A.R., Shor, P.W.: Good quantum error-correcting codes exist. Phys. Rev. A 54(2), 1098–
1105 (1996)
5. Ekert, A., Macchiavello, C.: Quantum error correction for communication. Phys. Rev. Lett. 77(12),
2585–2588 (1996)
6. Feng, K., Ma, Z.: A finite Gilbert-Varshamov bound for pure stabilizer quantum codes. IEEE Trans.
Inform. Theory 50(12), 3323–3325 (2004). DOI 10.1109/TIT.2004.838088
7. Galindo, C., Geil, O., Hernando, F., Ruano, D.: Improved constructions of nested code pairs. IEEE Trans.
Inform. Theory (2017). DOI 10.1109/TIT.2017.2755682
8. Gottesman, D.: Class of quantum error-correcting codes saturating the quantum Hamming bound. Phys.
Rev. A 54(3), 1862–1868 (1996)
9. Ioffe, L., Mézard, M.: Asymmetric quantum error-correcting codes. Phys. Rev. A 75, 032345 (2007).
DOI 10.1103/PhysRevA.75.032345
6
Ryutaroh Matsumoto
10. Jin, L., Xing, C.: Quantum Gilbert-Varshamov bound through symplectic self-orthogonal codes. In:
Proc. 2011 IEEE International Symposium on Information Theory, pp. 455–458. Sait Petersburg, Russia
(2011). DOI 10.1109/ISIT.2011.6034167
11. MacWilliams, F.J., Sloane, N.J.A.: The Theory of Error-Correcting Codes. Elsevier, Amsterdam (1977)
12. Matsumoto, R., Uyematsu, T.: Lower bound for the quantum capacity of a discrete memoryless quantum
channel. Journal of Mathematical Physics 43(9), 4391–4403 (2002). DOI 10.1063/1.1497999
13. Nielsen, M.A., Chuang, I.L.: Quantum Computation and Quantum Information. Cambridge University
Press, Cambridge, UK (2000)
14. Sarvepalli, P.K., Klappenecker, A., Rötteler, M.: Asymmetric quantum codes: constructions, bounds and
performance. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences
465(2105), 1645–1672 (2009). DOI 10.1098/rspa.2008.0439
15. Steane, A.M.: Multiple particle interference and quantum error correction. Proc. Roy. Soc. London Ser.
A 452(1954), 2551–2577 (1996)
16. Steane, A.M.: Simple quantum error-correcting codes. Phys. Rev. A 54(6), 4741–4751 (1996). DOI
10.1103/PhysRevA.54.4741
| 7cs.IT |
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, DECEMBER 2017
1
Automatic Classification of Functional Gait
Disorders
arXiv:1712.06405v2 [cs.CV] 24 Dec 2017
Djordje Slijepcevic, Matthias Zeppelzauer, Anna-Maria Gorgas, Caterine Schwab, Michael Schüller, Arnold Baca,
Christian Breiteneder, Brian Horsak
Abstract—This article proposes a comprehensive investigation
of the automatic classification of functional gait disorders based
solely on ground reaction force (GRF) measurements. The aim
of the study is twofold: (1) to investigate the suitability of stateof-the-art GRF parameterization techniques (representations) for
the discrimination of functional gait disorders; and (2) to provide
a first performance baseline for the automated classification of
functional gait disorders for a large-scale dataset. The utilized
database comprises GRF measurements from 279 patients with
gait disorders (GDs) and data from 161 healthy controls (N).
Patients were manually classified into four classes with different functional impairments associated with the “hip”, “knee”,
“ankle”, and “calcaneus”. Different parameterizations are investigated: GRF parameters, global principal component analysis
(PCA)-based representations and a combined representation
applying PCA on GRF parameters. The discriminative power
of each parameterization for different classes is investigated by
linear discriminant analysis (LDA). Based on this analysis, two
classification experiments are pursued: (1) distinction between
healthy and impaired gait (N vs. GD) and (2) multi-class
classification between healthy gait and all four GD classes.
Experiments show promising results and reveal among others
that several factors, such as imbalanced class cardinalities and
varying numbers of measurement sessions per patient have a
strong impact on the classification accuracy and therefore need to
be taken into account. The results represent a promising first step
towards the automated classification of gait disorders and a first
performance baseline for future developments in this direction.
Index Terms—Ground Reaction Force (GRF), Gait Classification, Principal Component Analysis (PCA), Gait Parameters,
Machine Learning
I. I NTRODUCTION
AIT analysis is a tool for clinicians to objectively
quantify human locomotion and to describe and analyze
a patient‘s gait performance. The primary aim is to identify
impairments that affect a patient’s gait pattern [1].
Recordings obtained during clinical gait analyses produce
a vast amount of data which are difficult to comprehend
and analyze due to their high-dimensionality, temporal dependences, strong variability, non-linear relationships and correlations within the data [2]. This makes data interpretation
challenging and requires an experienced clinician to draw valid
conclusions. Several automatic analysis approaches based on
G
Manuscript received –; revised –.
Djordje Slijepcevic and Matthias Zeppelzauer are with St. Pölten University
of Applied Sciences, Austria and TU Wien, Austria.
Anna-Maria Gorgas, Caterine Schwab, and Brian Horsak are with St. Pölten
University of Applied Sciences, Austria.
Michael Schüller and Arnold Baca are with University of Vienna, Austria.
Christian Breitender is with TU Wien, Austria.
machine learning have been published in recent years to
tackle these problems and to support clinicians in identifying
and categorizing specific gait patterns into clinically relevant
categories [2], [3]. Machine learning methods employed in
this context comprise neural networks [4]–[6], support vector
machines (SVMs) [7]–[9], nearest neighbor classifiers [10],
[11], and different clustering approaches (hierarchical, kmeans, etc.) [12]. The performance of such methods strongly
depends on the input data representation [13]. Frequently used
representations in gait analysis comprise discrete kinematic
gait parameters (e.g. local minima and maxima of gait signals
and time-distance parameters) [11], [14], [15]. Additionally,
previous research has shown that global signal representations
obtained by principal component analysis (PCA) [16], [17],
kernel-based PCA (KPCA) [18], [19] and discrete wavelet
transformation (DWT) [10], [11] are suitable for subsequent
classification [10], [16].
Typical use cases for automatic gait analysis described in the
literature show a moderate to high accuracy in distinguishing
between different pathologies or patient groups [4], [7]–[9],
[11], [16], [17]. However, most of the existing literature investigated rather simple cases such as the differentiation between
the affected/non-affected limb in hemiplegic patients [20], and
the distinction of healthy gait from people with neurological
disorders [5], [11], transfemoral amputation [16], and lower
limb fractures [4], [17]. A more complex study is presented
in [21], where several disorders associated with traumatic
brain injuries are classified. The majority of published articles
employed kinematic and kinetic data derived from threedimensional gait analysis (3DGA), which provide a vast
amount of kinematic and kinetic information for multiple
joints. Drawbacks of such 3DGA measurement systems are the
relatively time-consuming data recording, the need for highly
trained staff as well as high acquisition and maintenance costs.
Therefore, such analysis tools are often not suitable for daily
use in clinical practice.
To manage the high patient throughput in rehabilitation
centers, a frequently used approach is to combine simple visual
inspection or 2D video recordings with the quantification of
ground reaction forces (GRF) by force platforms, as changes
in the morphology of the GRF waveforms reflect pathological
gait [11], [17]. One major drawback of this approach is the
loss of clinically relevant and quantifiable information (e.g.
gait kinematics), causing a potential decrease in classification
accuracy [22]. However, such simple approaches are common
in clinical practice as they overcome the before-mentioned limitations of 3DGA. To date, few attempts have been published
©2017 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including
reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or
reuse of any copyrighted component of this work in other works. DOI: 10.1109/JBHI.2017.2785682
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, DECEMBER 2017
that use only GRF data for automated gait pattern classification
[16], [23]. Most of these gait classification approaches show
promising results. However, the majority of previous works
employed relatively small datasets. Alaqtash et al. [11], for
example, compared the data of 12 healthy adults to those of
patients with cerebral palsy and multiple sclerosis (4 patients
each), Muniz and Nadal [17] used data from 38 healthy
controls and 13 patients with lower limb fracture, and Soares
et al. [16] classified GRF data of 20 able-bodied and 12
patients with transfemoral amputation. Such small datasets
make it difficult to train robust and reliable classifiers that
are applicable in complex real-world scenarios. Furthermore,
a majority of studies [10], [17], [23] relies solely on the vertical ground reaction force for classification purposes, rather
than considering all available GRF components, including the
center of pressure (COP), for a more conclusive picture of the
underlying gait pattern. Previous classification attempts mainly
focused on the differentiation between specific diseases rather
than drawing a distinction between functional gait disorders.
The work of Köhle and Merkl [24], [25], who clustered
and classified GRF measurements into deficits of different
body regions, represents an exception in this regard. Their
dataset was about half of the size of the one presented in
this article and their work also focused on patients walking
with a prosthesis. In this article we define a functional gait
disorder as the cause of a gait impairment, which is reflected
by the individual gait patterns. These may be associated
with a patient’s condition after joint replacement surgery,
fractures, ligament ruptures, osteoarthritis or related disorders.
The classification of functional gait disorders is of particular
interest in clinical examinations, as it may play a key role
in detecting arthropathies or diseases at an early stage. In
addition, such a classification may also indicate secondary
disorders that otherwise might be easily overlooked during
clinical examination.
The aim of this article is to present a detailed investigation of the automated classification of several functional gait
disorders solely based on GRF data. The presented approach
builds upon the aforementioned studies, e.g. [16], [17], [23],
investigates the suitability of frequently used state-of-theart GRF parameterization techniques for gait classification
and analyzes their discriminative power. In the experiments
we evaluate the individual representations on a large-scale
and real-world dataset for different classification tasks. This
paper therefore presents a first performance baseline for the
automatic classification of different gait disorders in a realworld setting.
II. M ATERIAL AND M ETHODS
A. Patients and dataset
The presented retrospective study was approved by the
local Ethics Committee of Lower Austria (GS1-EK-4/2992014). The anonymized data used in this study are part of an
existing clinical gait database maintained by a rehabilitation
center of the Austrian Workers’ Compensation Board (AUVA).
The AUVA is the social insurance for occupational risks for
more than 3.3 million employees and 1.4 million pupils and
2
students in Austria. The utilized database comprises GRF
measurements from 279 patients with gait disorders (GD) and
data from 161 healthy controls (N), both of various physical
composition and gender (see Table I for details on the dataset).
Patients were manually classified into four classes - calcaneus
“C” (n = 82), ankle “A” (n = 62), knee “K” (n = 69), and
hip “H” (n = 66) - by a physical therapist, based on the
available medical diagnosis of each patient. Thus, GD refers to
C ∪ A ∪ K ∪ H. The individual GD classes include patients
after joint replacement surgery, fractures, ligament ruptures,
and related disorders associated with the above-mentioned
anatomical areas. The most common injuries present in the hip
class are fractures of the pelvis and thigh as well as luxation of
the hip joint, coxarthrosis, and total hip replacement. The knee
class comprises patients after patella, femur or tibia fractures,
ruptures of the cruciate or collateral ligaments or the meniscus
and total knee replacements. The ankle class includes patients
after fractures of the calcaneus, malleoli, talus, tibia or lower
leg, and ruptures of ligaments or the achilles tendon. The
calcaneus class comprises patients after calcaneus fractures or
ankle fusion surgery. All of the above-mentioned injuries may
occur individually or in combination within each class.
Each patient performed one or several measurement sessions. In each session, eight recordings for two consecutive
steps were performed. Each bilateral recording is referred to
as one trial in this paper. Thus, the utilized dataset contains
1,187 sessions comprising 9,496 individual trials (see Table I
for details).
B. Data recording and preprocessing
Gait analysis was performed on a 10m walkway with two
centrally embedded force plates (Kistler, Type 9281B12). The
force plates were placed in a consecutive order, allowing a
person to walk across by placing one foot on each plate. Both
plates were flush with the ground and covered with the same
walkway surface material, so that targeting was not an issue.
During a session, participants walked unassisted and without a
walking aid at a self-selected walking speed until a minimum
of eight valid recordings were available. These recordings were
defined as valid by a supervisor when the participant walked
naturally and there was a clean foot strike on the force plate.
Prior to the gait analysis session, each participant underwent
rigorous physical examination by a physician.
All processing steps and subsequent analyses were performed in Matlab 2016a (The MathWorks Inc., Natick, MA,
USA). The three analog GRF signals as well as the two COP
signals were converted to digital signals using a sampling rate
of 2000Hz and a 12-bit analog-digital converter (DT3010, Data
Translation Incorporation, Marlboro, MA, USA) with a signal
input range of ±10V. A threshold of 10N was used for step
detection and 30N for COP calculation. Raw signals were
filtered using a 2nd order low-pass butterworth filter with a cutoff frequency of 20Hz. All gait measurements were temporally
aligned so that they all started with the initial contact and
ended with toe-off. They were further time-normalized to
100% stance by re-sampling the data to 1000 points. The
processed signals are referred to as waveforms in this article.
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, DECEMBER 2017
3
TABLE I: Details of the dataset and classes
Class
Healthy Control (N)
Calcaneus (C)
Ankle (A)
Knee (K)
Hip (H)
SUM
Amount
161
82
62
69
66
440
Age (yrs.)
Mean ± SD
32.4 ± 13.6
42.4 ± 9.9
40.0 ± 11.5
41.5 ± 11.4
43.6 ± 14.7
38.4 ± 13.3
Body Mass (kg)
Mean ± SD
74.1 ± 16.2
84.5 ± 12.1
88.3 ± 16.9
83.7 ± 19.6
81.6 ± 18.3
80.7 ± 17.3
Sex
(m/f)
84/77
74/8
56/6
44/25
53/13
311/129
Num. Sessions
Num. Trials
161
320
259
258
189
1,187
1,288
2,560
2,072
2,064
1,512
9,496
Amplitude values of the three force components, i.e. vertical
(V), medio-lateral (ML), and anterior-posterior (AP), were
expressed as a multiple of body weight (BW ) by dividing
the force by the product of body mass times acceleration
due to gravity (g). The COP waveforms from each trial were
normalized by the foot length (F L) determined during each
session, expressed as a multiple of foot length.
C. Signal representation
The representations employed in our investigation comprise
(1) discrete GRF parameters (DP) in combination with timedistance parameters (TDP) [11], [14], [15]; (2) PCA-based
parameterizations of the entire GRF waveforms [4], [8], [16]
and (3) a combination of the first two approaches, i.e. PCA
applied to DPs and TDPs [7]. In the following, all three
approaches are described in detail.
DPs were calculated for the affected limb and extracted from
all three force components, FV (t) (vertical), FAP (t) (anteriorposterior), and FM L (t) (medio-lateral), as well as from the
COP displacement in the anterior-posterior (walking) direction
COPAP (t) and in the medio-lateral direction COPM L (t). An
example of the GRF and corresponding COP waveforms is
presented in Figure 1. Furthermore, a more detailed visualization of the mean GRF waveforms over each class is illustrated
in Figure S1 (supplementary material). DPs include a set of
predefined (most prominent) local minima and maxima of the
waveforms, which were extracted by peak detection in a fully
automatic way from each trial. Furthermore, impulses were
calculated over different segments of the waveform by multiplying the average force (in N ) by the time this force is active.
To account for differences in body mass between participants
[26], all impulses were divided by the product of body mass
times acceleration due to gravity (g) and then multiplied by
100 (%BW ·s). TDPs such as cadence (CAD), double support
time (DS), gait velocity (GV ), step length (ST EP LEN ), and
stance time (ST ) were calculated from two consecutive steps
(affected and unaffected limb) and averaged over the eight
valid trials. Table II lists all 52 extracted parameters.
In contrast to the GRF parameters (DPs and TDPs), the PCA
takes the entire waveforms1 of the affected limb into account
and provides a holistic representation of the data. Complementary information to the parameters is thus captured. The
main goal of PCA is to reduce the dimensionality of a dataset
by transforming the data into a set of uncorrelated variables,
i.e. the principal components (PCs) [27]. Each PC points in
1 For the purpose of the present study, every third sample was used in order
to reduce redundancy in the data, thereby improving the robustness of the
decomposition.
Fig. 1: (top) The characteristic shape of the three components
of the GRF: the vertical force (FV ), the anterior-posterior
shear (FAP ), and the medio-lateral shear (FM L ). (bottom) The
corresponding COP path for one step. Note that x and y axes
are scaled slightly differently for better visualization.
(and thus explains) one orthogonal direction of variance in
the data. The main intention is to obtain a lower-dimensional
representation of our time- and weight-normalized waveforms
similar to [4], [8], [16] by projecting the data onto those
PCs which explain most variance in the data. This dimensionality reduction fosters subsequent machine learning [3].
We performed PCA on each of the five signals separately and
concatenated the resulting PCs to obtain a feature vector for
classification. This approach proved to be superior to other
PCA-based representations in a preliminary study [28]. The
final dimensionality of the obtained representations is specified
by the amount of variance preserved in a particular projection,
i.e. 98%, 95%, and 90%. An exemplary visualization of
the different PCA representations is presented in Figure S2
(supplementary material). A preliminary evaluation indicated
that preserving 98% of the variance results in a good trade-off
between data reduction and classification performance. Thus,
all results presented in the following are based on the approach
in which 98% of the variance is preserved (the number of
resulting PCs is waveform-specific and ranges from four to
twelve. For all five signals there are 39 PCs in total).
As a third representation, PCA was applied to the previ-
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, DECEMBER 2017
4
TABLE II: Discrete and time-distance parameters, description, type of normalization and physical unit.
Abbreviation
Description
ST
Stance time is the duration of the stance phase of one foot
FV 1
Maximum value of FV within the breaking phase of stance
TV 1
Time of FV 1
FV 2
Minimum value of FV between TV 1 and TV 3
TV 2
Time of FV 2
FV 3
Maximum value of FV ) within the propulsive phase of stance
TV 3
Time of FV 3
FAP 1
Maximum value of FAP between initial contact and TAP 2
TAP 1
Time of FAP 1
FAP 2
Minimum value of FAP ) within the breaking phase of stance
TAP 2
Time of FAP 2
FAP 3
Maximum value of FAP ) within the propulsive phase of stance
TAP 3
Time of FAP 3
FM L1
Minimum value of FM L within the breaking phase of stance
TM L1
Time of FM L1
FM L2
Maximum value of FM L within the breaking phase of stance
TM L2
Time of FM L2
FM L3
Maximum value of FM L within the propulsive phase of stance
TM L3
Time of FM L3
FV AV G
Mean value of FV
FAP AV G
Mean value of FAP
FM LAV G
Mean value of FM L
IFV
Impulse of FV during stance
IFAP
Impulse of FAP during stance
IFM L
Impulse of FM L during stance
IFV 1
Impulse of FV between initial contact and TV 1
IFV 2
Impulse of FV between initial contact and TV 2
IFV 3
Impulse of FV between initial contact and TV 3
IFAP DEC
Impulse of FAP during the breaking phase
IFAP ACC
Impulse of FAP during the propulsive phase
IFLAT
Impulse of the lateral component of FM L
IFM ED
Impulse of the medial component of FM L
COP AN G
COP angle is the horizontal angle between the COP linear regression line and the x-axes (6= foot rotation)
COP DEV
COP deviation is the root mean square error of the COP linear regression
COPAP
COP range is the range in the anterior-posterior direction during stance phase
COP V
COP velocity is calculated as the ratio of foot length and stance time
COPM L
COP range is the range in the medio-lateral direction during stance phase
DECT
Deceleration time (breaking phase) is the duration of FAP being negative
ACCT
Acceleration time (propulsive phase) is the duration of FAP being positive
LR0080
Loading rate represented as the slope of FV from the initial contact to 80% of FV 1
LR2080
Loading rate represented as the slope of FV from 20% to 80% of FV 1
U R8000
Unloading rate represented as the slope of FV from 80% of FV 3 to the toe-off
U R8020
Unloading rate represented as the slope of FV from 80% to 20% of FV 3
DS
Double support time during one stride
ST EP LEN
Step length is the distance of the COP position from initial contact to following contralateral initial contact
ST EP V
Step velocity is calculated as the ratio of step length and step time
ST RIDET
Stride time is the duration from initial contact to initial contact of the ipsilateral foot
BF
Basic frequency is the mean number of strides per second (1/ST RIDET )
CAD
Cadence is the number of steps per minute
ST EP W D
Step width is the medio-lateral distance of the mean COP between both feet
ST RLEN
Stride length is the distance of the COP position from initial contact to following ipsilateral initial contact
GV
Gait velocity is calculated as the mean step velocity of both feet
Body weight (BW): product of body mass and acceleration due to gravity;
%ST: percentage of stance time; %BW: percentage of body weight; FL: multiple of foot length.
ously extracted DPs and TDPs (a vector comprising of 52
parameters), similarly to Wu et al. [7]. This approach combines
both methodologies and aims at extracting the most important
information from the (possibly redundant) parameters.
D. Statistical analysis
Our first aim was to investigate the suitability of different
parameterization techniques for subsequent gait classification.
For this purpose we analyzed the variance and discriminative
power of each DP and TDP across the different classes by
descriptive statistics in a first step. We calculated the median,
interquartile-range (IQR) and range of each parameter within
each class and visualized them as boxplots. This enabled us to
visually inspect variances and distributions in and across the
classes, thereby allowing a first estimation of the discriminative power of each parameter.
In a second step, we investigated the discriminative power
of the parameters and the global PCA-based representations
by linear discriminant analysis (LDA). A natural measure
to describe the separation of two distributions (classes) is
the Fisher criterion, which represents the core of LDA [29].
We applied (multi-class) LDA to assess the discriminative
Normalization
Body weight
Stance time
Body weight
Stance time
Body weight
Stance time
Body weight
Stance time
Body weight
Stance time
Body weight
Stance time
Body weight
Stance time
Body weight
Stance time
Body weight
Stance time
Body weight
Body weight
Body weight
Body weight
Body weight
Body weight
Body weight
Body weight
Body weight
Body weight
Body weight
Body weight
Body weight
Foot length
Foot length
Foot length
Foot length
Body weight
Body weight
Body weight
Body weight
-
Unit
s
BW
%ST
BW
%ST
BW
%ST
BW
%ST
BW
%ST
BW
%ST
BW
%ST
BW
%ST
BW
%ST
BW
BW
BW
%BW ·s
%BW ·s
%BW ·s
%BW ·s
%BW ·s
%BW ·s
%BW ·s
%BW ·s
%BW ·s
%BW ·s
deg
FL
FL
F L/s
FL
s
s
N/s
N/s
N/s
N/s
s
m
km/h
s
Hz
1/min
m
m
km/h
power of individual parameters for two (or more) classes. The
advantage of this approach is that the discriminative power of
a parameter (even across multiple classes) can be expressed
by one scalar value that directly reflects the statistical properties of the input data. Hence, there is no need to apply
additional modelling and data transformations (which may
influence results) prior to LDA, which would be necessary for
other methods such as SVM. Furthermore, this approach can
easily be extended to estimate the discriminative power of a
combination of several parameters by multi-dimensional LDA
(e.g. in case of PCA-based representations). We computed the
accuracy of LDA and reported the divergence from a random
baseline [30] to quantify to which degree an input parameter or
input representation is able to separate the underlying classes.
The random baseline was estimated by the zero rule (always
choosing the most frequent class in the dataset). Thus, in the
case of five classes where the largest class contains 30% of
the data the random baseline equals 30%.
E. Classification
We applied two classification tasks to the dataset by using SVMs as classifiers: (1) (binary) classification between
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, DECEMBER 2017
normal gait and all gait disorders (N/GD) and (2) (multiclass) classification between N and each of the four GD
classes (N/C/A/K/H). In the first task, the class priors
are imbalanced, i.e. there are many more observations in the
combined GD class than in the normal class (see Table I). The
second task separates each type of disorder from each other
and from the normal class.
For the classification experiments the dataset was split into
a training (65%) and a test set (35%), thereby mutually
disjoining the groups of patients in both sets. The training
dataset in combination with a k-fold cross-validation approach
served to train the classifiers and to optimize their parameters
(model selection), whereas the test dataset was used to evaluate
the generalization ability of the trained models (and was
not considered during model selection and hyper-parameter
optimization). The calculated DPs and TDPs as well as the
PCA-based representations served as input to classification.
The parameters (DPs and TDPs) were normalized (each independently) in a twofold way, by min-max normalization
and z-standardization, in order to determine the more suitable
approach. The PCA representations were z-standardized. We
employed SVMs for the classification with linear and radial
basis function (RBF) kernels, provided by the LIBSVM library
[31]. For hyper-parameter selection we applied a grid search
over the regularization parameter C ∈ [2−5 , 215 ] for the linear
SVM and over C ∈ [2−1 , 215 ] and the kernel hyper-parameter
γ ∈ [2−15 , 25 ] for the RBF SVM. During the grid search, a
5-fold cross-validation was performed on the training dataset.
Finally, an SVM with the best parameters estimated during
model selection was trained on the entire training set and evaluated on the test set. Additionally a k-nearest neighbor (k-NN)
classifier and a multi-layer perceptron (MLP) were employed
as a reference to compare their results to the performance of
the SVM. Grid search was performed over various values of
k for the k-NN. For the MLPs different numbers and sizes
of hidden layers were employed. As a performance measure
we use the classification accuracy, which is the percentage of
correct classifications among all classes and input samples.
Since in different experiments the random baseline varies,
the absolute values of accuracy are of limited expressiveness.
To enable a fair comparison, we employ the divergence from
a random baseline approach [30] and thus provide for each
experiment the difference between the random baseline and
the absolute classification accuracy.
III. R ESULTS AND D ISCUSSION
T
HIS section presents and discusses the results of the
statistical analysis and the classification experiments.
A. Statistical analysis
The statistical analysis aimed at assessing the suitability
of the individual GRF parameters (DPs and TDPs) for distinguishing different classes of gait disorders. In order to be
considered a ”good” parameter, intra-class variation should be
low (e.g. small IQR inside a given class), while the inter-class
variation should be high (e.g. significantly different means or
medians between the samples of different classes) [15].
5
The visual inspection of the boxplots for each parameter
enables a first assessment of the intra- and inter-class variation
and thereby gives an impression of the parameters’ potential
to differentiate between different classes. Figure 2 shows
boxplots for selected parameters. A presentation of boxplots
for all 52 investigated parameters for all classes is provided
in Figure S3 (supplementary material). Parameters such as
FV 3 (see Figure 2(a)) show a clear difference in the median
and the IQR between the healthy controls and all four GD
classes. This indicates a high potential to discriminate between
normal gait and arbitrary gait disorders. However, the overlap
of the distributions within the GD classes indicates a low
potential to discriminate between them. Other parameters such
as TAP 3 (see Figure 2(b)) vary strongly in the IQR and the
median across the classes. While the IQR is high for calcaneus
and ankle, the classes hip, knee, and the normal controls
exhibit a very similar distribution. Thus, such a parameter has
solely limited potential to separate normal gait from general
gait disorders. There may be, however, a certain potential
to separate individual classes (in this case calcaneus) from
other classes. Other parameters may lack in discriminative
power. An example is IFAP (see Figure 2(c)), which shows
a similar median and overlapping distributions with a similar
IQR across all classes. Several parameters are discriminative
for particular classes or a group of classes. However, none
of the observed parameters discriminates well between all
classes. Therefore, the combination of several parameters
for the distinction between classes seems advisable. These
assumptions are further corroborated by the LDA results.
Fig. 2: Example of boxplots for three parameters. Each boxplot
shows the median and the IQR (box) for each class (outliers
were removed for better visualization). Box-whiskers correspond to 1.5 of the box-length, thus show approximately ± 2.7
standard deviations. The overlap of distributions between the
classes gives an impression of the parameters’ discriminative
power.
LDA was applied to the individual parameters and their
combination, as well as to the higher-dimensional PCAbased representations. This analysis aimed at quantifying the
discriminative power of the investigated representations and
thereby evaluating their suitability for automated classification.
Figure 3 illustrates discriminativity scores obtained by LDA
in terms of deviation from the random baseline (zero rule).
In detail, results for different combinations of classes (rows)
are illustrated: rows 1-4 provide results for the discrimination
of normal gait vs. ankle, calcaneus, hip or knee (each class
separately). Row 5 shows how well all 5 classes can be
differentiated from each other. Row 6 illustrates how well
normal gait can be differentiated from all types of gait
disorders. Rows 7-12 show how all possible pairs of gait
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, DECEMBER 2017
disorder classes can be differentiated from each other. Positive
discriminativity scores are represented by a color scale from
blue (corresponding to low values) to yellow (representing
high values), whereas negative values are colored in gray.
Positive values mean that the random baseline is exceeded
and that the respective input parameter or input representation
exhibits a certain discriminative power (the higher the value
the better). Negative values indicate the absence of discriminative power, i.e. the random baseline is not reached. It has to
be noted that, since the different class partitions represented
by the rows of Figure 3 have different random baselines, the
values across rows cannot be compared directly. Comparisons
are solely valid along the rows. In general, however, columns
including a larger number of high values indicate parameters or
representations with a higher discriminative power. Similarly,
rows with higher values represent tasks that are easier to solve
than others.
The leftmost part of Figure 3 illustrates the discriminativity scores for the individual parameters. Several parameters achieve high scores for individual classes or combinations of classes, e.g. FAP 3 , FV 3 , FV AV G , FV 1 , TV 3 ,
TAP 3 , GV , ST EP V , DS, ST RLEN , FV 2 , ST EP W D,
CAD, BF , and ST RIDET . No parameter, however, performs well across all tasks. This indicates that individual
parameters are quite limited in expressiveness. The second
part (ALL P ARAM S) of Figure 3 illustrates the results
from the combination of all parameters. The combination
yields much better discrimination across all rows of Figure 3.
This demonstrates that the individual parameters contain complementary information and attain synergies when they are
combined. The third part of Figure 3 visualizes the results
of the PCA-based representations of the five input signals
FV , FAP , FM L , COPAP , and COPM L . The three GRF
components achieved higher scores compared to the COP
signals. The rightmost part of Figure 3 shows the discriminativity scores for combined PCA representations, i.e all
three GRF components combined (P CA FALL ), both COP
signals combined (P CA COPAP,M L ), and all five components combined (P CA ALL). In general, the combination
of components improved the results, which indicates that the
individual GRF components are complementary to each other.
The addition of the COP further improved the discriminative
power. Thus, adding COP to a classification may contribute
positively to the results. The representations (P CA ALL and
ALL P ARAM S) are combined able to contribute to all
evaluated tasks (rows) of Figure 3.
The evaluated representations are more suitable for differentiating between the healthy control group and a functional
gait disorder (rows 1-4) than between two functional gait
disorders (rows 7-12). Regarding the task N/GD, solely a few
parameters are able to exceed the random baseline. This is due
to the fact that the combined set of all gait disorders contains
much more samples than the class of healthy controls (i.e.
279 vs. 161 samples). This yields a random baseline around
87.1% which is more difficult to exceed than random baselines
in other tasks.
6
B. Classification
The results of the classification experiments, which were
performed on data from the test set, are summarized in Table
III. The test set was not presented to the classifier during the
training phase and the selection of its parameters. Results are
provided for the two classification tasks (N/C/A/K/H and
N/GD) and for three different parameterizations. The results
of the additional experiments with other classifiers such as
the multi-layer perceptron (MLP) and the k-nearest neighbors
algorithm (k-NN) were all outperformed by the SVM results,
which confirms also the results of Janssen, Schöllhorn et al.
[32]. Therefore, and due to the limited space available, these
results will not be discussed in detail.
The first evaluated parameterization comprises of 52 GRF
parameters (DPs and TDPs) that are extracted from all five
GRF input signals. Due to the strong variation in the parameters’ value ranges, a suitable normalization of the data
is essential. We evaluated min-max normalization as well as
z-standardization. The use of z-standardization resulted in a
slightly higher deviation from the RB for both tasks (except for
the RBF SVM in task N/C/A/K/H) compared to min-max
normalization. Furthermore, the RBF SVM failed to exceed
the random baseline for both methods in the task N/GD.
The second parameterization was obtained by PCA of the
raw GRF waveforms. PCAs obtained solely from the three
force components clearly outperform the GRF parameters
(DPs and TDPs). By adding the COP measurements the results
were further improved for both tasks. Normalization of the
PCA-based representations is crucial as performance otherwise
drops significantly.
The third parameterization applied PCA on the zstandardized and min-max normalized DPs and TDPs. The
dimensionality reduction resulted in a 28-dimensional vector
which was also z-standardized prior to classification. In this
case, results for both normalizations (last two rows of Table
III) were improved for the task N/GD compared to the
representation with the original GRF parameters (first two
rows of Table III). However, this is not the case for task
N/C/A/K/H, where the deviation from the RB slightly
decreased.
In summary, the best performance (marked in bold in
Table III) was achieved by applying PCA to all five GRF
signals. The linear SVM achieved the highest deviation from
the RB (22.5%) for task N/C/A/K/H as well as for task
N/GD (3.7%). Alternative classifiers which were also evaluated yielded a lower deviation from the RB: MLP 21.0%
and k-NN 13.4% for task N/C/A/K/H and MLP 2.6%
and k-NN 2.2% for task N/GD. In terms of accuracy and
deviation from the RB, the linear SVM performed better in
all experiments. The RBF SVM has an advantage solely in
terms of runtime.
The main reason for the great difference in the performance between the two tasks is the strong class imbalance
in task N/GD, which makes this task particularly difficult
to solve. One way of dealing with unbalanced datasets in
SVMs is the use of different weights for different classes,
thereby emphasizing the importance of the under-represented
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, DECEMBER 2017
7
Fig. 3: Discriminativity scores obtained by LDA for different selections of classes (rows). The figure is divided into four
blocks. Each column represents a different input parameter or higher-dimensional input representation. Best viewed in color.
classes. Therefore, additional class-weighted experiments were
performed. Results with different cost functions showed that
no further performance increase can be achieved. The uniform
cost function seems to work best on the data.
C. Discussion and further aspects
We presented a study on the classification of different
functional gait disorders, stemming from a wide range of
possible impairments, into categories that represent the main
affected body region. The motivation for selecting these broad
categories is that identifying the region of impairment is
essential for clinical practice and may allow to pinpoint impairments already at an early stage. In addition, it could indicate
secondary impairments which may easily be overlooked by
the physician during clinical examination. The present study
represents a first performance baseline for the classification
of gait disorders. Results are particularly promising for task
N/GD. However, an absolute classification accuracy of 91%
still lies below an acceptable threshold for clinical practice.
For the classification of individual disorder categories, the
results indicate that further improvements are necessary. To
date, the proposed approach could, however, already serve as
a support for clinicians indicating the presence of (additional)
arthropathies or diseases. In order to reduce the classification
complexity, while still providing support for clinicians, similar
classes could be merged, i.e. the hip and knee classes into a
thigh class and the ankle and calcaneus classes into a shank
class. The results of this additional experiment showed a deviation from the RB of 26.8% (using a linear SVM, RB: 51%,
absolute accuracy: 77.8%). Compared to the distinction of
all five classes (N/C/A/K/H), this is a clear increase in
accuracy and deviation from the RB.
Different influencing factors, i.e. the imbalance of the class
priors, the variability in the number of sessions per person
and gender-specific aspects may introduce a bias into the
aforementioned analyses. To investigate the effect of these
factors on classification performance, we performed additional
experiments. To this end, we used the best configuration found
so far as a baseline, i.e. PCA on all five signals with a
linear SVM (4th parametrization in Table III) and applied it
to different balanced subsets of our dataset. The results are
presented in Table IV and are discussed in the following.
For the experiments in Section III-B we decided to use
all available sessions of patients recorded in the course of
their rehabilitation to account for different progression stages
of impairments. This, however, may introduce a bias in the
experiments as more trials exist for some patients than for
others. To evaluate to which extent the varying number of
recorded sessions per patient influences the overall result, we
balanced the dataset by selecting only one random session per
person. Interestingly, the deviation from the RB improved for
task N/C/A/K/H to 23.7% (+1.2%) and for task N/GD to
20.6% (+16.9%), as presented in the first row of Table IV.
These results show that intra-patient variability needs to be
taken into account and requires additional modeling in a
classification approach.
Another factor causing an imbalance in the data are the
different class cardinalities, i.e. different numbers of persons
per class. In order to investigate the influence of this imbalance
we performed an experiment for both tasks with a dataset
containing the same number of participants per class (but
keeping all sessions in the dataset). For task N/C/A/K/H
the balanced dataset is composed of data from 62 persons from
each class (overall 310 persons, 7616 trials). For task N/GD
the balanced dataset contained data from 160 healthy controls
and 160 persons with a deficit (40 from each GD class, overall
320 persons and 6096 trials). The deviation from the RB
improved for task N/C/A/K/H to 28.3% (+5.8%) and for
task N/GD to 5.3% (+1.6%), as shown in the second row
of Table IV. Although the results show that balancing the
number of patients among classes is beneficial, the results of
task N/GD reveal the still existing imbalance in the dataset
(due to the fact that healthy controls have only one session
and patients up to several sessions).
The next question deals with the effect of balancing the
number of patients and the number of sessions at the same
time. We performed experiments with a completely balanced
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, DECEMBER 2017
8
TABLE III: Classification results (%) of two tasks - N/C/A/K/H and N/GD - and three different parameterization
approaches. Note that the random baseline (RB) is stated next to the task name and that the values in the table represent
the deviation from the random baseline (RB) and the corresponding absolute accuracy in brackets.
Parameterization
Norm.
GRF
GRF
PCA
PCA
PCA
PCA
z-score
min-max
z-score
z-score
z-score
z-score
Parameters (DPs and TDPs)
Parameters (DPs and TDPs)
on FV , FAP , FM L
on FV , FAP , FM L , COPAP , COPM L
on z-standardized GRF parameters
on min-max normalized GRF parameters
Dim.
TABLE IV: Results (%) of analyses assessing the influence of
several factors on the results of the two tasks - N/C/A/K/H
and N/GD. The experiments are performed with a PCA on
all five signals in combination with a linear SVM. Note that
the values represent the deviation from the random baseline
(RB) and the corresponding absolute accuracy in brackets.
Partitions of the dataset
Session are balanced
Persons are balanced
Persons & sessions are balanced
Male population
N/C/A/K/H
23.7 (60.2)
28.3 (59.5)
39.2 (59.2)
20.9 (51.3)
N/GD
20.6 (84.1)
5.3 (84.7)
35.4 (85.4)
0.6 (91.4)
version of our dataset for each task, containing only one
session per person and equal numbers of persons per class.
For task N/C/A/K/H the balanced dataset is composed of
data from 62 persons from each class (overall 310 persons,
2480 trials). For task N/GD the balanced dataset contained
data from 160 healthy controls and 160 persons with a deficit
(40 from each GD class, overall 320 persons and 2560 trials).
The results of our experiments showed clear performance improvements of +16.7% in the deviation from the RB compared
to the baseline for task N/C/A/K/H and +31% compared to
the baseline for task N/GD (see the third row in Table IV).
Other biases in the data may be introduced by variations
in gender, walking velocity, leg length and other parameters
[33] leading to a variability of GRF parameterizations in the
individual disorder classes. Additional normalization of the
input data may be necessary to reduce intra-class variation to
improve classification accuracy. Several studies have shown
that in particular gender causes strong variability in gait
signals [34], [35]. To assess the influence of gender on our
results, an experiment was performed on a reduced dataset
containing only data from male participants (note that the
number of female participants in our dataset is not sufficient
to perform separate experiments). Surprisingly, the results did
not improve (see the last row in Table IV). This indicates that
for our data, gender has rather little influence on the results,
which, however, does not imply that the influence of gender
can be neglected a priori. A detailed study on the influence of
gender is subject to future investigation.
IV. C ONCLUSIONS
T
HE present study aimed at classifying patients with different orthopedic gait impairments at the hip, knee, ankle,
and calcaneus from healthy controls using GRF measurements.
For this purpose a dataset of 9,496 gait measurements from
52
52
30
39
28
28
N/C/A/K/H
linear SVM
15.0 (46.8)
14.3 (46.1)
19.8 (51.6)
22.5 (54.3)
13.8 (45.6)
13.5 (45.3)
(RB: 31.8%)
RBF SVM
8.8 (40.6)
9.5 (41.3)
15.4 (47.2)
19.4 (51.2)
8.8 (40.6)
7.9 (39.7)
N/GD (RB: 87.1%)
linear SVM
RBF SVM
2.4 (89.5)
-0.8 (86.3)
1.6 (88.7)
-3.8 (83.3)
2.4 (89.5)
2.0 (89.1)
3.7 (90.8)
1.9 (89.0)
2.6 (89.7)
-0.6 (86.5)
2.8 (89.9)
0.1 (87.2)
clinical practice was utilized. In a first step we investigated
the suitability of state-of-the-art GRF parameterizations and
analyzed their statistical properties and discriminative power
among the classes. Based on these results, the use of entire
GRF waveform parameterizations as input (such as PCA),
rather than relying on GRF parameters (DPs and TDPs) seems
advisable. Furthermore, the use of GRF force components
paired with the respective COP measurements yielded the
best results. Our experiments further showed that balancing
the dataset significantly improves results (e.g. increasing the
deviation from the random baseline by +16.7% for the classification into healthy controls and all four GD classes and
by +31% for distinguishing between healthy controls and
patients).
The presented study shows that results heavily depend on
the employed GRF representation. Future work will investigate
and evaluate adaptively learned signal representations [36],
[37] to obtain more discriminative and expressive parameterizations of GRF measurements. Furthermore, we will focus
on establishing a large, open-source, and balanced data set to
foster further developments in this area. Our results thereby
provide a first performance baseline for the classification of
functional gait disorders and can serve as a reference for future
improvements.
ACKNOWLEDGMENT
This project is funded by the NFB Lower Austrian Research
and Education Company and the Provincial Government of
Lower Austria, Department of Science and Research (LSC14005). We want to thank Marianne Worisch and Szava Zoltán
for their great assistance in data preparation and their great
support in clinical and technical questions.
R EFERENCES
[1] R. Baker. Measuring walking: a handbook of clinical gait analysis. Mac
Keith Press, London, 2013.
[2] T. Chau. A review of analytical techniques for gait data. Part 1: fuzzy,
statistical and fractal methods. Gait & Posture, 13(1):49–66, 2001.
[3] T. Chau. A review of analytical techniques for gait data. Part 2: neural
network and wavelet methods. Gait & Posture, 13(2):102–120, 2001.
[4] C.A. Lozano-Ortiz, A.M.S. Muniz, and J. Nadal. Human gait classification after lower limb fracture using artificial neural networks and
principal component analysis. In Engineering in Medicine and Biology
Society, 2010 Annual International Conference of the IEEE, pages 1413–
1416. IEEE, 2010.
[5] W. Zeng, F. Liu, Q. Wang, Y. Wang, L. Ma, and Y. Zhang. Parkinson’s
disease classification using gait analysis via deterministic learning.
Neuroscience Letters, 633:268–278, 2016.
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, DECEMBER 2017
[6] A. Vieira, H. Sobral, J.P. Ferreira, P. Ferreira, S. Cruz, M. Crisostomo,
and A.P. Coimbra. Software for human gait analysis and classification.
In Bioengineering (ENBENG), 2015 IEEE 4th Portuguese Meeting on,
page 1. IEEE, 2015.
[7] J. Wu, J. Wang, and L. Liu. Feature extraction via kpca for classification
of gait patterns. Human movement science, 26(3):393–411, 2007.
[8] J. Wu and J. Wang. Pca-based svm for automatic recognition of gait
patterns. Journal of applied biomechanics, 24(1):83–87, 2008.
[9] P. Levinger, D.T.H. Lai, R.K. Begg, K.E. Webster, and J.A. Feller.
The application of support vector machines for detecting recovery from
knee replacement surgery using spatio-temporal gait parameters. Gait
& posture, 29(1):91–96, 2009.
[10] N. Mezghani, S. Husse, K. Boivin, K. Turcot, R. Aissaoui, N. Hagemeister, and J.A. de Guise. Automatic classification of asymptomatic
and osteoarthritis knee gait patterns using kinematic data features
and the nearest neighbor classifier. IEEE transactions on biomedical
engineering, 55(3):1230–1232, 2008.
[11] M. Alaqtash, T. Sarkodie-Gyan, H. Yu, O. Fuentes, R. Brower, and
A. Abdelgawad. Automatic classification of pathological gait patterns
using ground reaction forces and machine learning algorithms. In
Engineering in Medicine and Biology Society, 2011 Annual International
Conference of the IEEE, pages 453–457. IEEE, 2011.
[12] M. Ferrarin, G. Bovi, M. Rabuffetti, P. Mazzoleni, A. Montesano,
E. Pagliano, A. Marchi, A. Magro, C. Marchesi, D. Pareyson, and
I. Moroni. Gait pattern classification in children with charcot–marie–
tooth disease type 1a. Gait & posture, 35(1):131–137, 2012.
[13] Y. Bengio, A. Courville, and P. Vincent. Representation learning: A
review and new perspectives. IEEE transactions on pattern analysis
and machine intelligence, 35(8):1798–1828, 2013.
[14] G. Giakas and V. Baltzopoulos. Time and frequency domain analysis
of ground reaction forces during walking: an investigation of variability
and symmetry. Gait & Posture, 5(3):189–197, 1997.
[15] R. Lafuente, J.M. Belda, J. Snchez-Lacuesta, C. Soler, and J. Prat.
Design and test of neural networks and statistical classifiers in computeraided movement analysis: a case study on gait analysis. Clinical
Biomechanics, 13(3):216–229, 1998.
[16] D.P. Soares, M.P. de Castro, E.A. Mendes, and L. Machado. Principal
component analysis in ground reaction forces and center of pressure
gait waveforms of people with transfemoral amputation. Prosthetics
and orthotics international, 40(6):729–738, 2016.
[17] A.M.S. Muniz and J Nadal. Application of principal component analysis
in vertical ground reaction force to discriminate normal and abnormal
gait. Gait & posture, 29(1):31–35, 2009.
[18] Z. Peng, C. Cao, Q. Liu, and W. Pan. Human Walking Pattern
Recognition Based on KPCA and SVM with Ground Reflex Pressure
Signal. Mathematical Problems in Engineering, 2013.
[19] Y. Xu, D. Zhang, Z. Jin, M. Li, and J.Y. Yang. A fast kernelbased nonlinear discriminant analysis for multi-class problems. Pattern
Recognition, 39(6):1026–1033, 2006.
[20] R. LeMoyne, W. Kerr, T. Mastroianni, and A. Hessel. Implementation
of machine learning for classifying hemiplegic gait disparity through
use of a force plate. In Machine Learning and Applications, 2014 13th
International Conference on, pages 379–382. IEEE, 2014.
[21] G. Williams, D. Lai, A. Schache, and M.E. Morris. Classification of
gait disorders following traumatic brain injury. The Journal of Head
Trauma Rehabilitation, 30(2):E13–E23, 2015.
[22] W.I. Schöllhorn, B.M. Nigg, D.J. Stefanyshyn, and W. Liu. Identification
of individual walking patterns using time discrete and time continuous
data sets. Gait & Posture, 15(2):180–186, 2002.
[23] K.L. Goh, K.H. Lim, A.A. Gopalai, and Y.Z. Chong. Multilayer
perceptron neural network classification for human vertical ground
reaction forces. In 2014 IEEE Conference on Biomedical Engineering
and Sciences, pages 536–540, 2014.
[24] M. Köhle and D. Merkl. Analyzing human gait patterns for malfunction
detection. In Proceedings of the 2000 ACM Symposium on Applied
Computing, volume 1 of SAC ’00, pages 41–45. ACM, 2000.
[25] M. Köhle and D. Merkl. Identification of gait patterns with selforganizing maps based on ground reaction force. In ESANN, volume 96,
pages 24–26, 1996.
[26] K.C. Moisio, D.R. Sumner, S. Shott, and D.E. Hurwitz. Normalization
of joint moments during gait: a comparison of two techniques. Journal
of Biomechanics, 36(4):599–603, 2003.
[27] I. Jolliffe. Principal Component Analysis. Springer-Verlag New York,
2 edition, 2002.
[28] D. Slijepcevic, B. Horsak, C. Schwab, A.M. Raberger, M. Schüller,
A. Baca, C. Breiteneder, and M. Zeppelzauer. Ground reaction force
9
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
measurements for gait classification tasks: Effects of different PCAbased representations. Gait & Posture, 57:4–5, 2017.
R.O. Duda, P.E. Hart, and D.G. Stork. Pattern classification. John Wiley
& Sons, 2012.
C.M. De Vries, S. Geva, and A. Trotman. Document clustering
evaluation: Divergence from a random baseline.
arXiv preprint
arXiv:1208.5654, 2012.
C.C. Chang and C.J. Lin. LIBSVM: A library for support vector
machines. ACM Transactions on Intelligent Systems and Technology,
2:27:1–27:27, 2011.
D. Janssen, W.I. Schöllhorn, K.M. Newell, J.M. Jäger, F. Rost, and
K. Vehof. Diagnosing fatigue in gait patterns by support vector machines
and self-organizing maps. Human movement science, 30(5):966–975,
2011.
M.R. Pierrynowski and V. Galea. Enhancing the ability of gait analyses
to differentiate between groups: scaling gait data to body size. Gait &
posture, 13(3):193–201, 2001.
B.M. Eskofier, M. Kraus, J.T. Worobets, D.J. Stefanyshyn, and B.M.
Nigg. Pattern classification of kinematic and kinetic running data to
distinguish gender, shod/barefoot and injury groups with feature ranking. Computer methods in biomechanics and biomedical engineering,
15(5):467–474, 2012.
M.C. Chiu and M.J. Wang. The effect of gait speed and gender on
perceived exertion, muscle activity, joint motion of lower extremity,
ground reaction force and heart rate during normal walking. Gait &
posture, 25(3):385–392, 2007.
Y. Zhang, P.O. Ogunbona, W. Li, B. Munro, and G.G. Wallace. Pathological gait detection of parkinson’s disease using sparse representation.
In Digital Image Computing: Techniques and Applications, 2013 International Conference on, pages 1–8. IEEE, 2013.
J. Hannink, T. Kautz, C.F. Pasluosta, K.G. Gaßmann, J. Klucken,
and B.M. Eskofier. Sensor-based gait parameter extraction with deep
convolutional neural networks. IEEE journal of biomedical and health
informatics, 21(1):85–93, 2017.
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, DECEMBER 2017
10
(a) FV
(b) FAP
(c) FM L
Fig. 1: Mean pattern of the three ground reaction forces (GRF) enveloped by ± 1 standard deviations for each class. Data
normalized by body weight and 100% stance.
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, DECEMBER 2017
11
(a) FV
(b) FAP
(c) FM L
Fig. 2: Comparison of different PCA representations. The final dimensionality of the obtained representations is specified by
the amount of variance preserved in a particular projection, i.e. 98%, 95%, and 90%. Data normalized by body weight and
100% stance.
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, DECEMBER 2017
12
1.3
0.9
1.2
35
0.9
0.8
1.1
30
0.8
0.7
1
25
0.7
0.6
0.9
20
0.6
FV1 [BW]
ST [s]
80
5
0.06
4
70
FV2 [BW]
0.02
0.9
TV2 [%ST]
0.3
0.2
20
-0.2
15
0.1
1
10
-0.3
0
FAP1 [BW]
FV3 [BW]
25
-0.1
2
TV3 [%ST]
1
3
0.04
60
1.1
50
40
TV1 [%ST]
0.08
1.2
60
TAP1 [%ST]
FAP2 [BW]
TAP2 [%ST]
FAP3 [BW]
0
90
80
12
0.08
10
0.06
8
-0.05
60
-0.1
TAP3 [%ST]
90
0.04
2
0
TML1 [%ST]
20
TML2 [%ST]
FML3 [BW]
4
70
0.04
2
60
0.02
0
0.7
0.02
0
FML2 [BW]
0.02
0.75
60
0.02
0.04
0.8
80
70
4
FML1 [BW]
0.06
0.04
6
70
0.08
40
0
50
-0.02
0
0.65
50
TML3 [%ST]
16
2
14
FMLAVG [BW]
35
4
3
40
25
-3
IF V1 [%BW"s]
2
-4
20
8
IF ML [%BW"s]
30
IF V2 [%BW"s]
1
IF V3 [%BW"s]
IF APDEC [%BW"s]
0
-0.2
10
5
2
-0.4
0.03
1
1.6
0.02
0.95
1.4
0.9
1.2
0
1
0.01
-5
-0.6
1
0.85
IF LAT [%BW"s]
IF MED [%BW"s]
COPANG [deg]
COP AP [FL]
COPDEV [FL]
12000
0.5
0.25
10000
0.5
0.2
0.4
0.4
0.15
0.1
0.3
0.3
0.05
0.2
COP ML [FL]
DECT [s]
-2000
10000
6000
4000
4000
-8000
2000
2000
-10000
LR0080 [N/s]
-8000
0.65
LR2080 [N/s]
5
1.4
4
1.2
3
1
UR8000 [N/s]
1
0.9
0.8
0.15
0.6
-12000
0.7
0.55
0.1
UR8020 [N/s]
-6000
0.7
0.2
-10000
-4000
8000
6000
ACCT [s]
-4000
COPV [FL/s]
-2000
8000
0.75
-6000
IF APACC [%BW"s]
1.8
15
3
IF AP [%BW"s]
-2
10
0
IF V [%BW"s]
-1
50
30
12
1
FAPAVG [BW]
5
18
3
40
-0.04
FVAVG [BW]
-2
DS [s]
STEPLEN [m]
STEPV [km/h]
STRIDET [s]
BF [Hz]
1.5
120
5
0.15
100
0.1
80
0.05
4
1.3
1.2
CAD [1/min]
normal
calcaneus
ankle
knee
hip
1.4
3
1.1
STEPWD [m]
STRLEN [m]
GV [km/h]
Fig. 3: Boxplots for all 52 GRF parameters. Each boxplot shows the median and the IQR (box) for each class (outliers
were removed for better visualization). Box-whiskers correspond to 1.5 of the box-length, thus show approximately ± 2.7
standard deviations. The overlap of distributions between the classes gives an impression of the parameters’ discriminative
power (inter-class variation). Data normalized by body weight and 100% stance, prior to the calculation of the parameters.
| 1cs.CV |
arXiv:1510.08565v3 [cs.NE] 5 Nov 2015
Attention with Intention for a Neural Network
Conversation Model
Geoffrey Zweig
Kaisheng Yao ∗
Microsoft Research
Microsoft Research
gzweig@microsoft.com
kaisheny@microsoft.com
Baolin Peng
Chinese University of Hong Kong
blpeng@se.cuhk.edu.hk
Abstract
In a conversation or a dialogue process, attention and intention play intrinsic roles.
This paper proposes a neural network based approach that models the attention
and intention processes. It essentially consists of three recurrent networks. The
encoder network is a word-level model representing source side sentences. The
intention network is a recurrent network that models the dynamics of the intention
process. The decoder network is a recurrent network that produces responses
to the input from the source side. It is a language model that is dependent on
the intention and has an attention mechanism to attend to particular source side
words, when predicting a symbol in the response. The model is trained end-toend without labeling data. Experiments show that this model generates natural
responses to user inputs.
1
Introduction
A conversation process is a process of communication of thoughts through words. It may be considered as a structural process that stresses the role of purpose and processing in discourse [7].
Essentially, the discourse structure is intimately connected with two nonlinguistic notions: intention
and attention. In processing an utterance, attention explicates the processing of utterances, for example, paying attention to particular words in a sentence. On the other hand, intention is higher level
than attention and has its primary role of explaining discourse structure and coherence. Clearly, a
conversation process is inherently complicated because of the two levels of structures.
A conversation process may be cast as a sequence-to-sequence mapping task. In this task, the
source side of the conversation is from one person and the target side of the conversation is from
another person. The sequence-to-sequence mapping task includes machine translation, graphemeto-phoneme conversion, named entity tagging, etc. However, an apparent difference of a dialogue
process from these tasks is that a dialogue process involves multiple turns, whereas usually the above
tasks involve only one turn of mapping a source sequence to its target sequence.
Neural network based approaches have been successfully applied in sequence-to-sequence mapping
tasks. They have made significant progresses in machine translation [1, 6, 13], language understanding [8], and speech recognition [4]. Among those neural network-based approaches, one particular
approach, which is called encoder-decoder framework [1, 13], aims at relaxing much requirement
on human labeling.
Conversation models have been typically designed to be domain specific with much knowledge such
as rules [3, 18]. Recent methods [15] relax such requirement to some extent but their whole systems
∗
Presented at NIPS Workshop on Machine Learning for Spoken Language Understanding and Interaction
2015.
1
Table 1: An example of dialogue process.
user
agent
user
agent
user
agent
user
agent
my computer is infected
do you want to retrieve the files that was deleted?
the ones that the virus deleted , yes.
i can help you resolve the issue with our virus removal and protection service
ok
here is a link how to run system restore
thank you .
you welcome
are still trained with manual labels because of their sub-components that require so. Manual labels
are error prone and expensive. Therefore, it is appealing to train a system end-to-end without manual
labels. Recent works in [10, 12, 14] are in this approach.
In general, however, using knowledge is helpful. For example, the alignment information between
the source and target side is critical in grapheme-to-phoneme conversion [17] to outperform a strong
baseline using n-gram models [2]. In a neural network based machine translation system [6], the
alignment information is used to outperform a strong phrase-based baseline [5].
In the context of modeling conversation process, a neural network model may be built with the
knowledge of the structural information of conversation processes. In particular, the network may
incorporate the notion of intention and attention. To test this, we developed a model that consists
of three recurrent neural networks (RNNs). The source side RNN, or encoder network, encodes the
source side inputs. The target side RNN, or decoder network, uses an attention mechanism to attend
to particular words in the source side, when predicting a symbol in its response to the source side.
Importantly, this attention in the target side is conditioned on the output from an intention RNN.
This model, which has the structural knowledge of the conversation process, is trained end-to-end
without labels. We experimented with this model and observed that it generates natural responses to
user inputs.
2
Background
In the theory of discourse in [7], discourse structure is composed of three separate but related components. The first is the linguistic structure, which is the structure of the sequence of utterance.
The linguistic structure consists of segments of the discourse into which the utterances naturally
aggregate. The second structure is the intentional structure, which captures the discourse-relevant
purposes, expressed in each of the linguistic segments as well as relationships among them. The
third is the attentional state that is dynamic, and records the objects, properties, and relations that
are salient at each point of the discourse.
In many examples we observe, there are usually just one linguistic segment that consists of all the
utterances. Therefore, in the following, we consider a discourse with two structures: intention and
attention.
In the example in Table 1, there is a clear flow of intentions. The user states the problem, with the
user’s intention of conveying the problem to the agent. The agent receives the words, processes
them, and communicates back to the user. The user responds to the agent afterwards. Therefore,
the whole conversation process consists of three intentions processed sequentially. The first is the
intention of communication of the problem. The second intention is the process of resolving the
issue. The third is the intention of acknowledgment. In processing each of the intentions, the user
and the agent pay attention to particular words. For example, when resolving the issue, the agent
pays attention to words such as ”virus”.
2
Figure 1: The attention with intention (AWI) model. The model is unrolled into three turns. In
each turn has RNNs for encoder network and decoder network. Each session is represented by a
fixed-dimension vector, which is a hidden state of an intention RNN network.
3
3.1
The model
The attention with intention (AWI) model
We propose a model that attempts to represent the structural process of intentions and the associated
attentions. Figure 1 illustrates the model. It shows three layers of processing: encoder network,
intention network, and decoder network.
The encoder network has inputs from the current source side input. Because the source side in the
current turn is also dependent on the previous turn, the source side encoder network is linked with
the output from the previous target side. The encoder network creates a representation of the source
side in the current turn.
The intention network is dependent on its past state, so that it memories the history of intentions.
It therefore is a recurrent network, taking a representation of the source side in the current turn and
updating its hidden state.
The decoder is a recurrent network for language modeling that outputs symbol at each time. This
output is dependent on the current intention from the intention network. It also pays attention to
particular words in the source side.
In more details, a conversation has in totoal U turns. At turn u, a user in the source side, denoted in
(s,u)
superscript (s) , has an input sequence of ~x(s,u) = (xt
: t = 1, · · · , T ) with length T . An agent in
(t,u)
the target side, denoted in superscript (t) , responds to the user with ~y (t,u) = (yj
: j = 1, · · · J) with
length J. The proposed model is a conditional model of the target given the source, p(~y (t,u) |~x(s,u) ).
If there is no confusion, we may omit the session index u in the following.
3.2
Encoder network
The encoder network reads the input sentence ~x(s) , and converts them into a fixed-length or a variable
length representation of the source side sequence. There are many choices to encode the source side.
The approach we use is an RNN such that
(s)
(s)
(s)
ht = f xt , ht-1
(1)
(s)
(s)
where f (·) is an RNN. ht is the hidden state at time t in the source side. The initial state ht with
(t,k-1)
t = 0 is the last hidden state activity, hT
, of the decoder network in the previous turn k − 1.
(s)
(s)
One form of the output from this encoder is the last hidden state activity cT = hT . This is used
as a representation of the source side in the current turn to the intention network. The other form is
a variable-length representation, to be used in the attention model described in Sec. 3.4. A general
description of the variable length representation is as follows
(s)
ct
(s)
= q({ht , ∀t = 0, · · · T })
3
(2)
where q(·) might be a linear network or a nonlinear network.
3.3
Intention network
The signal from the encoder network is fed into an intention network to model the intention process.
Following [7], the intention process is a dynamic process to model the intrinsic dynamics of conversation, in which an intention in one turn is dependent on the intention in the previous turn. This
property might be modeled using a Markov model, but we choose an RNN.
Interestingly, the hidden state of an RNN in a certain turn may be considered as a distributed representation of the intention. Different from the usual process of training distributed representation of
words [9], the distribution representation of intentions are trained with previous turns as their context. We use a first order RNN model, in which a hidden state is dependent explicitly on its previous
state.
The intention model in AWI is therefore an RNN as follows
(s)
(t,k-1)
h(i,k) = f cT , h(i,k-1) , hT
(3)
(s)
where cT is the fixed dimension representation of the source side described in Sec. 3.2. k is the
(t,k-1)
is the last hidden layer activity of the decoder network in the
index of the current turn. hT
previous turn k − 1.
3.4
Decoder network
The last step is to decode the sequence in the target side, which is framed as a language model
over each symbol, generated left to right. In this framework, the decoder computes conditional
probability as
(t) (t)
(t)
(t)
(t) (t)
p(yj |y1 , · · · , yj-1 , ~x(s) ) = g(yj-1 , hj , cj )
(4)
where the hidden state in the decoder is computed using an RNN
(t)
(t)
(t) (t)
hj = f yj-1 , hj-1 , cj
(5)
(t)
The initial state hj at t = 0 is the last hidden state activity from the intention network.
(t)
(t)
The cj is a vector to represent the context to generate yj . It is dependent on the source side as
(t)
(t)
(s)
cj
= z hj-1 , {ct : ∀t = {1, · · · , T }}
(6)
(s)
where z(·) summerizes the variable-length source side representations {ct } using weighted average. The weight is computed using a content-based alignment model [1] that produces high scores
(s)
(t)
if the target side hidden state in previous time hj-1 and ct are similar. More formally, the weight
(s)
αjt for the context ct
is computed using
exp ejt
m exp(ejm )
αjt = P
where
(t)
(7)
(s)
ejt = a(hj-1 , ct )
(8)
(s)
The alignment model enables an attention to particular words, represented as a vector ct in the
source side. Since the decoder network generates responses on condition of the attention and also
the intention, our model is called attention with intention (AWI) model.
3.5
Implementation details
All of the recurrent networks are implemented using a recently proposed depth-gated long-short(s)
term memory (LSTM) network [16]. The context vector ct is an embedding vector of the source
side word at time t.
4
Table 2: Perplexity results with AWI model. Models have 50 or 200 hidden layer dimension, 25
alignment dimension, and one layer of LSTMs.
# hidden dimension
50
200
PPL
30.8
22.1
The alignment model in Eq. (8) follows the attention model in [1], in which ejt is calculated as
(t)
(s)
ejt = ~v > tanh W(ah) hj-1 + W(ae) ct
,
(9)
which is a neural network with one hidden layer of size A and a single output, parameterised by
W(ae) ∈ RA×H , W(ah) ∈ RA×H and ~v ∈ RA . H and A are the hidden layer dimension and
alignment dimension.
4
Evaluation
We used an in-house dialogue dataset. The dataset consists of dialogues from a helpdesk chat service. In this service, costumers seeks helps on computer related issues from human agents. Training
consists of 10000 dialogues with 96913 turns or conversations. Number of tokens is 2215047 in the
source side and 2378950 in the target side. The vocabulary size is 9085 including words from both
side. Development set data has 1000 dialogues with 9971 turns. Test set data has 500 dialogues with
5232 turns.
We use sentence-level SGD without momentum. Learning rate is initialized to 0.1. Development set
is used to control the learning rate. The learning rate is halved when perplexity on the development
is increased. One epoch of training has one pass of the training data. The order of training dialogues
is randomly shuffled in the beginning of each epoch. The order of turns in the dialogue is however
kept.
4.1
Performances measure in perplexity
An objective comparison of different models for conversation is still an open question. We report
perplexity (PPL), though it may have drawbacks, to compare different models. Table 2 presents
results in perplexity with two models with different hiden layer sizes. Results show that a larger
model with 200 hidden layer dimension has lower PPL than the model with 50 dimension.
4.2
Examples of outputs from the trained model
Table 3 lists an example of the conversation process between a human and the trained model. The
model has two layers of LSTMs and other setups are the same as used in Sec 4.1. Similarly as
observed in [11], the model produces natural responses to user inputs. The flow of intentions is
clearly seen in this example.
5
Related work
Our work is related to the recent works in [10, 12, 14], which use an encoder-decoder framework
to model conversation. The work in [10] is a model for single turn conversation. The work in [14]
is a simple encoder-decoder method using a fixed-dimension representation of the source side. The
work in [12] also uses a fixed-dimension representaiton of the source side but has an additional RNN
to model dialogue context. This additional RNN is similar to the intention RNN in AWI model.
However, AWI model differs from [12] in that it incorprates the concept of attention and intention
based on the theory in [7]. Therefore, attention mechanism is essential to AWI. The model in [12]
doesn’t have an attention model.
5
Table 3: The generated response from the trained model.
user
agent
user
agent
user
agent
user
agent
user
agent
user
agent
user
agent
user
agent
agent
user
agent
user
agent
user
agent
user
agent
user
agent
user
agent
user
agent
user
agent
agent
agent
user
agent
user
agent
user
agent
hi
how may i help you today ?
my computer is not responding
oh ! i sorry to hear that . are you getting an error message ?
yes
what recent changes did you made with this computer ?
i updated my computer
i see . is this on windows 7 ?
no it is on linux
ok . are you using the affected computer right now ?
no
i can definitely assist you on your issue .
thanks
we can go ahead and start working on this issue within our office premium software support service .
that is allright
if in case the issue comes back at all within the next 30 days we will perform the service again for free .
let me go ahead and prepare the offer and let start working on this issue now , alright ?
please
sure . may i have your phone number please .
xxx-xxx-xxxx
thank you .
no problem
thank you for waiting .
no problem
i see .
what is the next step
i see . let me go ahead and prepare the offers for you .
okay
thank you .
no problem
thank you for contacting answer desk . have a great day !
bye
thanks again for working with me . id like to send you a quick survey .
it starts with a yes/no question , with a few more questions about your experience with me today.
the survey takes about a minute and really helps us improve our service . can i send it to you now ?
sure
thank you .
no problem
thank you for contacting answer desk . have a great day !
bye
goodbye .
Because it is not yet clear what objective measure to use to compare different models, it is hard to
make claims of superiority of these models. We believe AWI model is an alternative to the models
in [12, 14].
6
Conclusions and discussions
We have presented a model that incorporates attention and intention processes in a neural network
model. Preliminary experiments show that this model generates natural responses to user inputs. Future works include experiments on common dataset to compare different models and incorporating
objective functions such as goals.
6
References
[1] D. Bahdanau, K. Cho, and Y. Bengio. Neural machine translation by jointly learning to align
and translate. In Proceedings of the International Conference on Learning Representations
(ICLR), San Diego, CA, 2015.
[2] M. Bisani and H. Ney. Joint-sequence models for grapheme-to-phoneme conversion. Speech
Communication, 50(5), 2008.
[3] D. Bohus and A. I. Rudnicky. The ravenclaw dialog management framework: architecture and
systems. Computer, Speech and Language, 23:332–361, 2009.
[4] W. Chan, N. Jaitly, Q. V. Le, and O. Vinyals. Listen, attend and spell. In arXiv:1508.01211
[cs.CL], 2015.
[5] D. Chiang. Hierarchical phrase-based translation. Computational Linguistics, 33(2):201–228,
1999.
[6] J. Devlin, R. Zbib, Z. Huang, T. Lamar, R. Schwartz, and J. Makhoul. Fast and robust neural
network joint models for statistical machine translation. In ACL, 2014.
[7] B. J. Grosz and C. L. Sidner. Attention, intentions, and the structure of discourse. Computational Linguistics, 12:175–204, 1986.
[8] G. Mesnil, Y. Dauphin, K. Yao, Y. Bengio, L. Deng, D. Hakkani-Tur, X. He, L. Heck, G. Tur,
D. Yu, and G. Zweig. Using recurrent neural networks for slot filling in spoken language
understanding. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2015.
[9] T. Mikolov, K. Chen, G. Corrado, and J. Dean. Effcient estimation of word representations in
vector space. In NIPS, 2013.
[10] L. Shang, Z. Lu, and H. Li. Neural responding machine for short-text conversation. In ACL,
2015.
[11] A. Sordoni, Y. Bengio, H. Vahabi, C. Lioma, J. G. Simonsen, and J.-Y. Nie. A hierarchical
recurrent encoder-decoder for generative context-aware query suggestion. In arXiv:1507–0222
[cd.NE], July 2015.
[12] A. Sordoni, M. Galley, M. Auli, C. Brockett, Y. Ji, M. Mitchell, J.-Y. Nie, J. Gao, and B. Dolan.
A neural network approach to context-sensitive generation of conversation responses. In
NAACL, 2015.
[13] I. Sutskever, O. Vinyals, and Q. V. Le. Sequence to sequence learning with neural networks.
In Neural Information Processing Systems (NIPS), pages 3104–3112, Montréal, 2014.
[14] O. Vinyals and Q. V. Le. A nerual converstion model. In ICML Deep Learning Workshop,
2015.
[15] T.-H. Wen, M. Gasic, D. Kim, N. Mrksic, P.-H. Su, D. Vandyke, and S. Young. Stochastic
language generation in dialogue using recurrent neural networks with convolutional sentence
reranking. Technical report, May 2015.
[16] K. Yao, T. Cohn, E. Vylomova, K. Duh, and C. Dyer. Depth-gated LSTM. In arXiv:1508.03790
[cs.NE], 2015.
[17] K. Yao and G. Zweig. Sequence-to-sequence neural net models for grapheme-to-phoneme
conversion. In INTERSPEECH, 2015.
[18] S. Young, M. Gasic, B. Thomson, and J. D. Williams. POMDP-based statistical spoken dialog
systems: A review. Proceedings of the IEEE, 101:1160–1179, 2013.
7
| 9cs.NE |
Bayes estimator for multinomial parameters and Bhattacharyya distances
Christopher Ferrie
Centre for Engineered Quantum Systems, School of Physics, The University of Sydney, Sydney, NSW, Australia
Robin Blume-Kohout
arXiv:1612.07946v1 [math.ST] 23 Dec 2016
Sandia National Laboratories, Albuquerque, New Mexico, 87185
(Dated: December 26, 2016)
We derive the Bayes estimator for the parameters of a multinomial distribution under two loss
√
functions (1 − B and 1 − B2 ) that are based on the Bhattacharyya coefficient B( p, q) = ∑ pk qk .
We formulate a non-commutative generalization relevant to quantum probability theory as an open
problem. As an example application, we use our solution to find minimax estimators for a binomial
parameter under Bhattacharyya loss (1 − B2 ).
I.
INTRODUCTION
In statistical decision theory, a Bayes estimator for a prior µ is a point estimator whose expected risk (with respect
to µ) is as small as possible [1]. Bayes estimators have utility beyond strictly Bayesian statistics; they also provide
lower bounds on minimax risk (a frequentist concept), and the greatest lower bound (That is, the maximum Bayes
risk over all priors) coincides with the minimax risk. This duality enables efficient numerical algorithms to find
minimax estimators by optimizing over Bayesian priors, rather than over the computationally intractable space of
estimators [2].
Different loss functions produce different risk functions, and the Bayes estimator is sensitive to this choice. For a
certain class of risks known as Bregman divergences the Bayes estimator is always the mean of the posterior distribution [3]. For the multinomial model that we consider, both mean-squared-error and Kullback-Leibler divergence
(relative entropy) are Bregman divergences. But not every useful risk function has this property. We consider two
loss functions defined through the Bhattacharyya coefficient B( p, q), a Fisher-adjusted measure of distinguishability
between probability distributions that finds use in machine learning (as a metric on probability distributions [4]) and
in quantum information theory (as fidelity [5, 6], the most commonly used measure of quantum state distinguishability in both theoretical and experimental studies).
In this paper, we find the Bayes estimators for 1 − B and 1 − B2 by mapping their respective optimizations to
simple linear programs. We also state, but do not solve, the non-commutative generalization of the problem, which
has applications in quantum information theory. Finally, we demonstrate the utility of the Bayes estimator by finding
the minimax estimator for a binomial parameter.
II.
PROBLEM SET-UP, REVIEW AND CHOICE OF LOSS FUNCTIONS
Suppose we have a coin, a K-sided die, or any process that generates i.i.d. samples from a finite set of size K. We
wish to estimate the probabilities { p0 , . . . pK −1 }. They form a parameter vector p which belongs to the K-simplex
(
)
4K =
( p0 , p1 . . . , pK−1 ) pk ≥ 0 ∀k, ∑ pk = 1 .
(1)
k
Estimation requires some data. From the data, we wish to produce an estimate of p, call it p̂. Although the data
are most commonly obtained by repeated sampling directly from p, other experiments are possible. We might have
“noisy” data (drawn from an ancillary distribution q = Σp, where the stochastic matrix Σ is known [7]), or side
information. Happily, the Bayes analysis turns out to be independent of how the data were obtained. Instead, it
depends only on the posterior probability density Pr( p|data)—which is to say that, given the posterior, the expected
risk is independent of the likelihood Pr(data| p).
The quality of an estimate p̂ is formalized by a loss function L( p, p̂) that quantifies how bad it is to report an
estimate p̂ when the truth is p. The best estimate should be p itself, so L( p, p) ≤ L( p, p̂) for all p̂. A “good”
estimator is one that minimizes expected loss, or risk:
r p = Edata [ L( p, p̂(data)].
(2)
2
The risk is a function of the underlying p, and it is generally impossible to minimize risk for all p simultaneously.
The Bayesian solution is to consider (and minimize) the average risk with respect to a given prior µ( p). This is called
the Bayes risk of the estimator p̂ with respect to µ:
r ( p̂) = E p,data [ L( p, p̂(data)].
(3)
An estimator p̂B that minimizes the Bayes risk is called a Bayes estimator for µ( p). The Bayes estimator can also be
defined (under normal circumstances; [8]) explicitly for each dataset as the minimizer of posterior risk. This identity
is extremely useful, because it means that a Bayes estimator can be constructed explicitly for any given value of data
without considering others.
Different loss functions yield different Bayes estimators. For example, for quadratic loss functions of the form
L( p, p̂) = ( p − p̂)T A( p − p̂),
(4)
where A > 0 is a positive definite matrix, the Bayes estimator is the mean of the posterior distribution [8]
p̂B (data) = E p|data [ p].
(5)
Another important loss function is the relative entropy or Kullback-Leibler divergence:
L( p, p̂) =
p
∑ pk log p̂kk .
(6)
k
Here, again, the Bayes estimator is the posterior mean [9]. This is not an accident; it holds true whenever the loss is
a Bregman divergence [3]. Quadratic loss and Kullback-Leibler loss are both Bregman divergences.
Unfortunately, not every useful and important loss function is a Bregman divergence. The loss functions we
consider here are not Bregman divergences, and their Bayes estimators are not the posterior mean. They are based
on the Bhattacharyya coefficient [10],
p
B( p, p̂) = ∑ pk p̂k ,
(7)
k
which quantifies the statistical indistinguishability or fidelity of two distributions. In different contexts, both L1 =
1 − B and L2 = 1 − B2 are useful loss functions, and so we find Bayes estimators for both of them. In the limit of
small deviations ( p̂ is close to p), they coincide up to a factor of 2. Both are Fisher-adjusted, meaning that their 2nd
order series expansions are proportional to the Fisher metric. However, L1 and L2 have different global behavior,
and therefore different Bayes estimators.
Since the minimizer of 1 − B p is the maximizer of B p , we define two Bayes estimators as follows:
p̂1 = argmax E p [ B( p, p̂)],
(8)
p̂2 = argmax E p [ B2 ( p, p̂)],
(9)
p̂
p̂
subject to p̂ ∈ 4K . Note that, since the Bayes estimator depends only on the posterior, with no explicit dependence
on the prior or the data, we hereafter drop the conditional notation on the data. That is, E p means expectation with
respect to an arbitrary distribution of p.
III.
BAYES ESTIMATORS
Let us now derive expressions for the Bayes estimators defined in Equations 8 and 9. A very useful way of writing
√
B is to define p as the element-wise square root of p, then write:
B( p, p̂) =
√
p·
p
p̂.
The posterior risk involves an average of B over p. Both this average and the dot product in B are linear in
means that they commute, so:
p
√
E p [ B( p, p̂)] = E p [ p] · p̂.
(10)
√
p, which
(11)
3
This also implies the normalization
1=
√ √
√
p k p k = k p k2 .
∑ pk = ∑
k
(12)
k
Thus, the problem in Equation 8 becomes
p
√
maximize E p [ p] · p̂
p p
subject to
p̂ · p̂ = 1,
p
p̂ ≥ 0.
This is a textbook [11] convex optimization problem, and the solution is
√
p
E p [ p]
,
p̂1 =
√
kE p [ p]k
(13)
(14)
implying
√
E p [ p ]2
p̂1 =
√
√ ,
(E p [ p]) · (E p [ p])
(15)
where the square in the numerator is element-wise.
Minimizing 1 − B2 , as in Equation 9, requires a bit more work, but we can again reduce it to a linear program. We
start by defining
P=
the projection operator onto
√
√ √ T
p p ,
(16)
p, and similarly for P̂. The normalization condition on P is
Tr[ P] = 1.
(17)
In terms of these objects, the squared Bhattacharyya coefficient is
B2 ( p, p̂) = Tr[ P P̂].
(18)
This expression is now linear in P, which allows us to apply the trick of commuting the expectation through:
E p [ B2 ( p, p̂)] = Tr[E p [ P] P̂].
The matrix elements of the expectation on the RHS can be written out explicitly as
p
E p [ P] ij = E p [ pi p j ].
(19)
(20)
Putting everything together, Equation 9 becomes
maximize
Tr [E p [ P] P̂]
subject to
Tr[ P̂] = 1,
(21)
P̂ ≥ 0,
where P̂ ≥ 0 means positive semi-definiteness as a matrix. This is another textbook [11] linear semidefinite program.
If we define a as the normalized eigenvector of E p [ P] with maximal eigenvalue, then the solution to Equation 21 is
P̂2 = aaT .
(22)
Explicitly, as a vector in 4K , we have
p̂2 = a2 ,
(23)
√
where a = ( a0 , a1 , . . . , aK −1 ) is the eigenvector of E p [ pi p j ] with maximal eigenvalue, and the square is elementwise as in Eq. 15.
4
IV.
NON-COMMUTATIVE GENERALIZATION: A QUANTUM (OPEN) PROBLEM
Quantum mechanics can be described as a non-commutative generalization of probability theory [12]. Quantum
random variables (That is, the “state” of a quantum system) are represented not by K-element vectors of probabilities
p, but by K × K self-adjoint complex matrices or density matrices. In this framework, observable events are also
represented by K × K self-adjoint matrices called effects. States can be sampled or observed in more than one way.
The quantum analogue to “sampling” is “measuring” the system, and the observer must choose how to measure
it. This means specifying a mutually exclusive and exhaustive set of possible events, each represented by a positive
semidefinite d × d matrix. The set of event matrices { Ek } is called a positive operator valued measure (POVM), and
satisfies the condition ∑k Ek = 1l. If an quantum system is state ρ is observed, then the probability of observing
the event represented by effect E is given by Pr( E|ρ) = Tr(ρE). As a result, the law of total probability (“exactly
one of the possible events occurs”) manifests itself as the constraint Trρ = 1, and the non-negativity of probabilities
corresponds to ρ ≥ 0. Thus, the set of quantum states (analogous to the simplex ∆ of valid probability vectors) is
S = {ρ|ρ ≥ 0, Tr(ρ) = 1} .
(24)
The subset of S containing only diagonal matrices (“classical” states) is in 1:1 correspondence with K-element probability vectors, since ρ ≥ 0 ensures positivity of the diagonal entries, and Tr(ρ) = 1 ensures their normalization.
Thus, if we consider only states that are simultaneously diagonalizable, we can identify p ≡ ρ and p̂ ≡ ρ̂, and the
Bhattacharyya coefficient has a simple expression in terms of ρ and ρ̂:
p
ρρ̂ .
(25)
B( p, p̂) = Tr
Its square—the B2 risk that we considered previously—is called classical fidelity in the quantum mechanics literature.
Hence, our solution to Equation 21 can be applied directly to estimation of quantum states, whenever the support of
the averaging measure (e.g. a posterior distribution) consists entirely of commuting quantum states:
p
maximize Eρ [Tr( ρρ̂)2 ]
(26)
subject to Tr(ρ̂) = 1,
ρ̂ ≥ 0.
However, quantum states are not usually restricted to be diagonal, which means that in the estimation of quantum
states (quantum tomography [13]), the posterior will have support on non-commuting density matrices. This requires
new definitions of loss and risk. Without getting into the details (see Ref. [6]), the general definition of quantum
fidelity between two states is:
q
√ √ 2
F (ρ, ρ̂) = Tr
ρρ̂ ρ ,
(27)
and the corresponding loss function is infidelity or L = 1 − F. For co-diagonal states ρ̂ and ρ, fidelity coincides with
B2 . Thus, if the posterior does happen to be supported only on mutually commuting states, then our solution to
Equation 21 gives the Bayes estimator. In general, however, the problem of finding a Bayes estimator for quantum
infidelity is:
maximize Eρ [ F (ρ, ρ̂)]
subject to
Tr(ρ̂) = 1,
ρ̂ ≥ 0.
(28)
This problem appears to be quite difficult (some bounds are given in [14]), and we do not attempt to solve it in full
generality here.
V.
EXAMPLE: BINOMIAL PARAMETER
A.
Canonical beta prior
As an interesting and useful application of our main result, we now compute (and examine) the Bayes estimator
p̂2 for L2 = 1 − B2 in the simple case of a binomial distribution—that is, a directly observed coin toss or Bernoulli
5
process. Here, p = ( p0 , p1 ), with p1 = 1 − p0 . The canonical (conjugate) prior is a beta distribution:
β −1 β −1
Pr( p) =
p0 p1
dp,
Beta( β, β)
(29)
where Beta() is the beta function. If the coin is flipped N times, yielding n heads and N − n tails, then the posterior
is
n + β −1 N − n + β −1
Pr( p|n) =
p0
p1
dp.
Beta(n + β, N − n + β)
(30)
√
The matrix E p [ p j pk ] is given explicitly in terms of beta functions by
Beta( β + n + 1, β + N − n)
Beta( β + n + 1/2, β + N − n + 1/2)
Beta( β + n + 1/2, β + N − n + 1/2)
Beta( β + n, β + N − n + 1)
/Beta( β + n, β + N − n).
(31)
To compute the Bayes estimator p̂2 , we diagonalize this matrix, extract the eigenvector corresponding to the larger
eigenvalue, and square its entries. This can be written down in closed form, but the result is lengthy, messy, and
opaque. Instead, we illustrate its properties for N = 10 flips and for conjugate priors with β = 1 (Laplace’s prior)
and β = 1/2 (Jeffrey’s prior), in Fig. 1. For comparison, we also show (1) the posterior mean estimator, and (2) the
maximum likelihood estimator.
Jeffreys Prior (β = 1/2)
0.5
Laplace Prior (β = 1)
0.5
0.3
0.3
p̂
0.4
p̂
0.4
0.2
0.2
MLE
Mean
Bayes
0.1
0.0
0
1
2
n
3
4
MLE
Mean
Bayes
0.1
5
0.0
0
1
2
n
3
4
5
FIG. 1: The maximum likelihood estimator (MLE), optimal Bayes estimator and the posterior mean for N = 10 coin flips. On
the left, the latter two estimators begin with a canonical beta prior with β = 1/2 (Jeffrey’s prior). On the right, the same but with
β = 1 (Laplace’s uniform prior). The fidelity performance of these estimators is shown in Figure 3.
The Bayes estimator is generally similar to the posterior mean estimator. Both “hedge” away from p = 0 and
p = 1, in contrast to the MLE, but the Bayes estimator is more aggressive (less hedged) than the mean. Since the
mean estimator is Bayes for Bregman divergences, this implies that the 1 − B2 loss function is more forgiving of
extreme estimates (That is, ones close to p = 0 or p = 1) than any Bregman divergence.
However, although the difference between the two estimators (Bayes and posterior mean) appears significant, the
difference in their performance is not. While the Bayes estimator does have lower Bayes risk, the difference is small
even for N = 10. Figure 2 shows the relative suboptimality of the mean estimator, defined as:
R=
rBayes − rmean
,
rBayes
(32)
where rmean = E p|n [1 − B2 ( p, p̂mean (n))] and likewise for rBayes . The relative suboptimality of the mean estimator—
even at N = 10—is always less than 10−3 , decreases rapidly as a function of N and is largely independent of β.
Given that the mean estimator is far easier to calculate (the Bayes estimator requires computing the matrix
√
E p [ p j pk ] and its eigenvectors), we suggest that in practice for most purposes, the mean estimator is a completely
satisfactory heuristic. For small N, there is a noticeable difference in the estimate itself (see Fig. 3), but not in performance (see Fig. 2).
6
0.025
Relative Loss
0.020
0.015
0.010
0.005
1.1
1.0
0.9
0.8
β
0.7
0.6
0.5 100
90
80
70
60
50
40
30
20
0.000
10
N
FIG. 2: The relative difference, R in Eq. (32), between the average loss of the Bayes estimator rBayes and that of the mean estimator
rmean .
B.
Least favorable prior
Frequentist analyses of estimators average the risk only over the data (That is with respect to Pr(data| p)), rather
than over the joint distribution Pr(data, p). This more restricted average defines pointwise risk,
R( p, p̂) = Edata| p [ L( p, p̂)],
(33)
which retains a dependence on p that has to be gotten rid of somehow in order to define an “optimal” estimator.
Instead of averaging over a prior (to get the Bayes risk in Eq. (3)), the most common frequentist way to remove this
dependence is to consider the “worst case”, and maximize the risk over p:
Rmax ( p̂) = max R( p, p̂).
p
(34)
An estimator that minimizes the maximum risk is called a minimax estimator:
p̂minimax = arg min Rmax ( p̂).
(35)
p̂
The minimax criterion seeks an estimator whose performance is “pretty good” for any value of the parameters, without reference to prior probability. In practice, minimax estimators achieve (approximately) equal risk for all p. The
Bayes estimators for the binomial parameter that we derived in the previous section are not minimax; Figure 3 shows
that their pointwise risk varies with p0 . However, they’re pretty close—the variations are small (e.g., compared with
those of the MLE).
However, while the minimax estimator is unique, each prior has a different Bayes estimator. We can rewrite Eq. (3),
making this explicit, letting µ denote the prior density:
r (µ, p̂) = E p [ R( p, p̂)]
(36)
Somewhat remarkably, there is (at least under weak regularity conditions [1]) a prior whose Bayes estimator is
minimax. The least favorable prior is the prior whose Bayes estimator has the highest Bayes risk (out of all priors), and
its Bayes estimator is minimax. Explicitly, the estimator is
p̂LFP = arg max min r (µ, p̂)
µ
p̂
(37)
This remarkable relation between Bayesian and frequentist optimality is called Bayes-minimax duality:
p̂LFP = p̂minimax .
(38)
7
Jeffreys Prior (β = 1/2)
MLE
Mean
Bayes
Average 1 − B 2
0.04
0.03
0.02
0.01
0.00
0.0
0.1
0.2
p
0.3
0.4
0.09
0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0.00
0.0
Laplace Prior (β = 1)
MLE
Mean
Bayes
Average 1 − B 2
0.05
0.5
0.1
0.2
p
0.3
0.4
0.5
FIG. 3: The pointwise risk (1 − B2 ) of the estimators shown in Figure 1, as a function of p0 .
One useful consequence of Bayes-minimax duality is that it provides a method for finding minimax estimators by
searching for the least favorable prior. While the space of priors is large, it is enormously smaller than the space of
estimators. However, this approach is only practical if we have a closed-form expression for Bayes estimators, since
this allows fast and efficient computation of the Bayes risk for each prior considered. Our result enables this sort of
analysis for 1 − B and 1 − B2 loss functions.
We used our result to construct minimax estimators for the binomial parameter. Unlike the Bayes estimator,
the minimax estimators do not have explicit closed forms; the least favorable priors are not elegant, and require
numerical optimization. We performed two numerical optimizations. First, we did a restricted optimization over
conjugate priors (beta distributions, of the form given in Eq. 29), to find the minimax value of β. Then, we performed
an unrestricted optimization over all priors to find the true minimax estimator
In the first (restricted) optimization over conjugate priors, we found that the optimal β varied only weakly with
N, and was given by β ≈ 0.44 for all N. This is extremely consistent with other answers to the general question
“What value of β works best?”; a variety of other analyses have yielded answers close to the β = 1/2 that defines
Jeffrey’s estimator. Just for reference, we considered what prior would yield the lowest maximum risk if we were to
use the posterior mean estimator rather than the Bayes estimator derived in this paper. For this ad-hoc procedure,
the optimal β is approximately β ≈ 0.26. This estimator is not minimax in any sense, but it does support our general
observation that the Bayes estimator can be approximated by the posterior mean without much damage (0.26 and
0.44 are not very different, and the achieved maximum risks are also not very different).
To find the true minimax estimator, we used the algorithm of Kempthorne [2] to find the least favorable prior.
Roughly speaking, this algorithm iteratively adds support points to a discrete prior, each time maximizing the Bayes
risk over the small finite search space of support points and weights. It proceeds until the average risk and the
maximum risk are within a pre-defined tolerance. We provide pseudo-code for our implementation of the algorithm
in Algorithm 1. The results for the binomial parameter for N = 10 are plotted in Figure 4.
VI.
CONCLUSION
Optimality of estimators is an always-relevant topic in statistics. Even when optimal estimators are intractable
or impractical, they provide a useful benchmark, and make it possible to show rigorously that some more tractable
estimator is “good enough”. Our main technical contributions in this paper are simple, constructive formulas to
compute Bayes estimators for “fidelity”-type loss functions based on the Bhattacharyya coefficient. This result may
be directly useful for Bayesian machine learning, quantum state estimation, and other inference problems where it
(1) provides a more accurate estimator, and (2) establishes a provable bound on performance.
We find the examples provided in the second half of our paper interesting because they suggest certain qualitative
properties of multinomial estimation. Two of the most widely used estimators for this problem are the MLE and the
posterior mean (which is Bayes-optimal for Bregman divergences). Our results indicate that the Bayes estimator for
Bhattacharya loss interpolates between the MLE and the posterior mean; it hedges away from p = 0 like the mean,
but not as much. Most interestingly, our analysis shows that all the estimators we considered have nearly identical
average risk—and therefore suggests that worrying about finding exact Bayes estimators may be unnecessary.
Superficially, this may appear to undercut our result. Who cares about deriving the Bayes estimator if it isn’t significantly better? But this is exactly the point: our result proves that the posterior mean is “good enough”, at least in this
8
Algorithm 1 Kempthorne’s numerical algorithm for finding the least favorable prior [2].
Input: Number of measurements N > 0.
Input: Support points of initial guess p.
Input: Probability weights of the support points w.
Input: Tolerance tol > 0 (default 10−3 ).
Input: Mixing parameter α (default 0.01).
Output: Least favorable prior { p, w}.
Output: Lower bound on the minimax risk avg risk.
Output: Upper bound on the minimax risk max risk.
function K EMPTHORNE(N, { p, w},tol)
diff ← tol + 1 # make sure we enter the loop
while diff > tol do
{ p, w} ← prior with same number of support points which maximizes the Bayes risk
avg risk ← the maximum value of the Bayes risk for the prior found above
max risk ← global maximum of risk using the Bayes estimator of { p, w}
diff ← |avg risk − max risk|/avg risk
if diff > tol then
Add a new support where the maximum risk is attained
wlength( p) ← α
for each i ≤ length( p) − 1, wi ← wi − α/(length( p) − 1)
end if
end while
return { p, w}, avg risk, max risk
end function
0.5
Average 1 − B 2
0.4
p̂
0.3
0.2
Bayes
Minimax
Mean
0.1
0.0
0
1
2
n
3
4
0.026
0.024
0.022
0.020
0.018
0.016
0.014
0.012
0.010
0.008
0.0
5
Bayes
Minimax
Mean
0.1
0.2
p
0.3
0.4
0.5
FIG. 4: In this figure, we compare three estimators: (1) the minimax estimator, which is the Bayes estimator for the least favorable
prior; (2) the Bayes estimator for the least favorable conjugate prior, given by β ≈ 0.44; and (3) the posterior mean estimator for
the least favorable [with respect to the posterior mean estimator] conjugate prior, given by β ≈ 0.26. On the left, we show the
estimators themselves. On the right, we show their pointwise fidelity (1-risk) as a function of p0 . N = 10 in all cases.
particular case. Furthermore, it shows that although the Bayes and mean estimators are visibly different, their performance is not. This suggests that—at least for Bhattacharya loss functions—a rather wide range of estimators achieve
near-optimal performance. This conclusion is reinforced by the behavior of the minimax (frequentist-optimal) estimators that we construct using our result, where we found that small changes in the precise definition of “minimax”
produced fairly large changes in the “optimal” estimator. And, while we did not fully solve the quantum problem
(by finding Bayes estimators for quantum fidelity), we hope that our partial solution (for commuting states) provides
a stepping stone to a full solution in the future.
Acknowledgments
CF acknowledges funding from the IARPA MQCO program, the ARC via EQuS project number CE11001013, and
by the US Army Research Office grant numbers W911NF-14-1-0098 and W911NF-14-1-0103. Sandia National Laboratories is a multi-mission laboratory managed and operated by Sandia Corporation, a wholly owned subsidiary of
9
Lockheed Martin Corporation, for the U.S. Department of Energy’s National Nuclear Security Administration under
contract DE-AC04-94AL85000.
[1] J. Berger, Statistical Decision Theory and Bayesian Analysis, Springer (1985).
[2] P. J. Kempthorne, Numerical specification of discrete least favorable prior distributions, SIAM Journal on Scientific and Statistical
Computing 8, 171 (1987).
[3] A. Banerjee, Xin Guo and Hui Wang, On the optimality of conditional expectation as a Bregman predictor, IEEE Transactions on
Information Theory 51, 2664 (2005) .
[4] A. Djouadi,O. Snorrason and F. D. Garber, The quality of training sample estimates of the Bhattacharyya coefficient, IEEE Transactions on Pattern Analysis and Machine Intelligence 12, 92 (1990).
[5] W. K. Wootters, Statistical distance and Hilbert space, Physical Review D 23, 357 (1981).
[6] R. Jozsa, Fidelity for mixed quantum states, Journal of Modern Optics 41, 2315 (1994).
[7] C. Ferrie and R. Blume-Kohout, Estimating the bias of a noisy coin, AIP Conference Proceedings 1443, 14 (2012).
[8] E. L. Lehmann and G. Casella, Theory of point estimation, Springer (1998).
[9] J. Aitchison, Goodness of prediction fit, Biometrika 62, 547 (1975).
[10] A. Bhattacharyya, On a measure of divergence between two statistical populations defined by their probability distributions, Bulletin
of the Calcutta Mathematical Society 35, 99 (1943).
[11] S. Boyd, Convex Optimization, Cambridge University Press (2004).
[12] M. A. Nielsen and I. L. Chuang, Quantum computation and quantum information. Cambridge University Press (2010).
[13] L. M. Artiles, R. D. Gill and M. I. Guta, An invitation to quantum tomography, Journal of the Royal Statistical Society: Series B
(Statistical Methodology), 67, 109 (2005).
[14] R. Kueng and C. Ferrie, Near-optimal quantum tomography: estimators and bounds, New Journal of Physics 17, 123013 (2015).
[15] B. S. Clarke and A. R. Barron, Jeffreys’ prior is asymptotically least favorable under entropy risk,Journal of Statistical planning and
Inference 41, 37 (1994).
| 10math.ST |
ON THE PRIME GRAPH QUESTION FOR INTEGRAL GROUP RINGS OF
CONWAY SIMPLE GROUPS
arXiv:1711.11322v1 [math.RA] 30 Nov 2017
LEO MARGOLIS
Abstract. The Prime Graph Question for integral group rings asks if it is true that if the
normalized unit group of the integral group ring of a finite group G contains an element of order
pq, for some primes p and q, also G contains an element of that order. We answer this question
for the three Conway sporadic simple groups after reducing it to a combinatorial question about
Young tableaus and Littlewood-Richardson coefficients. This finishes work of V. Bovdi, A.
Konovalov and S. Linton.
1. Introduction
The unit group of the integral group ring ZG of a finite group G has given rise to many interesting research questions and results, as recorded e.g. in the monographs [Seh93, JdR16]. Many of
these questions concern the connection between finite subgroups of units of ZG and the group base
G, e.g. the Isomorphism Problem which was open for over 50 years until answered by Hertweck
[Her01] or the recently answered Zassenhaus Conjecture [EM17] which was open for over 40 years.
For both questions there turned out to be counterexamples. In this light the weaker versions of
the former Zassenhaus Conjecture, which are still open, become more important. One of these
versions, the so-called Prime Graph Question, gained some attention after being introduced by
Kimmerle in [Kim06]. To formulate it denote by V(ZG) the group of the so-called normalized
units in ZG, i.e. the units whose coefficients sum up to 1.
Prime Graph Question: Let G be a finite group and p and q some primes. If V(ZG) contains
an element of order pq, does G contain an element of order pq?
The prime graph Γ(G) of a group G is a graph whose vertices are the primes appearing as orders
of elements in G and two vertices p and q are connected if and only if G contains an element of
order pq. Hence the Prime Graph Question can also be formulated as: For G a finite group, does
Γ(G) = Γ(V(ZG)) hold?
The Prime Graph Question seems particularly approachable, since, in contrast to other questions on the finite subgroups in V(ZG), a reduction result is available here: The Prime Graph
Question holds for a group G, if it holds for all almost simple images of G [KK17, Theorem
2.1]. Recall that a group A is called almost simple if it is isomorphic to a subgroup of the automorphism group of a non-abelian simple group S containing the inner automorphisms of S, i.e.
Inn(S) ∼
= S ≤ A ≤ Aut(S). In this case S is called the socle of A.
It is known that the Prime Graph Question has a positive answer for almost simple groups
with socle isomorphic to a projective special linear group PSL(2, p) or PSL(2, p2 ) for p a prime
[BM17b, Theorem A] or an alternating group of degree up to 17 [BC17]. The question also has
a positive answer for groups whose order is divisible by exactly three pairwise different primes
[KK17, BM17c] and many almost simple groups whose order is divisible by four pairwise different
primes [BM16].
Employing a computer implementation of a method introduced by Luthar and Passi [LP89] and
Hertweck [Her07], nowadays known as the HeLP method, in a series of papers between 2007 and
2012 Bovdi, Konovalov and different collaborators of them studied the Prime Graph Question for
This research was partially supported by the European Commission individual Marie Curie-Skłodowska Fellowship in the H2020 program under Grant 705112-ZC and the FWO (Research Foundation Flanders).
1
2
LEO MARGOLIS
sporadic simple groups [BKS07, BK07a, BK07b, BKL08, BKM08, BK08, BGK09, BK09, BK10,
BJK11, BKL11, BK12]. Overall they studied 17 sporadic simple groups and were able to prove
the Prime Graph Question for 13 of these groups, an overview of their results can be found in
[KK15, Section 5]. Also in [KK15, Corollary 5.4] it was recorded that the Prime Graph Question
holds for almost simple groups with a socle isomorphic to one of the 13 sporadic simple groups
for which Bovdi, Konovalov et al. proved the Prime Graph Question. The four groups studied by
Bovdi, Konovalov et al. for which they were not able to prove the Prime Graph Question were
the O’Nan simple group [BGK09] and the three Conway simple groups [BKL11]. Using a method
introduced in [BM17c] in this note we obtain a positive answer for the Prime Graph Question for
the latter groups. This is the first contribution to the Prime Graph Question for sporadic simple
groups since the papers of Bovdi, Konovalov et al.
Theorem 1.1. The Prime Graph Question has a positive answer for the sporadic Conway simple
groups Co3 , Co2 and Co1 .
Note that the outer automorphism group of each Conway simple group is trivial, so Theorem 1.1
proves the Prime Graph Question also for every almost simple group whose socle is isomorphic to
a Conway simple group.
To describe our main tool let us first introduce some notation. Let G be a finite group, u ∈
V(ZG) a unit of order n and χ an ordinary character of G with corresponding representation D.
Then we can extend D linearly to ZG and afterwards restrict it to V(ZG), obtaining an extension
of D and χ. In particular D(u) is then a matrix of finite order dividing n. Let ξ be an n-th root
of unity. Then we denote by µ(ξ, u, χ) the multiplicity of ξ as an eigenvalue of D(u). Moreover
for an integer k we will denote by ζk a primitive k-th root of unity. Then the results given above
will follow from an application of the following theorem.
Theorem 1.2. Let G be a finite group, p an odd prime and m a positive integer not divisible by
p. Assume that some p-block of G is a Brauer Tree Algebra with ordinary characters χ1 , ..., χp and
with the Brauer tree being a line of the form
χ1
χ2
χ3
χp−2
χp−1
χp
.
Moreover assume that in the rings of values of χ1 , ..., χp the prime p is unramified.
Let u ∈ V(ZG) be a unit of order pm and let ξ be some m-th root of unity. Then the inequality
p−2
X
(−1)i µ(ξ · ζp , u, χi )
µ(ξ · ζp , u, χp−1 ) − µ(ξ · ζp , u, χp ) ≤ µ(ξ, u, χ1 ) −
i=1
holds.
This theorem might be regarded as a generalization of [BM16, Proposition 3.2]. The condition
on the unit can also be formulated in terms of character values which makes the inequality easier
to check by hand, although it becomes much longer then, cf. Lemma 2.3.
Theorem 1.2 will be achieved using the so-called lattice method introduced in [BM17c] and
developed further in [BM16]. It is typical for the study of group rings that different fields of mathematics, such as group theory, ring theory, representation theory and number theory are combined
to achieve results. In the present paper the main tool is combinatorics, more precisely calculations
with Young tableaus and Littlewood-Richardson coefficients, cf. the proof of Theorem 1.2. It is
quite plausible that Theorem 1.2, which is the main tool to prove Theorem 1.1, can be applied
also to study the Prime Graph Question or related questions for many other groups. Also the
methods presented here could be used to prove variations of Theorem 1.2, in particular for other
forms of Brauer trees.
We will also need to use the HeLP method to obtain enough restrictions on torsion units which
allow the application of Theorem 1.2. We will start by recalling the needed methods in Section 2,
develop the necessary combinatorics in Section 3, continue to prove Theorem 1.2 and finally show
how this can be applied to prove Theorem 1.1 in Section 4.
ON THE PRIME GRAPH QUESTION FOR INTEGRAL GROUP RINGS OF CONWAY SIMPLE GROUPS
3
2. Preliminaries and methods
A fundamental notion when studying torsion units of
Pintegral group rings are so-called partial
augmentations. Let u ∈ ZG be an element of the form g∈G zg g and denote by xG the conjugacy
P
class of an element x in G. Then εxG (u) = g∈xG zg is called the partial augmentation of u at
x. A fundamental theorem of Hertweck [Her07, Theorem 2.3] states that if u ∈ V(ZG) is a unit
of order n then εxG (u) 6= 0 implies that the order of x divides n. This implies in particular that
the exponents of G and V(ZG) coincide, a result due originally to Cohn and Livingstone [CL65].
Moreover ε1 (u) = 0 unless u = 1 by the Berman-Higman Theorem [JdR16, Proposition 1.5.1].
2.1. The HeLP method. Partial augmentations are class functions and as such can be investigated using representation theory. Note
Pthat if χ is an ordinary character of G and u ∈ V(ZG)
a torsion unit of order n then χ(u) = xG εxG (u)χ(g), where the sum runs over the conjugacy
classes of G. It was shown by Hertweck [Her07, Theorem 3.2] that this also holds for p-Brauer
characters, if p is not a divisor of n and the sum is understood to run on p-regular conjugacy
classes of G. On the other hand if we are given the partial augmentations of u and its powers
we can compute the eigenvalues, including multiplicities, of u under any ordinary or p-modular
representation of G, again assuming p is not dividing n. These multiplicities can be expressed in
explicit formulas, depending on the partial augmentations of u and its powers, and that the results
of these formulas are non-negative integers is the basic idea of the HeLP method, cf. [BM17b,
Section 2] for a detailed explanation and [Her07] for the proofs.
Hence there is an algorithmic method which allows to obtain restrictions on the possible partial
augmentations of torsion units in V(ZG), provided we have some knowledge on the characters of
G. If we apply the HeLP method to the whole character table of G then, for any fixed n, we will
obtain a finite number of possibilities for the partial augmentations of units of order n in V(ZG).
If it happens that we obtain no possibility for units of some given order n, then we know that
there exist no units at all in V(ZG) of order n. This is the basic idea in the application of the
HeLP method for the study of the Prime Graph Question and in this way Bovdi, Konovalov et
al. gave their proofs for 13 sporadic groups. We will also rely on the HeLP method to produce
a finite number of possibilities for the partial augmentations of units of certain orders and then
Theorem 1.2 will do the rest. We will apply here a computer implementation of the HeLP method
as a package [BM17a] in the computer algebra system GAP [GAP17]. This package can also be
used to reproduce the results of Bovdi, Konovalov et al.
2.2. The lattice method. The lattice method is an other method which makes use of the multiplicities of eigenvalues of a torsion unit u ∈ V(ZG) of order n under ordinary representations of
G. This method was introduced in [BM17c] and the basic idea consists in obtaining restrictions
on the simple kG-modules when viewed as khui-modules, where k is a field of characteristic p
for a divisor p of n. These simple kG-modules are the p-modular composition factors of ordinary
representations of G and studying different ordinary representations with common composition
factors can finally produce a contradiction to the existence of u.
We will only recall those parts of the lattice method necessary for our means in this article.
We will cite the articles [BM17c, BM16] where this method is explained, but many facts can be
found in text books on representation theory. First of all let C = hci be a cyclic group of order
pm where p is a prime and m an integer not divisible by p. Let k be a field of characteristic
p containing a primitive m-th root of unity ξ. Then there are, up to isomorphism, exactly pm
indecomposable kC-modules determined by a pair (i, j) where 1 ≤ i ≤ p, 1 ≤ j ≤ m such that a
module M corresponding to (i, j) is i-dimensional as a k-vector space and cp acts as ξ j on M . We
denote this module by Iij and if the action of cp is clear from the context or if m = 1 we simply
write Ii . The module Iij is simple if and only if i = 1 and each Iij is uniserial [BM17c, Proposition
2.2].
Now assume that M is a kC-module, d-dimensional as k-vector space, such that cp acts as ξ j
on M for some fixed j. Then by the above we know that M ∼
= ap Ip ⊕ ap−1 Ip−1 ⊕ · · · ⊕ a1 I1 for
4
LEO MARGOLIS
certain non-negative integers a1 , ..., ap . Moreover ap p + ap−1 (p − 1) + ... + a1 = d and hence
λ = (p, ..., p, p − 1, ..., p − 1, . . . , 1, ..., 1)
| {z }
{z
}
| {z } |
ap
ap−1
a1
is a partition of d. We call λ the partition corresponding to M .
We recall the necessary background from combinatorics, cf. e.g. [Ful97]. A Young diagram
corresponding to a partition λ = (λ1 , ..., λr ) is a diagram consisting of boxes ordered in rows and
columns such that the first row contains λ1 boxes, the second row λ2 boxes etc. If µ = (µ1 , ..., µr )
is a subpartition of λ, i.e. µ1 ≤ λ1 , . . . , µr ≤ λr , then the skew diagram corresponding to λ/µ
is obtained by erasing from the Young diagram corresponding to λ, always from left to right, µ1
boxes in the first row, µ2 boxes in the second row etc. If we fill a skew diagram with entries from
an alphabet, in our case the alphabet always will be the positive integers, a skew diagram becomes
a skew tableau T . By "filling" we mean writing a letter in every box. T is called semistandard
if reading a row from left to right the entries do not decrease and reading a column from top to
bottom the entries strictly increase.
For a box b in a skew tableau T we denote by w(b) the word which we obtain reading T from
top to bottom and from right to left until b, including the entry in b. We denote by w(T ) the
word which we obtain reading all boxes of T in this way, i.e. if b is the lowest box in the first (i.e.
most left) column of T then w(b) = w(T ). We say that T satisfies the lattice property if for every
box b in T the word w(b) contains the letter 1 at least as many times as the letter 2, the letter
2 at least as many times as 3 etc. If the maximal letter of w(T ) is s and w(T ) contains ν1 times
the letter 1, ν2 times the letter 2 etc. then ν = (ν1 , ..., νs ) is called the content of T . Note that
if T is a skew tableau satisfying the lattice property then the content of T is a partition. If now
ν is some partition then the Littlewood-Richardson coefficient cλµ,ν is the number of semistandard
skew tableaus corresponding to λ/µ which satisfy the lattice property and have content ν. A fact
of fundamental importance to us is that the Littlewood-Richardson coefficient cλµ,ν is symmetric
in µ and ν.
These combinatorial objects play a role in the study of torsion units of group rings via the
following observation [BM16, Theorem 2.8].
Theorem 2.1. Let C = hci be a cyclic group of order p, for p a prime, and k a field of characteristic p. Let M, U and Q be kC-modules with corresponding partitions λ, µ and ν respectively.
Then M contains a submodule Ũ isomorphic to U such that M/Ũ ∼
= Q if and only if cλµ,ν 6= 0.
To pass from multiplicities of eigenvalues under ordinary representations to modules over modular group algebras we will employ the following [BM17c, Propositions 2.3, 2.4].
Proposition 2.2. Let C = hci be a cyclic group of order pm such that p does not divide m. Let
R be a complete local ring of characteristic 0 containing a primitive m-th root of unity ξ such that
p is contained in the maximal ideal of R and not ramified in R. Denote by k the residue class field
of R and adopt the bar-notation for reduction modulo the maximal ideal of R.
Let D be a representation of C such that the eigenvalues of D(u) in the algebraic closure of the
quotient field of R, with multiplicities, are ξA1 , ξ 2 A2 , . . . , ξ m Am for certain multisets Ai consisting
of p-th roots of unity. Here also Ai = ∅ is possible. Let ζ be a non-trivial p-th root of unity. Note
that since the sum of the eigenvalues of D(u) is an element in R we know for every i that if Ai
contains ζ exactly r times then Ai contains also ζ 2 , ..., ζ p−1 exactly r times.
Let M be an RC-lattice affording the representation D. Then
M∼
= M1 ⊕ M2 ⊕ ... ⊕ Mm
such that for every i we have: If Ai contains ζ exactly r times and 1 exactly s times then
Mi ∼
= aI i ⊕ (r − a)I i ⊕ (s − a)I i
p
p−1
1
for some non-negative integer a ≤ min{r, s}.
For u ∈ V(ZG) a unit of order pm for which we know the partial augmentations, including
those of its powers, these facts allow to derive information on the isomorphism type of simple
ON THE PRIME GRAPH QUESTION FOR INTEGRAL GROUP RINGS OF CONWAY SIMPLE GROUPS
5
kG-modules, for k a big enough field of characteristic p, from the eigenvalues of u under ordinary
representations of G, see the proof of Theorem 1.2, or [BM17c] for a more detailed sketch of the
method.
2.3. A reformulation using character values. The multiplicities in Theorem 1.2 can be computed from the partial augmentations of u and its proper powers. This can be done by hand or using
the GAP-package HeLP [BM15], more precisely the command HeLP_MultiplicitiesOfEigenvalues.
The theorem can however also be translated in a condition on character values, so that the condition of Theorem 1.2 can be more easily checked without using GAP (or just to look up the character
values). This is particularly handy if the characters in question only have integral values and the
following lemma provides a way to do this.
Lemma 2.3. Let G be a finite group and u ∈ V(ZG) a unit of order pq where p and q are different
primes. Let χ be a character of G which takes only integral values. Set χ(1) = d, χ(up ) = x,
χ(uq ) = y and χ(u) = z. Then the following formulas hold.
pq · µ(1, u, χ) = d + (q − 1)x + (p − 1)y + (p − 1)(q − 1)z,
pq · µ(ζp , u, χ) = d + (q − 1)x − y − (q − 1)z,
pq · µ(ζq , u, χ) = d − x + (p − 1)y − (p − 1)z,
pq · µ(ζpq , u, χ) = d − x − y + z.
Proof. Note that for i not divisible by p we have µ(ζp , u, χ) = µ(ζpi , u, χ), for i not divisible by q
i
, u, χ). So
we have µ(ζq , u, χ) = µ(ζqi , u, χ) and for i not divisible by pq also µ(ζpq , u, χ) = µ(ζpq
χ(1) = µ(1, u, χ) + (p − 1)µ(ζp , u, χ) + (q − 1)µ(ζq , u, χ) + (p − 1)(q − 1)µ(ζpq , u, χ),
χ(uq ) = µ(1, u, χ) − µ(ζp , u, χ) + (q − 1)µ(ζq , u, χ) − (q − 1)µ(ζpq , u, χ),
χ(up ) = µ(1, u, χ) + (p − 1)µ(ζp , u, χ) − µ(ζq , u, χ) − (p − 1)µ(ζpq , u, χ),
χ(u) = µ(1, u, χ) − µ(ζp , u, χ) − µ(ζq , u, χ) + µ(ζpq , u, χ).
The lemma now follows by linear transformations.
3. Combinatorics
The main result of this section, which is the main ingredient in the proof of Theorem 1.2,
is Proposition 3.5. It provides bounds on the possible entries in a semistandard skew tableau
satisfying the lattice property for a special form of skew tableaus which are typical for our intended
application.
Definition 3.1. We call a skew diagram of form A if it consists of p columns such that the second
to (p − 1)-th column all end at the same row, say the m-th row. This is equivalent to saying that
there is a number m such that for a box b in the first to (p − 2)-th column and in the k-th row,
for k ≤ m, there is always a box to the right of b and moreover the second column contains no
box in the (m + 1)-th row. Furthermore we also allow the p-th column to be empty.
We call the boxes in the first to (p − 1)-th column which lie in the m-th row and above the body
of the skew diagram. The boxes of the first column lying below the m-th row are called the tail
and the p-th column the head of the tableau.
For a semistandard skew tableau T satisfying the lattice property we denote by γi (T ) the
number of letters which appear at least i times in w(T ).
Example 3.2. In Figure 1 the upper skew diagram is of form A while the lower one is not. In the
first skew diagram the boxes in the head are marked with h, the boxes of the body with b and the
boxes of the tail with t.
Lemma 3.3. Let T be a semistandard skew tableau of form A with p columns satisfying the lattice
property. Let b be a box in the k-th column lying in the body of T with entry e where 1 ≤ k ≤ p − 1.
Then w(b) contains e at least (p − k) times.
6
LEO MARGOLIS
h
b
b
b
b
b
b
b
b
b
b
b
h
t
t
Figure 1. A skew diagram of form A and not of form A
Proof. We will argue by decreasing induction. For k = p − 1 the claim is clear, since being the
entry in b the letter e appears in w(b) at least once. So let k < p − 1 and let er be the entry in
the box br right from b. Note that br exists by the assumption that T is of form A and b a box in
the body of T . Then by induction the letter er appears at least p − k − 1 times in w(br ). Since
e ≤ er , since T is semistandard, and w(br ) satisfies the lattice property also e must appear at least
p − k − 1 times in w(br ). Hence e appears at least p − k times in w(b).
Lemma 3.4. Let T be a semistandard skew tableau of form A with p columns satisfying the lattice
property. Let 1 ≤ k ≤ p − 2 and let b be a box in the k-th column of T inside the body of T with
entry e. Assume that the box right from b is the h-th box in the (k + 1)-th column of T . Then
e ≤ h.
Proof. Assume that reading from top to bottom and right to left b is the first box contradicting
the claim. In particular, e > h. Let br be the box right from b with entry er . We will show that the
(k + 1)-th column contains every letter 1, ..., e − 1. This will imply the lemma since if the (k + 1)-th
column contains the letters 1, ..., e − 1 then er = h < e, contradicting that T is a semistandard
tableau. We will show also that when bk+1 is a box in the (k + 1)-th column containing an entry
es smaller than e then this entry appears exactly (p − k) times in w(bk+1 ) and if there is a box
left from bk+1 then the entry of this box is strictly smaller than es .
We will argue by decreasing induction. By Lemma 3.3 the word w(b) contains the letter e at
least p − k times and so w(b) also contains every letter smaller than e at least p − k times. By
assumption b is the first box violating the lemma, so if there exists a box above b with entry ea
then e − ea ≥ 2. In any case, if such a box exists or not, we have that the k-th column does not
contain the entry e − 1. Note that e − 1 is not the entry in a box lying in a row above e and to
the left of e, since this would be a box contradicting the fact that T is semistandard. Since w(b)
contains e − 1 at least p − k times we conclude that every column right from b must contain e − 1,
in particular the (k + 1)-th column and if bk+1 is the box in the (k + 1)-th column containing e − 1
then e − 1 appears exactly (p − k) times in w(bk+1 ). Moreover bk+1 lies above br since otherwise
e − 1 = er < e, which would contradict the fact that T is semistandard. Since the k-th column
does not contain e − 1 a box lying left from bk+1 , if it exists, must contain an entry smaller than
e − 1.
So assume that es ≤ e − 1 is an entry contained in the box bk+1 which lies in the (k + 1)-th
column, that es appears exactly (p − k) times in w(bk+1 ) and that a box left from bk+1 , if existent,
ON THE PRIME GRAPH QUESTION FOR INTEGRAL GROUP RINGS OF CONWAY SIMPLE GROUPS
7
contains an entry smaller than es . Then w(bk+1 ) contains the letter es at least p − k times and
so it contains es − 1 at least p − k times. Since the box above and left from bk+1 , if existent, has
maximal entry es − 2 we get that es − 1 must be an entry in the (k + 1)-th column, so it must lie
in the box b′k+1 above bk+1 . So es − 1 appears exactly (p − k) times in w(b′k+1 ) and the box left
from b′k+1 , if existent, contains an entry smaller than es − 1. This finishes the induction step and
thus the lemma is proved.
Proposition 3.5. Let T be a semistandard skew tableau of form A with p columns satisfying the
lattice property. Let 1 ≤ k ≤ p − 1 and let h be the hight of the k-th column inside the body of T ,
i.e. the number of boxes in the k-th column of T which lie in the body (so if k = 1 we do not count
the boxes in the tail).
Then γp−k (T ) ≥ h and if k ≥ 2 then γp−k+2 (T ) ≤ h.
Proof. Let b be the lowest box in the k-th column with entry e. Then just for the reason that the
entries in a column are strictly increasing we have h ≤ e. By Lemma 3.3 the entry e appears in
w(b) at least (p − k) times, so each letter 1, 2..., h appears at least p − k times in w(T ) and hence
γp−k (T ) ≥ h.
So assume k ≥ 2 and assume s = γp−k+2 (T ) > h. By Lemma 3.4 the maximal possible entry
in the box left from b is h. So inside the body an entry bigger than h does not appear in the 1st
to (k − 1)-th column. Hence each letter h + 1, ..., s must appear in the tail and in the k-th column
and every column right from the k-th column. We will show that the k-th column then contains
every letter 1, ..., s, which will yield the final contradiction, since clearly the k-th column can not
contain more than h entries. The proof is similar to the proof of Lemma 3.4.
Let b be the box in the k-th column containing s. Then s appears in w(b) exactly (p − k + 1)
times, since it appears at least overall p − k + 2 times in w(T ) and it does not appear in the 1st
to (k − 1)-th column inside the body. The box left from b, if existent, contains an entry smaller
than h, so also smaller than s, by Lemma 3.4. We will show by decreasing induction that the
k-th column contains every entry e between 1 and s such that if be is the box containing e than e
appears exactly (p − k + 1) times in w(be ) and the box left from be , if existent, contains an entry
smaller than e. Since this holds for e = s assume that this holds for e + 1 and let be+1 be the box
in the k-th column containing e + 1. Since e + 1 appears exactly (p − k + 1) times in w(be+1 ), also
e appears at least p + k − 1 times in w(be+1 ). Since the entry in the box which lies above and to
the left of be+1 is at most e − 1 the box above be+1 must contain e. So this box is be , the letter e
appears exactly (p − k + 1) times in w(be ) and the box to the left of be contains an entry smaller
than e. This finishes the induction step and hence the proof.
4. Proofs of main results
We proceed to prove Theorem 1.2. For the basic facts about Brauer trees we refer to [HL89].
Here we will only need that if (K, R, k) is a p-modular system such that the characters labelling
the vertices of the Brauer tree are afforded by RG-modules then each edge of the Brauer tree
corresponds to a simple kG-module, each vertex of the Brauer tree corresponds to a simple RGmodule with the given character and after reducing such a simple RG-module modulo the maximal
ideal of R the composition factors of this module are those corresponding to the adjacent edges,
each with multiplicity 1.
Proof of Theorem 1.2: Let R be a complete discrete valuation ring containing a primitive m-th
root of unity such that R is an algebraic extension of Zp , representations affording the characters
χ1 , ..., χp can be realized over R and p is unramified in R. Such a ring R exists by a result
of Fong [BM16, Remark 2.5]. Let k be the residue class field of R. We use the bar-notation
to denote reduction modulo the maximal ideal of R also with respect to modules. Let M be a
khui-module. Adapting the notation from skew tableaus we denote by γi (M ) the number of direct
indecomposable summands of M of dimension at least i.
Let M1′ , ..., Mp′ be RG-modules corresponding to the characters χ1 , ..., χp and let, as explained
in Proposition 2.2, M1 , ..., Mp be the direct summands of M1′ , ..., Mp′ respectively on which up acts
′
as ξ p . Let S1′ , ..., Sp−1
be the simple kG-modules corresponding to the edges of the Brauer tree
8
LEO MARGOLIS
in the natural order, i.e. Si′ corresponds to the edge with vertices χi and χi+1 . Let S1 , ..., Sp−1
′
be the direct summand of S1′ , ..., Sp−1
respectively on which up acts as ξ p . We will view Mi
and Si as khui-modules, unless explicitly stated otherwise. Let λi be a partition describing the
isomorphism type of Mi and µi the partition describing the isomorphism type of Si . Note that
by Proposition 2.2 if Sj is a submodule of Mi then the skew diagram corresponding to λi /µj is of
form A with p columns, since the indecomposable summands of Mi are all of dimension 1, p − 1
or p, i.e. the possible length of rows in a Young diagram corresponding to λi are 1, p − 1 and p.
Note that if we want to compute the possible isomorphism types of Si for some 1 < i < p − 1
using Theorem 2.1 we can consider a semistandard skew tableau satisfying the lattice property of
′
form λi /µi−1 or λi+1 /µi+1 . Indeed the composition factors of Mi′ as kG-module are Si−1
and Si′ .
′
If Si−1
is a submodule of Mi′ as kG-module then Si ∼
= Mi /Si−1 and cλµii−1 ,µi 6= 0. If on the other
′
λi
′
hand Si is a submodule of Mi as kG-module then cµi ,µi−1 6= 0, implying cλµii−1 ,µi 6= 0 and so we
can also consider λi /µi−1 . The same arguments apply for λi+1 /µi+1 .
We will show that
γp−2 (Sp−2 ) ≥ µ(ξ · ζp , u, χp−1 ) − µ(ξ · ζp , u, χp )
and also
γp−2 (Sp−2 ) ≤ µ(ξ, u, χ1 ) −
p−2
X
(−1)i µ(ξ · ζp , u, χi )
i=1
which will imply the theorem. First by Theorem 2.1 the isomorphism type of Sp−2 can be described
by a semistandard skew tableau T satisfying the lattice property of shape λp−1 /µp−1 . Note that
µp−1 = λp since Mp ∼
= Sp−1 . By Proposition 3.5 we know that γp−2 (T ) = γp−2 (Sp−2 ) is at least as
big as the height of the second column of T . Now by Proposition 2.2 the number of indecomposable
direct summands of dimension at least 2 in Mp−1 , i.e. the number of rows of length at least 2 in a
Young diagram corresponding to λp−1 , is µ(ξ · ζp , u, χp−1 ) and the analogues statement holds for
λp . So the height of the second column of T is µ(ξ · ζp , u, χp−1 ) − µ(ξ · ζp , u, χp ), proving the first
claim.
To prove the second claim we will prove by induction on r that if r is odd then
γr (Sr ) ≤ µ(ξ, u, χ1 ) −
r
X
(−1)i µ(ξ · ζp , u, χi )
i=1
and if r is even then
γp+1−r (Sr ) ≥ −µ(ξ, u, χ1 ) +
r
X
(−1)i µ(ξ · ζp , u, χi ).
i=1
Once this induction reaches r = p − 2 we will be done. If r = 1 then γ1 (S1 ) is the number of
indecomposable direct summands of S1 ∼
= M1 which by Proposition 2.2 is at most µ(ξ, u, χ1 ) +
µ(ξ · ζp , u, χ1 ).
For the induction step assume that the statement holds for all numbers smaller than r. First
assume that r is even. Now Sr corresponds to a semistandard skew tableau T satisfying the lattice
property of form λr /µr−1 . So by Proposition 3.5 the number γp+1−r (Sr ) = γp+1−r (T ) is at least
as big as the height of the (r − 1)-th column in the body of T . Since r ≤ p this numbers equals
the difference of the number of direct indecomposable summands of Mr which is at least (r − 1)dimensional, but not 1-dimensional, and the number of indecomposable direct summands of Sr−1
of dimension at least r−1, i.e. γr−1 (Sr−1 ). The number of indecomposable direct summands of Mr
which are at least (r − 1)-dimensional, but not 1-dimensional, is µ(ξ · ζp , u, χr ) by Proposition 2.2.
Pr−1
Since by induction γr−1 (Sr−1 ) ≤ µ(ξ, u, χ1 ) − i=1 (−1)i µ(ξ · ζp , u, χi ) we obtain that the height
of the (r − 1)-th column in the body of T is at least
µ(ξ · ζp , u, χr ) − (µ(ξ, u, χ1 ) −
r−1
X
i=1
(−1)i µ(ξ · ζp , u, χi )) = −µ(ξ, u, χ1 ) +
r
X
i=1
(−1)i µ(ξ · ζp , u, χi ).
ON THE PRIME GRAPH QUESTION FOR INTEGRAL GROUP RINGS OF CONWAY SIMPLE GROUPS
9
Now assume that r is odd, bigger than 2 and smaller than p−1. Also in this case Sr corresponds
to a semistandard skew tableau T satisfying the lattice property of form λr /µr−1 . So the number
γr (Sr ) = γr (T ) is at most as big as the hight of the (p + 2 − r)-th column in the body of T
by Proposition 3.5. This is the difference of the number of indecomposable direct summands of
dimension at least p + 2 − r in Mr , i.e. γp+2−r (Mr ), and the number of indecomposable direct
summands of Sr−1 of dimension at least p + 2 − r, i.e. γp+2−r (Sr−1 ) = γp+1−(r−1) (Sr−1 ). By
Proposition 2.2 we know γp+2−r (Mr ) = µ(ξ · ζp , u, χr ) and by induction γp+1−(r−1) (Sr−1 ) is at
Pr−1
least −µ(ξ, u, χ1 ) + i=1 (−1)i µ(ξ · ζp , u, χi ). So the height of the (p + 2 − r)-th column of T is
at most
r
r−1
X
X
(−1)i µ(ξ · ζp , u, χi ),
(−1)i µ(ξ · ζp , u, χi )) = µ(ξ, u, χ1 ) −
µ(ξ · ζp , u, χr ) − (−µ(ξ, u, χ1 ) +
i=1
i=1
finishing the proof of the induction claim.
The proof of the Prime Graph Question for the Conway simple groups now boils down to the
application of Theorem 1.2 for various cases described in [BKL11]. The information contained in
[BKL11] is not completely sufficient for our purposes, since to compute multiplicities of eigenvalues
we do not only need to know the partial augmentations of a unit u ∈ V(ZG), but also the partial
augmentations of its proper powers. These can be computed using the GAP-package HeLP [BM17a]
and we will indicate in all cases we need to consider which characters are sufficient to obtain these
possible partial augmentations using the command HeLP_WithGivenOrder. Also in the case of the
first Conway group we will use stronger results obtainable by the HeLP method than those given
in [BKL11].
We will denote by χi the i-th irreducible complex character of a group G as given in the GAP
character table library [Bre12]. We will also use names for conjugacy classes as in [Bre12]. The
statements we will need about the p-blocks of various groups, their defect and their corresponding
Brauer trees can all be derived from GAP, but they are also given in [HL89].
Proof of Theorem 1.1: We will study the three cases of interest separately.
Case G = Co3 : Assume first that G = Co3 . By [BKL11, Theorem 1] we only have to consider
units of order 35 in V(ZG). Using χ2 to compute partial augmentations for units of order 5 and
χ2 , χ3 to compute the partial augmentations of units of order 35 we are left with the following
possibilities:
(ε5a (u7 ), ε5b (u7 ), ε5a (u), ε5b (u), ε7a (u)) ∈ {(−4, 5, 3, 12, −14), (−3, 4, 4, 11, −14)}.
Note that there is only one class of elements of order 7 in G, so we do not need to consider
the partial augmentations of u5 , since the only class in which a partial augmentation at u5 is
non-vanishing is 7a.
G possesses a 5-block of defect 1 whose Brauer tree is a line of form
χ5
χ29
χ39
χ35
χ12
.
All irreducible ordinary characters in this block have only integral values and 5 is of course not
ramified in Z. So we can apply Theorem 1.2. We provide the multiplicities of the needed eigenvalues of u, for the two critical distributions of partial augmentations, in Table 1. Here the first
entry in each column contains the possible values of (ε5a (u7 ), ε5b (u7 ), ε5a (u), ε5b (u), ε7a (u)).
So using Theorem 1.2 with ξ = 1 we would have, e.g. in case (ε5a (u7 ), ε5b (u7 ), ε5a (u), ε5b (u), ε7a (u)) =
(−4, 5, 3, 12, −14) that
µ(ζ5 , u, χ35 ) − µ(ζ5 , u, χ12 ) = 4967 ≤ µ(1, u, χ5 ) + µ(ζ5 , u, χ5 ) − µ(ζ5 , u, χ29 ) + µ(ζ5 , u, χ39 ) = 4945,
a contradiction. The same argument applies to the other possibility of distributions of partial
augmentations of u, in which case we get 4965 ≤ 4944.
Case G = Co2 : Let G = Co2 . By [BKL11, Theorem 2] we only have to consider units of order
35. Using χ2 to compute partial augmentations of units of order 5 and χ2 , χ3 for order 35 we get
(ε5a (u7 ), ε5b (u7 ), ε5a (u), ε5b (u), ε7a (u)) ∈ {(−4, 5, 3, 12, −14), (−3, 4, 4, 11, −14)}.
10
LEO MARGOLIS
(−4, 5, 3, 12, −14)
(−3, 4, 4, 11, −14)
χ
µ(1, u, χ) µ(ζ5 , u, χ) µ(1, u, χ) µ(ζ5 , u, χ)
χ5 33
2
29
3
χ29
2119
2118
7029
7030
χ39
χ35
5071
5070
104
105
χ12
Table 1. Multiplicities of eigenvalues for G = Co3 for units of order 35.
Again there is only one class of elements of order 7, so we do not need to consider u5 .
Also in this case G possesses a 5-block of defect 1 with Brauer tree of the form
χ4
χ24
χ43
χ38
χ20
.
All ordinary characters in this block only have integral values. The necessary multiplicities are
provided in Table 2
(−4, 5, 3, 12, −14)
(−3, 4, 4, 11, −14)
χ
µ(1, u, χ) µ(ζ5 , u, χ) µ(1, u, χ) µ(ζ5 , u, χ)
χ5 33
2
29
3
3269
3268
χ29
χ39
13354
13355
11396
11395
χ35
χ12
1254
1255
Table 2. Multiplicities of eigenvalues for G = Co2 for units of order 35.
In the first case using Theorem 1.2 with ξ = 1 we get 10142 ≤ 10120 and in the second case
10140 ≤ 10119.
Case G = Co1 : Let G = Co1 . By [BKL11, Theorem 3] we only have to consider units of order
55 and 65. We will first apply the HeLP method here to obtain stronger results than in [BKL11].
The 2-, 3- and 5-modular Brauer tables are not available in GAP, but the Atlas, also implemented
as a GAP-package [WPN+ 11], contains several representations in these characteristics and we will
use a character ψ coming from a 24-dimensional representation of G over F2 . The values of ψ on
classes of interest can be found in Table 3.
1a 5a 5b 5c 11a 13a
ψ 24 −6 4 −1
2
−2
Table 3. A 2-modular Brauer character for Co1 .
The results we obtain using the HeLP method are stronger than those in [BKL11], since we
include ψ. Using χ2 , χ3 and ψ we obtain 98 possibilities for the partial augmentations of units of
order 5. Then using χ2 , χ4 , χ5 and ψ we can exclude the existence of units of order 65 in ZG. For
units of order 55 we apply the characters χ2 , χ3 , χ4 and ψ. Note that there is only one class of
elements of order 11 in G.
We get that (ε5a (u11 ), ε5b (u11 ), ε5c (u11 ), ε5a (u), ε5b (u), ε5c (u), ε11a (u)) is one of the four possibilities
(1, 5, −5, 1, −6, −5, 11), (1, 6, −6, 1, −5, −6, 11), (2, 6, −7, 2, −5, −7, 11), (2, 7, −8, 2, −4, −8, 11).
The main 11-block of G has a Brauer tree of the form
ON THE PRIME GRAPH QUESTION FOR INTEGRAL GROUP RINGS OF CONWAY SIMPLE GROUPS
χ1
χ12
χ34
χ41
χ64
χ79
χ85
χ73
χ60
11
χ38
χ14
.
Also in this block all ordinary characters only have integral values. The necessary multiplicities
are provided in Table 4.
χ
χ1
χ12
χ34
χ41
χ64
χ79
χ85
χ73
χ60
χ38
χ14
(1, 5, −5, 1, −6, −5, 11) (1, 5, −5, 1, −6, −5, 11)
µ(1, u, χ) µ(ζ11 , u, χ) µ(1, u, χ) µ(ζ11 , u, χ)
1
0
1
0
5668
5670
138600
138600
391385
391385
1929876
1929870
4495195
4495195
5326485
5326495
3734505
3734505
1522180
1522180
297293
297285
6885
6875
Table 4. Multiplicities of eigenvalues for
(1, 5, −5, 1, −6, −5, 11) (1, 5, −5, 1, −6, −5, 11)
µ(1, u, χ) µ(ζ11 , u, χ) µ(1, u, χ) µ(ζ11 , u, χ)
1
0
1
0
5638
5640
138520
138520
391350
391350
1929856
1929850
4495195
4495195
5326570
5326580
3734540
3734540
1522180
1522180
297303
297295
6880
6870
G = Co1 for units of order 55.
From the four possibilities we obtain by Theorem 1.2 the inequalities 290408 ≤ 290389, 290410 ≤
290391, 290423 ≤ 290304 and 290425 ≤ 290306, all of which do not hold.
Remark 4.1. In the proof of Theorem 1.1 to handle the first Conway group we used Theorem 1.2
with p = 11 instead of p = 5 for two reasons. One is to show that Theorem 1.2 can also be
applied with other primes than 5. The other, more importantly, is that the 11-Brauer trees can be
computed using GAP which can not be done as easily with 5-Brauer trees, since the 5-Brauer table
is not available in GAP. The 5-Brauer trees are given in [HL89], but the calculations with p = 11
are easier to verify for a reader who does not have [HL89] available. Using the first Brauer tree
given in [HL89, Section 6.22] and Theorem 1.2 with p = 5 and ξ = 1 one can also get the same
result.
Remark 4.2. Of the 17 sporadic groups studied by Bovdi, Konovalov et al. the only group for
which they could not prove the Prime Graph Question, apart from the Conway groups, was the
O’Nan simple group [BGK09]. Here it remains to exclude the existence of units of order 33 and
57. Theorem 1.2 can be applied here with p = 11 to handle units of order 33, but not for order
57. This is not possible since for p ∈ {3, 19} every p-block of G which is a Brauer Tree Algebra
contains characters such that p is ramified in the ring of values of these characters.
Acknowledgements
The setting handled in this paper was discussed at various stages with Alexander Konovalov,
Andreas Bächle, Alexander Thumm and Eugenio Giannelli. I am thankful to them all for their
input.
References
[BC17]
[BGK09]
[BJK11]
A. Bächle and M. Caicedo, On the prime graph question for almost simple groups with an alternating
socle, Internat. J. Algebra Comput. 27 (2017), no. 3, 333–347.
V. Bovdi, A. Grishkov, and A. Konovalov, Kimmerle conjecture for the Held and O’Nan sporadic simple
groups, Sci. Math. Jpn. 69 (2009), no. 3, 353–361.
V. Bovdi, E. Jespers, and A. Konovalov, Torsion units in integral group rings of Janko simple groups,
Math. Comp. 80 (2011), no. 273, 593–615.
12
[BK07a]
LEO MARGOLIS
V. Bovdi and A. Konovalov, Integral group ring of the first Mathieu simple group, Groups St. Andrews
2005. Vol. 1, London Math. Soc. Lecture Note Ser., vol. 339, Cambridge Univ. Press, Cambridge, 2007,
pp. 237–245.
, Integral group ring of the McLaughlin simple group, Algebra Discrete Math. (2007), no. 2,
[BK07b]
43–53.
[BK08]
, Integral group ring of the Mathieu simple group M23 , Comm. Algebra 36 (2008), no. 7, 2670–
2680.
[BK09]
, Integral group ring of Rudvalis simple group, Ukraïn. Mat. Zh. 61 (2009), no. 1, 3–13.
, Torsion units in integral group ring of Higman-Sims simple group, Studia Sci. Math. Hungar.
[BK10]
47 (2010), no. 1, 1–11.
[BK12]
, Integral group ring of the Mathieu simple group M24 , J. Algebra Appl. 11 (2012), no. 1,
1250016, 10.
[BKL08]
V. Bovdi, A. Konovalov, and S. Linton, Torsion units in integral group ring of the Mathieu simple
group M22 , LMS J. Comput. Math. 11 (2008), 28–39.
[BKL11]
, Torsion units in integral group rings of Conway simple groups, Internat. J. Algebra Comput.
21 (2011), no. 4, 615–634.
[BKM08] V. Bovdi, A. Konovalov, and E. Marcos, Integral group ring of the Suzuki sporadic simple group, Publ.
Math. Debrecen 72 (2008), no. 3-4, 487–503.
[BKS07]
V. Bovdi, A. Konovalov, and S. Siciliano, Integral group ring of the Mathieu simple group M12 , Rend.
Circ. Mat. Palermo (2) 56 (2007), no. 1, 125–136.
[BM15]
A. Bächle and L. Margolis, HeLP – A GAP-package for torsion units in integral group rings, preprint,
http://arxiv.org/abs/1507.08174 (2015), 6 pages.
, On the Prime Graph Question for Integral Group Rings of 4-primary groups II, preprint,
[BM16]
arxiv.org/abs/1606.01506 (2016), 17 pages.
[BM17a]
,
HeLP
–
Hertweck-Luthar-Passi
method,
GAP
package,
Version
3.1,
http://homepages.vub.ac.be/abachle/help/, 2017.
, On the prime graph question for integral group rings of 4-primary groups I, Internat. J. Algebra
[BM17b]
Comput. 27 (2017), no. 6, 731–767.
[BM17c]
, Rational Conjugacy of Torsion Units in Integral Group Rings of Non-Solvable Groups, Proc.
Edinb. Math. Soc. (2) 60 (2017), no. 4, 813–830.
[Bre12]
T.
Breuer,
The
GAP
Character
Table
Library,
Version
1.2.1,
http://www.math.rwth-aachen.de/~Thomas.Breuer/ctbllib, May 2012, GAP package.
[CL65]
J. A. Cohn and D. Livingstone, On the structure of group algebras. I, Canad. J. Math. 17 (1965),
583–593.
[EM17]
F. Eisele and L Margolis,
A counterexample to the first Zassenhaus conjecture,
http://arxiv.org/abs/1710.08780 (2017), 32 pages.
[Ful97]
W. Fulton, Young tableaux, London Mathematical Society Student Texts, vol. 35, Cambridge University
Press, Cambridge, 1997, With applications to representation theory and geometry.
[GAP17]
The GAP Group, GAP – Groups, Algorithms, and Programming, Version 4.8.8, 2017,
http://www.gap-system.org.
[Her01]
M. Hertweck, A counterexample to the isomorphism problem for integral group rings, Ann. of Math.
(2) 154 (2001), no. 1, 115–138.
[Her07]
, Partial augmentations and Brauer character values of torsion units in group rings,
arXiv:0612429v2 [math.RA], 2004 - 2007.
[HL89]
G. Hiss and K. Lux, Brauer trees of sporadic groups, Oxford Science Publications, The Clarendon
Press, Oxford University Press, New York, 1989.
[JdR16]
E. Jespers and Á. del Río, Group ring groups. Vol. 1. Orders and generic constructions of units, De
Gruyter Graduate, De Gruyter, Berlin, 2016.
[Kim06]
W. Kimmerle, On the prime graph of the unit group of integral group rings of finite groups, Groups,
rings and algebras, Contemp. Math., vol. 420, Amer. Math. Soc., Providence, RI, 2006, pp. 215–228.
[KK15]
W. Kimmerle and A. Konovalov, Recent advances on torsion subgroups of integral group rings, Groups
St Andrews 2013, London Math. Soc. Lecture Note Ser., vol. 422, Cambridge Univ. Press, Cambridge,
2015, pp. 331–347.
, On the Gruenberg–Kegel graph of integral group rings of finite groups, Internat. J. Algebra
[KK17]
Comput. 27 (2017), no. 6, 619–631.
[LP89]
I. S. Luthar and I. B. S. Passi, Zassenhaus conjecture for A5 , Proc. Indian Acad. Sci. Math. Sci. 99
(1989), no. 1, 1–5.
[Seh93]
S. K. Sehgal, Units in integral group rings, Pitman Monographs and Surveys in Pure and Applied
Mathematics, vol. 69, Longman Scientific & Technical, Harlow, 1993.
[WPN+ 11] R.A. Wilson,
R.A. Parker,
S. Nickerson,
J.N. Bray,
and T. Breuer,
AtlasRep,
A GAP Interface to the Atlas of
Group Representations,
Version 1.5,
http://www.math.rwth-aachen.de/~Thomas.Breuer/atlasrep, July 2011, Refereed GAP package.
ON THE PRIME GRAPH QUESTION FOR INTEGRAL GROUP RINGS OF CONWAY SIMPLE GROUPS
13
Vrije Universiteit Brussel, Faculty of Sciences, Department of Mathematics, Pleinlaan 2, B-1050
Brussels, Belgium
E-mail address: leo.margolis@vub.ac.be
| 4math.GR |
arXiv:1212.4590v1 [math.AP] 19 Dec 2012
RELATIVE PARAMETRIZATION OF
LINEAR MULTIDIMENSIONAL SYSTEMS
J.F. Pommaret
CERMICS, Ecole Nationale des Ponts et Chaussées,
6/8 Av. Blaise Pascal, 77455 Marne-la-Vallée Cedex 02, France
E-mail: jean-francois.pommaret@wanadoo.fr, pommaret@cermics.enpc.fr
URL: http://cermics.enpc.fr/∼pommaret/home.html
ABSTRACT :
In the last chapter of his book ”The Algebraic Theory of Modular Systems ” published in 1916,
F. S. Macaulay developped specific techniques for dealing with ”unmixed polynomial ideals ” by
introducing what he called ”inverse systems ”. The purpose of this paper is to extend such a
point of view to differential modules defined by linear multidimensional systems, that is by linear
systems of ordinary differential (OD) or partial differential (PD) equations of any order, with any
number of independent variables, any number of unknowns and even with variable coefficients.
The first and main idea is to replace unmixed polynomial ideals by pure differential modules.
The second idea is to notice that a module is 0-pure if and only if it is torsion-free and thus if and
only if it admits an ” absolute parametrization ” by means of arbitrary potential like functions, or,
equivalently, if it can be embedded into a free module by means of an ” absolute localization ”.
The third idea is to refer to a difficult theorem of algebraic analysis saying that an r-pure module
can be embedded into a module of projective dimension r, that is a module admitting a projective
resolution with exactly r operators.
The fourth and final idea is to establish a link between the use of extension modules for such a
purpose and specific formal properties of the underlying multidimensional system through the use
of involution and a ”relative localization ” leading to a ” relative parametrization ”.
The paper is written in a rather effective self-contained way and we provide many explicit examples
that should become test examples for a future use of computer algebra.
KEY WORDS :
Unmixed ideal, algebraic analysis, homological algebra, extension module, projective dimension,
torsion-free module, pure module, characteristic variety, formal integrability, involution, Spencer
operator, inverse system.
1) INTRODUCTION :
Let D = K[d1 , ..., dn ] = K[d] be the ring of differential operators with coefficients in a differential field K with n commuting derivations ∂1 , ..., ∂n and commutation relations di a = adi +∂i a, ∀a ∈
K. If y 1 , ..., y m are m differential indeterminates, we may identify Dy 1 + ... + Dy m = Dy with
Dm and consider the finitely presented left differential module M with presentation Dp → Dm →
M → 0 determined by a given linear multidimensional system with n independent variables,
m unknowns and p equations. Applying the functor homD (•, D), we get the exact sequence
0 → homD (M, D) → Dm → Dp −→ N −→ 0 of right differential modules that can be transformed
by a side-changing functor to an exact sequence of finitely generated left differential modules.
This new presentation corresponds to the formal adjoint ad(D) of the linear differential operator
D determined by the initial presentation but now with p unknowns and m equations, obtaining
therefore a new finitely generated left differential module N and we may consider homD (M, D)
as the module of equations of the compatibility conditions (CC) of ad(D), a result not evident
at all (Compare to [24]). Using now a maximum free submodule 0 −→ Dl −→ homD (M, D)
1
and repeating this standard procedure while using the well known fact that ad(ad(D)) = D,
we obtain therefore an embedding 0 → homD (homD (M, D), D) → Dl of left differential modules for a certain integer 1 ≤ l < m because K is a field and thus D is a noetherian bimodule over itself, a result leading to l = rkD (homD (M, D)) = rkD (M ) < m as in ([19], p
178,201)(See section 3 for the definition of the differential rankrkD ). Now, the kernel of the map
ǫ : M → homD (homD (M, D), D) : m → ǫ(m)(f ) = f (m), ∀f ∈ homD (M, D) is the torsion submodule t(M ) ⊆ M and ǫ is injective if and only if M is torsion-free, that is t(M ) = 0. In that case,
we obtain by composition an embedding 0 → M → Dl of M into a free module that can also be
obtained by localization if we introduce the ring of fractions S −1 D = DS −1 when S = D − {0}.
This result is quite important for applications as it provides a (minimum) parametrization of the
linear differential operator D and amounts to the controllability of a classical control system when
n = 1 ([24], p 258). This parametrization will be called an ”absolute parametrization ” as it only
involves arbitrary ”potential-like ” functions (See [1], [9], [18], [19], [20], [24], [25] and [32] for more
details and examples, in particular that of Einstein equations).
The purpose of this paper is to extend suh a result to a much more general situation, that
is when M is not torsion-free, by using unexpected results first found by F.S. Macaulay in 1916
through his study of ”inverse systems ” for ”unmixed polynomial ideals ”.
For this we define the purity filtration :
0 = tn (M ) ⊆ tn−1 (M ) ⊆ ... ⊆ t1 (M ) ⊆ t0 (M ) = t(M ) ⊆ M
by introducing tr (M ) = {m ∈ M | cd(Dm) > r} where the codimension of Dm is n minus
the dimension of the characteristic variety determined by m in the corresponding system for one
unknown. The module M is said to be r-pure if tr (M ) = 0, tr−1 (M ) = M or, equivalently, if
cd(M ) = cd(N ) = r, ∀N ⊂ M and a torsion-free module is a 0-pure module. Moreover, when
K = k = cst(K) is a field of constants and m = 1, a pure module is unmixed in the sense of
Macaulay, that is defined by an ideal having an equidimensional primary decomposition.
Example 1.1 : As an elementary example with K = k = Q, m = 1, n = 2, p = 2, the differential
module defined by d22 y = 0, d12 y = 0 is not pure because z ′ = d2 y satisfies d2 z ′ = 0, d1 z ′ = 0 while
z” = d1 y only satisfies d2 z” = 0 and ((χ2 )2 , χ1 χ2 ) = (χ1 ) ∩ (χ1 , χ2 )2 . We obtain therefore the
purity filtration 0 = t2 (M ) ⊂ t1 (M ) ⊂ t0 (M ) = t(M ) = M with strict inclusions as 0 6= z ′ ∈ t1 (M )
while z” ∈ t0 (M ) but z” ∈
/ t1 (M ).
From the few (difficult) references ([1],[9],[15],[18],[27]) dealing with extension modules extr (M ) =
and purity in the framework of algebraic analysis, it is known that M is r-pure if and
only if there is an embedding 0 → M → extrD (extrD (M, D), D). Indeed, the case r = 0 is exactly
the one already considered because ext0D (M, D) = homD (M, D) and the ker/coker exact sequence:
extrD (M, D)
0 −→ ext1 (N ) −→ M −→ ext0 (ext0 (M )) −→ ext2 (N ) −→ 0
allows to test the torsion-free property of M in actual practice by using the double-duality formula
t(M ) = ext1 (N ) as in ([19]). Also, when r ≥ 1, a similar construction that we shall recall and
illustrate in section 4 provides a finitely generated module L with projective dimension pdD (L) = r,
that is a minimum resolution of L with only r operators, and an embedding 0 → M → L that
allows to exhibit a relative parametrization of D because now the parametrizing potential-like functions are no longer arbitrary but must only depend on arbitrary functions of n − r variables.
Example 1.2 : With K = k = Q, m = 2, n = 3, r = 1, the differential module M defined by
the involutive system Φ1 ≡ d3 y 1 = 0, Φ2 ≡ d3 y 2 = 0, Φ3 ≡ d2 y 1 − d1 y 2 = 0 is 1-pure and admits
the resolution 0 −→ D −→ D3 −→ D2 −→ M −→ 0. The differential module L defined by
the system d3 z = 0 is also 1-pure and admits the resolution 0 −→ D −→ D −→ L −→ 0. We
finally obtain the relative parametrization y 1 = d1 z, y 2 = d2 z providing the strict inclusion M ⊂ L.
In a simple way, this result can be considered as a measure of how far a module is from being
projective, recalling that a module P is projective if there exists another (projective) module Q
2
and a free module F such that P ⊕ Q ≃ F .
We adapt the ”relative localization” technique used by Macaulay and combine it with the ”involution” technique used in the formal theory of systems of partial differential equations in order
to obtain an explicit procedure for determining L when M is given. Many examples will illustrate these new methods that avoid the previous abstract arguments based on ”double duality”.
In particular, original non-commutative examples will also be presented. However, we point out
the fact that the latter method can be adapted without any change to the case of systems with
variables coefficients as it only depends on the use of adjoint operators but the following example
will explain by itself the type of difficulty involved.
Example 1.3 : Starting now with K = Q(x1 , x2 ), m = 2, n = 3, r = 1, the new differential module
M defined by d3 y 1 = 0, d3 y 2 = 0, d2 y 1 − d1 y 2 + x2 y 2 = 0 is also 1-pure and the differential module
L is again defined by d3 z = 0 as in the previous example. However we obtain the totally different
relative parametrization y 1 = d12 z − x2 d2 z + z, y 2 = d22 z providing the strict inclusion M ⊂ L.
More generally, we may consider a constant parameter a ∈ k = Q and consider the new system
d3 y 1 = 0, d3 y 2 = 0, d2 y 1 − d1 y 2 + ax2 y 2 = 0 depending on a. For a = 0 we find back the case of
the previous example and we let the reader wonder why the situation only changes when a 6= 0.
The content of the paper is just following the introduction.
In section 2 we recall the definitions and results from the formal theory of systems of OD/PD
equations that will be crucially used in the sequel. We pay a particular emphasize to the definition
of involution and the way to introduce the Spencer operator in this framework. We also study the
possibility and difficulty to use computer algebra in this framework.
In section 3 we recall the basic tools needed from module theory and homological algebra in
a way adapted to our purpose, in particular the definition of the extension modules, and provide
a few of their properties which, though well known by specialists of algebraic analysis, cannot be
found easily in the literature. Meanwhile, we provide a few links with the preceding section which
are not so well known. Many explicit examples will illustrate the main concepts in the commutative
(constant field k) and the non-commutative (differential field K) framework.
In section 4 we shall recall the proof of the theorem already quoted showing how to embed an
r-pure module M into another module L with projrective dimension equal to r. We shall provide
for the first time explicit computations of this result in order to point out the difficulty encountered
in such a procedure as a motivation for avoiding it.
In section 5 we extend the work of Macaulay, showing why only pure modules can fit with
relative localization in a coherent way with what happens for torsion-free modules. Meanwhile, we
shall extend for the first time this work to the non-commutative framework, showing in particular
that the operator introduced by Macaulay ([11], §60) for studying inverse systems is nothing else
than the Spencer operator. Many explicit examples, including highly non-trivial ones provided by
Macaulay himself, will be fully treated in such a way that any engineer, even with a poor knowledge of homological algebra, will nevertheless become intuitively able to understand and apply
these new techniques without reading the previous sections, just comparing to the way the same
examples have been treated in section 4 by means of another approach.
2) TOOLS FROM SYSTEM THEORY :
If X is a manifold of dimension n with local coordinates (x) = (x1 , ...xn ), we denote as usual by
T = T (X) the tangent bundle of X, by T ∗ = T ∗ (X) the cotangent bundle, by ∧r T ∗ the bundle of
r-forms and by Sq T ∗ the bundle of q-symmetric tensors. More generally, let E be a vector bundle
over X, that is (roughly) a manifold with local coordinates (xi , y k ) for i = 1, ..., n and k = 1, ..., m
simply denoted by (x, y), projection π : E → X : (x, y) → (x) and changes of local coordinates
x̄ = ϕ(x), ȳ = A(x)y. If E and F are two vector bundles over X with respective local coordinates
(x, y) and (x, z), we denote by E×X F the fibered product of E and F over X as the new vector
3
bundle over X with local coordinates (x, y, z). We denote by f : X → E : (x) → (x, y = f (x)) a
global section of E, that is a map such that π ◦ f = idX but local sections over an open set U ⊂ X
may also be considered when needed. Under a change of coordinates, a section transforms like
f¯(ϕ(x)) = A(x)f (x) and the derivatives transform like:
∂ f¯l
(ϕ(x))∂i ϕr (x) = (∂i Alk (x))f k (x) + Alk (x)∂i f k (x)
∂ x̄r
We may introduce new coordinates (xi , y k , yik ) transforming like:
ȳrl ∂i ϕr (x) = (∂i Alk (x))y k + Alk (x)yik
k
We shall denote by Jq (E) the q-jet bundle of E with local coordinates (xi , y k , yik , yij
, ...) = (x, yq )
k
k
k
called jet coordinates and sections fq : (x) → (x, f (x), fi (x), fij (x), ...) = (x, fq (x)) transforming
like the sections jq (f ) : (x) → (x, f k (x), ∂i f k (x), ∂ij f k (x), ...) = (x, jq (f )(x)) where both fq and
jq (f ) are over the section f of E. Of course Jq (E) is a vector bundle over X with projection πq
while Jq+r (E) is a vector bundle over Jq (E) with projection πqq+r , ∀r ≥ 0.
DEFINITION 2.1: A linear system of order q on E is a vector sub-bundle Rq ⊂ Jq (E) and a
solution of Rq is a section f of E such that jq (f ) is a section of Rq .
Let µ = (µ1 , ..., µn ) be a multi-index with length |µ| = µ1 + ... + µn , class i if µ1 = ... = µi−1 =
0, µi 6= 0 and µ + 1i = (µ1 , ..., µi−1 , µi + 1, µi+1 , ..., µn ). We set yq = {yµk |1 ≤ k ≤ m, 0 ≤ |µ| ≤ q}
with yµk = y k when |µ| = 0. If E is a vector bundle over X with local coordinates (xi , y k ) for
i = 1, ..., n and k = 1, ..., m, we denote by Jq (E) the q-jet bundle of E with local coordinates
simply denoted by (x, yq ) and sections fq : (x) → (x, f k (x), fik (x), fijk (x), ...) transforming like
the section jq (f ) : (x) → (x, f k (x), ∂i f k (x), ∂ij f k (x), ...) when f is an arbitrary section of E.
Then both fq ∈ Jq (E) and jq (f ) ∈ Jq (E) are over f ∈ E and the Spencer operator just allows
to distinguish them by introducing a kind of ”difference” through the operator D : Jq+1 (E) →
T ∗ ⊗ Jq (E) : fq+1 → j1 (fq ) − fq+1 with local components (∂i f k (x) − fik (x), ∂i fjk (x) − fijk (x), ...)
k
(x). In a symbolic way, when changes of
and more generally (Dfq+1 )kµ,i (x) = ∂i fµk (x) − fµ+1
i
coordinates are not involved, it is sometimes useful to write down the components of D in the
form di = ∂i − δi and the restriction of D to the kernel Sq+1 T ∗ ⊗ E of the canonical projection
πqq+1 : Jq+1 (E) → Jq (E) is minus the Spencer map δ = dxi ∧ δi : Sq+1 T ∗ ⊗ E → T ∗ ⊗ Sq T ∗ ⊗ E.
The kernel of D is made by sections such that fq+1 = j1 (fq ) = j2 (fq−1 ) = ... = jq+1 (f ). Finally, if Rq ⊂ Jq (E) is a system of order q on E locally defined by linear equations Φτ (x, yq ) ≡
aτk µ (x)yµk = 0 and local coordinates (x, z) for the parametric jets up to order q, the r-prolongation
Rq+r = ρr (Rq ) = Jr (Rq ) ∩ Jq+r (E) ⊂ Jr (Jq (E)) is locally defined when r = 1 by the link
+ ∂i aτk µ (x)yµk = 0 and has symbol
ear equations Φτ (x, yq ) = 0, di Φτ (x, yq+1 ) ≡ aτk µ (x)yµ+1
i
gq+r = Rq+r ∩ Sq+r T ∗ ⊗ E ⊂ Jq+r (E) if one looks at the top order terms. If fq+1 ∈ Rq+1 is
over fq ∈ Rq , differentiating the identity aτk µ (x)fµk (x) ≡ 0 with respect to xi and substracting the
k
k
(x)) ≡ 0
(x)+∂i aτk µ (x)fµk (x) ≡ 0, we obtain the identity aτk µ (x)(∂i fµk (x)−fµ+1
identity aτk µ (x)fµ+1
i
i
∗
and thus the restriction D : Rq+1 → T ⊗ Rq ([17],[18],[30]).
q+r+1
DEFINITION 2.2: Rq is said to be formally integrable when the restriction πq+r
: Rq+r+1 →
Rq+r is an epimorphism ∀r ≥ 0 or, equivalently, when all the equations of order q + r are obtained
by r prolongations only ∀r ≥ 0. In that case, Rq+1 ⊂ J1 (Rq ) is a canonical equivalent formally
integrable first order system on Rq with no zero order equations, called the Spencer form.
Finding an intrinsic test has been achieved by D.C. Spencer in 1970 ([30]) along coordinate
dependent lines sketched by M. Janet in 1920 ([7]) and W. Gröbner in 1940 ([4],[6]). The key
ingredient, missing in the old approach, is provided by the following definition.
Let T ∗ be the cotangent vector bundle of 1-forms on X and ∧s T ∗ be the vector bundle of s-forms
on X with usual bases {dxI = dxi1 ∧ ... ∧ dxis } where we have set I = (i1 < ... < is ). Moreover, introducing the exterior derivative d : ∧s T ∗ −→ ∧s+1 T ∗ : ω = ωI (x)dxI −→ dω = ∂i ωI (x)dxi ∧ dxI ,
we have d2 = d ◦ d = 0 and may introduce the Poincaré sequence:
4
d
d
d
d
∧0 T ∗ −→ ∧1 T ∗ −→ ∧2 T ∗ −→ ... −→ ∧n T ∗ −→ 0
PROPOSITION 2.3: There exists a map δ : ∧s T ∗ ⊗ Sq+1 T ∗ ⊗ E → ∧s+1 T ∗ ⊗ Sq T ∗ ⊗ E which
restricts to δ : ∧s T ∗ ⊗ gq+1 → ∧s+1 T ∗ ⊗ gq and δ 2 = δ ◦ δ = 0.
k
k
.
Proof: Let us introduce the family of s-forms ω = {ωµk = vµ,i
dxI } and set (δω)kµ = dxi ∧ ωµ+1
i
2
k
i
j
k
We obtain at once (δ ω)µ = dx ∧ dx ∧ ωµ+1i +1j = 0.
Q.E.D.
The kernel of each δ in the first case is equal to the image of the preceding δ but this may no
longer be true in the restricted case and we set:
s
DEFINITION 2.4: We denote by Hq+r
(gq ) the cohomology at ∧s T ∗ ⊗ gq+r of the restricted
1
s
δ-sequence which only depends on gq . The symbol gq is said to be s-acyclic if Hq+r
= ... = Hq+r
=
0, ∀r ≥ 0, involutive if it is n-acyclic and finite type if gq+r = 0 becomes trivially involutive for r
large enough.
DEFINITION 2.5: Rq is said to be involutive when it is formally integrable and its symbol gq
δ
δ
is involutive, that is to say all the sequences ... → ∧s T ∗ ⊗ gq+r → ... are exact ∀0 ≤ s ≤ n, ∀r ≥ 0.
Equivalently, the following procedure, where one may have to change linearly the independent
variables if necessary, is the heart towards the next effective definition of involution. It is intrinsic
even though it must be checked in a particular coordinate system called δ-regular ([17],[18],[29])
and is particularly simple for first order systems without zero order equations.
• Equations of class n: Solve the maximum number βqn of equations with respect to the jets of
order q and class n. Then call (x1 , ..., xn ) multiplicative variables.
− − − − − − − − − − − − − − −−
• Equations of class i ≥ 1: Solve the maximum number βqi of remaining equations with respect
to the jets of order q and class i. Then call (x1 , ..., xi ) multiplicative variables and (xi+1 , ..., xn )
non-multiplicative variables.
−−−−−−−−−−−−−−−−−
• Remaining equations equations of order ≤ q − 1: Call (x1 , ..., xn ) non-multiplicative variables.
In actual practice, we shall use a multiplicative board where the multiplicative ”variables” are represented by their index in upper left position while the non-multiplicative variables are represented
by dots in lower right position.
DEFINITION 2.6: A system of PD equations is said to be involutive if its first prolongation
can be achieved by prolonging its equations only with respect to the corresponding multiplicative
(q+n−i−1)!
−βqi for i = 1, ..., n and
variables. In that case, we may introduce the characters αiq = m (q−1)!((n−i)!
we have dim(gq+1 ) = α1q + ... + αnq . Moreover, one can exhibit the Hilbert polynomial dim(Rq+r )
in r with leading term (α/d!)rd with d ≤ n when α is the smallest non-zero character in the case
of an involutive symbol. Such a prolongation allows to compute in a unique way the principal
(pri) jets from the parametric (par) other ones. This definition may also be applied to nonlinear
systems as well.
REMARK 2.7: For an involutive system with β = βqn < m, then (y β+1 , ..., y m ) can be given
arbitrarily and may constitute the input variables in control theory, though it is not necessary
to make such a choice. In this case, the intrinsic number α = αnq = m − β > 0 is called the
n-character and is the system counterpart of the so-called ”differential transcendence degree” in
differential algebra. As we shall see in the next section, the smallest non-zero character and the
number of zero characters are intrinsic numbers that cannot be known without bringing the system
5
to involution and we have α1q ≥ ... ≥ αnq ≥ 0.
EXAMPLE 2.8: ([11], §38, p 40 where one can find the first intuition of formal integrability) The primary ideal q = ((χ1 )2 , χ1 χ3 − χ2 ) provides the system y11 = 0, y13 − y2 = 0 which
is neither formally integrable nor involutive. Indeed, we get d3 y11 − d1 (y13 − y2 ) = y12 and
d3 y12 − d2 (y13 − y2 ) = y22 , that is to say each first and second prolongation does bring a new second
order PD equation. Considering the new system y22 = 0, y12 = 0, y13 −y2 = 0, y11 = 0, the question
is to decide whether this system is involutive or not. One could use Janet or Gröbner algorithm but
with no insight towards involution. In such a simple situation, as there is no PD equation of class 3,
two evident permutations of coordinates (1, 2, 3) → (3, 2, 1) or (1, 2, 3) → (2, 3, 1) both provide one
equation of class 3, 2 equations of class 2 and 1 equation of clas 1. It is then easy to check directly
that the first permutation brings the involutive system y33 = 0, y23 = 0, y22 = 0, y13 − y2 = 0 that
will be used in the sequel and we have α32 = 0, α22 = 0, α12 = 2.
EXAMPLE 2.9: With n = 4, m = 1, q = 1, K = Q(x1 , x2 , x3 , x4 ), let us consider the system R1 :
{y4 − x3 y2 − y = 0,
y3 − x4 y1 = 0
Again, the reader will check easily that the subsystem R1′ ⊂ R1 :
u≡
v≡
w≡
y4 − x3 y1 − y = 0
y3 − x4 y1 = 0
y2 − y1 = 0
1 2
1 2
1 2
3 4
3 •
• •
(1)
namely the projection R1 of R2 to R1 , is formally integrable and even involutive with one equation of class 4, one equation of class 3 and one equation of class 2.
In the situation of the last remark, the following theorem will generalizing for PD control systems the well known first order Kalman form of OD control systems where the derivatives of the
input do not appear ([27], VI,1.14, p 802). For this, we just need to modify the Spencer form and
we provide the procedure that must be followed in the case of a first order involutive system with
no zero order equation, for example an involutive Spencer form.
• Look at the equations of class n solved with respect to yn1 , ..., ynβ .
• Use integrations by part like:
yn1 − a(x)ynβ+1 = dn (y 1 − a(x)y β+1 ) + ∂n a(x)y β+1 = ȳn1 + ∂n a(x)y β+1
• Modify y 1 , ..., y β to ȳ 1 , ..., ȳ β in order to ”absorb” the various ynβ+1 , ..., ynm only appearing in the
equations of class n.
We have the following unexpected result providing what we shall call reduced Spencer form:
THEOREM 2.10: The new equations of class n only contain yiβ+1 , ..., yim with 0 ≤ i ≤ n − 1
while the equations of class 1, ..., n − 1 no longer contain y β+1 , ..., y m and their jets. Accordingly,
as we shall see in the next section, any torsion element, if it exists, only depends on ȳ 1 , ..., ȳ β .
Proof: The first assertion comes from the absorption procedure. Now, if y m or yim should appear in an equation of class ≤ n−1, prolonging this equation with respect to the non-multiplicative
m
variable xn should bring ynm or yin
and (here involution is essential) we should get a linear combination of equations of various classes prolonged with respect to x1 , ..., xn−1 only, but this is
impossible and we get the desired reduced form.
Q.E.D.
jq
Φ
When Rq is involutive, the linear differential operator D : E → Jq (E) → Jq (E)/Rq = F0 of order
q with space of solutions Θ ⊂ E is said to be involutive and one has the canonical linear Janet
sequence ([17], p 144):
D
D
D
D
n
2
1
Fn −→ 0
... −→
F1 −→
0 −→ Θ −→ T −→ F0 −→
6
where each other operator is first order involutive and generates the compatibility conditions (CC)
of the preceding one. As the Janet sequence can be cut at any place, the numbering of the Janet
bundles has nothing to do with that of the Poincaré sequence for the exterior derivative, contrary
to what many physicists believe. Moreover, the dimensions of the Janet bundles can be computed
at once inductively from the board of multiplicative and non-multiplicative variables that can be
exhibited for D by working out the board for D1 and so on. For this, the number of rows of this
new board is the number of dots appearing in the initial board while the number nb(i) of dots in
the column i just indicates the number of CC of class i for i = 1, ..., n with nb(i) < nb(j), ∀i < j.
It follows that the successive first order operators D1 , ..., Dn are automatically in reduced Spencer
form.
EXAMPLE 2.11: Coming back to Example 2.9 and changing slightly our usual notations, we
get for D1 the following first order involutive system of CC in reduced Spencer form:
3
φ ≡
φ2 ≡
1
φ ≡
d4 v − d3 u + x4 d1 u − x3 d1 v − v = 0
d4 w − d2 u + d1 u − x3 d1 w − w = 0
d3 w − d2 v + d1 v − x4 d1 w = 0
1 2
1 2
1 2
3 4
3 4
3 •
as d4 u does not appear in φ2 and φ3 while u does not appear in φ1 .
We finally obtain for D2 the only CC:
ψ ≡ d4 φ1 − d3 φ2 − d1 φ3 + x4 d1 φ2 − x3 d1 φ1 − φ1 = 0
DEFINITION 2.12: The Janet sequence is said to be locally exact at Fr if any local section of
Fr killed by Dr+1 is the image by Dr of a local section of Fr−1 . It is called locally exact if it is
locally exact at each Fr for 0 ≤ r ≤ n. The Poincaré sequence is locally exact, that is a closed
form is locally an exact form but counterexamples may exist ([18], p 373).
j1
Equivalently, we have the involutive first Spencer operator D1 : C0 = Rq → J1 (Rq ) →
J1 (Rq )/Rq+1 ≃ T ∗ ⊗ Rq /δ(gq+1 ) = C1 of order one induced by D : Rq+1 → T ∗ ⊗ Rq . Introducing
the Spencer bundles Cr = ∧r T ∗ ⊗ Rq /δ(∧r−1 T ∗ ⊗ gq+1 ), the first order involutive (r + 1)-Spencer
operator Dr+1 : Cr → Cr+1 is induced by D : ∧r T ∗ ⊗ Rq+1 → ∧r+1 T ∗ ⊗ Rq : α ⊗ ξq+1 →
dα ⊗ ξq + (−1)r α ∧ Dξq+1 and we obtain the canonical linear Spencer sequence ([17], p 150):
jq
D
D
D
D
n
3
2
1
Cn −→ 0
... −→
C2 −→
0 −→ Θ −→ C0 −→
C1 −→
as the canonical Janet sequence for the first order involutive system Rq+1 ⊂ J1 (Rq ).
The canonical Janet sequence and the canonical Spencer sequence can be connected by a
commutative diagram where the Spencer sequence is induced by the locally exact central horizontal sequence which is at the same time the Janet sequence for jq and the Spencer sequence for
Jq+1 (E) ⊂ J1 (Jq (E)) ([17], p 153) but this result will not be used in this paper (See [5],[20],[22],[23]
for more details on Cosserat and Maxwell equations, see ([16]-[21]) and in particular ([22],[23]) for
applications to engineering and mathematical physics).
REMARK 2.13: We shall revisit Example 2.8 in order to explain the word ”canonical ” that
has been used in the previous definitions. For this, starting with the inhomogeneous system
y33 = u, y13 − y2 = v, we obtain easily the following inhomogeneous involutive system with its
corresponding board of multiplicative and non-multiplicative variables:
4
Φ
3
Φ
Φ2
1
Φ
≡
≡
≡
≡
y33
y23
y22
y13 − y2
=
=
=
=
u
d1 u − d3 v
d11 u − d13 v − d2 v
v
1
1
1
1
2
2
2
•
3
•
•
•
Using prolongation with respect to the 4 non-multiplicative variables involved should bring 4 first
order CC for the right members and we could wait for 4 third order CC involving u and v.
7
Surprisingly, we need the only CC Ψ ≡ d33 v − d13 u + d2 u = 0 and obtain the differential sequence:
0 −→ Θ −→ 1 −→ 2 −→ 1 −→ 0
as a single CC has no CC for itself (See ([18],p365) for the effective general procedure).
Such a differential sequence is quite different from the canonical Janet sequence:
D
D
D
1
2
0 −→ Θ −→ 1 −→ 4 −→
4 −→
1 −→ 0
which is the only sequence that can provide the Spencer sequence as we already said and could not
be obtained by simply using Gröbner bases. This remark will become essential in mathematical
physics (foundations of continuum mechanics, gauge theory, general relativity) where only involutive operators must be used ([20],[22],[23]). We also check that the Euler-Poincaré characteristic,
namely the alternate sum of the circled dimensions of the vector bundles involved, does not depend
on the differential sequence used as we get 1 − 2 + 1 = 1 − 4 + 4 − 1 = 0 (See [18], p 378).
In the same spirit, using certain parametric jet variables as new unknowns, we may set z 1 = y, z 2 =
y1 , z 3 = y2 , z 4 = y3 in order to obtain the following involutive first order system with no zero order
equation:
class 3
class 2
class 1
d3 z 1 − z 4 = 0, d3 z 2 − z 3 = 0, d3 z 3 = 0, d3 z 4 = 0
d2 z 1 − z 3 = 0, d2 z 2 − d1 z 3 = 0, d2 z 3 = 0, d2 z 4 = 0
d1 z 1 − z 2 = 0, d1 z 4 − z 3 = 0
1 2
1 2
1 •
3
•
•
where we have separated the classes. Contrary to what could be believed, this operator does not
describe the Spencer sequence that could be obtained from the previous Janet sequence. Indeed, introducing the trivial vector bundle E with local coordinates (x1 , x2 , x3 , y), it follows that J1 (E) has
local coordinates (x1 , x2 , x3 , z 1 , z 2 , z 3 , z 4 ). Now, the involutive system R2 ⊂ J2 (E) ⊂ J1 (J1 (E))
with involutive symbol g2 ⊂ S2 T ∗ ⊗ E ⊂ T ∗ ⊗ T ∗ ⊗ E ⊂ T ∗ ⊗ J1 (E) projects onto J1 (E)
but dim(R2 ) = 6 because par(R2 ) = {y, y1 , y2 , y3 , y11 , y12 } while we have only 4 unknowns
(z 1 , z 2 , z 3 , z 4 ). Nevertheless, as R2 projects onto J1 (E), we may construct a canonical Janet
sequence for this new system where the successive Janet bundles involved will be the Spencer
bundles Cr = ∧r T ∗ ⊗ J1 (E)/δ(∧r−1 T ∗ ⊗ g2 ) with a shift by one step in the numbering of the
bundles as now C0 = J1 (E) and the successive operators are induced by the composition of the
inclusion R2 ⊂ J2 (E) with the Spencer operator D : J2 (E) −→ T ∗ ⊗ J1 (E) as in ([17],p144,150) or
([18],p356). In any case, it is essential to notice that, both in the canonical Spencer sequence and
in the canonical Janet sequence, any intermediate operator can be constructed explicitely without
knowing the previous ones.
EXAMPLE 2.14: With m = 1, n = 4, q = 2, one could treat similarly the involutive system:
y44 = 0, y34 = 0, y33 = 0, y24 − y13 = 0 with one equation of class 4, two equations of class 3 and
one equation of class 2.
EXAMPLE 2.15: Coming back to the involutive system of Example 2.9 with variable coefficients, we let the reader prove that the Janet sequence is:
D
D
D
1
2
0 −→ Θ −→ 1 −→ 3 −→
3 −→
1 −→ 0
EXAMPLE 2.16: Let us finally consider the following involutive system of PD equations with
two independent variables (x1 , x2 ) and three unknowns (y 1 , y 2 , y 3 ), where again a is an arbitrary
constant parameter and we have set for simplicity yik = di y k :
2
y2 + y12 + y23 − y13 − ay 3
y 1 − y12 − y23 − y13 − ay 3
21
y1 − y12 − 2y13
=0
=0
=0
1 2
1 2
1 •
Then the corresponding Janet sequence is:
D
D
1
1 −→ 0
0 −→ Θ −→ 3 −→ 3 −→
8
Finally, setting ȳ 1 = y 1 − y 3 , ȳ 2 = y 2 + y 3 , we obtain the new first order involutive system:
2
ȳ2 − ȳ12 − ay 3
ȳ 1 − ȳ12 − ay 3
21
ȳ1 − ȳ12 = 0
=0
=0
1
1
1
2
2
•
with two equations of class 2 and one equation of class 1 in which y 3 surprisingly no longer appears.
If χ1 , ..., χn are n algebraic indeterminates or, in a more intrinsic way, if χ = χi dxi ∈ T ∗ is
a covector and D : E −→ F : ξ −→ aτk µ (x)∂µ ξ k (x) is a linear involutive operator of order q,
we may introduce the characteristic matrix a(x, χ) = (aτk µ (x)χµ | µ = q) and the resulting map
σχ (D) : E −→ F is called the symbol of D at χ. Then there are two possibilities:
• If maxχ rk(σχ (D) < m ⇔ αnq > 0: the characteristic matrix fails to be injective for any covector.
• If maxχ rk(σχ (D) = m ⇔ αnq = 0: the characteristic matrix fails to be injective if and only if
all the determinants of the m × m submatrices vanish. However, one can prove that this algebraic
ideal a ∈ K[χ] is not intrinsically defined and must be replaced by its radical rad(a) made by all
polynomials having a power in a. This radical ideal is called the characteristic ideal of the operator.
DEFINITION 2.17: For each x ∈ X, the algebraic set defined by the characteristic ideal is
called the characteristic set of D at x and V = ∪x∈X Vx is called the characteristic set of D.
One has the following important theorem ([18],[29]) that will play an important part later on:
THEOREM 2.18: (Hilbert-Serre) The dimension d(V ) of the characteristic set, that is the maximum dimension of the irreducible components, is equal to the number of non-zero characters while
the codimension cd(V ) = n − d(V ) is equal to the number of zero characters, that is to the number
of ”full ” classes in the board of multiplicative variables of an involutive system.
EXAMPLE 2.19: Coming back to Remark 2.12, we obtain a = ((χ3 )2 , χ2 χ3 , (χ2 )2 , χ1 χ3 ) =⇒
rad(a) = (χ2 , χ3 ) and thus cd(V ) = 2. However, if we take only into account Example 2.8, we
should only get the radical ideal (χ3 ) and the wrong result cd(V ) = 1. The reason for using the
radical can be seen from the equivalent first order system that shoul provide b = ((χ3 )4 , ...) with homogeneous polynomials of degree 4 and thus b ⊂ a with a strict inclusion though rad(a) = rad(b).
A similar situation can be obtained with Examples 1.1 and 2.9.
3) TOOLS FROM MODULE THEORY :
We may roughly say that, if a reader familiar with Gröbner bases ([4],[6]) and computer algebra
looks at the previous section, he will feel embarassed because he will believe that ”intrinsicness
is always competing with complexity ” as can be seen from Examples 2.8 + 2.12. However, even
if he admits that it may be useful to have intrinsic and thus canonical procedures, then looking
at the existing literature on differential modules ([1],[9],12]), he will really feel to be on another
planet as the main difficulty involved in the theory of differentia modules is to understand why and
where formal integrability and involution become essential tools to apply quite before dealing with
the homological background of ”algebraic analysis ” involving extension modules. This is the main
reason for which the case of variable coefficients is rarely treated ”by itself ” always refering to Weyl
algebras for examples and the main difficulty we found when writing ([18], in particular Chapter
IV). The central concept, essential for applications but well hidden in the literature dealing with
filtred modules ([14],p 383) and totally absent from the use of Gröbner bases because it amounts
to formal integrability by duality, is that of a ”strict morphism ”. Accordingly, the purpose of this
section will be to explain why such a definition, which seems to be purely technical, will be so
important for studying extension modules and purity.
If P = aµ dµ ∈ D = K[d], the highest value of |µ| with aµ 6= 0 is called the order of the operator
P and the ring D with multiplication (P, Q) −→ P ◦ Q = P Q is filtred by the order q of the
operators. We have the filtration 0 ⊂ K = D0 ⊂ D1 ⊂ ... ⊂ Dq ⊂ ... ⊂ D∞ = D. Moreover, it
is clear that D, as an algebra, is generated by K = D0 and T = D1 /D0 with D1 = K ⊕ T if we
9
identify an element ξ = ξ i di ∈ T with the vector field ξ = ξ i (x)∂i of differential geometry, but
with ξ i ∈ K now. It follows that D = D DD is a bimodule over itself, being at the same time a left
D-module by the composition P −→ QP and a right D-module by the composition P −→ P Q.
We define the adjoint functor ad : D −→ Dop : P = aµ dµ −→ ad(P ) = (−1)|µ| dµ aµ and we have
ad(ad(P )) = P . It is easy to check that ad(P Q) = ad(Q)ad(P ), ∀P, Q ∈ D. Such a definition can
also be extended to any matrix of operators by using the transposed matrix of adjoint operators
(See [18],[19],[22] for more details and applications to control theory and mathematical physics).
Accordingly, if y = (y 1 , ..., y m ) are differential indeterminates, then D acts on y k by setting
k
and y0k = y k . We may therefore use he jet coordidi y = yik −→ dµ y k = yµk with di yµk = yµ+1
i
nates in a formal way as in the previous section. Therefore, if a system of OD/PD equations is
written in the form Φτ ≡ aτk µ yµk = 0 with coefficients a ∈ K, we may introduce the free differential
module Dy = Dy 1 + ... + Dy m ≃ Dm and consider the differential submodule I = DΦ ⊂ Dy
which is usually called the module of equations, both with the differential module M = Dy/DΦ
or D-module and we may set M = D M if we want to specify the ring of differential operators.
The work of Macaulay only covers the case m = 1 with K replaced by k ⊆ cst(K). Again, we
k
+ ∂i aτk µ yµk
may introduce the formal prolongation with respect to di by setting di Φτ ≡ aτk µ yµ+1
i
k
k
in order to induce maps di : M −→ M : ȳµ −→ ȳµ+1i by residue if we use to denote the residue
Dy −→ M : y k −→ ȳ k by a bar as in algebraic geometry. However, for simplicity, we shall not
write down the bar when the background will indicate clearly if we are in Dy or in M .
k
As a byproduct, the differential modules we shall consider will always be finitely generated
(k = 1, ..., m < ∞) and finitely presented (τ = 1, ..., p < ∞). Equivalently, introducing the
matrix of operators D = (aτk µ dµ ) with m columns and p rows, we may introduce the morphism
D
Dp −→ Dm : (Pτ ) −→ (Pτ Φτ ) : P −→ P Φ = P D over D by acting with D on the left of these
row vectors while acting with D on the right of these row vectors and the presentation of M is
defined by the exact cokernel sequence Dp −→ Dm −→ M −→ 0. It is essential to notice that
the presentation only depends on K, D and Φ or D, that is to say never refers to the concept of
(explicit or formal) solutions. It is at this moment that we have to take into account the results
of the previous section in order to understant that certain presentations will be much better than
others, in particular to establish a link with formal integrability and involution.
It follows from its definition that M can be endowed with a quotient filtration obtained from
that of Dm which is defined by the order of the jet coordinates yq in Dq y. We have therefore the
inductive limit 0 ⊆ M0 ⊆ M1 ⊆ ... ⊆ Mq ⊆ ... ⊆ M∞ = M with di Mq ⊆ Mq+1 and M = DMq for
q ≫ 0 with prolongations Dr Mq ⊆ Mq+r , ∀q, r ≥ 0.
DEFINITION 3.1: ([14],p 383) If M and N are two differential modules and f : M −→ N is
a morphism over D compatible with the filtration, that is if f (Mq ) ⊂ Nq with induced morphism
fq : Mq −→ Nq , then f is a strict morphism if fq (Mq ) = f (M ) ∩ Nq , ∀q ≥ 0.
Equivalently, chasing in the following diagram:
0
↓
0
↓
fq
Mq
↓
−→ Nq
↓
M
−→
f
N
0
↓
−→ coker(fq )
↓
−→ 0
−→
−→ 0
coker(f )
then f is strict if the induced morphism coker(fq ) −→ coker(f ) is a monomorphism ∀q ≥ 0.
DEFINITION 3.2: An exact sequence of morphisms finishing at M is said to be a resolution of
M . If the differential modules involved apart from M are free, we shall say that we have a free
resolution of M . Moreover, a sequence of strict morphisms is called a strict sequence.
LEMMA 3.3: If f is a strict morphism as in the last definition, there are exact sequences
10
0 −→ coker(fq ) −→ coker(fq+1 ), ∀q ≥ 0.
Proof: As f is compatible with the filtrations and Mq ⊆ Mq+1 , Nq ⊆ Nq+1 , we have an induced
morphism coker(fq ) −→ coker(fq+1 ). Now, as f is also strict, we have the following commutative
and exact diagram:
0 −→
coker(fq )
−→ coker(f )
↓
k
0 −→ coker(fq+1 ) −→ coker(f )
The lemma finally follows from an elementary chase.
Q.E.D.
Having in mind that K is a left D-module with the standard action (D, K) −→ K : (di , a) −→
∂i a and that D is a bimodule over itself, we have only two possible constructions:
DEFINITION 3.4: We define the system R = homK (M, K) = M ∗ and set Rq = homK (Mq , K) =
Mq∗ as the system of order q. We have the projective limit R = R∞ −→ ... −→ Rq −→ ... −→
R1 −→ R0 . It follows that fq ∈ Rq : yµk −→ fµk ∈ K with aτk µ fµk = 0 defines a section at order
q and we may set f∞ = f ∈ R for a section of R. For an arbitrary differential field K, such a
definition has nothing to do with the concept of a formal power series solution (care).
DEFINITION 3.5: We may define the right differential module homD (M, D).
PROPOSITION 3.6: When M is a left D-module, then R is also a left D-module.
Proof: As D is generated by K and T as we already said, let us define:
(af )(m) = af (m),
∀a ∈ K, ∀m ∈ M
(ξf )(m) = ξf (m) − f (ξm),
∀ξ = ai di ∈ T, ∀m ∈ M
In the operator sense, it is easy to check that di a = adi + ∂i a and that ξη − ηξ = [ξ, η] is the
k
and thus
standard bracket of vector fields. We finally get (di f )kµ = (di f )(yµk ) = ∂i fµk − fµ+1
i
recover exactly the Spencer operator of the previous section though this is not evident at all. We
k
k
k
=⇒ di dj = dj di , ∀i, j = 1, ..., n and thus
+ fµ+1
− ∂j fµ+1
also get (di dj f )kµ = ∂ij fµk − ∂i fµ+1
i +1j
i
j
di Rq+1 ⊆ Rq =⇒ di R ⊂ R induces a well defined operator R −→ T ∗ ⊗ R : f −→ dxi ⊗ di f . This
result has been discovered (up to sign) by Macaulay in 1916 ([11]). For more details on the Spencer
operator and its applications, the reader may look at ([22],[23]).
Q.E.D.
As D is a bimodule over itself, it follows from this proposition that that D∗ = homK (D, K)
is a left D-module. Moreover, using Baer’s criterion ([28]), it is known that D∗ is an injective
D-module as there is a canonical isomorphism:
M ∗ = homK (M, K) ≃ homD (M, D∗ )
where both sides are well defined ([2], Prop 11, p 18;)([28], p 37).
DEFINITION 3.7: With any differential module M we shall associate the graded module
G = gr(M ) over the polynomial ring gr(D) ≃ K[χ] by setting G = ⊕∞
q=0 Gq with Gq = Mq /Mq+1
and we get gq = G∗q where the symbol gq is defined by the short exact sequences:
0 −→ Mq−1 −→ Mq −→ Gq −→ 0
=⇒
0 −→ gq −→ Rq −→ Rq−1 −→ 0
We have the short exact sequences 0 −→ Dq−1 −→ Dq −→ Sq T −→ 0 leading to grq (D) ≃ Sq T
and we may set as usual T ∗ = homK (T, K) in a coherent way with differential geometry. Moreover
any compatible morphism f : M −→ N induces a morphism gr(f ) : gr(M ) −→ gr(N ).
11
EXAMPLE 3.8: If K = Q(x), m = 1, n = 1, let us consider the system yxxx − yx = 0 for which
we may exhibit the basis of sections {f = (1, 0, 0, 0, ...), f ′ = (0, 1, 0, 1, ...), f ” = (0, 0, 1, 0, ...)} as in
([11],§59,p 67) or ([21]). We obtain dx f = 0, dx f ′ = −f −f ”, dx f ” = −f ′ and check that all the sections can be generated by a single one, namely f ” which describes the power series of ch(x)−1. With
1
now m = 2, let us consider the module defined by the system yxx
= 0, yx2 = 0. Setting y = y 2 −xy 1 ,
2
1
1
1
1
we successively get yx = −xyx − y , yxx = −2yx, yxxx = 0 =⇒ y = x2 yxx − yx , y 2 = x2 yxx − xyx + y
and a differential isomorphism with the module defined by the new system yxxx = 0. All the sections of the second system are easily seen to be generated by the single section f = (0, 0, 1, ...), a
2
1
= 0, ..., f 2 (x) = x2 , fx2 = 0, ...
result leading to the only generating section f 1 (x) = x2 , fx1 = − 21 , fxx
of the initial system but these sections do not describe solutions because ∂x f 1 − fx1 = 1 6= 0 and
∂x f 2 − fx2 = x 6= 0. We do not know any reference in computer algebra dealing with sections (See
[21] for more details)
Coming back to the presentation of M under study, we notice that the morphism D involved
is not compatible unless we shift the index of the filtration by ord(D) = q. In that case, we obtain
im(D) = I ⊂ Dy and may set Iq+r = I ∩ Dq+r y but we have in general Drp D ⊆ Iq+r only, that
is the equations of order q+r may not be produced by r prolongations only. We have thus obtained:
PROPOSITION 3.9: The morphism induced by D is strict if and only if D is formally integrable. Accordingly, the module version of both the Janet sequence and the Spencer sequence are
strictly exact sequences.
Proof: Using q + r instead of q in Lemma 3.3 and applying homK (•, K), we obtain the epimorphisms Rq+r+1 −→ Rq+r −→ 0, ∀r ≥ 0.
Q.E.D.
The reader will find in ([18], IV,3) more details on the relations existing between G and M
which are needed in order to study the non-commutative situation, at least when K is a differential field as such a case is hard enough. We obtain in particular the Hilbert polynomial
α d
dimK (Mq+r ) = dimK (Rq+r ) = d!
r + ... where α is called the multiplicity of M and we use to set
cdD (M ) = cd(M ) = n − r, rkD (M ) = rk(M ) = α if cd(M ) = 0 and 0 otherwise.
EXAMPLE 3.10: Coming back to the Example 2.8 of Macaulay, we obtain from Remark 2.13
the free resolution 0 −→ D −→ D2 −→ D −→ M −→ 0 but only the morphism on the left is strict
as for the morphism on the right we know that its image is indeed I2 = {y33 , y23 , y22 , y13 − y2 } and
not just {y33 , y13 −y2 }. However, bringing the system to involution, we get the strict free resolution
0 −→ D −→ D4 −→ D4 −→ D −→ M −→ 0 as the module version of the Janet sequence and we
let the reader exhibit the module version of the corresponding Spencer sequence as an exercise.
If M is a differential module over the ring D = K[d] of differential operators and m ∈ M , then
the differential submodule Dm ⊂ M is defined by a system of OD or PD equations for one unknown
and we may look for its codimension cd(Dm). A similar comment can be done for any differential
submodule M ′ ⊂ M . Sometimes, a single element m ∈ M , called differentially primitive element,
may generate M if Dm = M .
EXAMPLE 3.11: With K = Q, let us consider the differential moduleM defined by the system
1
1
yxx
− y 1 = 0, yx2 = 0. Then, setting y = y 1 − y 2 , we get yx = yx1 , yxx = yxx
= y 1 , yxx − y = y 2 and
thus yxxx − yx = 0 provides another way to describe M by means of a single element as in ([18],
p435). We have the following commutative and exact diagram:
0 −→
0 −→
D2
↓
D
D2
↓
D
−→
−→
(P, Q)
−→
(P (d2 − 1), Qd)
↓
↓
(P d + Q) −→ ((P d + Q)(d3 − d))
12
−→ M
k
−→ M
−→ 0
−→ 0
1
If now we consider the differential module M defined by yxx
− ay 1 = 0, yx2 = 0 where a is a constant
parameter, we cannot find a differentially primitive element when K = Q if a = 0 but we can when
K = Q(x) for any value of a, as in Example 3.8.
We may check the following definition in a constructive way ([27]):
DEFINITION 3.12: tr (M ) = {m ∈ M | cd(Dm) > r} is the greatest differential submodule of
M having codimension > r.
PROPOSITION 3.13: cd(M ) = cd(V ) = r ⇐⇒ αqn−r 6= 0, αn−r+1
= ... = αnq = 0 ⇐⇒ tr (M ) 6=
q
M, tr−1 (M ) = ... = t0 (M ) = t(M ) = M and this intrinsic result can be most easily checked by
using the Spencer form of the system defining M .
We are now in a good position for defining and studying purity for differential modules.
DEFINITION 3.14: M is r-pure ⇐⇒ tr (M ) = 0, tr−1 (M ) = M ⇐⇒ cd(Dm) = r, ∀m ∈ M . In
particular, M is 0-pure if t(M ) = 0 and, if cd(M ) = r but M is not r-pure, we may call M/tr (M )
the pure part of M . It follows that tr−1 (M )/tr (M ) is equal to zero or is r-pure (See the picture in
[18], p 545). Finally, when tr−1 (M ) = tr (M ), we shall say that there is a gap in the purity filtration:
0 = tn (M ) ⊆ tn−1 (M ) ⊆ ... ⊆ t1 (M ) ⊆ t0 (M ) = t(M ) ⊆ M
PROPOSITION 3.15: tr (M ) does not depend on the presentation or on the filtration of M .
EXAMPLE 3.16: If K = Q and M is defined by the involutive system y33 = 0, y23 = 0, y13 = 0,
then z = y3 satifies d3 z = 0, d2 z = 0, d1 z = 0 and cd(Dz) = 3 while z ′ = y2 only satisfies d3 z ′ = 0
and cd(Dz ′ ) = 1. We have the purity filtration 0 = t3 (M ) ⊂ t2 (M ) = t1 (M ) ⊂ t0 (M ) = t(M ) = M
with one gap and two strict inclusions.
We now recall the definition of the extension modules extiD (M, D) that we shall simply denote
by exti (M ) and the way to use their dimension or codimension. We point out once more that
these numbers cannot be obtained without bringing the underlying systems to involution in order
to get informations on M from informations on G. We divide the procedure into four steps that
can be achieved by means of computer algebra ([27]):
• Construct a free resolution of M , say:
... −→ Fi −→ ... −→ F1 −→ F0 −→ M −→ 0
• Suppress M in order to obtain the deleted sequence:
... −→ Fi −→ ... −→ F1 −→ F0 −→ 0
• Apply homD (•, D) in order to obtain the dual sequence heading backwards:
... ←− homD (Fi , D) ←− ... ←− homD (F1 , D) ←− homD (F0 , D) ←− 0
• Define exti (M ) to be the cohomology at homD (Fi , D) in the dual sequence with ext0 (M ) =
homD (M, D).
The following nested chain of difficult propositions and theorems can be obtained, even in the
non-commutative case, by combining the use of extension modules and bidualizing complexes in the
framework of algebraic analysis. The main difficulty is to obtain first these results for the graded
module G = gr(M ) by using techniques from commutative algebra before extending them to the
filtred module M as in ([1],[9],[18],[19]).
THEOREM 3.17: The extension modules do not depend on the resolution of M used.
13
PROPOSITION 3.18: Applying homD (•, D) provides right D-modules that can be transformed
to left D-modules by means of the side changing functor and vice-versa. Namely, if ND is a right
D-module, then D N = ∧n T ⊗K ND is the converted left D-module while, if D N is a left D-module,
then ND = ∧n T ∗ ⊗K D N is the converted right D-module.
PROPOSITION 3.19: Instead of applying homD (•, D) and the side changing functor in the
module framework, we may use ad in the operator framework. Namely, to any operator D : E −→ F
we may associate the formal adjoint ad(D) : ∧n T ∗ ⊗ F ∗ −→ ∧n T ∗ ⊗ E ∗ with the useful though
striking relation rkD (ad(D)) = rkD (D).
PROPOSITION 3.20: exti (M ) is a torsion module ∀1 ≤ i ≤ n but ext0 (M ) = homD (M, D)
may not be a torsion module.
EXAMPLE 3.21: When M is a torsion module, we have homD (M, D) = 0 (exercise). When
n = 3 and the torsion-free module M is defined by the formally surjective div operator, the formal
adjoint of div is −grad which defines a torsion module. Also, when n = 1 as in classical control
theory, a controllable system allows to define a torsion-free module M which is free in that case
and homD (M, D) is thus also a free module.
THEOREM 3.22:
exti (M ) = 0, ∀i ≥ n + 1.
THEOREM 3.23:
cd(exti (M )) ≥ i.
PROPOSITION 3.24:
THEOREM 3.25:
exti (M ) = 0, ∀i < cd(M ).
cd(M ) ≥ r ⇔ exti (M ) = 0, ∀i < r.
PROPOSITION 3.26:
cd(M ) = r =⇒ cd(extr (M )) = r and extr (M )is r-pure.
PROPOSITION 3.27:
extr (extr (M )) is equal to 0 or is r-pure, ∀0 ≤ r ≤ n.
PROPOSITION 3.28: If we set t−1 (M ) = M , there are exact sequences ∀0 ≤ r ≤ n:
0 −→ tr (M ) −→ tr−1 (M ) −→ extr (extr (M ))
THEOREM 3.29: If cd(M ) = r, then M is r-pure if and only if there is a monomorphism
0 −→ M −→ extr (extr (M )) of left differential modules.
THEOREM 3.30: M is pure ⇐⇒ exts (exts (M )) = 0, ∀s 6= cd(M ).
The last two theorems are known to characterize purity but it is however evident that they are
not very useful in actual practice.
THEOREM 3.31: When M is r-pure, the characteristic ideal is thus unmixed, that is a finite
intersection of prime ideals having the same codimension r and the characteristic set is equidimensional, that is the union of irreducible algebraic varieties having the same codimension r.
REMARK 3.32: For the reader knowing more about commutative algebra, we add a few details about the localization used in the primary decomposition of a module which are not so well
known ([3],[18],[21],[31]). For simplicity, setting k = cst(K), we shall denote by A = k[χ] the
polynomial ring isomorphic to D = k[d] and consider a module M over A. We denote as usual
by spec(A) the set of proper prime ideals in A, by max(A) the subset of maximal ideals in A
and by ass(M ) = {p ∈ spec(A)|∃0 6= m ∈ M, p = annA (m)} the (finite) set {p1 , ..., pt } of
associated prime ideals, while we denote by {p1 , ...ps } the subset of minimum associated prime
ideals. It is well known that M 6= 0 =⇒ ass(M ) 6= ∅. We recall that an ideal q ⊂ A is pprimary if ab ∈ q, b ∈
/ q =⇒ a ∈ rad(q) = p ∈ spec(A). We say that a module Q is p-primary
if am = 0, 0 6= m ∈ Q =⇒ a ∈ p = rad(q) ∈ spec(A) when q = annA (Q) or, equivalently,
14
ass(Q) = {p}. Similarly, we say that a module P is p-prime if am = 0, 0 6= m ∈ P =⇒ a ∈
p ∈ spec(A) when p = annA (P ). It follows that any p-prime or p-primary module is r-pure with
n − r = trd(A/p), a result generalizing ([11],§4, p 43). Accordingly, a module M is r-pure if and
only if a = annA (M ) admits a primary decomposition a = q1 ∩...∩qs and rad(a) = p1 ∩...∩ps with
cd(A/pi ) = cd(M ) = r, ∀i = 1, ..., s. In that case, the monomorphism 0 −→ M −→ ⊕p∈ass(M) Mp
induces a monomorphism 0 −→ M −→ Q1 ⊕ ... ⊕ Qs called primary embedding where the primary
modules Qi are the images of the localization morphisms M −→ Mpi = S −1 M with S = A − p
inducing epimorphisms M −→ Qi −→ 0 for i = 1, ..., s. Macaulay was only considering the case
M = A/a with primary decomposition a = q1 ∩ ... ∩ qs .
EXAMPLE 3.33: With k = Q and n = 3, then a = rad(a) = (χ1 , χ2 χ3 ) = (χ1 , χ2 ) ∩ (χ1 , χ3 ) is
unmixed and M = A/a is 2-pure while a = rad(a) = (χ1 χ2 , χ1 χ3 ) = (χ1 )∩(χ2 , χ3 ) is mixed, though
an intersection of two minimum prime ideals and M = A/a is not 1-pure. On the contrary, if one
has the primary decomposition a = ((χ1 )2 , χ1 χ2 , χ1 χ3 , χ2 χ3 ) = (χ1 , χ2 ) ∩ (χ1 , χ3 ) ∩ (χ1 , χ2 , χ3 )2 =
p1 ∩ p2 ∩ m2 and M = A/a, then ass(M ) = {p1 , p2 , m} with pi ⊂ m for i = 1, 2, though
rad(a) = p1 ∩ p2 as before. In that case, there is an embedding 0 −→ M −→ Q1 ⊕ Q2 ⊕ Q3
where Qi = A/pi is the image of the localization morphism M −→ Mpi for i = 1, 2 because p1
is killed by χ3 ∈ A − p1 , p2 is killed by χ2 ∈ A − p2 and Q3 = A/m2 is m-primary because
rad(m2 ) = m ∈ max(A). We have also an embedding 0 −→ M −→ Mp1 ⊕ Mp2 ⊕ Mm but no
element of the multiplicative set A − m = {1 + a|a ∈ m} can kill any element of M and the image
of M into Mm is thus isomorphic to M which is not a primay module. It is important to notice
that the example of Macaulay q = ((χ3 )2 , χ2 χ3 , (χ2 )2 , χ1 χ3 − χ2 ) provides a p-primary module
A/q with p = (χ3 , χ2 ) even though the annihilating ideal of G = gr(M ) is the homogeneous
ideal a = ((χ3 )2 , χ2 χ3 , (χ2 )2 , χ1 χ3 ) = ((χ2 )2 , χ3 ) ∩ (χ1 , χ2 , χ3 )2 which is a mixed ideal because
ass(A/a) = {(χ2 , χ3 ), (χ1 , χ2 , χ3 )}. However, we get rad(a) = (χ2 , χ3 ) in a coherent way.
PROBLEM : Is it possible to have a test for checking whether a differential module is pure or
not without using the previous results ?
4) MOTIVATION :
As we already said in the introduction and in the previous section, a torsion-free module M
is 0-pure because in that case t0 (M ) = t(M ) = 0. Accordingly, M can be embedded into a free
module F and the inclusion, which may not be strict when n > 1, provides a parametrization by
means of a finite number of potential-like arbitrary functions in the classical language of elasticity (Airy function) or electromagnetism (EM 4-potential). As it is clear that such a situation is
only a very particular case of purity, it remains to wonder what can happen for an r-pure module
whenever r ≥ 1. One has the following result ([9], [18], compare to [1], p494):
THEOREM 4.1: If M is an r-pure differential module with r ≥ 1, there exists a differential
module L with pd(L) ≤ r and an embedding M ⊆ L.
Proof: First of all we notice that we have r > 0 and thus any element m ∈ M is surely a torsion element because cd(Dm) > 0, that is M = t(M ) is a torsion module with ext0 (M ) = homD (M, D) =
0 because D is an integral domain. Let now ... −→ Fr −→ ... −→ F1 −→ F0 −→ N −→ 0 be a free
resolution of the right differential module N = extr (M ). According to Proposition 3.26, we have
cd(N ) = r > 0 too and N is also a torsion module with ext0 (N ) = homD (N, D) = 0. Applying the
functor homD (•, D) to the previous sequence or , equivalently, constructing the adjoint sequence in
the operator framework while using the fact that exti (N ) = 0, ∀i < r according to Theorem 3.25,
we obtain the finite long exact sequence with exactly r morphisms because N is finitely presented
and extr (N ) 6= 0:
0 −→ homD (F0 , D) −→ ... −→ homD (Fr−1 , D) −→ homD (Fr , D) −→ L −→ 0
where the left differential module L is the cokernel of the last morphism on the right. As
homD (F, D) is free whenever F is free because of the bimodule structure of D = D DD , the
15
corresponding deleted complex is:
0 −→ homD (F0 , D) −→ ... −→ homD (Fr−1 , D) −→ homD (Fr , D) −→ 0
Applying again homD (•, D) and using the reflexivity of any free module F , that is the isomorphism
homD (homD (F, D), D) ≃ F , we obtain the dual sequence:
0 −→ Fr −→ ... −→ F1 −→ F0 −→ 0
and a similar procedure may be followed with operators as we shall see in the next illustrating
examples ([18],[27]). This sequence is exact everywhere but at Fr and at F0 where its cohomology
is just N by definition, that is to say extr (L) = N = extr (M ). Looking for the cohomology at
homD (Fr , D) in the sequence obtained by duality from the resolution of N with coboundry module
Br and cocycle module Zr , we obtain the following commutative and exact diagram:
0
0
↓
↓
0 −→
Br
=
Br
↓
↓
0 −→
Zr
−→ homD (Fr , D)
↓
↓
0 −→ extr (N ) −→
L
↓
↓
0
0
−→ 0
Finally, composing the bottom monomorphism with the monomorphism 0 −→ M −→ extr (N )
provided by Theorem 3.29, we get the desired embedding M ⊆ L. It must be noticed that such a
procedure can be followed equally well in the commutative and non-commutative framework, that
is when K is a field of constants or a true differential field.
Q.E.D.
EXAMPLE 4.2: With K = Q, m = 1, n = 4, q = 2, let us study the 2-pure differential module
M defined by the involutive system:
4
Φ
3
Φ
Φ2
1
Φ
≡
≡
≡
≡
y44
y34
y33
y24 − y13
=0
=0
=0
=0
1
1
1
1
2
2
2
2
3
3
3
•
4
•
•
•
From the board of multiplicative variables we may construct at once the Janet sequence:
D
D
D
1
2
4 −→
1 −→ 0
0 −→ Θ −→ 1 −→ 4 −→
where D1 is defined by the involutive system:
4
Ψ
3
Ψ
Ψ2
1
Ψ
≡
≡
≡
≡
d4 Φ3 − d3 Φ4
d4 Φ2 − d3Φ3
d4 Φ1 − d2 Φ4 + d1 Φ3
d3 Φ1 − d2 Φ3 + d1 Φ2
=0
=0
=0
=0
1
1
1
1
2
2
2
2
3
3
3
3
4
4
4
•
and D2 by the (trivially) involutive system:
Ω
≡
d4 Ψ1 − d3 Ψ2 + d2 Ψ4 − d1 Ψ3
=0
1 2
We have therefore the resolution:
0 −→ D −→ D4 −→ D4 −→ D −→ M −→ 0
16
3 4
leading to pd(M ) ≤ 3 and the deleted complex is:
0 −→ D −→ D4 −→ D4 −→ D −→ 0
Applying homD (•, D) to this sequence, we get the sequence:
0 ←− D ←− D4 ←− D4 ←− D ←− 0
which can be described by the following adjoint sequence:
ad(D)
←−
0 ←− 1
4
ad(D1 )
←−
4
ad(D2 )
←−
1 ←− 0
which is not a Janet sequence. As M is a torsion module, using now Theorem 3.25 we get
ext0 (M ) = 0, ext1 (M ) = 0 and we check that N = ext2 (M ) 6= 0. For this, dualizing Ψ by λ
and Ω by θ, we have to look for the CC of the inhomogeneous system:
1
Ψ −→ −d4 θ = λ1
2
Ψ −→
d3 θ = λ2
3
Ψ
−→
d1 θ = λ3
4
Ψ −→ −d2 θ = λ4
which are not already provided by
1
Φ −→
2
Φ −→
Φ3 −→
4
Φ −→
the system:
−d4 λ2 − d3 λ1
−d4 λ3 − d1 λ1
−(d4 λ4 − d2 λ1 ) + (d3 λ3 − d1 λ2 )
d3 λ4 + d2 λ2
=0
=0
=0
=0
One can check that the torsion module N can be generated by {u = d2 λ3 + d1 λ4 , v = d3 λ3 − d1 λ2 }
satisfying the involutive system:
4
φ
3
φ
φ2
1
φ
with the two CC:
ψ2
ψ1
≡
≡
≡
≡
d4 u − d1 v
d4 v
d3 u − d2 v
d3 v
=0
=0
=0
=0
≡ d4 φ2 − d3 φ4 + d2 φ3 − d1 φ1
≡ d4 φ1 − d3 φ3
1
1
1
1
2
2
2
2
3
3
3
3
=0
=0
4
4
•
•
1
1
2 3
2 3
4
4
Accordingly, we have the following strict free resolution of N :
0 −→ D2 −→ D4 −→ D2 −→ N −→ 0
with deleted complex:
0 −→ D2 −→ D4 −→ D2 −→ 0
Applying homD (•, D), we get the desired resolution of L, namely:
0 ←− L ←− D2 ←− D4 ←− D2 ←− 0
Dualizing ψ by z, we finally discover that L is defined by the involutive system:
−φ1
−φ2
−φ3
φ4
−→
−→
−→
−→
d4 z 1 − d1 z 2
d4 z 2
d3 z 1 − d2 z 2
d3 z 2
17
=0
=0
=0
=0
1
1
1
1
2
2
2
2
3
3
3
3
4
4
•
•
and is therefore 2-pure with pd(L) ≤ 2 and a strict inclusion M ⊂ L defined by y = z 1 .
.
REMARK 4.3: In this example, we discover that, if L were also r-pure, we should therefore
have an embedding 0 −→ L −→ extr (extr (L)) = extr (N ) and thus an isomorphism extr (N ) = L
leading to an isomorphism Zr = homD (Fr , D) and to Fr+1 = 0, as can be checked on this example
with r = 2. It has been a challenge for the author during many months to find the following
counter-example showing that, sometimes L may not even be a torsion module.
EXAMPLE 4.4: According to the proof of the theorem, N = extr (M ) does not depend on the
resolution of M used while L does indeed depend on the resolution of N used. Coming back to
the system studied in Example 2.8 and Remark 2.12 with r = 2, we may use the shortest finite
free resolution of M already presented, namely 0 −→ D −→ D2 −→ D −→ M −→ 0. Therefore,
taking the adjoint of the only CC found, we may define N by the system:
v −→ d33 θ
=0
−u −→ d13 θ + d2 θ = 0
and obtain the corresponding involutive system:
4
φ
3
φ
φ2
1
φ
≡
≡
≡
≡
d33 θ
d23 θ
d22 θ
d13 θ + d2 θ
=0
=0
=0
=0
1
1
1
1
2
2
2
•
3
•
•
•
1
1
1
1
2
2
2
2
We obtain the first order involutive system of CC:
4
ψ
3
ψ
ψ2
1
ψ
≡
≡
≡
≡
d3 φ3 − d2 φ4
d3 φ2 − d2 φ3
d3 φ1 − d1 φ4 − φ3
d2 φ1 − d1 φ3 − φ2
=0
=0
=0
=0
3
3
3
•
with the only CC :
ω ≡ d3 ψ 1 − d2 ψ 2 + d1 ψ 4 + ψ 3 = 0 .
We may therefore introduce in reverse order the corresponding adjoint
volved in the Janet sequence we have just constructed:
4
ψ −→ −d1 λ = ν 4
3
ψ −→
λ = ν3
2
ψ −→
d2 λ = ν 2
1
ψ −→ −d3 λ = ν 1
4
4
4
2
φ3 −→ µ3 ≡ d2 ν +4 d1 ν 1
φ −→ µ ≡ −(d3 ν − d1 ν ) + (d2 ν 3 − ν 2 )
φ2 −→ µ2 ≡ −d3 ν 3 − ν 1
1
φ −→ µ1 ≡ −d3 ν 2 − d2 ν 1
operators of the ones in-
=0
=0
=0
=0
This last operator is defining L but is not involutive. We have the two torsion elements:
µ5 ≡ d2 ν 3 − ν 2 ,
µ6 ≡ d1 ν 3 + ν 4
which are generating ext2 (N ) and are easily seen to satisfy the involutive system:
d3 µ6 − µ5 = 0,
d3 µ5 = 0,
d2 µ6 − d1 µ5 = 0,
d2 µ5 = 0
because d2 µ5 ≡ d2 µ3 + d3 µ4 + d1 µ1 = 0. Finally, using the first equation, we may eliminate µ5
and identify µ6 with y because we have indeed d33 µ6 = 0, d13 µ6 − d2 µ6 = 0 in order to obtain the
strict inclusion M ⊂ L. Equivalently, we may also eliminate ν 1 and ν 2 respectively from µ2 and
µ3 in order to obtain:
µ4 −→ d13 (d1 ν 3 + ν 4 ) − d2 (d1 ν 3 + ν 4 ) = 0,
18
µ1 −→ d33 (d1 ν 3 + ν 4 ) = 0
but we may notice that L is not 2-pure and thus a torsion module because ν 3 (similarly ν 4 ) is
not by itself a torsion element of L. Such a situation is well known in control theory with the
SISO (single input u, single output y) system ẏ − u̇ = 0 because u (similarly y) is not by itself a
torsion element but z = y − u is a torsion element because ż = 0 (See the pages 9 and 10 of the
introduction in [17] for more details on such a comment).
PROBLEM : Is it possible to find an analogue of the previous theorem or of the case r = 0,
where L should be also r-pure with a free resolution having exactly r morphisms ?.
5) ABSOLUTE AND RELATIVE LOCALIZATIONS :
Surprisingly, the positive answer to such a problem has been given by Macaulay in ([11]) for
differential modules defined by systems with constant coefficients and only one unknown. Our purpose in this section is to generalize this resul to arbitrary differential modules defined by systems
of PD equations with coefficients in a differential field.
Now we hope that, after reading the previous section, the reader is convinced that the use of
extension modules is a quite important though striking tool for studying linear multidimensional
systems. Of course, as for any new language, it is necessary to apply it on many explicit examples
before being familiar with it. However, it is evident that it should be even more important to have
a direct approach allowing to exhibit the purity filtration and, in particular, to recognize whether
a differential module is pure or not. The purpose of this section is to combine the module approach
with the system approach, while taking into account the specific properties of the Spencer form in
a way rather similar to the use of the Kalman form of a control system when testing controllability, namely to check that an ordinary differential module is 0-pure. For this, we shall divide the
procedure into a few successive constructive steps that will be illustrated on explicit examples.
• STEP 1: Whenever a system Rq ⊂ Jq (E) is given, there is no way to obtain informations on
the corresponding module without bringing this system to an involutive or at least formally integrable system by means of prolongations and projections as in the Example 2.8 of Macaulay where
(2)
only the projection R2 ⊂ R2 of R4 to R2 is involutive. Of course, an homogeneous system with
constant coefficients is automatically formally integrable and one only needs to use a finite number
of prolongations in order to obtain an involutive symbol, though it is known that 2-acyclicity is
sufficient to obtain first order generating CC ([17]). However, it is essential to notice that it is
only with an involutive system that we are sure that the CC system is first order both with the
following ones in the Janet sequence.
EXAMPLE 5.1: With K = Q, m = 1, n = 3, q = 2, the homogeneous second order systems
y33 = 0, y23 − y11 = 0, y22 = 0 or y33 − y11 = 0, y23 = 0, y22 − y11 = 0 both have a 2-acyclic symbol
g3 of dimension 1 at order 3 (exercise) and a trivially involutive symbol g4 = 0 at order 4, such a
result leading to only one CC of order 2 with cd(M ) = 3 in both cases. We let the reader treat the
system y3 = 0, y12 = 0 similarly and conclude (Hint: (χ3 , χ1 χ2 ) = (χ3 , χ1 ) ∩ (χ3 , χ2 ) is unmixed).
It is however not evident that the homogeneous system y11 = 0, y12 = 0, y13 = 0, y23 = 0 of Example 3.33 is involutive.
Finally, according to section 2 and 3, this first step provides the characters α1q ≥ ... ≥ αnq ≥ 0
and the smallest non-zero character α = αqn−r 6= 0 providing cd(M ) = r, a result leading at once
to tr (M ) ⊂ M with a strict inclusion while tr−1 (M ) = ... = t0 (M ) = t(M ) = M . Of course, if
α = αnq 6= 0, then M cannot be a torsion module and t(M ) ⊂ M with a strict inclusion. The
following example proves nevertheless that it is much more delicate to study systems with variable
coefficients.
EXAMPLE 5.2: With K = Q(x2 ), n = 3, m = 1, q = 1, let us consider the differential module M
defined by the trivially involutive system y3 − x2 y1 = 0. We have cd(M ) = 1 but we can only say
that cd(Dz) ≥ 1, ∀z ∈ M . If we set z = y2 , proceeding as in Remark 2.12, we get the involutive
system:
19
y3
y2
y1
=
=
=
x2 z3 − (x2 )2 z1
z
z3 − x2 z1
1 2
1 2
1 •
3
•
•
The differential submodule Dz ⊂ M is defined by the second order involutive system:
z33 − 2x2 z13 + (x2 )2 z11
z23 − x2 z12 − 2z1
=0
=0
1 2
1 2
3
•
and we get cd(Dz) = 1 exactly. However, even on such a very elementary example, it is not
evident that t0 (M ) = t(M ) = M is 1-pure. We also understand that the decoupling system for
any autonomous element in engineering sciences, like in magnetohydrodynamics, cannot be studied
without these new techniques if we want intrinsic results. Finally, if we denote by I the left ideal of
D = Dy generated by y3 −x2 y1 , we notice the relation ann(G) = (χ3 −x2 χ1 ) = gr(I) = rad(gr(I)).
However, we have ann(gr(Dz)) = ((χ3 − x2 χ1 )2 , χ2 (χ3 − x2 χ1 )) with radical equal to the prime
ideal (χ3 − x2 χ1 ) as before. Hence, in this example, the strict inclusion Dz ⊂ M does not imply
gr((Dz) ⊂ gr(M ) = G because otherwise we should get ann(G) ⊆ ann(gr(Dz) and this is the
reason for which only the radical must be considered as it does not depend on the filtration.
• STEP 2: Once we have obtained cd(M ) = r, in order to check that M is r-pure, it remains
to prove that tr (M ) = 0 as we already know that tr−1 (M ) = M . For this, the second step will
be to use the specific properties of the Spencer form Rq+1 ⊂ J1 (Rq ). More generally, it is possible to use any equivalent involutive first order system of the form R1 ⊂ J1 (E) with no zero
order equations, that is with an induced epimorphism R1 −→ E −→ 0 and such that the corresponding differential module is isomorphic and thus identified to the initial module as in Remark
2.13 . We have now the characters α11 ≥ ... ≥ αn1 ≥ 0 and the smallest non-zero character is
is still α = α1n−r 6= 0 providing of course the same codimension cd(M ) = r as in the first step.
Accordingly, the number r of non-zero characters and the number r of full classes is the same
as in the previous step. However, it must be noticed that the filtration may be different and the
following example explains once more why only the radical of the characteristic ideal must be used.
EXAMPLE 5.3: K = Q, n = 2, m = 1, q = 2.For the involutive system y22 = 0, y12 = 0, the characteristic ideal is a = ((χ2 )2 , χ1 χ2 ) =⇒ rad(a) = (χ2 ) =⇒ r = 1. Setting z 1 = y, z 2 = y1 , z 3 = y2 ,
we get the equivalent first order system d2 z 3 = 0, d2 z 2 = 0, d2 z 1 − z 3 = 0, d1 z 1 − z 2 = 0, d1 z 3 = 0
and the polynomial ideal generated by the 3 × 3 minors of the characteristic matrix is a =
((χ2 )3 , (χ2 )2 χ1 , χ2 (χ1 )2 ). Hence the characteristic ideal is rad(a) = (χ2 ) and r = 1 too.
EXAMPLE 5.4: For Example 2.8 we may set z 1 = y, z 2 = y1 , z 3 = y2 , z 4 = y3 and obtain the
first order involutive system :
d3 z 4 = 0, d3 z 3 = 0, d3 z 2 − z 3 = 0, d3 z 1 − z 4 = 0
d2 z 4 = 0, d2 z 3 = 0, d2 z 2 − d1 z 3 = 0, d2 z 1 − z 3 = 0
d1 z 4 − z 3 = 0, d1 z 1 − z 2 = 0
1
1
1
2 3
2 •
• •
with no zero-order equation. We have α31 = 4 − 4 = 0, α21 = 4 − 4 = 0, α11 = 4 − 2 = 2 =⇒ r = 2
too. We let the reader treat Example 4.2 similarly and obtain α41 = 0, α31 = 0, α21 = 2 =⇒ r = 2.
It is at this moment that we discover that such systems have particular properties not held by
other systems, apart from the fact that a canonical sequence may be constructed exactly like the
Spencer sequence or the first order part of the Janet sequence.
Shrinking the board of multiplicative variables, we obtain from the definition of involutiveness:
PROPOSITION 5.5: For an involutive first order system with no zero order equations and
solved with respect to the principal (pri) first order jets expressed by means of the parametric
20
(par) other first order jets, the system obtained by looking only at the PD equations of class 1+
... + class i only contains d1 , ..., di and is still involutive ∀i = 1, ..., n, after adopting the ordering
di+1 , ..., dn , d1 , ..., di .
EXAMPLE 5.6: Looking at Example 2.9, we notice that the systems:
y3 − x4 y1
y2 − y1
y2 − y1
=0
=0
=0
4
4
3
1
1
2 3
2 •
4 1
2
are both involutive. Also, looking at Example 5.4, we notice that the system:
d2 z 4 = 0, d2 z 3 = 0, d2 z 2 − d1 z 3 = 0, d2 z 1 − z 3 = 0
d1 z 4 − z 3 = 0, d1 z 1 − z 2 = 0
3
3
1
1
2
•
is still involutive and we let the reader treat Example 4.2 similarly.
We shall denote the corresponding differential module by Mn−i and we notice that M = M0
is defined by more equations than Mn−i . Accordingly, we have an epimorphism (specialization)
Mn−i −→ M −→ 0 and similarly epimorphisms Mn−i −→ Mn−i−1 −→ 0. Finally, as cd(M ) = r,
we notice that the classes n − r + 1, ..., n are full and we find therefore (χn )m , ..., (χn−r+1 )m among
the m × m minors with lower powers of χ1 , ..., χn−r for the other minors because the numbers of
equations of the lower classes are decreasing and thus strictly smaller than m. The characteristic
ideal is thus (χn , ..., χn−r+1 ) if we set χ1 , ..., χn−r to zero, in a coherent way. Finally, choosing
i = n − r, we get an epimorphism Mr −→ M −→ 0. The background will always indicate clearly
the meaning of the lower index and cannot be confused with the filtration index of M .
• STEP 3: We are now in position to study Mr with more details as a system in n − r variables
([11], §77, p 86). Its defining system has β1n−r = β = m − α equations of strict class n − r, a
smaller number of equations of class n − r − 1, ... , and eventually an even smaller number of
equations of class 1. Studying this system for itself, we may look for t(Mr ) exactly following the
known techniques working for any differential module, in particular double duality as described
in section 4. However, if one could find any (relative) torsion element z ∈ Mr , we could project
it to an element z ∈ M and we have N = Dz ⊆ M where we do not put a residue bar on the
new z for simplicity. Introducing the respective annihilator ideal a and b of M and N , we should
have a ⊆ rad(a) ⊆ rad(b) as it is the only result not depending on the filtration of the modules.
However, we know that z must be killed by at least one operator involving only d1 , ..., dn−r , in
addition to the operators involving separately (dn )m + ..., ..., (dn−r−1 )m + ... and we should obtain
tr−1 (M ) = M but tr (M ) 6= 0, that is M should not be pure. Hence M is r-pure if and only if Mr
is torsion-free as a differential module over K[d1 , ..., dn−r ]. In such a case, the system defining Mr
can be parametrized by α arbitrary unknowns among {y 1 , ..., y m } by using a so-called minimum
parametrization in the sense of ([24]). In actual practice, as shown on all the examples, one can use
a relative localization with respect to (d1 , ..., dn−r ) only by keeping (dn−r+1 , ..., dn ) untouched and
replacing (d1 , ..., dn−r ) by (χ1 , ..., χn−r ) considered as (constant) ”parameters” in the language of
Macaulay ([11], §43, p 45 and §77, p 86 with r instead of n − r and a different ordering). Such
a method provides therefore a quite useful and simple test for checking purity by linking it to
involutivity. An important intermediate result is provided by the next proposition.
PROPOSITION 5.7: The partial localization ”kills” the equations of class 1 up to class n− r − 1
(care) and finally only depends on the equations of strict class n − r.
Proof: Instead of using the column n − r in the multiplicative board, we provide the proof when
r = 0 by using the column n. Working out as usual the first order CC, we only look at the p < m
CC of class n for the equations Φ1 , ..., Φp of class 1 up to class n − 1 if we order the Φ starting
from the lower class involved. These p CC will be of a very specific form with a square p × p
operator matrix for (Φ1 , ..., Φp ) with diagonal operators of the form dn + ... where the dots denote
21
operators involving only d1 , ..., dn−1 in a quasi linear way with coefficients in K, the remaining
of the matrix only depending on d1 , ..., dn−1 for the (Φp+1 , ..., Φm ) of strict class n. Therefore, if
we have K = k = cst(K), the absolute localization is simply done by setting di −→ χi and the
determinant of the square matrix is equal to (χn )p + ... where the dots denote a polynomial of
degree < p in χp with coefficients involving only χ1 , ..., χn−1 . It follows that each Φ1 , ..., Φp can be
expressed as a linear combination over k(χ1 , ..., χn ) of the Φ of strict class n. The result is similar
for the variable coefficient case by using the graded machinery but needs much more work. In any
case, setting the Φ of strict class n equal to zero, we should obtain a zero graded module for the
Φ1 , ..., Φp which must be eequal to zero too. It must finally be noticed that the first order CC used
are in reduced Spencer form as the dn of the Φ of strict class n do not appear in the CC we have
used and these Φ do not appear in the other CC too.
Q.E.D
EXAMPLE 5.8: Coming back to Example 2.11, we notice that the only CC is an identity in
(u, v, w) and we may forget about u in order to obtain the new system for (v, w):
3
φ ≡ d4 v − x3 d1 v − v = 0
φ2 ≡ d4 w − x3 d1 w − w = 0
1
φ ≡ d3 w − x4 d1 w − d2 v + d1 v = 0
1
1
1
2 3
2 3
2 3
4
4
•
defining a new module M with cd(M ) = 1 as the class 4 is now full. The same CC as before can
be written again in the form:
(d4 − x3 d1 − 1)φ1 = (d3 − x4 d1 )φ2 − (d2 − d1 )φ3
that we can localize in (d1 , d2 , d3 ) with:
φ1 = 0 =⇒ (d3 − x4 d1 )w = (d2 − d1 )v =⇒ v = (d3 − x4 d1 )y, w = (d2 − d1 )y
We let the reader check that we have indeed:
2
φ ≡ (d4 − x3 d1 − 1)(d3 − x4 d1 )y
φ3 ≡
(d4 − x3 d1 − 1)(d2 − d1 )y
=
=
(d3 − x4 d1 )u = 0 =⇒ u = 0
(d2 − d1 )u = 0 =⇒ u = 0
We finally obtain for M a relative parametrization with the only constraint u ≡ (d4 −x3 d1 −1)y = 0
in a coherent way with Example 2.11.
• STEP 4: In this section we come back to the commutative situation with a field k of constants
and generalize the results of Macaulay in the following theorem which is recapitulating the results
so far obtained.
THEOREM 5.9: One has the commutative and exact diagram:
0 −→
t(Mr ) −→ Mr
↓
↓
0 −→ tr (M ) −→ M
↓
↓
0
0
−→ k(χ1 , ..., χn−r ) ⊗
↓
−→ k(χ1 , ..., χn−r ) ⊗
↓
0
Mr
M
Proof: For simplifying the notations in this diagram of modules over D, we have identified Mr as a
module over k[d1 , ..., dn−r ] with Mr as a module over k[dn−r+1 , ..., dn , d1 , ..., dn−r ] = k[d1 , ..., dn ] =
D while the localization of M just tells that the coefficients are now in the field k(χ1 , ..., χn−r ),
exactly following Macaulay. Moreover the central column is exact according to the definitionof Mr
and the right column is exact because localization preserves the exactness of a sequence.
For exampe, with k = Q, n = 2, m = 2, q = 2, r = 1, the differential module M defined by
the involutive system y22 = 0, y12 = 0 may also be defined by the first order involutive system
z 1 = y, z 2 = y1 , z 3 = y2 =⇒ d2 z 3 = 0, d2 z 2 = 0, d2 z 1 − z 3 = 0, d1 z 3 = 0, d1 z 1 − z 2 = 0. Then M1
is defined by the first order system d1 z 3 = 0, d1 z 1 − z 2 = 0 with torsion module t(M1 ) generated
22
by z 3 and the tensor product of M by k(χ1 ) is defined by d2 z 3 = 0, d2 z 2 = 0, d2 z 1 = 0, d1 z 3 =
0, z 3 = 0, χ1 z 1 − z 2 = 0 after division by χ1 in a coherent way with Example 1.1.
Q.E.D.
Finally, when Mr is torsion-free as a differential module over k[d1 , ..., dn−r ], then tr (M ) = 0
and we get the following generalization of a result provided by Macalualy ([11], §41, p 43):
COROLLARY 5.10: The differential module M is r-pure if and only if cd(M ) = r and there is
a monomorphism 0 −→ M −→ k(χ1 , ..., χn−r ) ⊗ M .
• STEP 5: The final idea is to embed Mr into a free module over K[d1 , ..., dn−r ] in order to
parametrize the corresponding system and substitute into the equations of class n−r+1, ..., n. However, if we look at Example 1.2, we should find after the substitution Φ1 ≡ z13 = 0, Φ2 ≡ z23 = 0
with one CC d2 Φ1 − d1 Φ2 = 0, that is on one side a module L which is not 1-pure and, on the
other side a module L having a finite free resolution with 2 operators. However, we forgot that
M , being pure, may be identified with its embedding into its localization. Hence, we get in fact
χ1 z3 = 0, χ2 z3 = 0 and thus only z3 = 0 is providing a convenient parametrizing module L.
Our purpose is to explain and illustrate this procedure for finding such an L in the general situation. Again, the main idea will be provided by this example. Indeed, we obtain the
only CC Ψ ≡ d3 Φ3 − d2 Φ1 + d1 Φ2 = 0. Substituting the parametrization, we get of course
Φ3 = 0 ⇐⇒ χ1 y 2 = χ2 y 1 , that is, among the two unknowns y 1 , y 2 we are left with only one,
say y 1 and, similarly, among the two equations Φ1 , Φ2 we are left with only one, say Φ1 , because
χ1 Φ2 = χ2 Φ1 from the CC which is of course compatible with the localization and we choose
z3 = 0 as χ1 z3 = 0 =⇒ z3 = 0.
The general situation may be treated similarly. Indeed, according to the previous step, we
are only concerned with the equations of class n − r + 1, ... , class n while the localization has
only to do with the β equations of strict class n − r (care) allowing to express β unknowns as
linear combinations of the α remaining unknowns with coefficients in k(χ1 , ..., χn−r ). To each such
equation are associated exactly r dots and each dot of index n − r + i provides a reduction of the
respective equations of class n − r + i for i = 1, ..., r. It follows that we are left with α equations of
each such class. When we ”delocalize”, replacing χi by di , we have to take into account the need to
take out the denominators and may find a few ”simplifications” as in the example just considered..
Finally, the maximum number r − 1 (care again) of dots found for one equation is obtained for the
equations of strict class n − r + 1 = n − (r − 1) and we have thus exhibited a system defining a
module L which is r-pure and admits a free resolution with exactly (r − 1) + 1 = r operators. In
any case, the reader must not forget that the localization of a module is useful only if we already
know that this module is torsion-free by means of the double-duality formula t(M ) = ext1 (N )
given in the introduction.
EXAMPLE 5.11: Let M be defined by the involutive system:
d3 y 4 + d1 y 2 − d1 y 1
d3 y 3 − d2 y 4 + d1 y 2 − d1 y 1
d3 y 2 + d1 y 2
d3 y 1 − d1 y 4 + d1 y 2
d2 y 2 − d1 y 4 + d1 y 1
d2 y 1 − d1 y 3 + d1 y 1
=0
=0
=0
=0
=0
=0
1
1
1
1
1
1
2
2
2
2
2
2
3
3
3
3
•
•
with characters α31 = 0, α21 = α = 2, α11 = 4. It follows that cd(M ) = 1 as only the class 4 is full
and we obtain the following relative localization showing that M is 1-pure:
y 1 = χ1 y, y 2 = χ1 z, y 3 = (χ1 + χ2 )y, y 4 = χ1 y + χ2 z
Substituting in the four equations of class 3, we only obtain the two equations:
23
d3 z + χ1 z
d3 y + (χ1 − χ2 )z − χ1 y
=0
=0
after a division by χ1 , χ2 and χ1 + χ2 . The parametrizing module L is thus defined by the two
equations:
d3 z + d1 z
=0
d3 y + (d1 − d2 )z − d1 y = 0
which are differentially independent and we have the relative parametrization:
y 1 = d1 y,
y 2 = d1 z,
y 3 = (d1 + d2 )y,
y 4 = d1 y + d2 z
Finally, M ⊂ L =⇒ ass(M ) ⊂ ass(L) =⇒ ass(M ) = {(d3 + d1 ), (d3 − d1 )} =⇒ annD (M ) =
(d3 + d1 ) ∩ (d3 − d1 ), a striking result showing that M can be embedded into the direct sum of two
primary differential modules according to Remark 3.32 (See [18] for more details).
EXAMPLE 5.12: With k = Q, m = 1, n = 3, let us consider the polynomial map χ1 = u5 , χ2 =
u3 , χ3 = u4 as in ([11], p 53). We have the exact sequence 0 −→ p −→ k[χ] −→ k[u] ⊂ k(u)
showing that p = ((χ2 )2 (χ3 ) − (χ1 )2 , (χ2 )3 − χ1 χ3 , (χ3 )2 − χ1 χ2 ) is a prime ideal ([18], p 126). The
corresponding prime differential module M is defined by the involutive system:
y333 − y123
y
233 − y122
y223 − y11
y222 − y13
y
133 − y112
y33 − y12
=0
=0
=0
=0
=0
=0
1
1
1
1
1
•
2
2
2
2
•
•
3
•
•
•
•
•
and is 2-pure. The localized system is finite type over k(χ1 )[d2 , d3 ] with par = {y, y2 , y3 , y22 , y23 }.
One can prove that p2 is not a primary ideal even though rad(p2 ) = p.
EXAMPLE 5.13: Similarly but now with k = Q, m = 1, n = 4, let us consider the polynomial map χ1 = uv, χ2 = u, χ3 = uv 3 , χ4 = uv 4 as in ([11], p 47). We have the exact sequence
0 −→ p −→ k[χ] −→ k[u, v] ⊂ k(u, v) showing that p = (χ2 χ4 − χ1 χ3 , (χ1 )3 − (χ2 )2 χ3 , (χ3 )3 −
χ1 (χ4 )2 , (χ1 )2 χ4 − χ2 (χ3 )2 ) is a prime ideal. It is not evident at all that the corresponding prime
differential module M can be defined by the homogeneous involutive system (exercise):
y444 − y224 − y134 − y123
y
344 − y111
y
334 − y114 − y112
y333 − y124 − y122 − y113
y244 + y224 − y123
y
234 − y133 + y111
y
144 + y124 − y113
y44 + y24 − y13
=0
=0
=0
=0
=0
=0
=0
=0
1
1
1
1
1
1
1
•
2
2
2
2
2
2
•
•
3
3
3
3
•
•
•
•
4
•
•
•
•
•
•
•
and is thus also 2-pure. The localized system is finite type over k(χ1 , χ2 )[d3 , d4 ] with par =
{y, y3 , y4 , y33 } and we have for example χ2 y34 − χ1 y33 + (χ1 )3 y = 0 in a coherent way with the
comments of Macaulay in ([11], §78, p 88, formula (A) and §88,89, p 98).
6) CONCLUSION :
In 1916, F.S. Macaulay discovered a new localization technique for studying unmixed polynomial ideals. We have been able to generalize this procedure for studying pure differential modules,
obtaining in particular a kind of relative parametrization generalizing the absolute parametrization already known for torsion-free modules and equivalent to controllability in classical control
theory. In the language of multidimensional systems theory, which is more intuitive, instead of
24
using arbitrary potential-like functions for the parametrization, the idea is now to use potentiallike functions which must satisfy a kind of minimum differential constraint limiting, in some sense,
the number of independent variables appearing in these functions, in a way similar to the situation met in the Cartan-Khäler theorem of analysis. For such a purpose, we have exhibited new
links between purity and involutivity, providing also a new insight into the primary decomposition
of modules and ideals by means of tools from the formal theory of linear multidimensional systems.
7) BIBLIOGRAPHY :
[1] J.E. BJORK: Analytic D-modules and Applications, Kluwer, 1993.
[2] N. BOURBAKI: Algèbre homologique, chap. X, Masson, Paris, 1980.
[3] N. BOURBAKI: Algèbre commutative, chap. I-IV, Masson, Paris, 1985.
[4] B. BUCHBERGER: Gröbner bases: an Algorithmic Methods in Polynomial Ideal Theory, in:
Recent Trends in Multidimensional System Theory, N.K. Bose ED., Reidel, Dordrecht, 1985, 184232.
[5] E. COSSERAT, F. COSSERAT: Théorie des Corps Déformables, Hermann, Paris, 1909.
[6] W. GRÖBNER: Über die Algebraischen Eigenschaften der Integrale von Linearen Differentialgleichungen mit Konstanten Koeffizienten, Monatsh. der Math., 47, 1939, 247-284.
[7] M. JANET: Sur les Systèmes aux dérivées partielles, Journal de Math., 8, 3, 1920, 65-151.
[8] E.R. KALMAN, Y.C. YO, K.S. NARENDA: Controllability of Linear Dynamical Systems, Contrib. Diff. Equations, 1, 2, 1963, 189-213.
[9] M. KASHIWARA: Algebraic Study of Systems of Partial Differential Equations, Mémoires de
la Société Mathématique de France 63, 1995, (Transl. from Japanese of his 1970 Master’s Thesis).
[10] E. KUNZ: Introduction to Commutative Algebra and Algebraic Geometry, Birkhäuser, 1985.
[11] F.S. MACAULAY: The Algebraic Theory of Modular Systems, Cambridge Tracts, vol. 19,
Cambridge University Press, London, 1916. Stechert-Hafner Service Agency, New-York, 1964.
[12] P. MAISONOBE, C. SABBAH: D-Modules Cohérents et Holonomes, Travaux en Cours, 45,
Hermann, Paris, 1993.
[13] D.G. NORTHCOTT: An Introduction to Homological Algebra, Cambridge University Press,
Cambridge, 1966.
[14] D.G. NORTHCOTT: Lessons on Rings, Modules and Multiplicities, Cambridge University
Press, Cambridge, 1968.
[15] V.P. PALAMODOV: Linear Differential Operators with Constant Coefficients, Grundlehren
der Mathematischen Wissenschaften 168, Springer, 1970.
[16] J.-F. POMMARET,:Dualité Différentielle et Applications, C. R. Acad. Sci. Paris, 320, Série
I, 1995, 1225-1230.
[17] J.-F. POMMARET: Partial Differential Equations and Group Theory: New Perspectives for
Applications, Kluwer, 1994.
[18] J.-F. POMMARET: Partial Differential Control Theory, Kluwer, 2001.
[19] J.-F. POMMARET: Algebraic Analysis of Control Systems Defined by Partial Differential
Equations, in Advanced Topics in Control Systems Theory, Lecture Notes in Control and Information Sciences LNCIS 311, Chapter 5, Springer, 2005, 155-223.
[20] J.-F. POMMARET: Parametrization of Cosserat Equations, Acta Mechanica, 215, 2010, 4355.
[21] J.-F. POMMARET: Macaulay Inverse Systems Revisited, Journal of Symbolic Computation,
46, 2011, 1049-1069.
[22] J.-F. POMMARET: Spencer Operator and Applications: From Continuum Mechanics to
Mathematical Physics, in ”Continuum Mechanics-Progress in Fundamentals and Engineering Applications”, Dr. Yong Gan (Ed.), ISBN: 978-953-51-0447–6, InTech, 2012, Available from:
http://www.intechopen.com/books/continuum-mechanics-progress-in-fundamentals-and-engineerin-applications/spen
[23] J.-F. POMMARET: A Pedestrian Approach to Cosserat/Maxwell/Weyl Theory, Preprint 2012
http://hal.archives-ouvertes.fr/hal-00740314
http://fr.arXiv.org/abs/1210.2675
[24] J.-F. POMMARET, A. QUADRAT: Localization and parametrization of linear multidimensional control systems, Systems & Control Letters 37 (1999) 247-260.
[25] J.-F. POMMARET, A. QUADRAT: Algebraic Analysis of Linear Multidimensional Control
Systems, IMA Journal of Mathematical Control and Informations, 16, 1999, 275-297.
25
[26] A. QUADRAT: Analyse Algébrique des Systèmes de Contrôle Linéaires Multidimensionnels,
Thèse de Docteur de l’Ecole Nationale des Ponts et Chaussées, 1999
(http://www-sop.inria.fr/cafe/Alban.Quadrat/index.html).
[27] A. QUADRAT: Une Introduction à l’Analyse Algébrique Constructive et à ses Applications,
INRIA Research Report 7354, AT-SOP Project, july 2010. Les Cours du CIRM, 1 no. 2: Journées
Nationales de Calcul Formel (2010), p281-471 (doi:10.5802/ccirm.11).
[28] J.J. ROTMAN: An Introduction to Homological Algebra, Pure and Applied Mathematics,
Academic Press, 1979.
[29] W. M. SEILER: Involution: The Formal Theory of Differential Equations and its Applications
to Computer Algebra, Springer, 2009, 660 pp. (See also doi:10.3842/SIGMA.2009.092 for a recent
presentation of involution, in particular sections 3 (p 6 and reference [11], [22]) and 4).
[30] D.C. SPENCER: Overdetermined Systems of Partial Differential Equations, Bull. Amer.
Math. Soc., 75, 1965, 1-114.
[31] O. ZARISKI, P. SAMUEL: Commutative Algebra, Van Nostrand, 1958.
[32] E. ZERZ: Topics in Multidimensional Linear Systems Theory, Lecture Notes in Control and
Information Sciences 256, Springer, 2000.
26
| 0math.AC |
Similarity-Aware Spectral Sparsification by Edge Filtering
Zhuo Feng
arXiv:1711.05135v2 [cs.DS] 15 Nov 2017
Department of Electrical and Computer Engineering
Michigan Technological University
Email: zhuofeng@mtu.edu
Abstract—In recent years, spectral graph sparsification techniques that
can compute ultra-sparse graph proxies have been extensively studied
for accelerating various numerical and graph-related applications. Prior
nearly-linear-time spectral sparsification methods first extract low-stretch
spanning tree of the original graph to form the backbone of the sparsifier,
and then recover small portions of spectrally-critical off-tree edges to the
spanning to significantly improve the approximation quality. However, it
is not clear how many off-tree edges should be recovered for achieving a
desired spectral similarity level within the sparsifier. Motivated by recent
graph signal processing techniques, this paper proposes a similarityaware spectral graph sparsification framework that leverages an efficient
off-tree edge filtering scheme to construct spectral sparsifiers with guaranteed spectral similarity (relative condition number) level. An iterative
graph densification framework and a generalized eigenvalue stability
checking scheme are introduced to facilitate efficient and effective filtering
of off-tree edges even for highly ill-conditioned problems. The proposed
method has been validated using various kinds of graphs obtained from
public domain sparse matrix collections relevant to VLSI CAD, finite
element analysis, as well as social and data networks frequently studied
in many machine learning and data mining applications.
I. I NTRODUCTION
Spectral methods are playing increasingly important roles in many
graph and numerical applications [25], such as scientific computing
[23], numerical optimization [2], data mining [14], graph analytics
[9], machine learning [6], graph signal processing [17], and VLSI
computer-aided design [7], [26]. For example, classical spectral graph
partitioning (data clustering) algorithms embed original graphs into
low-dimension space using the first few nontrivial eigenvectors of
graph Laplacians and subsequently perform graph partitioning (data
clustering) on the low-dimensional graphs to obtain high-quality
solution.
To further push the limit of spectral methods for large graphs,
mathematics and theoretical computer science researchers have extensively studied many theoretically-sound research problems related
to spectral graph theory [1], [3], [10], [14], [19], [22]. Recent spectral
graph sparsification research allows computing nearly-linear-sized
subgraphs (sparsifiers) that can robustly preserve the spectrum (i.e.,
eigenvalues and eigenvectors) of the original graph’s Laplacian,
which immediately leads to a series of “theoretically nearly-lineartime” numerical and graph algorithms for solving sparse matrices,
graph-based semi-supervised learning (SSL), spectral graph partitioning (data clustering), and max-flow problems [2], [8], [10], [11],
[18], [19], [22], [23]. For example, sparsified circuit networks allow
for developing more scalable computer-aided (CAD) design algorithms for designing large VLSI systems [7], [26]; sparsified social
(data) networks enable to more efficiently understand and analyze
large social (data) networks; sparsified matrices can be immediately
leveraged to accelerate the solution computation of large linear
system of equations [27]. To this end, a nearly-linear time spectral
sparsification algorithm leveraging an edge sampling scheme that sets
sampling probabilities proportional to edge effective resistances (of
the original graph) has been proposed in [19]. However, it becomes
a chicken-and-egg problem since calculating effective resistances
requires solving the original Laplacian matrix multiple times and thus
can be extremely expensive for large graphs.
A practically-efficient, nearly-linear complexity spectral graph
sparsification algorithm has been recently introduced in [7], which
first extracts a “spectrally critical” spanning tree subgraph as a
backbone of the sparsifier, and subsequently recovers a small portion
of dissimilar “spectrally critical” off-tree edges to the spanning tree.
This algorithm has been shown to be highly efficient for solving
symmetric diagonally dominant (SDD) matrices in many circuit and
finite element analysis tasks. However, in many scientific computing
and graph-related applications, it is important to generate spectral
graph sparsifiers of desired spectral similarity level with minimum
number of edges: introducing too few edges may lead to poor
approximation of the original graph, whereas too many edges can
result in high computational complexity. For example, considering
a preconditioned conjugate gradient (PCG) solver to be used for
solving multiple right-hand-side (RHS) vectors, it is hoped the PCG
solver should convergence to a good solution as fast as possible,
which usually requires the sparsifier (preconditioner) to have a high
spectral similarity; on the other hand, in many graph partitioning
tasks [20], only the Fielder vector (the first nontrivial eigenvector) of
graph Laplacian is needed, which usually only needs a much sparser
subgraph with relatively low spectral similarity.
This paper introduces a similarity-aware spectral graph sparsification framework that leverages an efficient off-tree edge filtering
scheme to construct spectral sparsifiers with guaranteed spectral
similarity. An incremental graph densification procedure and a generalized eigenvalue stability checking scheme are also proposed
for efficiently and effectively filtering off-tree edges. The proposed
methods will allow sparsifying real-world large graphs according
to desired spectral similarity levels, making it possible to flexibly
trade off computational complexity and solution quality by effectively
controlling the graph sparsity and spectral similarity of the spectral
sparsifiers. The major contributions of this work include:
1) We present a similarity-aware spectral graph sparsification
framework leveraging an iterative edge filtering scheme motivated by recent graph signal processing problems [16], [17].
2) An iterative graph densification procedure is introduced to
incrementally improve the approximation quality of the graph
sparsifier, which allows flexibly trading off the sparsity and
spectral similarity of the spectral sparsifier in various numerical
and graph-related applications.
3) Extensive experiments have been conducted to validate the
proposed method in various numerical and graph-related applications, such as solving sparse SDD matrices, and spectral
graph partitioning, and reduction of large social, and data
networks.
The rest of this paper is organized as follows. Section II provides
a brief introduction to prior work on spectral graph sparsification
problems. In Section III, a similarity-aware method for spectral
graph sparsification is described in details. Section IV demonstrates
extensive experiment results for a variety of real-world, large-scale
sparse Laplacian matrix and graph problems, which is followed by
the conclusion of this work in Section V.
0.04
0.03
0.02
0.01
0
-0.01
0.04
-0.02
-0.03
-0.02
0.02
0
0
-0.02
Fig. 1. Spectral drawing of two σ-spectrally similar (σ = 5) airfoil graphs.
II. S PECTRAL G RAPH S PARSIFICATION
Consider a graph G = (V, E, w) with V denoting the vertex (data
point) set of the graph, E denoting the edge set of the graph, and w
denoting a weight (similarity) function that assigns positive weights
to all edges. The graph Laplacian of G is a symmetric diagonally
dominant (SDD) matrix that can be constructed as follows:
−w(p, q)
if (p, q) ∈ E
P
w(p,
t)
if (p = q)
LG (p, q) =
(1)
(p,t)∈E
0
otherwise .
Spectral graph sparsification [22] aims to preserve original graph
spectrum within ultra-sparse subgraphs (sparsifiers), which allows
preserving not only eigenvalues and eigenvectors of the original graph
Laplacian, but also distances (e.g. effective resistances) between
vertices, low-dimensional graph embedding, as well cuts in the graph.
Two graphs G and P are σ−spectrally similar if their quadratic
forms satisfy the following conditions for all real vectors x ∈ RV :
xT LP x
≤ xT LG x ≤ σxT LP x.
(2)
σ
Define the relative condition number to be κ(LG , LP ) =
λmax /λmin , where λmax and λmin denote the largest and smallest
nonzero generalized eigenvalues satisfying:
LG u = λLP u,
(c) similarity-aware sparsifier improvement (graph densification) by
gradually adding small portions of “filtered” off-tree edges to the
spanning tree. Fig. 1 shows the spectral drawing [9] of an airfoil
graph [5] as well as its spectrally-similar subgraph computed by the
proposed similarity-aware spectral sparsification algorithm.
In the following sections, assume that G = (V, E, w) is a
weighted, undirected and connected graph, whereas P = (V, Es , ws )
is its graph sparsifier. To simplify the our analysis, we assume the
edge weights in the sparsifier remain the same as the original ones,
though edge weight re-scaling schemes [22] also can be applied to
further improve the approximation quality. The descending eigenvalues of L+
P LG are denoted by λmax = λ1 ≥ λ2 ≥ · · · ≥ λn ≥ 1,
where L+
P denotes the Moore-Penrose pseudoinverse of LP .
B. Off-Tree Edge Embedding via Spectral Perturbation Analysis
1) Spectral Distortion of Spanning-Tree Sparsifiers.: It has been
shown that there are not too many large generalized eigenvalues
for spanning tree sparsifiers [24]: L+
P LG has at most k generalized
eigenvalues greater than stPk(G) , where stP (G) is the total stretch of
the spanning-tree subgraph P with respect the original graph G that
measures the spectral distortion due to the tree approximation. In fact,
recent research shows that every graph has a low-stretch spanning tree
(LSST) such that [18]:
O(m log n log log n) ≥ stP (G) = T r(L+
P LG ) =
III. S IMILARITY-AWARE S PECTRAL S PARSIFICATION OF G RAPHS
A. Overview of Our Approach
The overview of the proposed method for similarity-aware spectral
sparsification of large graphs has been summarized as follows. For
a given input graph, the following key procedures are involved in
the proposed algorithm flow: (a) low-stretch spanning tree extraction
based on its original graph Laplacian; (b) spectral (generalized
eigenvalue) embedding and criticality ranking of off-tree edges
leveraging the recent spectral perturbation analysis framework [7];
λi ≥ σ 2 ,
i=1
(4)
where m = |E| and n = |V |. As a result, it is possible to construct an
ultra-sparse yet spectrally similar sparsifier by recovering only a small
portion of “spectrally critical” off-tree edges to the spanning tree: for
log n
example, σ-similar spectral sparsifiers with O( m log nσlog
) off2
tree edges can be constructed in nearly-linear time [7].
2) Identification of Spectrally Critical Off-Tree Edges.: To identify
“spectrally critical” off-tree edges, the following generalized eigenvalue perturbation analysis is considered:
LG (ui + δui ) = (λi + δλi ) (LP + δLP ) (ui + δui ) ,
(3)
with u being the generalized eigenvector of λ. It can be further
shown that κ(LG , LP ) = σ 2 , which indicates that a smaller relative
condition number or σ 2 corresponds to a higher spectral similarity.
Spectral graph sparsification allows to construct spectrally-similar
subgraphs for mimicking the original graph so that they can be
leveraged for developing much faster numerical and graph algorithms.
For complete graphs, Ramanujan graphs are the best spectral sparsifiers [18], whereas for general graphs the linear-sized sparsifiers
(e.g. Twice-Ramanujan sparsifier) are the best but will be very costly
to compute in practice [1], [12]. Recent nearly-linear-time spectral
sparsification methods aim to find spectral sparsifier through the
following two phases [7], [10], [11], [18], [21], [23]: phase (a)
extracts a low-stretch spanning tree (LSST) P from the original graph
G, while phase (b) adds a few “spectrally critical” off-tree edges to
the tree to form the final nearly-linear-sized spectral sparsifier.
n
X
(5)
where a perturbation δLP is applied for the inclusion of extra offtree edges into LP , resulting in perturbed generalized eigenvalues
and eigenvectors λi + δλi and ui + δui for i = 1, ..., n, respectively.
The key to effective spectral sparsification is to identify the most
“spectrally-critical” off-tree edges that will result in the greatest
reduction in δλi . To this end, the following scheme for embedding
generalized eigenvalues into each off-tree edge is adopted in this work
[7]:
n
P
Step 1: Start with an initial random vector h0 =
αi ui , where ui
i=1
for i = 1, ..., n are the LP -orthogonal generalized eigenvectors of
T
L+
P LG that satisfy ui LP uj = 1 for i = j, and 0 for i 6= j;
Step 2: Perform t-step generalized power iterations with h0 to obtain
n
t
P
ht = L+
h0 =
αi λti ui ;
P LG
i=1
Step 3: Compute the Laplacian quadratic form for δLP,max =
LG − LP with ht :
n
2
P
QδLP,max (ht ) = ht T δLP,max ht =
αi λti (λi − 1)
=
P
(p,q)∈E\Es
wp,q
n
P
i=1
i=1
2
αi2 λ2t
ui T ep,q
i
=
P
wp,q v p,q 2 ,
(p,q)∈E\Es
(6)
where δLP,max includes all off-tree edges, and ep,q ∈ Rn is a
vector with only the p-th element being 1, the q-th element being
−1 and others being 0. In fact, (6) reflects the spectral similarity
between graphs P and G: larger QδLP,max (ht ) values indicates
C. “Spectrally-Unique” Off-Tree Edges
To simplify the following analysis, a “spectrally unique” off-tree
edge ei is defined to be the edge that connects to vertices pi and
qi , and only impacts a single large generalized eigenvalue λi , though
each off-tree edge can usually influence more than one eigenvalue or
eigenvector according to (6). Then the following truncated expansion
of the Laplacian quadratic form of (6) including the top k dominant “spectrally unique” off-tree edges for fixing the top k largest
eigenvalues of L+
P LG can be written as:
v1
“Resonant Frequency”
greater σ 2 and thus lower spectral similarity. More importantly,
(6) allows embedding generalized eigenvalues into the Laplacian
quadratic form of each off-tree edge and thus ranking off-tree edges
according to their “spectral criticality” levels: recovering the off-tree
edges with largest edge Joule heat wp,q v p,q 2 or |v p,q | values will
most significantly decrease the largest generalized eigenvalues. In
practice, using a small number (e.g. 0 < t < 3) of power iterations
will suffice for embedding purpose.
W1
v2
0
W 2 ! W1
W3 ! W2
v3
v4
W4 ! W3
Fig. 2. The eigenvectors (vi ) of increasing eigenvalues (τi ) for a line graph.
translation, modulation, dilation, and down-sampling to the graph
setting has been provided in [17].
The proposed similarity-aware spectral sparsification method aims
to maintain a simplest subgraph sufficient for preserving the slowlyk
k
2
P
P
varying or “low-frequency” signals on graphs, which therefore can be
QδLP,max (ht ) ≈
wpi ,qi αi2 λ2t
ui T epi ,qi =
αi2 λ2t
i
i (λi − 1).
i=1
i=1
regarded as a “low-pass” graph filter. In other words, such spectrally
(7) sparsified graphs will be able to preserve the eigenvectors associate
Since each off-tree edge can only impact one generalized eigenvalue, with low eigenvalues more accurately than high eigenvalues, and
we can assume epi ,qi = γi LP ui according to (6), leading to:
thus perform well for representing “low-frequency” graph signals
but poorly for highly-oscillating (signal) components due to the
γi , i = j,
T
uj epi ,qi =
(8) significant amount of missing edges.
0, i 6= j.
In practice, preserving the spectral (structural) properties of the
Consequently, the effective resistance of edge ei in P becomes:
original graph within the spectral sparsifier is becoming key to design
2 T
2
(9) of many fast numerical and graph-related algorithms [2], [11], [18],
Reefi f = eTpi ,qi L+
P epi ,qi = γi ui LP ui = γi ,
[19], [23]. For example, when using spectral sparsifier as a preconwhich immediately leads to:
ditioner in preconditioned conjugate gradient (PCG) iterations, the
k
k
convergence rate (the number of PCG iterations n < O(σ log )) is
X
X
ef f
αi2 λ2t
=
αi2 λ2t
QδLP,max (ht ) ≈
only governed by the spectral similarity (or relative condition number)
i wpi ,qi Rei
i (λi − 1).
i=1
i=1
for achieving a desired accuracy level , while in spectral graph
(10)
partitioning and data clustering tasks only the first few eigenvectors
Since the stretch of off-tree edge ei is computed by stP (ei ) =
associated with the smallest nontrivial eigenvalues of graph Laplacian
wpi ,qi Reefi f , (10) also indicates that stP (ei ) = λi − 1 holds for
are important [14], [20].
“spectrally-unique” off-tree edges. Consequently, the most “spectrally
critical” off-tree edges identified by (6) or (10) will have the largest E. Off-Tree Edge Filtering with Joule Heat
stretch values and therefore most significantly impact the largest
To only include the edges that are most critical for achieving the
eigenvalues of L+
P LG . In fact, (10) also can be considered as a
+
desired
spectral similarity level, we propose the following scheme for
randomized version of the trace T r(LP LG ) that is further scaled
truncating “spectrally-unique” off-tree edges based on each edge’s
for
each
λ
1.
up by a factor of λ2t
i
i
Joule heat. The theorem in [24] indicates that for a spanning-tree
D. Spectral Sparsification: A Graph Filter Perspective
preconditioner, there will be at most k generalized eigenvalues that
Although (6) and (10) provide a “spectral-criticality” ranking for are greater than stP (G)/k. Therefore, the following simple yet
each off-tree edge, it is not clear how many off-tree edges should be nearly worst-case generalized eigenvalue distribution can be assumed
selected for achieving a desired spectral similarity level. To this end, when the total stretch stP (G) is dominated by the largest eigenvalue
we introduce a simple yet effective off-tree edge filtering strategy λmax = λ1 = stP (G)/2:
motivated by recent graph signal processing problems [16], [17]. To
stP (G)
2λmax
more efficiently analyze signals on general undirected graphs, graph
λi =
=
,i ≥ 1
(11)
i
+
1
i+1
signal processing techniques have been extensively studied recently
[16], [17]. There is a clear analogy between traditional signal process- To most economically select the top-k “spectrally-unique” off-tree
ing based on classical Fourier analysis and graph signal processing: edges that will dominantly impact the top-k largest eigenvalues, the
1) the signals at different time points in classical Fourier analysis following sum of quadratic forms (Joule heat levels) can be computed
correspond to the signals at different nodes in an undirected graph; using (10) by performing t-step generalized power iterations using r
2) the more slowly oscillating functions in time domain correspond to multiple random vectors ht,1 , ..., ht,r , for λi 1:
the graph Laplacian eigenvectors associated with lower eigenvalues
2t+1
r X
k
X
2λmax
and more slowly varying (smoother) components across the graph.
QδLP,max (ht,1 , ..., ht,r ) ≈
(αi,j )2
. (12)
For example, the first few nontrivial eigenvectors associated with the
i+1
j=1 i=1
smallest non-zero eigenvalues of a line graph Laplacian have been
illustrated in Fig. 2, where the increasing eigenvalues correspond to Consider that we are ready to select the k-th “spectrally-unique” offincreasing oscillation frequencies in the line graph. A comprehensive tree edge for fixing the k-th largest eigenvalue λk so that the final
review of fundamental signal processing operations, such as filtering, spectral similarity will become σ 2 = λ̃λ̃max with λ̃max and λ̃min
min
denoting the largest and smallest eigenvalues of L+
P LG after adding
top-k “spectrally-unique” off-tree edges, then we have:
k = 2λmax /λ̃max − 1.
(13)
When using multiple random vectors for computing (12), it is
r
r
P
P
2
2
expected that
αk,j
≈
α1,j
, which allows us to define the
j=1
j=1
normalized edge Joule heat θk for the k-th “spectrally-unique” offtree edge through the following simplifications:
P
r
2
αk,j
2
2t+1
j=1
λ̃max 2t+1
heatλk
σ λ̃min
=
≈
.
θk =
r
P
λmax
heatλ1
λmax
2
α1,j
original and sparsified Laplacians, λmin can be easily computed. In
our method, we will only assign value 1 or 0 to each node in graphs
G and P for minimizing the following ratio between quadratic forms:
P
wpq
T
x(p)6=x(q),(p,q)∈E
x LG x
P
,
= min
λmin ≈ min
|x|6=0
|x|6=0
wpq
xT LP x
x(i)∈[0,1]
x(i)∈[0,1] x(p)6=x(q),(p,q)∈Es
(17)
which allows approximately computing λmin . To this end, we initialize all nodes with 0 values and will identify only a small portion of
nodes to be colored with 1 values so that the ratio between quadratic
forms can be minimized. An extreme case is to only select a single
node p to be colored with 1 such that:
j=1
(14)
In practice, (14) can be leveraged as a truncation threshold for achieving a desired spectral similarity level (relative condition number)
during spectral sparsification. Although the above filtering scheme
is derived for “spectrally-unique” off-tree edges, general off-tree
edges can also be filtered using similar strategies. Since adding
“spectrally-critical” off-tree edges will mainly impact the largest
generalized eigenvalues but not the smallest ones, we can always
assume λ̃min ≈ λmin and thus use the following edge truncation
scheme for filtering general off-tree edges to achieve a final spectral
similarity level σ: the edge (p, q) will be included into the sparsifier if
its normalized Joule heat value is greater than the truncation threshold
formulated below:
2
2t+1
heat(p,q)
σ λmin
> θσ ≈
,
(15)
θ(p,q) =
heatmax
λmax
where heatmax is the maximum Joule heat of all off-tree edges
computed by (6) with multiple initial random vectors.
F. Estimation of Extreme Eigenvalues λmax and λmin
To achieve the above edge filtering scheme, we need to compute
(15) that requires to estimate the extreme eigenvalues λmax and λmin
of L+
P LG . In this work, we propose the following efficient methods
for computing these extreme generalized eigenvalues.
1) Estimating λmax Using Power Iterations : It has been shown
that the largest eigenvalues of L+
P LG are well separated from
each other [24], which allows computing rather accurate largest
generalized eigenvalue (λmax or λ1 ) by performing only a small
number of generalized power iterations with an initial random vector.
Since generalized power iterations can converge at a geometric rate
determined by λ2 /λ1 , the error of the estimated largest generalized
eigenvalue will quickly drop to |λ2 /λ1 |k e0 after k iterations for an
initial error e0 . Our results show that only performing five to ten
iterations will suffice for obtaining a good estimation of λ1 .
2) Estimating λmin Using Node Coloring: Since the smallest
eigenvalues of L+
P LG are crowded together [24], using (shifted)
inverse power iterations may not be feasible due to the extremely slow
convergence rate considering the very poor separation of smallest
eigenvalues. To the extent of our knowledge, none of existing
eigenvalue decomposition methods can efficiently compute λmin . To
more efficiently estimate the smallest generalized eigenvalue without
loss of accuracy, this work exploits the following Courant-Fischer
theorem for generalized eigenvalue problems:
T
λmin = min
|x|6=0
x LG x
,
xT LP x
(16)
where x is also required to be orthogonal to the all-one vector to
avoid zero-value denominator. (16) indicates that if we can find a
vector x that minimizes the ratio between the quadratic forms of the
λmin ≈ min
p∈V
D(p)
,
Ds (p)
(18)
where D and Ds are the diagonal vectors satisfying D(p) =
LG (p, p) and Ds (p) = LP (p, p). (18) motivates to sort nodes
and subsequently color
according to the node degree ratio DD(p)
s (p)
a few (e.g. 10% of total) nodes that have the smallest degree ratios
with value 1. The resultant vector x can be used to compute λmin
using (16). The proposed approach only involves a partial sort of
nodes according to their node degree ratios and thus will be highly
scalable for large problems. Our experimental results on real-world
graphs show that the proposed method is highly efficient and accurate
for estimating the smallest generalized eigenvalues when compared
with existing generalized eigenvalue methods [15].
G. Iterative Graph Densification
To achieve more effective edge filtering for similarity-aware spectral graph sparsification, we propose to iteratively recover off-tree
edges to the sparsifier through an incremental graph densification
procedure. Each densification iteration adds a small portion of “filtered” off-tree edges to the latest spectral sparsifier, while the spectral
similarity is estimated to determine if more off-tree edges are needed.
The i-th graph densification iteration includes the following steps:
1) Estimate the spectral similarity by computing λmax and λmin
using the methods in Section III-F;
2) If the spectral similarity is not satisfactory, continue with the
following steps; otherwise, stop the densification procedure.
3) Perform t-step generalized power iterations with a few (e.g.
O(log |V |)) random vectors to compute the sum of Laplacian
quadratic forms similar to (12);
4) Rank each off-tree edge according to its “spectral criticality”
based on (6);
5) Use the truncation threshold (15) to recover top critical yet
dissimilar off-tree edges to the sparsifier.
IV. E XPERIMENTAL RESULTS
The proposed similarity-aware spectral sparsification method has
been implemented in C++ 1 . Test cases demonstrated in this paper
cover a great variety of graph Laplacian problems obtained from
the sparse matrix collection of the University of Florida [5]. All
experiments are performed using a single CPU core of a computing
platform running 64-bit RHEW 7.2 with a 2.67GHz 12-core CPU.
A. Effectiveness of Similarity-Aware Spectral Graph Sparsification
The results of edge ranking according to Joule heat levels computed
by one-step generalized power iteration using (6) have been shown
in Fig. 3 for two SDD matrices in [5]. The off-tree edges critical
for different spectral similarity levels can be easily identified using
1 Available
at https://sites.google.com/mtu.edu/zhuofeng-graphspar
G2_matrix
Thermal1
TABLE I
R ESULTS OF PCG SOLVER FOR GRAPH L APLACIANS IN [5].
V 2 | 500
V 2 | 500
Graphs
G3 circuit
thermal2
ecology2
tmt sym
paraboli fem
V 2 | 100
V 2 | 100
Not too many large generalized eigenvalues!
Fig.
3. Spectral edge
ranking and filtering by Joule heat of off-tree edges
5HVXOWVRI6SHFWUDO
*UDSK6SDUVLILFDWLRQ
for G2 circuit (left) and T hermal1 (right) test cases [5] with top off-tree
'PHVKJULG Q ZLWKXQLWHGJHZHLJKW
edges
highlighted in red rectangles.
± ZUWRIQRGHV H[WUDHGJHVDGGHGWRWKH³KDLUFRPE´VSDQQLQJWUHH
±
\
OPD[ LVUHGXFHGIURP(WR( ;UHGXFWLRQ
,QLWLDO-RXOH
+HDW0DS
\
×10 -3
1
6
[
4
0.5
2
[
7RWDO6WUHWFK 2 Q Q
)LQDO-RXOH 0
+HDW0DS200
0
200
200
100
100
0
0
|E|
3.0E6
3.7E6
2.0E6
2.2E6
1.6E6
|E50 |
|V |
1.11
1.14
1.14
1.21
1.22
N50
21
20
20
19
18
T50
20s
23s
16s
16s
16s
|E200 |
|V |
1.05
1.06
1.06
1.14
1.09
N200
37
36
40
38
38
T200
8s
9s
5s
4s
3s
converging to the required solution accuracy level in much fewer PCG
iterations, but need to retain more edges in the subgraphs and thus
longer time to compute (set up).
C. A Nearly-Linear Time Spectral Graph Partitioning Algorithm
Q u Q 'PHVK
|V |
1.6E6
1.2E6
1.0E6
0.7E6
0.5E6
200
100
100
0
0
Fig. 4. Joule heat distributions before (middle) and after (rightmost) adding
off-tree edges to the “hair comb” spanning tree (leftmost).
the proposed edge filtering scheme. It is also observed in Fig. 3 that
there are not many large eigenvalues of L+
P LG in both cases, which
matches the theoretical results [24].
We demonstrate the results of spatial Joule heat levels of off-tree
edges computed by (6) for a random vector using one-step generalized
power iteration (t = 1) for a “hair comb” spanning tree shown in
Fig. 4. It is not difficult to show that such a spanning tree can not
well match the top part of the original 2D grid, so the Joule heat
levels of off-tree edges at the top should be much greater than the
ones at the bottom part.
√ In fact,
√ it has been shown that a “hair comb”
spanning
tree
for
a
n
×
n 2D mesh will have a total stretch of
√
O(n n) that is mainly contributed by the off-tree edges with largest
stretch values near the top part of the mesh grid [13]. Using the
proposed similarity-aware spectral graph sparsification framework,
we are able to efficiently identify and recover the most “spectrally
critical” off-tree edges, thereby dramatically reducing the largest
generalized eigenvalues. For example, the “hair comb” spanning tree
of a 200 × 200 mesh grid can be dramatically improved in terms
of spectral similarity by recovering top 400 “spectrally critical” offtree edges: σ 2 is reduced from about 64, 000 to about 100 (640×
reduction), which can also be indicated by the Joule heat distributions
before and after adding these off-tree edges as illustrated in Fig. 4.
B. A Nearly-Linear Time Sparse SDD Matrix Solver
The spectral sparsifier obtained by the proposed similarity-aware
algorithm is also leveraged as a preconditioner in a PCG solver. The
RHS input vector b is generated randomly and the solver is set to
converge to an accuracy level ||Ax − b|| < 10−3 ||b|| for all test
cases. “|V |” and “|E|” denote the numbers of nodes and edges in
the original graph, whereas “|Eσ2 |”, “Nσ2 ” and “Tσ2 ” denote the
number of edges in the sparsifier, the number of PCG iterations
required for converging to the desired accuracy level, and the total
time of graph sparsification for achieving the spectral similarity of
σ 2 , respectively. As observed in all test cases, there are very clear
trade-offs between the graph density, computation time, and spectral
similarity for all spectral sparsifiers extracted using the proposed
method: sparsifiers with higher spectral similarities (smaller σ 2 ) allow
It has been shown that by applying only a few inverse power
iterations, approximate Fiedler vector (uf ) that correspond to the
smallest nonzero eigenvalue of the (normalized) graph Laplacian
matrix can be obtained for spectral graph partitioning purpose with
high solution quality [23]. Therefore, using the spectral sparsifiers
computed by the proposed spectral graph sparsification algorithm can
immediately accelerate the PCG solver for inverse power iterations,
leading to nearly-linear runtime algorithms for graph partitioning
problems [23]. In fact, if the spectral graph sparsifier can already
well approximate the original graph (with a small relative condition
number), its Fiedler vector can be directly used for partitioning the
original graph with high fidelity.
To validate the accelerated graph partitioning algorithm, sparse
matrices from [5] and a few larger 2D mesh graphs synthesized
with random edge weights have been tested. As shown in Table II,
the graphs associated with sparse matrices are partitioned based on
the sign of each node according to the approximate Fiedler vectors
computed by a few steps of inverse power iterations. The direct solver
[4] and the preconditioned iterative solver are invoked within each
inverse power iteration for updating the approximate Fiedler vectors
|V+ |
” denotes the ratio of nodes assigned
uf and ũf , respectively. “ |V−
|
with positive and negative signs according to the approximate Fiedler
vector, and “Rel.Err.” denotes the relative error of the proposed solver
|Vdif |
compared to the direct solver computed by |V
, where |Vdif |
|
denotes the number of nodes with different signs in uf and ũf . “ TD ”
(“ TI ”) and “MD ” (“ MI ”) denote the total solution time (excluding
spectral sparsification time) and memory cost of the direct (iterative)
algorithm. To gain high runtime efficiency and solution quality for
all test cases, we extract sparsifiers with σ 2 ≤ 200.
It can be observed that the proposed preconditioned spectral graph
partitioner only results in a very small portion of nodes (0.07% to
4%) assigned with different signs when comparing with the original
spectral graph partitioner, while achieving significant runtime and
memory savings (4−10×). The approximate Fiedler vector computed
by our fast solver for the test case “mesh 1M” is also illustrated in
Fig. 5, showing rather good agreement with the true solution.
D. Spectral Sparsification of Other Complex networks
As shown in Table III, a few finite element, protein, data and
social networks have been spectrally sparsified to achieve σ 2 ≈ 100
using the proposed similarity-aware method. “Ttot ” is the total time
for extracting the sparsifier, “ λλ̃1 ” denotes the reduction ratio of the
1
largest generalized eigenvalues before and after adding off-tree edges
o
s
into the spanning tree sparsifier, and Teig
(Teig
) denotes the time for
computing the first few eigenvectors of the original (sparsified) graph
Laplacians using Matlab’s eigs function. Since spectral sparsifiers
can well approximate the spectral (structural) properties of the
original graph, the sparsified graphs can be leveraged for accelerating
Fig. 5. The approximate Fiedler vector (left) and its magnitude error (right)
for “mesh 1M”.
TABLE II
R ESULTS OF SPECTRAL GRAPH PARTITIONING WITH THE PROPOSED
SIMILARITY- AWARE SPECTRAL GRAPH SPARSIFICATION ALGORITHM .
Test Cases
|V |
|V+ |
|V− |
TD (MD )
TI (MI )
Rel.Err.
G3 circuit
thermal2
ecology2
tmt sym
paraboli fem
mesh 1M
mesh 4M
mesh 9M
1.6E6
1.2E6
1.0E6
0.7E6
0.5E6
1.0E6
4.5E6
9.0E6
1.35
1.00
1.03
0.99
0.98
1.01
0.99
0.99
52.3s (2.3G)
13.0s (0.9G)
12.1s (0.7G)
10.2s (0.6G)
8.8s (0.4G)
10.2s (0.7G)
49.6s (3.0G)
138.5s (6.9G)
7.6s (0.3G)
3.0s (0.2G)
3.4s (0.2G)
1.9s (0.1G)
2.4s (0.1G)
1.7s (0.2G)
8.2s (0.7G)
13.3s (1.5G)
2.2E-2
6.8E-4
8.9E-3
2.1E-2
3.9E-2
3.3E-3
7.5E-3
7.8E-4
many numerical and graph-related tasks. For example, the original
“RCV-80NN” 80-nearest-neighbor graph can not be clustered on our
server with 50GB memory whereas the sparsified one only requires
a few minutes without sacrificing solution quality.
V. C ONCLUSIONS
For the first time, this paper introduces a similarity-aware spectral
graph sparsification framework that can be immediately leveraged
to develop fast numerical and graph-related algorithms. Motivated
by recent graph signal processing concepts, an iterative graph densification procedure based on spectral filtering of off-tree edges
has been proposed for producing ultra-sparse yet spectrally similar
graph sparsifiers, which enables highly flexible trade-offs between
computation complexity and solution quality in various numerical
and graph-related applications. Extensive experiment results have
confirmed the effectiveness and scalability of an iterative matrix
solver and a spectral graph partitioning algorithm for a variety of
large-scale, real-world graph problems, such as VLSI and finite
element analysis problems, as well as data and social networks.
R EFERENCES
[1] J. Batson, D. Spielman, and N. Srivastava. Twice-Ramanujan Sparsifiers.
SIAM Journal on Computing, 41(6):1704–1721, 2012.
[2] P. Christiano, J. Kelner, A. Madry, D. Spielman, and S. Teng. Electrical
flows, laplacian systems, and faster approximation of maximum flow in
undirected graphs. In Proc. ACM STOC, pages 273–282, 2011.
TABLE III
R ESULTS OF SIMILARITY- AWARE NETWORK SPARSIFICATION .
Test Cases
|V |
|E|
Ttot
|Es |
|E|
|Es|
λ1
λ̃1
o
s
Teig
(Teig
)
Reuters911
pdb1HYS
fe tooth
coAuthorsDBLP
RCV-80NN
1.3E4
3.6E4
7.8E4
3.0E5
1.9E5
1.5E5
2.2E6
4.5E5
1.0E6
1.2E7
0.2s
3.8s
3.0s
7.2s
46.5s
1.6E4
7.1E4
9.8E4
3.6E5
3.3E5
9×
31×
5×
3×
36×
21×
8, 981×
8, 050×
1, 364×
28, 410×
2.1s (0.4s)
10.5s (1.3s)
10.3s (1.6s)
2,047s (36s)
N/A (170s)
[3] M. B. Cohen, J. Kelner, J. Peebles, R. Peng, A. B. Rao, A. Sidford,
and A. Vladu. Almost-linear-time algorithms for markov chains and
new spectral primitives for directed graphs. In Proceedings of the 49th
Annual ACM SIGACT Symposium on Theory of Computing, pages 410–
419. ACM, 2017.
[4] T. Davis.
CHOLMOD: sparse supernodal Cholesky
factorization and update/downdate.
[Online]. Available:
http://www.cise.ufl.edu/research/sparse/cholmod/, 2008.
[5] T. Davis and Y. Hu. The university of florida sparse matrix collection.
ACM Trans. on Math. Soft. (TOMS), 38(1):1, 2011.
[6] M. Defferrard, X. Bresson, and P. Vandergheynst. Convolutional neural
networks on graphs with fast localized spectral filtering. In Advances in
Neural Information Processing Systems, pages 3844–3852, 2016.
[7] Z. Feng. Spectral graph sparsification in nearly-linear time leveraging
efficient spectral perturbation analysis. In Design Automation Conference
(DAC), 2016 53nd ACM/EDAC/IEEE, pages 1–6. IEEE, 2016.
[8] W. Fung, R. Hariharan, N. Harvey, and D. Panigrahi. A general
framework for graph sparsification. In Proc. ACM STOC, pages 71–
80, 2011.
[9] X. Hu, A. Lu, and X. Wu. Spectrum-based network visualization
for topology analysis. IEEE Computer Graphics and Applications,
33(1):58–68, 2013.
[10] A. Kolla, Y. Makarychev, A. Saberi, and S. Teng. Subgraph sparsification
and nearly optimal ultrasparsifiers. In Proc. ACM STOC, pages 57–66,
2010.
[11] I. Koutis, G. Miller, and R. Peng. Approaching Optimality for Solving
SDD Linear Systems. In Proc. IEEE FOCS, pages 235–244, 2010.
[12] Y. T. Lee and H. Sun. An SDP-based Algorithm for Linear-sized
Spectral Sparsification. In Proceedings of the 49th Annual ACM SIGACT
Symposium on Theory of Computing, STOC 2017, pages 678–687, New
York, NY, USA, 2017. ACM.
[13] R. Peng. Algorithm Design Using Spectral Graph Theory. PhD thesis,
Carnegie Mellon University, 2013.
[14] R. Peng, H. Sun, and L. Zanetti. Partitioning well-clustered graphs:
Spectral clustering works. In Proceedings of The 28th Conference on
Learning Theory (COLT), pages 1423–1455, 2015.
[15] Y. Saad. Numerical Methods for Large Eigenvalue Problems: Revised
Edition, volume 66. Siam, 2011.
[16] A. Sandryhaila and J. M. Moura. Big data analysis with signal processing
on graphs: Representation and processing of massive data sets with
irregular structure. IEEE Signal Processing Magazine, 31(5):80–90,
2014.
[17] D. I. Shuman, S. K. Narang, P. Frossard, A. Ortega, and P. Vandergheynst. The emerging field of signal processing on graphs: Extending high-dimensional data analysis to networks and other irregular
domains. IEEE Signal Processing Magazine, 30(3):83–98, 2013.
[18] D. Spielman. Algorithms, graph theory, and linear equations in Laplacian
matrices. In Proceedings of the International Congress of Mathematicians (ICM), volume 4, pages 2698–2722, 2010.
[19] D. Spielman and N. Srivastava. Graph sparsification by effective
resistances. SIAM Journal on Computing, 40(6):1913–1926, 2011.
[20] D. Spielman and S. Teng. Spectral partitioning works: Planar graphs and
finite element meshes. In Foundations of Computer Science (FOCS),
1996. Proceedings., 37th Annual Symposium on, pages 96–105. IEEE,
1996.
[21] D. Spielman and S. Teng. Nearly-linear time algorithms for graph
partitioning, graph sparsification, and solving linear systems. In Proc.
ACM STOC, pages 81–90, 2004.
[22] D. Spielman and S. Teng. Spectral sparsification of graphs. SIAM
Journal on Computing, 40(4):981–1025, 2011.
[23] D. Spielman and S. Teng. Nearly linear time algorithms for preconditioning and solving symmetric, diagonally dominant linear systems. SIAM
Journal on Matrix Analysis and Applications, 35(3):835–885, 2014.
[24] D. Spielman and J. Woo. A note on preconditioning by low-stretch
spanning trees. arXiv preprint arXiv:0903.2816, 2009.
[25] S.-H. Teng. Scalable algorithms for data and network analysis. Foundations and Trends R in Theoretical Computer Science, 12(1–2):1–274,
2016.
[26] Z. Zhao and Z. Feng. A spectral graph sparsification approach to scalable
vectorless power grid integrity verification. In Proceedings of the 54th
Annual Design Automation Conference 2017, page 68. ACM, 2017.
[27] Z. Zhao, Y. Wang, and Z. Feng. SAMG: Sparsified Graph Theoretic
Algebraic Multigrid for Solving Large Symmetric Diagonally Dominant
(SDD) Matrices. In Proceedings of the 36th International Conference
on Computer-Aided Design (ICCAD). ACM, 2017.
| 8cs.DS |
arXiv:1710.00402v2 [math.AC] 13 Oct 2017
Ext and local cohomology modules of face rings of
simplicial posets
Connor Sawaske
Department of Mathematics
University of Washington
Seattle, WA 98195-4350, USA
sawaske@math.washington.edu
October 17, 2017
Abstract
There are a large number of theorems detailing the homological properties of the
Stanley–Reisner ring of a simplicial complex. Here we attempt to generalize some of
these results to the case of a simplicial poset. By investigating the combinatorics of
certain modules associated with the face ring of a simplicial poset from a topological
viewpoint, we extend some results of Miyazaki and Gräbe to a wider setting.
1
Introduction
The Stanley–Reisner ring of a simplicial complex has been one of the most useful tools in
the study of combinatorial properties of simplicial polytopes, spheres, and manifolds for
decades. From the upper bound theorem to the g-theorem, this algebraic construction has
played a central role in many of the main results in this area of research. Its structure is
well-understood, and the translation of geometric properties of a simplicial complex into
algebraic statements about the associated ring is often clear and elegant. For an excellent
overview of the historical advances allowed by these rings, the reader is referred to [Sta96,
Sections II and III].
When shifting from simplicial complexes to simplicial posets, the situation changes
rather dramatically. Stanley defined the face ring of a simplicial poset as a generalization
of the Stanley–Reisner ring ([Sta91]); in the case that the poset in question is the face lattice of a simplicial complex, the two rings are isomorphic. Unfortunately, in this broader
scope the structure of the face ring becomes much less tangible. Though many classification theorems for simplicial complexes have analogs in the world of simplicial posets (see
[Sta91, Theorem 3.10] for an analog of the g-theorem, or [BK14], [Mas05], and [Mur13]
for further abstractions), many algebraic statements about Stanley–Reisner rings do not
have generalizations to face rings of simplicial posets.
1
The purpose of this paper is to provide such generalizations. In particular, we wish to
extend the homological properties of Stanley–Reisner rings established by Gräbe, Hochster,
Miyazaki, Reisner, and Schenzel to the world of simplicial posets and their face rings, in
the combinatorial spirit of the work begun by Duval in [Duv97]. It is worth noting that
these face rings are also squarefree modules; these objects have been studied from a various
viewpoints by Yanagawa (see, e.g., [Yan11b] and [Yan11a]). Furthermore, similar topics
have been investigated from a sheaf-theoretic viewpoint by Brun and Römer in [BR08]
and by Brun, Bruns, and Römer in [BBR07].
Before stating our main results, we will fix some notation. Let P be a simplicial poset
with vertex set V = {x1 , . . . , xn } and let supp(z) = {i : xi z} denote the support of
an element z ∈ P . Define A = k[x1 , . . . , xn ], let m be the irrelevant ideal (x1 , . . . , xn ),
and let mℓ be the ideal (xℓ1 , . . . , xℓn ). Lastly, let AP be the face ring of P ; we consider AP
as a Zn -graded A-module (we defer all definitions until the next section). Our first new
result is an extension of Miyazaki’s calculation of the graded pieces of the Ext -modules of
a Stanley–Reisner ring ([Miy89, Theorem 1]):
Theorem 3.1. Let P be a simplicial poset with vertex set V , and let α ∈ Zn . Set B =
{i : −ℓ < αi < 0}, C = {i : αi = −ℓ}, and D = {i : αi > 0}. If −ℓ ≤ αi for all i, then
M
e i−|B|−|C|−1 [0̂, zD ] × lkP (z)V rC
H
Ext iA (A/mℓ , AP )α ∼
=
supp(z)=B∪D
and Ext iA (A/mℓ , AP )α = 0 otherwise. In particular, if D 6= ∅ then Ext iA (A/mℓ , AP )α = 0.
By considering Hmi (AP ) as the direct limit of the Ext iA (A/mℓ , AP ) modules, we will
re-prove the result of Duval ([Duv97, Theorem 5.9]) that calculates the dimensions of the
graded pieces of Hmi (AP ). By tracing the A-module structure of Ext iA (A/mℓ , AP ) through
this direct limit, we will further be able to establish an extension of Gräbe’s result detailing
the corresponding structure of Hmi (AP ) ([Grä84, Theorem 2]). More precisely, our second
main contribution is:
Theorem 4.4. Let α = (α1 , . . . , αn ) ∈ Zn . Then
M
H i−1 (P, costP (w))
Hmi (AP )α ∼
=
supp(w)={j:αj 6=0}
if α ∈ Zn≤0 and Hmi (AP )α = 0 otherwise. Under these isomorphisms, the A-module
structure of Hmi (AP ) is given as follows. Let γ = α + deg(xj ). If αj < −1, then
·xj : Hmi (AP )α → Hmi (AP )γ corresponds to the direct sum of identity maps
M
M
H i−1 (P, costP (w)) →
H i−1 (P, costP (w)).
supp(w)={j:αj 6=0}
supp(w)={j:αj 6=0}
If αj = −1, then ·xj corresponds to the direct sum of maps
M
M
H i−1 (P, costP (w)) →
H i−1 (P, costP (z))
supp(w)={j:αj 6=0}
supp(z)={j:γj 6=0}
2
induced by the inclusions of pairs (P, costP (w r {xj })) → (P, costP (w)). If αj ≥ 0, then
·xj is the zero map.
The structure of the paper is as follows. In Section 2 we will review notation and some
foundational results, then examine the structure of AP in preparation for later computations. In Section 3 and Section 4 we will prove Theorems 3.1 and 4.4, respectively. In
Section 5 we will apply these results in a discussion of Cohen-Macaulay and Buchsbaum
properties. Finally, we will close with some additional comments in Section 6.
2
2.1
Preliminaries
Combinatorics and topology
A simplicial poset is a partially ordered finite set (P, ) satisfying the following two
properties:
(i) There exists an element 0̂ ∈ P satisfying 0̂ y for all y ∈ P .
(ii) For every element y ∈ P , the interval [0̂, y] is a Boolean lattice.
The rank of an element y of P is the maximal length among chains from 0̂ to y in P and is
denoted rkP (y). When the poset P is understood, we sometimes abbreviate this to rk(y).
The atoms of P are the elements of rank 1.
If y is an element of a poset P , then we define the link of y in P by
lkP (y) := {w ∈ P : y w}.
Note that y plays the role of 0̂ in lkP (y), and, more generally, if w ∈ lkP (y) then
rklkP (y) (w) = rkP (w) − rkP (y). Similarly, we define the contrastar of y in P by
costP (y) := {w ∈ P : y w}.
If the atoms of P are labeled x1 , . . . , xn , then we define the support of an element
y ∈ P to be the set
supp(y) := {i : xi y}.
Given B ⊆ supp(y), we denote by yB the unique face of P satisfying yB y and supp(yB ) =
B (the existence and uniqueness of yB follows from the fact that [0̂, y] is a Boolean lattice).
Similarly, if W ⊆ {1, . . . , n}, then define
PW := {y ∈ P : supp(y) ⊆ W }.
If w and y are two elements in P , then we define
M(w, y) := {z ∈ P : z is a minimal upper bound for w and y}.
3
The properties defining simplicial posets are preserved under some important operations and constructions, as outlined in the following proposition ([Bjö84, Proposition 2.6]).
Recall that if P and Q are posets, then a partial order may be placed on the product P ×Q
by defining (p1 , q1 ) (p2 , q2 ) if p1 p2 in P and q1 q2 in Q.
Proposition 2.1. Let P and Q be simplicial posets, and let y ∈ P . Then P × Q and
lkP (y) are both simplicial posets. If R is an order ideal of P (that is, R is a subset of P
that is closed under ), then R is a simplicial poset as well; in particular, costP (y) is a
simplicial poset.
There is a wealth of geometric meaning behind simplicial posets. In particular, every
simplicial poset may be realized as the face poset of a regular CW complex ([Bjö84, Section
3]). Such a complex is much like a simplicial complex in that the cells are simplices.
However, these more general objects may differ from simplicial complexes in that the
intersection of two faces needs only to be a (possibly empty) subcomplex of each of the
two faces, instead of a single simplex. For example, the following figure depicts the face
poset of a cone formed by two triangles with two pairs of edges identified and vertices
labeled 1, 2, and 3, with the apex of the cone being the vertex 1. The elements of the
poset have been labeled according to their support.
{1, 2}
{1}
{1, 2, 3}
{1, 2, 3}
{1, 3}
{2, 3}
{2}
{2, 3}
{3}
∅
From this geometric viewpoint, we refer to the atoms of P as vertices and to all
elements of P as faces. The dimension of a face y ∈ P is dim(y) := rk(y) − 1, and
the dimension of P is dim(P ) := max{dim(y) : y ∈ P }. If all maximal faces of P have
the same dimension, then we say that P is pure. We denote the regular CW complex
associated to the poset P by |P |, and we call it the geometric realization of P .
In addition to the geometric realization of P , there is another associated geometric
object known as the order complex. First, let P denote the poset P r {0̂}. The order
complex of P is the simplicial complex ∆(P ) whose elements are the chains of P . In
fact, ∆(P ) can be realized as the face poset of the barycentric subdivision of |P | ([LW69,
4
Theorem 2.1.7]). Furthermore, ∆(P × Q) isomorphic as a simplicial complex to the join
∆(P ) ∗ ∆(Q).
Let P be a simplicial poset with vertices {x1 , . . . , xn }, and let k be a field. Denote by
i
e
C (P ) the vector space over k whose basis elements correspond to elements of rank i + 1
in P . Next let O be a partial ordering of the indices {1, . . . , n} of the vertices of P such
that the restriction of O to the support of any one face is a total order; we call such an
ei (P ) → C
ei+1 (P ) by
ordering an orientation. Define a coboundary operator δ : C
δ(y) =
n
X
(−1)ui
i=1
X
z
z∈M(xi ,y)
where ui = |{j ∈ supp(y) : i <O j}|. We define the reduced cohomology groups
e• (P ) with
e i (P ) of P with coefficients in k to be the homology of the chain complex C
H
respect to the coboundary δ. If Q is a non-empty order ideal in P , then we define the
relative cohomology groups H i (P, Q) of the pair (P, Q) with coefficients in k to be the
e• (P )/C
e • (Q) with respect to the coboundary
homology of the chain complex C • (P, Q) := C
induced by δ. This definition of cohomology is identical to the cellular cohomology of the
CW complex |P | (see, e.g., [Hat02, Sections 2.2 and 3.1]). From this fact (along with its
relative analog), the equivalence of cellular, singular, and simplicial (co)homology imply
the following theorem.
Theorem 2.2. Let P be a simplicial poset, and let Q be a non-empty order ideal of P .
e i (|P |) (resp. H i (|P |, |Q|)) denotes the singular cohomology of |P | (resp. (|P |, |Q|))
If H
e i (∆(P )) (resp. H i (∆(P ), ∆(Q)))denotes the simplicial cohomology of ∆(P ) (resp.
and H
(∆(P ), ∆(Q)), then
e i (P ) ∼
e i (|P |) ∼
e i (∆(P ))
H
=H
=H
and
2.2
H i (P, Q) ∼
= H i (|P |, |Q|) ∼
= H i (∆(P ), ∆(Q)).
Algebra
Let P be a simplicial poset with atoms (or vertices) x1 , . . . , xn . Fix k to be an infinite
field and define A to be the polynomial ring over k whose variables are the atoms of P :
A := k[x1 , . . . , xn ].
Next, define k[P ] to be the polynomial ring whose variables are the elements of P :
k[P ] := k[y : y ∈ P ].
Let IP be the ideal of k[P ] defined as follows. For each pair of elements w and y in P , if
M(w, y) = ∅ then add wy as a generator of IP . If M(w, y) 6= ∅, then add
X
wy − (w ∧ y)
z
z∈M(w,y)
5
as a generator of IP . Finally, add 0̂ − 1 to IP . Define the face ring AP to be the quotient
k[P ]/IP .
Given some
elements y1 , . . . , yk in P and positive integers p1 , . . . , pk , let m be the
Q
monomial ki=1 yipi in AP . We define the support of m to be the set
supp(m) := {i : xi yj for some j}.
We say that the monomial m is standard if y1 ≻ y2 ≻ · · · ≻ yk . In this case, we call y1
the leading variable of m. Of vital importance for our arguments will be the following
fact ([Sta91, Lemma 3.4]).
Lemma 2.3. If P is a simplicial poset, then AP is an algebra with straightening law (ASL)
on P . In particular, the standard monomials form a basis for AP as a vector space over
k.
Note that A is naturally Zn -graded by setting deg(xi ) = ei , the i-th standard basis
vector of Zn . We may extend this grading to k[P ] by defining the degree of a variable y
to be
X
deg(y) :=
ei .
i∈supp(y)
Since IP is homogeneous with respect to this grading, the quotient AP is Zn -graded as
well. We will usually view AP as a Zn -graded A-module.
Given some degree α = (α1 , . . . , αn ) in Zn , we will also speak of its support, defined
by
supp(α) := {i : αi 6= 0}.
At times it will be necessary to focus only on the “negative” part of α. This negative part
is the degree α− ∈ Zn defined by
αi αi < 0
−
αi :=
0 αi ≥ 0.
2.3
The structure of AP
Although the face ring of a simplicial poset has many subtleties and does not generally
behave as well as the Stanley–Reisner ring of a simplicial complex, it does still have some
nice properties. We will now highlight two of these properties with the following lemmas
that will be critical to later computations.
Lemma 2.4. Let P be a simplicial poset, let w, y ∈ P , and let q be a positive integer.
Then in AP ,
X
(wy)q = (w ∧ y)q
zq .
z∈M(w,y)
6
Proof. Set M(w, y) = {z1 , . . . , zk }. If q = 1 or k ≤ 1, then the statement is immediate. If
k > 1 and q > 1, then write
!q
k
X
zi
(wy)q = (w ∧ y)q
i=1
= (w ∧ y)q
k
X
ziq + (w ∧ y)q
i=1
X
fℓ
ℓ
for some monomials fℓ , where each fℓ contains at least one product zi zj with zi 6= zj .
Suppose that zi and zj have some upper bound b in P . Then [0̂, b] would be a Boolean
lattice containing w and y, in which w and y have at least two distinct minimal upper
bounds. This is a contradiction, so zi zj = 0 in AP and each fℓ = 0 as well.
Our next lemma shows that the product of a standard monomial and a variable corresponding to a single vertex in AP may be expressed as a sum of “nice” standard monomials.
Q
Lemma 2.5. If m = ki=1 yipi is a standard monomial in Ap with y1 ≻ y2 ≻ · · · ≻ yk and
ℓ is a positive integer, then (xj )ℓ m can be written in Ap as
X
(xj )ℓ m =
mz
z∈M(xj ,y1 )
where mz is a standard monomial with leading variable z.
P
Proof. Write qr = ri=1 pi for 0 ≤ r ≤ k. We have three cases.
Case 1: j 6∈ supp(y1 ) and ℓ ≥ qk . We can re-write xℓj m as
xℓj m =
k
Y
(xj yi )pi
i=1
!
k
xℓ−q
.
j
(2.1)
By the definition of AP ,
xj y1 = (xj ∧ y1 )
X
z =
X
z.
z∈M(xj ,y1 ) z∈M(xj ,y1 )
since j 6∈ supp(y1 ). If M(xj , y1 ) = ∅, then the entire expression (2.1) is zero in AP . If
M(xj , y1) 6= ∅, then for each z ∈ M(xj , y1) let zi be the unique element of [0̂, z] with
vertices consisting of those in yi along with xj (note z = z1 ). By Lemma 2.4, we have
!
k
k
X p
X
Y
X
Y
pi
k
z p1 xℓ−qk
wi i
xℓj m =
xℓ−q
z
+
j
i
j
.
z∈M(xj ,y1 )
i=1
z∈M(xj ,y1 )
7
i=2
wi ∈M(xj ,yi )
wi 6=zi
However, all terms in the second summand on the right are zero in AP . Indeed, if z and
some wi 6= zi were to have some upper bound z ′ , then both zi and wi would be a minimal
′
upper bound for xj and yi in the boolean lattice
Q[0̂, z ], a contradiction.
k
pi
k
Writing mz for the standard monomial z p1
xℓ−q
, we see that
j
i=2 zi
X
(xj )ℓ m =
mz
z∈M(xj ,y1 )
as desired.
Case 2: j 6∈ supp(y1 ) and ℓ < qk . Let r be such that qr ≤ ℓ while qr+1 > ℓ. Now
re-write xℓj m as
!
!
k
r
Y
Y
p
−(l−q
)
p
r
r+1
xℓj m =
yi i .
(xj yi )pi (xj yr+1 )ℓ−qr yr+1
i=r+2
i=1
After replacing the products xj yi using the definition of AP and using the notation of Case
1, our “cross terms” cancel again and this simplifies to
!
!!
r
k
X
Y
Y
ℓ−qr pr+1 −(ℓ−qr )
xℓj m =
zipi zr+1
yr+1
yipi
z∈M(xj ,y)
Now if
i=1
r
Y
mz = z p1
i=2
then again we have
zipi
i=r+2
!
k
Y
ℓ−qr pr+1 −(ℓ−qr )
zr+1
yr+1
yipi
i=r+2
X
(xj )ℓ m =
!
mz .
z∈M(xj ,y1 )
Case 3: j ∈ supp(y1 ). If j ∈ supp(yi ) for all i, then m(xj )ℓ is already a standard
monomial and we are done. So, suppose that j ∈ supp(yi ) for 1 ≤ i ≤ r while j 6∈ supp(yi )
for i > r. Then by Cases 1 and 2,
(xℓj )
k
Y
yipi =
i=r+1
X
mz
z∈M(xj ,yr+1 )
for some standard monomials mz with leading coefficient z. Note that there is one unique
z ∈ M(xj , yr+1) such that z ≺ yr . Now write
!
!
r
r
Y
X
Y
pi
ℓ
′ .
(xj )m =
yi mz +
m
yipi
z
i=1
i=1
z ′ ∈M(xj ,yr+1 )
z ′ 6=z
The first term is a standard monomial with leading coefficient y1 , the unique element of
M(xj , y), while all other terms are zero as in Case 1.
8
3
3.1
Ext modules
Koszul complexes and an important basis
All of our results will crucially depend upon calculating the A-modules Ext iA (A/mℓ , AP ),
where mℓ is the ideal (xℓ1 , . . . , xℓn ). Let K•ℓ denote the Koszul complex of A with respect to
the sequence (xℓ1 , . . . , xℓn ) (for the construction of K•ℓ and some of its associated properties,
the reader is referred to [HH11, Section A.3]). We will view each Ktℓ as the direct sum
M
Ktℓ =
A(xℓi1 ∧ xℓi2 ∧ · · · ∧ xℓit ).
1≤i1 <i2 <···<it ≤n
That is, a free A-module basis for Ktℓ is indexed by sets F = {i1 , . . . , it } with 1 ≤ i1 <
i2 < · · · < it ≤ n. Given such an F , define x̂ℓF to be the corresponding basis element
xℓi1 ∧ xℓi2 ∧ · · · ∧ xℓit .
Since K•ℓ provides a projective resolution for A/mℓ , we can calculate Ext iA (A/mℓ , AP )
as the homology of the chain complex Hom A (K•ℓ , AP ). Using the Zn -grading of A and
AP , this complex splits into a direct sum of chain complexes of the form Hom A (K•ℓ , AP )α ,
where α ∈ Zn (it is easily verified that the differential preserves the graded pieces). Recall
(Lemma 2.3) that the standard monomials form a basis for AP as a vector space over k.
Then Hom A (Ktℓ , AP )α has a basis consisting of homomorphisms fF,m defined by
m, if G = F,
ℓ
fF,m (x̂G ) =
0, if G 6= F,
where m is a standard monomial that satisfies
deg(m) = deg(x̂ℓF ) + α.
3.2
The main theorem
Having examined the structure of AP as well as that of Hom A (K•ℓ , AP ), we are now prepared to prove the first of our main results: a topological description of the graded pieces
of the A-modules Ext iA (A/mℓ , AP ).
Theorem 3.1. Let P be a simplicial poset with vertex set V and let α ∈ Zn . Set B = {i :
−ℓ < αi < 0}, C = {i : αi = −ℓ}, and D = {i : αi > 0}. If −ℓ ≤ αi for all i, then
M
e i−| supp(α− )|−1 [0̂, zD ] × lkP (z)V rC
Ext iA (A/mℓ , AP )α ∼
H
=
supp(z)=B∪D
and Ext iA (A/mℓ , AP )α = 0 otherwise. In particular, if D 6= ∅ then Ext iA (A/mℓ , AP )α = 0.
Proof. Fix α, and without loss of generality assume that
D = {1, . . . , r}
and
9
C = {s, . . . , n}
(3.1)
for some r and s. Let K•ℓ denote the Koszul complex of A with respect to the sequence
(xℓ1 , . . . , xℓn ). As in Section 3.1, we will compute Ext iA (A/mℓ , AP )α by computing the
homology of the chain complex Hom A (K•ℓ , AP )α .
Suppose first that αi < −ℓ for some i. Then there cannot exist F = {i1 , . . . , it } and a
standard monomial m satisfying
deg(m) = deg(x̂ℓF ) + α.
That is, Hom A (Ktℓ , AP )α = 0 and hence Ext tA (A/mℓ , AP )α = 0. Assume from now on that
αi ≥ −ℓ for all i.
If fF,m is a basis element of Hom A (Ktℓ , AP )α , let y be the leading variable of m and set
w = yD∩F . Then we can consider fF,m as an element of the poset
[0̂, yD ] × lkP (yB∪D )V rC
where fF,m corresponds to the element (w, y), since B ∪ D ⊂ supp(y) and supp(y) ∩ C = ∅.
Note that the rank of (w, y) in this poset is given by
rk(w, y) = rk[0̂,yD ] (w) + rklkP (yB∪D ) (y)
= | supp(w)| + | supp(y)| − |B| − |D|
= |D ∩ F | + |F r C| + |D r F | − |B| − |D|
= |D ∩ F | + t − | supp(α− )| − |D ∩ F |
= t − | supp(α− )|,
where the fourth line follows because |F r C| − |B| = t − | supp(α− )| (recall that |F | = t)
and |D r F | − |D| = −|D ∩ F |. So, we can define a map
M
et−| supp(α− )|−1 0̂, zD × lkP (z)V rC
ϕ : Hom A (K ℓ , AP )α →
C
t
supp(z)=B∪D
by ϕ(fF,m ) = (w, y) under the correspondence above.
To prove that ϕ is actually an isomorphism of vector spaces, we would like to construct
an inverse. Given some fixed z with supp(z) = B ∪ D, let (w, y) be an element of [0, zD ] ×
lkP (z)V rC of rank t − supp(α− ). First, set
F = C ∪ (supp(y) r (D r supp(w))) .
Then it is immediate that
yF ∩D = ysupp(w) = w.
Furthermore, since rk(w, y) = t − | supp(α− )| = t − |B| − |C|, we have
rk(w, y) = rk[0̂,yD ] (w) + rklkP (yB∪D ) (y)
t − |B| − |C| = | supp(w)| + | supp(y)| − |B| − |D|
t = |C| + | supp(y)| − (|D| − | supp(w)|),
10
from which it follows that |F | = t. So, if we can construct a monomial m such that the
leading variable of m is y while deg(m) = deg(x̂ℓF ) + α, then we will have constructed the
desired inverse.
Let deg(x̂ℓF ) + α = δ = (δ1 , . . . , δn ), and define p0 = min{δi : δi > 0} and S0 = {i : δi ≥
p0 }. If j > 0 and pj−1 6= max{δi : 1 ≤ i ≤ n}, then define
pj = min{δi : δi > pj−1 }
and
Sj = {i : δi ≥ pj }.
Eventually, this process terminates at some point and pℓ = max{δi : 1 ≤ i ≤ n} for some
ℓ ≥ 0. Now define
ℓ
Y
p −p
p0
m=y
ySii i−1 .
i=1
We will verify that deg(m) = δ. Fix i. Then δi = pj for some j, which means that i ∈ Sk
for k ≤ j and i 6∈ Sk for k > j. Then xi ySk for k ≤ j, while xi 6 ySk for k > j, so
deg(ySk )i = 1 for k ≤ j and deg(ySk ) = 0 for k > j. Hence,
deg(m)i = p0 deg(y)i +
ℓ
X
(pm − pm−1 ) deg(ySm )i = p0 +
j
X
(pm − pm−1 ) = pj .
m=1
m=1
We have now shown that fF,m ∈ Hom A (Ktℓ , AP )α , and it is clear by construction that
ϕ(fF,m ) = (w, y). Hence, ϕ is a vector space isomorphism.
It remains to show that ϕ provides an isomorphism of chain complexes by verifying
that the differentials in each complex are preserved under ϕ. Let fF,m be a basis element
Q
in Hom A (K•ℓ , AP )α with m = y p ki=1 yipi . The differential on Hom A (K•ℓ , AP )α sends the
ℓ
homomorphism fF,m to a homomorphism dfF,m that maps x̂F ∪{j} ∈ Kt+1
to (−1)uj xℓj m ∈
AP for any j 6∈ F , where uj = |{i ∈ F : i < j}|. By Lemma 2.5, we can write each
(−1)uj xℓj m as
X
(−1)uj xℓj m = (−1)uj
mz ,
z∈M(xj ,y)
where mz is a standard monomial with leading variable z. Hence, each value of j contributes to dfF,m a term of the form
X
(−1)uj
fF ∪{j},mz ,
z∈M(xj ,y)
where mz is a standard monomial with leading variable z. Then we may write d(fF,m) as
X
X
(−1)uj
fF ∪{j},mz
d(fF,m) =
j6∈F
=
X
j∈(DrF )
z∈M(xj ,y)
(−1)uj
X
fF ∪{j},mz +
z∈M(xj ,y)
11
X
j6∈(D∪F )
(−1)uj
X
fF ∪{j},mz .
z∈M(xj ,y)
Choose z ∈ M(xj , y). If j ∈ D, then xj y already because D ⊂ supp(y) and thus
z = y. Furthermore, zD∩F = yD∩F regardless of the value of j and z(D∩F )∪{j} = y(D∩F )∪{j}
if j ∈ D. Hence, applying ϕ to this expression allows for the following simplifications:
X
X
X
X
(−1)uj
(−1)uj
(ϕ ◦ d)(fF,m) =
(z(D∩F )∪{j} , z) +
(zD∩F , z)
j∈(DrF )
=
X
j6∈(D∪F )
z∈M(xj ,y)
(−1)uj (y(D∩F )∪{j} , y) +
j∈(DrF )
X
j6∈(D∪F )
(−1)uj
z∈M(xj ,y)
X
(yD∩F , z) .
z∈M(xj ,y)
(3.2)
e• ([0̂, yD ]×lkP (yB∪D )V rC )
Our final step is to show that the differential on (yD∩F , y) in C
acts the same way. First note that if w is an atom of lkP (yB∪D )V rC , then w is an element of M(xj , yB∪D ) for some j 6∈ C. Arbitrarily label the elements of M(xj , yB∪D ) by
xj 1 , xj 2 , . . . , xj |M(xj ,yB∪D )| for each j. Then by definition, we can write
|M(xj ,yB∪D )|
X
X
X X
(−1)wjk
(yD∩F , z) . (3.3)
(d ◦ ϕ)(fF,m ) =
(−1)vj (y(D∩F )∪{j} , y) +
j6∈supp(y) k=1
j6∈C
j∈(DrF )
z∈M(xj k ,y)
where vj and wj i depend upon some choice of orientations for [0̂, yD ] and lkP (yB∪D )V rC ,
respectively. Since supp(y) ∪ C = D ∪ F , the right-hand sums in (3.2) and (3.3) run over
the same values of j. On the other hand,
|M(xj ,y)|
n
n
[
[
[
M(xj , y) =
M(xj k , y) ,
j=1
j=1
k=1
so if the characteristic of k is 2 then we are done. The construction of consistent orientations for fields of other characteristics is tedious, but unavoidable. If O1 is an ordering of
the vertices in [0̂, yD ] and O2 is an ordering of the vertices in lkP (yB∪D )V rC , then
vj = |{i ∈ supp(yD∩F ) : i <O1 j}|
and
wj k = |{i ∈ supp(y) : i <O2 j k }|.
In [0̂, yD ] we will choose O1 to be the natural ordering of the vertices. To show that uj = vj
in the left sums of (3.2) and (3.3), first choose j ∈ D r F . If i ∈ D ∩ F satisfies i <O1 j,
then i < j in the natural order and i ∈ F . On the other hand, if i ∈ F satisfies i < j, then
i ∈ D by our assumption (3.1) because j ∈ D. Hence, uj = vj as desired.
We now wish to construct an ordering of the j i ’s such that wj i = wj k for any pair i, k
and such that uj = wj i for all j 6∈ (D ∪ F ). Once this is done, we will have that
j ,y)|
X
X
X
X
X |M(x
(−1)uj
(−1)wji
(yD∩F , z)
(yD∩F , z) =
j6∈supp(y)
j6∈C
i=1
j6∈supp(y)
j6∈C
z∈M(xj i ,y)
12
z∈M(xj ,y)
as desired. We choose the partial ordering O2 of the vertices of lkP (yB∪D )V rC as follows.
First, order the indices of the the vertices of V r C as
r + 1 < r + 2 < · · · < s − 1 < 1 < 2 < · · · < n,
and call this order O3 . Next order the vertices of lkP y such that (j1 )i1 < (j2 )i2 if j1 < j2
under the order O3 . In P , we have supp(y) = {j1 , j2 , . . . , j| supp(y)| } for some ji ’s with
i supp(y)|
ji ∈ (V r C). Then in lkP (yB∪D ), we have supp(y) = {j1i1 , j2i2 , . . . , j| |supp(y)|
} for some jkik ’s
i
i
with jk ∈ (V r C). In particular, no pair jkk11 , jkk22 ∈ supp(y) has k1 = k2 . Hence, the
restriction of O2 to any face of lkP (yB∪D )V rC is a linear order and O2 provides a valid
orientation for lkP (yB∪D )V rC . Under this orientation,
wj i = |{k m ∈ supp(y) : k m <O2 j i }| = |{k m ∈ supp(y) : k <O3 j}|.
Now choose j 6∈ D ∪ F . If i ∈ F is such that i < j, then i cannot be in C on account
of our assumption (3.1) and thus i ∈ supp(y) in P . Then in lkP (yB∪D )V rC , ik ∈ supp(y)
for some single value k. Furthermore, i <O3 j, so ik <O2 j m for all m. That is, uj ≤ wj m
for all m.
On the other hand, if ik ∈ supp(y) (in lkP (yB∪D )V rC ) is such that ik <O2 j m , then
i <O3 j and hence i cannot be in D. Then i ∈ F , and i < j in the natural order. Hence,
wjm ≤ uj for all m, so wjm = uj for all j and m and ϕ provides an isomorphism of chain
complexes.
Lastly, if D 6= ∅, then the CW complex corresponding to [0̂, zD ] × lkP (z)V rC is the join
of a simplex with the CW complex corresponding to lkP (z)V rC ([Bjö84, Section 3]). In
particular, it is contractible; hence, Ext iA (A/mℓ , AP )α = 0.
4
4.1
Local cohomology modules
Reproving Duval’s result
Let m = (x1 , . . . , xn ) denote the irrelevant ideal of A. From Theorem 3.1 we can quickly
obtain the dimensions of the graded pieces of the local cohomology modules Hmi (AP ), a
calculation originally completed by Duval in [Duv97, Theorem 5.9]. Here we will reprove
the result using the techniques employed by Miyazaki in the simplicial complex case in
[Miy89], allowing for an important corollary.
Theorem 4.1. If P is a simplicial poset and α ∈ Zn is such that αi ≤ 0 for all i then
M
e i−| supp(α)|−1 (lkP (w)) ,
H
(4.1)
Hmi (AP )α ∼
=
supp(w)=supp(α)
and Hmi (AP )α = 0 otherwise.
13
Proof. Since the chain of ideals m, m2 , m3 , . . . is cofinal with the chain m, m2 , m3 , . . . (that
is, for all p, there exists q and r such that mr ⊆ mp and ms ⊆ mp ), we can compute
Hmi (AP )α as the direct limit of the modules Ext iA (A/mℓ , AP )α (see [ILL+ 07, Chapter 7]).
Since Ext iA (A/mℓ , AP )α = 0 for all ℓ if αi > 0 for some i, it follows that Hmi (AP )α = 0 if
αi > 0 for some i.
Suppose now that αi ≤ 0 for all i, and choose ℓ such that 1 − ℓ ≤ αi for all i.
Then it suffices to show that the following diagram commutes, in which the diagonal
maps are the isomorphisms of Theorem 3.1 and π
e is induced by the canonical projection
π : A/mℓ+1 → A/mℓ :
π
e
Ext iA (A/mℓ , AP )α
L
Ext iA (A/mℓ+1 , AP )α
e i−| supp(α)|−1 (lkP (w)) .
H
supp(w)=supp(α)
To begin, define a map π̂ : K•ℓ+1 → K•ℓ by
ℓ
ℓ
ℓ+1
π̂(xℓ+1
i1 ∧ · · · ∧ xit ) = (xi1 · · · xit )(xi1 ∧ · · · ∧ xit ).
Then the following diagram of projective resolutions commutes
K•ℓ+1
A/mℓ+1
0
π
π̂
K•ℓ
A/mℓ
0,
so dualizing produces another commutative diagram:
Hom A (K•ℓ+1 , AP )α
Hom A (A/mℓ+1 , AP )α
π̂ ∗
0
π∗
Hom A (K•ℓ , AP )α
Hom A (A/mℓ , AP )α
0.
When taking homology to compute Ext iA (A/mℓ , AP )α and Ext iA (A/mℓ+1 , AP )α , the map
π̂ ∗ may be used to compute π
e. That is, we only need to check that the following diagram
commutes:
π̂ ∗
Hom A (Ktℓ , AP )α
Hom A (Ktℓ+1 , AP )α
(4.2)
L et−| supp(α)|−1
C
(lkP (w)) .
supp(w)=supp(α)
14
Q
Let fF,m be a basis element for Hom A (Ktℓ , AP )α , where F = {i1 , . . . , it } and m = y p ki=1 yipi
et−| supp(α)|−1 (lkP (ysupp(α) ))
with y ≻ y1 ≻ y2 ≻ · · · ≻ yk . Then fF,m corresponds to the face y in C
under the map ϕ of Theorem 3.1.
On the other hand, fF,m ◦ π̂ is the homomorphism that sends x̂ℓ+1
to (xi1 · · · xit )m and
F
all other basis elements of Ktℓ+1 to zero. Note that xij ∈ supp(y) for 1 ≤ j ≤ t because
1 − ℓ ≤ αi ≤ 0 for all i. By iterating Lemma 2.5 t times, we have (xi1 · · · xit )m = m′ where
m′ is a standard monomial in AP with leading variable y. Hence, ϕ(fF,m ) = (ϕ ◦ π̂ ∗)(fF,m )
and Diagram (4.2) commutes.
The proof of Theorem 4.1 immediately provides the following corollary, which we will
use in Section 5 to characterize Buchsbaum simplicial posets.
Corollary 4.2. If α ∈ Zn is such that 1 − ℓ ≤ αi for all i, then the canonical map
Ext iA (A/mℓ , AP )α → Hmi (AP )α
is an isomorphism for all i.
4.2
The A-module structure of Hmi (AP )
The simplicial complex analog of Theorem 4.1 counting the dimensions of the local cohomology modules Hmi (k[∆]) of a Stanley–Reisner ring k[∆] was proved first by Hochster
([Sta96, Section II.4]). Later, Gräbe ([Grä84, Theorem 2]) examined the graded module
structure of Hmi (k[∆]). Here we will prove a similar result for Hmi (AP ). First, we will need
the following simplicial poset analog of [Grä84, Lemma, p.273].
Lemma 4.3. Let P be a simplicial poset and let y ∈ P . Then
e i−| supp(y)|−1 (lkP (y)) ∼
H
= H i−1 (P, costP (y)).
ei−| supp(y)| (lkP (y)) ∼
Proof. We clearly have an isomorphism of vector spaces C
= C i (P, costP (y))
under the map ϕ : w 7→ w. To show that this is an isomorphism of chain complexes, we
must proceed as in the proof of Theorem 3.1.
Label the vertices of P by x1 , . . . , xn , and let O1 be any ordering of the vertices of P
in which the vertices contained in supp(y) come last. If w is a vertex of lkP (y), then w
is an element of M(xj , y) for some j with j 6∈ supp(y). Arbitrarily label the elements of
M(xj , y) by xj 1 , xj 2 , . . . , xj |M(xj ,y)| . Let O2 be the partial ordering of the vertices of lkP (y)
given by setting j1k1 < j2k2 if j1 <O1 j2 .
If w ∈ (P, costP (y)), then write supp(w) = {j1 , . . . , jt }. In lkP (y), we have supp(w) =
k1
{j1 , . . . , jtkt } for some ki ’s. That is, O2 is a total order when restricted to supp(w) for any
w ∈ lkP (y), so O2 provides a valid orientation for lkP (y).
In C t (P, costP (y)),
X
X
(−1)uj
δ(w) =
z
j6∈supp(w)
z∈M(xj ,w)
15
where
uj = |{i ∈ supp(w) : i <O1 j}|.
et−| supp(y)| (lkP (y)),
On the other hand, in C
δ(w) =
|M(xj ,w)|
X
X
j6∈supp(w)
(−1)vjk
k=1
X
z
z∈M(xj k ,w)
where
vj k = |{i ∈ supp(w) : i <O2 j k }|.
Note that
[
M(xj , w) =
j6∈supp(w)
[
j6∈supp(w)
|M(x ,w)|
jk
[
M(xj k , w)
k=1
and that uj = vj k for fixed j and arbitrary k under our choice of ordering for O1 . The
result now follows.
With this lemma in hand, we are now prepared to prove our second main result.
Theorem 4.4. Let α ∈ Zn . Then
Hmi (AP )α ∼
=
M
H i−1 (P, costP (w))
supp(w)=supp(α)
if α ∈ Zn≤0 and Hmi (AP )α = 0 otherwise. Under these isomorphisms, the A-module
structure of Hmi (AP ) is given as follows. Let γ = α + deg(xj ). If αj < −1, then
·xj : Hmi (AP )α → Hmi (AP )γ corresponds to the direct sum of identity maps
M
H i−1 (P, costP (w)) →
supp(w)=supp(α)
M
H i−1 (P, costP (w)).
supp(w)=supp(γ)
If αj = −1, then ·xj corresponds to the direct sum of maps
M
H i−1 (P, costP (w)) →
supp(w)=supp(α)
M
H i−1 (P, costP (z))
supp(z)=supp(γ)
induced by the inclusions of pairs (P, costP (w r {xj })) → (P, costP (w)). If xj 6∈ supp(α),
then ·xj is the zero map.
Proof. The first statement is immediate, as it is a reformulation of Theorem 4.1 using the
isomorphisms of Lemma 4.3. Let ℓ be such that 1 − ℓ ≤ αi ≤ 0 for all i. By Corollary 4.2,
16
we only need to show that the multiplication maps above are valid for Ext tA (A/mℓ , AP )α ,
as the following diagram commutes:
Ext tA (A/mℓ , AP )α
∼
Hmt (AP )α
·xj
·xj
Ext tA (A/mℓ , AP )γ
∼
Hmt (AP )γ .
First note that if xj 6∈ supp(α), then γj = 1 so that Ext tA (A/mℓ , AP )γ = Hmt (AP )γ = 0.
Assume now that xj ∈ supp(α) and let fF,m be a basis element for Hom tA (A/mℓ , AP )α in
Q
which m = y p ki=1 yipi with y ≻ y1 ≻ y2 ≻ · · · ≻ yk .
et−| supp(α)|−1 (lkP (ysupp(α) )).
By the proof of Theorem 3.1, fF,m corresponds to the face y of C
On the other hand, xj · fF,m = fF,xj m and xj ≺ y, so by Lemma 2.5 we can write xj m as a
standard monomial m′ in which the leading variable of m′ is still y. Then fF,m′ corresponds
et−1 (P, costP (ysupp(γ) )).
to the face y of C
Suppose first that γj < 0. Then supp(α) = supp(γ), so
e t−1 (P, costP (ysupp(α) )) = C
et−1 (P, costP (ysupp(γ) ))
C
and ·xv corresponds to the identity map, as desired.
If γj = 0, then we have a map
·xv e t−1
et−1 (P, costP (ysupp(α) )) −
C
→ C (P, costP (ysupp(α)r{xj } ))
taking y to y, which is identical to the map induced by the inclusion of pairs
(P, costP (ysupp(α)r{xj } )) → (P, costP (ysupp(α) )).
5
Characterizing Buchsbaumness
Central to the historical study of simplicial complexes have been the notions of a complex
being Cohen-Macaulay or Buchsbaum. In analogy to the results of Reisner ([Rei76])
and Schenzel ([Sch81]) for simplicial complexes, we will say that a poset P is CohenMacaulay (over k) if its order complex is Cohen-Macaulay (over k) in the topological
sense; that is,
e i (lk
H
∆(P ) (F )) = 0
for all faces F ∈ ∆(P ) and all i < dim(lk∆(P ) (F )). We say that P is Buchsbaum (over k)
if it is pure and the link of every vertex of ∆(P ) is Cohen-Macaulay (over k).
There are also algebraic notions of Cohen-Macaulayness and Buchsbaumness. Let R
be a ring with maximal ideal m and let M be a noetherian R-module of dimension d. We
call a sequence of elements θ1 , . . . , θd ∈ m a system of parameters if M/(θ1 , . . . , θd )M
is a finite-dimensional vector space over k. If
(θ1 , . . . , θi−1 )M : θi = 0
17
for i = 1, . . . , d, then we say that θ1 , . . . , θd is an M-sequence. If every homogeneous
system of parameters for M is an M-sequence, then we say that M is a Cohen-Macaulay
R-module. If
(θ1 , . . . , θi−1 )M : θi = (θ1 , . . . , θi−1 )M : m
for i = 1, . . . , d, then we say that θ1 , . . . , θd is a weak M-sequence. If every homogeneous system of parameters for M is a weak M-sequence, then we call M a Buchsbaum R-module. We say that the ring R is Cohen-Macaulay (resp. Buchsbaum) if it is
Cohen-Macaulay (resp. Buchsbaum) as a module over itself. Stanley showed ([BMSW94,
Theorem 3.4, p. 314]) that the topological notion of Cohen-Macaulayness is tied to this
algebraic property (proofs of the forward direction appear frequently in the literature, but
we have had trouble finding a proof of the reverse; however, it does follow quickly from
Duval’s results, e.g., [Duv97, Corollary 6.1]):
Theorem 5.1. Let P be a simplicial poset. Then P is Cohen-Macaulay as a poset over
if and only if AP is a Cohen-Macaulay ring.
k
In fact, an additional characterization is that AP is a Cohen-Macaulay A-module.
Our next goal is obtaining an analogous characterization for the Buchsbaum property. To
prove our result, we will need the following homological characterization of Buchsbaumness
([SV86, Theorem I.3.7]):
Theorem 5.2. Let k be an infinite field and let M be a graded A-module with dim(M) > 0.
Then M is a Buchsbaum A-module if and only if the canonical maps
Ext iA (k, M) → Hmi (M)
are surjective for i < dim(M).
Along with Theorem 4.4 and Corollary 4.2, this result will provide a quick proof of the
following generalization of [SV86, Theorem II.2.4], the forward direction of which appeared
first as [NS09, Proposition 6.2]:
Theorem 5.3. Let P be a simplicial poset. Then P is Buchsbaum as a poset if and only
if AP is a Buchsbaum A-module.
Proof. Suppose first that P is a Buchsbaum poset. Under the isomorphisms
∼ ei
∼ ei
e i (lk
H
∆(P ) (F )) = H (∆(lkP (w))) = H (lkP (w))
(5.1)
for F a saturated chain in (0̂, w], by Theorem 4.1 along with the purity of P , Hmi (AP )
is concentrated in degree zero for i < dim(AP ). By Corollary 4.2, the canonical map
Ext iA (k, AP )0 → Hmi (AP )0 is always an isomorphism. Hence, AP satisfies the conditions
of Theorem 5.2 and AP is a Buchsbaum A-module.
18
Now suppose that AP is a Buchsbaum A-module, and for the moment assume that P
is pure. By [SV86, Corollary I.2.4], m · Hmi (AP ) = 0 for all i 6= dim AP . By Theorem 4.4
and purity, this is only possible if
e i (lkP (y)) = 0
H
for all y 6= ∅ in P and all i < dim(lkP (y)). Then once more, the isomorphisms in (5.1)
e i (lk
imply that H
∆(P ) (F )) = 0 for all faces F 6= ∅ of ∆(P ) and all i < dim(lk∆(P ) (F )). In
particular, lk∆(P ) (v) is Cohen-Macaulay for each vertex v of ∆(P ).
It remains to show that P is pure. Suppose to the contrary that there exists some
maximal face y ∈ P with dim(y) = k, while dim(P ) = d − 1 and k < d − 1. Note first that
e −1 (lkP (y)) 6= 0.
H
Let α ∈ Zn be such that supp(α) = supp(α) + ei = supp(y) for some ei . By Theorem
e −1 (lkP (y)) is a submodule of H k+1(AP )α . Furthermore, the multiplication map
4.1, H
m
·xi : Hmk+1(AP )α → Hmk+1(AP )α+ei is the identity map on this summand by Theorem 4.4.
Then m · Hmk+1(AP ) 6= 0 and k + 1 < dim(AP ), which contradicts [SV86, Corollary I.2.4].
Hence, P must be pure.
6
Comments and further applications
Our Theorem 4.4 detailing the A-module structure of Hmi (AP ) has the potential to extend
many results about simplicial complexes relying upon Gräbe’s theorem to the simplicial
poset case. For example, many of the theorems for face rings of simplicial complexes with
isolated singularities found in [MNS11], [NS12], and [Saw] appear likely to be true in this
greater generality.
On the other hand, Theorems 3.1 and 4.4 may be worth further specializing in their
own right. Simplicial posets come about frequently as quotients of simplicial complexes by
free group actions (see [GS84, Section 6]). It has been shown in [Saw17] that the Ext- and
local cohomology modules of a simplicial complex admitting a free action by an abelian
group inherit some nice symmetry properties. It would be interesting to see to what extent
these results continue to hold in the simplicial poset case.
Acknowledgments
The author was partially supported by NSF grants DMS-1361423 and DMS-1664865. I
am very grateful to Isabella Novik for many helpful suggestions and guidance during the
writing of this paper, as well as to Satoshi Murai for additional comments. Finally, I would
like to thank Tim Römer for pointing me toward his previous related research with Morton
Brun and Winfried Bruns.
19
References
[BBR07]
Morten Brun, Winfried Bruns, and Tim Römer. Cohomology of partially
ordered sets and local cohomology of section rings. Adv. Math., 208(1):210–
235, 2007.
[Bjö84]
A. Björner. Posets, regular CW complexes and Bruhat order. European J.
Combin., 5(1):7–16, 1984.
[BK14]
Jonathan Browder and Steven Klee. A classification of the face numbers of
Buchsbaum simplicial posets. Math. Z., 277(3-4):937–952, 2014.
[BMSW94] T. Bisztriczky, P. McMullen, R. Schneider, and A. Ivić Weiss, editors. Polytopes: abstract, convex and computational, volume 440 of NATO Advanced
Science Institutes Series C: Mathematical and Physical Sciences. Kluwer Academic Publishers Group, Dordrecht, 1994.
[BR08]
Morten Brun and Tim Römer. On algebras associated to partially ordered
sets. Math. Scand., 103(2):169–185, 2008.
[Duv97]
Art M. Duval. Free resolutions of simplicial posets. J. Algebra, 188(1):363–399,
1997.
[Grä84]
Hans-Gert Gräbe. The canonical module of a Stanley-Reisner ring. J. Algebra,
86(1):272–281, 1984.
[GS84]
A. M. Garsia and D. Stanton. Group actions of Stanley-Reisner rings and
invariants of permutation groups. Adv. in Math., 51(2):107–201, 1984.
[Hat02]
Allen Hatcher. Algebraic topology. Cambridge University Press, Cambridge,
2002.
[HH11]
Jürgen Herzog and Takayuki Hibi. Monomial ideals, volume 260 of Graduate
Texts in Mathematics. Springer-Verlag London, Ltd., London, 2011.
[ILL+ 07]
Srikanth B. Iyengar, Graham J. Leuschke, Anton Leykin, Claudia Miller, Ezra
Miller, Anurag K. Singh, and Uli Walther. Twenty-four hours of local cohomology, volume 87 of Graduate Studies in Mathematics. American Mathematical
Society, Providence, RI, 2007.
[LW69]
A.T. Lundell and S. Weingram. The topology of CW complexes. The University Series in Higher Mathematics. New York etc.: Van Nostrand Reinhold
Company. VIII, 216 p. (1969)., 1969.
[Mas05]
Mikiya Masuda. h-vectors of Gorenstein∗ simplicial posets.
194(2):332–344, 2005.
20
Adv. Math.,
[Miy89]
Mitsuhiro Miyazaki. Characterizations of Buchsbaum complexes. Manuscripta
Math., 63(2):245–254, 1989.
[MNS11]
Ezra Miller, Isabella Novik, and Ed Swartz. Face rings of simplicial complexes
with singularities. Math. Ann., 351(4):857–875, 2011.
[Mur13]
Satoshi Murai. Face vectors of simplicial cell decompositions of manifolds.
Israel J. Math., 195(1):187–213, 2013.
[NS09]
Isabella Novik and Ed Swartz. Socles of Buchsbaum modules, complexes and
posets. Adv. Math., 222(6):2059–2084, 2009.
[NS12]
Isabella Novik and Ed Swartz. Face numbers of pseudomanifolds with isolated
singularities. Math. Scand., 110(2):198–222, 2012.
[Rei76]
Gerald Allen Reisner. Cohen-Macaulay quotients of polynomial rings. Advances in Math., 21(1):30–49, 1976.
[Saw]
Connor Sawaske. Almost Buchsbaumness of some rings arising from complexes
with isolated singularities. J. Commut. Algebra. to appear.
[Saw17]
Connor Sawaske. Stanley-Reisner rings of Buchsbaum complexes with a free
action by an abelian group. ArXiv e-prints, June 2017. arXiv:1706.06506.
[Sch81]
Peter Schenzel. On the number of faces of simplicial complexes and the purity
of Frobenius. Math. Z., 178(1):125–142, 1981.
[Sta91]
Richard P. Stanley. f -vectors and h-vectors of simplicial posets. J. Pure Appl.
Algebra, 71(2-3):319–331, 1991.
[Sta96]
Richard P. Stanley. Combinatorics and commutative algebra, volume 41 of
Progress in Mathematics. Birkhäuser Boston, Inc., Boston, MA, second edition, 1996.
[SV86]
Jürgen Stückrad and Wolfgang Vogel. Buchsbaum rings and applications, volume 21 of Mathematische Monographien [Mathematical Monographs]. VEB
Deutscher Verlag der Wissenschaften, Berlin, 1986. An interaction between
algebra, geometry, and topology.
[Yan11a]
Kohji Yanagawa. Dualizing complex of the face ring of a simplicial poset. J.
Pure Appl. Algebra, 215(9):2231–2241, 2011.
[Yan11b]
Kohji Yanagawa. Higher Cohen-Macaulay property of squarefree modules and
simplicial posets. Proc. Amer. Math. Soc., 139(9):3057–3066, 2011.
21
| 0math.AC |
Overcoming Exploration in Reinforcement Learning
with Demonstrations
arXiv:1709.10089v2 [cs.LG] 25 Feb 2018
Ashvin Nair12 , Bob McGrew1 , Marcin Andrychowicz1 , Wojciech Zaremba1 , Pieter Abbeel12
Abstract— Exploration in environments with sparse rewards
has been a persistent problem in reinforcement learning (RL).
Many tasks are natural to specify with a sparse reward, and
manually shaping a reward function can result in suboptimal
performance. However, finding a non-zero reward is exponentially more difficult with increasing task horizon or action
dimensionality. This puts many real-world tasks out of practical
reach of RL methods. In this work, we use demonstrations
to overcome the exploration problem and successfully learn to
perform long-horizon, multi-step robotics tasks with continuous
control such as stacking blocks with a robot arm. Our method,
which builds on top of Deep Deterministic Policy Gradients and
Hindsight Experience Replay, provides an order of magnitude
of speedup over RL on simulated robotics tasks. It is simple
to implement and makes only the additional assumption that
we can collect a small set of demonstrations. Furthermore, our
method is able to solve tasks not solvable by either RL or
behavior cloning alone, and often ends up outperforming the
demonstrator policy.
I. I NTRODUCTION
RL has found significant success in decision making for
solving games, so what makes it more challenging to apply
in robotics? A key difference is the difficulty of exploration,
which comes from the choice of reward function and complicated environment dynamics. In games, the reward function
is usually given and can be directly optimized. In robotics,
we often desire behavior to achieve some binary objective
(e.g., move an object to a desired location or achieve a certain
state of the system) which naturally induces a sparse reward.
Sparse reward functions are easier to specify and recent work
suggests that learning with a sparse reward results in learned
policies that perform the desired objective instead of getting
stuck in local optima [1], [2]. However, exploration in an
environment with sparse reward is difficult since with random
exploration, the agent rarely sees a reward signal.
The difficulty posed by a sparse reward is exacerbated
by the complicated environment dynamics in robotics. For
example, system dynamics around contacts are difficult to
model and induce a sensitivity in the system to small errors.
Many robotics tasks also require executing multiple steps
successfully over a long horizon, involve high dimensional
control, and require generalization to varying task instances.
These conditions further result in a situation where the agent
so rarely sees a reward initially that it is not able to learn at
all.
All of the above means that random exploration is not a
tenable solution. Instead, in this work we show that we can
use demonstrations as a guide for our exploration. To test our
1
OpenAI,
2
University of California, Berkeley.
method, we solve the problem of stacking several blocks at
a given location from a random initial state. Stacking blocks
has been studied before in the literature [3], [4] and exhibits
many of the difficulties mentioned: long horizons, contacts,
and requires generalizing to each instance of the task. We
limit ourselves to 100 human demonstrations collected via
teleoperation in virtual reality. Using these demonstrations,
we are able to solve a complex robotics task in simulation
that is beyond the capability of both reinforcement learning
and imitation learning.
The primary contribution of this paper is to show that
demonstrations can be used with reinforcement learning
to solve complex tasks where exploration is difficult. We
introduce a simple auxiliary objective on demonstrations, a
method of annealing away the effect of the demonstrations
when the learned policy is better than the demonstrations,
and a method of resetting from demonstration states that
significantly improves and speeds up training policies. By
effectively incorporating demonstrations into RL, we shortcircuit the random exploration phase of RL and reach
nonzero rewards and a reasonable policy early on in training.
Finally, we extensively evaluate our method against other
commonly used methods, such as initialization with learning
from demonstrations and fine-tuning with RL, and show that
our method significantly outperforms them.
II. R ELATED W ORK
Learning methods for decision making problems such as
robotics largely divide into two classes: imitation learning
and reinforcement learning (RL). In imitation learning (also
called learning from demonstrations) the agent receives behavior examples from an expert and attempts to solve a task
by copying the expert’s behavior. In RL, an agent attempts
to maximize expected reward through interaction with the
environment. Our work combines aspects of both to solve
complex tasks.
Imitation Learning: Perhaps the most common form of
imitation learning is behavior cloning (BC), which learns a
policy through supervised learning on demonstration stateaction pairs. BC has seen success in autonomous driving
[5], [6], quadcopter navigation [7], locomotion [8], [9]. BC
struggles outside the manifold of demonstration data. Dataset
Aggregation (DAGGER) augments the dataset by interleaving
the learned and expert policy to address this problem of
accumulating errors [10]. However, DAGGER is difficult to
use in practice as it requires access to an expert during all
of training, instead of just a set of demonstrations.
Fig. 1: We present a method using reinforcement learning to solve the task of block stacking shown above. The robot starts
with 6 blocks labelled A through F on a table in random positions and a target position for each block. The task is to move each
block to its target position. The targets are marked in the above visualization with red spheres which do not interact with the
environment. These targets are placed in order on top of block A so that the robot forms a tower of blocks. This is a complex,
multi-step task where the agent needs to learn to successfully manage multiple contacts to succeed. Frames from rollouts of
the learned policy are shown. A video of our experiments can be found at: http://ashvin.me/demoddpg-website
Fundamentally, BC approaches are limited because they
do not take into account the task or environment. Inverse
reinforcement learning (IRL) [11] is another form of imitation learning where a reward function is inferred from the
demonstrations. Among other tasks, IRL has been applied
to navigation [12], autonomous helicopter flight [13], and
manipulation [14]. Since our work assumes knowledge of a
reward function, we omit comparisons to IRL approaches.
Reinforcement Learning: Reinforcement learning methods have been harder to apply in robotics, but are heavily
investigated because of the autonomy they could enable.
Through RL, robots have learned to play table tennis [15],
swing up a cartpole, and balance a unicycle [16]. A renewal
of interest in RL cascaded from success in games [17], [18],
especially because of the ability of RL with large function
approximators (ie. deep RL) to learn control from raw pixels.
Robotics has been more challenging in general but there
has been significant progress. Deep RL has been applied to
manipulation tasks [19], grasping [20], [21], opening a door
[22], and locomotion [23], [24], [25]. However, results have
been attained predominantly in simulation per high sample
complexity, typically caused by exploration challenges.
Robotic Block Stacking: Block stacking has been studied
from the early days of AI and robotics as a task that
encapsulates many difficulties of more complicated tasks we
want to solve, including multi-step planning and complex
contacts. SHRDLU [26] was one of the pioneering works,
but studied block arrangements only in terms of logic and
natural language understanding. More recent work on task
and motion planning considers both logical and physical
aspects of the task [27], [28], [29], but requires domainspecific engineering. In this work we study how an agent
can learn this task without the need of domain-specific
engineering.
One RL method, PILCO [16] has been applied to a simple
version of stacking blocks where the task is to place a
block on a tower [3]. Methods such as PILCO based on
learning forward models naturally have trouble modelling
the sharply discontinuous dynamics of contacts; although
they can learn to place a block, it is a much harder problem
to grasp the block in the first place. One-shot Imitation [4]
learns to stack blocks in a way that generalizes to new target
configurations, but uses more than 100,000 demonstrations
to train the system. A heavily shaped reward can be used
to learn to stack a Lego block on another with RL [30]. In
contrast, our method can succeed from fully sparse rewards
and handle stacking several blocks.
Combining RL and Imitation Learning: Previous work
has combined reinforcement learning with demonstrations.
Demonstrations have been used to accelerate learning on
classical tasks such as cart-pole swing-up and balance [31].
This work initialized policies and (in model-based methods)
initialized forward models with demonstrations. Initializing
policies from demonstrations for RL has been used for
learning to hit a baseball [32] and for underactuated swingup [33]. Beyond initialization, we show how to extract more
knowledge from demonstrations by using them effectively
throughout the entire training process.
Our method is closest to two recent approaches —
Deep Q-Learning From Demonstrations (DQfD) [34] and
DDPG From Demonstrations (DDPGfD) [2] which combine
demonstrations with reinforcement learning. DQfD improves
learning speed on Atari, including a margin loss which
encourages the expert actions to have higher Q-values than
all other actions. This loss can make improving upon the
demonstrator policy impossible which is not the case for
our method. Prior work has previously explored improving
beyond the demonstrator policy in simple environments by
introducing slack variables [35], but our method uses a
learned value to actively inform the improvement. DDPGfD
solves simple robotics tasks akin to peg insertion using
DDPG with demonstrations in the replay buffer. In contrast
to this prior work, the tasks we consider exhibit additional
difficulties that are of key interest in robotics: multi-step
behaviours, and generalization to varying goal states. While
previous work focuses on speeding up already solvable tasks,
we show that we can extend the state of the art in RL with
demonstrations by introducing new methods to incorporate
demonstrations.
III. BACKGROUND
A. Reinforcement Learning
We consider the standard Markov Decision Process framework for picking optimal actions to maximize rewards over
discrete timesteps in an environment E. We assume that the
environment is fully observable. At every timestep t, an agent
is in a state xt , takes an action at , receives a reward rt ,
and E evolves to state xt+1 . In reinforcement learning, the
agent must learn a policy at = π(xt ) toPmaximize expected
T
returns. We denote the return by Rt = i=t γ (i−t) ri where
T is the horizon that the agent optimizes over and γ is
a discount factor for future rewards. The agent’s objective
is to maximize expected return from the start distribution
J = Eri ,si ∼E,ai ∼π [R0 ].
A variety of reinforcement learning algorithms have been
developed to solve this problem. Many involve constructing
an estimate of the expected return from a given state after
taking an action:
Qπ (st , at ) = Eri ,si ∼E,ai ∼π [Rt |st , at ]
= Ert ,st+1 ∼E [rt + γ Eat+1 ∼π [Qπ (st+1 , at+1 )]]
(1)
(2)
We call Qπ the action-value function. Equation 2 is a
recursive version of equation 1, and is known as the Bellman equation. The Bellman equation allows for methods to
estimate Q that resemble dynamic programming.
B. DDPG
Our method combines demonstrations with one such
method: Deep Deterministic Policy Gradients (DDPG) [23].
DDPG is an off-policy model-free reinforcement learning
algorithm for continuous control which can utilize large
function approximators such as neural networks. DDPG
is an actor-critic method, which bridges the gap between
policy gradient methods and value approximation methods
for RL. At a high level, DDPG learns an action-value
function (critic) by minimizing the Bellman error, while
simultaneously learning a policy (actor) by directly maximizing the estimated action-value function with respect to
the parameters of the policy.
Concretely, DDPG maintains an actor function π(s) with
parameters θπ , a critic function Q(s, a) with parameters θQ ,
and a replay buffer R as a set of tuples (st , at , rt , st+1 )
for each transition experienced. DDPG alternates between
running the policy to collect experience and updating the
parameters. Training rollouts are collected with extra noise
for exploration: at = π(s) + N , where N is a noise process.
During each training step, DDPG samples a minibatch
consisting of N tuples from R to update the actor and critic
networks. DDPG minimizes the following loss L w.r.t. θQ
to update the critic:
yi = ri + γQ(si+1 , π(si+1 ))
(3)
L=
1 X
(yi − Q(si , ai |θQ ))2
N i
(4)
The actor parameters θπ are updated using the policy
gradient:
1 X
∇a Q(s, a|θQ )|s=si ,a=π(s) ∇θπ π(s|θπ )|si
∇θ π J =
N i
(5)
To stabilize learning, the Q value in equation 3 is usually
computed using a separate network (called the target network) whose weights are an exponential average over time
of the critic network. This results in smoother target values.
Note that DDPG is a natural fit for using demonstrations. Since DDPG can be trained off-policy, we can use
demonstration data as off-policy training data. We also take
advantage of the action-value function Q(s, a) learned by
DDPG to better use demonstrations.
C. Multi-Goal RL
Instead of the standard RL setting, we train agents with
parametrized goals, which lead to more general policies
[36] and have recently been shown to make learning with
sparse rewards easier [1]. Goals describe the task we expect
the agent to perform in the given episode, in our case
they specify the desired positions of all objects. We sample
the goal g at he beginning of every episode. The function
approximators, here π and Q, take the current goal as an
additional input.
D. Hindsight Experience Replay (HER)
To handle varying task instances and parametrized goals,
we use Hindsight Experience Replay (HER) [1]. The key
insight of HER is that even in failed rollouts where no
reward was obtained, the agent can transform them into
successful ones by assuming that a state it saw in the rollout
was the actual goal. HER can be used with any off-policy
RL algorithm assuming that for every state we can find a
goal corresponding to this state (i.e. a goal which leads to a
positive reward in this state).
For every episode the agent experiences, we store it in
the replay buffer twice: once with the original goal pursued
in the episode and once with the goal corresponding to the
final state achieved in the episode, as if the agent intended
on reaching this state from the very beginning.
IV. M ETHOD
Our method combines DDPG and demonstrations in several ways to maximally use demonstrations to improve
learning. We describe our method below and evaluate these
ideas in our experiments.
A. Demonstration Buffer
First, we maintain a second replay buffer RD where we
store our demonstration data in the same format as R. In
each minibatch, we draw an extra ND examples from RD
to use as off-policy replay data for the update step. These
examples are included in both the actor and critic update.
This idea has been introduced in [2].
B. Behavior Cloning Loss
Second, we introduce a new loss computed only on the
demonstration examples for training the actor.
LBC =
ND
X
2
kπ(si |θπ ) − ai k
(6)
i=1
This loss is a standard loss in imitation learning, but we show
that using it as an auxiliary loss for RL improves learning
significantly. The gradient applied to the actor parameters θπ
is:
λ1 ∇θπ J − λ2 ∇θπ LBC
(7)
(Note that we maximize J and minimize LBC .) Using this
loss directly prevents the learned policy from improving
significantly beyond the demonstration policy, as the actor is
always tied back to the demonstrations. Next, we show how
to account for suboptimal demonstrations using the learned
action-value function.
C. Q-Filter
We account for the possibility that demonstrations can be
suboptimal by applying the behavior cloning loss only to
states where the critic Q(s, a) determines that the demonstrator action is better than the actor action:
LBC =
ND
X
2
kπ(si |θπ ) − ai k 1Q(si ,ai )>Q(si ,π(si ))
(8)
i=1
The gradient applied to the actor parameters is as in equation
7. We label this method using the behavior cloning loss and
Q-filter “Ours” in the following experiments.
D. Resets to demonstration states
To overcome the problem of sparse rewards in very long
horizon tasks, we reset some training episodes using states
and goals from demonstration episodes. Restarts from within
demonstrations expose the agent to higher reward states during training. This method makes the additional assumption
that we can restart episodes from a given state, as is true in
simulation.
To reset to a demonstration state, we first sample a
demonstration D = (x0 , u0 , x1 , u1 , ...xN , uN ) from the set
of demonstrations. We then uniformly sample a state xi
from D. As in HER, we use the final state achieved in the
demonstration as the goal. We roll out the trajectory with the
given initial state and goal for the usual number of timesteps.
At evaluation time, we do not use this procedure.
We label our method with the behavior cloning loss, Qfilter, and resets from demonstration states as “Ours, Resets”
in the following experiments.
V. E XPERIMENTAL S ETUP
A. Environments
We evaluate our method on several simulated MuJoCo [37]
environments. In all experiments, we use a simulated 7-DOF
Fetch Robotics arm with parallel grippers to manipulate one
or more objects placed on a table in front of the robot.
The agent receives the positions of the relevant objects
on the table as its observations. The control for the agent is
continuous and 4-dimensional: 3 dimensions that specify the
desired end-effector position1 and 1 dimension that specifies
the desired distance between the robot fingers. The agent is
controlled at 50Hz frequency.
We collect demonstrations in a virtual reality environment.
The demonstrator sees a rendering of the same observations
as the agent, and records actions through a HTC Vive
interface at the same frequency as the agent. We have
the option to accept or reject a demonstration; we only
accept demonstrations we judge to be mostly correct. The
demonstrations are not optimal. The most extreme example
is the “sliding” task, where only 7 of the 100 demonstrations
are successful, but the agent still sees rewards for these
demonstrations with HER.
B. Training Details
To train our models, we use Adam [38] as the optimizer
with learning rate 10−3 . We use N = 1024, ND = 128, λ1 =
10−3 , λ2 = 1.0/ND . The discount factor γ is 0.98. We use
100 demonstrations to initialize RD . The function approximators π and Q are deep neural networks with ReLU activations and L2 regularization with the coefficient 5×10−3 . The
final activation function for π is tanh, and the output value
is scaled to the range of each action dimension. To explore
during training, we sample random actions uniformly within
the action space with probability 0.1 at every step, and the
noise process N is uniform over ±10% of the maximum
value of each action dimension. Task-specific information,
including network architectures, are provided in the next
section.
C. Overview of Experiments
We perform three sets of experiments. In Sec. VI, we
provide a comparison to previous work. In Sec. VII we
solve block stacking, a difficult multi-step task with complex
contacts that the baselines struggle to solve. In Sec. VIII
we do ablations of our own method to show the effect of
individual components.
VI. C OMPARISON W ITH P RIOR W ORK
A. Tasks
We first show the results of our method on the simulated
tasks presented in the Hindsight Experience Replay paper
[1]. We apply our method to three tasks:
1) Pushing. A block placed randomly on the table must
be moved to a target location on the table by the robot
(fingers are blocked to avoid grasping).
2) Sliding. A puck placed randomly on the table must be
moved to a given target location. The target is outside
the robot’s reach so it must apply enough force that
the puck reaches the target and stops due to friction.
3) Pick-and-place. A block placed randomly on the table
must be moved to a target location in the air. Note
1 In
the 10cm x 10cm x 10cm cube around the current gripper position
Pushing
1.0
0.4
0.2
0.8
Success Rate
0.6
0.6
0.4
0.2
0.0
2M
4M
6M
Timesteps
8M
10M
0.6
0.4
0.2
0.0
0M
Pick and Place
1.0
Ours
HER
BC
0.8
Success Rate
0.8
Success Rate
Sliding
1.0
0.0
0M
2M
4M
6M
8M
10M
0M
Timesteps
2M
4M
6M
8M
10M
Timesteps
Fig. 2: Baseline comparisons on tasks from [1]. Frames from the learned policy are shown above each task. Our method
significantly outperforms the baselines. On the right plot, the HER baseline always fails.
that the original paper used a form of initializing from
favorable states to solve this task. We omit this for our
experiment but discuss and evaluate the initialization
idea in an ablation.
As in the prior work, we use a fully sparse reward for this
task. The agent is penalized if the object is not at its goal
position:
(
0,
if ||xi − gi || < δ
rt =
(9)
−1, otherwise
where the threshold δ is 5cm.
B. Results
Fig. 2 compares our method to HER without demonstrations and behavior cloning. Our method is significantly faster
at learning these tasks than HER, and achieves significantly
better policies than behavior cloning does. Measuring the
number of timesteps to get to convergence, we exhibit a 4x
speedup over HER in pushing, a 2x speedup over HER in
sliding, and our method solves the pick-and-place task while
HER baseline cannot solve it at all.
The pick-and-place task showcases the shortcoming of RL
in sparse reward settings, even with HER. In pick-and-place,
the key action is to grasp the block. If the robot could manage
to grasp it a small fraction of the time, HER discovers
how to achieve goals in the air and reinforces the grasping
behavior. However, grasping the block with random actions
is extremely unlikely. Our method pushes the policy towards
demonstration actions, which are more likely to succeed.
In the HER paper, HER solves the pick-and-place task
by initializing half of the rollouts with the gripper grasping
the block. With this addition, pick-and-place becomes the
easiest of the three tasks tested. This initialization is similar
in spirit to our initialization idea, but takes advantage of the
fact that pick-and-place with any goal can be solved starting
from a block grasped at a certain location. This is not always
true (for example, if there are multiple objects to be moved)
and finding such a keyframe for other tasks would be difficult, requiring some engineering and sacrificing autonomy.
Instead, our method guides the exploration towards grasping
the block through demonstrations. Providing demonstrations
does not require expert knowledge of the learning system,
which makes it a more compelling way to provide prior
information.
VII. M ULTI -S TEP E XPERIMENTS
A. Block Stacking Task
To show that our method can solve more complex tasks
with longer horizon and sparser reward, we study the task
of block stacking in a simulated environment as shown in
Fig. 1 with the same physical properties as the previous
experiments. Our experiments show that our approach can
solve the task in full and learn a policy to stack 6 blocks
with demonstrations and RL. To measure and communicate
various properties of our method, we also show experiments
on stacking fewer blocks, a subset of the full task.
We initialize the task with blocks at 6 random locations
x1 ...x6 . We also provide 6 goal locations g1 ...g6 . To form a
tower of blocks, we let g1 = x1 and gi = gi−1 + (0, 0, 5cm)
for i ∈ 2, 3, 4, 5.
By stacking N blocks, we mean N blocks reach their
target locations. Since the target locations are always on top
of x1 , we start with the first block already in position. So
stacking N blocks involves N −1 pick-and-place actions. To
solve stacking N , we allow the agent 50 ∗ (N − 1) timesteps.
This means that to stack 6 blocks, the robot executes 250
actions or 5 seconds.
We recorded 100 demonstrations to stack 6 blocks, and
use subsets of these demonstrations as demonstrations for
stacking fewer blocks. The demonstrations are not perfect;
they include occasionally dropping blocks, but our method
can handle suboptimal demonstrations. We still rejected more
than half the demonstrations and excluded them from the
demonstration data because we knocked down the tower of
blocks when releasing a block.
B. Rewards
Two different reward functions are used. To test the
performance of our method under fully sparse reward, we
Ours
Stack 2, Sparse
Stack 3, Sparse
Stack 4, Sparse
Stack 4, Step
Stack 5, Step
Stack 6, Step
99%
99%
1%
91%
49%
4%
Ours,
Resets
97%
89%
54%
73%
50%
32%
BC
HER
65%
1%
0%
-
0%
0%
0%
-
BC+
HER
65%
1%
0%
-
Fig. 3: Comparison of our method against baselines. The
value reported is the median of the best performance (success
rate) of all randomly seeded runs of each method.
reward the agent only if all blocks are at their goal positions:
rt = min 1||xi −gi ||<δ
i
(10)
The threshold δ is the size of a block, 5cm. Throughout the
paper we call this the “sparse” reward.
To enable solving the longer horizon tasks of stacking 4
or more blocks, we use the “step” reward :
X
rt = −1 +
1||xi −gi ||<δ
(11)
i
Note the step reward is still very sparse; the robot only
sees the reward change when it moves a block into its
target location. We subtract 1 only to make the reward more
interpretable, as in the initial state the first block is already
at its target.
Regardless of the reward type, an episode is considered
successful for computing success rate if all blocks are at
their goal position in their final state.
C. Network architectures
We use a 4 layer networks with 256 hidden units per layer
for π and Q for the HER tasks and stacking 3 or fewer
blocks. For stacking 4 blocks or more, we use an attention
mechanism [39] for the actor and a larger network. The
attention mechanism uses a 3 layer network with 128 hidden
units per layer to query the states and goals with one shared
head. Once a state and goal is extracted, we use a 5 layer
network with 256 hidden units per layer after the attention
mechanism. Attention speeds up training slightly but does
not change training outcomes.
D. Baselines
We include the following methods to compare our method
to baselines on stacking 2 to 6 blocks. 2
Ours: Refers to our method as described in section IV-C.
Ours, Resets: Refers to our method as described in section
IV-C with resets from demonstration states (Sec. IV-D).
BC: This method uses behavior cloning to learn a policy.
Given the set of demonstration transitions RD , we train the
2 Because of computational constraints, we were limited to 5 random seeds
per method for stacking 3 blocks, 2 random seeds per method for stacking 4
and 5 blocks, and 1 random seed per method for stacking 6 blocks. Although
we are careful to draw conclusions from few random seeds, the results are
consistent with our collective experience training these models. We report
the median of the random seeds everywhere applicable.
Stack 3, Sparse Reward
1.0
Ours
Ours, Resets
No Q-Filter
No BC
No HER
0.8
Success Rate
Task
0.6
0.4
0.2
0.0
0M
50M
100M
150M
200M
250M
300M
350M
400M
Timesteps
Fig. 4: Ablation results on stacking 3 blocks with a fully
sparse reward. We run each method 5 times with random
seeds. The bold line shows the median of the 5 runs while
each training run is plotted in a lighter color. Note “No
HER” is always at 0% success rate. Our method without
resets learns faster than the ablations. Our method with resets
initially learns faster but converges to a worse success rate.
policy π by supervised learning. Behavior cloning requires
much less computation than RL. For a fairer comparison,
we performed a large hyperparameter sweep over various
networks sizes, attention hyperparameters, and learning rates
and report the success rate achieved by the best policy found.
HER: This method is exactly the one described in Hindsight
Experience Replay [1], using HER and DDPG.
BC+HER: This method first initializes a policy (actor) with
BC, then finetunes the policy with RL as described above.
E. Results
We are able to learn much longer horizon tasks than
the other methods, as shown in Fig. 3. The stacking task
is extremely difficult using HER without demonstrations
because the chance of grasping an object using random
actions is close to 0. Initializing a policy with demonstrations
and then running RL also fails since the actor updates depend
on a reasonable critic and although the actor is pretrained,
the critic is not. The pretrained actor weights are therefore
destroyed in the very first epoch, and the result is no better
than BC alone. We attempted variants of this method where
initially the critic was trained from replay data. However,
this also fails without seeing on-policy data.
The results with sparse rewards are very encouraging. We
are able to stack 3 blocks with a fully sparse reward without
resetting to the states from demonstrations, and 4 blocks with
a fully sparse reward if we use resetting. With resets from
demonstration states and the step reward, we are able to learn
a policy to stack 6 blocks.
VIII. A BLATION E XPERIMENTS
In this section we perform a series of ablation experiments
to measure the importance of various components of our
method. We evaluate our method on stacking 3 to 6 blocks.
We perform the following ablations on the best performing
of our models on each task:
No BC Loss: This method does not apply the behavior
cloning gradient during training. It still has access to demonstrations through the demonstration replay buffer.
Stack 4, Step Reward
0.5
0.6
0.4
0.35
Stack 6, Step Reward
Ours
Ours, Resets
0.30
0.4
0.3
0.2
0.25
0.20
0.15
0.10
0.2
0.1
0.0
0.05
0.0
100M
200M
300M
400M
500M
600M
700M
800M
0M
3.0
2.5
Reward
2.0
1.5
1.0
500M
1000M
1500M
4.0
4.0
3.5
3.5
3.0
3.0
2.5
2.5
2.0
1.5
1.0
0.5
100M
200M
300M
400M
500M
Timesteps
600M
700M
800M
1000M
1500M
2000M
1000M
1500M
2000M
2.0
1.5
0.5
0.0
0M
500M
1.0
0.5
0.0
0.00
0M
Reward
0M
Reward
0.40
Ours
Ours, Resets
No Q-Filter
0.6
Success Rate
0.8
Success Rate
Stack 5, Step Reward
Ours
Ours, Resets
No Q-Filter
No BC
Success Rate
1.0
0.0
0M
500M
1000M
1500M
0M
500M
Timesteps
Timesteps
Fig. 5: Ablation results on longer horizon tasks with a step reward. The upper row shows the success rate while the lower
row shows the average reward at the final step of each episode obtained by different algorithms. For stacking 4 and 5 blocks,
we use 2 random seeds per method. The median of the runs is shown in bold and each training run is plotted in a lighter
color. Note that for stacking 4 blocks, the “No BC” method is always at 0% success rate. As the number of blocks increases,
resets from demonstrations becomes more important to learn the task.
No Q-Filter: This method uses standard behavioral cloning
loss instead of the loss from equation Eq. 8, which means
that the actor tries to mimic the demonstrator’s behaviour
regardless of the critic.
No HER: Hindsight Experience Replay is not used.
A. Behavior Cloning Loss
Without the behavior cloning loss, the method is significantly worse in every task we try. Fig. 4 shows the training
curve for learning to stack 3 blocks with a fully sparse
reward. Without the behavior cloning loss, the system is
about 2x slower to learn. On longer horizon tasks, we do
not achieve any success without this loss.
To see why, consider the training curves for stacking 4
blocks shown in Fig. 5. The “No BC” policy learns to stack
only one additional block. Without the behavior cloning loss,
the agent only has access to the demonstrations through the
demonstration replay buffer. This allows it to view highreward states and incentivizes the agent to stack more blocks,
but there is a stronger disincentive: stacking the tower higher
is risky and could result in lower reward if the agent knocks
over a block that is already correctly placed. Because of
this risk, which is fundamentally just another instance of
the agent finding a local optimum in a shaped reward, the
agent learns the safer behavior of pausing after achieving a
certain reward. Explicitly weighting behavior cloning steps
into gradient updates forces the policy to continue the task.
B. Q-Filter
The Q-Filter is effective in accelerating learning and
achieving optimal performance. Fig. 4 shows that the method
without filtering is slower to learn. One issue with the
behavior cloning loss is that if the demonstrations are suboptimal, the learned policy will also be suboptimal. Filtering
by Q-value gives a natural way to anneal the effect of the
demonstrations as it automatically disables the BC loss when
a better action is found. However, it gives mixed results
on the longer horizon tasks. One explanation is that in the
step reward case, learning relies less on the demonstrations
because the reward signal is stronger. Therefore, the training
is less affected by suboptimal demonstrations.
C. Resets From Demonstrations
We find that initializing rollouts from within demonstration states greatly helps to learn to stack 5 and 6 blocks but
hurts training with fewer blocks, as shown in Fig. 5. Note that
even where resets from demonstration states helps the final
success rate, learning takes off faster when this technique
is not used. However, since stacking the tower higher is
risky, the agent learns the safer behavior of stopping after
achieving a certain reward. Resetting from demonstration
states alleviates this problem because the agent regularly
experiences higher rewards.
This method changes the sampled state distribution, biasing it towards later states. It also inflates the Q values
unrealistically. Therefore, on tasks where the RL algorithm
does not get stuck in solving a subset of the full problem, it
could hurt performance.
IX. D ISCUSSION AND F UTURE W ORK
We present a system to utilize demonstrations along with
reinforcement learning to solve complicated multi-step tasks.
We believe this can accelerate learning of many tasks,
especially those with sparse rewards or other difficulties in
exploration. Our method is very general, and can be applied
on any continuous control task where a success criterion can
be specified and demonstrations obtained.
An exciting future direction is to train policies directly on
a physical robot. Fig. 2 shows that learning the pick-andplace task takes about 1 million timesteps, which is about
6 hours of real world interaction time. This can realistically
be trained on a physical robot, short-cutting the simulationreality gap entirely. Many automation tasks found in factories
and warehouses are similar to pick-and-place but without the
variation in initial and goal states, so the samples required
could be much lower. With our method, no expert needs
to be in the loop to train these systems: demonstrations
can be collected by users without knowledge about machine
learning or robotics and rewards could be directly obtained
from human feedback.
A major limitation of this work is sample efficiency
on solving harder tasks. While we could not solve these
tasks with other learning methods, our method requires a
large amount of experience which is impractical outside
of simulation. To run these tasks on physical robots, the
sample efficiency will have to improved considerably. We
also require demonstrations which are not easy to collect
for all tasks. If demonstrations are not available but the
environment can be reset to arbitrary states, one way to learn
goal-reaching but avoid using demonstrations is to reuse
successful rollouts as in [40].
Finally, our method of resets from demonstration states
requires the ability to reset to arbitrary states. Although we
can solve many long-horizon tasks without this ability, it is
very effective for the hardest tasks. Resetting from demonstration rollouts resembles curriculum learning: we solve a
hard task by first solving easier tasks. If the environment
does not afford setting arbitrary states, then other curriculum
methods will have to be used.
X. ACKNOWLEDGEMENTS
We thank Vikash Kumar and Aravind Rajeswaran for valuable discussions. We thank Sergey Levine, Chelsea Finn, and
Carlos Florensa for feedback on initial versions of this paper.
Finally, we thank OpenAI for providing a supportive research
environment.
R EFERENCES
[1] M. Andrychowicz et al., “Hindsight experience replay,” in Advances
in neural information processing systems, 2017.
[2] M. Večerı́k et al., “Leveraging Demonstrations for Deep Reinforcement Learning on Robotics Problems with Sparse Rewards,” arXiv
preprint arxiv:1707.08817, 2017.
[3] M. P. Deisenroth, C. E. Rasmussen, and D. Fox, “Learning to Control a
Low-Cost Manipulator using Data-Efficient Reinforcement Learning,”
Robotics: Science and Systems, vol. VII, pp. 57–64, 2011.
[4] Y. Duan et al., “One-shot imitation learning,” in NIPS, 2017.
[5] D. A. Pomerleau, “Alvinn: An autonomous land vehicle in a neural
network,” NIPS, pp. 305–313, 1989.
[6] M. Bojarski et al., “End to End Learning for Self-Driving Cars,” arXiv
preprint arXiv:1604.07316, 2016.
[7] A. Giusti et al., “A Machine Learning Approach to Visual Perception
of Forest Trails for Mobile Robots,” in IEEE Robotics and Automation
Letters., 2015, pp. 2377–3766.
[8] J. Nakanishi et al., “Learning from demonstration and adaptation of
biped locomotion,” in Robotics and Autonomous Systems, vol. 47, no.
2-3, 2004, pp. 79–91.
[9] M. Kalakrishnan et al., “Learning Locomotion over Rough Terrain
using Terrain Templates,” in The 2009 IEEE/RSJ International Conference on Intelligent Robots and Systems, 2009.
[10] S. Ross, G. J. Gordon, and J. A. Bagnell, “A Reduction of Imitation
Learning and Structured Prediction to No-Regret Online Learning,”
in Proceedings of the 14th International Conference on Artificial
Intelligence and Statistics (AISTATS), 2011.
[11] A. Ng and S. Russell, “Algorithms for Inverse Reinforcement Learning,” International Conference on Machine Learning (ICML), 2000.
[12] B. D. Ziebart et al., “Maximum Entropy Inverse Reinforcement
Learning.” in AAAI Conference on Artificial Intelligence, 2008, pp.
1433–1438.
[13] P. Abbeel and A. Y. Ng, “Apprenticeship learning via inverse reinforcement learning,” in ICML, 2004, p. 1.
[14] C. Finn, S. Levine, and P. Abbeel, “Guided Cost Learning: Deep
Inverse Optimal Control via Policy Optimization,” in ICML, 2016.
[15] J. Peters, K. Mülling, and Y. Altün, “Relative Entropy Policy Search,”
Artificial Intelligence, pp. 1607–1612, 2010.
[16] M. P. Deisenroth and C. E. Rasmussen, “Pilco: A model-based and
data-efficient approach to policy search,” in ICML, 2011, pp. 465–472.
[17] V. Mnih et al., “Human-level control through deep reinforcement
learning,” Nature, vol. 518, no. 7540, pp. 529–533, 2015.
[18] D. Silver et al., “Mastering the game of Go with deep neural networks
and tree search,” Nature, vol. 529, no. 7587, pp. 484–489, Jan 2016.
[19] S. Levine et al., “End-to-end training of deep visuomotor policies,”
CoRR, vol. abs/1504.00702, 2015.
[20] L. Pinto and A. Gupta, “Supersizing self-supervision: Learning
to grasp from 50k tries and 700 robot hours,” arXiv preprint
arXiv:1509.06825, 2015.
[21] S. Levine et al., “Learning hand-eye coordination for robotic grasping
with deep learning and large-scale data collection,” arXiv preprint
arXiv:1603.02199, 2016.
[22] S. Gu et al., “Deep Reinforcement Learning for Robotic Manipulation with Asynchronous Off-Policy Updates,” arXiv preprint
arXiv:1610.00633, 2016.
[23] T. P. Lillicrap et al., “Continuous control with deep reinforcement
learning,” arXiv preprint arXiv:1509.02971, 2015.
[24] V. Mnih et al., “Asynchronous methods for deep reinforcement learning,” in ICML, 2016.
[25] J. Schulman et al., “Trust region policy optimization,” in Proceedings
of the twenty-first international conference on Machine learning, 2015.
[26] T. Winograd, Understanding Natural Language. Academic Press,
1972.
[27] L. P. Kaelbling and T. Lozano-Perez, “Hierarchical task and motion
planning in the now,” IEEE International Conference on Robotics and
Automation, pp. 1470–1477, 2011.
[28] L. Kavraki et al., “Probabilistic roadmaps for path planning in highdimensional configuration spaces,” IEEE transactions on Robotics and
Automation, vol. 12, no. 4, pp. 566–580, 1996.
[29] S. Srivastava et al., “Combined Task and Motion Planning Through
an Extensible Planner-Independent Interface Layer,” in International
Conference on Robotics and Automation, 2014.
[30] I. Popov et al., “Data-efficient Deep Reinforcement Learning for
Dexterous Manipulation,” arXiv preprint arXiv:1704.03073, 2017.
[31] S. Schaal, “Robot learning from demonstration,” Advances in Neural
Information Processing Systems, no. 9, pp. 1040–1046, 1997.
[32] J. Peters and S. Schaal, “Reinforcement learning of motor skills with
policy gradients,” Neural Networks, vol. 21, no. 4, pp. 682–697, 2008.
[33] J. Kober and J. Peter, “Policy search for motor primitives in robotics,”
in Advances in neural information processing systems, 2008.
[34] T. Hester et al., “Learning from Demonstrations for Real World
Reinforcement Learning,” arXiv preprint arxiv:1704.03732, 2017.
[35] B. Kim et al., “Learning from Limited Demonstrations,” Neural
Information Processing Systems., 2013.
[36] T. Schaul et al., “Universal Value Function Approximators,” Proceedings of The 32nd International Conference on Machine Learning, pp.
1312–1320, 2015.
[37] E. Todorov, T. Erez, and Y. Tassa, “MuJoCo: A physics engine for
model-based control,” in The IEEE/RSJ International Conference on
Intelligent Robots and Systems, 2012.
[38] D. Kingma and J. Ba, “Adam: A method for stochastic optimization,”
International Conference on Learning Representations (ICLR), 2015.
[39] D. Bahdanau, K. Cho, and Y. Bengio, “Neural Machine Translation
by Jointly Learning to Align and Translate,” in ICLR, 2015.
[40] C. Florensa et al., “Reverse Curriculum Generation for Reinforcement
Learning,” in Conference on robot learning, 2017.
| 2cs.AI |
Polymorphic Types in ACL2
Benjamin Selfridge
Eric Smith
University of Texas at Austin
Austin, TX
benself@cs.utexas.edu
Kestrel Institute
Palo Alto, CA
eric.smith@kestrel.edu
This paper describes a tool suite for the ACL2 programming language which incorporates certain
ideas from the Hindley-Milner paradigm of functional programming (as exemplified in popular languages like ML and Haskell), including a “typed” style of programming with the ability to define
polymorphic types. These ideas are introduced via macros into the language of ACL2, taking advantage of ACL2’s guard-checking mechanism to perform type checking on both function definitions
and theorems. Finally, we discuss how these macros were used to implement features of Specware
[1], a software specification and implementation system.
1
Introduction
Specware is “a next-generation environment supporting the design, development and automated synthesis
of scalable, correct-by-construction software.” [1] The language of Specware is a high-level programming and specification language used to create program specifications, and to refine them into executable
code.1 One of the selling points of a high-powered system like Specware is its robust type system, which
accommodates sum type definitions, pattern-matching, and polymorphic types.
Figure 1 contains a snippet of Specware code that will be used as a running example throughout
this work. It includes two type definitions, two function definitions (these are called ops in Specware),
and several theorems. The first type definition listed, SeqInt, represents a list of integers. It consists
of two type cases: SeqNil, representing an empty list, and SeqCons, consisting of a integer/SeqInt
pair. In Specware we call these sum types “coproduct” types, and we will stick to this terminology in the
remainder of this work.
The next type definition, Seq a, is a generalization of the SeqInt type to be polymorphic in the
type of its contents. We can instantiate the type variable a to be any of Specware’s built-in base types,
or any other type we wish to define. We can also define functions on this new type, like SeqAppend
and SeqRev, and we can state theorems about such functions, like SeqAppend_Associative and
SeqRev_of_SeqRev.
Polymorphic typing greatly enhances the expressibility of Specware, and it is one of many sophisticated language features that ACL2, a first-order untyped programming language and theorem prover,
does not support natively. The goal of this work is to present a partial implementation of some of these
features in ACL2, which includes a mechanism for defining “typed” ACL2 functions and theorems in
a style that mimics Specware syntax. This work is some of the fruits of a larger effort to use ACL2 as
a back-end prover for Specware. Since Specware emits proof obligations but does not directly verify
them, all the proofs must be translated into an external proof environment and verified independently.
The ACL2 constructs presented in this work were essential in the translation process, and they also have
1 Throughout this paper, we use the term Specware to refer to both the full Specware system, and to Metaslang, the programming language used by Specware.
F. Verbeek and J. Schmaltz (Eds.): ACL2 Workshop 2014 (ACL2’14).
EPTCS 152, 2014, pp. 49–59, doi:10.4204/EPTCS.152.4
Polymorphic Types in ACL2
50
type SeqInt =
| SeqNil
| SeqCons Int * SeqInt
type Seq a =
| SeqNil
| SeqCons a * (Seq a)
op [a] SeqAppend (x:Seq a, y:Seq a) : Seq a =
case x of
| SeqNil -> y
| SeqCons (hd,tl) -> SeqCons (hd, SeqAppend (tl, y))
op [b] SeqRev (x:Seq b) : Seq b =
case x of
| SeqNil -> SeqNil
| SeqCons (hd,tl) -> SeqAppend (SeqRev tl, SeqCons (hd,SeqNil))
theorem SeqAppend_Associative is [a]
fa(x:Seq a,y:Seq a,z:Seq a)
SeqAppend(SeqAppend(x,y),z) = SeqAppend(x,SeqAppend(y,z))
theorem SeqAppend_of_SeqNil_1 is [a]
fa (x:Seq a) SeqAppend(SeqNil,x) = x
theorem SeqAppend_of_SeqNil_2 is [a]
fa (x:Seq a) SeqAppend(x,SeqNil) = x
theorem SeqRev_of_SeqAppend is [a]
fa (x:Seq a,y:Seq a) SeqRev (SeqAppend (x,y)) = SeqAppend (SeqRev y, SeqRev x)
theorem SeqRev_of_SeqRev is [a]
fa (x:Seq a) (SeqRev (SeqRev x)) = x
Figure 1: Our example Specware program.
the potential to be quite useful on their own, independent of Specware. Although our work does not address type inference, this could be achieved by maintaining an ACL2 table mapping functions to types;
we discuss this more in the conclusion.
The organization of the remainder of this paper is as follows. Sections 2, 3, and 4 describe how
the program in Figure 1 is translated into ACL2 (via the use of some new macros), addressing the
coproduct type definitions, function definitions, and theorems respectively. Section 5 concludes the
paper by summarizing what has been accomplished and what remains to be done (type inference).
2
ACL2 “types” and polymorphic type definitions
Before we can begin thinking about defining polymorphic types in ACL2, we must first understand how
to implement types in the first place. ACL2 is an untyped language; in the ACL2 logic, every function
B. Selfridge & E. Smith
51
must accept all types of arguments. However, we can use predicates (functions of arity 1 that return either
T or NIL) to define a type. An example of such a function in ACL2 is integerp, which returns T if its
argument is an integer, and NIL otherwise. In this work, we use Int to designate this type, and Int-p
to designate the ACL2 predicate that recognizes Ints.2 Throughout this work, we maintain a distinction
between the name of a type, Type, and its recognizer function, Type-p. This enables better automation
of macros that operate on these types; if the user refers to a type Type, the system will automatically
append -p to its name.
In order to implement coproduct types in ACL2, we started with the pre-existing ACL2 book defsum.
The defsum macro3 uses a combination of guard checking and theorems about the output type of functions (along with many other useful theorems) in order to implement an ad-hoc sum type. It also includes
a pattern-matching mechanism, pm, which is a more sophisticated version of ACL2’s case macro. We
introduced a new macro, defcoproduct, which is defined in terms of defsum but also accommodates
polymorphism, and a slightly modified version of pm, which we named case-of in order to more closely
reflect the syntax of Specware. We can define the “concrete” (non-polymorphic) type SeqInt as follows:
(defcoproduct SeqInt
(SeqNil)
(SeqCons Int SeqInt))
This macro-expands to a simple call to defsum:
(DEFSUM SEQINT
(SEQNIL)
(SEQCONS (INT-P ARG-1) (SEQINT-P ARG-2)))
As we can see, this type is simple enough to have been defined using defsum alone. To define a polymorphic Seq data structure, however, we need to use defcoproduct:
(defcoproduct Seq
:type-vars (a)
(SeqNil)
(SeqCons a (:inst Seq a)))
;; type variables
;; two type cases - SeqNil and SeqCons
This code, although its syntax is obviously ACL2, still resembles the original Specware definition of Seq.
The defcoproduct macro defines a new type, Seq, by introducing two type constructors, SeqCons and
SeqNil. The type is defined using a single type variable a, which is a placeholder for the type of Seq’s
contents. The :inst tag is necessary to inform the macro that we are using a particular instance of the
Seq type.
Because the logic of ACL2 does not have true polymorphic types (indeed, it does not have a type
system at all), the ACL2 definition of Seq is not a true type definition, like it is in Specware. Instead, it
serves as a template for creating instantiations of Seq with a specific type replacing the type variable a.
The above definition actually introduces a macro, Seq-instantiate, which we can use to instantiate
Seq on integers as follows:
(Seq-instantiate int)
This macro-expands (as before) to a call to defsum:
2 We
define this function as a synonym for integerp, and likewise, we define Bool-p as a synonym for booleanp.
macro was introduced and first used in Swords and Cook [2]. We added some slight modifications to the original
macro in order to accommodate a bookkeeping mechanism that would enable automatic functional instantiation for polymorphic
theorems.
3 This
Polymorphic Types in ACL2
52
(DEFSUM SEQ-INT
(SEQNIL-INT)
(SEQCONS-INT (INT-P ARG-1)
(SEQ-INT-P ARG-2)))
We can see this looks nearly identical to the definition of the concrete coproduct type SeqInt above.
(We have deliberately omitted some bookkeeping information from this defsum call - this information
will be necessary when we wish to instantiate theorems about the Seq type on specific instances, but the
details of this are not important.)
We can also define polymorphic types with more than one type variable, or in terms of previously
defined polymorphic types:
(defcoproduct EitherSeq
:type-vars (a b)
(LeftSeq (:inst Seq a))
(RightSeq (:inst Seq b)))
This defines a new polymorphic type, EitherSeq, parameterized by variables a and b. We can now
instantiate it with concrete types:
(EitherSeq-instantiate int bool)
The above call expands to
(PROGN (SEQ-INSTANTIATE INT)
(SEQ-INSTANTIATE BOOL)
(DEFSUM EITHERSEQ-INT-BOOL
(LEFTSEQ-INT-BOOL (SEQ-INT-P ARG-1))
(RIGHTSEQ-INT-BOOL (SEQ-BOOL-P ARG-1))))
Notice that before EitherSeq is instantiated on integers and booleans, the Seq type must be instantiated
on these two types. This is automatically checked by the defcoproduct macro, and these instantiations
are collected and included before defining EitherSeq-Int-Bool. It is notable that none of the macros
presented in this paper query the ACL2 world. If some types have already been defined, the preliminary
type instantiations will be dismissed by ACL2 as redundant definitions, and the macro will still succeed.
The polymorphism supported by defcoproduct is somewhat limited; the macro does not support
mutually recursive datatypes, and all instantiation must happen in one step (i.e., we cannot instantiate the
a variable of EitherSeq and leave b as a type variable). However, these are not fundamental limitations,
and could be implemented with more work.
3
Polymorphic functions and theorems
Consider the Specware function SeqAppend defined in Figure 1. This is the “append” function on
lists, defined for the polymorphic type Seq. We can translate this definition into ACL2 using our new
defun-typed macro:
(defun-typed
:type-vars
((x (:inst
(:inst Seq
(case-of x
SeqAppend
(a)
Seq a)) (y (:inst Seq a)))
a)
;;
;;
;;
;;
type variables
typed argument list
output type
function body
B. Selfridge & E. Smith
53
(((:inst SeqNil a))
;; case 1: SeqNil
y)
(((:inst SeqCons a) hd tl)
;; case 2: SeqCons
((:inst SeqCons a) hd ((:inst SeqAppend a) tl y)))))
The defun-typed macro is a version of defun that requires type annotations for all its input values,
as well as a type annotation for its own return value. We supply a list of type variables (which can be
omitted if there are none), a list of the arguments with their associated types, an output type, and the
body of the function.
One obvious weakness of this definition is its verbosity. Every time we pattern match on a polymorphic type, we must include the :inst keyword in each pattern to indicate which instantiation for a
given constructor we are using. This could, of course, be solved by implementing type inference; for our
immediate task of translating Specware code to ACL2, this was not necessary, but it could be done with
some more work.
In order to use this function, we must first instantiate it:
(SeqAppend-instantiate int)
This macro-expands to
(PROGN (SEQ-INSTANTIATE INT)
(SEQNIL-INSTANTIATE INT)
(SEQCONS-INSTANTIATE INT)
(DEFUN SEQAPPEND-INT (X Y)
(DECLARE (XARGS :GUARD (AND (SEQ-INT-P X) (SEQ-INT-P Y))
:VERIFY-GUARDS NIL))
(IF (MBT (AND (SEQ-INT-P X) (SEQ-INT-P Y)))
(CASE-OF X ((SEQNIL-INT) Y)
((SEQCONS-INT HD TL)
(SEQCONS-INT HD (SEQAPPEND-INT TL Y))))
NIL))
(DEFTHM SEQAPPEND-INT-TYPE
(IMPLIES (AND (SEQ-INT-P X) (SEQ-INT-P Y))
(SEQ-INT-P (SEQAPPEND-INT X Y)))
:RULE-CLASSES (:TYPE-PRESCRIPTION :REWRITE))
(VERIFY-GUARDS SEQAPPEND-INT)
As before, we first instantiate the polymorphic Seq type for ints before defining our SeqAppend-Int
function. The SEQ-INSTANTIATE, SEQNIL-INSTANTIATE, and SEQCONS-INSTANTIATE are all redundant; they have the exact same definition. The defun-typed macro scans its argument list, output type,
and body for the :inst keyword, and calls the associated instantiation macro for each occurrence. This
was a brute-force way to guarantee that all the necessary functions and types will be defined before the
current function definition is submitted. Notice how a combination of ACL2 guards and theorems are
used to ensure that SeqAppend-Int satisfies all the typing requirements; we require that the guards of
the function calls made in the definition of SeqAppend-Int are never violated given our assumptions
about the types of x and y, and we also require that, assuming both input variables are Seq-Ints, the
output of this function is also a SeqInt. Notice how we check the latter first; for recursive definitions, it
is often the case that we need to know the output type of the function before we can verify the guards.
Of course, we can also define polymorphic functions in terms of other, previously defined ones.
Polymorphic Types in ACL2
54
(defun-typed SeqRev
:type-vars (a)
((x (:inst Seq a)))
(:inst Seq a)
(case-of x
(((:inst SeqNil a)) ((:inst SeqNil a)))
(((:inst SeqCons a) hd tl)
((:inst SeqAppend a)
((:inst SeqRev a) tl)
((:inst SeqCons a) hd ((:inst SeqNil a)))))))
If we instantiate SeqRev with the concrete type bool via
(SeqRev-instantiate bool)
this will macro-expand to
(PROGN
(SEQ-INSTANTIATE BOOL)
(SEQAPPEND-INSTANTIATE BOOL)
(SEQCONS-INSTANTIATE BOOL)
(SEQNIL-INSTANTIATE BOOL)
(DEFUN SEQREV-BOOL (X)
(DECLARE (XARGS :GUARD (SEQ-BOOL-P X)
:VERIFY-GUARDS NIL))
(IF (MBT (SEQ-BOOL-P X))
(CASE-OF X ((SEQNIL-BOOL) (SEQNIL-BOOL))
((SEQCONS-BOOL HD TL)
(SEQAPPEND-BOOL (SEQREV-BOOL TL)
(SEQCONS-BOOL HD (SEQNIL-BOOL)))))
NIL))
(DEFTHM SEQREV-BOOL-TYPE
(IMPLIES (SEQ-BOOL-P X)
(SEQ-BOOL-P (SEQREV-BOOL X)))
:RULE-CLASSES (:TYPE-PRESCRIPTION :REWRITE))
(VERIFY-GUARDS SEQREV-BOOL))
Notice that both the Seq type and the SeqAppend function are instantiated for bool before defining
SeqRev-Bool. Of course, it would have sufficed to only invoke (SEQAPPEND-INSTANTIATE BOOL),
but our defun-typed macro is not smart enough to figure that out; everywhere it sees an :inst keyword,
it calls the associated instantiation macro.
We can also state (and prove) theorems about functions involving polymorphic types using our new
macro defthm-typed. For instance, we can translate the SeqAppend_Associative theorem from the
introduction into ACL2 like so:
(defthm-typed SeqAppend_Associative
:type-vars (a)
;; type variables
((x (:inst Seq a))
;; type annotations for free variables
(y (:inst Seq a))
(z (:inst Seq a)))
B. Selfridge & E. Smith
55
(equal
;; theorem body
((:inst SeqAppend a) ((:inst SeqAppend a) x y) z)
((:inst SeqAppend a) x ((:inst SeqAppend a) y z))))
This macro-expands to
(PROGN
(ENCAPSULATE (((A-P *) => *))
(LOCAL (DEFUN A-P (X) (DECLARE (IGNORE X)) T))
(DEFTHM A-TYPE (BOOLEANP (A-P X))
:RULE-CLASSES :TYPE-PRESCRIPTION))
(SEQ-INSTANTIATE A)
(SEQAPPEND-INSTANTIATE A)
(DEFUND-TYPED SEQAPPEND_ASSOCIATIVE-A-BODY
((X SEQ-A) (Y SEQ-A) (Z SEQ-A))
BOOL
(EQUAL (SEQAPPEND-A (SEQAPPEND-A X Y) Z)
(SEQAPPEND-A X (SEQAPPEND-A Y Z))))
(DEFTHM SEQAPPEND_ASSOCIATIVE-A
(IMPLIES (AND (SEQ-A-P X)
(SEQ-A-P Y)
(SEQ-A-P Z))
(EQUAL (SEQAPPEND-A (SEQAPPEND-A X Y) Z)
(SEQAPPEND-A X (SEQAPPEND-A Y Z)))))
(DEFMACRO SEQAPPEND_ASSOCIATIVE-INSTANTIATE (A)
;; ... macro definition omitted
)
The defthm-typed macro does several things. First, it defines an encapsulated predicate, A-P, which
will be used to represent the type variable a. Then, after instantiating all the needed types and functions, we type check the body of the theorem by defining it as a function with output type bool (if the
theorem doesn’t even type check, then we don’t need to bother to try and prove it). Then, it proves
a version of SeqAppend_Associative where the Seq type has been instantiated on an encapsulated
predicate A-P. In theory, this proves the theorem in general for any type instantiation. Finally, a new
macro, SEQAPPEND_ASSOCIATIVE-INSTANTIATE, is introduced, which allows us to prove this theorem for a specific instantiation of the Seq type. This macro uses functional instantiation (along with
a substantial amount of bookkeeping) to prove the theorem automatically from the original theorem,
SEQAPPEND_ASSOCIATIVE-A. If we instantiate this theorem for integers via
(SeqAppend_Associative-instantiate int)
we get
(PROGN
(SEQ-INSTANTIATE INT)
(SEQAPPEND-INSTANTIATE INT)
(DEFTHM-TYPED
SEQAPPEND_ASSOCIATIVE-INT
((X SEQ-INT) (Y SEQ-INT) (Z SEQ-INT))
(EQUAL (SEQAPPEND-INT (SEQAPPEND-INT X Y) Z)
Polymorphic Types in ACL2
56
(SEQAPPEND-INT X (SEQAPPEND-INT Y Z)))
:HINTS
(("Goal" :DO-NOT-INDUCT T
:IN-THEORY (ENABLE SEQ-INT-FUNCTIONS)
:USE ((:FUNCTIONAL-INSTANCE
SEQAPPEND_ASSOCIATIVE-A (A-P INT-P)
(SEQ-A-P SEQ-INT-P)
(SEQCONS-A-P SEQCONS-INT-P)
(SEQNIL-A-P SEQNIL-INT-P)
(SEQCONS-A SEQCONS-INT)
(SEQNIL-A SEQNIL-INT)
(SEQCONS-A-ARG-2 SEQCONS-INT-ARG-2)
(SEQCONS-A-ARG-1 SEQCONS-INT-ARG-1)
(SEQAPPEND-A SEQAPPEND-INT)))))))
Notice how the theorem is proved using functional instantiation on the more general theorem,
SeqAppend_Associative-A.
We can use the macros described above to implement the entire Specware program of Figure 1 using
our four new macros, defcoproduct, defun-typed, defthm-typed, and case-of. The full listing
for the ACL2 version of this program is given in Figure 2.
4
Conclusions
These macros were introduced in an ACL2 book in order to facilitate the translation process from
Specware to ACL2. Instead of having our translator produce the raw ACL2 code, we hide the messy
details of implementing these high-level language features with ACL2 macros, which has the advantage
of both making the translation process easier and making the automatically generated ACL2 code much
more readable. The gen-acl2 tool was added to Specware in order to automatically translate Specware
programs into ACL2 code that uses the macros defined here (the ACL2 code in Figure 2 was generated
by this tool).
A byproduct of this effort was the ACL2 macros themselves, which are quite useful in their own right,
and suggest the possibility of implementing more of these high-level features in ACL2. Limitations of
the polymorphism presented here include the inability to define mutually recursive polymorphic types,
as well as the lack of partial type instantiation. These macros could be extended in a straightforward way
to include these features.
Type inference could be implemented by maintaining an ACL2 table mapping function names to their
types (essentially, the “typing” theorems exported by the defun-typed macro). When the user defines a
new function, the Hindley-Milner algorithm can be used to deduce the necessary input and output types
of the function (assuming all the function calls in the body already exist in the table), and we can export
a theorem capturing this which could be proved in a theory that only includes the typing theorems of
previously defined functions.
We also believe that the techniques used in this paper could be used to introduce “higher-order”
functions; in particular, we can modify our technique of extracting away a type variable by introducing
an encapsulated predicate, to extracting away a function by introducing an encapsulated function. We
have not thoroughly investigated the viability of this idea, but it seems to us like a fruitful avenue for
further research.
B. Selfridge & E. Smith
57
(in-package "ACL2")
(include-book "~/Desktop/specware-files/code/specware-book")
(set-ignore-ok t)
(set-bogus-defun-hints-ok t)
(defcoproduct SeqInt
(SeqNil)
(SeqCons Int SeqInt))
(defcoproduct Seq
:type-vars (a)
(SeqCons a (:inst Seq a))
(SeqNil))
(defun-typed
:type-vars
((x (:inst
(:inst Seq
(case-of x
(((:inst
(((:inst
((:inst
SeqAppend
(a)
Seq a)) (y (:inst Seq a)))
a)
SeqNil a)) y)
SeqCons a) hd tl)
SeqCons a) hd ((:inst SeqAppend a) tl y)))))
(defthm-typed SeqAppend_Associative
:type-vars (a)
((x (:inst Seq a))
(y (:inst Seq a))
(z (:inst Seq a)))
(equal ((:inst SeqAppend a) ((:inst SeqAppend a) x y) z)
((:inst SeqAppend a) x ((:inst SeqAppend a) y z))))
(defthm-typed SeqAppend_of_SeqNil_1
:type-vars (a)
((x (:inst Seq a)))
(equal ((:inst SeqAppend a) ((:inst SeqNil a)) x) x))
(defthm-typed SeqAppend_of_SeqNil_2
:type-vars (a)
((x (:inst Seq a)))
(equal ((:inst SeqAppend a) x ((:inst SeqNil a))) x))
58
Polymorphic Types in ACL2
(defun-typed SeqRev
:type-vars (a)
((x (:inst Seq a)))
(:inst Seq a)
(case-of x
(((:inst SeqNil a)) ((:inst SeqNil a)))
(((:inst SeqCons a) hd tl)
((:inst SeqAppend a)
((:inst SeqRev a) tl)
((:inst SeqCons a) hd ((:inst SeqNil a)))))))
(defthm-typed SeqRev_of_SeqAppend
:type-vars (a)
((x (:inst Seq a))
(y (:inst Seq a)))
(equal ((:inst SeqRev a) ((:inst SeqAppend a) x y))
((:inst SeqAppend a) ((:inst SeqRev a) y)
((:inst SeqRev a) x))))
(defthm-typed SeqRev_of_SeqRev
:type-vars (a)
((x (:inst Seq a)))
(equal ((:inst SeqRev a) ((:inst SeqRev a) x)) x)
:hints (("Goal" :in-theory (enable SeqAppend-a SeqRev-a))))
Figure 2: Our example Specware program, defined in ACL2.
B. Selfridge & E. Smith
59
References
[1] James McDonald & John Anton (2001): SPECWARE - Producing Software Correct by Construction.
[2] Sol Swords & William R. Cook (2006): Soundness of the Simply Typed Lambda Calculus in ACL2. In:
Proceedings of the Sixth International Workshop on the ACL2 Theorem Prover and Its Applications, ACL2
’06, ACM, New York, NY, USA, pp. 35–39, doi:10.1145/1217975.1217982.
| 6cs.PL |
Targeted Random Projection for Prediction
arXiv:1712.02445v1 [math.ST] 6 Dec 2017
from High-Dimensional Features
Minerva Mukhopadhyay
David B. Dunson
Abstract
We consider the problem of computationally-efficient prediction from high dimensional and highly correlated predictors in challenging settings where accurate variable selection is effectively impossible. Direct application of penalization or Bayesian
methods implemented with Markov chain Monte Carlo can be computationally daunting and unstable. Hence, some type of dimensionality reduction prior to statistical
analysis is in order. Common solutions include application of screening algorithms to
reduce the regressors, or dimension reduction using projections of the design matrix.
The former approach can be highly sensitive to threshold choice in finite samples,
while the later can have poor performance in very high-dimensional settings. We
propose a TArgeted Random Projection (TARP) approach that combines positive
aspects of both strategies to boost performance. In particular, we propose to use information from independent screening to order the inclusion probabilities of the features in the projection matrix used for dimension reduction, leading to data-informed
1
sparsity. We provide theoretical support for a Bayesian predictive algorithm based
on TARP, including both statistical and computational complexity guarantees. Examples for simulated and real data applications illustrate gains relative to a variety
of competitors.
Some key words: Bayesian; Dimension reduction; High-dimensional; Large p, small n;
Random projection; Screening.
Short title: Targeted Random Projection
1
Introduction
In many applications, the focus is on prediction of a response variable y given a massivedimensional vector of predictors x = (x1 , x2 , . . . xp )0 . Often enormous numbers of possibly
collinear predictors x are collected, and the sample size n is modest relative to p, so that
p n. In such situations, it is common to assume that x can be replaced by a much
lower-dimensional feature vector comprised of sparse linear combinations of the original
predictors. However, accurate learning of the precise lower-dimensional structure is often
not possible, as the data simply do not contain sufficient information even putting aside
the intractable computational problem.
There is a large literature on variable selection in p n settings. Much of the focus has
been on penalized optimization-based approaches, with some popular methods including
LASSO (Tibshirani, 1996), SCAD (Fan and Li, 2001)), elastic net (Zou and Hastie, 2005),
the Dantzig selector (Candes and Tao, 2007), and MCP (Zhang et al., 2010). There is
2
also a large literature on Bayesian approaches that attempt to characterize uncertainty
in variable selection. Most approaches use some variation on the spike and slab prior
(e.g. Ishwaran and Rao (2005)) or continuous shrinkage priors concentrated at zero with
heavy tails (e.g. Carvalho et al. (2009)). There is an increasingly rich theoretical literature
providing asymptotic support for such methods in the p → ∞ as n → ∞ setting. However,
positive results rely on strong assumptions in terms of a high signal-to-noise ratio and low
linear dependence in the columns of the design matrix.
We are interested in settings where practical performance of the above methods is poor,
due to a combination of statistical and computational intractability. In such settings, it
is common to use variable screening as a pre-processing step. In particular, independent
screening tests for association between y and each xj separately, and selects predictors with
the largest or most significant associations for second stage analysis. In general, screening
can be guaranteed asymptotically to select a superset of the ‘true’ predictors (Fan et al.,
2009). When the number of predictors is sufficiently reduced, one can apply a simple
maximum likelihood approach, penalized optimization, or Bayesian Markov chain Monte
Carlo (MCMC) algorithms in the second stage. However, when the predictors are highly
correlated and/or the true data generating process does not exhibit strong sparsity with
a high signal-to-noise ratio, it may be necessary to use a very conservative threshold for
the measure of marginal association, limiting the dimensionality reduction occurring in the
first stage.
As an alternative to variable screening, there is a rich literature on using random projections (RPs) to reduce data dimensionality prior to statistical analysis. For example,
3
compressive sensing uses RPs to reduce storage and communication costs in signal processing. By exploiting sparsity, the original signal can be reconstructed from the compressive
measurements with high accuracy (see, e.g., Candes and Tao (2005), Donoho (2006), Davenport et al. (2010)). Usual compressive sensing acts in a row-wise manner, reducing the
dimensionality of the design matrix from n × p to m × p, with m n. This does not
solve the big p problem. There is a relatively smaller literature on column-wise compression, which instead reduces the design matrix from n × p to n × m, with m p, while
providing bounds on predictive errors (see, e.g., Maillard and Munos (2009), Fard et al.
(2012), Kabán (2014), Thanei et al. (2017), Guhaniyogi and Dunson (2015), Pettenuzzo
et al. (2016)). Guhaniyogi and Dunson (2015) concentrate on approximating predictive
distributions in Bayesian regression. The above literature on RPs focuses primarily on
random matrices with i.i.d elements.
When predictors are very high-dimensional, existing RP methods can fail as they tend
to include many unimportant predictors in each linear combination, diluting the signal.
Potentially, one can attempt to improve performance by estimating the projection matrix,
but this results in a daunting computational and statistical problem. Alternatively, we
propose a TArgeted Random Projection (TARP) approach, which includes predictors in
the RP matrix with probability proportional to their marginal utilities. These utilities are
estimated quickly in a first stage using an independent screening-type approach. To reduce
sensitivity of the results to the different realizations of the RP matrices, we aggregate over
multiple realizations. TARP can be viewed as a type of rapid preconditioning, enabling
improved predictive performance in high-dimensional settings. Compared with applying
4
RPs after screening out predictors, TARP has the advantage of removing sensitivity to
threshold choice by using a soft probabilistic approach.
In Section 2, we propose the methodology including the computational algorithm, choice
of different tuning parameters and an analysis of computational complexity. We focus on
generalized linear models (GLMs) for ease in presentation and development of a strong
theoretical justification, although TARP can be applied directly in general settings to provide lower-dimensional features that can then be used in any predictive algorithm (random
forests, Gaussian processes, neural nets, etc). Section 3 provides theory on convergence
rates for the predictive distribution of y. Section 4 contains a simulation study comparing
performance with a variety of competitors. In Section 5, we apply TARP to a variety of
real data applications including a genomics dataset with millions of predictors. Section 6
contains a brief discussion, and proofs are included in an Appendix.
2
The Proposed Method
Let Dn = {(yn ; Xn ) : yn ∈ Rn , Xn ∈ Rn×pn } denote the dataset consisting of n observations
on pn predictors x1 , x2 , . . . , xpn and a response y, and (yi ; xi ) denote the ith data point,
i = 1, 2, . . . , n. Suppose that the data can be characterized by a generalized linear model
(GLM). The density of y is related to the predictors as
1
0
0
{yi a(xi β) + b(xi β) + c(yi )} ,
f (yi |β, σ ) = exp
d(σ 2 )
2
5
(1)
where a(·) and b(·) are continuously differentiable functions, a(·) has non-zero derivative
and d(·) is a non-zero function. The vector of regression coefficients β ∈ Rpn , and σ 2 ∈ R+ is
the scale parameter. We approximate the density of y in a compressed regression framework
as follows:
1
0
0
{yi a ((Rn xi ) θ) + b ((Rn xi ) θ) + c(yi )} .
f (yi |θ, Rn , σ ) = exp
d(σ 2 )
2
Here Rn ∈ Rmn ×pn is a random projection matrix, θ ∈ Rmn is the vector of compressed
regression coefficients, and mn pn . We discuss the choice of the random matrix Rn in
Section 2.1, and illustrate the method in detail in Section 2.2.
Priors. We assume that the covariates are standardized. Taking a Bayesian approach
to inference, we assign priors to θ and σ 2 . The vector of compressed regression parameters
θ is assigned a Nmn (0, σ 2 I) prior given σ 2 , where 0 is a vector of zeros, and I is the identity
matrix. The scale parameter σ 2 is assigned an Inv-Gamma (aσ , bσ ) prior, with aσ , bσ > 0.
The Normal-Inverse Gamma (ING) prior is a common choice of prior for GLMs. In the
special case of a Gaussian likelihood, this prior is conjugate, and the posterior and predictive
distributions are available in analytic forms.
2.1
Choice of the projection matrix
The projection matrix Rn embeds Xn to a lower dimensional subspace. If pn is not large,
the best linear embedding can be estimated using a singular value decomposition (SVD)
of Xn . The projection of Xn to the space spanned by the singular vectors associated with
6
the first mn singular values is guaranteed to be closer to Xn in an appropriate sense than
any other mn -dimensional matrix. However, if pn is very large with pn n, then it is
problematic to estimate the projection, both computational and statistically, and random
projection (RP) provides a practical alternative. If an appropriate RP matrix is chosen, due
to Johnson-Lindenstrauss (JL) type embedding results, distances between sample points
are maintained (see Dasgupta and Gupta (2003), Achlioptas (2003)).
Our focus is on modifying current approaches by constructing RP matrices that incorporate sparsity in a way that the predictors xj having relatively weak marginal relationships
with y are less likely to be included in the matrix. In particular, the TArgeted Random
Projection (TARP) matrices are constructed as follows:
i.i.d.
γ = (γ1 , γ2 , . . . , γpn )0
qj ∝ |rxj ,y |δ
Rγ = Omn ×(pn −pγ )
and γj ∼ Bernoulli (qj )
for some constant δ > 0,
and
where
(2)
Rγ = Rn∗ ,
where rxj ,y is the Pearson’s correlation coefficient of xj and yn , qj is the inclusion probability
of xj , Rγ and Rγ are the sub-matrices of Rn with columns corresponding to non-zero and
P
zero values of γ, respectively, and Rn∗ is the mn × pγ projection matrix where pγ = j γj .
We prioritize predictors based on their marginal utilities, q = (q1 , q2 , . . . , qpn )0 , and
consider a random subset of the predictors with inclusion probabilities q. This can be
viewed as a randomized version of independent screening. The selected subset is further
projected to a lower dimensional sub-space using Rn∗ . There are many possible choices of Rn∗
7
which can successfully reduce the dimension of the selected variables while having minimal
impact on prediction accuracy. Two predominant classes of such projection matrices are
based on partial SVD and random projections facilitating JL type embedding. We consider
both these choices, as described below.
∗
, of Rn∗ is sampled independently from a three
Random projection. Each element, Rk,j
point distribution as
∗
Rk,j
±1/√2ψ
=
0
with probability ψ,
(3)
with probability 1 − 2ψ,
where ψ ∈ (0, 0.5) is a constant.
Projection matrices of this form are widely used due to their inter point distance preservation property. Incorporating zero values facilitates data reduction and improves computational efficiency. We refer to the method that generates Rn∗ in (2) from (3) as RIS-RP
(Randomized Independent Screening-Random Projection).
Remark 1. The choice of projection matrix in (3) can be replaced by a wide variety of
matrices having i.i.d. components with mean zero and finite fourth moments. One of the
√
∗
sparsest choices is of the form Rk,j
= ±nκ/2 / mn with probability 1/2nκ , 0 with probability
(1 − 1/nκ ), where mn ∼ nκ (see Li et al. (2006)). This choice of projection matrix reduces
the data to a great extent, and is useful in compressing extremely large dimensional data.
Our theoretical results would hold also if we consider a random matrix Rn∗ as above.
Principal component projection. Another alternative is to use the matrix of principal
8
component scores. Let Xγ be the sub-matrix of Xn with columns corresponding to non-zero
0
values of γ. Consider the partial spectral decomposition Xγ0 Xγ = Vγ,m
Dγ,mn Vγ,mn . The
n
projection matrix Rn∗ we consider is
0
.
Rn∗ = Vγ,m
n
(4)
The method corresponding to this choice of projection matrix combines a randomized
version of independence screening with principal component regression (PCR). Therefore,
we refer to this method as RIS-PCR (Randomized Independence Screening-PCR).
The performance of TARP depends on tuning parameters mn , δ and ψ. In addition, for
any given choice of tuning parameters, different realizations of the random projection matrix will vary, leading to some corresponding variation in the results. To limit dependence
of the results on the choice of tuning parameters and random variation in the projection,
we take the approach of generating multiple realizations of the matrix for different choices
of tuning parameters and aggregating these results together. Potentially, one could estimate weights for aggregation using Bayesian methods (see Hoeting et al. (1999)) or other
ensemble learning approaches, but we focus on simple averaging due to its computational
and conceptual simplicity.
2.2
Posteriors and Predictive Distribution
We illustrate the proposed method in normal linear regression for simplicity, although it is
applicable to more general settings. In the compressed regression framework, we replace
9
the normal linear model yi = x0i β + ei by yi = (Rn xi )0 θ + ei , where ei ∼ N (0, σ 2 ).
Given the NIG prior structure stated above, the posterior distribution of θ follows a scaled
mn -variate t distribution with degrees of freedom (d.f.) n + 2aσ , location vector µt , and
scale matrix Σt , where µt = Z(Xn Rn0 )0 yn , Σt = (yn0 yn − µ0t Z −1 µt + 2bσ )Z/(n + 2aσ ) and
Z = (Rn Xn0 Xn Rn0 + I)−1 . Moreover, the posterior distribution of σ 2 , given Dn and Rn , is
inverse gamma with parameters aσ + n/2 and (yn0 yn − µ0t Σ−1
t µt + bσ )/2.
Consider the problem of point prediction of y when nnew new data points on the predictors are obtained, given the dataset Dn . The predicted values of y, say ynew , can be
obtained using the Bayes estimator of θ under squared error loss as follows
ŷnew = Xnew Rn0 θ̂ Bayes where θ̂ Bayes = (Zn0 Zn + I)−1 Zn0 yn and Zn = Xn Rn .
Here Xnew is the new design matrix. Moreover, the posterior predictive distribution of ynew
is a nnew -variate t distribution with degrees of freedom n + 2aσ , location vector ŷnew and
0
scale parameter (yn0 yn − µ0t Z −1 µt + 2bσ )(I + Xnew ZXnew
)/(n + 2aσ ).
When the distribution yn is non-normal, analytical expressions of the posteriors of θ,
σ 2 and predictive distribution of ynew are not available. In such cases, it is common to rely
on a Laplace approximation (see Tierney and Kadane (1986)) or sampling algorithm, such
as MCMC. In the compressed regression framework, as the pn -dimensional variables are
projected to a much lower-dimensional mn -hyperplane, with mn pn , we are no longer in
a high-dimensional setting. Hence, MCMC is computationally tractable.
2.3
Tuning Parameter Choice
Next, we describe the choices of tuning parameters that are involved in TARP.
10
Choice of mn . The parameter mn determines the number of linear combinations of
predictors we consider. Instead of choosing a fixed value of mn , we consider a range of
values. In particular, we suggest choosing values in the range (2 log pn , min{3n/4, pn }),
consistent with our theoretical results in Section 3 requiring that mn < n and with our
experiments assessing performance for different choices in a variety of scenarios.
Choice of δ and q. The parameter δ plays an important role in screening. Higher values
of δ lead to fewer variables selected in the screening step. If pn n, one would like to select
a small proportion of predictors, and if pn ∼ n then selection of a moderate proportion is
desirable. We recommend δ = max{0, (1+log(pn /n))/2} as a default. This function selects
all the predictors if pn n, and becomes more restrictive as pn becomes larger than n. The
selection probabilities in the RIS stage are then qj = |rxj ,y |δ / maxj |rxj ,y |δ , j = 1, 2, . . . , pn .
Hence, the variable with highest marginal correlation is definitely included in the model.
Choice of ψ. The value of ψ controls sparsity in the random matrix, and it is necessary
to let ψ ∈ (0, 0.5). Achlioptas (2003) suggest choosing ψ = 1/6 as a default value. To avoid
sensitivity of the results for a particular choice of ψ, we choose ψ ∈ (0.1, 0.4) avoiding the
very sparse and dense cases.
2.4
Computational Algorithm and Complexity
We now illustrate the time complexity of TARP along with the algorithm for computation
in normal linear models.
RIS-RP. For a specific choice of (mn , δ, ψ), calculation of ŷnew using RIS-RP involves the
following steps:
11
1: Calculate rxj ,y for j = 1, . . . , pn .
2: Generate γj ∼ Bernoulli(qj ) where qj = |rxj ,y |δ / max{|rxj ,y |δ }, j = 1, . . . , pn .
IF γj = 1, generate Rn with Ri,j as in (3).
ELSE set Ri,j = 0.
3: Post-multiply Rn with Xn . Set Zn = Xn Rn .
4: For a given Xnew , compute Znew = Xnew Rn and ŷnew = Znew θ̂.
The complexity of steps 1, 2-3, 4 and 5 are O(pn ), O(npγ mn ), O(nm2n ) and O (nnew pγ mn ),
P
respectively, where pγ =
γj . Thus, if nnew ≤ n, the total complexity for a single choice
of (mn , δ, ψ) is O(pn ) + 2O(nmn pγ ) + O(nm2n ) without using parallelization.
RIS-PCR. RIS-PCR differs from RIS-RP in step 2 of the algorithm. After generation
of γ RIS-PCR requires SVD of Xγ involving complexity O (npγ min{n, pγ }). Thus, total
complexity of RIS-PCR for a single choice of (mn , δ, ψ) can similarly be derived. Therefore,
the two methods have comparable time complexity unless either n or pγ is much larger
than mn . Although theoretically we do not impose any restriction on pγ , in practice when
pn = exp{o(n)} and δ ≥ 2, pγ is usually of order n.
Increment of complexity due to aggregation. Suppose N different choices of (mn , ψ, Rn )
are considered. Each choice yields a model Ml : y ∼ f (y|x, mn,l , ψl , Rn,l ) along with a corresponding estimate of ynew (say ŷnew,l ), where l ∈ {1, 2, . . . , N }. The proposed estimate
is the simple average of these N estimates of ynew .
Step 1 in the TARP algorithm is not repeated over the aggregration replicates, while
the remaining steps are repeated N times. In addition, the first step of screening and
12
aggregration are embarassingly parallelizable. Hence, given k processors, if nnew ≤ n, the
total complexity is O(pn /k) + 2O(N nmn pγ /k) + O(N nm2n /k) for RIS-RP and O(pn /k) +
O(N npγ min{n, pγ }/k) + O(N nmn pγ /k) + O(N nm2n /k) for RIS-PCR.
3
Theory on Predictive Accuracy
We study the asymptotic performance of the predictive distribution produced by TARP
for a single random projection matrix without considering the aggregation step. We focus
on weakly sparse and dense cases where the absolute sum of the true regression coefficients
is bounded. This condition also includes strong sparsity where only a few covariates have
non-zero coefficients.
The projection matrix in TARP depends on the random variable γ, and therefore is
denoted by Rγ . We denote a particular realization of the response variable as y, and a
particular realization of the variables (x1 , x2 , . . . , xpn )0 as x. Let f0 be the true density of
y given the predictors, and f (y|x, γ, Rγ , θ) be the conditional density of y given the model
induced by (γ, Rγ ), the corresponding vector of regression coefficients θ and x. We follow
Jiang (2007) in showing that the predictive density under our procedure is close to the true
predictive density in an appropriate sense.
We assume that each covariate xj is standardized so that |xj | < M , for j = 1, 2, . . . , pn ,
with M a constant. We also assume that the scale parameter σ 2 in (1) is known. We
require the following two assumptions on the design matrix.
Assumption (A1) Let rxj ,y denote the correlation coefficient of the observed values
13
of xj and y, 1 ≤ j ≤ pn . Then for each data point (y, x) and constant δ in (2), there exists
a positive constant αδ such that
pn
1 X 2
xj |rxj ,y |δ → αδ .
lim
pn →∞ pn
j=1
Assumption (A2) Let q(γ) =
Qn
γj
(1−γj )
,
i=1 qj (1 − qj )
with qj defined in (2), denote the
probability of obtaining a particular γ = (γ1 , . . . , γpn )0 in the random screening step. Let
Pn
γj = l, and let Ml ⊂ Γl
Γl ⊂ {0, 1}pn denote the set of γ vectors such that pγ = pj=1
denote the first pknn elements of Γl ordered in their q(γ) values. Let An denote the event
that γ ∈ Ml for some l. Then, P (Acn ) = P {γ : γ ∈
/ ∪l Ml } ≤ exp(−nε2n /4), for some
increasing sequence of integers {kn } and sequence {εn } satisfying 0 < ε2n < 1 and nε2n → ∞.
Remark 2. As the probability of selection in the random screening step depends on the
empirical correlation between the predictor and the response, assumption (A2) is on the
data generating process. If l ≤ kn or l ≥ pn − kn , then all models of dimension l belong to
Ml , and none of the corresponding γ vectors belong to Acn . If kn < l < pn − kn , there are
more than pknn models of dimension l, but as the models are ordered in terms of their q(γ)
values, the models not falling in Ml should have extremely small probabilities of selection,
hence satisfying (A2). Violations of (A2) would imply that large numbers of predictors have
empirical correlations that are not close to zero.
Measure of closeness: Let νx (dx) be the probability measure for x, and νy (dy) be the
dominating measure for conditional densities f and f0 . The dominating measure of (y, x)
is taken to be the product of νy (dy)νx (dx).
14
The Hellinger distance between f and f0 is given by
rZ
p
p 2
f − f0 νx (dx)νy (dy).
d(f, f0 ) =
The Kullback-Leibler divergence between f and f0 is given by
Z
f0
νx (dx)νy (dy).
d0 (f, f0 ) = f0 ln
f
t
Z
f0
−1
Define dt (f, f0 ) = t
f0
νx (dx)νy (dy) − 1 , for any t > 0.
f
Consider the following two facts: (i) d(f, f0 ) ≤ (d0 (f, f0 ))1/2 , and (ii) dt (f, f0 ) decreases
to d0 (f, f0 ) as t decreases to 0 (see Jiang (2007)).
Let Pn be a sequence of sets of probability densities, and εn be a sequence of positive
numbers. Let N (εn , Pn ) be the εn -covering number, i.e., the minimal number of Hellinger
balls of radius εn needed to cover Pn .
RIS-RP. The result showing asymptotic accuracy in approximating the predictive density using RIS-RP is stated below.
Theorem 1. Let θ ∼ N (0, σθ2 I), and f (y|x, γ, Rγ , θ) be the conditional density of y given
the model induced by (γ, Rγ ), where Rγ is as in (2) and (3). Let β 0 be the true regression
P
parameter with
j |β0,j | < K for some constant K, and assumptions (A1)-(A2) hold.
Consider the sequence {εn } as in assumption (A2) satisfying 0 < ε2n < 1 and nε2n → ∞,
and assume that the following statements hold for sufficiently large n:
(i) mn | log ε2n | < nε2n /4,
(ii) kn log pn < nε2n /4, and
p
< nε2n /4, where
(iii) mn log 1 + D σθ 6nε2n pn mn
15
D(h∗ ) = h∗ suph≤h∗ |a0 (h)| suph≤h∗ |a0 (h)/b0 (h)|, b(·) as in (1). Then,
h
i
2
2
Pf0 π {d(f, f0 ) > 4εn |Dn } > 2e−nεn /4 ≤ 2e−nεn /5 ,
where π{·|Dn } is the posterior measure.
The proof of Theorem 1 is given in the Appendix.
Remark 3. If we consider the sparse choice of Rγ , as described in Remark 1, the same
line of proof (as that of Theorem 1) would go though. For both the choices of Rγ , each
component of the projection matrix has expectation zero and finite fourth moment. For the
random matrix in (3), the probability of choosing a non-zero element, P (Ri,j 6= 0), is fixed,
while the probability is decaying with n for the sparse choice of Rγ . However, the rate of
decay is such that distances are preserved between the sample-points, a critical property for
proving consistency.
RIS-PCR. Asymptotic guarantees on predictive approximation for RIS-PCR requires an
additional assumption.
Assumption (A3) Let Xγ be the sub-matrix of Xn with columns corresponding to nonzero values of γ, and xγ be a row of Xγ . Let Vγ be the mn × pγ matrix of mn eigenvectors
corresponding to the first mn eigenvalues of Xγ0 Xγ . Then, for each γ and data point xγ ,
kVγ xγ k2
≥ αn ,
kxγ k2
16
where αn ∼ (nε2n )−1 , where the sequence {ε2n } is as in assumption (A2).
Remark 4. If the matrix Xγ0 Xγ has rank less than mn , then αn = 1 by Perseval’s identity.
Consider the situation where rank of the gram matrix, say rn (≤ n), is bigger than mn . Then
the row space of Xγ , or that of Xγ0 Xγ , is spanned by a set of rn basis vectors v1 , v2 , . . . , vrn .
Therefore, any data point x can be written as a linear combination of these rn vectors as
x = a1 v1 + a2 v2 + · · · + arn vrn , where a1 , a2 , . . . , arn are constants not all equal to zero. As
the vectors vj are orthonormal, vj0 x = aj for all j = 1, 2, . . . , rn , which in turn implies that
Pn 2
x0 x = rj=1
aj . Also, note that the first mn among these rn vectors constitute Vγ0 , which
P n 2
Pmn 2 Prn 2
2
2
implies kVγ0 xk2 = m
j=1 aj . Thus kVγ xk /kxk =
j=1 aj /
j=1 aj , and magnitude of the
ratio depends on the part of x explained by the last few principal component directions.
The lower bound αn ∼ (nε2n )−1 is weaker than many real data scenarios where most of the
variation is explained by the first few principal components.
Theorem 2. Let θ ∼ N (0, σθ2 I), and f (y|x, γ, Rγ , θ) be the conditional density of y given
the model induced by (γ, Rγ ), where Rγ is as in (2) and (4). Let β 0 be the true regression
P
parameter with j |β0,j | < K for some constant K, and assumptions (A1)-(A3) hold. Assume that the conditions (i)-(iii) of Theorem 1 hold for the sequence {εn } as in assumption
(A2) satisfying 0 < ε2n < 1 and nε2n → ∞. Then,
h
i
2
2
Pf0 π {d(f, f0 ) > 4εn |Dn } > 2e−nεn /4 ≤ 2e−nεn /5 ,
where π{·|Dn } is the posterior measure.
The proof of Theorem 2 is given in the Appendix.
17
Remark 5. The conditions (i)-(iii) in Theorems 1 and 2 are related to the sizes of pn ,
mn and kn in comparison with nε2n . A sufficient condition for (i) is mn log n < nε2n /4,
providing an upper bound on the dimension of the subspace mn . Condition (ii) restricts the
permissible number of regressors pn , and the number of possible models of each dimension
kn . If there is a strict ordering in the marginal correlation coefficients |rxj ,y |, so that
kn ≤ κ for some large number κ (see assumption (A2)), then the condition reduces to
log pn < nε2n /4. To illustrate that condition (iii) tends to be weak, consider distributions of
y corresponding to Bernoulli, Poisson and normal. For these cases, the quantity D(h∗ ) is
at most order O(h∗ ). Therefore, condition (iii) does not impose much additional restriction
over (i)-(ii), except mn log pn < nε2n /4, inducing a stronger upper-bound to mn .
4
Simulation Study
In this section, we consider different simulation schemes (Scheme I – IV ) to compare
TARP with a variety of methods. We mainly focus on high-dimensional and weakly sparse
regression problems with a variety of correlation structures in the predictors. The sample
size is taken to be 200, while pn varies. Additional results for different choices of n are
provided in the Supplement.
Competitors. We compare with: SCAD screened by iterative SIS (ISIS), ISIS-SCAD;
minimax concave penalty (MCP) method screened by ISIS, ISIS-MCP ; LASSO screened
by sequential strong rule (SSR, Tibshirani et al. (2012)), SSR-LASSO; ridge regression
screened by SSR, SSR-Ridge; elastic net screened by SSR, SSR-EN ; principal component
18
regression (PCR); sparse PCR, SPCR (see Witten et al. (2009)); robust PCR, RPCR (see
Candès et al. (2011)); and Bayesian compressed regression (BCR). ISIS-SCAD and ISISMCP are available in the ‘SIS’ package, and LASSO, ridge and elastic net are available in
the ‘biglasso’ package (Zeng and Breheny (2017)). SPCR and RPCR are performed using
‘PMA’ and ‘rsvd’ packages in R, respectively. To estimate PC scores, we rely on approximate SVD using fast.svd in the ‘corpcor’ package. For BCR, we average over 100 different
random projections with varying mn values within the range [2 log pn , 3n/4]. We use the
qr function in R to apply QR factorization in place of Gram-Schmidt orthogonalization of
the random matrix, which is computationally prohibitive for large pn .
The proposed method. We select the tuning parameters of TARP as described in Section
2.3. The parameter mn is chosen in the range [2 log pn , 3n/4]. We assign δ = 2 as the
function max{0, (1 + log(pn /n))/2} is close to 2 for all the choices of (n, pn ). Further, the
hyperparameters of the inverse gamma priors are set to 0.02 to correspond to a minimally
informative prior.
Simulation Schemes. In the first three simulation schemes, the predictors were generated
from N (0, Σ), with different choices of pn , Σ and the regression coefficients. In Scheme IV
we consider a functional regression setup. Different methods are compared with respect to
their performance in out of sample prediction. We calculate mean square prediction error
(MSPE), empirical coverage probability (ECP) of a 50% prediction interval (PI) and the
width of the PI for each of a 100 replicate datasets in each simulation case.
Scheme I: First order autoregressive structure. Σi,j = (0.9)|i−j| , i, j = 1, . . . , pn , with
19
Figure 1: Box-plots of MSPE for pn = 2000 in Scheme I.
pn ∈ {2000, 3000}, and βj = 0 for all but a randomly selected set of 30 predictors having
βj = 1.
Scheme II: Block diagonal covariance structure. We choose (pn /100 − 2) blocks of 100
predictors each, along with 200 independent predictors, with pn ∈ {104 , 2 × 104 }. The
within-block correlation is ρ and the across-block correlation is zero, with ρ = 0.3 for half
of the blocks and ρ = 0.9 for the remaining half. There are 21 non-zero βj s having βj = 1,
with 20 of the corresponding predictors in the ρ = 0.9 blocks and the remaining in the
independent block.
Scheme III: Principal Component Regression. We first choose a matrix P with orthonormal
columns, and a 3 × 3 matrix D = diag(152 , 102 , 72 ). We set Σ = P DP 0 and choose β = P·,1 ,
where P·,1 is the first column of P . This produces an Xn with three dominant principal
components, with the response yn dependent on the first and pn ∈ {104 , 5 × 104 }.
Scheme IV: Functional Regression. Finally, we consider a functional regression setup, where
the covariates are generated from Brownian bridge Bt with t ∈ (0, 5) and values ranging
from (0, 10). A set of 20 covariates is randomly selected as active, each having regression
20
Figure 2: Box-plots of MSPE for pn = 10000 in Scheme II.
Figure 3: Box-plot of MSPEs for pn = 5000 in Scheme III.
parameters in the range (2, 2.5), and pn ∈ {104 , 2 × 104 }.
Results. For each of the simulation schemes, we present the results corresponding to the
first choice of pn , with the other results provided in the Supplement.
In Scheme I (see Figure 1) SSR-Ridge, PCR, RIS-RP and RIS-PCR show competitive
performance, with RIS-PCR showing the best overall result. Performance of RIS-PCR is
closely followed by RIS-RP, which in turn is followed by PCR and SSR-Ridge. Apart from
these four methods, ISIS based approaches and SSR-EN exhibit reasonable performance.
Although ISIS based methods have lower average MSPE, the variances of MSPEs are high
indicating less stability. SPCR, RPCR, BCR and SSR-LASSO fail to perform adequately.
21
(a) MSPE of all the methods.
(b) MSPE of selected methods.
Figure 4: Box-plot of MSPEs for pn = 10000 in Scheme IV.
For Scheme II, PCR, RIS-RP and RIS-PCR again have the best performance (see Figure
2). RIS-PCR yields lowest overall MSPE, closely followed by RIS-RP and PCR. SSR-EN,
SSR-Ridge and RPCR exhibit moderate performance. Performance of these three methods
are followed by LASSO, SPCR and BCR. Although SPCR enjoys better average MSPE, it
yields higher dispersion as well. ISIS based methods fail to perform adequately.
In Scheme III (see Figure 3) SSR based methods fail to exhibit competitive performance.
Among the SSR based methods, performance of SSR-LASSO is better than the others. All
the other methods perform well, although ISIS based methods occasionally show higher
MSPEs.
In Scheme IV (see Figure 4(a)), SPCR and ISIS based methods fail completely, making
the other methods indistinguishable in the figure. Hence, we separately show these methods
in Figure 4(b). Among the other methods, PCR, RIS-RP and RIS-PCR have the best
overall performance closely followed by BCR and then by SSR-EN. The other three methods
fail to exhibit competitive performance.
22
We next consider the empirical coverage probabilities (ECPs) of 50% prediction intervals
(PI), and the width of PIs. For the Bayesian methods, the PI is obtained from the highest
credible region of the predictive distribution of ynew given Dn , Xnew . The PIs for the freq
quentist methods can be obtained as ynew ±tn−pγ −1, α2 M SP E 1 + x0γ,new (Xγ0 Xγ )−1 xγ,new ,
where tn−pγ −1, α2 is the upper α/2 point of t distribution with (n − pγ − 1) degrees of freedom, the suffix γ indicates consideration of the regressors selected by the corresponding
method, and pγ is the number of selected regressors. The PIs for PCR based methods
can be obtained similarly, except xγ,new and Xγ are replaced by the principal component
scores. For SSR-Ridge and SSR-EN, pγ is much larger than n. Therefore, the PIs of these
methods could not be calculated using the above formula, instead we obtain the interval as
p
√
ynew ± tn, α2 M SP E + se(ŷ|Dn , xnew )2 , or ynew ± tn, α2 2M SP E, whichever gives better re-
sults, where se(ŷ|Dn , xnew )2 is the variance of the fitted values. The results are summarized
in Table 1.
Summary. Considering all the simulation schemes, RIS-RP, RPCR, SPCR and SSRLASSO have the best performance with respect to mean ECP, with RIS-RP having lowest
average width among these four methods. The average width of PI is much bigger for
SPCR, particularly in Scheme IV. ISIS based methods and RIS-PCR have relatively lower
coverage probabilities in general, although among these methods, RIS-PCR has higher coverage with much lower average width than others, especially in Scheme IV. The average
ECP for PCR is satisfactory, although the corresponding widths of PIs have large variances
in almost all the simulation schemes. This indicates instability in the overall performance
23
Table 1: Mean and standard deviation(sd) of empirical coverage probabilities of 50% prediction intervals, and mean and sd of width of 50% prediction intervals.
Methods→
Scheme, pn
ISISSACD
ISISSACD
SSRSSRLASSO EN
SSRRidge
PCR
SPCR
RPCR BCR
RISRP
RISPCR
Average and standard deviation (in braces) of empirical coverage probability
I
2 × 103
0.309
(.052)
0.311
(.058)
0.432
(.051)
0.549
(.040)
0.692
(.055)
0.498
(.051)
0.457
(.055)
0.433
(.290)
0.286
(.053)
0.493
(.059)
0.431
(.057)
II
104
0.327
(.056)
0.324
(.055)
0.429
(.049)
0.607
(.062)
0.702
(.075)
0.455
(.308)
0.499
(.058)
0.445
(.050)
0.278
(.049)
0.502
(.056)
0.362
(.050)
III
5 × 103
0.494
(.053)
0.494
(.053)
0.503
(.058)
0.657
(.031)
0.678
(.127)
0.503
(.058)
0.494
(.049)
0.494
(.049)
0.527
(.054)
0.494
(.055)
0.507
(.054)
0.425
(.053)
0.416
(.056)
0.477
(.065)
0.660
(.031)
0.665
0.448
(0.102) (.300)
0.488
(.048)
0.491
(.053)
0.487
(.058)
0.696
(.052)
0.418
(.055)
IV
104
Average and standard deviation (in braces) of width of the 50% prediction interval
I
2 × 103
3.894
(.500)
3.857
(.480)
5.939
(.424)
7.410
(.578)
8.192
(.528)
7.739
7.964
(8.309) (.700)
6.483
(.451)
5.260
(.416)
5.029
(.194)
4.253
(.365)
II
104
5.868
(.657)
5.836
(.684)
5.113
(.527)
7.335
(.615)
8.139
(.830)
8.407
6.066
4.635
(25.713) (1.103) (.407)
4.894
(.433)
4.204
(.198)
2.744
(.232)
III
5 × 103
1.424
(.174)
1.385
(.144)
1.362
(.072)
1.945
(.145)
5.755
1.362
(1.928) (.072)
1.366
(.069)
1.463
(.073)
1.351
(.079)
1.391
(.077)
13.63
13.34
2.170
(2.705) (2.540) (.422)
1.537
(.139)
3.582
(.501)
2.508
26.512 2.480
(3.169) (6.209) (.345)
1.792
(.115)
2.284
(.268)
1.154
(.123)
IV
104
1.366
(.069)
of PCR. BCR shows under-coverage in the first two simulation schemes, but performs well
with respect to both the measures in Schemes III and IV. Finally the other two methods,
viz., SSR-Ridge and SSR-EN have higher values of ECP, along with higher width of PIs.
SSR-Ridge has highest average width of PI in Schemes I and III. In all the simulation
schemes SSR-EN outperforms SSR-Ridge with respect to width of PI.
4.1
Computational Time
The computational time of a method may depend on the simulation scheme due to varying
level of complexity in the dataset. We only present the computational time for Scheme
24
IV as an example. Figure 5 presents the time (in minutes) taken by different methods to
compute ŷnew using a single core, as pn grows and n = nnew = 100. We run all the methods
in R 3.4.1 in a 64 bit Dell-Inspiron desktop with Ubuntu 16.04 LTS operating system, 15.6
GB random access memory and Intel R CoreTM i5-4460 CPU @ 3.20GHz processor.
Results. When pn is below 104 all the methods require comparable computational
time. When pn is increased, SSR based methods, except SSR-Ridge, RPCR, and fast.svd
based PCR continue to require low computational time. ISIS-SCAD, SPCR and RIS-RP
also have reasonable computational expense (approximately, 5 minutes for pn = 5 × 105 ).
Computational time of BCR, ISIS-MCP and RIS-PCR tends to increase rapidly after pn =
105 . Among these three methods, RIS-PCR requires highest system time (approximately
27 minutes for pn = 5 × 105 ). The computational time required by SSR-Ridge exceeds all
other methods for pn > 5×104 , and for pn = 5×105 it becomes computationally prohibitive
(it takes more than 2 hours).
The increment of computational time of RIS-PCR is due to the computation of exact
SVD of the screened design matrix Xγ . However, this burden would immediately be reduced
if one uses some approximation of the SVD. In that case the computational time would be
comparable to RIS-RP.
5
Real Data Analysis
In this section, we study the performance of TARP using three real datasets, viz., Golub
dataset, GTEx dataset and Eye dataset. The Golub dataset is available at GitHub https:
25
Figure 5: System time required by different methods to predict y as pn grows.
//github.com/ramhiser/datamicroarray/wiki, the GTEx dataset is available at GTEx
portal https://www.gtexportal.org/home/ and the eye data is available in the flare
package of R. In each case, we assess out-of-sample predictive performance averaging over
multiple training-test splits of the data.
Golub data. The Golub data consist of 47 patients with acute lymphoblastic leukemia
(ALL) and 25 patients with acute myeloid leukemia (AML). Each of the 72 (= n) patients
had bone marrow samples obtained at the time of diagnosis (see Golub et al. (1999)).
Expression levels of 7129 (= pn ) genes have been measured for each patient. We consider
a training set of size 60 with 20 AML patients, and 40 ALL patients. The test set consists
of the remaining 12 samples.
GTEx Data. To understand the functional consequences of genetic variation, Consortium
26
et al. (2015) presented an analysis of RNA sequencing data from 1641 samples across 43
tissues from 175 individuals, generated as part of the pilot phase of the Genotype-Tissue
Expression (GTEx) project. We selected RNA-seq data on two normal tissues, viz., ArteryAorta and Artery-Tibial. The dataset contains RNA-seq expressions on 36115 (= pn ) genes
and 556 (= n) samples, among which 224 are from Artery-Aorta, and 332 are from ArteryTibial. A training set of 100 samples from each of the tissue types is considered, and the
remaining 446 samples are used as test set.
Eye Data. The Eye dataset consists of gene expressions for 200 (= pn ) gene probes
from the microarray experiments of mammalian-eye tissue samples of 120 (= n) rats (see
Scheetz et al. (2006)). The response variable is the expression level of the TRIM32 gene.
We consider 100 sample points as the training set, and the remaining 20 samples as the
test set.
Golub and GTEx datasets have nominal response, and therefore the methods are evaluated by the misclassification rate (in %) and the area under the receiver operating characteristic (ROC) curve. Table 2 provides the average and standard deviation (sd) of percentages
of misclassification, and those for the area under the ROC curve over 100 random subsets
of the same size chosen from the dataset for the competing methods. We further compare
the predictive performance of the methods in terms of mean squared difference of predictive and empirical probabilities for these two datasets. Most methods (except SSR based
methods and SPCR) exhibit similar performance in this aspect. We provide the details of
the predictive calibration in the Supplement.
The eye dataset has continuous response, and therefore we evaluate the methods by
27
Table 2: Mean and standard deviation (in braces) of percentage of misclassification and
area under the ROC curve for Golub and GTEx datasets, and those of MSPE, ECP of 50%
PI and width of PI for Eye dataset for all the competing methods.
Methods→
Dataset ↓
ISISSACD
ISISSACD
SSRSSRLASSO EN
SSRRidge
PCR
SPCR
RPCR BCR
RISRP
RISPCR
Misclassification rate and Area under ROC curve for Datasets with Categorical response
Mean and SD of Misclassification Rate (in %)
Golub
11.82
(6.90)
11.50
(7.06)
45.45
(0.00)
45.45
(0.00)
45.45
(0.00)
7.09
(5.68)
41.36
9.73
(13.31) (7.28)
19.36
(9.79)
5.54
(4.36)
5.77
(4.52)
GTEx
0.00
(0.00)
0.00
(0.00)
34.83
(0.00)
34.83
(0.00)
34.83
(0.00)
0.06
(0.13)
3.53
(3.31)
13.28
(3.79)
0.39
(0.20)
0.49
(0.32)
Golub
0.876
(.073)
0.879
(.074)
0.500
(.000)
0.500
(.000)
0.500
(.000)
0.923
(.062)
0.582
(.134)
0.895
(.078)
0.816
(.093)
0.978
(.027)
0.943
(.044)
GTEx
1.00
(.000)
1.00
(.000)
0.500
(.000)
0.500
(.000)
0.500
(.000)
0.999
(.001)
0.964
(.033)
0.998
(.001)
0.877
(.041)
1.00
(.000)
.996
(.002)
10.01
(4.04)
8.54
(3.09)
8.29
(2.99)
0.22
(0.18)
Mean and SD of Area under Receiver Operating Characteristic curve
MSPE, ECP of 50% PI and Width of 50% PI
Mean and SD of Mean Square Prediction Error
Eye
11.66
(4.06)
11.66
(4.06)
20.92
20.92
7.31
(19.33) (19.33) (2.91)
13.84
(3.94)
8.65
(3.08)
7.67
(3.30)
Mean and SD of Empirical Coverage Probability and Width of the Prediction Interval
Empirical
coverage
0.502
(.138)
0.502
(.138)
0.634
(.130)
0.709
(.106)
0.700
(.076)
0.423
(.325)
0.508
(.123)
0.522
(.114)
0.564
(.117)
0.598
(.101)
0.507
(.107)
Width of
interval
1.208
(.057)
1.208
(.057)
1.970
(.190)
2.033
(.917)
1.539
(.303)
1.884
1.202
(1.612) (.079)
1.055
(.049)
1.249
(.056)
1.341
(.038)
1.056
(.036)
MSPE and empirical coverage probabilities (ECP) of 50% prediction intervals (PI) as in
Section 4. As variation in the expression levels of the TRIM32 gene is very small (the
range is 1.37) we multiply the MSPEs of different methods by 10 to increase the variability.
Table 2 provides the mean and sd of MSPEs, ECPs of 50% PIs, and widths of the PIs over
100 different training and test sets selected from the dataset, for the competing methods.
28
Results. For the Golub data set, both the lowest misclassification rate and the highest area
under ROC curve are achieved by RIS-RP, which is closely followed by RIS-PCR. TARP
based methods attain lower sd than other methods as well. PCR also yields reasonable
performance with 7% average misclassification rate and area under ROC more than 0.9.
RPCR and ISIS-based methods produce average rates of misclassification of at least 10%,
with area under the ROC of ∼ 0.9. BCR possesses high misclassification rate (about 19%),
although area under ROC is more than 0.8. Neither the MSPE, nor the area under ROC
curve, is satisfactory for SPCR. Finally, for all the 100 repetitions, SSR based methods
invariably select the intercept-only model. Thus, the MSPEs of these methods depend
entirely on the proportion of test samples obtained from the two classes.
For the GTEx dataset, perfect classification is achieved by ISIS based methods. These
methods along with RIS-RP have the highest area under the ROC curve. PCR, RPCR,
RIS-RP and RIS-PCR also yield satisfactory results, having less than 0.5% average misclassification rate and more than 99% area under the ROC curve. SPCR is comparable with an
average MSPE of less than 4%. BCR attains 13.3% average misclassification rate, with the
area under the ROC curve almost 0.9. SSR based methods fail to show any discriminatory
power in the GTEx dataset.
SSR-Ridge, RPCR, RIS-PCR, RIS-RP and SPCR yield excellent performance in terms
of MSPE in the eye dataset with an average MSPE of less than 0.9. SSR-Ridge has an
average ECP of about 0.7. RIS-PCR shows more stable performance in terms of ECP,
followed by SPCR. BCR and ISIS based methods have similar overall performance. In
terms of MSPE, BCR outperforms ISIS based methods but is outperformed by ISIS based
29
methods in terms of ECP. PCR is not quite as good in terms of either measure. SSR-LASSO
and SSR-EN again fail to perform adequately for the eye dataset.
The GEUVADIS cis-eQTL dataset
We conclude this section by illustrating the
TARP approach on a massive dataset. The GEUVADIS cis-eQTL dataset (Lappalainen
et al. (2013)) is publicly available at http://www.ebi.ac.uk/Tools/geuvadis-das/. This
dataset consists of messenger RNA and microRNA on lymphoblastoid cell line (LCL) samples from 462 individuals provided by the 1000 Genomes Project along with roughly 38
million SNPs. E2F2 plays a key role in the control of the cell cycle. Hence, as in Chen
and Dunson (2017), we choose the gene E2F2 (Ensemble ID: ENSG00000000003) as the
response. A total of 8.2 million (= pn ) SNPs are pre-selected as candidate predictors on
the basis of having at least 30 non-zero expressions. The total number of subjects included in the dataset is about 450 (= n). The genotype of each SNP is coded as 0, 1 or 2
corresponding to the number of copies of the minor allele.
TARP is applied on this dataset. We consider four different training sample sizes, viz.,
nt = 200, 250, 300 and 350, and test sample size 100 in each case. As pn is huge, we
applied three different values of δ, namely, 2, 5 and 8, to analyze the effect of a conservative
screening. The recommended choice of δ lies within (5, 6) when pn = 8.2 × 106 and n ∈
[200, 400]. To perform SVD for RIS-PCR, we use fast.svd instead of the usual svd to cope
with the massive number of regressors. Table 3 provides the MSPE, the ECP of 50% PI
and width of the PI, obtained by two different variants of TARP.
Results: The MSPEs of RIS-RP and RIS-PCR are comparable for all the choices on
n. However, RIS-RP yields much better empirical coverage probabilities than RIS-PCR,
30
Table 3: MSPE, ECP and width of PI (in order) obtained by RIS-RP and RIS-PCR for
three values of δ and different training sample sizes (nt ).
nt
200
250
300
350
δ=2
MSPE ECP Width
0.800 0.39 1.059
0.852 0.39 1.102
0.860 0.36 1.126
0.778 0.45 1.210
RIS-RP
δ=5
MSPE ECP Width
0.872 0.42 0.983
0.920 0.42 1.023
0.855 0.44 1.075
0.779 0.48 1.221
δ=8
MSPE
0.855
0.921
0.866
0.829
ECP
0.34
0.35
0.36
0.46
Width
0.928
1.013
1.069
1.219
nt
200
250
300
350
δ=2
MSPE
0.834
0.858
0.845
0.757
RIS-PCR
δ=5
MSPE ECP Width
0.838 0.12 0.192
0.882 0.12 0.289
0.867 0.20 0.511
0.786 0.36 0.886
δ=8
MSPE
0.831
0.896
0.865
0.826
ECP
0.10
0.19
0.20
0.41
Width
0.252
0.420
0.487
0.984
ECP
0.06
0.14
0.14
0.35
Width
0.177
0.355
0.399
0.893
especially when n ≤ 300. The three choices of δ yield comparable results in terms of all
the measures in general. For RIS-RP, δ = 5 results in higher ECP and for RIS-PCR higher
ECP is obtained using δ = 8. Moreover, the choice δ = 8 makes both the procedures
much faster compared to other choices of δ. When the training sample is 350, δ = 2, 5
and 8 select about 290800, 12600 and 7960 variables, respectively, on an average in the
screening stage out of 8.2 × 106 variables. In view of the results in this massive dimensional
dataset, it seems reasonable to use a higher value of δ for filtering out noisy regressors, and
computational convenience.
31
6
Appendix
This section contains proofs of the theorems stated in the paper. We use a generic notation c for
the constants, although all of them may not be equal.
Some Useful Results
Lemma 1. Let εn be a sequence of positive numbers such that nε2n 1. Then under conditions
(a) lnN (εn , Pn ) ≤ nε2n for all sufficiently large n.
2
(b) π (Pnc ) ≤ e−2nεn for all sufficiently large n.
(c) π f : dt (f, f0 ) ≤ ε2n /4 ≥ exp{−nε2n /4} for all sufficiently large n and for some t > 0.
Then
i
h
2
2
Pf0 π {d(f, f0 ) > 4εn |(yn , X)} > 2e−nεn (0.5∧(t/4)) ≤ 2e−nεn (0.5∧(t/4)) .
The proof is given in Jiang (2007).
Lemma 2. Suppose assumption (A1) holds. Let αδ be such that
P
j
x2j |rxj ,y |δ /pn → αδ as n → ∞
and x be a pn × 1 sample vector of the regressors. Then the following holds:
a. The random matrix Rγ described in (2) and (3) satisfies
p
kRγ xk2 /pn →
− cαδ .
b. Let kxγ k2 =
Ppn
2
j=1 xj I(γj
= 1) where γj is the j th element of the vector γ described in (2) and
I(·) is the indicator function, then
p
kxγ k2 /pn →
− cαδ ,
32
where c is the proportionality constant in (2).
The proof is given in the Supplement.
Lemma 3. Let θ ∼ N (0, σθ2 I), then for a given Rγ , x and y the following holds
24 ∆4
(x0 β )2 + ∆2
.
P (|(Rγ x)0 θ − x0 β 0 | < ∆) > exp − 2 0
2
σθ kRγ xk2
σθ kRγ xk2
The proof is given in Guhaniyogi and Dunson (2015).
Proofs of the Theorems. Without loss of generality we consider |xj | < M with M = 1,
j = 1, 2, . . . , pn and d(σ 2 ) = 1, although the proofs go through for any fixed value of M and σ 2 .
Proof of Theorem 1. Define the sequence of events
n
o
2
2
Bn = π {d(f, f0 ) > 4εn |(yn , X)} > 2e−nεn /4 and we need to show P (Bnc ) > 1 − 2e−nε /5 .
We first consider the sequence of events An in assumption (A2), and show that P (Bnc |An ) >
2 /4
1 − 2e−nε
. The proof then follows from assumption (A2) for moderately large n.
The proof of P (Bnc |An ) > 1 − 2e−nε
2 /4
hinges on showing the three conditions of Lemma 1
for the approximating distribution
f (y) = exp{ya(h) + b(h) + c(y)} with h = (Rγ x)0 θ,
(5)
and the true distribution f0 , where Rγ is as given in (4).
Checking condition (a). Let Pn be the set of densities f (y) stated above with parameter |θj | < cn ,
√
j = 1, 2, . . . , mn , where {cn } = {σθ 5nεn } and the model γ is such that γ ∈ Mk , for some
k ∈ {0, 1, . . . , pn }, given An . For any γ the corresponding set of regression parameters can be
n
covered by l∞ balls of the form B = (vj − , vj + )m
j=1 of radius > 0 and center vj . It takes
33
(cn / + 1)mn balls to cover the parameter space for each of model γ in Pn . There are at most
min{ pkn , pknn } models for each γ under consideration as we are only concerned with models in
An (see assumption (A2)), and there are (pn + 1) possible choices of k. Hence it requires at most
N (, k) ≤ cpnkn +1 (cn / + 1)mn l∞ balls to cover the space of regression parameters Pn , for some
constant c.
Next we find the number of Hellinger balls required to cover Pn . We first consider the KL
distance between f and f0 , then use the fact d(f, f0 ) ≤ (d0 (f, f0 ))1/2 . Given any density in Pn , it
n
can be represented by a set of regression parameters (uj )m
j=1 falling in one of these N (, k) balls
n
B = (vj − , vj + )m
j=1 and pγ = k. More specifically, let fu and fv be two densities in Pn of the
form (5), where u = (Rγ x)0 θ 1 , v = (Rγ x)0 θ 2 with |θi,j | < cn , i = 1, 2 and pγ = k, then
d0 (fu , fv ) =
Z Z
Z Z
fv log
fv
fu
νy (dy)νx (dx)
{y(a(u) − a(v)) + (b(u) − b(v))} fv νy (dy)νx (dx)
0
Z
b (v)
0
0
=
(u − v) a (uv ) − 0
+ b (uv ) νx (dx).
a (v)
=
The last expression is achieved by integrating with respect to y and using mean value theorem,
where uv is an intermediate point between u and v. Next consider
|u − v| = |(Rγ x)0 θ 1 − (Rγ x)0 θ 2 | ≤ kRγ xkkθ 1 − θ 2 k,
p
using the Cauchy-Schwartz inequality. Now, by Lemma 2 we have kRγ xk2 /pn →
− αδ as
n → ∞ for some constant 0 < αδ < 1. Therefore we can assume that for sufficiently large pn ,
√
√
√
kRγ xk ≤ pn . Also, kθ 1 − θ 2 k ≤ mn . Combining these facts we have |u − v| ≤ mn pn .
√
Similarly max{|u|, |v|} ≤ cn mn pn . These together imply that
(
)
0 (h)|
|b
√
d0 (fu , fv ) ≤ mn pn
sup
|a0 (h)|
sup
.
0
√
√
|h|≤cn mn pn
|h|≤cn mn pn |a (h)|
34
Therefore d(fu , fv ) ≤ εn if we choose
o
n√
mn pn sup|h|≤cn √mn pn |a0 (h)| sup|h|≤cn √mn pn (|b0 (h)|/|a0 (h)|) .
= ε2n /
Therefore, density fu falls in a Hellinger ball of size εn , centered at fv . As shown earlier, there
are at most N (, k) such balls. Thus, the Hellinger covering number
c
mn
n
N (εn , Pn ) ≤ N (, k) = cpknn +1
+1
(
)
!#mn
"
cn √
|b0 (h)|
kn +1
0
mn pn
sup
+1
= cpn
|a (h)|
sup
0
√
√
ε2n
|h|≤cn mn pn
|h|≤cn mn pn |a (h)|
mn
1
√
kn +1
≤ cpn
D(cn mn pn ) + 1
,
ε2n
where D(R) = R suph≤R |a0 h| suph≤R |b0 (h)/a0 (h)|. The logarithm of the above quantity is no
√
more than log c + (kn + 1) log pn − mn log(ε2n ) + mn log 1 + D(cn mn pn ) , as 0 < ε2n < 1. Using
the assumptions in Theorem 1 condition (a) follows.
n
Checking condition (b) For the Pn defined in condition (a), π(Pnc ) ≤ π(∪m
j=1 |θj | > cn ).
q
Observe that π(|θj | > cn ) ≤ 2 exp{−c2n /(2σθ2 )}/ 2πc2n /σθ2 by Mills ratio. Now for the choice that
p
√
cn = σθ 5nεn the above quantity is 2 exp{−5nε2n /2}/ 10πnε2n . Therefore
π(Pnc ) ≤
mn
X
j=1
π(|θj | > cn ) ≤ 2mn exp{−5nε2n /2}/
p
2
10πnε2n ≤ e−2nεn
for sufficiently large n. Thus condition (b) follows.
Checking condition (c) Condition (c) is verified for t = 1. Observe that
dt=1 (f, f0 ) =
Integrating out y we would get
R
Z Z
f0
f0
− 1 νy (dy)νx (dx).
f
Ey|x [{(f0 /f )(Y ) − 1}] νx (dx). Note that under f and f0 we
35
have same function of y as given in (5) with h = x0 β 0 for f0 .
Therefore, the above can
be written as Ex [{(Rγ x)0 θ − x0 β 0 } g (u∗ )] using mean value theorem where g is a continuous
derivative function, and u∗ is an intermediate point between (Rγ x)0 θ and x0 β 0 . Therefore, if
|(Rγ x)0 θ − x0 β 0 | < ∆n , then |u∗ | < |x0 β 0 | + ∆n . This in turn implies that for sufficiently small
∆n , |g(u∗ )| will be bounded, say by M . Consider a positive constant ∆n . From Lemma 3 we have
P (|(Rγ x)0 θ − x0 β 0 | < ∆n ) =
X
γ
P (|(Rγ x)0 θ − x0 β 0 | < ∆n |γ)π(γ)
24 ∆4
(x0 β 0 )2 + ∆2n
≥ Eγ exp − 2
σθ kRγ xk2
σθ2 kRγ xk2
Zγ
Zγ
24 ∆4n
=
,
Eγ
exp −
(x0 β 0 )2 + ∆2n
pn
pn
(6)
where Zγ = (x0 β 0 )2 + ∆2n / σθ2 kRγ xk2 /pn . By part (a) of Lemma 2, and continuous mapping
p
theorem Zγ →
− (x0 β 0 )2 + ∆2n / σθ2 cαδ > ∆2n / σθ2 cαδ .
For some non-negative random variable Z and non-random positive numbers p, a and b,
consider the following fact
E
Z
Z
Z
Z
exp −
≥ aP
exp −
>a
p
p
p
p
Z
a
Z
≥ aP
> , exp −
> ab
p
b
p
ap
= aP Z >
, Z < −p log(ab)
b
ap
< Z < −p log(ab) .
= aP
b
(7)
Replacing Z by Zγ , p by pn and taking a = ∆2n exp{−nε2n /3}/(σθ2 cαδ ), and b = pn exp{−nε2n /3}.
Thus −pn log(ab) = −pn log ∆2n pn exp{−2nε2n /3}/(σθ2 cαδ ) > pn nε2n /2 and apn /b = ∆2n / σθ2 cαδ
36
for sufficiently large n. Therefore the expression in (7) is greater than
∆2n −nε2n /3
e
P
σθ2 cαδ
Note that (x0 β 0 )2 <
Ppn
j=1 |β0,j |
∆2n
1
≤ Zγ ≤ pn nε2n
2
2
σθ cαδ
< K, and the probability involved in the above expression can be
shown to be bigger than some positive constant p for sufficiently large n. Using these facts along
with equation (6), we have P (|(Rγ x)0 θ − x0 β 0 | < ∆n ) > exp{−nε2 /4}. Choosing ∆n < ε2 /(4M )
condition (c) follows.
Proof of Theorem 2. The outline of the proof of Theorem 2 closely follows the arguments given
in the proof of Theorem 1. Therefore we only present those parts of the proof which are different.
2 /4
As in Theorem 1, we show that P (Bnc |An ) > 1 − 2e−nε
by checking the three conditions of
Lemma 1.
The proof of Condition (a) is the same as for Theorem 1, except for the places involving
the projection matrix Rγ . Observe that given a dataset Dn and other tuning parameters we fix
a particular projection matrix Rγ . The only property of Rγ needed to prove condition (a) is
kRγ xk2 ≤ pn for sufficiently large n. To show this we use that fact that Rγ is a matrix with
orthonormal row vectors, and therefore kRγ xk2 ≤ kxγ k2 ≤ pn .
The proof of Condition (b) depends only on the prior assigned on θ, and therefore remains
the same under the conditions of Theorem 2.
The proof of Condition (c) differs from that of Theorem 1 in showing P (|(Rγ x)0 θ − x0 β 0 | <
∆n ) > exp{−nε2 /4} for some constant ∆n . To see this consider a positive constant ∆n . As
before, from Lemma 3 we have
37
(x0 β 0 )2 + ∆2n
24 ∆4
P (|(Rγ x) θ − x β 0 | < ∆n ) ≥ Eγ exp − 2
σθ kRγ xk2
σθ2 kRγ xk2
(x0 β 0 )2 + ∆2n
24 ∆4
≥ Eγ exp − 2
σθ αn kxγ k2
σθ2 kxγ k2
Zγ
Zγ
24 ∆4n
= 0
exp
−
,
E
γ
(x β 0 )2 + ∆2n
pn
αn pn
0
0
(8)
where Zγ = (x0 β 0 )2 + ∆2n / σθ2 kxγ k2 /pn , and αn is as in (A3). From part (b) of Lemma 2,
p
and continuous mapping theorem we have Zγ →
− (x0 β 0 )2 + ∆2n / σθ2 cαδ > ∆2n / σθ2 cαδ .
For some positive random variable Z and non-random positive numbers p, a and b, consider
the following
E
Z
Z
Z
Z
exp −
≥ aP
exp −
≥a
p
αp
p
αp
Z
a
Z
≥ aP
≥ , exp −
≥ ab
p
b
αp
ap
= aP
< Z < −αp log(ab) .
b
(9)
(10)
Replacing Z by Zγ , p by pn , α by αn and taking a = ∆2n exp{−nε2n /3}/(σθ2 cαδ ), and b =
pn exp{−nε2n /3}. Thus,
−αn pn log(ab) = −αn pn log ∆2n pn exp{−2nε2n /3}/(σθ2 cαδ ) ∼ 2pn log ∆2n pn /(σθ2 cαδ ) /3 > 2pn /3
for sufficiently large n and apn /b = ∆2n / σθ2 cαδ . Therefore the expression in (10) is greater than
∆2n
2
∆2n −nε2n /3
e
P
≤ Zγ ≤ p n .
3
σθ2 cαδ
σθ2 cαδ
P
n
Note that (x0 β 0 )2 < pj=1
|β0,j | < K, and the probability involved in the above expression
can be shown to be bigger than some positive constant p for sufficiently large n. Using these
facts along with equation (8), we have P (|(Rγ x)0 θ − x0 β 0 | < ∆n ) > exp{−nε2 /4}. Choosing
38
∆n < ε2 /(4M ) condition (c) follows.
References
Achlioptas, D. (2003). Database-friendly random projections: Johnson-Lindenstrauss with binary
coins. Journal of Computer and System Sciences, 66(4):671–687.
Candes, E. and Tao, T. (2007). The Dantzig selector: statistical estimation when p is much larger
than n. The Annals of Statistics, 35(6):2313–2351.
Candès, E. J., Li, X., Ma, Y., and Wright, J. (2011). Robust principal component analysis?
Journal of the ACM, 58(3):Art. 11, 37.
Candes, E. J. and Tao, T. (2005). Decoding by linear programming. IEEE Transactions on
Information Theory, 51(12):4203–4215.
Carvalho, C. M., Polson, N. G., and Scott, J. G. (2009). Handling sparsity via the horseshoe.
In International Conference on Artificial Intelligence and Statistics, volume 5 of Proceedings of
Machine Learning Research, pages 73–80.
Chen, Y. and Dunson, D. B. (2017). Modular Bayes screening for high-dimensional predictors.
ArXiv preprint arXiv:1703.09906.
Consortium, G. et al. (2015). The genotype-tissue expression (GTEx) pilot analysis: Multitissue
gene regulation in humans. Science, 348(6235):648–660.
Dasgupta, S. and Gupta, A. (2003). An elementary proof of a theorem of Johnson and Lindenstrauss. Random Structures & Algorithms, 22(1):60–65.
39
Davenport, M. A., Boufounos, P. T., Wakin, M. B., and Baraniuk, R. G. (2010). Signal processing with compressive measurements. IEEE Journal of Selected Topics in Signal Processing,
4(2):445–460.
Donoho, D. L. (2006).
Compressed sensing.
IEEE Transactions on Information Theory,
52(4):1289–1306.
Fan, J. and Li, R. (2001). Variable selection via nonconcave penalized likelihood and its oracle
properties. Journal of the American Statistical Association, 96(456):1348–1360.
Fan, J., Samworth, R., and Wu, Y. (2009). Ultrahigh dimensional feature selection: Beyond the
linear model. Journal of Machine Learning Research, 10(Sep):2013–2038.
Fard, M. M., Grinberg, Y., Pineau, J., and Precup, D. (2012). Compressed least-squares regression
on sparse spaces. In Association for the Advancement of Artificial Intelligence, pages 1055–1060.
Golub, T. R., Slonim, D. K., Tamayo, P., Huard, C., Gaasenbeek, M., Mesirov, J. P., Coller, H.,
Loh, M. L., Downing, J. R., Caligiuri, M. A., Bloomfield, C. D., and Lander, E. S. (1999).
Molecular classification of cancer: Class discovery and class prediction by gene expression
monitoring. Science, 286(5439):531–537.
Guhaniyogi, R. and Dunson, D. B. (2015). Bayesian compressed regression. Journal of the
American Statistical Association, 110(512):1500–1514.
Hoeting, J. A., Madigan, D., Raftery, A. E., and Volinsky, C. T. (1999). Bayesian model averaging:
A tutorial. Statistical Science. A Review Journal of the Institute of Mathematical Statistics,
14(4):382–417.
40
Ishwaran, H. and Rao, J. S. (2005). Spike and slab variable selection: Frequentist and Bayesian
strategies. The Annals of Statistics, 33(2):730–773.
Jiang, W. (2007). Bayesian variable selection for high dimensional generalized linear models:
Convergence rates of the fitted densities. The Annals of Statistics, 35(4):1487–1511.
Kabán, A. (2014). New bounds on compressive linear least squares regression. In Artificial
Intelligence and Statistics, pages 448–456.
Lappalainen, T., Sammeth, M., Friedländer, M. R., AC‘t Hoen, P., Monlong, J., Rivas, M. A.,
Gonzalez-Porta, M., Kurbatova, N., Griebel, T., Ferreira, P. G., et al. (2013). Transcriptome
and genome sequencing uncovers functional variation in humans. Nature, 501(7468):506.
Li, P., Hastie, T. J., and Church, K. W. (2006). Very sparse random projections. In Proceedings of
the 12th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining,
KDD ’06, pages 287–296.
Maillard, O. and Munos, R. (2009). Compressed least-squares regression. In Advances in Neural
Information Processing Systems, pages 1213–1221.
Pettenuzzo, D., Koop, G., Korobilis, D., et al. (2016). Bayesian compressed vector autoregressions. Working Papers 103, Brandeis University, Department of Economics and International
Businesss School.
Scheetz, T. E., Kim, K.-Y. A., Swiderski, R. E., Philp, A. R., Braun, T. A., Knudtson, K. L.,
Dorrance, A. M., DiBona, G. F., Huang, J., Casavant, T. L., et al. (2006). Regulation of gene
expression in the mammalian eye and its relevance to eye disease. Proceedings of the National
Academy of Sciences, 103(39):14429–14434.
41
Thanei, G.-A., Heinze, C., and Meinshausen, N. (2017). Random projections for large-scale
regression. In Big and Complex Data Analysis, pages 51–68.
Tibshirani, R. (1996). Regression shrinkage and selection via the lasso. Journal of the Royal
Statistical Society. Series B. Methodological, 58(1):267–288.
Tibshirani, R., Bien, J., Friedman, J., Hastie, T., Simon, N., Taylor, J., and Tibshirani, R. J.
(2012). Strong rules for discarding predictors in lasso-type problems. Journal of the Royal
Statistical Society: Series B (Statistical Methodology), 74(2):245–266.
Tierney, L. and Kadane, J. B. (1986). Accurate approximations for posterior moments and
marginal densities. Journal of the American Statistical Association, 81(393):82–86.
Witten, D. M., Tibshirani, R., and Hastie, T. (2009). A penalized matrix decomposition, with
applications to sparse principal components and canonical correlation analysis. Biostatistics,
10(3):515–534.
Zeng, Y. and Breheny, P. (2017). The biglasso package: A memory-and computation-efficient
solver for lasso model fitting with big data in R. ArXiv preprint arXiv:1701.05936.
Zhang, C.-H. et al. (2010). Nearly unbiased variable selection under minimax concave penalty.
The Annals of Statistics, 38(2):894–942.
Zou, H. and Hastie, T. (2005). Regularization and variable selection via the elastic net. Journal
of the Royal Statistical Society: Series B (Statistical Methodology), 67(2):301–320.
42
arXiv:1712.02445v1 [math.ST] 6 Dec 2017
Supplementary Material
1
Additional Simulation Results
1.1
Results for Larger pn
Here we present the performance of the different methods with respect to mean square prediction error (MSPE) and empirical coverage probability (ECP) of 50% prediction intervals
(PI), along with the width of PIs for a larger choice of pn in each scheme.
Mean Square Prediction Error (MSPE). In Scheme I (see Figure 1) for pn = 3000,
TARP based methods yield the best results, with RIS-PCR having marginally better performance than RIS-RP. These methods are followed by SSR-Ridge and PCR. ISIS based
methods and SSR-EN have overall comparable performance with SSR-EN showing more
stable results, and ISIS-MCP having minimum average MSPE among these 3 methods.
Performance of these six methods are followed by BCR, and that in turn is followed by
RPCR. SSR-LASSO and SPCR have very poor performance in this scenario.
As in the main paper simulation case, TARP based methods have the best overall
performance in Scheme II for pn = 2 ∗ 104 , immediately followed by PCR (see Figure 2).
These three methods are followed by SSR-EN and SSR-Ridge. Unlike for pn = 104 , here
1
Figure 1: Box-plot of MSPE of eleven competing methods for pn = 3000 in Scheme I.
SSR-EN outperforms SSR-Ridge. RPCR is substantially worse but is still better than SSRLASSO, BCR and SPCR (in that order). ISIS based methods had very poor performance.
In Scheme III (see Figure 3), the performance of all the methods are almost similar,
except SSR based methods. Among the SSR based methods, SSR-EN and SSR-LASSO
have very similar results. SSR-EN has lower average MSPE, however SSR-LASSO has a
more stable performance. Unlike for pn = 5000, ISIS based methods often yield large values
of MSPE for pn = 104 .
The relative performance of SPCR and ISIS based methods are unsatisfactory in Scheme
IV (see Figure 4(a)). Among the other methods (see Figure 4(b)), TARP based methods
and PCR have the best performance, followed by BCR and SSR-EN. The relative performance of RPCR and SSR-LASSO improve significantly for pn = 2 ∗ 104 . These two
methods outperform SSR-Ridge, and among them, RPCR yields much better results than
SSR-LASSO.
Empirical Coverage Probability (ECP) and Prediction Interval (PI) Widths.
The relative performance of all the methods is quite similar to that observed in the main
2
Figure 2: Box-plot of MSPE of eleven competing methods for pn = 20000 in Scheme II.
Figure 3: Box-plot of MSPE of eleven competing methods for pn = 10000 in Scheme III.
paper (see Table 1 of the paper and Table 4) for all the simulation schemes. The detailed
description is given below.
The average ECP of SPCR is closest to 0.5 in Scheme I, followed by RPCR, RIS-RP
and SSR-EN, with RIS-RP showing the smallest average width among these methods. The
sd of ECPs and widths of the PIs is highest for PCR, indicating a lack of stability in it’s
performance. RIS-PCR yields some under coverage, however the average width of PI for
RIS-PCR is lowest among all the competing methods. SSR-Ridge has an average ECP of
3
(a) MSPE of all the methods.
(b) MSPE of selected methods.
Figure 4: Box-plot of MSPE of the competing methods for pn = 20000 in Scheme IV.
4
Table 4: Mean and standard deviation(sd) of empirical coverage probabilities of 50% prediction intervals, and mean and sd of width of 50% prediction intervals.
Methods→
Scheme, pn
ISISISISSSR- SSRSACD SACD LASSO EN
SSRRidge
PCR
SPCR RPCR BCR
RISRP
RISPCR
Average and standard deviation (in braces) of empirical coverage probability
I
3∗
103
0.307
(.052)
0.306
(.050)
0.434
(.059)
0.536
(.043)
0.631
(.048)
0.437
(.288)
0.498
(.059)
0.471
(.055)
0.283
(.053)
0.464
(.052)
0.382
(.054)
II
2 ∗ 104
0.334
(.051)
0.333
(.048)
0.422
(.055)
0.577
(.045)
0.637
(.049)
0.365
(.260)
0.505
(.059)
0.454
(.052)
0.280
(.058)
0.488
(.056)
0.312
(.055)
III
104
0.495
(.053)
0.494
(.053)
0.502
(.058)
0.653
(.033)
0.671
(.144)
0.503
(.058)
0.498
(.052)
0.498
(.052)
0.550
(.054)
0.487
(.055)
0.499
(.054)
IV
2 ∗ 104
0.432
(.063)
0.433 0.485
(0.060) (.099)
0.658
(.034)
0.664
(.108)
0.489
(.308)
0.492
(.052)
0.483
(.058)
0.494
(.051)
0.675
(.060)
0.408
(.052)
Average and standard deviation (in braces) of width of the 50% prediction interval
I
3 ∗ 103
4.096
(.522)
3.942
(.531)
6.148
(.410)
7.388
(.543)
8.098
(.484)
8.082 8.441
(7.354) (.608)
7.023
(.492)
5.720
(.411)
5.192
(.152)
4.090
(.335)
6.589 7.649
(7.380) (.929)
6.270
(.565)
5.687
(.465)
4.594
(.204)
2.635
(.209)
1.373
(.070)
1.551
(.073)
1.343
(.079)
1.383
(.077)
2.895 32.430 1.925
(3.808) (7.647) (.234)
1.794
(.144)
2.066
(.142)
1.070
(.071)
II
2 ∗ 104
5.964
(.554)
5.954
(.554)
5.276
(.525)
7.039
(.479)
8.604
(.674)
III
104
1.483
(.174)
1.452
(.144)
1.36
(.072)
1.942
(.122)
5.439 1.362
(2.018) (.072)
IV
2 ∗ 104
18.00 17.98 3.428 1.621
(2.876) (2.789) (2.522) (.167)
4.087
(.634)
1.373
(.070)
about 0.63 and a high width of the PI. ISIS based methods and BCR have lower-coverage
probabilities on average.
For Scheme II, SSR-RP and SPCR have the best overall results, with SPCR having
higher PI widths than SSR-RP. The ECPs of SSR-LASSO, SSR-EN and RPCR are closer
to 0.5 than the remaining methods with SSR-LASSO having a lower average width of the
PIs. The average ECP of PCR is close to 0.5, however there are huge fluctuations in the
widths of PIs (sd= 7.4). Among the other methods, SSR-Ridge shows higher coverage with
a comparatively large width of PI. The remaining four methods show under coverage for
both the choices of pn . Among these four methods, RIS-PCR attains the lowest average
width of PI.
5
All the methods, except SSR-EN and SSR-Ridge have comparable performance in terms
of both the criteria in Scheme III. SSR-EN and SSR-Ridge tend to have higher ECP and
larger width of PI in this setup.
SSR-LASSO, RPCR and BCR yield the best performance in Scheme IV, and BCR
achieves the lowest average width of PI among them. Although the average ECP of SPCR
is close to 0.5, it possesses much higher width than all other methods. SSR-Ridge, SSR-EN
and RIS-RP show higher ECP in general. In terms of PI widths of these three methods,
performance of RIS-RP is the best, followed by SSR-EN. In Scheme IV as well, PCR has
a higher sd of the width of PIs indicating instability in its performance. ISIS-SCAD, ISISMCP and RIS-PCR show some under coverage, and averages of ECPs for all of them are
around 0.4. RIS-RCR yields the lowest overall width of PI among all the methods. Average
width of PI for ISIS based methods and SPCR are much higher than all the other methods
in Scheme IV.
1.2
Effect of sample size on different competing methods
We consider three different choices of n, viz., 150, 200, 250. The results for n = 200 are
presented in the paper and previous sections of Supplementary file. The results corresponding two other choices of n are presented in this section. The relative performance of
the competing methods for different values of n remains similar for the two choices of pn .
Therefore, we present results corresponding to higher values of pn only.
Comparison with respect to MSPE. In Scheme I (see Figures 5(a)-5(b)), SSR-Ridge,
SSR-EN, PCR and TARP based methods work well for lower sample size. Among these
methods, RIS-RP has the best overall performance, immediately followed by RIS-PCR.
6
(a) n = 150
(b) n = 250
Figure 5: Box-plot of MSPE of the competing methods for pn = 3 × 103 in Scheme I.
Among the other methods, BCR and SSR-LASSO yield better performance than others.
The overall scenario changes reasonably for higher sample size, when n = 250. ISIS-MCP
based methods outperform other methods with respect to average MSPE, closely followed
by RIS-RP, ISIS-SCAD and RIS-PCR which also have satisfactory performance. However,
the ISIS based methods have much larger dispersion than TARP based methods, indicating
lack of stability in their results. The other methods do not show much improvement with
increment of sample size. Among these methods, SSR-Ridge yields relatively better results
followed by SSR-EN and PCR.
The relative performance of all these methods does not vary much for different sample
sizes in Scheme II (see Figures 6(a)-6(b)). For both the choices of n, TARP outperforms
all other methods. SSR based methods show reasonable results, although the differences of
their performance with TARP increases with increment of sample size. The other methods
do not perform well for either choices of n. Notably, PCR yields a low median MSPE in
general, however, it results in very high values of MSPE occasionally. For n = 150, the
7
(a) n = 150
(b) n = 250
Figure 6: Box-plot of MSPE of the competing methods for pn = 2 × 104 in Scheme II.
MSPE is as large as 1038, while for n = 250 the MSPE increases up to 3089.
Projection based methods work extremely well in Scheme III for both choices of n (see
Figures 7(a)-7(b)). SSR based methods fail to perform well in this scheme, with SSR-Ridge
showing largest overall MSPE. ISIS based methods show low MSPE in general, however
they occasionally yield higher values of MSPEs.
As the ISIS based method and SPCR yield extremely large values of MSPE in Scheme
IV, we avoid presenting further results on those methods (see Figures 8(a)-8(b)). TARP
yields the best performance for both choices of n in this scheme. SSR-EN also gives comparable performance. Among the other methods, SSR-LASSO and BCR yield reasonable
results. The other three methods tend to have higher values of MSPE in this scheme. As
in Scheme II, PCR tends to have extremely large MSPEs. For n = 250, the highest MSPE
obtained by PCR is around 125, whereas the average MSPE of all other methods is below
7.
8
(a) n = 150
(b) n = 250
Figure 7: Box-plot of MSPE of the competing methods for pn = 104 in Scheme III.
Table 5: Mean and standard deviation (in small font) of empirical coverage probabilities
of 50% prediction intervals.
n→
Schemes →
n = 150
n = 250
I
II
III
VI
I
II
III
VI
Methods↓ pn →
3 × 103
2 × 104
104
2 × 104
3 × 103
2 × 104
104
2 × 104
ISIS-SCAD
0.290.060
0.340.068
0.488.064
0.429.057
0.344.060
0.349.057
0.507.061
0.447.050
ISIS-MCP
0.294.052
0.336.070
0.489.063
0.427.058
0.345.065
0.354.059
0.503.061
0.443.053
SSR-LASSO
0.406.059
0.386.053
0.494.064
0.660.031
0.441.052
0.436.050
0.502.061
0.656.049
ISIS-EN
0.658.035
0.660.035
0.655.031
0.657.030
0.656.031
0.661.035
0.657.033
0.656.030
ISIS-Ridge
0.660.031
0.652.034
0.652.032
0.656.030
0.661.031
0.660.034
0.659.032
0.657.033
PCR
0.407.200
0.450.285
0.494.064
0.484.287
0.421.193
0.416.275
0.502.061
0.504.316
SPCR
0.507.064
0.499.055
0.494.064
0.490.056
0.491.054
0.497.050
0.502.061
0.501.054
RPCR
0.487.052
0.475.056
0.494.065
0.490.053
0.489.057
0.484.051
0.502.061
0.494.058
BCR
0.519.056
0.353.062
0.626.061
0.525.058
0.555.057
0.353.054
0.557.059
0.510.052
RIS-RP
0.508.063
0.460.059
0.492.065
0.723.050
0.598.058
0.504.057
0.499.059
0.704.047
RIS-PCR
0.434.062
0.270.049
0.508.066
0.395.058
0.549.054
0.338.051
0.510.060
0.420.052
Empirical Coverage Probability (ECP) and Prediction Interval (PI) Widths.
Tables 5 and 6 provide the means and standard deviations (sd) of empirical coverage
probabilities (ECP) of 50% prediction intervals (PI) and the widths of the PIs, respectively,
9
(a) n = 150
(b) n = 250
Figure 8: Box-plot of MSPE of the competing methods for pn = 2 × 104 in Scheme IV.
for two different choices of n.
As n increases, ISIS based methods tend to improve rapidly with respect to both the
measures. The other methods show little difference with an increment of n. ISIS based
methods show under coverage for all the simulation schemes except Scheme III. However,
the PI widths are low in general for these methods, except for Scheme IV, where they
result in the highest average width. SSR-EN and SSR-Ridge tend to have higher coverage
probabilities, and higher average PI width. For most of the schemes, these two methods
possess the highest average width. PCR tends to show under coverage for Schemes I and II,
and yield reasonable ECPs for the other two schemes. However, the widths of PIs for PCR
attain the highest variance in all the schemes, except Scheme III, indicating lack of stability
in the results. SPCR and RPCR yield the best performance in terms of ECP. The width of
PI is higher on an average for SPCR than RPCR. For Scheme IV, SPCR yields extremely
large widths of PI. Except for Scheme II, BCR and SSR-LASSO perform reasonably well
in terms of ECP. However, these two methods result in some over-coverage in Scheme III
10
and Scheme IV. With respect to PI width, SSR-LASSO outperforms BCR. TARP based
methods yield the lowest PI width in general. RIS-RP yields an average ECP close to 0.5
in most of the cases, and tend to show high coverage probabilities occasionally. However,
RIS-PCR tends to show low coverage probabilities in general, especially for n = 150.
2
Predictive Calibration in Binary Response Datasets
Apart from measuring the misclassification rates and the area under ROC curve, we validate TARP in terms of it’s ability to quantify uncertainly in real datasets with binary
responses. To this end, we partition the interval [0, 1] into ten equal sub-intervals, viz.,
[0, 0.1), [0.1, 0.2) and so on, and classify the test data points (x, y)i,new to the k th class if
predictive probability of yi,new falls in that class. Next, we consider the squared difference
of the empirical proportion of yi,new = 1 among the data points classified in a given interval
with the middle point of the interval, and consider the mean of these squared differences
(MSD) of all the intervals. If a method is well calibrated, then the MSD would be small.
The following table shows means and standard deviations of MSDs of eleven competing
methods for Golub and GTEx datasets.
Table 7 indicates that the competing methods do not show much differences in MSD
of empirical and predictive probabilities. ISIS based methods, PCR, RPCR and TARP
perform relatively well compared to the other methods. Among these methods, RIS-PCR
and RIS-RP have lowest MSD for Golub and GTEx, respectively. Relative performance of
SSR based methods are deficient in terms of MSD as well. SPCR results in high MSD for
both the datasets too.
11
Table 6: Mean and standard deviation (in small font) of width of 50% prediction intervals.
n→
n = 150
Schemes →
n = 250
I
II
III
VI
I
II
III
VI
Methods↓ pn →
3 × 103
2 × 104
104
2 × 104
3 × 103
2 × 104
104
2 × 104
ISIS-SCAD
4.390.50
6.130.85
7.8744.45
13.762.60
2.990.87
5.870.54
1.490.25
9.674.11
ISIS-MCP
4.410.54
6.120.86
1.460.30
14.202.24
2.780.95
5.900.55
1.400.14
9.714.01
SSR-LASSO
5.800.41
4.950.38
1.360.08
3.920.40
6.450.42
5.490.50
1.360.06
3.760.47
ISIS-EN
9.660.77
8.870.79
3.350.69
2.650.34
9.350.78
8.110.74
3.270.53
2.670.38
ISIS-Ridge
9.310.75
9.641.19
5.392.05
6.320.99
8.120.64
8.830.95
5.731.97
4.490.63
PCR
5.642.89
7.978.33
1.360.08
2.602.82
5.912.97
6.776.41
1.360.06
2.823.53
SPCR
8.440.61
8.220.80
1.360.08
31.089.79
8.310.53
7.930.70
1.360.06
30.2110.41
RPCR
7.760.50
7.730.64
1.360.08
2.820.49
7.720.45
7.540.54
1.360.06
2.750.46
BCR
7.440.48
6.280.52
1.810.07
2.740.23
8.180.42
5.970.38
1.550.05
2.050.14
RIS-RP
6.120.26
4.910.22
1.350.08
2.460.16
6.260.21
4.350.17
1.360.06
1.970.13
RIS-PCR
5.810.33
2.530.20
1.40.08
1.090.08
6.100.37
2.630.20
1.390.06
0.980.06
Table 7: Mean and standard deviation (in bracket) of mean square differences (MSD) of
empirical and predictive probabilities.
3
Methods→
Dataset ↓
ISISSACD
ISISSACD
SSRSSRLASSO EN
SSRRidge
PCR
SPCR
RPCR BCR
RISRP
RISPCR
Golub
2.605
(.036)
2.606
(.035)
2.804
(.000)
2.804
(.000)
2.742
(.028)
2.589
(.012)
3.429
(.073)
2.587
(.017)
2.886
(.032)
2.611
(.045)
2.555
(.044)
GTEx
2.784
(.000)
2.784
(.000)
3.325
(.000)
3.325
(.000)
3.325
(.000)
2.784
(.000)
3.216
(.102)
2.784
(.000)
2.873
(.007)
2.782
(.001)
2.783
(.001)
Mathematical Details
Proof of Lemma 2 a. Consider the conditional expectation and variance of kRγ xk2 given
(γ, x) as follows:
E kRγ xk2 |γ
var kRγ xk2 |γ
= kxγ k2
= 4kxγ k2
(
12
1+
Ppγ 4 )
1
j=1 xγ,j
−1
,
ψ
2kxγ k2
where xγ includes the regressors j for which γj = 1. The detailed proof is given in Result
1 below.
Next consider the conditional expectation of kRγ xk2 given x is given by
Eγ E kRγ xk2 |γ = Eγ
X
!
x2j I(γj = 1)
j
=c
X
x2j |rxj ,yn |δ ,
j
(1)
where c > 0 is the proportionality constant. Also the conditional variance of kRγ xk2 given
x is given by
varγ E kRγ xk2 |γ + Eγ var kRγ xk2 |γ .
(2)
Considering both the terms in (2) separately as follows:
2
varγ E kRγ xk |γ
= varγ
= c
X
j
X
x2j I(γj
j
!
= 1)
x4j |rxj ,yn |δ 1 − c|rxj ,yn |δ ≤ pn ,
(3)
as given x, γj s are independent, and each |xj | ≤ 1, and qj = c|rj |δ < 1. Again
Eγ var kRγ xk2 |γ
"
= Eγ 4kxγ k2
(
1+
≤ cEγ kxγ k2
X
≤ c
x2j |rxj ,yn |δ ,
Ppγ 4 )#
1
j=1 xγ,j
−1
ψ
2kxγ k2
(4)
j
for some constant c, as
Ppγ
j=1
x4γ,j < kxγ k2 .
Therefore, from (1), (3) and (4) it can be shown that the expectation of kRγ xk2 /pn
13
converges to the limit cαδ , and variance of the same converges to 0.
Proof of Lemma 2 b. The proof follows from observing that
X
Eγ kxγ k2 = c
x2j |rxj ,yn |δ
j
and
X 4
varγ E kxγ k2 |γ c
xj |rxj ,yn |δ 1 − c|rxj ,yn |δ ≤ pn .
j
Therefore it can be shown that the expectation of kxγ k2 /pn converges to the limit cαδ , and
variance of the same converges to 0.
Result 1. Consider a random matrix Rγ which depends on another random vector γ
distributed as in (2). Then the conditional distribution of Rγ satisfies the following:
a. E (kRγ xk2 |γ) = kxγ k2 , and
Ppγ 4
b. var (kRγ xk2 |γ) = 4kxγ k2 1 + (ψ −1 − 1) j=1
xγ,j / (2kxγ k2 ) .
Proof of part a. Observe that
2
kRγ xk
=
X
r1,j γj xj ,
j
=
X
r1,j γj xj
j
X
r2,j γj xj , . . . ,
+
r2,j γj xj
!2
j
X
j
rmn ,j γj xj
!0
+ ... +
X
X
j
!2
2
rmn ,j γj xj
j
!2
.
(5)
Now
E
X
j
r1,j γj xj
!2
=E
(
X
j
2
r1,j
γj x2j +
X
r1,j r1,j 0 γj γj 0 xj xj 0
j6=j 0
14
)
=2
X
j
γj x2j = 2kxγ k2 ,
2
as E(ri,j
) = 1 and E(ri,j ri,j 0 ) = 0 as i = 1, 2, . . . , mn , j, j 0 = 1, 2, . . . , pn , and j 6= j 0 .
Proof of part b. From (5) we have
2
var kRγ xk |γ
= var
X X
+
i
ri,j γj xj
j
X
cov
i6=i0
X
!2
=
X
var
,
X
i
ri,j γj xj
j
!2
X
ri,j γj xj
j
ri0 ,j γj xj
j
!2
!2
.
We will consider each term of (6) one by one. Consider the first term. Note that
var
X
ri,j γj xj
j
!2
= var
(
X
2
ri,j
γj x2j +
j
= var
(
X
X
2
ri,j
γj x2j
j
+cov
(
X
)
ri,j ri,j 0 γj γk xj xj 0
j6=k
+ var
(
X
)
ri,j ri,j 0 γj γk xj xj 0
j6=j 0
2
ri,j
γj x2j ,
X
ri,j ri,j 0 γj γj 0 xj xj 0
j6=j 0
j
)
)
.
Consider the first term in (6).
var
(
X
j
2
ri,j
γj x2j
)
=
X
j
=
X
X
2
2
2
2
cov ri,j
γj x2j , ri,j
var ri,j
γj x2j +
0 γj 0 xj 0
2
γj x4j var ri,j
+
j
=
X
j
j6=j 0
γj x4j
E
4
ri,j
15
X
2
2
γj x2j γj 0 x2j 0 cov ri,j
, ri,j
0
j6=j 0
−E
2
2
ri,j
=2
X
1
−1
γj x4j .
ψ
j
(6)
Again,
var
=
X
j6=j 0
(
X
ri,j ri,j 0 γj γk xj xj 0
j6=j 0
)
2 2
γj γk x2j x2j 0 E ri,j
ri,j 0 +
=4
X
=E
X
X
ri,j ri,j 0 γj γk xj xj 0
j6=j 0
!2
,
γj γk γj 0 γk0 x2j x2j 0 x2k x2k0 E (ri,j ri,j 0 ri,k ri,k0 )
(j, j 0 ) 6= (k, k0 )
j 6= j 0 , k 6= k0
γj γk x2j x2j 0 ,
j6=j 0
as the other term will be zero. Next
o
P
2
2
0 γj γj 0 xj xj 0
,
r
r
γ
x
r
0
i,j
i,j
j
j
j6=j
j i,j
P P
2
= j k6=k0 γj x2j , γk γk0 xk xk0 cov ri,j
, ri,k ri,k0 = 0.
cov
nP
Therefore the first term in (6) is
X
var
i
X
ri,j γj xj
j
!2
#
"
X
X
1
κ
4
2 2
=
γj γk xj xj 0
(n − 1)
γj xj +
mn
j
j6=j 0
!2
X
X
1
= 2
−1
γj x4j + 2
γj x2j .
ψ
j
j
(7)
Now consider the last term in (6).
cov
X
= cov
j
(
X
j
ri,j γj xj
!2
2
ri,j
γj x2j +
,
X
j
X
ri0 ,j γj xj
!2
ri,j ri,j 0 γj γj 0 x2j x2j 0 ,
j6=j 0
X
k
16
ri20 ,k γk x2k +
X
k6=k0
ri0 ,k ri0 ,k0 γk γk0 x2k x2k0
)
= cov
(
X
2
γj x2j ,
ri,j
j
+cov
(
X
X
ri20 ,k γk x2k
k
ri,j ri,j 0 γj γj 0 x2j x2j 0 ,
j6=j 0
+cov
(
X
X
)
+ cov
ri20 ,k γk x2k
k
ri,j ri,j 0 γj γj 0 x2j x2j 0 ,
j6=j 0
X
(
X
)
2
γj x2j ,
ri,j
k6=k0
ri0 ,k ri0 ,k0 γk γk0 x2k x2k0
k6=k0
j
ri0 ,k ri0 ,k0 γk γk0 x2k x2k0
X
)
.
)
(8)
Consider the first term of (8).
cov
(
X
2
ri,j
γj x2j ,
j
X
k
ri20 ,k γk x2k
)
=
XX
j
k
2
γj x2j γk x2k cov ri,j
, ri20 ,k = 0.
Similarly, it can be shown that all other terms in (8) are zero. Combining the above result
and (7) the proof follows.
17
| 10math.ST |
On the Approximation of Constrained Linear Quadratic
Regulator Problems and their Application to Model
Predictive Control - Supplementary Notes
Michael Muehlebach and Raffaello D’Andrea∗
arXiv:1803.05510v1 [cs.SY] 23 Feb 2018
Abstract
By parametrizing input and state trajectories with basis functions different approximations to the
constrained linear quadratic regulator problem are obtained. These notes present and discuss technical
results that are intended to supplement a corresponding journal article. The results can be applied in a
model predictive control context.
1
Outline
Sec. 2 summarizes the notation. The resulting approximations are given by convex finite-dimensional optimization problems that have a quadratic cost, linear equality constraints, and linear semi-infinite inequality
constraints. Sec. 3 suggests several strategies for dealing with these semi-infinite constraints. One of these
strategies is discussed in more details in Sec. 4, as it is found to be computationally efficient in a model
predictive control context. The remaining sections discuss several technical results. Sec. 5 and Sec. 6 are
related to an upper, respectively lower bound approximation of the cost of the underlying constrained linear
quadratic regulator problem. Sec. 7 presents a result that is related to the convergence of the different approximations. Sec. 8 shows recursive feasibility and closed-loop stability of the resulting model predictive
control algorithm. Sec. 9 and Sec. 10 highlight important properties of the basis functions parametrizing
input and state trajectories.
2
Notation
We are concerned with the approximation of the following optimal control problem
1
||x||2 + ||u||2 + ψ(xT )
2
s.t. ẋ = Ax + Bu, x(0) = x0 , x(T ) = xT ,
Cx x + Cu u ≤ b, xT ∈ X ,
x ∈ L2n , u ∈ L2m ,
J∞ := min
(1)
where the space of square integrable functions mapping from the interval I := (0, T ) to Rq is denoted by
L2q , where q is a positive integer; and the L2q -norm is defined as the map L2q → R,
Z
2
x → ||x||, ||x|| := xT x dt,
(2)
I
∗
Michael Muehlebach and Raffaello D’Andrea are with the Institute for Dynamic Systems and Control, ETH Zurich. The
contact author is Michael Muehlebach, michaemu@ethz.ch. This work was supported by ETH-Grant ETH-48 15-1.
1
with dt the Lebesgue measure. The function ψ : Rn → R is assumed to be positive definite and strongly
convex, A ∈ Rn×n , B ∈ Rn×m , Cx ∈ Rnc ×n , Cu ∈ Rnc ×m , and b ∈ Rnc are constant, and the set X is closed
and convex. The dynamics as well as the stage constraints are assumed to be fulfilled almost everywhere.
Thus we simply write
f = g, f ≤ g,
(3)
when we mean f (t) = g(t), respectively f (t) ≤ g(t) for all t ∈ I almost everywhere, with f, g ∈ L2q .
The weak derivative of x is denoted by ẋ.1 To simplify notation we abbreviate the domain of the objective
function by
X := L2n × L2m × Rn .
(4)
We assume throughout the article that the constraints in (1) are nonempty, i.e. there exists trajectories x
and u, fulfilling the dynamics, the initial condition, the constraints, and thus achieve a finite cost.
The main motivation for studying problem (1) comes from the fact that (1) often serves as a starting
point for model predictive control (MPC).
The problem (1) can be written in compact form as
J∞ = min ||x||2 + ||u||2 + ψ(xT )
s.t. (x, u, xT ) ∈ C ∩ D,
(5)
where the set C contains all trajectories x ∈ L2n , u ∈ L2m that fulfill the constraints and D the trajectories
that fulfill the dynamics and the initial condition, that is,
C := {(x, u, xT ) ∈ X | Cx x + Cu u ≤ b, xT ∈ X },
D := {(x, u, xT ) ∈ X|ẋ = Ax + Bu, x(0) = x0 , x(T ) = xT }.
(6)
(7)
It was shown that (5) can be approximated by two auxiliary optimization problems, where input and state
trajectories are parameterized with basis functions,
x̃(t) = (In ⊗ τ s (t))T ηx ,
ũ(t) = (Im ⊗ τ s (t))T ηu .
The parameter vectors are denoted by ηx ∈ Rns and ηu ∈ Rms , whereas τ s (t) := (τ1 (t), τ2 (t), . . . , τs (t)) ∈
Rs contains the first s basis functions, ⊗ denotes the Kronecker product, and Iq ∈ Rq×q refers to the
identity matrix for any integer q > 0. The superscript s refers to the number of basis functions used for
the approximation. For ease of notation the superscript s will be dropped whenever the number of basis
functions is clear from context. The basis functions are required to satisfy the following assumptions:
A1) The basis functions τi ∈ L21 , i = 1, 2, . . . , s are linearly independent and orthonormal with respect to
the standard L21 -scalar product.
A2) The basis functions fulfill τ̇ s (t) = Ms τ s (t) for all t ∈ I, for some Ms ∈ Rs×s .
The finite-dimensional subspace spanned by the first s basis functions will be denoted by X s ,
X s := {(x, u, xT ) ∈ X | ηx ∈ Rns , ηu ∈ Rms , x = (In ⊗ τ s )T ηx , u = (Im ⊗ τ s )T ηu }.
(8)
We can think of an element in X s not only as an element in X (i.e. a tuple of a finite-dimensional vector
and two square integrable functions), but also as a finite-dimensional vector given by the corresponding
parameter vectors ηx and ηu . To make this distinction explicit, we introduce the map π qs : L2q → Rqs ,
defined as
Z
x → (Iq ⊗ τ s )xdt,
(9)
I
The equations of motion imply that ẋ ∈ L2n , which can be used to conclude that x has a unique absolutely continuous
representative defined on the closure of I (a classical solution of the equations of motion). With x(t) we refer to the value this
unique absolutely continuous representative takes at time t ∈ [0, T ].
1
2
which maps an arbitrary element x ∈ L2n to its first s basis function coefficients. Similarly, we define
π s : X → Rns × Rms × Rn as
(x, u, xT ) → (π ns (x), π ms (u), xT ).
(10)
As a consequence, we write π s (x) for describing the finite dimensional representation of x ∈ X s , that is,
its representation in terms of the parameter vectors ηx and ηu . The adjoint map (π qs )∗ : Rqs → L2q is given
by
η → (Iq ⊗ τ )T η,
(11)
and is used to obtain the trajectory corresponding to the vector η ∈ Rqs , containing the first s basis function
coefficients. Similarly, we define (π s )∗ : Rns × Rms × Rn → X as
(ηx , ηu , xT ) → ((π ns )∗ (ηx ), (π ms )∗ (ηu ), xT ).
(12)
The composition (π s )∗ π s : X → X yields the projection of an element x ∈ X onto the subspace X s ⊂ X.
The sets C and D are approximated in two different ways,
CUs := C ∩ X s ,
(13)
CLs := {(x, u, xT ) ∈ X |
Z
δ p̃T (−Cx x − Cu u + b)dt ≥ 0
I
∀δ p̃ = (Inc ⊗ τ )T δηp : δ p̃ ≥ 0, δηp ∈ Rnc s ; xT ∈ X },
DUs
DLs
s
:= D ∩ X ,
:= {(x, u, xT ) ∈ X | δ p̃ = (In ⊗ τ )T δηp ,
Z
Z
T
˙
− δ p̃ xdt − δ p̃T (Ax + Bu)dt − δ p̃(0)T x0 + δ p̃(T )T xT = 0, ∀δηp ∈ Rns },
I
(14)
(15)
(16)
I
resulting in the following two approximations to (5)
Js := inf ||x̃||2 + ||ũ||2 + ψ(xT )
s.t. (x̃, ũ, xT ) ∈ CUs ∩ DUs ,
(17)
J˜s := min ||x̃||2 + ||ũ||2 + ψ(xT )
s.t. (x̃, ũ, xT ) ∈ CLs ∩ DLs .
(18)
and
By definition of the constraints CUs and DUs , the minimizer of (17) (for s ≥ s0 ) is required to be an element
of X s . Consequently, the problem (17) is equivalent to
Js = inf |ηx |2 + |ηu |2 + ψ(xT )
s.t. (ηx , ηu , xT ) ∈ π s (CUs ) ∩ π s (DUs ),
(19)
which corresponds to a convex finite-dimensional optimization problem. Likewise, it is found that the
convex finite-dimensional problem
J˜s = min |ηx |2 + |ηu |2 + ψ(xT )
s.t. (ηx , ηu , xT ) ∈ π s (CLs ) ∩ π s (DLs ),
(20)
is equivalent to (18) in the sense that its (unique) minimizer (ηx , ηu , x̄T ) is related to the minimizer (x, u, xT ) ∈
X of (18) by x = (π ns )∗ (ηx ), u = (π ms )∗ (ηu ), x̄T = xT , and achieves the same cost.
The proposed approximations can be applied in the context of MPC. We will show that by repeatedly
solving the infinite-horizon optimal control problem (19) (with I = (0, ∞), ψ = 0, xT = limt→∞ x(t) = 0),
3
recursive feasibility and closed-loop stability are inherent to the resulting MPC algorithm. For the sake of
completeness, (19) is written out as
min ηxT (Q ⊗ Is )ηx + ηuT (R ⊗ Is )ηu
s.t. (In ⊗
MsT
(21)
− A ⊗ Is )ηx − (B ⊗ Is )ηu = 0,
(In ⊗ τ (0))T ηx = x0 ,
(Cx ⊗ τ (t))T ηx + (Cu ⊗ τ (t))T ηu ≤ b, ∀t ∈ [0, ∞),
where the input and state costs are weighted with the matrices Q > 0 and R > 0, which is common in
MPC.
In the following an efficient numerical solution algorithm for (21) will be discussed.
3
Implementation of the semi-infinite constraint
The optimization (21) is a convex finite-dimensional optimization problem. However, it is not a quadratic
program, as it includes the semi-infinite constraint
(Cx ⊗ τ (t))T ηx + (Cu ⊗ τ (t))T ηu ≤ b, ∀t ∈ [0, ∞).
(22)
As a result, (21) cannot be solved by a standard quadratic programming solver. Three different approaches
to deal with the semi-infinite constraint are immediate:
1) global polyhedral approximation
2) sum-of-squares approximation
3) local polyhedral approximation (active-set approach)
The first is based on a fixed polyhedral approximation, leading to a quadratic program. The second is based
on exploiting polynomial basis functions for obtaining a characterization using linear matrix inequalities,
whereas the third is based on an iterative constraint sampling scheme, resulting in a local polyhedral approximation. In the following subsections we will focus on approach 1 and 2. We will describe an efficient
solution algorithm based on approach 3 in detail in Sec. 4.
It turns out that it is enough to check the constraint (22) over a compact time interval, instead of the
unbounded interval t ∈ [0, ∞). This is because the basis functions are assumed to be linearly independent
and exponentially decaying according to Assumption A1 and A2. A formal proof of this claim can be found
in Sec. 9. The compact time interval for which the constraint (22) has to be checked is denoted by [0, Tc ]
and depends on the choice of basis functions, on the order s, and in some cases also on the bound b (see
Sec. 9).
3.1
Global polyhedral approximation
In order to construct a global polyhedral approximation of the set CUs we assume that an upper bound on the
achievable cost Js , denoted by J¯s , is available. In order to simplify the discussion, we assume further that
for now Cx and Cu are row vectors, and that b is a scalar. The proposed scheme can be readily extended
to the case where Cx and Cu are matrices, and b is a vector. The approximation is based on constraint
sampling, where we tighten the constraint slightly to
Cx x̃(ti ) + Cu ũ(ti ) ≤ (1 − )b,
(23)
with > 0, and where ti , denotes the constraint sampling instances, i = 1, 2, . . . , ni , which are to be
determined. The algorithm is based on the following two steps:
4
1) Compute
h(t) := max Cx x̃(t) + Cu ũ(t) − b
s.t. Cx x̃(ti ) + Cu ũ(ti ) ≤ (1 − )b, i = 1, . . . , ni ,
(x̃, ũ, lim x̃(te )) ∈ DUs ,
te →∞
||x̃|| + ||ũ||2 ≤ J¯s ,
2
for all times t ∈ Is , where Is contains a finite number of sampling instances (to be made precise
below).
2) Find the local peaks of h(t), denoted by t∗i . Add each t∗i to the constraint sampling points if h(t∗i ) >
−b/2. Repeat the procedure until h(t) ≤ −b/2 for all t ∈ Is .
Note that the function h(t) is again only evaluated at the discrete time points t ∈ Is . The index set Is has
to be chosen such that h(t) ≤ −b/2 for all t ∈ Is implies that h(t) ≤ 0 for all t ∈ [0, ∞). As remarked
earlier, due to the fact that the basis functions are exponentially decaying and linearly independent it is
enough to check h(t) ≤ 0 for all t ∈ [0, Tc ], for a fixed time Tc , as h(t) ≤ 0 for all t ∈ (Tc , ∞) will be
fulfilled automatically. Moreover, a Lipschitz constant of
Cx x̃(t) + Cu ũ(t) − b
(24)
can be found by using an upper bound on its time-derivative, that is, for example,
|τ (t)T M T (Cx ηx + Cu ηu )| ≤ |τ (t)||M T (Cx ηx + Cu ηu )|,
(25)
where the first term can be bounded for all t ∈ [0, ∞) due to the fact that the basis functions are exponentially decaying and the second term can be bounded using the fact that the cost Js is below J¯s . We therefore
choose the index set Is as
2L
Is = {tk < Tc | tk = k , k = 0, 1, 2, . . . },
(26)
b
where L denotes a Lipschitz constant of (24).
It is important to note that the optimization in step 1 imposes the dynamics and the upper bound J¯s on
the cost. Both constraints tend to reduce the number of constraint sampling points ti greatly. The initial
condition x0 enters the optimization as an optimization variable. The optimization in step 1 represents a
quadratically constrained linear program for each time instant t ∈ Is , and as such, it can be solved using
standard software packages. The whole procedure for determining the constraint sampling points is done
offline. Once these time instances are found, the optimization problem that is solved online reduces to a
quadratic program. The number of constraint sampling points ti is upper bounded by the cardinality of the
index set Is , and thus guaranteed to be finite. Due to the fact that the above procedure is greedy, it will not
necessarily lead to the smallest number of constraint sampling points.
3.2
Sum-of-squares approximation
In case exponentially decaying polynomials are used as basis functions, sum-of-squares techniques can be
applied. In particular, it is shown in [3] that the set
{η ∈ Rs | η T (1, t, . . . , ts−1 ) ≥ 0, ∀t ∈ [0, ∞)}
(27)
can be expressed using matrix inequalities that are linear in the coefficients η. In the case of exponentially
decaying polynomials it is therefore enough to approximate the exponential decay by a polynomial upper
bound (for example by appropriately truncating a Taylor series expansion at 0), in order to approximate
the constraint (22) in a slightly conservative manner. As a result, by applying the results from [3], the
optimization problem (21) is approximated by a semidefinite program that can be solved using standard
optimization routines.
5
4
A dedicated active-set approach
In the following section we present an efficient optimization routine for solving (19) (and likewise (21)).
The method is an extension of traditional active set methods and generalizes to optimization problems with
a linear quadratic cost function, linear equality constraints, and linear semi-infinite inequality constraints,
i.e.
ˆ c ) := min z T Hz
J(I
s.t. Aeq z = beq ,
(28)
(29)
lb ≤ (Inc ⊗ τ (t))T Cz z ≤ lu ,
∀t ∈ Ic ,
(30)
where Ic is any subset of the non-negative real line. In case of (21), the interval Ic is taken to be [0, ∞).
Note that an optimization problem, whose objective function has a linear part, can be brought to the form
(28) by completing the squares. It is assumed that the optimization problem (28) is feasible, that lb < 0
and lu > 0, and that the Hessian H is positive definite, which guarantees existence and uniqueness of the
corresponding minimizer2 .
The method is based on the observation that if the set Ic consists merely of a collection of time instants
(constraint sampling instances) ti , (28) reduces to a quadratic program that can be solved efficiently. Moreover, due to the fact that the basis functions fulfill Assumptions A1 and A2, a trajectory parametrized with
the basis functions has a finite number of maxima and minima, as is shown in Sec. 10. Consequently, (28)
has only a finite number of active constraints. The collection of the time instants corresponding to these
active constraints will be denoted by Ic∗ . If this finite collection of constraint sampling instants is known
ˆ c∗ ) = J(I
ˆ c ). In
ahead of time, one could simply solve (28) with respect to Ic∗ instead of Ic , resulting in J(I
k+1
k
k
k
ˆ
ˆ
addition, for any subset Ic of Ic it holds that J(Ic ) ≥ J(Ic ), and likewise, if Ic is a subset of Ic we have
ˆ c ) can be
ˆ k ). Hence, a monotonically increasing sequence J(I k ), bounded above by J(I
ˆ k+1 ) ≥ J(I
that J(I
c
c
c
k+1
k
k
constructed using any sequence of sets Ic that fulfill Ic ⊂ Ic ⊂ · · · ⊂ Ic . In particular, such sets are obtained by starting with an arbitrary initial guess Ic0 containing a finite number of constraint sampling points
(or even the empty set), and by adding at least one constraint violation point at each iteration. Moreover,
at each iteration the inactive constraints contained in the set Ick can be removed, as this will not alter the
ˆ k ). In that way, the number of constraint sampling instances contained
optimizer nor the optimal value J(I
c
k
in Ic remains finite. This motivates Alg. 1, which solves (28) up to a given tolerance, by constructing an
approximation to the set of active constraints Ic∗ .
ˆ k ) = J(I
ˆ ∗ ) = J(I
ˆ c ). In order to achieve constraint
Proposition 4.1 Alg. 1 converges, that is, limk→∞ J(I
c
c
violations smaller than at most
ˆ c ) − J(I
ˆ 0 ))
4cτ (J(I
c
(31)
2
σ
steps are required, where cτ is defined as
cτ := sup |τ (t)|2 ,
(32)
t∈Ic
and σ denotes the smallest eigenvalue of the Hessian H.
ˆ k ) is monotonically increasing whenever
Proof From the above arguments it can be concluded that J(I
c
ˆ k ) < J(I
ˆ c ) and upper bounded by J(I
ˆ c ). Therefore the sequence J(I
ˆ k ) converges. It remains to show
J(I
c
c
ˆ ck ) converges to J(I
ˆ c ). The strong convexity of the objective function can be used to establish
that J(I
ˆ ck+1 ) − J(I
ˆ ck )),
|z k+1 − z k |2 ≤ 4σ −1 (J(I
2
(33)
This is due to the fact that the constraints describe a closed convex set and due to the strong convexity of the objective
function.
6
Algorithm 1 Iterative constraint sampling
Initialize: initial guess for the constraint sampling points: Ic0 = {t1 , t2 , . . . , tN }; maximum number of
iterations: MAXITER; constraint satisfaction tolerance: ;
1: k = 0
2: for k <MAXITER do
ˆ k)
3:
solve (28) for Ick → z k , J(I
c
4:
if infeasible then
5:
abort
6:
else if z k fulfills (30) for all t ∈ I (with tol. ) then
7:
algorithm converged
8:
return z k
9:
end if
10:
find at least one constraint violation instant → tc
11:
remove inactive time instants in Ick
12:
Ick+1 = Ick ∪ {tc }, k = k + 1
13: end for
ˆ k ) and J(I
ˆ k+1 ). As a result we can conclude
where z k+1 and z k are the minimizer corresponding to J(I
c
c
k
k
that z converges, and that limk→∞ z is well-defined and satisfies the constraint (30) (by a contradiction
ˆ c ), which,
ˆ k ) ≥ J(I
argument). It is therefore a feasible candidate for (28), implying that limk→∞ J(I
c
ˆ ck ) ≤ J(I
ˆ c ) for all integers k, leads to limk→∞ J(I
ˆ ck ) = J(I
ˆ c ).
combined with J(I
It remains to show that (31) is fulfilled. To that extent, let > 0 denote the smallest constraint violation
that occurs within the first N steps. For all k ≤ N − 1, there exists the constraint violation point tkj , which
will be added to Ick , and therefore
≤ |(e(tkj ) ⊗ τ (tkj ))T (z k − z k+1 )|,
(34)
where e(tkj ) is a canonical unit vector. Combining the Cauchy-Schwarz inequality and the above bound on
|z k+1 − z k | results in
ˆ ck+1 ) − J(I
ˆ k )).
2 ≤ 4cτ σ −1 (J(I
(35)
c
By summing over the first N steps we arrive at
ˆ cN ) − J(I
ˆ c0 )) ≤ 4σ −1 cτ (J(I
ˆ c∗ ) − J(I
ˆ c0 )),
N 2 ≤ 4σ −1 cτ (J(I
(36)
ˆ ck ) is strictly increasing and bounded above by J(I
ˆ c∗ ). Dividing by 2 on both sides
since the sequence J(I
concludes the proof.
4.1
Implementation details
Alg. 1 can be naturally embedded in an active-set method. An introduction to active-set methods for solving
quadratic programs can be found for example in [1, Ch. 10]. Starting from an initial guess of the active conˆ c0 ) is solved:
straint sampling instants, which is denoted by Ic0 , the quadratic program with optimal cost J(I
0
This is done by initially assuming that all constraints in the set Ic are active. The resulting optimization
problem reduces to an equality constrained quadratic program, whose solution can be calculated by solving
a linear system of equations. The Lagrange multipliers corresponding to (30), which are denoted by µ(t),
t ∈ Ic0 , are evaluated subsequently. If all the constraints are indeed active, then the optimizer to the quadratic
ˆ c0 ) has been found. If, however, not all the constraints are found to be active, that is,
program with cost J(I
if there are some Lagrange multipliers that are zero, the standard active-set procedure, see [1, Ch. 10] is
used to find the subset of active constraints Ia ⊂ Ic0 . Provided that the active set Ia ⊂ Ic0 and the optimizer
ˆ 0 ) has been found, the constraint (30) is then checked for all t ∈ Ic .
to the quadratic program with cost J(I
c
7
If no constraint violations occur, the solution to (28) has been found. If constraint violations occur, the time
instant tc for which a violation occurs, is added to the set of active constraints Ia resulting in Ic1 = Ia ∪ {tc }.
The above procedure is then repeated until convergence.
Each iteration requires solving equality constrained quadratic programs of the type
min z T Hz
(37)
s.t. Aeq z = beq ,
(c(t) ⊗ τ (t))T Cz z = la (t),
∀t ∈ Ia ,
(38)
where c(t) ∈ Rnc , la (t) ∈ R, t ∈ Ia , and Ia ⊂ Ick describe the active constraints corresponding to (30). Due
to the fact that very few constraints are expected to be active, we use a range space approach, [1, p. 238].
To that extent, the equality constraint is eliminated and the optimizer z ∗ corresponding to (37) is rewritten
as
X
z ∗ = b̂ + ĤCzT
(c(t) ⊗ τ (t))T µ(t),
(39)
t∈Ia
where the dual variable µ(t) ∈ R, defined for t ∈ Ick , satisfies
X
(c(t) ⊗ τ (t))µ(t) = la (tj ) − (c(tj ) ⊗ τ (tj ))T Cz b̂,
(c(tj ) ⊗ τ (tj ))T Cz ĤCzT
(40)
t∈Ia
for all tj ∈ Ia ⊂ Ick , and µ(t) = 0 for all t ∈ Ick \ Ia , with
Ĥ := H −1 ATeq (Aeq H −1 ATeq )−1 Aeq H −1 − H −1 ,
b̂ := H −1 ATeq (Aeq H −1 ATeq )−1 beq .
(41)
(42)
The dual variable µ(t) is therefore obtained by solving (40), and the optimizer z ∗ is then determined via
(39). At each iteration, a single constraint is either added or removed. Therefore, in order to efficiently find
a solution to (40), a LDLT -decomposition of the matrix
{(c(tj ) ⊗ τ (tj ))T Cz ĤCzT (c(ti ) ⊗ τ (ti ))}(tj ,ti )∈Ia ×Ia
(43)
is computed and adapted at each step by performing rank-1 updates. The matrix Ĥ and the vector b̂ are
precomputed. For additional details regarding the regularity of (43), and issues related to cycling and
stalling we refer to [1, Ch. 10] and [4, p. 467].
4.2
Constraint check
It remains to explain how to efficiently check whether the constraint (30) is fulfilled for a given solution
candidate z. We assume that the interval Ic has the form Ic = [0, Tc ]. As it has been explained in Sec. 3,
the constraint check over the interval [0, ∞) reduces to the check over a compact interval, provided that the
basis functions fulfill Assumptions A1 and A2.
A straightforward approach would be to exploit the specific structure of the basis functions. For example, if the basis functions consist of exponentially decaying polynomials having a degree of at most 4,
determining the stationary points of (30) amounts to solving a quartic equation, which can be done analytically. As a result, it would be enough to check the constraints at these stationary points in order to determine
if the constraint is satisfied or not.
We propose a more general approach that is based on local Taylor approximations, and thus valid for
arbitrary basis functions compatible with Assumptions A1 and A2. In order to simplify the discussion, we
consider the special case of (30), where Cz is the identity and nc = 1. The resulting algorithm extends
naturally to the more general case. According to Taylor’s theorem we obtain the following identity
τ (t)T z = τ (0)T z + τ̇ (0)T zt + τ̈ (0)T z
8
t2
t3
+ τ (3) (t̄)T z ,
2
6
(44)
where t̄ ∈ [0, t]. As will become clear in the following, a third-order Taylor expansion represents a good
compromise between approximation quality and computational effort. The last term of the previous equation can be bounded by the Cauchy-Schwarz inequality leading to
|τ (3) (s)T z| ≤ sup |τ (t̄)||(M T )3 z| =: R(z),
∀s ∈ [0, t].
(45)
t̄∈Ic
As a consequence, the following upper and lower bounds are obtained
bl (t) ≤ τ (t)T z ≤ bu (t),
(46)
for all t ∈ Ic , with
t2
t3
− R(z) ,
2
6
2
t3
t
bu (t) := τ (0)T z + τ̇ (0)T zt + τ̈ (0)T z + R(z) .
2
6
bl (t) := τ (0)T z + τ̇ (0)T zt + τ̈ (0)T z
(47)
(48)
The situation is exemplarily depicted in Fig. 1 (right). Given that τ̇ (0)T z ≥ 0 the lower bound attains its
maximum at time
p
τ̈ (0)T z + (τ̈ (0)T z)2 + 2R(z)τ̇ (0)T z
tu :=
> 0,
(49)
R(z)
whereas if τ̇ (0)T z < 0 the upper bound attains its minimum at time
p
−τ̈ (0)T z + (τ̈ (0)T z)2 − 2R(z)τ̇ (0)T z
tl :=
> 0.
R(z)
(50)
Thus, if the lower bound exceeds lu or the upper bound drops below ll , that is, if
bl (tu ) > lu ,
or
bu (tl ) < ll ,
(51)
the constraint (30) is guaranteed to be violated at time tu , respectively at time tl . If this is not the case, we
are guaranteed that the constraint (30) is satisfied in the interval [0, ts ], with
ts := min{t1 , t2 | bu (t1 ) = lu , bl (t2 ) = ll }.
(52)
Thus, finding the value ts requires the solution of two cubic equations, which stems from the fact that a
third order Taylor approximation was used as a starting point. By shifting the parameter vector in time
by ts and repeating the above procedure, the constraint is either found to be satisfied for all t ∈ Ic or
a constraint violation is detected. Shifting the parameter vector by ts amounts in multiplying z with the
matrix exponential
T
eM ts z → z.
(53)
The procedure is illustrated by the flow chart depicted in Fig. 1 (left).
The efficiency of the proposed strategy can be improved via the following observation. A constraint
violation occurring close to t = 0, is often found within the first few iterations. However, if no constraint
violation occurs, the whole interval Ic needs to be traversed, which tends to increase computation. The
computational effort may be reduced by including additional conservative constraint satisfaction checks.
For example, upper and lower bounds can be tightened by a factor γ ∈ (0, 1) such that the satisfaction of
γll ≤ τ (ti )T z ≤ γlu
(54)
for certain time instances ti implies (30) (for all t ∈ Ic ). As a result, at each iteration, the above inequality
is checked. If it is found to be fulfilled, then constraint satisfaction can be guaranteed.
9
Ic = [0, Tc ], z
0→t
constraint
satisfaction
true
check (54)
ts
lu
compute tu , tl
bu (t)
τ (t)T z
tu
t
constraint
violation
true
check (51)
ll
compute ts
constraint
satisfaction
true
bl (t)
T
eM ts z → z
t + ts → t
t + ts > Tc
Figure 1: Left: Flow chart that illustrates the proposed constraint satisfaction check. Right: Illustration of
the upper and lower bounds bu (t) and bl (t) obtained from the Taylor expansion of τ (t)T z. In that case the
test (51) is indecisive, and constraint satisfaction over the interval [0, ts ] can be guaranteed.
5
Properties B1-B5
The sets C, CUs , and CLs have the following properties
B1) the sets CUs , CLs , and C are closed and convex.
B2) the sets π s (CUs ), π s (CLs ) are closed and convex.
B3) CUs ⊂ CUs+1 ⊂ C.
B4) CLs ⊃ CLs+1 ⊃ C.
B5) (π s )∗ π s (CUs ) ⊂ CUs , (π s )∗ π s (CLs ) ⊂ CLs .
We will sketch the proofs of Properties B1-B5 in the following. It follows from the linearity of the stage
constraint, and the fact that the terminal constraint X is convex, that the sets C, CUs , and CLs are likewise
convex. We will sketch the proof that the set CLs is closed. The argument can be translated to the set C by
using the following variational formulation
Z
C = {(x, u, xT ) ∈ X |
δpT (−Cx x − Cu u + b)dt ≥ 0
I
Z
2
∀δp ∈ Lnc : δp ≥ 0,
δpT bdt < ∞; xT ∈ X }. (55)
I
CLs
We will argue indirectly, i.e. that the complement of
is open. To that extent we choose (x, u, xT ) ∈
s
X \ CL . As a result, there exists a test function δ p̃, with δ p̃ ≥ 0, which is spanned by the first s basis
functions and is such that
Z
δ p̃T (−Cx x − Cu u + b)dt < 0.
(56)
I
10
For any x̂ ∈ L2n , û ∈ L2m , and x̂T ∈ Rn , with ||x̂ − x|| < , ||û − u|| < , and |x̂T − xT | < , it follows that
Z
Z
Z
T
T
δ p̃ (−Cx x̂ − Cu û + b)dt = δ p̃ (−Cx x − Cu u + b)dt + δ p̃T (Cx (x − x̂) + Cu (u − û))dt, (57)
I
I
I
where the last integral can be bounded by (using the Cauchy-Schwarz inequality)
||δ p̃||(|Cx |||x − x̂|| + |Cu |||u − û||) < ||δ p̃||(|Cx | + |Cu |).
(58)
We can infer from X being closed, that there exists an open ball centered at xT , which does not intersect
X . As a result, by choosing small enough, it can be concluded that
Z
δ p̃T (−Cx x̂ − Cu û + b)dt < 0, x̂T 6∈ X ,
(59)
I
and therefore (x̂, û, xT ) ∈ X \ CLs , for all x̂ ∈ L2n , û ∈ L2m , x̂T ∈ Rn with ||x − x̂|| < , ||u − û|| < , and
|x̂T − xT | < . Hence, the complement of CLs is open, and therefore CLs is closed.
Note that the set CUs is given by the intersection of the set C with the linear subspace X s spanned by the
first s basis functions. Both of these sets are closed3 implying that CUs is closed as well.
The projection π s is linear, which asserts the convexity of the sets π s (CUs ) and π s (CLs ). Moreover, it is
surjective, and hence, by the open mapping theorem, it follows directly from CLs and CUs being closed that
π s (CLs ) and π s (CUs ) are closed as well.
The inclusion CUs ⊂ CUs+1 follows directly from CUs = C ∩ X s and the inclusion X s ⊂ X s+1 . In
other words, given (x, u, xT ) ∈ CUs , the parameter vectors ηx and ηu , corresponding to the state and input
trajectories x and u, can be extended with zeros resulting in trajectories x̂, û spanned by s+1 basis functions.
But x̂ = x and û = u and therefore (x, u, xT ) ∈ CUs+1 . The inclusion CLs ⊃ CLs+1 follows from the fact that
the dimension of the subspace to which the test functions δ p̃ are constrained increases with s.
The claim (π s )∗ π s (CUs ) ⊂ CUs follows from the fact that (π s )∗ π s is a projection from X onto X s ⊂ X
and that CUs ⊂ X s . The claim that (π s )∗ π s (CLs ) ⊂ CLs follows from the linearity of the stage constraints.
More precisely, it follows by noting that for any δ p̃ = (Inc ⊗ τ )T δηp and any x ∈ L2n ,
Z
δ p̃T Cx xdt = δηpT Cx π ns (x)
I
Z
= δ p̃T Cx (π ns )∗ π ns (x)dt
(60)
I
holds.
6
Properties C1-C5
The sets D, DUs , and DLs have the following properties
C1) the sets DUs , DLs , and D are closed and convex.
C2) the sets π s (DUs ) and π s (DLs ) are closed and convex.
C3) DUs ⊂ DUs+1 ⊂ D.
C4) DLs ⊃ DLs+1 ⊃ D.
C5) (π s )∗ π s (DUs ) ⊂ DUs , (π s )∗ π s (DLs ) ⊂ DLs .
3
X s is finite-dimensional, thus complete, and hence also closed.
11
We will sketch the proof of Properties C1-C5 below.
The convexity of the sets DUs , DLs , and D follows directly from the linearity of the dynamics.
The fact that the set DLs is closed can be seen by a similar argument used for showing closedness of C˜s
in Sec. 5, that is, showing that X \ DLs is open. The set D can be characterized using the variational equality,
Z
Z
T
− δ ṗ xdt − δpT (Ax + Bu)dt − δp(0)T x0 + δp(T )T xT = 0, ∀δp ∈ Hn ,
(61)
I
I
where Hn denotes the set of functions in L2n having a weak derivative in L2n . Thus, again a similar argument
can be applied to show that D is closed. It follows that DUs is closed, since DUs is defined as the intersection
of D with the closed set X s .
The linearity and surjectivity of the map π s implies that π s (DUs ) and π s (DLs ) are indeed closed (by the
open mapping theorem) and convex.
The inclusion DUs ⊂ DUs+1 ⊂ D for all s follows directly from the fact that X s ⊂ X s+1 ⊂ X and
s
DU = D ∩ X s . The inclusion DLs+1 ⊂ DLs for all s can be seen by noting that the variational equality
in (16) has to hold for variations spanned by more and more basis functions as s increases. The claim
that D is contained in DLs follows from the equivalence of (61) with the formulation in (7). The properties
(π s )∗ π s (DUs ) ⊂ DUs and (π s )∗ π s (DLs ) ⊂ DLs can be shown using the same arguments as in 5, where the latter
relies on the linearity of the dynamics.
7
Convergence results
We prove the following result for the case where I has infinite measure.
Lemma 7.1 Given that the basis functions form an algebra and that the basis functions are dense in the set
of smooth functions with compact support in I, it holds that
lim CUs = lim CLs .
s→∞
s→∞
Proof The idea of the proof is the following: We claim that lims→∞ CUs ⊃ lims→∞ CLs . We assume that the
claim is incorrect and show that this leads to a contradiction. Thus, we choose (x, u, xT ) ∈ lims→∞ CLs ,
such that there exists an open set U (bounded) and a k ∈ {1, 2, . . . , nc }, for which
Z
δv(−Cxk x − Cuk u + bk )dt < 0
(62)
I
holds for all smooth test functions δv : I → R, δv ≥ 0, with support in U , and δv(t0 ) > 0 for some t0 ∈ U .
Due to the smoothness of the test functions, δv(t0 ) > 0 readily implies that there is an open neighborhood
of t0 , denoted by N (t0 ), such that δv(t) > 0, ∀t ∈ N (t0 ). The above integral exists, since δv is bounded,
has compact support and x ∈ L2n , u ∈ L2m . We fix t0 ∈ U and pick one of these variations that is positive,
strictly positive at time t0 , and has support in U , which we name δp. Due to the fact that√the basis functions
are dense in the
√ set of smooth functions with compact support, there exists a sequence δ p̃i that converges
uniformly to δp. Due to the fact that the basis functions form an algebra, δ p̃i lies likewise in the span of
the basis functions. Moreover, for a given > 0 (small enough) there exits an integer N > 0 such that
||δ p̃i − δp||∞ < C1
holds for all integers i > N , where C1 > 0 is constant.
We claim that there is an integer p > 0 such that the basis function τp : I → R is nonzero for t0 .
This can be shown by a contradiction argument: If the claim was not true, then all basis functions τ p =
(τ1 , τ2 , . . . , τp ) would be zero at time t0 . The basis functions fulfill Assumption A2, and hence the first order
differential equation τ̇ p = Mp τ p , which is guaranteed to have unique solutions. From τ p (t0 ) = 0, τ̇ p (t0 ) =
12
0 we can infer that τ p (t) = 0 for all t ∈ (0, ∞) is the (unique) solution to τ̇ p = Mp τ p , contradicting
Assumption A1.
Thus, we can choose the basis function τp that is nonzero for t0 and hence also nonzero in a neighborhood around t0 , due to the smoothness of the basis functions. The same applies to the function τp2 , which
is likewise contained in the set of basis functions (the basis functions form an algebra that is closed under
multiplication). Moreover, due to the fact that the basis functions are orthonormal, it follows that
Z
τp2 dt = 1.
(63)
I
In the same way, the function δ p̃i (t)τp2 (t) is also contained in the set of basis functions, is non-negative for
all t ∈ I, and is bounded and integrable for all integers i > 0 (|τ p | is bounded, see 10). Thus, it follows that
the integral
Z
τp2 δ p̃i (−Cxk x − Cuk u + bk )dt
(64)
I
exists. By assumption (x, u, xT ) ∈ lims→∞ CLs , and therefore
Z
0 ≤ τp2 δ p̃i (−Cxk x − Cuk u + bk )dt
ZI
= τp2 δp(−Cxk x − Cuk u + bk )dt
ZI
+ τp2 (δ p̃i − δp)(−Cxk x − Cuk u + bk )dt,
(65)
(66)
(67)
I
where the last term can be bounded by
Z
C1
τp2 |Cxk x + Cuk u − bk |dt.
(68)
I
The above integral is bounded due to the fact that τp2 is bounded and integrable, x ∈ L2n , u ∈ L2m , and that
bk is bounded, and therefore (67) can be made arbitrarily small by sufficiently increasing i. However, this
leads to a contradiction, since (66) is strictly negative, according to (62). This proves that lims→∞ CUs ⊃
lims→∞ CLs . The desired result is then established due to the fact that CUs ⊂ CLs holds for all integers s > 0.
8
Recursive feasibility and closed-loop stability
Proposition 8.1 Provided that the optimization (21) is feasible at time t = 0, it remains feasible for all
times t > 0, and the resulting closed-loop system is guaranteed to be asymptotically stable.
Proof The proof is taken from [2] and included for completeness. The following notation is introduced:
The closed-loop state and input trajectories are denoted by x(t) and u(t). The predicted trajectories are
referred to as x̃(tp |t), ũ(tp |t), where tp > 0 denotes the prediction horizon. For tp = 0, the prediction
matches the true trajectory, that is x̃(0|t) = x(t), ũ(0|t) = u(t) for all t ∈ [0, ∞). The predictions x̃(tp |t),
ũ(tp |t) are obtained by solving (21) subject to the initial condition x0 = x(t), which yields the parameters
ηx and ηu defining x̃(tp |t) and ũ(tp |t) by
x̃(tp |t) = (In ⊗ τ (tp ))T ηx ,
ũ(tp |t) = (Im ⊗ τ (tp ))T ηu .
(69)
In order to highlight the dependence on the initial condition, the resulting optimal cost of (21) is denoted
by J MPC (x(t)).
13
By assumption, (21) is feasible at time t = 0. The resulting trajectories x̃(tp |0), ũ(tp |0) fulfill the
equations of motion, the initial condition x̃(0|0) = x(0), and the constraints and hence, the system evolves
according to x(t) = x̃(t|0), u(t) = ũ(t|0), ∀t ∈ [0, Td ). Due to the time-shift property of the basis functions
implied by Assumption A2, the feasible candidates
x̃(tp + Td |0) = (In ⊗ τ (tp )T )(In ⊗ exp(Ms Td )T )ηx ,
ũ(tp + Td |0) = (Im ⊗ τ (tp )T )(Im ⊗ exp(Ms Td )T )ηu
(70)
for the optimization at time Td can be constructed from the optimizer ηx and ηu at time 0. As a result,
recursive feasibility of (21) follows by induction.
We will show that the optimal cost J MPC acts as a Lyapunov function. The function J MPC is a valid
Lyapunov candidate since J MPC (x) > 0 for all x 6= 0 and J MPC (x) = 0 if and only if x = 0. Due to the fact
that the shifted trajectories x̃(tp +Td |0) and ũ(tp +Td |0) (as defined in (70)) are feasible for the optimization
at time Td , the following upper bound on J MPC (x(Td )) can be established
Z ∞
MPC
J
(x(Td )) ≤
x̃(tp |0)T Qx̃(tp |0) + ũ(tp |0)T Rũ(tp |0)dtp .
(71)
Td
The right-hand side can be rewritten as
Z Td
MPC
J
(x(0)) − x̃(tp |0)T Qx̃(tp |0) + ũ(tp |0)T Rũ(tp |0)dtp ,
(72)
0
resulting in
J MPC (x(Td )) − J MPC (x(0)) ≤
Z Td
x̃(tp |0)T Qx̃(tp |0) + ũ(tp |0)T Rũ(tp |0)dtp .
−
(73)
0
The right-hand side of (73) is guaranteed to be strictly negative, except for x(0) = 0, and thus, by induction,
J MPC (x(kTd )) is strictly decreasing, which concludes the proof.
9
Reduction of the semi-infinite constraint
The following section discusses the reduction of the semi-infinite constraint
al ≤ τ (t)T η ≤ au
(74)
over the interval t ∈ [0, ∞), with al , au ∈ R, al < 0, au > 0 to a compact interval. Thereby we consider
the symmetric case, where |al | = |au | first, before discussing the asymmetric case |al | 6= |au |. It is shown
that the length of this compact time interval depends only on the properties of the basis functions and on
the ratio between |au | and |al |. Both al and au are assumed to be finite.
9.1
The symmetric case
Proposition 9.1 Provided that the basis functions fulfill Assumptions A1 and A2 for all t ∈ [0, ∞) there
exists a positive real number Tc such that
sup |τ (t)T η| = max |τ (t)T η|
t∈[0,Tc ]
t∈[0,∞)
holds for all parameter vectors η ∈ Rs .
14
(75)
Proof Without loss of generality we restrict the parameter vectors to have unit magnitude, that is, |η| = 1.4
We prove the claim in 4 steps. We first derive an exponentially decaying upper bound on the Euclidean
norm of the basis function vector τ . We then use this bound to argue that the basis functions are linearly
independent over the interval [0, Ti ] (the scalar Ti will be determined). The third step consists of constructing
a lower bound on
max |τ (t)T η|
(76)
t∈[0,Ti ]
that holds for all parameter vectors η with |η| = 1. Linear independence of the basis functions on [0, Ti ] will
be used to argue that this lower bound is strictly positive. In the last step, we show that if t is sufficiently
large, |τ (t)T η| will be below this lower bound, which concludes the proof.
Step 1): The fact that Ms is asymptotically stable implies that there exists a quadratic Lyapunov function
that decays exponentially. This provides a means to establish the following bound
|τ (t)|2 ≤ C2 e−c2 t ,
∀t ∈ [0, ∞),
(77)
where C2 > 0, c2 > 0 are constant.
Step 2): We use the Gram matrix to argue that the basis functions are linearly independent over the
interval [0, Ti ]. According to [5, p. 2, Thm. 3] it holds that the basis functions are linearly independent in
the set [0, Ti ] if and only if the matrix
Z Ti
τ τ T dt
(78)
0
has full rank. This is the case if the bilinear form
Z
T
v
Ti
τ τ T dt v
(79)
0
is strictly positive for all v ∈ Rs with |v| = 1. Combining the fact that the basis functions are orthonormal
with the Cauchy-Schwarz inequality, leads to the following lower bound of the above bilinear form,
Z ∞
Z ∞
T 2
|τ |2 dt.
(80)
(v τ ) dt ≥ 1 −
1−
Ti
Ti
Using the upper bound (77) we therefore obtain
Z Ti
Z
T
T
v
τ τ dtv ≥ 1 −
0
∞
C2 e−c2 t dt = 1 −
Ti
C2 −c2 Ti
e
,
c2
(81)
for all v ∈ Rs with |v| = 1. Thus, we fix Ti > 0, such that
1>
C2 −c2 Ti
e
,
c2
(82)
implying that the matrix (78) is positive definite and has therefore full rank. Consequently, the basis functions are guaranteed to be linearly independent on the interval [0, Ti ].
Step 3): We claim that
c∗ := inf max |τ (t)T η|
(83)
|η|=1 t∈[0,Ti ]
is well-defined and strictly positive. To that extent we first prove that the function g : Rs → [0, ∞),
g(η) := max |τ (t)T η|
t∈[0,Ti ]
4
The claim holds trivially for η = 0; in case η 6= 0 we can always normalize η.
15
(84)
is continuous (in fact Lipschitz-continuous). Therefore we consider two parameter vectors η1 and η2 with
g(η1 ) ≤ g(η2 ) (without loss of generality). From the fact that g(η2 ) can be rewritten as g(η1 + (η2 − η1 ))
and by invoking the triangle inequality it can be inferred that
g(η1 + (η2 − η1 )) ≤ max |τ (t)T η1 | + |τ (t)T (η2 − η1 )|.
t∈[0,Ti ]
(85)
By noting that the maximum of the above sum is smaller than the sum of the summand’s maxima it can be
concluded that
g(η2 ) ≤ g(η1 ) + max |τ (t)T (η2 − η1 )|.
(86)
t∈[0,Ti ]
Moreover, by combining the Cauchy-Schwarz inequality and the bound (77) we obtain
p
|g(η2 ) − g(η1 )| ≤ C2 |η2 − η1 |,
(87)
showing that the function g is indeed (Lipschitz) continuous. As a result, the Bolzano-Weierstrass theorem
asserts that the infimum in (83), is attained and well-defined. It remains to argue that c∗ > 0. For the sake
of contradiction, we assume c∗ = 0. This implies the existence of the minimizer η ∗ , with |η ∗ | = 1, which
fulfills
max |τ (t)T η ∗ | = 0.
(88)
t∈[0,Ti ]
As a result, it follows that τ (t)T η ∗ is zero for all t ∈ [0, Ti ], which contradicts the fact that the basis functions
τ are linearly independent on [0, Ti ].
Step 4): From the upper bound (77) and the Cauchy-Schwarz inequality it follows that
p
c2 t
|τ (t)T η| ≤ C2 e− 2 , ∀t ∈ [0, ∞),
(89)
√
and for all η ∈ Rs with |η| = 1. Clearly, c∗ ≤ C2 and therefore we can choose the time Tc such that
c∗ =
p
c2 Tc
C2 e− 2 ,
(90)
implying
sup |τ (t)T η| < c∗ ≤ max |τ (t)T η|,
t∈[0,Ti ]
t∈(Tc ,∞)
(91)
for all η ∈ Rs with |η| = 1. This proves the claim.
9.2
The asymmetric case
Proposition 9.2 Provided that the basis functions fulfill Assumptions A1 and A2 for all t ∈ [0, ∞) there
exists a positive real number Tc such that
al ≤ τ (t)T η ≤ au ,
∀t ∈ [0, Tc ]
(92)
al ≤ τ (t)T η ≤ au ,
∀t ∈ [0, ∞),
(93)
implies
for any η ∈ Rs , where al , au ∈ R, al < 0, au > 0.
Proof We define f˜ := τ T η and establish upper and lower bounds on f˜. Without loss of generality we
assume η 6= 0. Combining Assumption A2 with the Caley-Hamilton theorem leads to
f˜(s) (t) + a1 f˜(s−1) (t) + · · · + as f˜(t) = η T τ (s) (t) + a1 η T τ (s−1) (t) + · · · + as η T τ (t)
= η T M s + a1 M s−1 + · · · + as M τ (t) = 0,
16
(94)
(95)
where a1 , a2 , . . . as are the coefficients of the characteristic polynomial of the matrix M . Thus, the trajectory
f˜ and its time derivatives fulfill the following set of differential equations
τ (0)T
0
1
0
... 0
0
τ (0)T M T
0
1
... 0
f,
f
(0)
=
(96)
f˙ = ..
η,
..
..
..
..
.
.
.
.
.
−as −as−1 −as−2 . . . −a1
τ (0)T (M (s−1) )T
{z
}
{z
}
|
|
:=M̂
:=Ĥ
where f := (f˜, f˜(1) , . . . , f˜(s−1) ). The matrix M̂ is Hurwitz and therefore, due to the Lyapunov theorem,
there exists a symmetric matrix P > 0, P ∈ Rs×s that satisfies
P M̂ + M̂ T P + Q = 0
(97)
for any symmetric matrix Q > 0, Q ∈ Rs×s . We fix the positive definite matrix Q and consider the
quadratic Lyapunov function V (t) = f (t)T P f (t), where P satisfies (97). The time derivative of V can be
upper bounded by
λQ T
f Pf
λP
λQ
≤ − P V,
λ
V̇ = −f T Qf ≤ −λQ |f |2 ≤ −
(98)
(99)
where the minimum eigenvalue of Q is denoted by λQ and the maximum eigenvalue of P is denoted by λP .
As a result, this yields the upper bound
λQ
V (t) ≤ V (0)e− λP t ≤ λP λĤ
T Ĥ
λQ
|η|2 e− λP t ,
(100)
T
where λĤ Ĥ denotes the maximum eigenvalue of the matrix Ĥ T Ĥ. According to the proof of Prop. 9.1, we
may choose Ti such that
C2 −c2 Ti
1>
e
,
(101)
c2
implying that the basis functions are linearly independent on the interval [0, Ti ] (see (82) and Prop. 9.1 for
the definition of the constants c2 and C2 ). Linear independence can be used to establish the following lower
bound, c.f. Prop. 9.1:
Z
Z
Ti
Ti
|f˜|2 dt = η T
τ τ T dt η ≥ c3 |η|2 ,
(102)
where the constant c3 > 0 denotes the minimum eigenvalue of the matrix
Z Ti
τ τ T dt.
(103)
0
0
0
Without loss of generality it is assumed that |al | ≤ |au |. Choosing the real number Tc > Ti implies that the
constraint is imposed over the interval [0, Ti ] and therefore
Z Ti
|f˜|2 dt ≤ Ti |au |2 .
(104)
0
Combined with the above upper bound on |η|2 it follows that
|η|2 ≤
Ti |au |2
.
c3
17
(105)
This can be used to upper bound the squared magnitude of f˜(t), that is,
T
λP λĤ Ĥ |au |2 Ti − λQP t
1
2
˜
V (t) ≤
e λ ,
|f (t)| ≤
λP
λ P c3
(106)
where λP denotes the minimum eigenvalue of the matrix P . As a result, we may choose Tc such that the
above upper bound is below |al |2 . This may be achieved by choosing Tc > max{t̂, Ti }, where
|au |
c3 λ P
λP
2 ln
+ ln
t̂ := −
(107)
λQ
|al |
Ti λP λĤ T Ĥ
and ln refers to the natural logarithm.
10
Additional properties
In the following we will discuss some of the properties of the basis functions that fulfill Assumptions A1
and A2.
10.1
Conditions imposed on Ms
The fact that the basis functions are orthonormal on the interval I = (0, T ) implies
Z
d
T
T
τ (T )τ (T ) − τ (0)τ (0) =
(τ τ T )dt
dt
I
= Ms + MsT .
(108)
(109)
In case the interval I has infinite measure, that is, I = (0, ∞), the above formula reduces naturally to
−τ (0)τ (0)T = Ms + MsT .
10.2
(110)
Bounds on the Euclidean norm
In case the interval I has infinite measure, that is, I = (0, ∞), it holds that
|τ (t)| ≤ |τ (0)|,
∀t ∈ [0, ∞).
(111)
This results from
d
(τ (t)T τ (t)) = τ (t)T (Ms + MsT )τ (t)
dt
= −|τ (0)T τ (t)|2 ≤ 0,
(112)
(113)
for all t ∈ [0, ∞), which follows from (110).
10.3
Finite number of minima and maxima
In the following we will argue that a function that is not everywhere zero and spanned by basis functions
fulfilling Assumptions A1 and A2, has a finite number of minima and maxima. The function will be
denoted by f˜ := τ T η, η ∈ Rs , η 6= 0, where the function and the basis functions are defined over the
interval t ∈ I = (0, T ). If I has infinite measure, it follows from Sec. 9 that f˜ takes is maxima and minima
within a compact interval, and hence, without loss of generality, we assume in the following that I has finite
measure. Moreover, due to the Caley-Hamilton theorem it holds that
f˜(s) (t) + a1 f˜(s−1) (t) + · · · + as f˜(t) = 0,
18
(114)
for all t ∈ I, where the ak , k = 1, 2, . . . , s are the coefficients of the characteristic polynomial of the matrix
Ms . According to [6], the time derivative of f˜ has at most s − 1 zeros on any subinterval of length l, where
s
X
ak l k
k=1
k!
< 1.
(115)
This proves readily that f˜ has a finite number of minima and maxima in the interval I.
References
[1] R. Fletcher. Practical Methods of Optimization. John Wiley & Sons, second edition, 1987.
[2] M. Muehlebach and R. D’Andrea. Parametrized infinite-horizon model predictive control for linear
time-invariant systems with input and state constraints. Proceedings of the American Control Conference, pages 2669–2674, 2016.
[3] Y. Nesterov. Squared functional systems and optimization problems. In H. Frenk, K. Roos, T. Terlaky,
and S. Zhang, editors, High Performance Optimization, chapter 17, pages 405–440. Springer, 2000.
[4] J. Nocedal and S. J. Wright. Numerical Optimization. Springer, second edition, 2006.
[5] G. Sansone. Orthogonal Functions. Interscience Publishers, 1959.
[6] S. Yakovenko. On functions and curves defined by ordinary differential equations. Fields Institute
Communications, 24:497–525, 1999.
19
| 3cs.SY |
Computing local zeta functions of groups,
algebras, and modules
Tobias Rossmann
arXiv:1602.00919v1 [math.GR] 2 Feb 2016
Fakultät für Mathematik, Universität Bielefeld, D-33501 Bielefeld, Germany
We develop a practical method for computing local zeta functions of groups,
algebras, and modules in fortunate cases. Using our method, we obtain a complete
classification of generic local representation zeta functions associated with unipotent
algebraic groups of dimension at most six. We also determine the generic local
subalgebra zeta functions associated with gl2 (Q). Finally, we introduce and compute
examples of graded subobject zeta functions.
1. Introduction
Zeta functions counting subobjects and representations. By considering associated
Dirichlet series, various algebraic counting problems give rise toQa global zeta function
Z(s) which admits a natural Euler product factorisation Z(s) = p Zp (s) into local zeta
functions Zp (s) indexed by rational primes p. For example, Z(s) could be the Dirichlet
series enumerating subgroups of finite index within a finitely generated nilpotent group
and Zp (s) might enumerate those subgroups of p-power index only (see [26]); in the
special Q
case of the infinite cyclic group, we then recover the classical Euler factorisation
ζ(s) = p 1/(1 − p−s ) of the Riemann zeta function.
This article is concerned with three types of counting problems and associated zeta
functions; all of these problems arose from (and remain closely related to) enumerative
problems for nilpotent groups.
• ([26]) Enumerate subalgebras of finite additive index of a possibly non-associative
algebra, e.g. a Lie algebra (possibly taking into account an additive grading).
• ([42]) Enumerate submodules of finite additive index under the action of an integral
matrix algebra.
• ([27, 46]) Enumerate twist-isoclasses of finite-dimensional complex representations
of a finitely generated nilpotent group.
2010 Mathematics Subject Classification. 11M41, 20F69, 20G30, 20F18, 20C15.
Keywords. Subgroup growth, representation growth, zeta functions, unipotent groups, Lie algebras.
This work is supported by the DFG Priority Programme “Algorithmic and Experimental Methods
in Algebra, Geometry and Number Theory” (SPP 1489).
Generic local zeta functions. Each of the preceding three counting problems provides us
with a global
Q zeta function Z(s) (namely the associated Dirichlet series) and a factorisation
Z(s) = p Zp (s) as above. The goal of this article is to compute the generic local zeta
functions Zp (s) at least in favourable situations—that is, we seek to simultaneously
determine Zp (s) for almost all p using a single finite computation. To see why this is a
sensible problem, we first recall some theory.
In the cases of interest to us, each Zp (s) will be a rational function in p−s over Q. In
particular, the task of computing one local zeta function Zp (s) using exact arithmetic
is well-defined. Regarding the behaviour of Zp (s) under variation of p, in all three
cases from above, sophisticated results from p-adic integration imply the existence of
schemes V1 , . . . , Vr and rational functions W1 , . . . , Wr ∈ Q(X, Y ) such that for almost
all primes p,
r
X
Zp (s) =
#Vi (Fp ) · Wi (p, p−s );
(1.1)
i=1
for more details, see Theorem 4.1 below. While constructive proofs of (1.1) are known,
they are generally impractical due to their reliance on resolution of singularities.
Previous work: computing topological zeta functions. In [35, 36, 38], the author
developed practical methods for computing so-called topological zeta functions associated
with the above counting problems; these zeta functions are derived from generic local
ones by means of a termwise limit “p → 1” applied to a formula (1.1). Due to their
reliance on non-degeneracy conditions for associated families of polynomials, the author’s
methods for computing topological zeta functions do not apply in all cases. However,
whenever they are applicable, as we will explain below, they come close to producing an
explicit formula (1.1).
Computing generic local zeta functions. In general, we understand the task of computing Zp (s) for almost p to be the explicit construction of Vi and Wi as in (1.1). While
this seems to be the only adequate general notion of “computing” generic local zeta
functions, we will often be more ambitious in practice.
Uniformity Problem. Decide if there exists W ∈ Q(X, Y ) such that Zp (s) = W (p, p−s )
for almost all primes p; in that case, we call (Zp (s))p prime uniform. Find W if it exists.
The term “uniformity” is taken from [18, §1.2.4]. In practice, a weaker, non-constructive
form of the Uniformity Problem which merely asks for the existence of W as above is
often easier to solve. For example, if Zp (s) is the zeta function enumerating subgroups
(or normal subgroups) of finite index in the free nilpotent pro-p group of some fixed
finite rank (independent of p) and class 2, then (Zp (s))p prime is shown to be uniform in
[26, Thm 2] even though no explicit construction of a rational function W is given.
For many cases of interest, a rational function W as in the Uniformity Problem exists,
see e.g. most examples in [18]. However, no conceptual explanation as to why this is so
seems to be known beyond explicit computations.
2
Woodward [49] used computer-assisted calculations to solve the Uniformity Problem
for a large number of subalgebra and ideal zeta functions of nilpotent Lie algebras.
Unfortunately, few details on his computations are available, rendering them rather
difficult to reproduce.
Results. While explicit formulae (1.1) have been obtained for specific examples and even
certain infinite families of these, all known general constructions of Vi and Wi as in (1.1)
are impractical. In full generality, we thus regard the Uniformity Problem as too ambitious
a task. In the present article, we extend the author’s work on explicit, combinatorially
defined formulae (1.1) (see [35, 36, 38]) in order to provide practical solutions to the
Uniformity Problem in fortunate cases. We will also consider computations of generic
local zeta functions in cases where no W as above exists.
As the following list illustrates, the method developed here can be used to compute a
substantial number of interesting new examples of generic local zeta functions:
• We completely determine the generic local representation zeta functions associated
with unipotent algebraic groups of dimension at most six (§8, Table 1).
• We compute the generic local subalgebra zeta functions associated with gl2 (Q);
this constitutes only the second instance (after sl2 (Q)) where such zeta functions
associated with an insoluble Lie algebra have been computed (§9.1, Theorem 9.1).
• We compute the generic local submodule zeta functions for the natural action of the
group of upper unitriangular integral n × n-matrices (or, equivalently, the nilpotent
associative algebra of strictly upper triangular integral n × n-matrices) for n 6 5
(§9.4, Theorem 9.5).
• We compute the graded subalgebra and ideal zeta functions associated with Q-forms
of each of the 26 “fundamental graded” Lie algebras of dimension at most six over
C (§10, Tables 2–3).
Outline. In §2, we recall definitions of the subobject and representation zeta functions
of concern to us. In §3, as a variation of established subalgebra and ideal zeta functions,
we discuss graded versions of these zeta functions. In §4, we consider formulae such as
(1.1) both in theory and as provided by the author’s previous work. Our work on the
Uniformity Problem then proceeds in two steps. First, in §5, we consider the symbolic
determination of numbers such as the #Vi (Fp ) in (1.1) as a function of p. Thereafter,
in §6, we discuss the explicit computation of the rational functions Wi as provided by
[35, 36, 38]; a key role will be played by algorithms of Barvinok et al. [5–7] surrounding
generating functions of rational polyhedra. In §7, we consider “reduced representation zeta
functions” in the spirit of Evseev’s work [20]; while these functions turn out to be trivial,
they provide us with a simple necessary condition for the correctness of calculations.
Finally, examples of generic local zeta functions are the subject of §§8–10.
3
Acknowledgement
I would like to thank Christopher Voll for interesting discussions.
Notation
The symbol “⊂” indicates not necessarily proper inclusion. For the remainder of this
article, let k be a number field with ring of integers o. We write Vk for the set of
non-Archimedean places of k. For v ∈ Vk , we denote by kv the v-adic completion of k
and by ov the valuation ring of kv . We further let pv ∈ Spec(o) denote the prime ideal
corresponding to v ∈ Vk and write qv = |o/pv |. Finally, we let | · |v denote the absolute
value on kv with |π|v = qv−1 for π ∈ pv \ p2v .
We let Qp and Zp denote the field of p-adic numbers and ring of p-adic integers,
respectively. By a p-adic field, we mean a finite extension of Qp . For a p-adic field K, let
OK denote the valuation ring of K and let PK denote the maximal ideal of OK . We
write qK = |OK /PK |.
2. Established zeta functions of groups, algebras, and modules
2.1. Subalgebra and ideal zeta functions
Following [26] (cf. [35, §2.1]), for a commutative ring R and a (possibly non-associative)
R-algebra A, we formally define the subalgebra zeta function of A to be
X
ζA6 (s) =
|A : U|−s ,
U
where U ranges over the R-subalgebras of A such that the R-module quotient A/U has
finite cardinality |A : U|. Additional hypotheses (which are satisfied in our applications
below) ensure that the number an (A) of R-subalgebras of index n of A is finite for every
n > 1 and, in addition, an (A) grows at most polynomially as a function of n. Under
these assumptions, ζA6 (s) defines an analytic function in some complex right half-plane.
Now let A be a finite-dimensional possibly non-associative k-algebra, where k is a
number field as above. Choose an o-form A of A whose underlying o-module is free.
For v ∈ Vk , let AQ
v := A ⊗o ov , regarded as an ov -algebra. We then have an Euler
6
product ζA (s) = v∈Vk ζA6v (s); see [35, Lem. 2.3]. While the global zeta function
ζA6 (s) is an analytic object, as we will recall below, the local zeta functions ζA6v (s)
are algebro-geometric
in nature. Note that up to discarding finitely many elements, the
family ζA6v (s) v∈V of local zeta functions only depends on A and not on the o-form A.
k
If, instead of enumerating subalgebras, we consider ideals, we obtain the global and
local ideal zeta functions ζA/ (s) and ζA/ v (s) of A, respectively; these are also linked by
an Euler product as above.
4
2.2. Submodule zeta functions
Submodule zeta functions were introduced by Solomon [42] in the context of semisimple
associative algebras. In the following generality (based upon [35, §2.1]), they also
generalise ideal zeta functions of algebras. For a commutative ring R, an R-module V,
and a set Ω ⊂ EndR (V), we formally define the submodule zeta function of Ω acting
on V to be
X
ζΩyV (s) =
|V : U|−s ,
U
where U ranges over the Ω-invariant R-submodules of V with finite R-module quotients V/U. The name “submodule zeta function” is justified by the observation that we
are free to replace Ω by its enveloping unital associative algebra within EndR (V).
Let V be a finite-dimensional vector space over k and let Ω ⊂ Endk (V ) be given.
Choose an o-form V of V which is free as an o-module. Furthermore, choose a finite
set Ω ⊂ Endo (V) which generates the same unital subalgebra
of Endk (V ) as Ω. Writing
Q
Vv = V ⊗o ov , we obtain an Euler product ζΩyV (s) = v∈Vk ζΩyVv (s); as in §2.1, up to
discarding finitely many factors, the collection of local zeta functions on the right-hand
side of this product only depends on (Ω, V ) and not on the choice of (Ω, V).
2.3. Representation zeta functions associated with unipotent groups
Following [27,43], for a topological group G, we let r̃n (G) denote the number of continuous
irreducible representations G → GLn (C) counted up to equivalence and tensoring with
continuous 1-dimensional complex representations. We formally define the (twist)
representation zeta function of G to be
f
irr
ζG
(s)
=
∞
X
r̃n (G)n−s .
n=1
Let G be a unipotent algebraic group over k; see [11, Ch. IV] for background. Let
Un denote the group scheme of upper unitriangular n × n-matrices. We choose an
embedding of G into some Un ⊗k and let G 6 Un ⊗o be the associated o-form of G (viz.
the scheme-theoretic closure of G within Un ⊗o). By [43, Prop. 2.2], the Euler product
Q
f
f
irr
irr
ζG(o)
(s) = v∈Vk ζG(o
(s) connects the representation zeta function of the discrete group
v)
G(o) and those of the pro-pv groups G(ov ), where pv is the rational prime contained in pv .
2.4. Motivation: zeta functions of nilpotent groups
We briefly recall the original motivation for the study of subalgebra and ideal zeta
function from [26] and representation zeta functions in [27, 46] (cf. [43]). For any
6
topological group G, the subgroup zeta function ζG
(s) (resp. the normal subgroup
P
/
zeta function ζG (s)) of G is formally defined to be H |G : H|−s , where H ranges of
the closed subgroups (resp. closed normal subgroups) of G of finite index.
Let G be a
Q
6
discrete torsion-free finitely generated nilpotent group. Then ζG
(s) = p ζ 6 (s), where
Ĝp
5
p ranges over primes and Ĝp denotes the pro-p completion of G. Moreover, the global
6
and local zeta functions ζG
(s) and ζ 6 (s) all converge in some complex right half-plane.
Ĝp
Analogous statements hold for the normal subgroup and representation zeta functions of
G.
Apart from finitely many exceptions, the local subobject and representation zeta
functions attached to G are special cases of those in §§2.1–2.2. Recall that the Mal’cev
correspondence attaches a finite-dimensional nilpotent Lie Q-algebra, L say, to G. As
explained in [26], if L is a Z-form of L which is finitely generated as a Z-module, then
6
/ (s) = ζ /
ζ 6 (s) = ζL⊗Z
(s) and ζĜ
L⊗Zp (s) for almost all p. Moreover, if G is the unipotent
p
Ĝp
p
algebraic group over Q with Lie algebra L and if G is a Z-form of G arising from an
embedding G 6 Un ⊗Q, then Ĝp = G(Zp ) for almost all primes p (see [43]).
3. Graded subalgebra and ideal zeta functions
In this section, we introduce variations of the subalgebra and ideal zeta functions from
§2.1 which take into account a given additive grading of the algebra under consideration.
3.1. Definitions
Let R be a commutative ring and let A be a possibly non-associative R-algebra. Further
suppose that we are given a direct sum decomposition
A = A1 ⊕ · · · ⊕ Ar
(3.1)
of R-modules. As usual, an R-submodule U 6 A is homogeneous if it decomposes as
U = U1 ⊕ · · · ⊕ Ur for R-submodules Ui 6 Ai for i = 1, . . . , r. We formally define the
graded subalgebra zeta function of A with respect to the decomposition (3.1) to be
X
ζAgr6 (s) =
|A : U|−s ,
U
where U ranges over the homogeneous R-subalgebras of A such that the R-module quotient
A/U is finite. We also define the graded ideal zeta function ζAgr / (s) in the evident
way. Note that we do not require (3.1) to be compatible with the given multiplication
in A. As in the non-graded context, given a finite-dimensional possibly non-associative
k-algebra A together with a vector space decomposition A = A1 ⊕ · · · ⊕ Ar , we obtain
associated global and local graded subalgebra and ideal zeta functions generalising those
from §2.1 by choosing appropriate o-forms.
Example 3.1. Let A = Zn1 ⊕ · · · ⊕ Znr be regarded as an abelian Lie Z-algebra for
n1 , . . . , nr > 1. Q
It follows
from the well-known non-graded case (r = 1; see [26, Prop. 1.1])
Q i −1
that ζAgr6 (s) = ri=1 nj=0
ζ(s − j), where ζ denotes the Riemann zeta function.
Remark 3.2. Let R, V, and Ω ⊂ EndR (V) be as in §2.2. Fix an R-module decomposition
V = V1 ⊕ · · · ⊕ Vr . In analogy to the above, we define the graded submodule zeta
gr
function ζΩyV
(s) of Ω by enumerating homogeneous Ω-invariant R-submodules of V.
6
3.2. Reminder: graded Lie algebras
Let R be a commutative Noetherian ring. All Lie R-algebras in the following are assumed
to be finitely generated as R-modules. Recall that an (N-)graded
Lie algebra over
L
g
R is a Lie R-algebra g together with a decomposition g = ∞
into
R-submodules
i=1 i
gi 6 g such that [gi , gj ] 6 gi+j for all i, j > 1. Since g is Noetherian as an R-module,
gi = 0 for sufficiently large i whence such an algebra g is nilpotent. Following [31, §2,
Def. 1], we say that g is fundamental if [g1 , gi ] = gi+1 for all i > 1. If R = R or R = C,
then the fundamental graded Lie R-algebras of dimension at most 7 have been classified
in [31]. In the case of dimension at most 5, the classification in [31] is in fact valid over
any field of characteristic zero; see [31, §2.2, Rem. 1].
Let g be a finite-dimensional Lie algebra over a field. Let g = g1 ⊃ g2 ⊃ · · ·L
be the lower
i i+1
central series of g. As is well-known, commutation in g endows gr(g) := ∞
i=1 g /g
with the structure of a graded Lie algebra; note that gr(g) is fundamental by construction.
We call gr(g) the graded Lie algebra associated with g.
The study of graded zeta functions seems quite natural in the context of nilpotent Lie
algebras. It would be interesting to find group-theoretic interpretations, in the spirit
of §2.4, of such zeta functions associated with graded nilpotent Lie algebras.
3.3. Graded subobject zeta functions as p-adic integrals
In order to carry out explicit computations of local graded subobject zeta functions,
we will use the following straightforward variation of [15, §5]; we only spell out the
enumeration of graded subalgebras, the case of ideals being analogous.
Theorem 3.3. Let O be the valuation ring of a non-Archimedean local field. Let A
be a (possibly non-associative) O-algebra whose underlying O-module is free with basis
a = (a1 , . . . , ad ). Let 0 = β1 < · · · < βr+1 = d and decompose A = A1 ⊕ · · · ⊕ Ar by
setting Ai = Oa1+βi ⊕ · · · ⊕ Oaβi+1 .
Let T denote the O-module of block diagonal upper triangular d × d-matrices over O
with block sizes β2 − β1 , . . . , βr+1 − βr . Let M (X) be the generic matrix of the same
shape over O; in other words,
X1,1 . . . X1,β2
X1+βr ,1+βr . . . X1+βr ,d
.. , . . . ,
.. .
..
..
M (X) = diag
.
.
.
.
Xβ2 ,β2
Xd,d
Let R = O[X] and let ? : Rd × Rd → Rd be induced via base extension by multiplication
in A with respect to a. Let F ⊂ R consist of all entries of all d-tuples (Mi (X) ?
Mj (X)) adj(M (X)) for 1 6 i, j 6 d, where adj(M (X)) denotes the adjugate matrix
of M (X) and Mi (X) the ith row of M (X). Define V = {x ∈ T : det(M (x)) |
f (x) for all f ∈ F }. Let q denote the residue field size of O, let µ denote the normalised
Pr
βi+1 −βi +1
) , and let | · | denote the absolute value on K such
2
Haar measure on T ≈ O i=1 (
7
that |π| = q −1 for any uniformiser π. Then
Z Y
−βi
r βi+1
Y
gr6
−1 −d
ζA (s) = (1 − q )
|xj+βi ,j+βi |s−j dµ(x).
V
i=1
(3.2)
j=1
Remark 3.4. As in [15, §5], a matrix x ∈ T belongs to the set V in Theorem 3.3 if
and only if its row span is a subalgebra of Od , regarded as an algebra via the given
identification A = Od .
The following illustrates Theorem 3.3 for an infinite family of graded algebras.
Proposition 3.5. Let n > 1 and let m(n) = m1 (n) ⊕ · · · ⊕ mn (n) be the graded Lie Zalgebra of additive rank n + 1 and nilpotency class n with m1 (n) = Ze0 ⊕ Ze1 , mi (n) = Zei
for i = 2, . . . , n, and non-trivial commutators [e0 , ei ] = ei+1 for 1 6 i 6 n − 1. Let k be a
number field with ring of integers o. Then for each v ∈ Vk ,
gr /
−s
1−s
−3s
−4s
−(n+1)s
ζm(n)⊗o
(s)
=
1/
(1
−
q
)(1
−
q
)(1
−
q
)(1
−
q
)
·
·
·
(1
−
q
)
,
v
v
v
v
v
v
where m(n) ⊗ ov is regarded as an ov -algebra. Denoting the Dedekind zeta function of k
by ζk (s), we thus have
gr /
ζm(n)⊗o
(s) = ζk (s)ζk (s − 1)ζk (3s)ζk (4s) · · · ζk (n + 1)s .
Proof. It is an elementary consequence of Theorem 3.3 and Remark 3.4 (both applied to
the enumeration of ideals instead of subalgebras) that for any v ∈ Vk ,
gr /
ζm(n)⊗o
(s) = (1 − qv−1 )−n−1
v
Z
s−2
s−1
× |x1 |s−1
· · · |yn−1 |s−1
dµ(x1 , x2 , x3 , y1 , . . . , yn−1 ),
v
v |x3 |v |y1 |v
V
where V = (x1 , x2 , x3 , y1 , . . . , yn−1 ) ∈ on+2
: yn−1 |yn | ·· · | y1 | x1 , x2 , x3 ; indeed,
v
(x1 , . . . , yn−1 ) ∈ V if and only if the row span of diag( x1 xx23 , y1 , . . . , yn−1 ) is an ideal of
m(n) ⊗ ov (identified with on+1
via (e0 , . . . , en )). Define a bianalytic bijection
v
ϕ : (kv× )n+2 → (kv× )n+2 ,
(x1 , x2 , x3 , y1 , . . . , yn−1 ) 7→ (x1 y1 · · · yn−1 , x2 y1 · · · yn−1 ,
x3 y1 · · · yn−1 ,
y1 · · · yn−1 ,
y2 · · · yn−1 , . . . , yn−1 );
n+1
note that the Jacobian determinant of ϕ is det ϕ0 (x1 , x2 , x3 , y1 , . . . , yn−1 ) = y13 y24 · · · yn−1
.
×
n+2
×
n
×
n+2
n+2
n
Since V ∩ (kv )
= ϕ(ov ∩ (kv )
) and µ(kv \R (kv ) ) = 0, by performing a change
of variables using ϕ and using the well-known fact ov |z|sv dµ(z) = (1 − qv−1 )/(1 − qv−1−s ),
Z
gr /
s−2
3s−1
−1 −n−1
ζm(n)⊗ov (s) = (1 − qv )
|x1 |s−1
· · · |yn−1 |v(n+1)s−1 dµ(x1 , . . . , yn−1 )
v |x3 |v |y1 |v
on+2
v
= 1/ (1 − qv−s )(1 − qv1−s )(1 − qv−3s )(1 − qv−4s ) · · · (1 − qv−(n+1)s ) .
The final claim follows by taking the product over all v ∈ Vk .
8
Remark. To the author’s knowledge, not a single example of a non-graded subobject
zeta function of a nilpotent Lie algebra of nilpotency class > 5 is known explicitly.
Integrals such as those in (3.2) are special cases of those associated with “toric data”
in [36, §3]. Hence, the author’s methods for manipulating such integrals as developed
in [36] apply directly without modification, as do the techniques explained below.
4. Explicit formulae
4.1. Theory: local zeta functions of Denef type
The following is a variation of the terminology employed in [35, §5.2]. As before, we assume
that k is a fixed number field. Suppose that we are given a collection Z = (ZK (s))K
of analytic functions of a complex variable s (each defined in some right half-plane)
indexed by p-adic fields K ⊃ k (up to k-isomorphism). We say that Z is of Denef
type if there exist a finite set S ⊂ Vk , k-varieties V1 , . . . , Vr , and rational functions
W1 , . . . , Wr ∈ Q(X, Y ) such that for all v ∈ Vk \ S and all finite extensions K/kv ,
ZK (s) =
r
X
−s
#V̄i (OK /PK ) · Wi (qK , qK
)
(4.1)
i=1
is an identity of analytic functions; here, we wrote V̄i = Vi ⊗o o/pv for a fixed but arbitrary
o-model Vi of Vi .
The following result formalises our discussion surrounding (1.1) from the introduction;
it summarises [15, §§2–3] (cf. [35, Thm 5.16]) and [43, Thm A].
Theorem 4.1. Let ZK (s) K be one of the following collections of local zeta functions
indexed by p-adic fields K ⊃ k (up to k-isomorphism).
gr6
6
/
(i) ZK (s) = ζA⊗
(s) or ZK (s) = ζA⊗
(s) (resp. ZK (s) = ζA⊗
(s) or ZK (s) =
o OK
o OK
o OK
gr /
ζA⊗
(s)),
where
A
is
an
o-form
of
a
finite-dimensional
(possibly
non-associative)
o OK
k-algebra as in §2.1 or §3.1, respectively.
(ii) ZK (s) = ζΩy(V⊗o OK ) (s), where Ω and V are as in §2.2.
irr
(iii) ZK (s) = ζG(O
(s), where G is an o-form of a unipotent algebraic group over k as
K)
in §2.3.
Then ZK (s) K is of Denef type.
f
The known proofs of Theorem 4.1 are constructive but impractical due to their reliance
on resolution of singularities. We note that the exclusion of finitely many primes implicit
in Theorem 4.1 is one of the main reasons for our focus on generic local zeta functions.
9
4.2. By-products of the computation of topological zeta functions
The computation of topological zeta functions is often considerably easier than that
of local ones. In [35, 36, 38], the author developed practical methods for computing
topological zeta functions associated with the local zeta functions in Theorem 4.1; these
methods are not algorithms because they may fail if certain non-degeneracy conditions
are violated. From now on, we will assume the validity of the following.
Assumption 4.2. In the setting of Theorem 4.1, the method from [36, §4] (resp. [38, §5.4])
for computing topological subalgebra and submodule zeta functions (resp. topological
representation zeta functions) succeeds.
Remark 4.3. The author is unaware of a useful intrinsic characterisation of those groups,
algebras, and modules such that Assumption 4.2 is satisfied. The local zeta functions in
Theorem 4.1 can be described in terms of p-adic integrals associated with a collection of
polynomials. A sufficient condition for the validity of Assumption 4.2 is “non-degeneracy”
of said collection of polynomials in the sense of [35, §4.2]; cf. [36, Lem. 5.7] and [38, §5.4.1].
The first stages of the methods for computing topological zeta functions associated
with the local zeta functions in Theorem 4.1, as described in [36, 38], come close to
constructing an explicit formula (4.1). In detail, using [35, Thm 4.10] (see [36, Thm 5.8]
and [38, Thm 5.9]), whenever they succeed, these methods derive a formula (4.1) such
that the following two assumptions are satisfied.
Assumption 4.4. The Vi in (4.1) are given as explicit subvarieties of algebraic tori
over k, defined by the vanishing of a finite number of Laurent polynomials and the
non-vanishing of a single Laurent polynomial.
Assumption 4.5. Up to multiplication by explicitly given rational functions of the form
(X − 1)a X b (for suitable a, b ∈ Z), each Wi in (4.1) is described explicitly in terms of
generating functions associated with half-open cones and convex polytopes.
We will clarify the deliberately vague formulation of Assumption 4.5 in §6.
In summary, whenever they apply, the methods for computing topological zeta functions
in [36, 38] fall short of “constructing” an explicit formula (4.1) only in the sense that the
Wi are characterised combinatorially instead of being explicitly given, say as fractions of
polynomials.
In the following sections, assuming the validity of Assumptions 4.2–4.5, we will develop
techniques for performing further computations with a formula of the form (4.1) with a
view towards solving the Uniformity Problem from the introduction in fortunate cases.
5. Counting rational points on subvarieties of tori
Assuming the validity of Assumption 4.4, this section is devoted to “computing” the
numbers #V̄i (OK /PK ) in (4.1). Using the inclusion-exclusion principle, we may reduce
10
to the case that the Vi are all closed subvarieties of algebraic tori over k. Note that
the non-constructive version of the Uniformity Problem from §1 has a positive solution
whenever each #V̄i (OK /PK ) is a polynomial in qK (after excluding finitely many places
of k). The following method is based on the heuristic observation that the latter condition
is often satisfied for examples of interest.
Setup. Let Tn := Spec(Z[X1±1 , . . . , Xn±1 ]) and, for a commutative ring R, write TnR :=
Tn ⊗ R. For a finite set S ⊂ Vk , let oS = {x ∈ k : x ∈ ov for all v ∈ Vk \ S} denote the
usual ring of S-integers of k. For f1 , . . . , fr ∈ oS [X1±1 , . . . , Xn±1 ], define
(f1 , . . . , fr )nS := Spec(oS [X1±1 , . . . , Xn±1 ]/hf1 , . . . , fr i) ⊂ TnoS .
For v ∈ Vk \ S and a finite extension K of o/pv , let |f1 , . . . , fr |nK denote the number of
K-rational points of (f1 , . . . , fr )nS .
Objective: symbolic enumeration. From now on, let f1 , . . . , fr ∈ oS [X1±1 , . . . , Xn±1 ] be
given as above. Our goal in the following is to symbolically “compute” the numbers
|f1 , . . . , fr |nK as a function of K. More precisely, the procedure described below constructs
a polynomial, H(X, c1 , . . . , c` ) say, over Z such that, after possibly enlarging S, for all
v ∈ Vk \ S and all finite extensions K of o/pv ,
(f1 , . . . , fr )nK = H(|K|, #U1 (K), . . . , #U` (K)),
where each Ui is an explicitly given closed subscheme of some TnoSi . We could of course
simply take H = c1 and U1 = (f1 , . . . , fr )nS but we seek to do better. Indeed, in many
cases of interest, H can be taken to be a polynomial in X only. In the following, we
describe a method which has proven to be quite useful for handling such cases.
Dimension 6 1. We first describe two base cases of our method. Namely, if n = 0,
then, after possibly enlarging S, (f1 , . . . , fr )nS is either ∅ or T0oS = Spec(oS ) depending
on whether some fi 6= 0 or not; thus, |f1 , . . . , fr |nK ∈ {0, 1} for K as above.
Secondly, if n = 1, then we use the Euclidean algorithm over k (thus possibly enlarging S) to compute a single square-free polynomial f ∈ oS [X1 ] such that (f1 , . . . , fr )1S =
(f )1S . If f splits completely over k, then, after possibly enlarging S once again, |f |1K =
deg(f ) for all K as above. If f does not split completely over k, then we introduce a new
variable, cf say, corresponding to the number of solutions of f = 0 in K× .
Simplification. It is often useful to “simplify” the given Laurent polynomials f1 , . . . , fr ;
while this step was sketched in [36, §6.6], here we provide some further details. As
before, the set S may need to be enlarged at various points in the following. First, we
discard any zero polynomials among the fi . We then clear denominators so that each
fi ∈ k[X1 , . . . , Xn ] is an actual (not just Laurent) polynomial. Next, we replace each fi
by its square-free part in k[X1 , . . . , Xn ]. For each pair (i, j) of distinct indices, we then
compute the (square-free part of the) remainder, r say, of fi after multivariate polynomial
11
division by fj with respect to some term order (see e.g. [1, §1.5]). If r consists of fewer
terms than fi , we replace fi by r. Next, for each pair (i, j) as above and each term
t
ti of fi and tj of fj , we are free to replace fi by (the square-free part of) gj fi − tgi fj ,
where g = gcd(ti , tj ) (computed over k), which we again do whenever it reduces the
total number of terms. After finitely many iterations of the above steps, f1 , . . . , fr will
stabilise at which point we conclude the simplification step.
We next describe two procedures which, if applicable, allow us to express |f1 , . . . , fr |nK
in terms of the numbers of rational points of subschemes of lower-dimensional tori. We
then recursively attempt to solve the symbolic enumeration problem from above for these.
Reduction of dimension I: torus factors. As explained in [36, §6.3], using the natural
action of GLn (Z) on Tn , a Smith normal form computation allows us to effectively
construct g1 , . . . , gr ∈ oS [X1±1 , . . . , Xd±1 ] and an explicit isomorphism (f1 , . . . , fr )nS ≈
(g1 , . . . , gr )dS ×oS Tn−d
oS , where d is the dimension of the Newton polytope of f1 · · · fr . It
follows that for all K as above, |f1 , . . . , fr |nK = |g1 , . . . , gr |dK · (|K| − 1)n−d . In the following,
we may thus assume that n = d.
Reduction of dimension II: solving for variables. Whenever it is applicable, the following lemma allows us to replace the problem of symbolically computing |f1 , . . . , fr |nK
by four instances of the same problem in dimension n − 1.
±1
Lemma 5.1. Let F ⊂ oS [X1±1 , . . . , Xn−1
]. Further let f = u − wXn for non-zero
±1
±1
u, w ∈ oS [X1 , . . . , Xn−1 ]. Then for all v ∈ Vk \ S and all finite extensions K of o/pv ,
|F, f |nK = |F |n−1
− |F, u|n−1
− |F, w|n−1
+ |K| · |F, u, w|Kn−1 .
K
K
K
Proof. Projection onto the first n − 1 coordinates induces an isomorphism of oS -schemes
n−1
n
n
\ (F, uw)n−1
×oS
(F, f )nS \ (F, f, w)nS ≈ (F )n−1
S
S . As (F, f, w)S = (F, u, w)S ≈ (F, u, w)S
1
ToS , the claim follows since for all v ∈ Vk \ S and all finite extensions K of o/pv ,
n−1
|((F )n−1
\ (F, uw)n−1
− |F, u|n−1
− |F, w|n−1
+ |F, u, w|n−1
K
K
K .
S
S )(K)| = |F |K
Remark 5.2. The evident analogue of Lemma 5.1 for Euler characteristics of closed
subvarieties of algebraic tori over k has already been used in the author’s software package
Zeta [39] for computing topological zeta functions. However, only the special case that
n−1
w ∈ o×
= (F, u, w)n−1
= ∅) was spelled out explicitly in [36, §6.6].
S (so that (F, w)S
S
Final case. Finally, if none of the above techniques for computing or decomposing
(f1 , . . . , fr )nS applies, then we introduce a new variable corresponding to |f1 , . . . , fr |nK . In
order to avoid this step whenever possibly, we first attempt to apply the above steps
(including all possible applications of Lemma 5.1) without ever invoking this final case.
12
6. Local zeta functions as sums of rational functions
Suppose that Assumptions 4.2–4.5 are satisfied. Our first task in this section is to
rewrite (4.1) as a sum of explicitly given rational functions. With the method from §5 at
our disposal, this problem reduces to finding such an expression for each Wi . We will see
that Barvinok’s algorithm from convex geometry solves this problem. Our second task
then concludes the computation of the generic local zeta functions in Theorem 4.1; it is
concerned with adding a potentially large number of multivariate rational functions. We
describe a method aimed towards improving the practicality of this step which, while
mathematically trivial, often vastly dominates the run-time of our computations.
6.1. Barvinok’s algorithm: generating functions and substitutions
Let P ⊂ Rn>0 be a rational polyhedron and let λ = (λ1 , . . . , λn ) be algebraically
indeP
pendent over Q. It is well-known that the generating function |P| := α∈P∩Zn λα is
rational in the sense that within the field of fractions of Q[[λ1 , . . . , λn ]], it belongs to
Q(λ1 , . . . , λn ). The standard proof of this fact (see e.g. [4, Ch. 13]) proceeds by reducing
to the case that P is a cone, in which case an explicit formula for |P| can be derived from
a triangulation of P via the inclusion-exclusion principle. This strategy for computing |P|
is, however, of rather limited practical use.
A far more sophisticated approach is given by “Barvinok’s algorithm”; see [5, 7].
Barvinok’s algorithm computes |P| for each (suitably encoded) rational polyhedron
P ⊂ Rn>0 as a sum of rational functions of the form cλα0 /((1 − λα1 ) · · · (1 − λαn )) for
suitable α0 , . . . , αn ∈ Zn and c ∈ Q. For a fixed ambient dimension n > 1, his algorithm
runs in polynomial time so that |P| is computed as a short sum of short rational functions
in a precise technical sense. Beyond its theoretical strength, Barvinok’s algorithm is also
powerful in practice as demonstrated by the software implementation LattE [3].
In the setting of Assumption 4.5, we are not primarily interested in generating functions
associated with polyhedra themselves but in rational functions derived from such generating functions via monomial substitutions. In detail, let ξ = (ξ1 , . . . , ξm ) be algebraically
independent over Q and let σ1 , . . . , σn ∈ Zm . Suppose that P ⊂ Rn>0 is a rational
polyhedron such that W := |P|(ξ σ1 , . . . , ξ σn ) is well-defined on the level of rational
functions. In principle, we could compute W by first using the output of Barvinok’s
algorithm in order to write |P| in lowest terms, followed by an application of the given
substitution. This method is, however, often impractical due to the computational cost
of (multivariate) rational function arithmetic.
A theoretically favourable and also practical alternative is developed in [6, §2] (cf. [5, §5])
There, a polynomial time algorithm is described which takes as input a short representation
of |P| (as, in particular, provided by Barvinok’s algorithm) and constructs a similar
short representation for W . The important point to note here is that while we assumed
the substitution λi 7→ ξ σi to be valid for |P| itself, it may be undefined for some of the
summands in the expression provided by Barvinok’s algorithm.
13
6.2. Computing the Wi in (4.1)
We may now clarify the vague formulation of Assumption 4.5. Namely, up to a factor
(X − 1)a X b , the Wi in (4.1) are obtained by applying suitable monomial substitutions (see
[35, Rem. 4.12] and [38, Thm 5.5]) to rational functions of the form Z C0 ,P1 ,...,Pm (ξ0 , . . . , ξm )
from [35, Def. 3.6]. The latter functions can, by their definitions, be written as sums of
rational functions obtained by applying suitable monomial substitutions to generating
functions enumerating lattice points inside rational half-open cones; as explained in
[36, §8.4], we may replace these half-open cones by rational polyhedra. We may thus use
Barvinok’s algorithm as well as the techniques for efficient monomial substitutions from
[6, §2] in order to write each Wi as a sum of bivariate rational functions of the form
f (X, Y )/ (1 − X a1 Y b1 ) · · · (1 − X am Y bm )
(6.1)
for suitable integers ai , bi ∈ Z, m > 0, and f (X, Y ) ∈ Q[X, Y ].
6.3. Final summation
In the following, we allow f (X, Y ) in (6.1) to be an element of Q[X, Y, c1 , c2 , . . . ]. By
taking into account the polynomials obtained using §5, at this point, we may thus assume
that we constructed a finite sum of expressions (6.1) such that, after excluding finitely
many places of k, the local zeta functions in Theorem 4.1 are obtained by specialising
−s
X 7→ qK , Y 7→ qK
, and ci 7→ #Ui (OK /PK ) for certain explicit subschemes Ui of tori
over o (or over oS ). All that remains to be done in order to recover the local zeta functions
of interest is to write the given sum of expressions (6.1) in lowest terms.
While our intended applications of Barvinok’s algorithm lie well within the practical
scope of LattE [3], it will often be infeasible to pass the rational functions (6.1) to a
computer algebra system in order carry out the final summation. In addition to the
sheer number of rational functions to be considered, a key problem is due to the fact
that the number of distinct pairs (ai , bi ) arising from summands (6.1) often obscures the
relatively simple shape of the final sum (i.e. the local zeta function to be computed).
This is consistent with the well-known observation (see e.g. [12, §2.3]) that few candidate
poles of local zeta functions as provided by explicit formulae (4.1) survive cancellation.
In order to carry out the final summation, we proceed in two stages. First, we use
an idea due to Woodward [49, §2.5] and add and simplify those summands (6.1) such
that distinguished pairs 1 − X c Y d occur in their written denominators; our hope here is
that some rays (ai , bi ) will be removed via cancellations. While this step is not essential,
it might improve the performance and memory requirements of the final stage. Here,
we first construct a common denominator of all the remaining rational functions (6.1).
We then compute the final result by summing the (6.1) rewritten over our common
denominator, followed by one final division. In addition to being trivially parallelisable,
by only adding numerators, we largely avoid costly rational function arithmetic.
14
6.4. Implementation issues
The method for computing generic local subobject or representation zeta functions
described above has been implemented (for k = Q) by the author as part of his package
Zeta [39] for Sage [44]. The program LattE [3] (which implements Barvinok’s algorithm)
plays an indispensable role. Moreover, the computer algebra system Singular [25] features
essentially in the initial stages of our method (as described in [36, 38]).
The author’s implementation is primarily designed to find instances of positive solutions
to the Uniformity Problem; its functionality and practicality are both quite restricted in
non-uniform cases. Furthermore, the author’s method supplements Woodward’s approach
[49] for computing local (subalgebra and ideal) zeta functions as well as various ad
hoc computations carried out by others without replacing them. In particular, various
examples of local zeta functions computed by Woodward cannot be reproduced using
the present method. In addition to the theoretical limitations of the techniques from
[35, 36, 38], this is also partially due to practical obstructions: while some computations of
topological zeta functions in [35, 36, 38] were already fairly involved, the present method
is orders of magnitude more demanding.
7. Interlude: reduced representation zeta functions
Reduced zeta functions arising from the enumeration of subalgebras and ideals were
introduced by Evseev [20]. They constitute a limit “p → 1” of suitable local zeta functions
distinct from but related to the topological zeta functions of Denef and Loeser [13] (which
were later adapted to the case of subobject zeta functions by du Sautoy and Loeser [16]).
Informally, Evseev’s definition can be summarised as follows in our setting. Let A be an
6
o-form of a k-algebra as in §2.1. For each v ∈ Vk , we may regard ζA⊗
(s) as a (rational)
o ov
−s
formal power series in Y = qv . The reduced subalgebra zeta function of A (an invariant
of A ⊗o C, in fact) is obtained by taking a limit “qv → 1” applied to the coefficients
6
of ζA⊗
(s) as a series in Y . The rigorous definition of reduced zeta function in [20]
o ov
involves the motivic subobject zeta functions introduced by du Sautoy and Loeser [16].
In this section, we show that “reduced representation zeta functions” associated with
unipotent groups are always identically 1. In addition to imposing restrictions on the
shapes of generic local representation zeta functions of such groups, this fact provides
a simple necessary condition for the correctness of explicit calculations of local zeta
functions such as those documented below.
We begin with a variation of a result from [37]. Let V be a separated k-scheme of
finite type. For any embedding k ⊂ C, the topological Euler characteristic χ(V (C)) is
defined and well-known to be independent of the embedding; cf. [29].
Lemma 7.1. Let V1 , . . . , Vr be separated o-schemes of finite type and W1 , . . . , Wr ∈
Q(X, Y1 . . . . , Ym ). Suppose that for almost all v ∈ Vk and all integers f > 0, each Wi is
regular at (qvf , Y1 , . . . , Ym ). Let P ⊂ Vk have natural density 1 and suppose that
r
X
#Vi (o/pv ) · Wi (qv , Y1 , . . . , Ym ) = 0
i=1
15
for all v ∈ P . Then
r
P
χ(Vi (C)) · Wi (1, Y1 , . . . , Ym ) = 0.
i=1
Proof. Using [40, Ch. 4], in the setting of [37, Thm 3.7], we may assume that α(1ΓS ) =
χ(V (C)). The claim is now an immediate consequence of [37, Thm 3.2] and its proof.
Remark 7.2. Given a formula (4.1) for local subalgebra or ideal zeta functions such
that the regularity conditions
P in Lemma 7.1 are satisfied, we may read off the associated
reduced zeta function as ri=1 χ(Vi (C)) · Wi (1, Y ) without using motivic zeta functions.
The following is a consequence of the explicit formulae in [19].
Theorem 7.3. Let G be a unipotent algebraic group over k. Let G be an o-form of G as
an affine group scheme of finite type. There are separated o-schemes U1 , . . . , U` of finite
type and rational functions W1 , . . . , W` ∈ Q(X, Y ) such that
irr (s) − 1 =
(i) for almost all v ∈ Vk , ζG(o
v)
f
P̀
#Ui (o/pv ) · Wi (qv , qv−s ),
i=1
(ii) each Wi is regular at each point (q, Y ) for q > 1, and
(iii) Wi (1, Y ) = 0 for i = 1, . . . , `.
Proof. In the setting of [19, Prop. 3.4], the rational numbers Aj and Bj can actually be
assumed to be integers; this follows e.g. by taking square roots of principal minors and
rewriting [19, (2.3)] as in [38, (4.3)]. Next, using the same notation as in [19, Prop. 3.4],
|Mi | 6 |Ui | + 1 whence the claim follows easily from [19, Rem. 3.6].
Remark 7.4. Theorem 7.3 refines the simple observation that for almost all v ∈ Vk ,
f
irr
the coefficients of ζG(o
(s) − 1 as a series in qv−s are non-negative integers divisible by
v)
qv − 1, a simple consequence of the Kirillov orbit method. (Indeed, (o/pv )× acts freely
on non-trivial characters while preserving the two types of radicals in [43, Thm 2.6].)
By combining Lemma 7.1 and Theorem 7.3, we obtain the following.
Corollary 7.5. Let G be as in Theorem 7.3. Let V1 , . . . , Vr be separated o-schemes of
finite type and let W1 , . . . , Wr ∈ Q(X, Y ) such that for almost all v ∈ Vk ,
irr
ζG(o
(s) =
v)
f
r
X
#Vi (o/pv ) · Wi (qv , qv−s ).
i=1
If each Wi is regular at (q, Y ) for each q > 1, then
r
P
χ(Vi (C)) · Wi (1, Y ) = 1.
i=1
Corollary 7.6. Let G be as in Theorem 7.3. Let W (X, Y ) ∈ Q(X, Y ) such that
(i) W (X, Y ) can be written over a denominator which is a product of non-zero factors
of the form 1 − X a Y b for integers a > 0 and b > 1 and
16
irr (s) = W (q , q −s ) for almost all v ∈ V .
(ii) ζG(o
k
v v
v)
f
Then W (1, Y ) = 1.
The assumptions in Corollary 7.6 are satisfied for many examples of interest; see
Table 1. In fact, even the following much stronger assumptions are often satisfied.
Corollary 7.7. Let G be as in Theorem 7.3. Suppose that there are integers ai > 0,
bi > 1, and εi ∈ {±1} for i = 1, . . . , m such that for almost all v ∈ Vk ,
irr
ζG(o
(s) =
v)
f
m
Y
(1 − qvai −bi s )εi .
i=1
Then
m
P
εi = 0 and the multisets {{bi : εi = 1}} and {{bi : εi = −1}} coincide.
i=1
Q
bi εi
Proof. Corollary 7.6 shows that 1 = m
i=1 (1 −
PYm ) . By considering the vanishing
order of this function in Y at 1, we see that i=1 εi = 0. Let c = max(bi : εi = 1)
and d = max(bi : εi = −1). If ξ ∈ C is a primitive cth root of unity, then 1 − ξ bi = 0
for some i with εi = −1 whence c 6 bi 6 d; dually, d 6 c and the final claim follows
by induction.
Remark 7.8. The above results carry over verbatim to the case of representation zeta
functions of “principal congruence subgroups” Gm (ov ) := exp(pm
v g ⊗o ov ) attached to
an o-form of a perfect Lie k-algebra in [2]. For example, by [2, Thm E], the ordinary
representation zeta function of SL13 (Zp ) (p 6= 3) is W (p, p−s ) for
W (X, Y ) =
(X 2 Y 2 + XY 2 + Y 3 + X 2 + XY + Y ) × (X 2 − Y )(X − Y )X 3
(1 − X 2 Y 3 )(1 − XY 2 )
and indeed W (1, Y ) = 1.
8. Applications I: representation zeta functions of unipotent
groups
In this and the following two sections, we record explicit examples of generic local zeta
functions of groups, algebras, and modules of interest which were computed using the
method developed in the present article and its implementation Zeta [39]. The explicit
formulae given below, as well as others, are also included with Zeta.
It is well-known that, up to isomorphism, unipotent algebraic groups over k correspond
1–1 to finite-dimensional nilpotent Lie k-algebras, see [11, Ch. IV]. Nilpotent Lie algebras
of dimension at most 6 over any field of characteristic zero were first classified by
Morozov [32]; various alternative versions of this classification have been obtained.
As one of the main applications of the techniques developed in the present article, for
an arbitrary number field k, we can compute the generic local (twist) representation
17
zeta functions associated with all unipotent algebraic groups of dimension at most 6
over k. The results of these computations are documented in Table 1 (p. 20). The
structure of Table 1 mimics the list of associated topological representation zeta functions
in [38, Table 1]. In detail, the first column lists the relevant Lie algebras using de Graaf’s
notation [10]; an algebra Ld,i has dimension d. For each Lie algebra g, we choose an
o-form G of the unipotent algebraic group over k associated with g. The second column
in Table 1 contains formulae for the representation zeta functions of the groups G(ov )
which are valid for almost all v ∈ Vk (depending on G). Note that Corollary 7.7 applies
to the majority of examples in Table 1. As we previously documented in [38, §6], generic
local representation zeta functions associated with various Lie algebras in Table 1 were
previously known (but sometimes only recorded for k = Q), as indicated in the third
column. For the convenience of the reader, the more detailed references to the literature
from [38, Tab. 1] are reproduced in Remark 8.3.
Remark 8.1 (From Q to k). Apart from the four infinite families (see the following
remark), all Lie algebras in Table 1 are defined over Q. By the invariance of (4.1) under
local base extensions (Theorem 4.1), it thus suffices to compute associated generic local
representation zeta functions for k = Q.
Remark 8.2 (Computations for infinite families). The method for computing generic
local zeta functions developed in this article takes as input a global object such as a
nilpotent Lie k-algebra. In order to carry out computations for the four infinite families
L6,19 (a), L6,21 (a), L6,22 (a), and L6,24 (a) in Table 1, additional arguments are required.
First, as explained in [10], we are free to multiply the parameters a from above by
elements of (k × )2 6 k × without changing the k-isomorphism type of the Lie algebra, g(a)
say, in question. We may thus assume that 0 6= a ∈ o in the following. The definition of
g(a) in [10] then provides us with a canonical o-form, g(a) say, of g(a) which is in fact
defined over Z[a]. Let Ga be an o-form of the unipotent algebraic group over k associated
with g(a). As explained in [43, §2], the structure constants of g(a) (with respect to its
f
defining basis from [10]) give rise to a formula for ζGirra (ov ) (s) in terms of certain explicit
o-defined p-adic integrals (see [43, Cor. 2.11]); this formula is valid for almost all v ∈ Vk .
It is an elementary exercise to verify that if g = L6,19 (a) or g = L6,21 (a), then
f
the polynomials featuring in the aforementioned integral formulae for ζGirra (ov ) (s) are all
monomials in a and the variables Y1 , . . . , Yd (in the notation of [43, §2.2] and up to signs).
f
It follows that up to excluding finitely many v ∈ Vk , ζGirra (ov ) (s) does not depend on a. We
may therefore simply carry out our calculation for k = Q and a = 1, say.
Let g(a) be L6,22 (a) or L6,24 (a). Another simple calculation reveals that (again up
to signs) a single non-monomial polynomial occurs in the associated integral formulae
from above, namely Y12 − aY22 . For any fixed a, by applying the procedure from [38, §5.4]
as well as the steps described in the present article, we produce a rational function
f
Wa (X, Y, Z) such that ζGirra (ov ) (s) = Wa (qv , qv−s , ca (v)) for almost all v ∈ Vk , where ca (v)
denotes the number of roots of X 2 − a in o/pv ; it is well-known that if a 6∈ (k × )2 , then for
almost all v ∈ Vk , ca (v) = 0 or ca (v) = 2 according to whether pv remains inert or splits
√
in k( a), respectively. The critical observation (which follows easily from [38, §5.4]) is
18
that W := Wa is independent of a and also of k. We may thus compute W explicitly by
e.g. taking k = Q and a = 2.
Remark 8.3. Explicit references for the known instances of generic local representation
zeta functions in Table 1 are as follows (cf. [38, Tab. 1]):
algebra
L3,2
L5,4
L5,7
L5,9
L6,10
L6,19 (0)
L6,22 (0)
L6,25
reference
[34, Thm 5])
B4 [41, Ex. 6.3])
M4 [21, (4.2.24)]
F3,2 [21, Tab. 5.2])
G6,12 [21, Tab. 5.2])
G6,7 [21, Tab. 5.2])
[41, Ex. 6.5]
M4,3 [21, (5.3.7)]
algebra
L4,3
L5,5
L5,8
reference
M3 [21, (4.2.24)]
G5,3 [21, Tab. 5.2])
M3,3 [21, (5.3.7)]
= G3 [41, Ex. 6.2])
L6,18
M5 [21, (4.2.24)]
L6,19 (1)
G6,14 [21, Tab. 5.2]
L6,22 (a) (a ∈ k × \(k × )2 )[22]
L6,26
F1,1 [43, Thm B].
The author would like to emphasise that all the formulae in Table 1 were obtained
using the method developed here. In particular, our computations provide independent
confirmation of the aforementioned (sometimes computer-assisted but predominantly
manual and ad hoc) calculations found in the literature.
For an example in dimension > 6, recall from §2.3 that Un denotes the group scheme
of upper unitriangular n × n-matrices. Using the notation from [10] as in Table 1, U3 ⊗Q
(the Heisenberg group) and U4 ⊗Q are the unipotent algebraic groups over Q associated
with the Lie algebras L3,2 and L6,19 (1), respectively. The following result obtained using
the method from the present article illustrates that the simple shapes of the corresponding
local representation zeta functions in Table 1 may mislead.
Theorem 8.4. For almost all primes p and all finite extensions K/Qp ,
−s
irr
ζU
(s) = W (qK , qK
),
5 (OK )
f
where
W = X 10 Y 10 − X 9 Y 9 − 2X 9 Y 8 + X 9 Y 7 + X 8 Y 8 − X 7 Y 7 − 2X 7 Y 6 + X 7 Y 5
+ 6X 6 Y 6 − 4X 5 Y 6 − 4X 5 Y 4 + 6X 4 Y 4 + X 3 Y 5 − 2X 3 Y 4 − X 3 Y 3
+ X 2 Y 2 + XY 3 − 2XY 2 − XY + 1 × 1 − Y 3 × 1 − Y
/ (1 − X 6 Y 4 )(1 − X 3 Y 3 )(1 − XY 3 )(1 − X 2 Y 2 )(1 − X 2 Y )2 .
The topological representation zeta function of U6 cannot be computed using [38].
Consequently, the corresponding local zeta functions cannot be computed using the
method developed here.
Observe that the numerator of each W (X, Y ) in Table 1 is divisible by a polynomial
of the form 1 − Y e . Experimental evidence provided by these examples and those in Zeta
suggests that the following p-adic version of [38, Qu. 7.4] might have a positive answer.
19
Question 8.5. Let G be an o-form of a non-abelian unipotent algebraic group over k.
f
irr
Does the meromorphic continuation of ζG(o
(s) always vanish at zero for almost all
v)
v ∈ Vk ?
Remark. By [24, Cor. 2], if p is odd, then the meromorphic continuation of the ordinary
(= non-twisted) representation zeta function of a compact FAb p-adic analytic group
vanishes at −2.
Lie algebra
abelian
L3,2
L4,3
L5,4
L5,5
L5,6
L5,7
L5,8
L5,9
L6,10
L6,11
L6,12
L6,13
L6,14
L6,15
L6,16
L6,17
L6,18
L6,19 (0)
L6,19 (a) (a ∈ k × )
L6,20
L6,21 (0)
L6,21 (a) (a ∈ k × )
L6,22 (0)
L6,22 (a)
(a ∈ k × \(k × )2 )1
L6,23
L6,24 (0)
L6,24 (a)
(a ∈ k × )
L6,25
L6,26
irr
W (X, Y ) s.t. ζG(o
(s) = W (qv , qv−s ) for almost all v ∈ Vk
v)
1
(1 − Y )/(1 − XY )
(1 − Y )2 /(1 − XY )2
(1 − Y 2 )/(1 − XY 2 )
(1 − XY 2 )(1 − Y )/ (1 − X 2 Y 2 )(1 − XY )
(1 − X 2 Y 2 )(1 − Y )2 / (1 − X 3 Y 2 )(1 − XY )2
(1 − Y )2 / (1 − X 2 Y )(1 − XY )
(1 − Y )/(1 − X 2 Y )
(1 − Y )2 / (1 − X 2 Y )(1 − XY )
(1 − Y 2 )(1 − Y )/ (1 − XY 2 )(1 − XY )
f
(−X 3 Y 4 +X 3 Y 3 −2X 2 Y 3 +3X 2 Y 2 −3XY 2 +2XY −Y +1)(1−Y )
(1−X 4 Y 3 )(1−X 2 Y 2 )
2 2
2
3 2
2
known
X
X
X
X
X
X
X
X
X
(1 − X Y )(1 − Y ) / (1 − X Y )(1 − XY )
(X 4 Y 6 +X 4 Y 5 −X 3 Y 4 −2X 2 Y 3 −XY 2 +Y +1)(1−Y )2
(1−X 3 Y 3 )(1−X 2 Y 2 )(1−XY 2 )(1−XY )
(X 4 Y 6 −X 4 Y 4 +X 3 Y 5 −2X 2 Y 3 +XY −Y 2 +1)(1−Y )2
(1−X 3 Y 3 )(1−X 3 Y 2 )(1−XY 2 )(1−XY )
(−X 5 Y 4 −X 4 Y 3 +X 3 Y 2 −X 2 Y 2 +XY +1)(1−Y )2
(1−X 5 Y 3 )(1−X 3 Y 2 )(1−XY )
2
2
2
2
(1 − Y )(1 − Y ) / (1 − X Y )(1 − XY )(1 − XY )
(1 − X 3 Y 2 )(1 − Y )2 / (1 − X 4 Y 2 )(1 − X 2 Y )(1 − XY )
(1 − Y )2 / (1 − X 3 Y )(1 − XY )
(1 − Y )2 / (1 − X 2 Y )(1 − XY )
(1 − Y 2 )(1 − Y )/ (1 − X 2 Y )(1 − XY 2 )
(1 − XY 2 )(1 − Y )/ (1 − X 2 Y )(1 − X 2 Y 2 )
(1 − Y )2 /(1 − X 2 Y )2
(1 − X 2 Y 2 )(1 − Y )2 / (1 − X 3 Y 2 )(1 − X 2 Y )(1 − XY )
(1 − X 2 Y 2 )(1 −√
Y )/ (1 − X 3 Y 2 )(1 − XY )
if pv splits in k( √
a): (1 − Y )2 /(1 − XY )2
if pv is inert in k( a): (1 − Y 2 )/(1 − X 2 Y 2)
(1 − X 3 Y 2 )(1 − Y )/ (1 − X 4 Y 2 )(1 − X 2 Y )
(X 4 Y 4 −X 4 Y 3 +X 3 Y 3 −2X 2 Y 2 +XY −Y +1)(1−Y )
(1−X 3 Y 2 )2 (1−XY )
2
√
+2XY −2Y +1)(1−Y )
× 2
if a ∈ (k ) or pv splits in k( a): (−XY(1−X
3 Y 2 )(1−XY )
√
(1−XY 2 )(1−Y )
× 2
if a 6∈ (k ) and pv is inert in k( a): (1−X 3 Y 2 )(1−XY )
2
2
(1 − XY )(1 − Y )/(1 − X Y )
(1 − Y )/(1 − X 3 Y )
X
X
X(a = 1)
X
X
X
X
X
Table 1: Generic local representation zeta functions associated with all indecomposable
unipotent algebraic groups of dimension at most six over a number field
20
9. Applications II: classical subobject zeta functions
9.1. Subalgebras: gl2 (Q)
The first computations of the subalgebra zeta functions of sl2 (Zp ) are due, independently,
to du Sautoy [14] (for p 6= 2, relying heavily on [28]) and White [48]. These zeta functions
have later been confirmed by different means in [17], [30, §4.2], and [35, §7.1] (for p 6= 2).
Up until now, sl2 (Q) has remained the sole example of an insoluble Lie Q-algebra whose
generic local subalgebra zeta functions have been computed. Using the method developed
in the present article, we obtain the following.
Theorem 9.1. For almost all primes p and all finite extensions K/Qp ,
ζgl6
2 (OK )
−s
(s) = W (qK , qK
),
where
W (X, Y ) = −X 8 Y 10 − X 8 Y 9 − X 7 Y 9 − 2X 7 Y 8 + X 7 Y 7 − X 6 Y 8
− X 6 Y 7 + 2X 6 Y 6 − 2X 5 Y 7 + 2X 5 Y 5 − 3X 4 Y 6 + 3X 4 Y 4
− 2X 3 Y 5 + 2X 3 Y 3 − 2X 2 Y 4 + X 2 Y 3 + X 2 Y 2 − XY 3
+ 2XY 2 + XY + Y + 1 / (1 − X 7 Y 6 )(1 − X 3 Y 3 )(1 − X 2 Y 2 )2 (1 − Y ) .
The topological subalgebra zeta function ζgl6 (Q),top (s) = (27s − 14)/(6(6s − 7)(s − 1)3 s)
2
of gl2 (Q) was first recorded in [35, §7.3] (relying on techniques from [36]); the result
given there is consistent with Theorem 9.1. Theorem 9.1 is particularly interesting since
the simple shape of ζgl6 (Q),top (s) might seem indicative of a local zeta function which is a
2
a−bs
product of “cyclotomic factors” 1 − qK
or their inverses, which is in fact not the case.
We note that the computations underpinning Theorem 9.1 used that gl2 (R) ≈ sl2 (R)⊕R
for any commutative ring R in which 2 is invertible; here we regarded R as an abelian
Lie R-algebra. Theorem 9.1 therefore also illustrates the potentially wild effect of direct
sums on subalgebra zeta functions; in contrast, [18] contains examples of subalgebra and
ideal zeta functions associated with nilpotent Lie algebras which are very well-behaved
under this operation.
The rational function W (X, Y ) in Theorem 9.1 satisfies the functional equation
W (X −1 , Y −1 ) = X 6 Y 4 W (X, Y )
predicted by [46, Thm A] (cf. [37, §5]). Moreover, the
reduced subalgebra zeta function of
gl2 (Q) is W (1, Y ) = (1 − Y 3 )/ (1 − Y )3 (1 − Y 2 )2 , as predicted by [20, Thm 3.3] (using
the fact that the reduced subalgebra zeta function of sl2 (Z) is (1−Y 3 )/ (1−Y )2 (1−Y 2 )2
(by [20, Prop. 4.1])).
1
For a ∈ (k× )2 , L6,22 (a) ≈ L23,2 decomposes.
21
9.2. Subalgebras: k[T ]/T n for n 6 4
Most examples of local subalgebra zeta functions in the literature are concerned with
(often nilpotent) Lie algebras. An important exception is given by the subalgebra zeta
functions of Znp endowed with component-wise multiplication; explicit formulae for these
zeta functions are known for n 6 3 (see [33]). In the following, we consider another
natural family of associative, commutative algebras, k[T ]/T n , for n 6 4.
Due to the simplicity of the associated “cone integrals” as in [15], the formulae for
n = 2, 3 recorded in the following can be obtained by hand with little difficulty. Using a
substantially more involved computation, the techniques developed in the present article
also allow us to consider the case n = 4. For n = 5, the author’s techniques for computing
topological subalgebra zeta functions do not apply, i.e. Assumption 4.2 is violated.
Theorem 9.2. For almost all primes p and all finite extensions K/Qp , writing q = qK ,
6
ζO
2 (s) =
K [T ]/T
1 − q −2s
,
(1 − q −s )2 (1 − q 1−2s )
6
−s
ζO
)×
3 (s) = FQ[T ]/T 3 (q, q
K [T ]/T
1 − q 2−4s
, and
(1 − q 4−5s )(1 − q 2−3s )2 (1 − q 1−2s )(1 − q −s )
6
−s
ζO
)/ (1 − q 13−13s )(1 − q 9−9s )(1 − q 8−8s )(1 − q 6−6s )2
4 (s) = FQ[T ]/T 4 (q, q
K [T ]/T
× (1 − q 5−6s )(1 − q 5−5s )(1 − q 3−4s )(1 − q −s ) ,
where FQ[T ]/T 3 = −X 4 Y 7 − X 4 Y 6 − X 3 Y 5 + X 3 Y 4 − X 2 Y 4 + X 2 Y 3 − XY 3 + XY 2 + Y + 1
and FQ[T ]/T 4 = 1 + · · · − X 49 Y 54 ∈ Q[X, Y ] is given in Appendix A.
The topological subalgebra zeta function of Q[T ]/T 4 can be found in [36, §9.2]. As
in §9.1, the zeta functions in Theorem 9.2 satisfy the functional equations predicted by
[46, Thm A] and the associated reduced subalgebra zeta functions coincide with those
computed using [20]; while Evseev only considered reduced zeta functions of Lie algebras,
his reasoning also applies to more general, possibly non-associative, algebras. For example,
using Theorem 9.2, after considerable cancellation, we find the reduced subalgebra zeta
function of Q[T ]/T 4 to be (Y 6 + Y 4 + 2Y 3 + Y 2 + 1)/ (1 − Y 6 )(1 − Y 2 )(1 − Y )2 , as
predicted by Evseev’s results.
9.3. Subalgebras: soluble, non-nilpotent Lie algebras
Taylor [45, Ch. 6] computed local subalgebra zeta functions associated with soluble,
non-nilpotent Lie algebras of the form k d o k (semidirect sum) for d = 2, 3, where k d
and k are regarded as abelian Lie algebras. In particular, he (implicitly) computed the
subalgebra zeta function of the Lie algebra tr2 (Zp ) of upper triangular 2 × 2-matrices
over Zp (see [18, §3.4.2]). Klopsch and Voll [30] computed subalgebra zeta functions of
arbitrary 3-dimensional Lie Zp -algebras in terms of Igusa’s local zeta functions attached to
associated quadratic forms. Regarding the enumeration of ideals of soluble, non-nilpotent
Lie algebras, Woodward [50] computed local ideal zeta functions of trd (Zp ) and certain
combinatorially defined quotients of these algebras.
22
Since, to the author’s knowledge, no examples of generic local subalgebra zeta functions
associated with soluble, non-nilpotent Lie algebras of dimension 4 have been recorded in
the literature, we now include some examples.
Theorem 9.3. Let Mi denote an arbitrary but fixed Z-form of the soluble Lie Q-algebra
M i of dimension 4 from [9]. Then for almost all primes p and all finite extensions K/Qp ,
writing q = qK ,
6
ζM
6
0,0 ⊗OK
= q 8−7s − q 7−5s + q 6−5s − 2q 5−4s + q 4−4s + q 4−3s − 2q 3−3s + q 2−2s
− q 1−2s + 1 / (1 − q 6−4s )(1 − q 3−2s )2 (1 − q 1−s )2 (1 − q −s ) ,
6
ζM
= q 5−7s − 3q 4−5s + q 4−4s + 2q 3−5s − 2q 3−4s + q 3−3s + q 2−4s
8 ⊗O
K
− 2q 2−3s + 2q 2−2s + q 1−3s − 3q 1−2s + 1
/ (1 − q 6−5s )(1 − q 2−2s )(1 − q 1−s )3 (1 − q −s ) ,
1 − q 2−3s
,
(1 − q 3−2s )(1 − q 2−2s )(1 − q 2−s )(1 − q 1−s )(1 − q −s )
−q 4−5s − q 3−4s + q 3−3s − 2q 2−3s + 2q 2−2s − q 1−2s + q 1−s + 1
.
=
(1 − q 4−3s )(1 − q 3−2s )(1 − q 2−2s )(1 − q 1−s )(1 − q −s )
6
ζM
=
12 ⊗O
K
6
ζM
13 ⊗O
0
K
Remark 9.4. Let g be the non-abelian Lie Q-algebra of dimension 2. Define a Z-form
g of g by g = Zx ⊕ Zy and [x, y] = y.
Then it is easy to see that for all p-adic fields K,
6
−s
1−s
ζg⊗OK (s) = 1/ (1 − qK )(1 − qK ) . Using the notation from [9] as in Theorem 9.3,
M 8 ≈ g ⊕ g and M013 ≈ g ⊗Q Q[X]/X 2 .
9.4. Submodules: Un for n 6 5 and relatives
For any commutative ring R, we consider
1 R
0 . . .
Un (R) =
.. . .
.
.
0 ···
···
..
.
..
.
0
R
..
.
R
1
together with its natural action on Rn by right-multiplication. For n 6 4, the determination of submodule zeta functions associated with Un in the following is quite
straightforward, even without the techniques developed here; the case n = 5, however, is
rather more complicated, as is the resulting formula.
23
Theorem 9.5. For almost all primes p and all finite extensions K/Qp , writing q = qK ,
1
,
(1 − q 1−2s )(1 − q −s )
1 − q 1−4s
,
ζU3 (OK )yO4 (s) =
K
(1 − q 2−4s )(1 − q 1−3s )(1 − q 1−2s )(1 − q −s )
ζU4 (OK )yO4 (s) = FU4 (q, q −s )/ (1 − q 4−8s )(1 − q 3−7s )(1 − q 2−6s )(1 − q 2−5s )
ζU2 (OK )yO2 (s) =
K
K
× (1 − q 2−4s )(1 − q 1−4s )(1 − q 1−2s )(1 − q 1−3s )(1 − q −s ) ,
ζU5 (OK )yO4 (s) = FU5 (q, q −s )/ (1 − q 6−13s )(1 − q 6−12s )(1 − q 4−11s )
K
× (1 − q 4−10s )(1 − q 3−10s )(1 − q 4−9s )(1 − q 3−9s )(1 − q 4−8s )
× (1 − q 3−8s )(1 − q 2−8s )(1 − q 3−7s )(1 − q 2−7s )(1 − q 2−6s )
× (1 − q 2−5s )(1 − q 1−5s )(1 − q 2−4s )(1 − q 1−4s )(1 − q 1−2s )
× (1 − q −s ) ,
where
FU4 = − X 10 Y 30 + X 9 Y 26 + X 9 Y 25 + X 9 Y 24 − X 9 Y 23 + 2X 8 Y 23 − X 8 Y 22 + 2X 7 Y 22
− 2X 7 Y 21 − 2X 7 Y 20 + X 6 Y 21 − 2X 7 Y 19 + X 6 Y 20 − X 6 Y 18 − X 6 Y 17 − X 5 Y 18
− X 5 Y 17 + 2X 6 Y 15 − X 5 Y 16 + X 5 Y 14 − 2X 4 Y 15 + X 5 Y 13 + X 5 Y 12 + X 4 Y 13
+ X 4 Y 12 − X 4 Y 10 + 2X 3 Y 11 − X 4 Y 9 + 2X 3 Y 10 + 2X 3 Y 9 − 2X 3 Y 8 + X 2 Y 8
− 2X 2 Y 7 + XY 7 − XY 6 − XY 5 − XY 4 + 1
and FU5 = 1 + · · · + X 43 Y 124 is given in Appendix A. These formulae for n 6 5 satisfy
the functional equation
ζUn (OK )yOnK (s)
n
q→q −1
n+1
2
= (−1)n q ( 2 )−(
)s · ζ
(s).
Un (OK )yOn
K
Despite the increasing complexity of the formulae in Theorem 9.5, we note that the
“reduced submodule zeta function” of Un (Z) acting on Zn (defined and computed using a
simple variation of [20]) is given by the simple formula 1/((1 − Y )(1 − Y 2 ) · · · (1 − Y n ))
for all n > 1.
Remark 9.6. Let g be an n-dimensional nilpotent Lie k-algebra.
(i) By Engel’s theorem, after choosing a suitable basis, we may regard ad(g) as a
subset of the enveloping associative algebra k[Un (k)] of Un (k) within Mn (k). In
particular, the submodule growth of Un (ov ) acting on onv provides a lower bound
for the ideal growth of nilpotent Lie ov -algebras of additive rank n (and without
ov -torsion).
(ii) Suppose that n > 1. It is easy to see that the minimal number of generators of
k[Un (k)] as a unital, associative k-algebra is n − 1 (use, for instance, [23, p. 263]).
24
Let z denote the centre of g. Then, as a Lie algebra ad(g) ≈ g/z is generated
by dimk (g/([g, g] + z)) many elements. Hence, if g has class > 3, then ad(g) is
generated by fewer than n − 1 elements. If, on the other hand, g has class 2, then
n > 3 and ad(g) is an abelian Lie algebra while k[Un (k)] is non-commutative. We
conclude that ad(g) never generates all of k[Un (k)] for n > 1.
Question 9.7. Is the abscissa of convergence of ζUn (o)yon (s) always 1 for n > 1?
In view of Remark 9.6(i), Question 9.7 is particularly interesting since the abscissa of
convergence of a subalgebra zeta function derived from a k-algebra of dimension n, say,
is bounded from below by a linear function of n (cf. [8, Thm 5.1]).
Let n > 2. If Question 9.7 has a positive answer, then there does not exist a nilpotent Lie
/
o-algebra g which is finitely generated as an o-module such that ζUn (ov )yonv (s) = ζg⊗
(s)
o ov
for almost all v ∈ Q
Vk . Indeed, it is easy to see that for every finite S ⊂ Vk , the abscissa
/
of convergence of v∈Vk \S ζg⊗
(s) is at least d := dimk (g/[g, g] ⊗o k) (cf. [26, Prop. 1])
o ov
and we may clearly assume d > 1. A positive answer to Question 9.7 would thus refine
Remark 9.6(ii).
For another illustration of the generally wild effect of direct products of algebraic
structures on associated zeta functions, we now consider generic local submodule zeta
functions associated with products Un1 × · · · × Unr , diagonally embedded into Un1 +···+nr .
Theorem 9.8. For almost all primes p and all finite extensions K/Qp , writing q = qK ,
ζU22 (OK )yO4 (s) = (1 − q 2−3s )/ (1 − q 3−3s )(1 − q 2−2s )2 (1 − q 1−s )(1 − q −s ) ,
K
ζU32 (OK )yO6 (s) = FU32 (q, q −s )/ (1 − q 8−5s )(1 − q 5−4s )(1 − q 4−3s )(1 − q 3−2s )3
K
× (1 − q 2−s )(1 − q 1−s )(1 − q −s ) ,
ζ(U3 × U2 )(OK )yO5 (s) = FU3 × U2 (q, q −s )/ (1 − q 6−6s )(1 − q 4−5s )(1 − q 3−4s )(1 − q 3−3s )
K
× (1 − q 2−3s )(1 − q 2−2s )2 (1 − q 1−s )(1 − q −s ) ,
ζU23 (OK )yO6 (s) = FU23 (q, q −s )/ (1 − q 9−9s )(1 − q 8−8s )(1 − q 6−7s )(1 − q 5−7s )
K
× (1 − q 6−6s )(1 − q 4−6s )(1 − q 4−5s )(1 − q 3−5s )(1 − q 3−4s )
× (1 − q 3−3s )(1 − q 2−3s )(1 − q 2−2s )2 (1 − q 1−s )(1 − q −s ) ,
where
FU32 = −X 14 Y 12 + 3X 11 Y 9 − X 11 Y 8 − 2X 10 Y 9 + 2X 10 Y 8 − X 8 Y 7 + 2X 7 Y 7
− 2X 7 Y 5 + X 6 Y 5 − 2X 4 Y 4 + 2X 4 Y 3 + X 3 Y 4 − 3X 3 Y 3 + 1,
FU3 × U2 = X 13 Y 18 − X 11 Y 15 − 2X 11 Y 14 + X 11 Y 13 + X 10 Y 14 − 2X 10 Y 13 + X 9 Y 12
− 2X 8 Y 12 + 3X 8 Y 11 − 2X 7 Y 11 + X 8 Y 9 + X 7 Y 10 + X 6 Y 8 + X 5 Y 9
− 2X 6 Y 7 + 3X 5 Y 7 − 2X 5 Y 6 + X 4 Y 6 − 2X 3 Y 5 + X 3 Y 4 + X 2 Y 5
− 2X 2 Y 4 − X 2 Y 3 + 1,
and FU23 = −X 43 Y 57 + · · · + 1 is given in Appendix A.
25
These generic local zeta functions satisfy the following functional equations:
ζU22 (OK )yO4 (s)
q→q −1
= q 6−6s · ζU22 (OK )yO4 (s),
ζU32 (OK )yO6 (s)
q→q −1
= q 15−9s · ζU32 (OK )yO6 (s),
ζ(U3 × U2 )(OK )yO5 (s)
q→q −1
= −q 10−9s · ζ(U3 × U2 )(OK )yO5 (s),
ζU23 (OK )yO6 (s)
q→q −1
= q 15−12s · ζU23 (OK )yO6 (s).
K
K
K
K
K
K
K
K
Further examples of the above form are included with Zeta; here, we only record the
following functional equations.
Theorem 9.9. For almost all primes p and all finite extensions K/Qp , writing q = qK ,
ζ(U5 × U1 )(OK )yO6 (s)
q→q −1
= q 15−16s · ζ(U5 × U1 )(OK )yO6 (s),
ζ(U3 × U2 × U1 )(OK )yO6 (s)
q→q −1
= q 15−10s · ζ(U3 × U2 × U1 )(OK )yO6 (s),
ζ(U4 × U2 )(OK )yO6 (s)
q→q −1
= q 15−13s · ζ(U4 × U2 )(OK )yO6 (s),
K
K
K
ζ(U3 × U22 )(OK )yO7 (s)
K
q→q −1
K
K
K
= −q
21−12s
· ζ(U3 × U22 )(OK )yO7 (s).
K
We note that the Uniformity Problem has a positive solution for each of the four
families of local zeta functions in Theorem 9.9.
10. Applications III: graded subobject zeta functions
By [31, §5.1], up to isomorphism, there are exactly 26 non-abelian fundamental graded
Lie C-algebras (see §3.2) of dimension at most six. All of these algebras are defined
in terms of integral structure constants which thus provide us with “natural” Q-forms.
It turns out that for each of the resulting 26 graded Lie Q-algebras, we can use the
techniques developed here to compute their associated generic local graded subalgebra
and graded ideal zeta functions. We note that for various of these Lie algebras, the
associated non-graded subalgebra and ideal zeta functions are unknown.
Examples of graded ideal zeta functions. Table 2 lists the generic local ideal zeta
functions associated with the aforementioned 26 graded Lie Q-algebras. The first column
contains the names of the associated C-algebras as in [31]; here, an algebra called
“md_c_i” has dimension d and nilpotency class c.
Given a Z-form g of a graded Lie algebra g as indicated by an entry in the first column,
the rational function W (X, Y ) in the corresponding entry of the second column satisfies
the following property: for almost all rational primes p and all finite extensions K/Qp ,
gr /
−s
ζg⊗O
(s) = W (qK , qK
). An entry ±X a Y b in the third column of Table 2 indicates that
K
26
the corresponding W (X, Y ) satisfies W (X −1 , Y −1 ) = ±X a Y b · W (X, Y ); an entry “7”
signifies the absence of such a functional equation.
The algebras m6_3_2 and m6_3_3 are precisely the graded Lie algebras associated
with L(3,2) in [18, Thm 2.32] (also called LW [49, Thm 3.4] and L6,25 [10]) and g6,7 in
[18, Thm 2.45] (called L6,19 (0) in [10]), respectively. The non-graded local ideal zeta
functions of these algebras do not satisfy functional equations of the above form either.
The algebra m6_4_1 is the graded Lie algebra associated with L6,21 (0) from [10]; to the
author’s knowledge, the non-graded local (and topological) subalgebra and ideal zeta
functions of this algebra are unknown.
We note that the formulae for m3_2, m4_3, m5_4_1, and m6_5_1 in Table 2 are
consistent with and explained by Proposition 3.5.
Examples of graded subalgebra zeta functions. While the methods developed here
can be used to compute the generic local graded subalgebra zeta functions of all 26
algebras in Table 2, we chose to only include the smaller ones of these examples in Table 3
(and Appendix B); for a complete list, we refer to Zeta [39].
Open questions. Voll [46, Thm A] established local functional equations under “inversion
of p” for generic local subalgebra zeta functions without any further assumptions on the
algebra in question. It is reasonable to expect the following question to have a positive
answer; the precise form of (10.1) below was suggested to the author by Voll.
Question 10.1. Let A = A1 ⊕ · · · ⊕ Ar be an o-form of a possibly non-associative finitedimensional k-algebra together with a direct sum decomposition into free o-submodules.
r
P
rko (Ai )
Let n = rko (A) and m =
. Is it always the case that
2
i=1
gr6
ζA⊗
(s)
o ov
qv →qv−1
= (−1)n qvm−ns · ζA⊗o ov (s)
(10.1)
for almost all v ∈ Vk ?
The following three questions are graded analogues of conjectures due to Voll [47].
Question 10.2. Let g = g1 ⊕ · · · ⊕ gc be a finite-dimensional graded Lie k-algebra of
class c. Let di = dim(gi ) and d = dim(g). Let 0 = z0 ⊂ · · · ⊂ zc = g be the upper central
series of g and write ei = dim(g/zi ). Let g be an o-form of g as a graded Lie algebra.
gr /
(i) Does ζg⊗
(s) have degree e1 + · · · + ec in qv−s for almost all v ∈ Vk ?
o ov
gr /
(ii) Suppose that there exists W ∈ Q(X, Y ) such that ζg⊗
(s) = W (qv , qv−s ) for almost
o ov
d1
dc
all v ∈ Vk . Does W have degree 2 + · · · + 2 in X?
(iii) Suppose that for almost all v ∈ Vk ,
gr /
ζg⊗
(s)
o ov
qv →qv−1
27
gr /
= εqva−bs · ζg⊗
(s),
o ov
where ε = ±1 and a, b ∈ Z. Do we have ε = (−1)d , a =
b = e1 + · · · + ec ?
d1
2
+ ··· +
dc
2
, and
Finally, the following is closely related to the questions raised in [35, §8.2].
Question 10.3. Let g = g1 ⊕ · · · ⊕ gc be a graded nilpotent Lie o-algebra of class c,
gr6
gr /
where each gi is free and of finite rank as an o-module. Do ζg⊗
(s) and ζg⊗
(s) always
o ov
o ov
have a pole of order c at zero for v ∈ Vk ?
As in [35, §8.2], a natural follow-up question would be to interpret or predict the
leading coefficients of the zeta functions in Question 10.3 expanded as Laurent series
in s; however, perhaps unexpectedly, the examples in Tables 2–3 show that these leading
coefficients are not functions of v and the numbers (rko (g1 ), . . . , rko (gc )) alone.
g
m3_2
m4_2
m4_3
m5_2_1
m5_2_2
m5_2_3
m5_3_1
m5_3_2
m5_4_1
m6_2_1
m6_2_2
m6_2_3
m6_2_4
m6_2_5
m6_2_6
m6_3_1
m6_3_2
m6_3_3
m6_3_4
m6_3_5
m6_3_6
m6_4_1
m6_4_2
m6_4_3
m6_5_1
m6_5_2
gr /
−s
W (X, Y ) s.t. ζg⊗O
(s) = W (qK , qK
)
K
3
1/ (1 − XY ) 1 − Y (1 − Y )
1/ 1 − X 2 Y (1 − XY
) 1 − Y3 (1 − Y )
1/ (1 − XY ) 1 − Y 4 1 − Y 3 (1 − Y )
1−Y 6
(1−XY 3 )(1−X 2 Y
)(1−XY )(1−Y
5 )(1−Y 3 )(1−Y )
1/ 1 − X 3 Y 1 − X 2 Y (1 − XY ) 1 −
1/ 1 − X 3 Y
Y 3 (1 − Y )
1 − X 2 Y (1 − XY ) 1 − Y 5 (1 − Y )
1−Y 8
(1−XY 4 )(1−XY)(1−Y 5 )(1−Y 4 )(1−Y 3 )(1−Y
)
1/ 1 − X 2 Y (1 − XY ) 1 − Y 4 1 − Y 3 (1
1/ (1 − XY ) 1 − Y
5
1−Y
4
1−Y
3
− Y )
(1 − Y )
XY 8 +XY 5 +Y 5 +XY 3 +Y 3 +1
(1−X 2 Y 5 )(1−X 2 Y 3 )(1−X 2 Y )(1−XY )(1−Y 6 )(1−Y )
1−Y 6
(1−XY 3 )(1−X 3 Y )(1−X 2 Y )(1−XY )(1−Y 5 )(1−Y 3 )(1−Y )
Y 4 +Y 3 +Y 2 +Y +1
(1−XY 5 )(1−X 3 Y )(1−X 2 Y )(1−XY )(1−Y 3 )2
1−XY 8
(1−XY 6 )(1−XY 5 )(1−X 3 Y )(1−X 2 Y )(1−XY )(1−Y 3 )(1−Y )
1
(1−X 4 Y )(1−X 3 Y )(1−X 2 Y )(1−XY )(1−Y 3 )(1−Y )
1
(1−X 4 Y )(1−X 3 Y )(1−X 2 Y )(1−XY )(1−Y 5 )(1−Y )
1−Y 8
(1−XY 4 )(1−X 2 Y )(1−XY )(1−Y 5 )(1−Y 4 )(1−Y 3 )(1−Y )
Y 8 +Y 7 +2 Y 6 +2 Y 5 +2 Y 4 +2 Y 3 +Y 2 +Y +1
(1−XY 3 )(1−X 2 Y )(1−XY )(1−Y 6 )(1−Y 5 )(1−Y 4 )
X 7Y 9
X 7 Y 10
X 7 Y 10
X 10 Y 8
X 10 Y 10
X 4 Y 11
same as for m6_3_2
1/ 1 − XY 3 1 − X 2 Y (1 − XY ) 1 − Y 5 1 − Y 3 (1 − Y )
same as for m6_3_4
1/ 1 − X 3 Y 1 − X 2 Y (1 − XY ) 1 − Y 4 1 − Y 3 (1 − Y )
Y 3 −Y +1
(1−XY 4 )(1−XY )(1−Y 6 )(1−Y 5 )(1−Y )2
1−Y 8
3 )(1−Y )
(1−XY 4 )(1−XY)(1−Y 6 )(1−Y 5 )(1−Y 4 )(1−Y
1/ 1 − X 2 Y (1 − XY ) 1 − Y 5 1 − Y 4 1
3
− Y (1 − Y )
1/ (1 − XY ) 1 − Y 6 1 − Y 5 1 − Y 4 1 − Y 3 (1 − Y )
same as for m6_5_1
Table 2: Examples of generic local graded ideal zeta functions
28
FEqn
−XY 5
X 3Y 6
XY 9
−X 4 Y 8
−X 6 Y 7
−X 6 Y 9
−X 2 Y 10
−X 3 Y 10
−XY 14
X 6Y 9
7
7
X 4 Y 14
X 4 Y 14
X 6 Y 11
7
X 2 Y 16
X 3 Y 15
XY 20
XY 20
g
m3_2
m4_2
m4_3
m5_2_1
m5_2_2
m5_2_3
m5_3_1
m5_3_2
m5_4_1
m6_2_1
m6_2_2
m6_2_3
m6_2_5
m6_2_6
m6_3_1
m6_3_2
m6_3_3
m6_3_6
m6_4_3
gr6
−s
W (X, Y ) s.t. ζg⊗O
(s) = W (qK , qK
)
K
1−XY 3
(1−XY 2 )(1−XY )(1−Y 2 )(1−Y )
1−XY 3
(1−X 2 Y )(1−XY 2 )(1−XY )(1−Y 2 )(1−Y )
X 2 Y 9 +X 2 Y 7 +X 2 Y 6 −XY 6 −2 XY 5 −2 XY 4 −XY 3 +Y 3 +Y 2 +1
(1−XY 3 )(1−XY 2 )(1−XY )(1−Y 4 )(1−Y 2 )(1−Y )
−X 2 Y 5 −X 2 Y 3 −XY 3 +XY 2 +Y 2 +1
(1−X 2 Y )(1−XY 2 )(1−X 2 Y 2 )(1−XY )(1−Y 3 )(1−Y )
1−XY 3
(1−XY 2 )(1−X 3 Y )(1−X 2 Y )(1−XY )(1−Y 2 )(1−Y )
−X 4 Y 7 −X 3 Y 6 −X 3 Y 4 −X 2 Y 5 +X 3 Y 3 −XY 4 +X 2 Y 2 +XY 3 +XY +1
(1−X 3 Y 3 )(1−X 2 Y 3 )(1−X 3 Y )(1−X 2 Y )(1−Y 3 )(1−Y )
W531 (B.1)
X 2 Y 9 +X 2 Y 7 +X 2 Y 6 −XY 6 −2 XY 5 −2 XY 4 −XY 3 +Y 3 +Y 2 +1
(1−XY 3 )(1−XY 2 )(1−X 2 Y )(1−XY )(1−Y 4 )(1−Y 2 )(1−Y )
W541 (B.2)
W621 (B.3)
−X 2 Y 5 −X 2 Y 3 −XY 3 +XY 2 +Y 2 +1
(1−X 2 Y 2 )(1−XY 2 )(1−X 3 Y )(1−X 2 Y )(1−XY )(1−Y 3 )(1−Y )
W623 (B.4)
1−XY 3
(1−XY 2 )(1−X 4 Y )(1−X 3 Y )(1−X 2 Y )(1−XY )(1−Y 2 )(1−Y )
−X 4 Y 7 −X 3 Y 6 −X 3 Y 4 −X 2 Y 5 +X 3 Y 3 −XY 4 +X 2 Y 2 +XY 3 +XY +1
(1−X 3 Y 3 )(1−X 2 Y 3 )(1−X 4 Y )(1−X 3 Y )(1−X 2 Y )(1−Y 3 )(1−Y )
W631 (B.5)
W632 (B.6)
W633 (B.7)
X 2 Y 9 +X 2 Y 7 +X 2 Y 6 −XY 6 −2 XY 5 −2 XY 4 −XY 3 +Y 3 +Y 2 +1
(1−XY 3 )(1−XY 2 )(1−X 3 Y )(1−X 2 Y )(1−XY )(1−Y 4 )(1−Y 2 )(1−Y )
W643 (B.8)
FEqn
−XY 3
X 3Y 4
XY 4
−X 4 Y 5
−X 6 Y 5
−X 6 Y 5
−X 2 Y 5
−X 3 Y 5
−XY 5
X 6Y 6
X 7Y 6
X 7Y 6
X 10 Y 6
X 10 Y 6
X 4Y 6
X 4Y 6
X 4Y 6
X 6Y 6
X 3Y 6
Table 3: Examples of generic local graded subalgebra zeta functions
A. Large numerators of local subobject zeta functions
FU5 = X 43 Y 124 + X 42 Y 121 − X 42 Y 120 − X 42 Y 119 − 2X 42 Y 118 + 2X 41 Y 118 − 3X 41 Y 117
+ X 42 Y 115 − 2X 41 Y 116 + X 42 Y 114 − 3X 41 Y 115 − 2X 40 Y 116 − X 42 Y 113 − X 41 Y 114
+ 2X 40 Y 115 + 4X 41 Y 113 − 2X 40 Y 114 − X 39 Y 115 − 2X 40 Y 113 − 2X 39 Y 114 + X 41 Y 111
+ 6X 40 Y 112 − 3X 39 Y 113 − X 41 Y 110 + X 40 Y 111 − 4X 39 Y 112 + 5X 40 Y 110 + 6X 39 Y 111
+ X 38 Y 112 + 3X 39 Y 110 − 6X 38 Y 111 − X 40 Y 108 + 8X 39 Y 109 + 2X 38 Y 110 − 2X 40 Y 107
+ 4X 39 Y 108 + 5X 38 Y 109 − 3X 37 Y 110 + X 39 Y 107 + 9X 38 Y 108 − 4X 39 Y 106 + 8X 38 Y 107
+ X 37 Y 108 − 4X 39 Y 105 + 3X 38 Y 106 + 6X 37 Y 107 − 2X 36 Y 108 − X 39 Y 104 − 5X 38 Y 105
+ 13X 37 Y 106 + 2X 36 Y 107 − 8X 38 Y 104 + 9X 37 Y 105 + 3X 36 Y 106 − X 35 Y 107 + 2X 39 Y 102
− 7X 38 Y 103 − 3X 37 Y 104 + 8X 36 Y 105 + X 35 Y 106 − 6X 38 Y 102 − 15X 37 Y 103 + 5X 36 Y 104
− X 35 Y 105 + 4X 38 Y 101 − 15X 37 Y 102 + 6X 36 Y 103 + 12X 35 Y 104 − X 38 Y 100 − 16X 37 Y 101
− 10X 36 Y 102 + 7X 35 Y 103 − 2X 34 Y 104 + 3X 38 Y 99 + 4X 37 Y 100 − 22X 36 Y 101 + 8X 35 Y 102
+ 5X 34 Y 103 + X 38 Y 98 − 28X 36 Y 100 − 8X 35 Y 101 + 8X 37 Y 98 − 8X 36 Y 99 − 19X 35 Y 100
+ 13X 34 Y 101 + 2X 33 Y 102 + X 37 Y 97 − 2X 36 Y 98 − 30X 35 Y 99 + X 34 Y 100 + X 33 Y 101
− X 37 Y 96 + 17X 36 Y 97 − 19X 35 Y 98 − 16X 34 Y 99 + 6X 33 Y 100 + 17X 36 Y 96 − 6X 35 Y 97
29
− 34X 34 Y 98 + 2X 33 Y 99 + X 32 Y 100 − X 37 Y 94 + 7X 36 Y 95 + 16X 35 Y 96 − 32X 34 Y 97
+ 2X 32 Y 99 − X 36 Y 94 + 32X 35 Y 95 − 15X 34 Y 96 − 21X 33 Y 97 + 2X 32 Y 98 − 5X 36 Y 93
+ 16X 35 Y 94 + 4X 34 Y 95 − 32X 33 Y 96 − 2X 32 Y 97 − X 36 Y 92 + 17X 35 Y 93 + 56X 34 Y 94
− 22X 33 Y 95 − 13X 32 Y 96 + 3X 31 Y 97 − X 36 Y 91 − 5X 35 Y 92 + 30X 34 Y 93 − 13X 33 Y 94
− 22X 32 Y 95 − 7X 35 Y 91 + 37X 34 Y 92 + 59X 33 Y 93 − 25X 32 Y 94 − 5X 31 Y 95 − 3X 35 Y 90
− 5X 34 Y 91 + 41X 33 Y 92 − 29X 32 Y 93 − 17X 31 Y 94 − X 35 Y 89 − 17X 34 Y 90 + 69X 33 Y 91
+ 41X 32 Y 92 − 22X 31 Y 93 + X 30 Y 94 − 11X 34 Y 89 + 18X 33 Y 90 + 48X 32 Y 91 − 26X 31 Y 92
− 5X 30 Y 93 − 12X 34 Y 88 − 26X 33 Y 89 + 89X 32 Y 90 + 13X 31 Y 91 − 10X 30 Y 92 − X 29 Y 93
+ 2X 34 Y 87 − 26X 33 Y 88 + 52X 32 Y 89 + 33X 31 Y 90 − 23X 30 Y 91 − 3X 29 Y 92 + 3X 34 Y 86
− 34X 33 Y 87 − 24X 32 Y 88 + 82X 31 Y 89 − 4X 29 Y 91 + X 34 Y 85 − 10X 33 Y 86 − 38X 32 Y 87
+ 88X 31 Y 88 + 24X 30 Y 89 − 11X 29 Y 90 − 3X 33 Y 85 − 69X 32 Y 86 + 4X 31 Y 87 + 66X 30 Y 88
− 6X 29 Y 89 − 2X 28 Y 90 + 7X 33 Y 84 − 30X 32 Y 85 − 31X 31 Y 86 + 101X 30 Y 87 − 5X 28 Y 89
+ 4X 33 Y 83 − 11X 32 Y 84 − 103X 31 Y 85 + 28X 30 Y 86 + 37X 29 Y 87 − 4X 28 Y 88 + 10X 32 Y 83
− 77X 31 Y 84 − 5X 30 Y 85 + 91X 29 Y 86 − 2X 28 Y 87 + X 33 Y 81 + 9X 32 Y 82 − 40X 31 Y 83
− 99X 30 Y 84 + 53X 29 Y 85 + 21X 28 Y 86 − 2X 27 Y 87 + 4X 32 Y 81 + 9X 31 Y 82 − 115X 30 Y 83
+ 20X 29 Y 84 + 53X 28 Y 85 − 6X 27 Y 86 + 4X 32 Y 80 + 32X 31 Y 81 − 78X 30 Y 82 − 80X 29 Y 83
+ 56X 28 Y 84 + 5X 27 Y 85 − 2X 32 Y 79 + 22X 31 Y 80 − 18X 30 Y 81 − 148X 29 Y 82 + 44X 28 Y 83
+ 30X 27 Y 84 − X 26 Y 85 + 11X 31 Y 79 + 42X 30 Y 80 − 114X 29 Y 81 − 25X 28 Y 82 + 46X 27 Y 83
+ 3X 31 Y 78 + 56X 30 Y 79 − 64X 29 Y 80 − 138X 28 Y 81 + 32X 27 Y 82 + 6X 26 Y 83 − X 31 Y 77
+ 36X 30 Y 78 + 35X 29 Y 79 − 143X 28 Y 80 + 10X 27 Y 81 + 22X 26 Y 82 − 3X 31 Y 76 + 14X 30 Y 77
+ 89X 29 Y 78 − 118X 28 Y 79 − 100X 27 Y 80 + 29X 26 Y 81 + 4X 25 Y 82 + X 31 Y 75 − 3X 30 Y 76
+ 76X 29 Y 77 − 130X 27 Y 79 + 33X 26 Y 80 + 10X 25 Y 81 − 7X 30 Y 75 + 50X 29 Y 76 + 107X 28 Y 77
− 152X 27 Y 78 − 62X 26 Y 79 + 13X 25 Y 80 + X 30 Y 74 + 6X 29 Y 75 + 121X 28 Y 76 − 71X 27 Y 77
− 99X 26 Y 78 + 23X 25 Y 79 + X 24 Y 80 − 3X 30 Y 73 − 19X 29 Y 74 + 100X 28 Y 75 + 88X 27 Y 76
− 137X 26 Y 77 − 16X 25 Y 78 + 8X 24 Y 79 − 12X 29 Y 73 + 34X 28 Y 74 + 145X 27 Y 75 − 119X 26 Y 76
− 53X 25 Y 77 + 16X 24 Y 78 − 7X 29 Y 72 − 16X 28 Y 73 + 167X 27 Y 74 + 47X 26 Y 75 − 107X 25 Y 76
− 3X 24 Y 77 + X 23 Y 78 − X 29 Y 71 − 28X 28 Y 72 + 84X 27 Y 73 + 118X 26 Y 74 − 132X 25 Y 75
− 29X 24 Y 76 + 4X 23 Y 77 − 2X 29 Y 70 − 27X 28 Y 71 − 12X 27 Y 72 + 209X 26 Y 73 + 6X 25 Y 74
− 55X 24 Y 75 + 5X 23 Y 76 − 13X 28 Y 70 − 54X 27 Y 71 + 156X 26 Y 72 + 80X 25 Y 73 − 117X 24 Y 74
− 6X 23 Y 75 + X 22 Y 76 − X 28 Y 69 − 57X 27 Y 70 + 38X 26 Y 71 + 213X 25 Y 72 − 39X 24 Y 73
− 19X 23 Y 74 + X 22 Y 75 + 2X 28 Y 68 − 33X 27 Y 69 − 51X 26 Y 70 + 208X 25 Y 71 + 28X 24 Y 72
− 77X 23 Y 73 − 4X 22 Y 74 − 13X 27 Y 68 − 114X 26 Y 69 + 107X 25 Y 70 + 164X 24 Y 71 − 41X 23 Y 72
− 2X 22 Y 73 − 80X 26 Y 68 − 11X 25 Y 69 + 221X 24 Y 70 − 5X 23 Y 71 − 35X 22 Y 72 − X 21 Y 73
+ 6X 27 Y 66 − 35X 26 Y 67 − 132X 25 Y 68 + 154X 24 Y 69 + 87X 23 Y 70 − 36X 22 Y 71 − X 21 Y 72
+ 3X 27 Y 65 − 131X 25 Y 67 + 44X 24 Y 68 + 200X 23 Y 69 − 22X 22 Y 70 − 14X 21 Y 71 + 10X 26 Y 65
− 97X 25 Y 66 − 124X 24 Y 67 + 176X 23 Y 68 + 33X 22 Y 69 − 21X 21 Y 70 + 11X 26 Y 64 − 18X 25 Y 65
− 177X 24 Y 66 + 109X 23 Y 67 + 135X 22 Y 68 − 15X 21 Y 69 − X 20 Y 70 + 7X 26 Y 63 + 21X 25 Y 64
− 162X 24 Y 65 − 73X 23 Y 66 + 144X 22 Y 67 − 3X 21 Y 68 − 8X 20 Y 69 + 2X 26 Y 62 + 35X 25 Y 63
− 62X 24 Y 64 − 189X 23 Y 65 + 145X 22 Y 66 + 71X 21 Y 67 − 8X 20 Y 68 + X 26 Y 61 + 16X 25 Y 62
− X 24 Y 63 − 216X 23 Y 64 − 7X 22 Y 65 + 101X 21 Y 66 − 9X 20 Y 67 − 2X 19 Y 68 + 5X 25 Y 61
+ 63X 24 Y 62 − 131X 23 Y 63 − 155X 22 Y 64 + 132X 21 Y 65 + 36X 20 Y 66 + 5X 25 Y 60 + 51X 24 Y 61
− 32X 23 Y 62 − 241X 22 Y 63 + 48X 21 Y 64 + 53X 20 Y 65 − 8X 19 Y 66 − 2X 25 Y 59 + 21X 24 Y 60
30
+ 76X 23 Y 61 − 205X 22 Y 62 − 92X 21 Y 63 + 86X 20 Y 64 + 10X 19 Y 65 + 10X 24 Y 59 + 86X 23 Y 60
− 92X 22 Y 61 − 205X 21 Y 62 + 76X 20 Y 63 + 21X 19 Y 64 − 2X 18 Y 65 − 8X 24 Y 58 + 53X 23 Y 59
+ 48X 22 Y 60 − 241X 21 Y 61 − 32X 20 Y 62 + 51X 19 Y 63 + 5X 18 Y 64 + 36X 23 Y 58 + 132X 22 Y 59
− 155X 21 Y 60 − 131X 20 Y 61 + 63X 19 Y 62 + 5X 18 Y 63 − 2X 24 Y 56 − 9X 23 Y 57 + 101X 22 Y 58
− 7X 21 Y 59 − 216X 20 Y 60 − X 19 Y 61 + 16X 18 Y 62 + X 17 Y 63 − 8X 23 Y 56 + 71X 22 Y 57
+ 145X 21 Y 58 − 189X 20 Y 59 − 62X 19 Y 60 + 35X 18 Y 61 + 2X 17 Y 62 − 8X 23 Y 55 − 3X 22 Y 56
+ 144X 21 Y 57 − 73X 20 Y 58 − 162X 19 Y 59 + 21X 18 Y 60 + 7X 17 Y 61 − X 23 Y 54 − 15X 22 Y 55
+ 135X 21 Y 56 + 109X 20 Y 57 − 177X 19 Y 58 − 18X 18 Y 59 + 11X 17 Y 60 − 21X 22 Y 54 + 33X 21 Y 55
+ 176X 20 Y 56 − 124X 19 Y 57 − 97X 18 Y 58 + 10X 17 Y 59 − 14X 22 Y 53 − 22X 21 Y 54 + 200X 20 Y 55
+ 44X 19 Y 56 − 131X 18 Y 57 + 3X 16 Y 59 − X 22 Y 52 − 36X 21 Y 53 + 87X 20 Y 54 + 154X 19 Y 55
− 132X 18 Y 56 − 35X 17 Y 57 + 6X 16 Y 58 − X 22 Y 51 − 35X 21 Y 52 − 5X 20 Y 53 + 221X 19 Y 54
− 11X 18 Y 55 − 80X 17 Y 56 − 2X 21 Y 51 − 41X 20 Y 52 + 164X 19 Y 53 + 107X 18 Y 54 − 114X 17 Y 55
− 13X 16 Y 56 − 4X 21 Y 50 − 77X 20 Y 51 + 28X 19 Y 52 + 208X 18 Y 53 − 51X 17 Y 54 − 33X 16 Y 55
+ 2X 15 Y 56 + X 21 Y 49 − 19X 20 Y 50 − 39X 19 Y 51 + 213X 18 Y 52 + 38X 17 Y 53 − 57X 16 Y 54
− X 15 Y 55 + X 21 Y 48 − 6X 20 Y 49 − 117X 19 Y 50 + 80X 18 Y 51 + 156X 17 Y 52 − 54X 16 Y 53
− 13X 15 Y 54 + 5X 20 Y 48 − 55X 19 Y 49 + 6X 18 Y 50 + 209X 17 Y 51 − 12X 16 Y 52 − 27X 15 Y 53
− 2X 14 Y 54 + 4X 20 Y 47 − 29X 19 Y 48 − 132X 18 Y 49 + 118X 17 Y 50 + 84X 16 Y 51 − 28X 15 Y 52
− X 14 Y 53 + X 20 Y 46 − 3X 19 Y 47 − 107X 18 Y 48 + 47X 17 Y 49 + 167X 16 Y 50 − 16X 15 Y 51
− 7X 14 Y 52 + 16X 19 Y 46 − 53X 18 Y 47 − 119X 17 Y 48 + 145X 16 Y 49 + 34X 15 Y 50 − 12X 14 Y 51
+ 8X 19 Y 45 − 16X 18 Y 46 − 137X 17 Y 47 + 88X 16 Y 48 + 100X 15 Y 49 − 19X 14 Y 50 − 3X 13 Y 51
+ X 19 Y 44 + 23X 18 Y 45 − 99X 17 Y 46 − 71X 16 Y 47 + 121X 15 Y 48 + 6X 14 Y 49 + X 13 Y 50
+ 13X 18 Y 44 − 62X 17 Y 45 − 152X 16 Y 46 + 107X 15 Y 47 + 50X 14 Y 48 − 7X 13 Y 49 + 10X 18 Y 43
+ 33X 17 Y 44 − 130X 16 Y 45 + 76X 14 Y 47 − 3X 13 Y 48 + X 12 Y 49 + 4X 18 Y 42 + 29X 17 Y 43
− 100X 16 Y 44 − 118X 15 Y 45 + 89X 14 Y 46 + 14X 13 Y 47 − 3X 12 Y 48 + 22X 17 Y 42 + 10X 16 Y 43
− 143X 15 Y 44 + 35X 14 Y 45 + 36X 13 Y 46 − X 12 Y 47 + 6X 17 Y 41 + 32X 16 Y 42 − 138X 15 Y 43
− 64X 14 Y 44 + 56X 13 Y 45 + 3X 12 Y 46 + 46X 16 Y 41 − 25X 15 Y 42 − 114X 14 Y 43 + 42X 13 Y 44
+ 11X 12 Y 45 − X 17 Y 39 + 30X 16 Y 40 + 44X 15 Y 41 − 148X 14 Y 42 − 18X 13 Y 43 + 22X 12 Y 44
− 2X 11 Y 45 + 5X 16 Y 39 + 56X 15 Y 40 − 80X 14 Y 41 − 78X 13 Y 42 + 32X 12 Y 43 + 4X 11 Y 44
− 6X 16 Y 38 + 53X 15 Y 39 + 20X 14 Y 40 − 115X 13 Y 41 + 9X 12 Y 42 + 4X 11 Y 43 − 2X 16 Y 37
+ 21X 15 Y 38 + 53X 14 Y 39 − 99X 13 Y 40 − 40X 12 Y 41 + 9X 11 Y 42 + X 10 Y 43 − 2X 15 Y 37
+ 91X 14 Y 38 − 5X 13 Y 39 − 77X 12 Y 40 + 10X 11 Y 41 − 4X 15 Y 36 + 37X 14 Y 37 + 28X 13 Y 38
− 103X 12 Y 39 − 11X 11 Y 40 + 4X 10 Y 41 − 5X 15 Y 35 + 101X 13 Y 37 − 31X 12 Y 38 − 30X 11 Y 39
+ 7X 10 Y 40 − 2X 15 Y 34 − 6X 14 Y 35 + 66X 13 Y 36 + 4X 12 Y 37 − 69X 11 Y 38 − 3X 10 Y 39
− 11X 14 Y 34 + 24X 13 Y 35 + 88X 12 Y 36 − 38X 11 Y 37 − 10X 10 Y 38 + X 9 Y 39 − 4X 14 Y 33
+ 82X 12 Y 35 − 24X 11 Y 36 − 34X 10 Y 37 + 3X 9 Y 38 − 3X 14 Y 32 − 23X 13 Y 33 + 33X 12 Y 34
+ 52X 11 Y 35 − 26X 10 Y 36 + 2X 9 Y 37 − X 14 Y 31 − 10X 13 Y 32 + 13X 12 Y 33 + 89X 11 Y 34
− 26X 10 Y 35 − 12X 9 Y 36 − 5X 13 Y 31 − 26X 12 Y 32 + 48X 11 Y 33 + 18X 10 Y 34 − 11X 9 Y 35
+ X 13 Y 30 − 22X 12 Y 31 + 41X 11 Y 32 + 69X 10 Y 33 − 17X 9 Y 34 − X 8 Y 35 − 17X 12 Y 30
− 29X 11 Y 31 + 41X 10 Y 32 − 5X 9 Y 33 − 3X 8 Y 34 − 5X 12 Y 29 − 25X 11 Y 30 + 59X 10 Y 31
+ 37X 9 Y 32 − 7X 8 Y 33 − 22X 11 Y 29 − 13X 10 Y 30 + 30X 9 Y 31 − 5X 8 Y 32 − X 7 Y 33
+ 3X 12 Y 27 − 13X 11 Y 28 − 22X 10 Y 29 + 56X 9 Y 30 + 17X 8 Y 31 − X 7 Y 32 − 2X 11 Y 27
− 32X 10 Y 28 + 4X 9 Y 29 + 16X 8 Y 30 − 5X 7 Y 31 + 2X 11 Y 26 − 21X 10 Y 27 − 15X 9 Y 28
31
+ 32X 8 Y 29 − X 7 Y 30 + 2X 11 Y 25 − 32X 9 Y 27 + 16X 8 Y 28 + 7X 7 Y 29 − X 6 Y 30
+ X 11 Y 24 + 2X 10 Y 25 − 34X 9 Y 26 − 6X 8 Y 27 + 17X 7 Y 28 + 6X 10 Y 24 − 16X 9 Y 25
− 19X 8 Y 26 + 17X 7 Y 27 − X 6 Y 28 + X 10 Y 23 + X 9 Y 24 − 30X 8 Y 25 − 2X 7 Y 26
+ X 6 Y 27 + 2X 10 Y 22 + 13X 9 Y 23 − 19X 8 Y 24 − 8X 7 Y 25 + 8X 6 Y 26 − 8X 8 Y 23
− 28X 7 Y 24 + X 5 Y 26 + 5X 9 Y 21 + 8X 8 Y 22 − 22X 7 Y 23 + 4X 6 Y 24 + 3X 5 Y 25
− 2X 9 Y 20 + 7X 8 Y 21 − 10X 7 Y 22 − 16X 6 Y 23 − X 5 Y 24 + 12X 8 Y 20 + 6X 7 Y 21
− 15X 6 Y 22 + 4X 5 Y 23 − X 8 Y 19 + 5X 7 Y 20 − 15X 6 Y 21 − 6X 5 Y 22 + X 8 Y 18
+ 8X 7 Y 19 − 3X 6 Y 20 − 7X 5 Y 21 + 2X 4 Y 22 − X 8 Y 17 + 3X 7 Y 18 + 9X 6 Y 19
− 8X 5 Y 20 + 2X 7 Y 17 + 13X 6 Y 18 − 5X 5 Y 19 − X 4 Y 20 − 2X 7 Y 16 + 6X 6 Y 17
+ 3X 5 Y 18 − 4X 4 Y 19 + X 6 Y 16 + 8X 5 Y 17 − 4X 4 Y 18 + 9X 5 Y 16 + X 4 Y 17 − 3X 6 Y 14
+ 5X 5 Y 15 + 4X 4 Y 16 − 2X 3 Y 17 + 2X 5 Y 14 + 8X 4 Y 15 − X 3 Y 16 − 6X 5 Y 13 + 3X 4 Y 14
+ X 5 Y 12 + 6X 4 Y 13 + 5X 3 Y 14 − 4X 4 Y 12 + X 3 Y 13 − X 2 Y 14 − 3X 4 Y 11 + 6X 3 Y 12 + X 2 Y 13
− 2X 4 Y 10 − 2X 3 Y 11 − X 4 Y 9 − 2X 3 Y 10 + 4X 2 Y 11 + 2X 3 Y 9 − X 2 Y 10 − XY 11 − 2X 3 Y 8
− 3X 2 Y 9 + XY 10 − 2X 2 Y 8 + XY 9 − 3X 2 Y 7 + 2X 2 Y 6 − 2XY 6 − XY 5 − XY 4 + XY 3 + 1
FQ[T ]/T 4 = − X 49 Y 54 − X 49 Y 53 − 2X 48 Y 52 − X 47 Y 52 − 3X 47 Y 51 − 2X 46 Y 51 − 3X 46 Y 50
− 4X 45 Y 50 + X 46 Y 48 − 4X 45 Y 49 + X 45 Y 48 − 4X 44 Y 49 + X 45 Y 47 − 2X 44 Y 48
+ 3X 44 Y 47 − 8X 43 Y 48 + X 44 Y 46 − X 43 Y 47 + 8X 43 Y 46 − 9X 42 Y 47 + X 43 Y 45
+ 3X 42 Y 46 + 9X 42 Y 45 − 12X 41 Y 46 + X 42 Y 44 + 10X 41 Y 45 + 10X 41 Y 44 − 13X 40 Y 45
+ 23X 40 Y 44 + 7X 40 Y 43 − 19X 39 Y 44 − 3X 40 Y 42 + 35X 39 Y 43 + 3X 39 Y 42 − 19X 38 Y 43
− 3X 39 Y 41 + 54X 38 Y 42 − 15X 38 Y 41 − 24X 37 Y 42 − 6X 38 Y 40 + 74X 37 Y 41 − 31X 37 Y 40
− 25X 36 Y 41 − 5X 37 Y 39 + 95X 36 Y 40 − 55X 36 Y 39 − 30X 35 Y 40 − 4X 36 Y 38 + 110X 35 Y 39
− X 36 Y 37 − 85X 35 Y 38 − 28X 34 Y 39 + 10X 35 Y 37 + 131X 34 Y 38 − 3X 35 Y 36 − 127X 34 Y 37
− 31X 33 Y 38 + 22X 34 Y 36 + 143X 33 Y 37 − 4X 34 Y 35 − 160X 33 Y 36 − 29X 32 Y 37 + 46X 33 Y 35
+ 154X 32 Y 36 − 8X 33 Y 34 − 204X 32 Y 35 − 30X 31 Y 36 + 73X 32 Y 34 + 159X 31 Y 35 − 11X 32 Y 33
− 246X 31 Y 34 − 26X 30 Y 35 + X 32 Y 32 + 113X 31 Y 33 + 169X 30 Y 34 − 19X 31 Y 32 − 290X 30 Y 33
− 27X 29 Y 34 + X 31 Y 31 + 148X 30 Y 32 + 166X 29 Y 33 − 26X 30 Y 31 − 314X 29 Y 32 − 23X 28 Y 33
+ 3X 30 Y 30 + 193X 29 Y 31 + 162X 28 Y 32 − 39X 29 Y 30 − 344X 28 Y 31 − 22X 27 Y 32 + 3X 29 Y 29
+ 230X 28 Y 30 + 153X 27 Y 31 − 49X 28 Y 29 − 354X 27 Y 30 − 17X 26 Y 31 + 6X 28 Y 28 + 271X 27 Y 29
+ 142X 26 Y 30 − 68X 27 Y 28 − 359X 26 Y 29 − 16X 25 Y 30 + 6X 27 Y 27 + 301X 26 Y 28 + 121X 25 Y 29
− 85X 26 Y 27 − 344X 25 Y 28 − 11X 24 Y 29 + 10X 26 Y 26 + 332X 25 Y 27 + 104X 24 Y 28 − 104X 25 Y 26
− 332X 24 Y 27 − 10X 23 Y 28 + 11X 25 Y 25 + 344X 24 Y 26 + 85X 23 Y 27 − 121X 24 Y 25 − 301X 23 Y 26
− 6X 22 Y 27 + 16X 24 Y 24 + 359X 23 Y 25 + 68X 22 Y 26 − 142X 23 Y 24 − 271X 22 Y 25 − 6X 21 Y 26
+ 17X 23 Y 23 + 354X 22 Y 24 + 49X 21 Y 25 − 153X 22 Y 23 − 230X 21 Y 24 − 3X 20 Y 25 + 22X 22 Y 22
+ 344X 21 Y 23 + 39X 20 Y 24 − 162X 21 Y 22 − 193X 20 Y 23 − 3X 19 Y 24 + 23X 21 Y 21 + 314X 20 Y 22
+ 26X 19 Y 23 − 166X 20 Y 21 − 148X 19 Y 22 − X 18 Y 23 + 27X 20 Y 20 + 290X 19 Y 21 + 19X 18 Y 22
− 169X 19 Y 20 − 113X 18 Y 21 − X 17 Y 22 + 26X 19 Y 19 + 246X 18 Y 20 + 11X 17 Y 21 − 159X 18 Y 19
− 73X 17 Y 20 + 30X 18 Y 18 + 204X 17 Y 19 + 8X 16 Y 20 − 154X 17 Y 18 − 46X 16 Y 19 + 29X 17 Y 17
+ 160X 16 Y 18 + 4X 15 Y 19 − 143X 16 Y 17 − 22X 15 Y 18 + 31X 16 Y 16 + 127X 15 Y 17 + 3X 14 Y 18
− 131X 15 Y 16 − 10X 14 Y 17 + 28X 15 Y 15 + 85X 14 Y 16 + X 13 Y 17 − 110X 14 Y 15 + 4X 13 Y 16
+ 30X 14 Y 14 + 55X 13 Y 15 − 95X 13 Y 14 + 5X 12 Y 15 + 25X 13 Y 13 + 31X 12 Y 14 − 74X 12 Y 13
32
+ 6X 11 Y 14 + 24X 12 Y 12 + 15X 11 Y 13 − 54X 11 Y 12 + 3X 10 Y 13 + 19X 11 Y 11 − 3X 10 Y 12
− 35X 10 Y 11 + 3X 9 Y 12 + 19X 10 Y 10 − 7X 9 Y 11 − 23X 9 Y 10 + 13X 9 Y 9 − 10X 8 Y 10
− 10X 8 Y 9 − X 7 Y 10 + 12X 8 Y 8 − 9X 7 Y 9 − 3X 7 Y 8 − X 6 Y 9 + 9X 7 Y 7
− 8X 6 Y 8 + X 6 Y 7 − X 5 Y 8 + 8X 6 Y 6 − 3X 5 Y 7 + 2X 5 Y 6 − X 4 Y 7
+ 4X 5 Y 5 − X 4 Y 6 + 4X 4 Y 5 − X 3 Y 6 + 4X 4 Y 4 + 3X 3 Y 4 + 2X 3 Y 3
+ 3X 2 Y 3 + X 2 Y 2 + 2XY 2 + Y + 1
FU2 = − X 43 Y 57 + 4X 41 Y 53 − X 41 Y 52 − 3X 40 Y 53 − 2X 41 Y 51 + 4X 40 Y 52 + X 41 Y 50 + 4X 40 Y 51
3
− X 39 Y 52 − 3X 40 Y 50 − 3X 39 Y 51 + 5X 39 Y 50 + 4X 38 Y 51 − 4X 38 Y 50 − X 37 Y 51 − 2X 39 Y 48
− 3X 38 Y 49 + 3X 37 Y 50 + X 39 Y 47 − X 38 Y 48 + 5X 37 Y 49 − 3X 38 Y 47 − 3X 37 Y 48 − X 36 Y 49
+ 3X 38 Y 46 + 4X 36 Y 48 − 6X 37 Y 46 − 2X 36 Y 47 + X 35 Y 48 − X 37 Y 45 − X 36 Y 46 − 2X 35 Y 47
+ 2X 37 Y 44 + 4X 36 Y 45 − 4X 35 Y 46 + 3X 34 Y 47 − 3X 36 Y 44 − 6X 35 Y 45 + X 34 Y 46 − X 36 Y 43
+ 8X 35 Y 44 − 6X 34 Y 45 + X 33 Y 46 + 6X 35 Y 43 − 9X 34 Y 44 + 4X 33 Y 45 − X 35 Y 42 − 4X 34 Y 43
− 6X 33 Y 44 − X 32 Y 45 − 2X 35 Y 41 + 13X 34 Y 42 + 4X 33 Y 43 − 2X 32 Y 44 + 7X 34 Y 41 − 11X 33 Y 42
− 6X 32 Y 43 + 3X 31 Y 44 − 4X 34 Y 40 − 2X 33 Y 41 + 10X 32 Y 42 − 4X 31 Y 43 + 10X 33 Y 40 + 19X 32 Y 41
− 8X 31 Y 42 + X 30 Y 43 − 2X 33 Y 39 − 16X 32 Y 40 − 21X 31 Y 41 − 2X 30 Y 42 − 7X 32 Y 39 + 40X 31 Y 40
+ 6X 30 Y 41 − X 29 Y 42 + 8X 32 Y 38 + 19X 31 Y 39 − 14X 30 Y 40 − 3X 32 Y 37 − 30X 31 Y 38 − 9X 30 Y 39
− 9X 29 Y 40 − 2X 28 Y 41 + 2X 31 Y 37 + 32X 30 Y 38 + 32X 29 Y 39 + 2X 28 Y 40 + 2X 30 Y 37 − 10X 29 Y 38
− 19X 28 Y 39 − 13X 30 Y 36 − 26X 29 Y 37 + 18X 28 Y 38 − X 27 Y 39 + 18X 29 Y 36 + 42X 28 Y 37 + X 26 Y 39
+ X 30 Y 34 − 3X 29 Y 35 − 31X 28 Y 36 − 15X 27 Y 37 − 5X 26 Y 38 − 23X 28 Y 35 + 29X 27 Y 36 + 14X 26 Y 37
+ X 29 Y 33 + 2X 28 Y 34 + 17X 27 Y 35 − 5X 25 Y 37 + 3X 28 Y 33 − 30X 27 Y 34 − 20X 26 Y 35 + 5X 25 Y 36
+ 3X 28 Y 32 − X 27 Y 33 + 20X 26 Y 34 + 33X 25 Y 35 − 2X 24 Y 36 − 12X 26 Y 33 − 20X 25 Y 34 − 9X 24 Y 35
− X 27 Y 31 − 4X 26 Y 32 − 34X 25 Y 33 + 15X 24 Y 34 + 2X 23 Y 35 + 3X 27 Y 30 + 12X 26 Y 31 + 9X 25 Y 32
+ 34X 24 Y 33 + X 23 Y 34 − 7X 26 Y 30 + X 25 Y 31 − 48X 24 Y 32 − 17X 23 Y 33 + X 22 Y 34 + X 26 Y 29
+ 13X 25 Y 30 − 14X 24 Y 31 + 18X 23 Y 32 + 10X 22 Y 33 + 4X 25 Y 29 + 13X 24 Y 30 − 4X 23 Y 31
+ 5X 22 Y 32 − 5X 25 Y 28 − 4X 24 Y 29 − 20X 23 Y 30 − 36X 22 Y 31 − 3X 21 Y 32 + 15X 24 Y 28
+ 20X 23 Y 29 + 3X 22 Y 30 + 26X 21 Y 31 + X 20 Y 32 − 4X 24 Y 27 − 8X 23 Y 28 − 2X 22 Y 29
− 21X 21 Y 30 − 3X 20 Y 31 + 22X 22 Y 28 − 22X 21 Y 29 + 3X 23 Y 26 + 21X 22 Y 27 + 2X 21 Y 28
+ 8X 20 Y 29 + 4X 19 Y 30 − X 23 Y 25 − 26X 22 Y 26 − 3X 21 Y 27 − 20X 20 Y 28 − 15X 19 Y 29
+ 3X 22 Y 25 + 36X 21 Y 26 + 20X 20 Y 27 + 4X 19 Y 28 + 5X 18 Y 29 − 5X 21 Y 25 + 4X 20 Y 26
− 13X 19 Y 27 − 4X 18 Y 28 − 10X 21 Y 24 − 18X 20 Y 25 + 14X 19 Y 26 − 13X 18 Y 27 − X 17 Y 28
− X 21 Y 23 + 17X 20 Y 24 + 48X 19 Y 25 − X 18 Y 26 + 7X 17 Y 27 − X 20 Y 23 − 34X 19 Y 24
− 9X 18 Y 25 − 12X 17 Y 26 − 3X 16 Y 27 − 2X 20 Y 22 − 15X 19 Y 23 + 34X 18 Y 24 + 4X 17 Y 25
+ X 16 Y 26 + 9X 19 Y 22 + 20X 18 Y 23 + 12X 17 Y 24 + 2X 19 Y 21 − 33X 18 Y 22 − 20X 17 Y 23
+ X 16 Y 24 − 3X 15 Y 25 − 5X 18 Y 21 + 20X 17 Y 22 + 30X 16 Y 23 − 3X 15 Y 24 + 5X 18 Y 20
− 17X 16 Y 22 − 2X 15 Y 23 − X 14 Y 24 − 14X 17 Y 20 − 29X 16 Y 21 + 23X 15 Y 22 + 5X 17 Y 19
+ 15X 16 Y 20 + 31X 15 Y 21 + 3X 14 Y 22 − X 13 Y 23 − X 17 Y 18 − 42X 15 Y 20 − 18X 14 Y 21
+ X 16 Y 18 − 18X 15 Y 19 + 26X 14 Y 20 + 13X 13 Y 21 + 19X 15 Y 18 + 10X 14 Y 19 − 2X 13 Y 20
− 2X 15 Y 17 − 32X 14 Y 18 − 32X 13 Y 19 − 2X 12 Y 20 + 2X 15 Y 16 + 9X 14 Y 17 + 9X 13 Y 18
+ 30X 12 Y 19 + 3X 11 Y 20 + 14X 13 Y 17 − 19X 12 Y 18 − 8X 11 Y 19 + X 14 Y 15 − 6X 13 Y 16
− 40X 12 Y 17 + 7X 11 Y 18 + 2X 13 Y 15 + 21X 12 Y 16 + 16X 11 Y 17 + 2X 10 Y 18 − X 13 Y 14 + 8X 12 Y 15
33
− 19X 11 Y 16 − 10X 10 Y 17 + 4X 12 Y 14 − 10X 11 Y 15 + 2X 10 Y 16 + 4X 9 Y 17 − 3X 12 Y 13
+ 6X 11 Y 14 + 11X 10 Y 15 − 7X 9 Y 16 + 2X 11 Y 13 − 4X 10 Y 14 − 13X 9 Y 15 + 2X 8 Y 16 + X 11 Y 12
+ 6X 10 Y 13 + 4X 9 Y 14 + X 8 Y 15 − 4X 10 Y 12 + 9X 9 Y 13 − 6X 8 Y 14 − X 10 Y 11 + 6X 9 Y 12 − 8X 8 Y 13
+ X 7 Y 14 − X 9 Y 11 + 6X 8 Y 12 + 3X 7 Y 13 − 3X 9 Y 10 + 4X 8 Y 11 − 4X 7 Y 12 − 2X 6 Y 13 + 2X 8 Y 10
+ X 7 Y 11 + X 6 Y 12 − X 8 Y 9 + 2X 7 Y 10 + 6X 6 Y 11 − 4X 7 Y 9 − 3X 5 Y 11 + X 7 Y 8 + 3X 6 Y 9 + 3X 5 Y 10
− 5X 6 Y 8 + X 5 Y 9 − X 4 Y 10 − 3X 6 Y 7 + 3X 5 Y 8 + 2X 4 Y 9 + X 6 Y 6 + 4X 5 Y 7 − 4X 5 Y 6 − 5X 4 Y 7
+ 3X 4 Y 6 + 3X 3 Y 7 + X 4 Y 5 − 4X 3 Y 6 − X 2 Y 7 − 4X 3 Y 5 + 2X 2 Y 6 + 3X 3 Y 4 + X 2 Y 5 − 4X 2 Y 4 + 1
B. Formulae for local graded subalgebra zeta functions
W531 = −X 5 Y 18 − X 5 Y 16 − X 5 Y 15 − X 4 Y 16 − X 5 Y 14 − X 4 Y 15 + 2X 4 Y 13 + X 3 Y 14
+ X 4 Y 12 + 2X 3 Y 13 + X 4 Y 11 + X 3 Y 12 + X 2 Y 13 + X 4 Y 10 + X 3 Y 11 + X 4 Y 9
+ 3X 3 Y 10 + 2X 3 Y 9 + X 2 Y 10 − X 3 Y 8 − 2X 2 Y 9 − 3X 2 Y 8 − XY 9 − X 2 Y 7
− XY 8 − X 3 Y 5 − X 2 Y 6 − XY 7 − 2X 2 Y 5 − XY 6 − X 2 Y 4 − 2XY 5 + XY 3
+ Y 4 + XY 2 + Y 3 + Y 2 + 1
/ 1 − XY 5 1 − X 2 Y 3 1 − X 2 Y 4 1 − XY 2 (1 − XY ) 1 − Y 5 1 − Y 2 (1 − Y )
(B.1)
W541 = −X 3 Y 21 − X 3 Y 20 − 3X 3 Y 19 − 5X 3 Y 18 − 7X 3 Y 17 + X 2 Y 18 − 8X 3 Y 16 + 4X 2 Y 17
− 7X 3 Y 15 + 9X 2 Y 16 − 6X 3 Y 14 + 16X 2 Y 15 − 6X 3 Y 13 + 19X 2 Y 14 − XY 15 − 4X 3 Y 12
+ 21X 2 Y 13 − 4XY 14 − 3X 3 Y 11 + 21X 2 Y 12 − 8XY 13 − X 3 Y 10 + 20X 2 Y 11 − 14XY 12
+ 18X 2 Y 10 − 18XY 11 + 14X 2 Y 9 − 20XY 10 + Y 11 + 8X 2 Y 8 − 21XY 9 + 3Y 10
+ 4X 2 Y 7 − 21XY 8 + 4Y 9 + X 2 Y 6 − 19XY 7 + 6Y 8 − 16XY 6 + 6Y 7 − 9XY 5
+ 7Y 6 − 4XY 4 + 8Y 5 − XY 3 + 7Y 4 + 5Y 3 + 3Y 2 + Y + 1
/ 1 − XY 4 1 − XY 3 1 − XY 2 (1 − XY ) 1 − Y 7 1 − Y 4 1 − Y 3 1 − Y 2
(B.2)
W621 = X 4 Y 8 + X 4 Y 6 + X 3 Y 6 − X 3 Y 5 + X 2 Y 6 − X 3 Y 4 − X 2 Y 5 − X 3 Y 3 − XY 5
− X 2 Y 3 − XY 4 + X 2 Y 2 − XY 3 + XY 2 + Y 2 + 1 XY 2 + 1
/ 1 − X 2 Y 3 1 − XY 3 1 − X 3 Y 2 1 − X 2 Y 2 1 − X 2 Y (1 − XY ) 1 − Y 3 (1 − Y )
(B.3)
34
W623 = X 9 Y 14 + X 9 Y 13 + X 9 Y 12 − 3X 8 Y 11 − 2X 8 Y 10 − X 6 Y 11 − X 6 Y 10 + X 7 Y 8 − 3X 6 Y 9
− X 5 Y 9 − X 4 Y 10 + X 6 Y 7 + 2X 5 Y 8 + X 6 Y 6 + 2X 5 Y 7 + 2X 4 Y 8 + 2X 5 Y 6 + 2X 4 Y 7
+ X 3 Y 8 + 2X 4 Y 6 + X 3 Y 7 − X 5 Y 4 − X 4 Y 5 − 3X 3 Y 5 + X 2 Y 6 − X 3 Y 4 − X 3 Y 3
2
− 2XY 4 − 3XY 3 + Y 2 + Y + 1 / 1 − X 4 Y 3 1 − X 3 Y 3 1 − XY 3 1 − XY 2
2
× 1 − X 3 Y 1 − X 2 Y (1 − XY ) 1 − Y 2
(B.4)
W631 = −X 5 Y 18 − X 5 Y 16 − X 5 Y 15 − X 4 Y 16 − X 5 Y 14 − X 4 Y 15 + 2X 4 Y 13 + X 3 Y 14
+ X 4 Y 12 + 2X 3 Y 13 + X 4 Y 11 + X 3 Y 12 + X 2 Y 13 + X 4 Y 10 + X 3 Y 11 + X 4 Y 9
+ 3X 3 Y 10 + 2X 3 Y 9 + X 2 Y 10 − X 3 Y 8 − 2X 2 Y 9 − 3X 2 Y 8 − XY 9 − X 2 Y 7 − XY 8
− X 3 Y 5 − X 2 Y 6 − XY 7 − 2X 2 Y 5 − XY 6 − X 2 Y 4 − 2XY 5 + XY 3 + Y 4 + XY 2
+ Y 3 + Y 2 + 1 / 1 − XY 5 1 − X 2 Y 4 1 − X 2 Y 3 1 − XY 2 1 − X 2 Y
× (1 − XY ) 1 − Y 5 1 − Y 2 (1 − Y )
(B.5)
W632 = X 3 Y 16 + 2X 3 Y 15 + 4X 3 Y 14 + 7X 3 Y 13 + X 2 Y 14 + 10X 3 Y 12 + 2X 2 Y 13 + 11X 3 Y 11
− XY 13 + 10X 3 Y 10 − 3X 2 Y 11 − 3XY 12 + 7X 3 Y 9 − 8X 2 Y 10 − 6XY 11 + 4X 3 Y 8
− 11X 2 Y 9 − 9XY 10 + 2X 3 Y 7 − 11X 2 Y 8 − 10XY 9 + Y 10 + X 3 Y 6 − 10X 2 Y 7
− 11XY 8 + 2Y 9 − 9X 2 Y 6 − 11XY 7 + 4Y 8 − 6X 2 Y 5 − 8XY 6 + 7Y 7 − 3X 2 Y 4
− 3XY 5 + 10Y 6 − X 2 Y 3 + 11Y 5 + 2XY 3 + 10Y 4 + XY 2 + 7Y 3 + 4Y 2 + 2Y + 1
× (1 − Y )/ 1 − XY 3 1 − X 2 Y 2 1 − XY 2 1 − X 2 Y (1 − XY )
× 1−Y5 1−Y4 1−Y3 1−Y2
(B.6)
W633 = X 3 Y 14 + X 3 Y 13 + 3X 3 Y 12 + 4X 3 Y 11 + X 2 Y 12 + 5X 3 Y 10 + 4X 3 Y 9 − X 2 Y 10
− XY 11 + 3X 3 Y 8 − 3X 2 Y 9 − XY 10 + X 3 Y 7 − 5X 2 Y 8 − 4XY 9 + X 3 Y 6 − 5X 2 Y 7
− 4XY 8 − 4X 2 Y 6 − 5XY 7 + Y 8 − 4X 2 Y 5 − 5XY 6 + Y 7 − X 2 Y 4 − 3XY 5 + 3Y 6
− X 2 Y 3 − XY 4 + 4Y 5 + 5Y 4 + XY 2 + 4Y 3 + 3Y 2 + Y + 1
2
/ 1 − XY 3 1 − X 2 Y 2 1 − XY 2 1 − X 2 Y (1 − XY ) 1 − Y 3 1 − Y 4
(B.7)
35
W643 = −X 3 Y 21 − X 3 Y 20 − 3X 3 Y 19 − 5X 3 Y 18 − 7X 3 Y 17 + X 2 Y 18 − 8X 3 Y 16 + 4X 2 Y 17
− 7X 3 Y 15 + 9X 2 Y 16 − 6X 3 Y 14 + 16X 2 Y 15 − 6X 3 Y 13 + 19X 2 Y 14 − XY 15 − 4X 3 Y 12
+ 21X 2 Y 13 − 4XY 14 − 3X 3 Y 11 + 21X 2 Y 12 − 8XY 13 − X 3 Y 10 + 20X 2 Y 11 − 14XY 12
+ 18X 2 Y 10 − 18XY 11 + 14X 2 Y 9 − 20XY 10 + Y 11 + 8X 2 Y 8 − 21XY 9 + 3Y 10
+ 4X 2 Y 7 − 21XY 8 + 4Y 9 + X 2 Y 6 − 19XY 7 + 6Y 8 − 16XY 6 + 6Y 7 − 9XY 5 + 7Y 6
− 4XY 4 + 8Y 5 − XY 3 + 7Y 4 + 5Y 3 + 3Y 2 + Y + 1
/ 1 − XY 4 1 − XY 3 1 − XY 2 1 − X 2 Y (1 − XY )
× 1−Y7 1−Y4 1−Y3 1−Y2
(B.8)
References
[1] W. W. Adams and P. Loustaunau, An introduction to Gröbner bases, Graduate Studies in Mathematics, vol. 3, American Mathematical Society, Providence, RI, 1994.
[2] N. Avni, B. Klopsch, U. Onn, and C. Voll, Representation zeta functions of compact p-adic analytic
groups and arithmetic groups, Duke Math. J. 162 (2013), no. 1, 111–197.
[3] V. Baldoni, N. Berline, J. A. De Loera, B. Dutra, M. Köppe, S. Moreinis, G. Pinto, M. Vergne,
and J. Wu, A user’s guide for LattE integrale v1.7.3, 2015. Software package. LattE is available at
http://www.math.ucdavis.edu/~latte/.
[4] A. Barvinok, Integer points in polyhedra, Zurich Lectures in Advanced Mathematics, European
Mathematical Society (EMS), Zürich, 2008.
[5] A. Barvinok and J. E. Pommersheim, An algorithmic theory of lattice points in polyhedra, New
perspectives in algebraic combinatorics (Berkeley, CA, 1996–97), 1999, pp. 91–147.
[6] A. Barvinok and K. Woods, Short rational generating functions for lattice point problems, J. Amer.
Math. Soc. 16 (2003), no. 4, 957–979 (electronic).
[7] A. I. Barvinok, A polynomial time algorithm for counting integral points in polyhedra when the
dimension is fixed, Math. Oper. Res. 19 (1994), no. 4, 769–779.
[8] J. F. Brakenhoff, Counting problems for number rings. PhD Thesis, Universiteit Leiden, 2009. See
http://hdl.handle.net/1887/14539.
[9] W. A. de Graaf, Classification of solvable Lie algebras, Experiment. Math. 14 (2005), no. 1, 15–25.
[10] W. A. de Graaf, Classification of 6-dimensional nilpotent Lie algebras over fields of characteristic
not 2, J. Algebra 309 (2007), no. 2, 640–653.
[11] M. Demazure and P. Gabriel, Groupes algébriques. Tome I: Géométrie algébrique, généralités, groupes
commutatifs, Masson & Cie, Éditeur, Paris; North-Holland Publishing Co., Amsterdam, 1970. Avec
un appendice ıt Corps de classes local par Michiel Hazewinkel.
[12] J. Denef, Report on Igusa’s local zeta function, Astérisque 201-203 (1991), Exp. No. 741, 359–386
(1992). Séminaire Bourbaki, Vol. 1990/91.
[13] J. Denef and F. Loeser, Caractéristiques d’Euler-Poincaré, fonctions zêta locales et modifications
analytiques, J. Amer. Math. Soc. 5 (1992), no. 4, 705–720.
[14] M. P. F. du Sautoy, The zeta function of sl2 (Z), Forum Math. 12 (2000), no. 2, 197–221.
[15] M. P. F. du Sautoy and F. J. Grunewald, Analytic properties of zeta functions and subgroup growth,
Ann. of Math. (2) 152 (2000), no. 3, 793–833.
36
[16] M. P. F. du Sautoy and F. Loeser, Motivic zeta functions of infinite-dimensional Lie algebras, Selecta
Math. (N.S.) 10 (2004), no. 2, 253–303.
[17] M. P. F. du Sautoy and G. Taylor, The zeta function of sl2 and resolution of singularities, Math.
Proc. Cambridge Philos. Soc. 132 (2002), no. 1, 57–73.
[18] M. P. F. du Sautoy and L. Woodward, Zeta functions of groups and rings, Lecture Notes in
Mathematics, vol. 1925, Springer-Verlag, Berlin, 2008.
[19] D. H. Dung and C. Voll, Uniform analytic properties of representation zeta functions of finitely
generated nilpotent groups (2015). To appear in Trans. Amer. Math. Soc. See arXiv:1503.06947.
[20] A. Evseev, Reduced zeta functions of Lie algebras, J. Reine Angew. Math. 633 (2009), 197–211.
[21] S. Ezzat, Representation growth of finitely generated torsion-free nilpotent groups: Methods and
examples. PhD Thesis, University of Canterbury, 2012. See http://hdl.handle.net/10092/7235.
[22]
, Counting irreducible representations of the Heisenberg group over the integers of a quadratic
number field, J. Algebra 397 (2014), 609–624.
[23] E. S. Golod and I. R. Šafarevič, On the class field tower (Russian), Izv. Akad. Nauk SSSR Ser. Mat.
28 (1964), 261–272.
[24] J. González-Sánchez, A. Jaikin-Zapirain, and B. Klopsch, The representation zeta function of a
FAb compact p-adic Lie group vanishes at −2, Bull. Lond. Math. Soc. 46 (2014), no. 2, 239–244.
MR3194743
[25] G.-M. Greuel, G. Pfister, and H. Schönemann, Singular 3.0, Centre for Computer Algebra,
University of Kaiserslautern, 2005. Available from http://www.singular.uni-kl.de/.
[26] F. J. Grunewald, D. Segal, and G. C. Smith, Subgroups of finite index in nilpotent groups, Invent.
Math. 93 (1988), no. 1, 185–223.
[27] E. Hrushovski, B. Martin, S. Rideau, and R. Cluckers, Definable equivalence relations and zeta
functions of groups (preprint) (2015). arXiv:0701011.
[28] I. Ilani, Zeta functions related to the group SL2 (Zp ), Israel J. Math. 109 (1999), 157–172.
[29] N. M. Katz, Review of `-adic cohomology, Motives (Seattle, WA, 1991), 1994, pp. 21–30.
[30] B. Klopsch and C. Voll, Zeta functions of three-dimensional p-adic Lie algebras, Math. Z. 263 (2009),
no. 1, 195–210.
[31] O. Kuzmich, Graded nilpotent Lie algebras in low dimensions, Lobachevskii J. Math. 3 (1999),
147–184 (electronic). Towards 100 years after Sophus Lie (Kazan, 1998).
[32] V. V. Morozov, Classification of nilpotent Lie algebras of sixth order, Izv. Vysš. Učebn. Zaved.
Matematika 1958 (1958), no. 4 (5), 161–171.
[33] J. Nakagawa, Orders of a quartic field, Mem. Amer. Math. Soc. 122 (1996), no. 583, viii+75.
[34] C. Nunley and A. Magid, Simple representations of the integral Heisenberg group, Classical groups
and related topics (Beijing, 1987), 1989, pp. 89–96.
[35] T. Rossmann, Computing topological zeta functions of groups, algebras, and modules, I, Proc. Lond.
Math. Soc. (3) 110 (2015), no. 5, 1099–1134.
[36]
, Computing topological zeta functions of groups, algebras, and modules, II, J. Algebra 444
(2015), 567–605.
[37]
, Stability results for local zeta functions of groups and related structures (preprint) (2015).
arXiv:1504.04164.
, Topological representation zeta functions of unipotent groups, J. Algebra 448 (2016), 210–
[38]
237.
[39]
, Zeta, version 0.3, 2016. See http://www.math.uni-bielefeld.de/~rossmann/Zeta/.
[40] J.-P. Serre, Lectures on NX (p), Chapman & Hall/CRC Research Notes in Mathematics, vol. 11,
CRC Press, Boca Raton, FL, 2012.
37
[41] R. Snocken, Zeta functions of groups and rings. PhD Thesis, University of Southampton, 2012. See
http://eprints.soton.ac.uk/id/eprint/372833.
[42] L. Solomon, Zeta functions and integral representation theory, Advances in Math. 26 (1977), no. 3,
306–326.
[43] A. Stasinski and C. Voll, Representation zeta functions of nilpotent groups and generating functions
for Weyl groups of type B, Amer. J. Math. 136 (2014), no. 2, 501–550.
[44] W. A. Stein et al., Sage Mathematics Software (Version 6.9), The Sage Development Team, 2015.
Available from http://www.sagemath.org/.
[45] G. Taylor, Zeta functions of algebras and resolution of singularities, Ph.D. Thesis, 2001.
[46] C. Voll, Functional equations for zeta functions of groups and rings, Ann. of Math. (2) 172 (2010),
no. 2, 1181–1218.
[47]
, Local functional equations for submodule zeta functions associated to nilpotent algebras of
endomorphisms (in preparation) (2016).
[48] J. White, Zeta functions of groups. DPhil Thesis, University of Oxford, 2000.
[49] L. Woodward, Zeta functions of groups: computer calculations and functional equations. DPhil Thesis,
University of Oxford, 2005.
[50]
, Zeta functions of Lie rings of upper-triangular matrices, J. Lond. Math. Soc. (2) 77 (2008),
no. 1, 69–82.
38
| 4math.GR |
ZU064-05-FPR
paper-jfp
6 February 2018
3:38
Under consideration for publication in J. Functional Programming
1
arXiv:1406.5106v1 [cs.PL] 19 Jun 2014
Pushdown flow analysis with abstract garbage
collection
J. IAN JOHNSON
Northeastern University
ILYA SERGEY
IMDEA Software Institute
CHRISTOPHER EARL
University of Utah
MATTHEW MIGHT
University of Utah
DAVID VAN HORN
University of Maryland
Abstract
In the static analysis of functional programs, pushdown flow analysis and abstract garbage collection
push the boundaries of what we can learn about programs statically. This work illuminates and poses
solutions to theoretical and practical challenges that stand in the way of combining the power of these
techniques. Pushdown flow analysis grants unbounded yet computable polyvariance to the analysis
of return-flow in higher-order programs. Abstract garbage collection grants unbounded polyvariance
to abstract addresses which become unreachable between invocations of the abstract contexts in
which they were created. Pushdown analysis solves the problem of precisely analyzing recursion in
higher-order languages; abstract garbage collection is essential in solving the “stickiness” problem.
Alone, our benchmarks demonstrate that each method can reduce analysis times and boost precision
by orders of magnitude.
We combine these methods. The challenge in marrying these techniques is not subtle: computing
the reachable control states of a pushdown system relies on limiting access during transition to the
top of the stack; abstract garbage collection, on the other hand, needs full access to the entire stack to
compute a root set, just as concrete collection does. Conditional pushdown systems were developed
for just such a conundrum, but existing methods are ill-suited for the dynamic nature of garbage
collection.
We show fully precise and approximate solutions to the feasible paths problem for pushdown
garbage-collecting control-flow analysis. Experiments reveal synergistic interplay between garbage
collection and pushdown techniques, and the fusion demonstrates “better-than-both-worlds” precision.
ZU064-05-FPR
paper-jfp
2
6 February 2018
3:38
J.I. Johnson, I. Sergey, C. Earl, M. Might, and D. Van Horn
(define (id x) x)
(define (f n)
(cond [(<= n 1)
[else
1]
(* n (f (- n 1)))]))
(define (g n)
(cond [(<= n 1)
[else
1]
(+ (* n n) (g (- n 1)))]))
(print (+ ((id f) 3) ((id g) 4)))
Fig. 1: A small example to illuminate the strengths and weaknesses of both pushdown
analysis and abstract garbage collection.
1 Introduction
The development of a context-free1 approach to control-flow analysis (CFA2) by Vardoulakis and Shivers (2010) provoked a shift in the static analysis of higher-order programs. Prior to CFA2, a precise analysis of recursive behavior had been a challenge—even
though flow analyses have an important role to play in optimization for functional languages, such as flow-driven inlining (Might and Shivers 2006a), interprocedural constant
propagation (Shivers 1991) and type-check elimination (Wright and Jagannathan 1998).
While it had been possible to statically analyze recursion soundly, CFA2 made it possible
to analyze recursion precisely by matching calls and returns without approximating the
stack as k-CFA does. The approximation is only in the binding structure, and not the control
structure of the program. In its pursuit of recursion, clever engineering steered CFA2 to
a theoretically intractable complexity, though in practice it performs well. Its payoff is
significant reductions in analysis time as a result of corresponding increases in precision.
For a visual measure of the impact, Figure 2 renders the abstract transition graph (a
model of all possible traces through the program) for the toy program in Figure 1. For this
example, pushdown analysis eliminates spurious return-flow from the use of recursion. But,
recursion is just one problem of many for flow analysis. For instance, pushdown analysis
still gets tripped up by the spurious cross-flow problem; at calls to (id f) and (id g) in
the previous example, it thinks (id g) could be f or g. CFA2 is not confused in this due
to its precise stack frames, but can be confused by unreachable heap-allocated bindings.
Powerful techniques such as abstract garbage collection (Might and Shivers 2006b) were
developed to address the cross-flow problem (here in a way complementary to CFA2’s
stack frames). The cross-flow problem arises because monotonicity prevents revoking a
judgment like “procedure f flows to x,” or “procedure g flows to x,” once it’s been made.
1
As in context-free language, not a context-insensitive analysis.
ZU064-05-FPR
paper-jfp
6 February 2018
3:38
PUSHDOWN FLOW ANALYSIS WITH ABSTRACT GARBAGE COLLECTION
3
398
250
76
401
607
344
100
369
414
54
304
287
350
218
407
360
453
133
150
79
647
45
47
300
103
186
508
635
382
377
562
495
433
346
58
267
310
122
138
392
145
560
273
329
499
163
340
482
90
406
615
447
68
245
484
596
28
35
338
628
175
631
408
581
332
48
586
40
569
25
526
254
496
306
480
253
603
271
176
233
216
371
529
439
405
213
53
98
134
400
390
200
509
627
217
142
610
410
555
74
244
327
208
240
549
374
432
566
351
320
323
189
541
72
540
474
352
363
42
278
417
430
147
459
489
110
165
367
85
302
498
152
514
305
373
117
201
313
265
99
595
234
604
516
316
443
416
94
5
10
18
543
177
641
112
570
241
551
275
146
512
299
559
645
281
486
108
246
1
17
121
101
455
397
527
450
16
501
442
257
404
523
222
29
599
524
578
162
141
183
235
106
465
358
425
381
114
609
36
468
343
359
477
620
118
418
8
411
536
161
533
178
312
308
435
545
553
283
212
303
84
11
368
438
120
574
629
269
193
479
298
485
567
552
136
614
573
585
347
571
650
255
6
292
554
587
184
215
296
575
148
111
196
317
105
181
621
471
249
355
135
230
137
214
339
258
159
78
606
630
589
409
648
561
440
594
331
263
535
59
451
309
353
539
153
113
444
31
481
372
92
365
494
598
396
488
619
584
556
104
262
341
590
144
274
280
210
384
168
179
272
37
426
89
593
314
95
649
9
515
402
518
46
119
534
276
356
232
446
39
399
15
470
64
483
259
67
191
502
82
558
140
592
188
487
223
293
131
531
572
225
419
463
33
307
23
199
375
57
12
333
457
335
125
167
154
107
319
361
182
14
321
203
616
66
424
376
532
591
88
205
221
170
624
530
83
506
362
456
563
568
420
537
348
580
441
513
602
458
517
421
285
452
336
622
102
252
123
172
242
493
43
600
227
207
391
194
130
427
4
27
55
236
507
143
370
185
237
467
431
260
318
180
224
187
87
290
73
129
546
38
155
49
284
34
202
192
608
190
434
206
415
448
195
636
50
550
460
156
633
476
466
383
2
77
472
63
21
229
605
289
646
115
65
548
547
7
510
387
132
62
149
75
286
139
228
324
625
454
357
279
251
174
490
97
116
429
634
395
579
577
126
297
198
542
60
330
128
612
231
93
52
393
473
642
389
294
436
295
412
475
51
626
525
157
30
582
328
564
337
91
462
504
56
311
261
644
322
70
96
69
379
597
449
611
437
342
288
538
617
544
380
268
211
623
291
325
301
209
428
26
557
423
601
264
413
22
497
588
20
520
19
13
491
500
528
86
511
386
394
226
3
173
81
197
632
613
366
469
270
505
640
326
204
239
583
464
354
219
461
521
277
315
109
24
522
124
334
248
61
639
345
637
71
403
618
638
422
151
385
643
576
32
378
364
349
164
243
565
256
445
80
44
158
282
492
220
388
160
238
127
503
41
519
247
171
166
266
478
169
(1) without pushdown analysis or abstract GC: 653 states
85
60
16
123
110
48
100
50
84
70
94
119
42
59
112
7
43
67
96
105
80
64
4
102
109
69
76
47
18
20
66
82
92
88
65
133
115
56
54
136
45
121
1
74
22
120
78
40
128
57
124
134
19
97
68
55
95
28
93
98
8
139
49
73
26
135
63
17
75
113
13
32
108
11
99
104
36
10
116
3
27
114
77
101
23
2
106
90
33
87
118
107
12
62
9
51
132
15
91
37
81
52
21
5
111
130
29
125
53
83
41
103
14
38
126
58
25
129
31
89
86
72
61
79
39
137
131
24
117
6
71
46
35
138
44
122
30
34
127
(2) with pushdown only: 139 states
3
84
50
94
21
19
102
10
36
55
28
65
74
34
97
98
13
44
24
43
104
72
70
100
1
48
85
78
82
41
103
47
12
18
11
8
15
64
40
52
73
53
58
45
93
81
96
79
105
87
16
59
61
20
60
62
17
88
71
6
99
5
57
83
63
49
7
26
38
32
80
14
66
42
86
69
101
22
35
33
27
54
76
4
67
37
75
89
77
29
9
51
31
46
92
2
25
23
56
91
39
68
90
30
95
(3) with GC only: 105 states
31
26
64
8
47
56
75
24
68
50
28
73
19
15
34
14
63
76
32
20
57
16
38
41
59
52
33
49
3
58
74
12
9
10
39
42
22
5
7
36
48
54
51
35
72
66
44
46
6
43
70
65
18
53
2
11
23
40
77
37
62
29
17
13
61
27
45
69
60
55
4
30
67
25
21
1
71
(4) with pushdown analysis and abstract GC: 77 states
Fig. 2: We generated an abstract transition graph for the same program from Figure 1 four
times: (1) without pushdown analysis or abstract garbage collection; (2) with only abstract
garbage collection; (3) with only pushdown analysis; (4) with both pushdown analysis and
abstract garbage collection. With only pushdown or abstract GC, the abstract transition
graph shrinks by an order of magnitude, but in different ways. The pushdown-only analysis
is confused by variables that are bound to several different higher-order functions, but
for short durations. The abstract-GC-only is confused by non-tail-recursive loop structure.
With both techniques enabled, the graph shrinks by nearly half yet again and fully recovers
the control structure of the original program.
ZU064-05-FPR
paper-jfp
4
6 February 2018
3:38
J.I. Johnson, I. Sergey, C. Earl, M. Might, and D. Van Horn
In fact, abstract garbage collection, by itself, also delivers significant improvements to
analytic speed and precision in many benchmarks. (See Figure 2 again for a visualization
of that impact.)
It is natural to ask: can abstract garbage collection and pushdown analysis work together? Can their strengths be multiplied? At first, the answer appears to be a disheartening “No.”
1.1 The problem: The whole stack versus just the top
Abstract garbage collection seems to require more than pushdown analysis can decidably
provide: access to the full stack. Abstract garbage collection, like its name implies, discards unreachable values from an abstract store during the analysis. Like concrete garbage
collection, abstract garbage collection also begins its sweep with a root set, and like concrete garbage collection, it must traverse the abstract stack to compute that root set. But,
pushdown systems are restricted to viewing the top of the stack (or a bounded depth)—a
condition violated by this traversal.
Fortunately, abstract garbage collection does not need to arbitrarily modify the stack. It
only needs to know the root set of addresses in the stack. This kind of system has been
studied before in the context of compilers that build a symbol table (a so-called “oneway stack automaton” (Ginsburg et al. 1967)),in the context of first-order model-checking
(pushdown systems with checkpoints (Esparza et al. 2003)),and also in the context of
points-to analysis for Java (conditional weighted pushdown systems (CWPDS) (Li and
Ogawa 2010)). We borrow the definition of (unweighted) conditional pushdown system
(CPDS) in this work, though our analysis does not take CPDSs as inputs.
Higher-order flow analyses typically do not take a control-flow graph, or similar preabstracted object, as input and produce an annotated graph as output. Instead, they take a
program as input and “run it on all possible inputs” (abstractly) to build an approximation
of the language’s reduction relation (semantics), specialized to the given program. This
semantics may be non-standard in such a way that extra-semantic information might be
accumulated for later analyses’ consumption. The important distinction between higherorder and first-order analyses is that the model to analyze is built during the analysis,
which involves interpreting the program (abstractly).
When a language’s semantics treats the control stack as an actual stack, i.e., it does not
have features such as first-class continuations, an interpreter can be split into two parts: a
function that takes the current state and returns all next states along with a pushed activation
frame or a marker that the stack is unchanged; and a function that takes the current state, a
possible “top frame” of the stack, and returns the next states after popping this frame. This
separation is crucial for an effective algorithm, since pushed frames are understood from
program text, and popped frames need only be enumerated from a (usually small) set that
we compute along the way.
Control-state reachability for the straightforward formulation of stack introspection ends
up being uncomputable. Conditional pushdown systems introduce a relatively weak regularity constraint on transitions’ introspection: a CPDS may match the current stack against
a choice of finitely many regular languages of stacks in order to transition from one state to
the next along with the stack action. The general solutions to feasible paths in conditional
ZU064-05-FPR
paper-jfp
6 February 2018
3:38
PUSHDOWN FLOW ANALYSIS WITH ABSTRACT GARBAGE COLLECTION
5
pushdown systems enumerate all languages of stacks that a transition may be conditioned
on. This strategy is a non-starter for garbage collection, since we delineate stacks by the
addresses they keep live; this is exponential in the number of addresses. The abstraction
step that finitizes the address space is what makes the problem fall within the realm of
CPDSs, even if the target is so big it barely fits. But abstract garbage collection is special
— we can compute which languages of stacks we need to check against, given the current
state of the analysis. It is therefore possible to fuse the full benefits of abstract garbage
collection with pushdown analysis. The dramatic reduction in abstract transition graph
size from the top to the bottom in Figure 2 (and echoed by later benchmarks) conveys the
impact of this fusion.
Secondary motivations There are four secondary motivations for this work:
1.
2.
3.
4.
bringing context-sensitivity to pushdown analysis;
exposing the context-freedom of the analysis;
enabling pushdown analysis without continuation-passing style; and
defining an alternative algorithm for computing pushdown analysis, introspectively
or otherwise.
In CFA2, monovariant (0CFA-like) context-sensitivity is etched directly into the abstract
“local” semantics, which is in turn phrased in terms of an explicit (imperative) summarization algorithm for a partitioned continuation-passing style. Our development exposes
the classical parameters (exposed as allocation functions in a semantics) that allow one to
tune the context-sensitivity and polyvariance (accomplishing (1)), thanks to the semantics
of the analysys formulated in the form of an “abstracted abstract machine” (Van Horn and
Might 2012).
In addition, the context-freedom of CFA2 is buried implicitly inside an imperative summarization algorithm. No pushdown system or context-free grammar is explicitly identified. Thus, a motivating factor for our work was to make the pushdown system in CFA2
explicit, and to make the control-state reachability algorithm purely functional (accomplishing (2)).
A third motivation was to show that a transformation to continuation-passing style is
unnecessary for pushdown analysis. In fact, pushdown analysis is arguably more natural
over direct-style programs. By abstracting all machine components except for the program
stack, it converts naturally and readily into a pushdown system (accomplishing (3)). In
his dissertation, Vardoulakis showed a direct-style version of CFA2 that exploits the metalanguage’s runtime stack to get precise call-return matching. The approach is promising,
but its correctness remains unproven, and it does not apply to generic pushdown systems.
Finally, to bring much-needed clarity to algorithmic formulation of pushdown analysis,
we have included an appendix containing a reference implementation in Haskell (accomplishing (4)). We have kept the code as close in form to the mathematics as possible, so
that where concessions are made to the implementation, they are obvious.
1.2 Overview
We first review preliminaries to set a consistent feel for terminology and notation, particularly with respect to pushdown systems. The derivation of the analysis begins with a con-
ZU064-05-FPR
paper-jfp
6
6 February 2018
3:38
J.I. Johnson, I. Sergey, C. Earl, M. Might, and D. Van Horn
crete CESK-machine-style semantics for A-Normal Form λ -calculus. The next step is an
infinite-state abstract interpretation, constructed by bounding the C(ontrol), E(nvironment)
and S(tore) portions of the machine. Uncharacteristically, we leave the stack component—
the K(ontinuation)—unbounded.
A shift in perspective reveals that this abstract interpretation is a pushdown system.
We encode it as a pushdown automaton explicitly, and pose control state reachability as
a decidable language intersection problem. We then extract a rooted pushdown system
from the pushdown automaton. For completeness, we fully develop pushdown analysis for
higher-order programs, including an efficient algorithm for computing reachable control
states. We go further by characterizing complexity and demonstrating the approximations
necessary to get to a polynomial-time algorithm.
We then introduce abstract garbage collection and quickly find that it violates the pushdown model with its traversals of the stack. To prove the decidability of control-state
reachability, we formulate introspective pushdown systems, and recast abstract garbage
collection within this framework. We then review that control-state reachability is decidable for introspective pushdown systems as well when subjected to a straightforward
regularity constraint.
We conclude with an implementation and empirical evaluation that shows strong synergies between pushdown analysis and abstract garbage collection, including significant
reductions in the size of the abstract state transition graph.
1.3 Contributions
We make the following contributions:
1. Our primary contribution is an online decision procedure for reachability in introspective pushdown systems, with a more efficient specialization to abstract garbage
collection.
2. We show that classical notions of context-sensitivity, such as k-CFA and poly/CFA,
have direct generalizations in a pushdown setting. CFA2 was presented as a monovariant analysis,2 whereas we show polyvariance is a natural extension.
3. We make the context-free aspect of CFA2 explicit: we clearly define and identify
the pushdown system. We do so by starting with a classical CESK machine and
systematically abstracting until a pushdown system emerges. We also remove the orthogonal frame-local-bindings aspect of CFA2, so as to focus solely on the pushdown
nature of the analysis.
4. (*) We remove the requirement for a global CPS-conversion by synthesizing the
analysis directly for direct-style (in the form of A-normal form lambda-calculus —
a local transformation).
5. We empirically validate claims of improved precision on a suite of benchmarks.
We find synergies between pushdown analysis and abstract garbage collection that
makes the whole greater that the sum of its parts.
2
Monovariance refers to an abstraction that groups all bindings to the same variable together: there
is one abstract variant for all bindings to each variable.
ZU064-05-FPR
paper-jfp
6 February 2018
3:38
PUSHDOWN FLOW ANALYSIS WITH ABSTRACT GARBAGE COLLECTION
7
6. We provide a mirror of the major formal development as working Haskell code in the
appendix. This code illuminates dark corners of pushdown analysis and it provides a
concise formal reference implementation.
(*) The CPS requirement distracts from the connection between continuations and stacks.
We do not discuss call/cc in detail, since we believe there are no significant barriers to
adapting the techniques of Vardoulakis and Shivers (2011) to the direct-style setting, given
related work in Johnson and Van Horn (2013). Languages with exceptions fit within the
pushdown model since a throw can be modeled as “pop until first catch.”
2 Pushdown Preliminaries
The literature contains many equivalent definitions of pushdown machines, so we adapt
our own definitions from Sipser (2005). Readers familiar with pushdown theory may wish
to skip ahead.
2.1 Syntactic sugar
When a triple (x, `, x0 ) is an edge in a labeled graph:
`
x x0 ≡ (x, `, x0 ).
Similarly, when a pair (x, x0 ) is a graph edge:
x x0 ≡ (x, x0 ).
We use both string and vector notation for sequences:
a1 a2 . . . an ≡ ha1 , a2 , . . . , an i ≡ ~a.
2.2 Stack actions, stack change and stack manipulation
Stacks are sequences over a stack alphabet Γ. To reason about stack manipulation concisely, we first turn stack alphabets into “stack-action” sets; each character represents a
change to the stack: push, pop or no change.
For each character γ in a stack alphabet Γ, the stack-action set Γ± contains a push
character γ+ ; a pop character γ− ; and a no-stack-change indicator, ε:
g ∈ Γ± ::= ε
[stack unchanged]
| γ+ for each γ ∈ Γ
[pushed γ]
| γ− for each γ ∈ Γ
[popped γ].
In this paper, the symbol g represents some stack action.
When we develop introspective pushdown systems, we are going to need formalisms
for easily manipulating stack-action strings and stacks. Given a string of stack actions, we
can compact it into a minimal string describing net stack change. We do so through the
operator b·c : Γ∗± → Γ∗± , which cancels out opposing adjacent push-pop stack actions:
b~g γ+ γ− ~g 0 c = b~g ~g 0 c
b~g ε ~g 0 c = b~g ~g 0 c,
ZU064-05-FPR
paper-jfp
8
6 February 2018
3:38
J.I. Johnson, I. Sergey, C. Earl, M. Might, and D. Van Horn
so that b~gc = ~g, if there are no cancellations to be made in the string ~g.
We can convert a net string back into a stack by stripping off the push symbols with the
stackify operator, d·e : Γ∗± * Γ∗ :
(n)
dγ+ γ+0 . . . γ+ e = hγ (n) , . . . , γ 0 , γi,
and for convenience, [~g] = db~gce. Notice the stackify operator is defined for strings containing only push actions.
2.3 Pushdown systems
A pushdown system is a triple M = (Q, Γ, δ ) where:
1. Q is a finite set of control states;
2. Γ is a stack alphabet; and
3. δ ⊆ Q × Γ± × Q is a transition relation.
The set Q × Γ∗ is called the configuration-space of this pushdown system. We use PDS to
denote the class of all pushdown systems.
For the following definitions, let M = (Q, Γ, δ ).
• The labeled transition relation (7−→M ) ⊆ (Q × Γ∗ ) × Γ± × (Q × Γ∗ ) determines
whether one configuration may transition to another while performing the given
stack action:
ε
(q,~γ) 7−→ (q0 ,~γ) iff q q0 ∈ δ
ε
[no change]
M
γ−
γ−
(q, γ : ~γ) 7−→ (q0 ,~γ) iff q q0 ∈ δ
[pop]
M
γ+
γ+
(q,~γ) 7−→ (q0 , γ : ~γ) iff q q0 ∈ δ
[push].
M
• If unlabelled, the transition relation (7−→) checks whether any stack action can
enable the transition:
g
c 7−→ c0 iff c 7−→ c0 for some stack action g.
M
M
• For a string of stack actions g1 . . . gn :
g ...gn
g
g
gn−1
gn
M
M
M
M
M
1
2
c0 −
7 1 → cn iff c0 7−→
c1 7−→
· · · 7−→ cn−1 7−→ cn ,
for some configurations c0 , . . . , cn .
• For the transitive closure:
~g
c 7−→∗ c0 iff c 7−→ c0 for some action string ~g.
M
M
Note Some texts define the transition relation δ so that δ ⊆ Q × Γ × Q × Γ∗ . In these
texts, (q, γ, q0 ,~γ) ∈ δ means, “if in control state q while the character γ is on top, pop the
stack, transition to control state q0 and push ~γ.” Clearly, we can convert between these two
representations by introducing extra control states to our representation when it needs to
push multiple characters.
ZU064-05-FPR
paper-jfp
6 February 2018
3:38
PUSHDOWN FLOW ANALYSIS WITH ABSTRACT GARBAGE COLLECTION
9
2.4 Rooted pushdown systems
A rooted pushdown system is a quadruple (Q, Γ, δ , q0 ) in which (Q, Γ, δ ) is a pushdown
system and q0 ∈ Q is an initial (root) state. RPDS is the class of all rooted pushdown
systems. For a rooted pushdown system M = (Q, Γ, δ , q0 ), we define the reachable-fromroot transition relation:
g
g
c7−
−→
→ c0 iff (q0 , hi) 7−→∗ c and c 7−→ c0 .
M
M
M
In other words, the root-reachable transition relation also makes sure that the root control
state can actually reach the transition.
We overload the root-reachable transition relation to operate on control states:
g
g
M
M
q 7−
−→
→ q0 iff (q,~γ) 7−
−→
→ (q0 ,~γ 0 ) for some stacks ~γ,~γ 0 .
For both root-reachable relations, if we elide the stack-action label, then, as in the un-rooted
case, the transition holds if there exists some stack action that enables the transition:
g
−→
→ q0 for some action g.
q 7−
−→
→ q0 iff q 7−
M
M
2.5 Computing reachability in pushdown systems
A pushdown flow analysis can be construed as computing the root-reachable subset of
control states in a rooted pushdown system, M = (Q, Γ, δ , q0 ):
n
o
q : q0 7−
−→
→ q .
M
Reps et. al and many others provide a straightforward “summarization” algorithm to compute this set (Bouajjani et al. 1997; Kodumal and Aiken 2004; Reps 1998; Reps et al.
2005). We will develop a complete alternative to summarization, and then instrument this
development for introspective pushdown systems. Summarization builds two large tables:
• One maps “calling contexts” to “return sites” (AKA “local continuations”) so that a
returning function steps to all the places it must return to.
• The other maps “calling contexts” to “return states,” so that any place performing a
call with an already analyzed calling context can jump straight to the returns.
This setup requires intimate knowledge of the language in question for where continuations
should be segmented to be “local” and is strongly tied to function call and return. Our
algorithm is based on graph traversals of the transition relation for a generic pushdown
system. It requires no specialized knowledge of the analyzed language, and it avoids the
memory footprint of summary tables.
2.6 Pushdown automata
A pushdown automaton is an input-accepting generalization of a rooted pushdown system, a 7-tuple (Q, Σ, Γ, δ , q0 , F,~γ) in which:
1. Σ is an input alphabet;
2. δ ⊆ Q × Γ± × (Σ ∪ {ε}) × Q is a transition relation;
ZU064-05-FPR
paper-jfp
10
6 February 2018
3:38
J.I. Johnson, I. Sergey, C. Earl, M. Might, and D. Van Horn
3. F ⊆ Q is a set of accepting states; and
4. ~γ ∈ Γ∗ is the initial stack.
We use PDA to denote the class of all pushdown automata.
Pushdown automata recognize languages over their input alphabet. To do so, their transition relation may optionally consume an input character upon transition. Formally, a PDA
M = (Q, Σ, Γ, δ , q0 , F,~γ) recognizes the language L (M) ⊆ Σ∗ :
ε ∈ L (M) if q0 ∈ F
aw ∈ L (M) if δ (q0 , γ+ , a, q0 ) and w ∈ L (Q, Σ, Γ, δ , q0 , F, γ : ~γ)
aw ∈ L (M) if δ (q0 , ε, a, q0 ) and w ∈ L (Q, Σ, Γ, δ , q0 , F,~γ)
aw ∈ L (M) if δ (q0 , γ− , a, q0 ) and w ∈ L (Q, Σ, Γ, δ , q0 , F,~γ 0 )
where ~γ = hγ, γ2 , . . . , γn i and ~γ 0 = hγ2 , . . . , γn i,
where a is either the empty string ε or a single character.
2.7 Nondeterministic finite automata
In this work, we will need a finite description of all possible stacks at a given control state
within a rooted pushdown system. We will exploit the fact that the set of stacks at a given
control point is a regular language. Specifically, we will extract a nondeterministic finite
automaton accepting that language from the structure of a rooted pushdown system. A
nondeterministic finite automaton (NFA) is a quintuple M = (Q, Σ, δ , q0 , F):
• Q is a finite set of control states;
• Σ is an input alphabet;
• δ ⊆ Q × (Σ ∪ {ε}) × Q is a transition relation.
• q0 is a distinguished start state.
• F ⊆ Q is a set of accepting states.
We denote the class of all NFAs as NFA.
3 Setting: A-Normal Form λ -Calculus
Since our goal is analysis of higher-order languages, we operate on the λ -calculus. To
simplify presentation of the concrete and abstract semantics, we choose A-Normal Form
λ -calculus. (This is a strictly cosmetic choice: all of our results can be replayed mutatis
mutandis in the standard direct-style setting as well. This differs from CFA2’s requirement
of CPS, since ANF can be applied locally whereas CPS requires a global transformation.)
ANF enforces an order of evaluation and it requires that all arguments to a function be
ZU064-05-FPR
paper-jfp
6 February 2018
3:38
PUSHDOWN FLOW ANALYSIS WITH ABSTRACT GARBAGE COLLECTION
c ∈ Conf = Exp × Env × Store × Kont
11
[configurations]
ρ ∈ Env = Var * Addr
[environments]
σ ∈ Store = Addr → Clo
[stores]
clo ∈ Clo = Lam × Env
[closures]
κ ∈ Kont = Frame∗
[continuations]
φ ∈ Frame = Var × Exp × Env
[stack frames]
a ∈ Addr is an infinite set of addresses
[addresses].
Fig. 3: The concrete configuration-space.
atomic:
e ∈ Exp ::= (let ((v call)) e)
[non-tail call]
| call
[tail call]
| æ
[return]
f , æ ∈ Atom ::= v | lam
[atomic expressions]
lam ∈ Lam ::= (λ (v) e)
[lambda terms]
call ∈ Call ::= ( f æ)
[applications]
v ∈ Var is a set of identifiers
[variables].
We use the CESK machine of Felleisen and Friedman (1987) to specify a small-step
semantics for ANF. The CESK machine has an explicit stack, and under a structural abstraction, the stack component of this machine directly becomes the stack component of a
pushdown system. The set of configurations (Conf ) for this machine has the four expected
components (Figure 3).
3.1 Semantics
To define the semantics, we need five items:
1. I : Exp → Conf injects an expression into a configuration:
c0 = I (e) = (e, [], [], hi).
2. A : Atom × Env × Store * Clo evaluates atomic expressions:
A (lam, ρ, σ ) = (lam, ρ)
A (v, ρ, σ ) = σ (ρ(v))
[closure creation]
[variable look-up].
3. (⇒) ⊆ Conf × Conf transitions between configurations. (Defined below.)
4. E : Exp → P (Conf ) computes the set of reachable machine configurations for a
given program:
E (e) = {c : I (e) ⇒∗ c} .
ZU064-05-FPR
paper-jfp
6 February 2018
12
3:38
J.I. Johnson, I. Sergey, C. Earl, M. Might, and D. Van Horn
5. alloc : Var × Conf → Addr chooses fresh store addresses for newly bound variables.
The address-allocation function is an opaque parameter in this semantics, so that the
forthcoming abstract semantics may also parameterize allocation. The nondeterministic nature of the semantics makes any choice of alloc sound (Might and Manolios
2009). This parameterization provides the knob to tune the polyvariance and contextsensitivity of the resulting analysis. For the sake of defining the concrete semantics,
letting addresses be natural numbers suffices. The allocator can then choose the
lowest unused address:
Addr = N
alloc(v, (e, ρ, σ , κ)) = 1 + max(dom(σ )).
Transition relation To define the transition c ⇒ c0 , we need three rules. The first rule
handles tail calls by evaluating the function into a closure, evaluating the argument into a
value and then moving to the body of the closure’s λ -term:
c0
c
}|
{
}|
{ z
z
([[( f æ)]], ρ, σ , κ) ⇒ (e, ρ 00 , σ 0 , κ) , where
0
([[(λ (v) e)]], ρ ) = A ( f , ρ, σ )
a = alloc(v, c)
ρ 00 = ρ 0 [v 7→ a]
σ 0 = σ [a 7→ A (æ, ρ, σ )].
Non-tail calls push a frame onto the stack and evaluate the call:
c0
c
z
}|
{ z
}|
{
([[(let ((v call)) e)]], ρ, σ , κ) ⇒ (call, ρ, σ , (v, e, ρ) : κ) .
Function return pops a stack frame:
c
c0
z
}|
{ z
}|
{
(æ, ρ, σ , (v, e, ρ 0 ) : κ) ⇒ (e, ρ 00 , σ 0 , κ) , where
a = alloc(v, c)
ρ 00 = ρ 0 [v 7→ a]
σ 0 = σ [a 7→ A (æ, ρ, σ )].
4 An Infinite-State Abstract Interpretation
Our first step toward a static analysis is an abstract interpretation into an infinite statespace. To achieve a pushdown analysis, we simply abstract away less than we normally
would. Specifically, we leave the stack height unbounded.
Figure 4 details the abstract configuration-space. To synthesize it, we force addresses to
be a finite set, but crucially, we leave the stack untouched. When we compact the set of
addresses into a finite set, the machine may run out of addresses to allocate, and when it
does, the pigeon-hole principle will force multiple closures to reside at the same address.
As a result, to remain sound we change the range of the store to become a power set in the
abstract configuration-space. The abstract transition relation has components analogous to
those from the concrete semantics:
ZU064-05-FPR
paper-jfp
6 February 2018
3:38
PUSHDOWN FLOW ANALYSIS WITH ABSTRACT GARBAGE COLLECTION
d × Store
d
[ = Exp × Env
[ × Kont
ĉ ∈ Conf
[configurations]
d = Var * Addr
[
ρ̂ ∈ Env
d
[ → P Clo
[ = Addr
σ̂ ∈ Store
[environments]
[stores]
c ∈ Clo
d = Lam × Env
d
clo
[closures]
d = Frame
\
κ̂ ∈ Kont
∗
d
\ = Var × Exp × Env
φ̂ ∈ Frame
[ is a finite set of addresses
â ∈ Addr
13
[continuations]
[stack frames]
[addresses].
Fig. 4: The abstract configuration-space.
[ pairs an expression
Program injection The abstract injection function Iˆ : Exp → Conf
with an empty environment, an empty store and an empty stack to create the initial abstract
configuration:
ĉ0 = Iˆ(e) = (e, [], [], hi).
Atomic expression evaluation The abstract atomic expression evaluator, Aˆ : Atom ×
d returns the value of an atomic expression in the context of an
d × Store
[ → P(Clo),
Env
environment and a store; it returns a set of abstract closures:
Aˆ(lam, ρ̂, σ̂ ) = {(lam, ρ̂)}
Aˆ(v, ρ̂, σ̂ ) = σ̂ (ρ̂(v))
[closure creation]
[variable look-up].
[ ) returns
Reachable configurations The abstract program evaluator Eˆ : Exp → P(Conf
all of the configurations reachable from the initial configuration:
Eˆ (e) = ĉ : Iˆ(e) c
⇒∗ ĉ .
Because there are an infinite number of abstract configurations, a naı̈ve implementation of
this function may not terminate. Pushdown analysis provides a way of precisely computing
this set and both finitely and compactly representing the result.
[ × Conf
[ has three rules,
Transition relation The abstract transition relation (c
⇒) ⊆ Conf
one of which has become nondeterministic. A tail call may fork because there could be
multiple abstract closures that it is invoking:
ĉ
ĉ0
}|
{
}|
{ z
z
([[( f æ)]], ρ̂, σ̂ , κ̂) c
⇒ (e, ρ̂ 00 , σ̂ 0 , κ̂) , where
([[(λ (v) e)]], ρ̂ 0 ) ∈ Aˆ( f , ρ̂, σ̂ )
[ ĉ)
â = alloc(v,
ρ̂ 00 = ρ̂ 0 [v 7→ â]
σ̂ 0 = σ̂ t [â 7→ Aˆ(æ, ρ̂, σ̂ )].
ZU064-05-FPR
paper-jfp
6 February 2018
14
3:38
J.I. Johnson, I. Sergey, C. Earl, M. Might, and D. Van Horn
We define all of the partial orders shortly, but for stores:
(σ̂ t σ̂ 0 )(â) = σ̂ (â) ∪ σ̂ 0 (â).
A non-tail call pushes a frame onto the stack and evaluates the call:
ĉ0
ĉ
}|
{ z
}|
{
z
([[(let ((v call)) e)]], ρ̂, σ̂ , κ̂) c
⇒ (call, ρ̂, σ̂ , (v, e, ρ̂) : κ̂) .
A function return pops a stack frame:
ĉ
ĉ0
}|
{ z
}|
{
z
(æ, ρ̂, σ̂ , (v, e, ρ̂ 0 ) : κ̂) c
⇒ (e, ρ̂ 00 , σ̂ 0 , κ̂) , where
[ ĉ)
â = alloc(v,
ρ̂ 00 = ρ̂ 0 [v 7→ â]
σ̂ 0 = σ̂ t [â 7→ Aˆ(æ, ρ̂, σ̂ )].
Allocation: Polyvariance and context-sensitivity In the abstract semantics, the abstract
[ : Var × Conf
[ → Addr
[ determines the polyvariance of the analysis.
allocation function alloc
In a control-flow analysis, polyvariance literally refers to the number of abstract addresses
(variants) there are for each variable. An advantage of this framework over CFA2 is that
varying this abstract allocation function instantiates pushdown versions of classical flow
analyses. All of the following allocation approaches can be used with the abstract semantics. Note, though only a technical detail, that the concrete address space and allocation
would change as well for the abstraction function to still work. The abstract allocation
function is a parameter to the analysis.
Monovariance: Pushdown 0CFA Pushdown 0CFA uses variables themselves for abstract
addresses:
[ = Var
Addr
alloc(v, ĉ) = v.
For better precision, a program would be transformed to have unique binders.
Context-sensitive: Pushdown 1CFA Pushdown 1CFA pairs the variable with the current
expression to get an abstract address:
[ = Var × Exp
Addr
alloc(v, (e, ρ̂, σ̂ , κ̂)) = (v, e).
For better precision, expressions are often uniquely labeled so that textually equal expressions at different points in the program are distinguished.
Polymorphic splitting: Pushdown poly/CFA Assuming we compiled the program from
a programming language with let expressions and we marked which identifiers were letbound, we can enable polymorphic splitting:
ZU064-05-FPR
paper-jfp
6 February 2018
3:38
PUSHDOWN FLOW ANALYSIS WITH ABSTRACT GARBAGE COLLECTION
15
[ = Var + Var × Exp
Addr
(
(v, [[( f æ)]]) f is let-bound
alloc(v, ([[( f æ)]], ρ̂, σ̂ , κ̂)) =
v
otherwise.
Pushdown k-CFA For pushdown k-CFA, we need to look beyond the current state and at
∗
[ to Var × Conf
[ → Addr.
[ By
the last k states, necessarily changing the signature of alloc
concatenating the expressions in the last k states together, and pairing this sequence with a
variable we get pushdown k-CFA:
[ = Var × Expk
Addr
[ h(e1 , ρ̂1 , σ̂1 , κ̂1 ), . . .i) = (v, he1 , . . . , ek i).
alloc(v,
4.1 Partial orders
For each set X̂ inside the abstract configuration-space, we use the natural partial order,
(vX̂ ) ⊆ X̂ × X̂. Abstract addresses and syntactic sets have flat partial orders. For the other
sets, the partial order lifts:
• point-wise over environments:
ρ̂ v ρ̂ 0 iff ρ̂(v) = ρ̂ 0 (v) for all v ∈ dom(ρ̂);
• component-wise over closures:
(lam, ρ̂) v (lam, ρ̂ 0 ) iff ρ̂ v ρ̂ 0 ;
• point-wise over stores:
σ̂ v σ̂ 0 iff σ̂ (â) v σ̂ 0 (â) for all â ∈ dom(σ̂ );
• component-wise over frames:
(v, e, ρ̂) v (v, e, ρ̂ 0 ) iff ρ̂ v ρ̂ 0 ;
• element-wise over continuations:
hφ̂1 , . . . , φ̂n i v hφ̂10 , . . . , φ̂n0 i iff φ̂i v φ̂i0 ; and
• component-wise across configurations:
(e, ρ̂, σ̂ , κ̂) v (e, ρ̂ 0 , σ̂ 0 , κ̂ 0 ) iff ρ̂ v ρ̂ 0 and σ̂ v σ̂ 0 and κ̂ v κ̂ 0 .
ZU064-05-FPR
paper-jfp
16
6 February 2018
3:38
J.I. Johnson, I. Sergey, C. Earl, M. Might, and D. Van Horn
4.2 Soundness
To prove soundness, an abstraction map α connects the concrete and abstract configurationspaces:
α(e, ρ, σ , κ) = (e, α(ρ), α(σ ), α(κ))
α(ρ) = λ v.α(ρ(v))
α(σ ) = λ â.
G
{α(σ (a))}
α(a)=â
αhφ1 , . . . , φn i = hα(φ1 ), . . . , α(φn )i
α(v, e, ρ) = (v, e, α(ρ))
α(a) is determined by the allocation functions.
It is then easy to prove that the abstract transition relation simulates the concrete transition
relation:
Theorem 4.1
[ such that α(c0 ) v ĉ0 and ĉ c
If α(c) v ĉ and c ⇒ c0 , then there exists ĉ0 ∈ Conf
⇒ ĉ0 .
Proof
The proof follows by case analysis on the expression in the configuration. It is a straightforward adaptation of similar proofs, such as that of Might (2007) for k-CFA.
5 From the Abstracted CESK Machine to a PDA
In the previous section, we constructed an infinite-state abstract interpretation of the CESK
machine. The infinite-state nature of the abstraction makes it difficult to see how to answer
static analysis questions. Consider, for instance, a control flow-question:
At the call site ( f æ), may a closure over lam be called?
If the abstracted CESK machine were a finite-state machine, an algorithm could answer this
question by enumerating all reachable configurations and looking for an abstract configuration ([[( f æ)]], ρ̂, σ̂ , κ̂) in which (lam, ) ∈ Aˆ( f , ρ̂, σ̂ ). However, because the abstracted
CESK machine may contain an infinite number of reachable configurations, an algorithm
cannot enumerate them.
Fortunately, we can recast the abstracted CESK as a special kind of infinite-state system:
a pushdown automaton (PDA). Pushdown automata occupy a sweet spot in the theory of
computation: they have an infinite configuration-space, yet many useful properties (e.g.,
word membership, non-emptiness, control-state reachability) remain decidable. Once the
abstracted CESK machine becomes a PDA, we can answer the control-flow question by
checking whether a specific regular language, accounting for the states of interest, when
intersected with the language of the PDA, is nonempty.
The recasting as a PDA is a shift in perspective. A configuration has an expression, an
environment and a store. A stack character is a frame. We choose to make the alphabet the
set of control states, so that the language accepted by the PDA will be sequences of controlstates visited by the abstracted CESK machine. Thus, every transition will consume the
ZU064-05-FPR
paper-jfp
6 February 2018
3:38
PUSHDOWN FLOW ANALYSIS WITH ABSTRACT GARBAGE COLLECTION
\ (e) = (Q, Σ, Γ, δ , q0 , F, hi), where
PDA
17
(q, ε, q0 , q0 ) ∈ δ iff (q, κ̂) c
⇒ (q0 , κ̂) for all κ̂
d × Store
[
Q = Exp × Env
(q, φ̂− , q0 , q0 ) ∈ δ iff (q, φ̂ : κ̂) c
⇒ (q0 , κ̂) for all κ̂
Σ=Q
(q, φ̂+0 , q0 , q0 ) ∈ δ iff (q, κ̂) c
⇒ (q0 , φ̂ 0 : κ̂) for all κ̂
(q0 , hi) = Iˆ(e)
\
Γ = Frame
F = Q.
\ : Exp → PDA.
Fig. 5: PDA
control-state to which it transitioned as an input character. Figure 5 defines the program-to\ : Exp → PDA. (Note the implicit use of the isomorphism
PDA conversion function PDA
∼
d
[
Q × Kont = Conf .)
At this point, we can answer questions about whether a specified control state is reachable by formulating a question about the intersection of a regular language with a contextfree language described by the PDA. That is, if we want to know whether the control state
(e0 , ρ̂, σ̂ ) is reachable in a program e, we can reduce the problem to determining:
\ (e)) 6= 0,
Σ∗ · (e0 , ρ̂, σ̂ ) · Σ∗ ∩ L (PDA
/
where L1 · L2 is the concatenation of formal languages L1 and L2 .
Theorem 5.1
Control-state reachability is decidable.
Proof
The intersection of a regular language and a context-free language is context-free (simple
machine product of PDA with DFA). The emptiness of a context-free language is decidable. The decision procedure is easiest for CFGs: mark terminals, mark non-terminals that
reduce to marked (non)terminals until we reach a fixed point. If the start symbol is marked,
then the language is nonempty. The PDA to CFG translation is a standard construction.
Now, consider how to use control-state reachability to answer the control-flow question
from earlier. There are a finite number of possible control states in which the λ -term lam
may flow to the function f in call site ( f æ); let’s call this set of states Ŝ:
Ŝ = ([[( f æ)]], ρ̂, σ̂ ) : (lam, ρ̂ 0 ) ∈ Aˆ( f , ρ̂, σ̂ ) for some ρ̂ 0 .
What we want to know is whether any state in the set Ŝ is reachable in the PDA. In effect
what we are asking is whether there exists a control state q ∈ Ŝ such that:
\ (e)) 6= 0.
Σ∗ · {q} · Σ∗ ∩ L (PDA
/
If this is true, then lam may flow to f ; if false, then it does not.
Problem: Doubly exponential complexity The non-emptiness-of-intersection approach
establishes decidability of pushdown control-flow analysis. But, two exponential complexity barriers make this technique impractical.
ZU064-05-FPR
paper-jfp
6 February 2018
18
3:38
J.I. Johnson, I. Sergey, C. Earl, M. Might, and D. Van Horn
[ |Var| ) and stores
First, there are an exponential number of both environments (|Addr|
d [
d [
((2|Clo| )|Addr| = 2|Clo|×|Addr| ) to consider for the set Ŝ. On top of that, computing the intersection of a regular language with a context-free language will require enumeration of
the (exponential) control-state-space of the PDA. The size of the control-state-space of the
PDA is clearly doubly exponential:
d × Store|
[
|Q| = |Exp × Env
d × |Store|
[
= |Exp| × |Env|
d Addr|
[
[ |Var| × 2|Clo|×|
= |Exp| × |Addr|
d Addr|
[
[ |Var| × 2|Lam×Env|×|
= |Exp| × |Addr|
[ |Var| ×|Addr|
[
[ |Var| × 2|Lam|×|Addr|
= |Exp| × |Addr|
As a result, this approach is doubly exponential. For the next few sections, our goal will be
to lower the complexity of pushdown control-flow analysis.
6 Focusing on Reachability
In the previous section, we saw that control-flow analysis reduces to the reachability of
certain control states within a pushdown system. We also determined reachability by converting the abstracted CESK machine into a PDA, and using emptiness-testing on a language derived from that PDA. Unfortunately, we also found that this approach is deeply
exponential.
Since control-flow analysis reduced to the reachability of control-states in the PDA,
we skip the language problems and go directly to reachability algorithms of Bouajjani
et al. (1997); Kodumal and Aiken (2004); Reps (1998) and Reps et al. (2005) that determine the reachable configurations within a pushdown system. These algorithms are
even polynomial-time. Unfortunately, some of them are polynomial-time in the number
of control states, and in the abstracted CESK machine, there are an exponential number of
control states. We don’t want to enumerate the entire control state-space, or else the search
becomes exponential in even the best case.
To avoid this worst-case behavior, we present a straightforward pushdown-reachability
algorithm that considers only the reachable control states. We cast our reachability algorithm as a fixed-point iteration, in which we incrementally construct the reachable subset
of a pushdown system. A rooted pushdown system M = (Q, Γ, δ , q0 ) is compact if for any
(q, g, q0 ) ∈ δ , it is the case that:
(q0 , hi) 7−→∗ (q,~γ) for some stack ~γ,
M
and the domain of states and stack characters are exactly those that appear in δ :
Q = {{q, q0 } : (q, g, q0 ) ∈ δ }
Γ = {γ : (q, γ+ , q0 ) ∈ δ or (q, γ− , q0 ) ∈ δ }
S
In other words, a rooted pushdown system is compact when its states, transitions and stack
characters appear on legal paths from the initial control state. We will refer to the class of
compact rooted pushdown systems as CRPDS.
ZU064-05-FPR
paper-jfp
6 February 2018
3:38
PUSHDOWN FLOW ANALYSIS WITH ABSTRACT GARBAGE COLLECTION
19
We can compact a rooted pushdown system with a map:
C : RPDS → CRPDS
M
z
}|
{
C (Q, Γ, δ , q0 ) = (Q0 , Γ0 , δ 0 , q0 )
n
o
where Q0 = q : (q0 , hi) 7−→∗ (q,~γ)
M
n
o
0
Γ = γi : (q0 , hi) 7−→∗ (q,~γ)
M
~g
g
0
0
0
δ = (q, g, q ) : (q0 , hi) 7−→ (q, [~g]) 7−→ (q , [~gg]) .
M
M
In practice, the real difference between a rooted pushdown system and its compact form
is that our original system will be defined intensionally (having come from the components
of an abstracted CESK machine), whereas the compact system will be defined extensionally, with the contents of each component explicitly enumerated during its construction.
Our near-term goals are (1) to convert our abstracted CESK machine into a rooted
pushdown system and (2) to find an efficient method to compact it.
To convert the abstracted CESK machine into a rooted pushdown system, we use the
\ : Exp → RPDS:
function RPDS
\ (e) = (Q, Γ, δ , q0 )
RPDS
ε
q q0 ∈ δ iff (q, κ̂) c
⇒ (q0 , κ̂) for all κ̂
d × Store
[
Q = Exp × Env
\
Γ = Frame
(q0 , hi) = Iˆ(e)
φ̂−
q q0 ∈ δ iff (q, φ̂ : κ̂) c
⇒ (q0 , κ̂) for all κ̂
φ̂+
q q0 ∈ δ iff (q, κ̂) c
⇒ (q0 , φ̂ : κ̂) for all κ̂.
7 Compacting a Rooted Pushdown System
We now turn our attention to compacting a rooted pushdown system (defined intensionally)
into its compact form (defined extensionally). That is, we want to find an implementation
of the function C . To do so, we first phrase the construction as the least fixed point of a
monotonic function. This will provide a method (albeit an inefficient one) for computing
the function C . In the next section, we look at an optimized work-set driven algorithm that
avoids the inefficiencies of this section’s algorithm.
The function F : RPDS → (CRPDS → CRPDS) generates the monotonic iteration
function we need:
F (M) = f , where
M = (Q, Γ, δ , q0 )
f (S, Γ, E, s0 ) = (S0 , Γ, E 0 , s0 ), where
n
o
S0 = S ∪ s0 : s ∈ S and s 7−
−→
→ s0 ∪ {s0 }
M
n g
o
g
0
0
E = E ∪ s s : s ∈ S and s 7−
−→
→ s0 .
M
Given a rooted pushdown system M, each application of the function F (M) accretes new
edges at the frontier of the system. Once the algorithm reaches a fixed point, the system is
complete:
ZU064-05-FPR
paper-jfp
20
6 February 2018
3:38
J.I. Johnson, I. Sergey, C. Earl, M. Might, and D. Van Horn
Theorem 7.1
C (M) = lfp(F (M)).
Proof
Let M = (Q, Γ, δ , q0 ). Let f = F (M). Observe that lfp( f ) = f n (0,
/ Γ, 0,
/ q0 ) for some n.
When N ⊆ C (M), then it easy to show that f (N) ⊆ C (M). Hence, C (M) ⊇ lfp(F (M)).
To show C (M) ⊆ lfp(F (M)), suppose this is not the case. Then, there must be at least
one edge in C (M) that is not in lfp(F (M)). Since these edges must be root reachable, let
(s, g, s0 ) be the first such edge in some path from the root. This means that the state s is
in lfp(F (M)). Let m be the lowest natural number such that s appears in f m (M). By the
definition of f , this edge must appear in f m+1 (M), which means it must also appear in
lfp(F (M)), which is a contradiction. Hence, C (M) ⊆ lfp(F (M)).
7.1 Complexity: Polynomial and exponential
To determine the complexity of this algorithm, we ask two questions: how many times
would the algorithm invoke the iteration function in the worst case, and how much does
each invocation cost in the worst case? The size of the final system bounds the run-time
of the algorithm. Suppose the final system has m states. In the worst case, the iteration
function adds only a single edge each time. Since there are at most 2|Γ|m2 + m2 edges in
the final graph, the maximum number of iterations is 2|Γ|m2 + m2 .
The cost of computing each iteration is harder to bound. The cost of determining whether
to add a push edge is proportional to the size of the stack alphabet, while the cost of
determining whether to add an ε-edge is constant, so the cost of determining all new
push and ε edges to add is proportional to |Γ|m + m. Determining whether or not to add
γ−
a pop edge is expensive. To add the pop edge s s0 , we must prove that there exists a
configuration-path to the control state s, in which the character γ is on the top of the stack.
This reduces to a CFL-reachability query (Melski and Reps 2000) at each node, the cost of
which is O(|Γ± |3 m3 ) (Kodumal and Aiken 2004).
To summarize, in terms of the number of reachable control states, the complexity of the
most recent algorithm is:
O((2|Γ|m2 + m2 ) × (|Γ|m + m + |Γ± |3 m3 )) = O(|Γ|4 m5 ).
While this approach is polynomial in the number of reachable control states, it is far from
efficient. In the next section, we provide an optimized version of this fixed-point algorithm
that maintains a work-set and an ε-closure graph to avoid spurious recomputation.
Moreover, we have carefully phrased the complexity in terms of “reachable” control
states because, in practice, compact rooted pushdown systems will be extremely sparse,
and because the maximum number of control states is exponential in the size of the input
program. After the subsequent refinement, we will be able to develop a hierarchy of pushdown control-flow analyses that employs widening to achieve a polynomial-time algorithm
at its foundation.
ZU064-05-FPR
paper-jfp
6 February 2018
3:38
PUSHDOWN FLOW ANALYSIS WITH ABSTRACT GARBAGE COLLECTION
21
8 An Efficient Algorithm: Work-sets and ε-Closure Graphs
We have developed a fixed-point formulation of the rooted pushdown system compaction
algorithm, but found that, in each iteration, it wasted effort by passing over all discovered
states and edges, even though most will not contribute new states or edges. Taking a cue
from graph search, we can adapt the fixed-point algorithm with a work-set. That is, our next
algorithm will keep a work-set of new states and edges to consider, instead of reconsidering
all of them. We will refer to the compact rooted pushdown system we are constructing as
a graph, since that is how we represent it (Q is the set of nodes, and δ is a set of labeled
edges).
In each iteration, it will pull new states and edges from the work list, insert them into the
graph and then populate the work-set with new states and edges that have to be added as a
consequence of the recent additions.
8.1 ε-closure graphs
Figuring out what edges to add as a consequence of another edge requires care, for adding
an edge can have ramifications on distant control states. Consider, for example, adding the
ε-edge q ε q0 into the following graph:
q0
γ+
/q
q0
γ−
/ q1
As soon this edge drops in, an ε-edge “implicitly” appears between q0 and q1 because the
net stack change between them is empty; the resulting graph looks like:
ε
q0
γ+
/q
ε
/ q0
γ−
/ q 1
where we have illustrated the implicit ε-edge as a dotted line.
To keep track of these implicit edges, we will construct a second graph in conjunction
with the graph: an ε-closure graph. In the ε-closure graph, every edge indicates the existence of a no-net-stack-change path between control states. The ε-closure graph simplifies
the task of figuring out which states and edges are impacted by the addition of a new edge.
Formally, an ε-closure graph, H ⊆ N × N, is a set of edges. Of course, all ε-closure
graphs are reflexive: every node has a self loop. We use the symbol ECG to denote the
class of all ε-closure graphs.
We have two notations for finding ancestors and descendants of a state in an ε-closure
graph:
←
−
G ε [s] = s0 : (s0 , s) ∈ H ∪ {s}
[ancestors]
0
→
−
G ε [s] = s : (s, s0 ) ∈ H ∪ {s}
[descendants].
8.2 Integrating a work-set
Since we only want to consider new states and edges in each iteration, we need a work-set,
or in this case, three work-sets:
ZU064-05-FPR
paper-jfp
22
6 February 2018
3:38
J.I. Johnson, I. Sergey, C. Earl, M. Might, and D. Van Horn
• ∆S contains states to add,
• ∆E contains edges to add,
• ∆H contains new ε-edges.
Let WS ::= (∆S, ∆E, ∆H) be the space of work-sets.
8.3 A new fixed-point iteration-space
Instead of consuming a graph and producing a graph, the new fixed-point iteration function
will consume and produce a graph, an ε-closure graph, and the work-sets. Hence, the
iteration space of the new algorithm is:
ICRPDS = (℘(Q) ×℘(Q × Γ± × Q)) × ECG × WS.
The I in ICRPDS stands for intermediate.
8.4 The ε-closure graph work-list algorithm
The function F 0 : RPDS → (ICRPDS → ICRPDS) generates the required iteration function (Figure 6). Please note that we implicitly distribute union across tuples:
(X,Y ) ∪ (X 0 ,Y 0 ) = (X ∪ X,Y ∪Y 0 ).
The functions sprout, addPush, addPop, addEmpty (defined shortly) calculate the additional the graph edges and ε-closure graph edges (potentially) introduced by a new state or
edge.
Sprouting Whenever a new state gets added to the graph, the algorithm must check whether
that state has any new edges to contribute. Both push edges and ε-edges do not depend on
the current stack, so any such edges for a state in the pushdown system’s transition function
belong in the graph. The sprout function:
sprout(Q,Γ,δ ,s0 ) : Q → (P (δ ) × P (Q × Q)),
checks whether a new state could produce any new push edges or no-change edges. We
can represent its behavior diagrammatically (as previously, the dotted arrows correspond
to the corresponding additions to the work-graph and ε-closure work graph):
s
γ+
ε
δ
q0
δ
q00
which means if adding control state s:
ε
add edge s q0 if it exists in δ (hence the arrow subscript δ ), and
γ+
add edge s q00 if it exists in δ .
Formally:
sprout(Q,Γ,δ ,s0 ) (s) = (∆E, ∆H), where
ZU064-05-FPR
paper-jfp
6 February 2018
3:38
PUSHDOWN FLOW ANALYSIS WITH ABSTRACT GARBAGE COLLECTION
23
F 0 (M) = f , where
M = (Q, Γ, δ , q0 )
f (G, H, (∆S, ∆E, ∆H)) = (G0 , H 0 , (∆S0 − S0 , ∆E 0 − E 0 , ∆H 0 − H)), where
(S, Γ, E, s0 ) = G
(∆E0 , ∆H0 ) =
[
sproutM (s)
s∈∆S
[
(∆E1 , ∆H1 ) =
(s,γ+
(∆E2 , ∆H2 ) =
addPushM (G, H)(s, γ+ , s0 )
,s0 )∈∆E
[
addPopM (G, H)(s, γ− , s0 )
(s,γ− ,s0 )∈∆E
(∆E3 , ∆H3 ) =
[
addEmptyM (G, H)(s, s0 )
(s,ε,s0 )∈∆E
(∆E4 , ∆H4 ) =
[
addEmptyM (G, H)(s, s0 )
(s,s0 )∈∆H
S0 = S ∪ ∆S
E 0 = E ∪ ∆E
H 0 = H ∪ ∆H
∆E 0 = ∆E0 ∪ ∆E1 ∪ ∆E2 ∪ ∆E3 ∪ ∆E4
∆S0 = s0 : (s, g, s0 ) ∈ ∆E 0 ∪ {s0 }
∆H 0 = ∆H0 ∪ ∆H1 ∪ ∆H2 ∪ ∆H3 ∪ ∆H4
∆Γ = γ : (s, γ+ , s0 ) ∈ ∆E 0
G0 = (S ∪ ∆S, Γ ∪ ∆Γ, E 0 , q0 ).
Fig. 6: The fixed point of the function F 0 (M) contains the compact form of the rooted
pushdown system M.
n ε
o γ+
γ+
ε
∆E = s q : s q ∈ δ ∪ s q : s q ∈ δ
n
o
ε
∆H = s q : s q ∈ δ .
Considering the consequences of a new push edge Once our algorithm adds a new push
edge to a graph, there is a chance that it will enable new pop edges for the same stack frame
somewhere downstream. If and when it does enable pops, it will also add new edges to the
ε-closure graph. The addPush function:
addPush(Q,Γ,δ ,s0 ) : RPDS × ECG → δ → (P (δ ) × P (Q × Q)),
ZU064-05-FPR
paper-jfp
24
6 February 2018
3:38
J.I. Johnson, I. Sergey, C. Earl, M. Might, and D. Van Horn
checks for ε-reachable states that could produce a pop. We can represent this action by the
following diagram (the arrow subscript ε indicates edges in the ε-closure graph):
s
/q
γ+
ε
ε
/ q0
γ−
δ
/ q00
G
ε
ε
γ+
which means if adding push-edge s q:
γ−
if pop-edge q0 q00 is in δ , then
γ−
add edge q0 q00 , and
add ε-edge s q00 .
Formally:
γ+
addPush(Q,Γ,δ ,s0 ) (G, H)(s q) = (∆E, ∆H), where
γ
γ
→
−
0 − 00
0
0 − 00
∆E = q q : q ∈ G ε [q] and q q ∈ δ
γ
→
−
00
0
0 − 00
∆H = s q : q ∈ G ε [q] and q q ∈ δ .
Considering the consequences of a new pop edge Once the algorithm adds a new pop
edge to a graph, it will create at least one new ε-closure graph edge and possibly more by
matching up with upstream pushes. The addPop function:
addPop(Q,Γ,δ ,s0 ) : RPDS × ECG → δ → (P (δ ) × P (Q × Q)),
checks for ε-reachable push-edges that could match this pop-edge. This action is illustrated
by the following diagram:
s
γ+
/ s0
ε
ε
/ s00
γ−
δ
/q
I
ε
ε
γ−
which means if adding pop-edge s00 q:
γ+
if push-edge s s0 is already in the graph, then
add ε-edge s q.
Formally:
γ−
addPop(Q,Γ,δ ,s0 ) (G, H)(s00 q) = (∆E, ∆H), where
γ+
←
−
∆E = 0/ and ∆H = s q : s0 ∈ G ε [s00 ] and s s0 ∈ G .
Considering the consequences of a new ε-edge Once the algorithm adds a new ε-closure
graph edge, it may transitively have to add more ε-closure graph edges, and it may connect
ZU064-05-FPR
paper-jfp
6 February 2018
3:38
PUSHDOWN FLOW ANALYSIS WITH ABSTRACT GARBAGE COLLECTION
25
an old push to (perhaps newly enabled) pop edges. The addEmpty function:
addEmpty(Q,Γ,δ ,s0 ) : RPDS × ECG → (Q × Q) → (P (δ ) × P (Q × Q)),
checks for newly enabled pops and ε-closure graph edges: Once again, we can represent
this action diagrammatically:
ε
ε
s
γ+
/ s0
ε
ε
ε
ε
/ s00
/ s000
ε
ε
ε
/ s0000
I
γ−
δ
/q
O
ε
ε
ε
ε
which means if adding ε-edge s00 s000 :
γ−
if pop-edge s0000 q is in δ , then
add ε-edge s q; and
γ−
add edge s0000 q;
add ε-edges s0 s000 , s00 s0000 , and s0 s0000 .
Formally:
addEmpty(Q,Γ,δ ,s0 ) (G, H)(s00 s000 ) = (∆E, ∆H), where
γ−
←
−
→
−
∆E = s0000 q : s0 ∈ G ε [s00 ] and s0000 ∈ G ε [s000 ] and
γ+
γ−
s s0 ∈ G and s0000 q ∈ δ
←
−
→
−
∆H = s q : s0 ∈ G ε [s00 ] and s0000 ∈ G ε [s000 ] and
γ+
γ−
s s0 ∈ G and s0000 q ∈ δ
n
o
←
−
∪ s0 s000 : s0 ∈ G ε [s00 ]
n
o
→
−
∪ s00 s0000 : s0000 ∈ G ε [s000 ]
n
o
←
−
→
−
∪ s0 s0000 : s0 ∈ G ε [s00 ] and s0000 ∈ G ε [s000 ] .
8.5 Termination and correctness
To prove that a fixed point exists, we show the iteration function is monotonic. The key
observation is that ∆G and ∆H drive all additions to, and are disjoint from, G and H. Since
G and H monotonically increase in a finite space, ∆G and ∆H run out of room (full details
in 18.1). Once the graph reaches a fixed point, all work-sets will be empty, and the εclosure graph will also be saturated. We can also show that this algorithm is correct by
defining first E C G : RPDS → ECG as
~g
0
0
E C G (M) = s s : s 7−
−→
→ s and [~g] = ε
M
ZU064-05-FPR
paper-jfp
26
6 February 2018
3:38
J.I. Johnson, I. Sergey, C. Earl, M. Might, and D. Van Horn
and stating the following theorem:
Theorem 8.1
For all M ∈ RPDS, C (M) = G and E C G (M) = H, where (G, H, (0,
/ 0,
/ 0))
/ = lfp(F 0 (M)).
In the proof of Theorem 8.1, the ⊆ case comes from an invariant lemma we have on F 0 :
Lemma 8.1
g
g
inv((S, E), H, (∆S, ∆E, ∆H)) = ∀s s0 ∈ E ∪ ∆E.s 7−
−→
→ s0
M
~g
∧ ∀s s0 ∈ H ∪ ∆H.∃~g.[~g] = ε ∧ ∀κ̂.(s, κ̂) 7−→∗ (s, κ̂)
M
The ⊇ case follows from
Lemma 8.2
~g
~g
−→
→ s, there is both a corresponding path s0 7−
−→
→ s and for all nonFor all traces π ≡ s0 7−
M
G
~g0
empty subtraces of π, sb 7−
−→
→ sa , if [~g0 ] = ε then sb sa ∈ H.
M
Since all edges in a compact rooted pushdown system must be in a path from the initial
state, we can extract the edges from said paths using this lemma.
8.6 Complexity: Still exponential, but more efficient
As in the previous case (Section 7.1), to determine the complexity of this algorithm, we
ask two questions: how many times would the algorithm invoke the iteration function in
the worst case, and how much does each invocation cost in the worst case? The run-time of
the algorithm is bounded by the size of the final graph plus the size of the ε-closure graph.
Suppose the final graph has m states. In the worst case, the iteration function adds only a
single edge each time. There are at most 2|Γ|m2 +m2 edges in the graph (|Γ|m2 push edges,
just as many pop edges, and m2 no-change edges) and at most m2 edges in the ε-closure
graph, which bounds the number of iterations. Recall that m can be exponential in the size
of the program, since m ≤ |Q| (and Section 5 derived the exponential size of |Q|).
Next, we must reason about the worst-case cost of adding an edge: how many edges
might an individual iteration consider? In the worst case, the algorithm will consider every
edge in every iteration, leading to an asymptotic time-complexity of:
O((2|Γ|m2 + 2m2 )2 ) = O(|Γ|2 m4 ).
While still high, this is a an improvement upon the previous algorithm. For sparse graphs,
this is a reasonable algorithm.
9 Polynomial-Time Complexity from Widening
In the previous section, we developed a more efficient fixed-point algorithm for computing a compact rooted pushdown system. Even with the core improvements we made,
the algorithm remained exponential in the worst case, owing to the fact that there could
ZU064-05-FPR
paper-jfp
6 February 2018
3:38
PUSHDOWN FLOW ANALYSIS WITH ABSTRACT GARBAGE COLLECTION
27
be an exponential number of reachable control states. When an abstract interpretation
is intolerably complex, the standard approach for reducing complexity and accelerating
convergence is widening (Cousot and Cousot 1977). Of course, widening techniques trade
away some precision to gain this speed. It turns out that the small-step variants of finitestate CFAs are exponential without some sort of widening as well (Van Horn and Mairson
2008).
To achieve polynomial time complexity for pushdown control-flow analysis requires the
same two steps as the classical case: (1) widening the abstract interpretation to use a global,
“single-threaded” store and (2) selecting a monovariant allocation function to collapse the
abstract configuration-space. Widening eliminates a source of exponentiality in the size
of the store; monovariance eliminates a source of exponentiality from environments. In
this section, we redevelop the pushdown control-flow analysis framework with a singlethreaded store and calculate its complexity.
9.1 Step 1: Refactor the concrete semantics
First, consider defining the reachable states of the concrete semantics using fixed points.
That is, let the system-space of the evaluation function be sets of configurations:
C ∈ System = P (Conf ) = P (Exp × Env × Store × Kont).
We can redefine the concrete evaluation function:
E (e) = lfp( fe ), where fe : System → System and
fe (C) = {I (e)} ∪ c0 : c ∈ C and c ⇒ c0 .
9.2 Step 2: Refactor the abstract semantics
We can take the same approach with the abstract evaluation function, first redefining the
abstract system-space:
[
\ = P Conf
Ĉ ∈ System
d × Store
d ,
[ × Kont
= P Exp × Env
and then the abstract evaluation function:
\ → System
\ and
Eˆ (e) = lfp( fˆe ), where fˆe : System
0
fˆe (Ĉ) = Iˆ(e) ∪ ĉ : ĉ ∈ Ĉ and ĉ c
⇒ ĉ0 .
What we’d like to do is shrink the abstract system-space with a refactoring that corresponds
to a widening.
9.3 Step 3: Single-thread the abstract store
We can approximate a set of abstract stores {σ̂1 , . . . , σ̂n } with the least-upper-bound of
those stores: σ̂1 t · · · t σ̂n . We can exploit this by creating a new abstract system space in
which the store is factored out of every configuration. Thus, the system-space contains a
ZU064-05-FPR
paper-jfp
28
6 February 2018
3:38
J.I. Johnson, I. Sergey, C. Earl, M. Might, and D. Van Horn
set of partial configurations and a single global store:
0
\ × Store
\ = P PConf
[
System
d × Kont.
d
\ = Exp × Env
π̂ ∈ PConf
We can factor the store out of the abstract transition relation as well, so that (_σ̂ ) ⊆
\ × (PConf
\ × Store):
[
PConf
σ̂
(e, ρ̂, κ̂) _ ((e0 , ρ̂ 0 , κ̂ 0 ), σ̂ 0 ) iff (e, ρ̂, σ̂ , κ̂) c
⇒ (e0 , ρ̂ 0 , σ̂ 0 , κ̂ 0 ),
0
0
\ → System
\ ,
which gives us a new iteration function, fˆe0 : System
fˆe0 (P̂, σ̂ ) = (P̂0 , σ̂ 0 ), where
σ̂
0
0
0
00
P̂ = π̂ : π̂ _ (π̂ , σ̂ ) ∪ {π̂0 }
G
σ̂
0
00
0
00
σ̂ =
σ̂ : π̂ _ (π̂ , σ̂ )
(π̂0 , hi) = Iˆ(e).
9.4 Step 4: Pushdown control-flow graphs
Following the earlier graph reformulation of the compact rooted pushdown system, we
can reformulate the set of partial configurations as a pushdown control-flow graph. A
pushdown control-flow graph is a frame-action-labeled graph over partial control states,
and a partial control state is an expression paired with an environment:
00
\ = PDCFG
\ × Store
[
System
\ = P(PState)
\ × P(PState
\ × Frame
\
\ ± × PState)
PDCFG
d
\ = Exp × Env.
ψ̂ ∈ PState
In a pushdown control-flow graph, the partial control states are partial configurations which
have dropped the continuation component; the continuations are encoded as paths through
the graph.
A preliminary analysis of complexity Even without defining the system-space iteration
function, we can ask, How many iterations will it take to reach a fixed point in the worst
case? This question is really asking, How many edges can we add? And, How many entries
are there in the store? Summing these together, we arrive at the worst-case number of
iterations:
PDCFG edges
store entries
}|
{
}|
{ z
z
d.
[ × |Clo|
\ × |Frame
\ + |Addr|
\ ± | × |PState|
|PState|
With a monovariant allocation scheme that eliminates abstract environments, the number
of iterations ultimately reduces to:
d + 1) × |Exp| + |Var| × |Lam|,
|Exp| × (2|Var|
ZU064-05-FPR
paper-jfp
6 February 2018
3:38
PUSHDOWN FLOW ANALYSIS WITH ABSTRACT GARBAGE COLLECTION
29
which means that, in the worst case, the algorithm makes a cubic number of iterations with
respect to the size of the input program.3
The worst-case cost of the each iteration would be dominated by a CFL-reachability
calculation, which, in the worst case, must consider every state and every edge:
O(|Var|3 × |Exp|3 ).
Thus, each iteration takes O(n6 ) and there are a maximum of O(n3 ) iterations, where n is
the size of the program. So, total complexity would be O(n9 ) for a monovariant pushdown
control-flow analysis with this scheme, where n is again the size of the program. Although
this algorithm is polynomial-time, we can do better.
9.5 Step 5: Reintroduce ε-closure graphs
Replicating the evolution from Section 8 for this store-widened analysis, we arrive at a
more efficient polynomial-time analysis. An ε-closure graph in this setting is a set of pairs
of store-less, continuation-less partial states:
d = P PState
\ × PState
\ .
ECG
Then, we can set the system space to include ε-closure graphs:
000
d × Store.
\ = CRPDS
\ × ECG
[
System
Before we redefine the iteration function, we need another factored transition relation.
σ̂
\ × PState
\ × Store deterThe stack- and action-factored transition relation (+) ⊆ PState
g
mines if a transition is possible under the specified store and stack-action:
(e, ρ̂) + ((e0 , ρ̂ 0 ), σ̂ 0 ) iff (e, ρ̂, σ̂ , κ̂) c
⇒ (e0 , ρ̂ 0 , σ̂ 0 , φ̂ : κ̂)
σ̂
φ̂+
(e, ρ̂) + ((e0 , ρ̂ 0 ), σ̂ 0 ) iff (e, ρ̂, σ̂ , φ̂ : κ̂) c
⇒ (e0 , ρ̂ 0 , σ̂ 0 , κ̂)
σ̂
φ̂−
(e, ρ̂) + ((e0 , ρ̂ 0 ), σ̂ 0 ) iff (e, ρ̂, σ̂ , κ̂) c
⇒ (e0 , ρ̂ 0 , σ̂ 0 , κ̂).
σ̂
ε
Now, we can redefine the iteration function (Figure 7).
Theorem 9.1
Pushdown 0CFA with single-threaded store (PDCFA) can be computed in O(n6 )-time,
where n is the size of the program.
Proof
As before, the maximum number of iterations is cubic in the size of the program for a
monovariant analysis. Fortunately, the cost of each iteration is also now bounded by the
number of edges in the graph, which is also cubic.
3
In computing the number of frames, we note that in every continuation, the variable and the
expression uniquely determine each other based on the let-expression from which they both came.
As a result, the number of abstract frames available in a monovariant analysis is bounded by both
\ = |Var|.
the number of variables and the number of expressions, i.e., |Frame|
ZU064-05-FPR
paper-jfp
30
6 February 2018
3:38
J.I. Johnson, I. Sergey, C. Earl, M. Might, and D. Van Horn
fˆe ((P̂, Ê), Ĥ, σ̂ ) = ((P̂0 , Ê 0 ), Ĥ 0 , σ̂ 00 ), where
(
T̂+ =
φ̂+
)
0
0
0
σ̂
0
(ψ̂ ψ̂ , σ̂ ) : ψ̂ + (ψ̂ , σ̂ )
φ̂+
n
o
ε
σ̂
T̂ε = (ψ̂ ψ̂ 0 , σ̂ 0 ) : ψ̂ + (ψ̂ 0 , σ̂ 0 )
ε
φ̂−
σ̂
T̂− = (ψ̂ 00 ψ̂ 000 , σ̂ 0 ) : ψ̂ 00 + (ψ̂ 000 , σ̂ 0 ) and
φ̂−
φ̂+
ψ̂ ψ̂ 0 ∈ Ê and
ψ̂ 0 ψ̂ 00 ∈ Ĥ
T̂ 0 = T̂+ ∪ T̂ε ∪ T̂−
Ê 0 = ê : (ê, ) ∈ T̂ 0
G
σ̂ 00 =
σ̂ 0 : ( , σ̂ 0 ) ∈ T̂ 0
Ĥε = ψ̂ ψ̂ 00 : ψ̂ ψ̂ 0 ∈ Ĥ and ψ̂ 0 ψ̂ 00 ∈ Ĥ
φ̂+
Ĥ+− = ψ̂ ψ̂ 000 : ψ̂ ψ̂ 0 ∈ Ê and ψ̂ 0 ψ̂ 00 ∈ Ĥ
φ̂−
and ψ̂ 00 ψ̂ 000 ∈ Ê
Ĥ 0 = Ĥε ∪ Ĥ+−
n
o
g
P̂0 = P̂ ∪ ψ̂ 0 : ψ̂ ψ̂ 0 ∪ {(e, ⊥)} .
Fig. 7: An ε-closure graph-powered iteration function for pushdown control-flow analysis
with a single-threaded store.
10 Introspection for Abstract Garbage Collection
Abstract garbage collection (Might and Shivers 2006b) yields large improvements in precision by using the abstract interpretation of garbage collection to make more efficient
use of the finite address space available during analysis. Because of the way abstract
garbage collection operates, it grants exact precision to the flow analysis of variables whose
bindings die between invocations of the same abstract context. Because pushdown analysis
grants exact precision in tracking return-flow, it is clearly advantageous to combine these
techniques. Unfortunately, as we shall demonstrate, abstract garbage collection breaks the
pushdown model by requiring a full traversal of the stack to discover the root set.
Abstract garbage collection modifies the transition relation to conduct a “stop-and-copy”
garbage collection before each transition. To do this, we define a garbage collection func[ → Conf
[ on configurations:
tion Ĝ : Conf
ĉ
z }| {
Ĝ(e, ρ̂, σ̂ , κ̂) = (e, ρ̂, σ̂ |Reachable(ĉ), κ̂),
ZU064-05-FPR
paper-jfp
6 February 2018
3:38
PUSHDOWN FLOW ANALYSIS WITH ABSTRACT GARBAGE COLLECTION
31
where the pipe operation f |S yields the function f , but with inputs not in the set S mapped
[ → P(Addr)
[ first
to bottom—the empty set. The reachability function Reachable : Conf
computes the root set, and then the transitive closure of an address-to-address adjacency
relation:
ĉ
z }| {
∗
Reachable(e, ρ̂, σ̂ , κ̂) = â : â0 ∈ Root(ĉ) and â0 _ â ,
σ̂
[ → P(Addr)
[ finds the root addresses:
where the function Root : Conf
Root(e, ρ̂, σ̂ , κ̂) = range(ρ̂) ∪ StackRoot(κ̂),
d → P(Addr)
[ function finds roots down the stack:
and the StackRoot : Kont
StackRoothφ1 , . . . , φn i =
[
T (φi ),
i
[
\ → P(Addr):
using a “touches” function, T : Frame
T (v, e, ρ̂) = range(ρ̂),
[ × Store
[ connects adjacent addresses:
[ × Addr
and the relation (_) ⊆ Addr
â _ â0 iff there exists (lam, ρ̂) ∈ σ̂ (â) such that â0 ∈ range(ρ̂).
σ̂
The new abstract transition relation is thus the composition of abstract garbage collection
with the old transition relation:
(c
⇒GC ) = (c
⇒) ◦ Ĝ.
Problem: Stack traversal violates pushdown constraint In the formulation of pushdown systems, the transition relation is restricted to looking at the top frame, and in less
restricted formulations that may read the stack, the reachability decision procedures need
the entire system up-front. Thus, the relation (c
⇒GC ) cannot be computed as a straightforward pushdown analysis using summarization.
Solution: Introspective pushdown systems To accommodate the richer structure of the
relation (c
⇒GC ), we now define introspective pushdown systems. Once defined, we can
embed the garbage-collecting abstract interpretation within this framework, and then focus
on developing a control-state reachability algorithm for these systems.
An introspective pushdown system is a quadruple M = (Q, Γ, δ , q0 ):
1. Q is a finite set of control states;
2. Γ is a stack alphabet;
3. δ ⊆ Q × Γ∗ × Γ± × Q is a transition relation (where (q, κ, φ− , q0 ) ∈ δ implies κ ≡ φ :
κ 0 ); and
4. q0 is a distinguished root control state.
The second component in the transition relation is a realizable stack at the given controlstate. This realizable stack distinguishes an introspective pushdown system from a general
pushdown system. IPDS denotes the class of all introspective pushdown systems.
ZU064-05-FPR
paper-jfp
32
6 February 2018
3:38
J.I. Johnson, I. Sergey, C. Earl, M. Might, and D. Van Horn
Determining how (or if) a control state q transitions to a control state q0 , requires knowing a path taken to the state q. We concern ourselves with root-reachable states. When
M = (Q, Γ, δ , q0 ), if there is a κ̂ such that (q0 , hi) 7−→∗ (q, κ̂) we say q is reachable via κ̂,
M
where
(q, κ̂) 7−→∗ (q0 , κ̂ 0 ) (q0 , κ̂ 0 , g0 , q00 ) ∈ δ
M
(q, κ̂) 7−→∗ (q, κ̂)
M
(q, κ̂) 7−→∗ (q00 , [κ̂+0 g0 ])
M
10.1 Garbage collection in introspective pushdown systems
To convert the garbage-collecting, abstracted CESK machine into an introspective pushdown system, we use the function I \
PDS : Exp → IPDS:
I\
PDS (e) = (Q, Γ, δ , q0 )
d × Store
[
Q = Exp × Env
\
Γ = Frame
(q0 , hi) = Iˆ(e)
(q, κ̂, ε, q0 ) ∈ δ iff Ĝ(q, κ̂) c
⇒ (q0 , κ̂)
(q, φ̂ : κ̂, φ̂− , q0 ) ∈ δ iff Ĝ(q, φ̂ : κ̂) c
⇒ (q0 , κ̂)
(q, κ̂, φ̂+ , q0 ) ∈ δ iff Ĝ(q, κ̂) c
⇒ (q0 , φ̂ : κ̂).
11 Problem: Reachability for Introspective Pushdown Systems is Uncomputable
As currently formulated, computing control-state reachability for introspective pushdown
systems is uncomputable. The problem is that the transition relation expects to enumerate
every possible stack for every control point at every transition, without restriction.
Theorem 11.1
Reachability in introspective pushdown systems is uncomputable.
Proof
Consider an IPDS with two states — searching (start state) and valid — and a singleton
stack alphabet of unit (>). For any first-order logic proposition, φ , we can define a reduction relation that interprets the length of the stack as an encoding of a proof of φ . If the
length encodes an ill-formed proof object, or is not a proof of φ , searching pushes > on
the stack and transitions to itself. If the length encodes a proof of φ , transition to valid.
By the completeness of first-order logic, if φ is valid, there is a finite proof, making the
pushdown system terminate in valid. If it is not valid, then there is no proof and valid
is unreachable. Due to the undecidability of first-order logic, we definitely cannot have a
decision procedure for reachability of IPDSs.
To make introspective pushdown systems computable, we must first refine our definition
of introspective pushdown systems to operate on sets of stacks and insist these sets be
regular.
ZU064-05-FPR
paper-jfp
6 February 2018
3:38
PUSHDOWN FLOW ANALYSIS WITH ABSTRACT GARBAGE COLLECTION
33
A conditional pushdown system (CPDS) is a quadruple M = (Q, Γ, δ , q0 ):
1. Q is a finite set of control states;
2. Γ is a stack alphabet;
3. δ ⊆fin Q × RE G (Γ∗ ) × Γ± × Q is a transition relation (same restriction on stacks);
and
4. q0 is a distinguished root control state,
where RE G (S) is the set of all regular languages formable with strings in S.
The regularity constraint on the transition relations guarantees that we can decide applicability of transition rules at each state, since (as we will see) the set of all stacks that reach
a state in a CPDS has decidable overlap with regular languages. Let CPDS denote the set
of all conditional pushdown systems.
The rules for reachability with respect to sets of stacks are similar to those for IPDSs.
(q, κ̂) 7−→∗ (q0 , κ̂ 0 ) κ̂ 0 ∈ K̂ 0
M
(q, κ̂) 7−→∗ (q, κ̂)
(q0 , K̂ 0 , g0 , q00 ) ∈ δ
(q, κ̂) 7−→∗ (q00 , [κ̂+0 g0 ])
M
M
K̂,g
We will write q 7−
−→
→ q0 to mean there are κ̂, K̂ such that q is reachable via κ̂, κ̂ ∈ K̂ and
M
(q, K̂, g, q0 ) ∈ δ . We will omit the labels above if they merely exist.
11.1 Garbage collection in conditional pushdown systems
Of course, we must adapt abstract garbage collection to this refined framework. To convert
the garbage-collecting, abstracted CESK machine into a conditional pushdown system, we
0
use the function I \
PDS : Exp → CPDS:
0
I\
PDS (e) = (Q, Γ, δ , q0 )
d × Store
[
Q = Exp × Env
\
Γ = Frame
[ let K̂ = {κ̂ : StackRoot(κ̂) = A}
For all sets of addresses A ⊆ Addr
(q, K̂, ε, q0 ) ∈ δ iff Ĝ(q, κ̂) c
⇒ (q0 , κ̂) for any κ̂ ∈ K̂
(q, K̂, φ̂− , q0 ) ∈ δ iff Ĝ(q, φ̂ : κ̂) c
⇒ (q0 , κ̂) for any φ̂ : κ̂ ∈ K̂
(q, K̂, φ̂+ , q0 ) ∈ δ iff Ĝ(q, κ̂) c
⇒ (q0 , φ̂ : κ̂) for any κ̂ ∈ K̂
(q0 , hi) = Iˆ(e).
Assuming we can overcome the difficulty of computing with some representation of a
set of stacks, the intuition for the decidability of control-state reachability with garbage
collection stems from two observations: garbage collection operates on sets of addresses,
and for any given control point there is a finite number of sets of sets of addresses. The
finiteness makes the definition of δ fit the finiteness restriction of CPDSs. The regularity of
K̂ (for any given A, which we recall are finite sets) is apparent from a simple construction:
let the DFA control states represent the subsets of A, with 0/ the start state and A the
accepting state. Transition from A0 ⊆ A to A ∪ T (φ̂ ) for each φ̂ (no transition if the result
ZU064-05-FPR
paper-jfp
34
6 February 2018
3:38
J.I. Johnson, I. Sergey, C. Earl, M. Might, and D. Van Horn
is not a subset of A). Thus any string of frames that has a stack root of A (and only A) gets
accepted.
The last challenge to consider before we can delve into the mechanics of computing
reachable control states is how to represent the sets of stacks that may be paired with
each control state. Fortunately, a regular language can describe the stacks that share the
same root addresses, the set of stacks at a control point are recognized by a one-way
non-deterministic stack automaton (1NSA), and, fortuitously, non-empty overlap of these
two is decidable (but NP-hard (Rounds 1973)). The 1NSA describing the set of stacks
at a control point is already encoded in the structure of the (augmented) CRPDS that we
will accumulate while computing reachable control states. As we develop an algorithm for
control-state reachability, we will exploit this insight (Section 13).
12 Reachability in Conditional Pushdown Systems
We will show a progression of constructions that take us along the following line:
specialize
CPDS −−−→ CCPDS −−−−−→ PDCFA with GC → approx. PDCFA with GC
§12.2
§12.3
In the first construction, we show that a CCPDS is finitely constructible in a similar
fashion as in Section 7. The key is to take the current introspective CRPDS and “read off”
an automaton that describes the stacks accepted at each state. For traditional pushdown
systems, this is always an NFA, but introspection adds another feature: transition if the
string accepted so far is accepted by a given NFA. Such power falls outside of standard
NFAs and into one-way non-deterministic stack automata (1NSA)4 . These automata enjoy
closure under finite intersection with regular languages and decidable emptiness checking (Ginsburg et al. 1967), which we use to decide applicability of transition rules. If
the stacks realizable at q have a non-empty intersection with a set of stacks K̂ in a rule
(q, K̂, g, q0 ) ∈ δ , then there are paths from the start state to q that further reach q0 .
The structure of the GC problem allows us to sidestep the 1NSA constructions and more
directly compute state reachability. We specialize to garbage collection in subsection 12.3.
We finally show a space-saving approximation that our implementation uses.
12.1 One-way non-deterministic stack automata
The machinery we use for describing the realizable stacks at a state is a generalized
pushdown automaton itself. A stack automaton is permitted to move a cursor up and down
the stack and read frames (left and right on the input if two-way, only right if one-way),
but only push and pop when the stack cursor is at the top. Formally, a one-way stack
automaton is a 6-tuple A = (Q, Σ, Γ, δ , q0 , F) where
1.
2.
3.
4.
4
Q is a finite nonempty set of states,
Σ is a finite nonempty input alphabet,
Γ is a finite nonempty stack alphabet,
δ ⊆ Q × (Γ ∪ {ε}) × (Σ ∪ {ε}) × {↑, ·, ↓} × Γ± × Q is the transition relation,
The reachable states of a 1NSA is known to be regular, but the paths are not.
ZU064-05-FPR
paper-jfp
6 February 2018
3:38
PUSHDOWN FLOW ANALYSIS WITH ABSTRACT GARBAGE COLLECTION
35
5. q0 ∈ Q is the start state, and
6. F ⊆ Q the set of final states
An element of the transition relation, (q, φε , a, d, φ± , q0 ), should be read as, “if at q the
right of the stack cursor is prefixed by φε and the input is prefixed by a, then consume
a of the input, transition to state q0 , move the stack cursor in direction d, and if at the
top of the stack, perform stack action φ± .” This reading translates into a run relation
on instantaneous descriptions, Q × (Γ∗ × Γ∗ ) × Σ∗ . These descriptions are essentially
machine states that hold the current control state, the stack split around the cursor, and the
rest of the input.
(q, φε , a, d, φ± , q0 ) ∈ δ
φε v ΓT
w ≡ aw0
[Γ0B , Γ0T ] = P(φ± , D(d, [ΓB , ΓT ]))
(q, [ΓB , ΓT ], w) 7−→ (q0 , [Γ0B , Γ0T ], w0 )
where
P(φ+ , [ΓB , φε0 ]) = [ΓB φε0 , φ ]
D(↑, [ΓB , φ ΓT ]) = [ΓB φ , ΓT ]
P(φ− , [ΓB φ 0 , φ ]) = [ΓB , φ 0 ]
D(↓, [ΓB φ , ΓT ]) = [ΓB , φ ΓT ]
P(φ− , [ε, φ ]) = [ε, ε]
P(φ± , ΓB,T ) = ΓB,T
D(d, ΓB,T ) = ΓB,T
otherwise
otherwise
The meta-functions P and D perform the stack actions and direct the stack cursor,
respectively. A string w is thus accepted by a 1NSA A iff there are q ∈ F, ΓB , ΓT ∈ Γ∗
such that
(q0 , [ε, ε], w) 7−→∗ (q, [ΓB , ΓT ], ε)
Next we develop an introspective form of compact rooted pushdown systems that use
1NSAs for realizable stacks, and prove a correspondence with conditional pushdown systems.
12.2 Compact conditional pushdown systems
Similar to rooted pushdown systems, we say a conditional pushdown system G = (S, Γ, E, s0 )
is compact if all states, frames and edges are on some path from the root. We will refer to
this class of conditional pushdown systems as CCPDS. Assuming we have a way to decide
overlap between the set of realizable stacks at a state and a regular language of stacks, we
can compute the CCPDS in much the same way as in section 7.
F (M) = f , where
M = (Q, Γ, δ , q0 )
G
z }| {
f (S, Γ, E, s0 ) = (S0 , Γ, E 0 , s0 ), where
n
o
S0 = S ∪ s0 : s ∈ S and s 7−
−→
→ s0 ∪ {s0 }
M
K̂,g
K̂,g 0
0
0
E = E ∪ s s : s ∈ S and s 7−
−→
→ s and Stacks(G)(s) ∩ K̂ 6= 0/ .
M
ZU064-05-FPR
paper-jfp
36
6 February 2018
3:38
J.I. Johnson, I. Sergey, C. Earl, M. Might, and D. Van Horn
The function Stacks : CCPDS → S → 1NSA performs the stack extraction with a construction that inserts the stack-checking NFA for each reduction rule after it has run the
cursor to the bottom of the stack, and continues from the final states to the state dictated
by the rule (added by meta-function gadget). All the stack manipulations from s0 to s are
ε-transitions in terms of reading input; only once control reaches s do we check if the stack
is the same as the input, which captures the notion of a stack realizable at s. Once control
reaches s, we run down to the bottom of the stack again, and then match the stack against
the input; complete matches are accepted. To determine the bottom and top of the stack,
we add distinct sentinel symbols to the stack alphabet, ¢ and $.
G
z }| {
Stacks(S, Γ, E, s0 )(s) = (S ∪ S0 , Γ, Γ ∪ {¢, $} , δ , sstart , {sfinal }), where
sstart , sdown , scheck , sfinal fresh, and S0 , δ the smallest sets such that
{sstart , sdown , scheck , sfinal } ⊆ S0
(sstart , ε, ε, ·, ¢+ , s0 ) ∈ δ
0
gadget(s , K̂, γ± , s00 ) v (δ , S0 ) if (s0 , K̂, γ± , s00 ) ∈ E
(s, ε, ε, ·, $+ , sdown ) ∈ δ
(sdown , ε, ε, ↓, ε, sdown ) ∈ δ
(sdown , ¢, ε, ↑, ε, scheck ) ∈ δ
(scheck , a, a, ↑, ε, scheck ) ∈ δ , a ∈ Γ ∪ {ε}
(scheck , $, ε, ↑, ε, sfinal ) ∈ δ
The first rule changes the initial state to initialize the stack with the “bottom” sentinel.
Every reduction of the CPDS is given the gadget discussed above and explained below.
The last five rules are what implement the final checking of stack against input. When at
the state we are recognizing realizable stacks for, the machine will have the cursor at the
top of the stack, so we push the “top” sentinel before moving the cursor all the way down
to the bottom. When sdown finds the bottom sentinel at the cursor, it moves the cursor past
it to start the exact matching in scheck . If the cursor matches the input exactly, we consume
the input and move the cursor past the matched character to start again. When scheck finds
the top sentinel, it transitions to the final state; if the input is not completely exhausted, the
machine will get stuck and not accept.
ZU064-05-FPR
paper-jfp
6 February 2018
3:38
PUSHDOWN FLOW ANALYSIS WITH ABSTRACT GARBAGE COLLECTION
37
gadget(s, K̂, γ± , s0 ) = (δ 0 , Q ∪ {qdown , qout }) where
Let N = (Q, Σ, δ , q0 , F) be a fresh NFA recognizing K̂, qdown , qout fresh states
(q, a, ε, ↑, ε, q0 ) ∈ δ 0 if (q, a, q0 ) ∈ δ , a ∈ Σ
(q, ε, ε, ·, ε, q0 ) ∈ δ 0 if (q, ε, q0 ) ∈ δ
(q, $, ε, ·, $− , qout ) ∈ δ 0 if q ∈ F
(qout , ε, ε, ·, γ± , s0 ) ∈ δ 0
(s, ε, ε, ·, $+ , qdown ) ∈ δ 0
(qdown , ε, ε, ↓, ε, qdown ) ∈ δ 0
(qdown , ¢, ε, ↑, ε, q0 ) ∈ δ 0
We explain each rule in order. When the NFA that recognizes K̂ consumes a character,
the stack automaton should similarly read the character on the stack and move the cursor
along. If the NFA makes an ε-transition, the stack automaton should also, without moving
the stack cursor. When this sub-machine N is in a final state, the cursor should be at the
top of the stack (if indeed it matched), so we pop off the top sentinel and proceed to do
the stack action the IPDS does when transitioning to the next state. The last three rules
implement the same “run down to the bottom” gadget used before, when matching the
stack against the input.
Finally, we can show that states are reachable in a conditional pushdown system iff they
are reached in their corresponding CCPDS. Consider a map
C C : CPDS → CCPDS
such that given a conditional pushdown system M = (Q, Γ, δ , q0 ), its equivalent CCPDS is
C C (M) = (S, Γ, E, q0 ) where S contains reachable nodes:
o
n
S = q : (q0 , hi) 7−→∗ (q, κ̂)
M
and the set E contains reachable edges:
K̂,g
K̂,g
0
0
E = q q : q 7−
−→
→ q
M
Theorem 12.1 (Computable reachability)
For all M ∈ CPDS, C C (M) = lfp(F (M))
Proof in Appendix 18.2.
Corollary 12.1 (Realizable stacks of CPDSs are recognized by 1NSAs)
For all M = (Q, Γ, δ , q0 ) ∈ CPDS, and (S, Γ, E, q0 ) = lfp(F (M)), (q0 , hi) 7−→∗ (q, κ̂) iff
M
q ∈ S and Stacks(G)(q) accepts κ̂.
12.3 Simplifying garbage collection in conditional pushdown systems
The decision problems on 1NSAs are computationally intractable in general, but luckily
GC is a special problem where we do not need the full power of 1NSAs. There are equally
ZU064-05-FPR
paper-jfp
6 February 2018
38
3:38
J.I. Johnson, I. Sergey, C. Earl, M. Might, and D. Van Horn
fˆe ((P̂, Ê), Ĥ) = ((P̂0 , Ê 0 ), Ĥ 0 ), where
(
Ê+ =
φ̂+
)
A
0
(ψ̂, A) (ψ̂ , A ∪ T (φ̂ )) : ψ̂ + ψ̂
0
φ̂+
n
o
ε
A
Êε = (ψ̂, A) (ψ̂ 0 , A) : ψ̂ + ψ̂ 0
ε
φ̂+
φ̂−
A
Ê− = (ψ̂ 00 , A) (ψ̂ 000 , A0 ) : ψ̂ 00 + ψ̂ 000 and (ψ̂, A0 ) (ψ̂ 0 , A) ∈ Ê and (ψ̂ 0 , A) (ψ̂ 00 , A) ∈ Ĥ
φ̂−
0
Ê = Ê+ ∪ Êε ∪ Ê−
Ĥε = Ω̂ Ω̂00 : Ω̂ Ω̂0 ∈ Ĥ and Ω̂0 Ω̂00 ∈ Ĥ
φ̂+
Ĥ+− = Ω̂ Ω̂000 : Ω̂ Ω̂0 ∈ Ê and Ω̂0 Ω̂00 ∈ Ĥ
φ̂−
and Ω̂00 Ω̂000 ∈ Ê
Ĥ 0 = Ĥε ∪ Ĥ+−
n
o
g
P̂0 = P̂ ∪ Ω̂0 : Ω̂ Ω̂0 ∪ {((e, ⊥, ⊥), 0)}
/ .
Fig. 8: An ε-closure graph-powered iteration function for pushdown garbage-collecting
control-flow analysis
precise techniques at much lower cost, and less precise techniques that can shrink the
explored state space.5 The transition relation we build does not enumerate all sets of
addresses, but instead queries the graph for the sets of addresses it should consider in
order to apply GC. A fully precise method to manage the stack root addresses is to add
the root addresses to the representation of each state, and update it incrementally. The
root addresses can be seen as the representation of K̂ in edge labels, but to maintain the
precision, the set must also distinguish control states. This addition to the state space is an
effective reification of the stack filtering that conditional performs.
An approximative method is to not distinguish control states, but rather to traverse the
graph backward through ε-closure edges and push edges, and collect the root addresses
through the pushed frames. As more paths are discovered to control states, more stacks
will be realizable there, which add more to the stack root addresses to consider as the
relation steps forward. For soundness, edges still must be labeled with the language of
stacks they are valid for, since they can become invalid as more stacks reach control states.
Notice that the root sets of addresses are isomorphic to languages of stacks that have the
given root set, so we can use sets of addresses as the language representation.
We consider both methods in turn, augmenting the compaction algorithm from subsection 9.5. Each have program states that consist of the expression, environment, and store;
d × Store.
\ = Exp × Env
[ Since GC is a non-monotonic operation, stores cannot
ψ̂ ∈ PState
5
The added precision of GC with tighter working sets makes the state space comparison between
the two approaches non-binary. Neither approach is clearly better in terms of performance.
ZU064-05-FPR
paper-jfp
6 February 2018
3:38
39
PUSHDOWN FLOW ANALYSIS WITH ABSTRACT GARBAGE COLLECTION
be shared globally without sacrificing the precision benefits of GC. For the first method,
program states additionally include the stack root set of addresses; we will call these
[ We show the non-worklist
\ = PState
\ × P(Addr).
ornamented program states, Ω̂ ∈ OPState
solution to computing reachability by employing the function fˆe defined in Figure 8.
A
In order to define the iteration function, we need a refactored transition relation + ⊆
g
\ × PState,
\ defined as follows:
PState
A
(e, ρ̂, σ̂ ) + (e0 , ρ̂ 0 , σ̂ 0 ) iff StackRoot(κ̂) = A and Ĝ(e, ρ̂, σ̂ , κ̂) c
⇒ (e0 , ρ̂ 0 , σ̂ 0 , φ̂ : κ̂)
φ̂+
A
(e, ρ̂, σ̂ ) + (e0 , ρ̂ 0 , σ̂ 0 ) iff StackRoot(φ̂ : κ̂) = A and Ĝ(e, ρ̂, σ̂ , φ̂ : κ̂) c
⇒ (e0 , ρ̂ 0 , σ̂ 0 , κ̂)
φ̂−
A
ε
(e, ρ̂, σ̂ ) + (e0 , ρ̂ 0 , σ̂ 0 ) iff StackRoot(κ̂) = A and Ĝ(e, ρ̂, σ̂ , κ̂) c
⇒ (e0 , ρ̂ 0 , σ̂ 0 , κ̂)
Theorem 12.2 (Correctness of GC specialization)
0
lfp( fˆe ) completely abstracts C C (I \
PDS (e))
The approximative method changes
to contain
the representation of edges in the graph
[ = P PState
[ × Frame
\ × P(Addr)
\ . We also add
\ ± × PState
sets of addresses, Ê ∈ Edge
[ → (PState
[ to
\ → P(Addr))
\ → P(Addr)),
in a sub-fixed-point computation for tˆ : (PState
traverse the graph and collect the union of all stack roots for stacks realizable at a state.
Although we show a non-worklist solution here (in Figure 9) to not be distracting, this
solution will not compute the same reachable states as a worklist solution due to the evergrowing stack roots at each state. Only states in the worklist would need to be analyzed at
the larger stack root sets. In other words, the non-worklist solution potentially throws in
more live addresses at states that would otherwise not need to be re-examined.
This approximation is not an easily described introspective pushdown system since the
root sets it uses depend on the iteration state — particularly what frames have reached a
state so far, regardless of the stack filtering the original CPDS performs. The regular sets
of stacks acceptable at some state can be extracted a posteriori from the fixed point of the
function fˆe0 defined in Figure 9, if so desired. The next theorem follows from the fact that
R(ψ̂) ⊇ A for any represented (ψ̂, A).
Theorem 12.3 (Approximate GC is sound)
lfp( fˆe0 ) approximates lfp( fˆe ).
The last thing to notice is that by disregarding the filtering, the stack root set can get
larger and render previous GCs unsound, since more addresses can end up live than were
previously considered. Thus we label edges with the root set for which the GC was considered, in order to not make false predictions.
13 Implementing Introspective Pushdown Analysis with Garbage Collection
The reachability-based analysis for a pushdown system described in the previous section
requires two mutually-dependent pieces of information in order to add another edge:
ZU064-05-FPR
paper-jfp
6 February 2018
40
3:38
J.I. Johnson, I. Sergey, C. Earl, M. Might, and D. Van Horn
fˆe0 ((P̂, Ê), Ĥ) = ((P̂0 , Ê 0 ), Ĥ 0 ), where
(
)
A
A
0
ψ̂ ψ̂ : A = R(ψ̂), ψ̂ + ψ̂
Ê+ =
φ̂+
φ̂+
0
A
0
Êε = ψ̂ ψ̂ : A = R(ψ̂), ψ̂
ε
A
+ ψ̂ 0
ε
A
A
Ê− = ψ̂ 00 ψ̂ 000 : A = R(ψ̂ 00 ), ψ̂ 00 + ψ̂ 000 and
φ̂−
φ̂−
A0
ψ̂ ψ̂ 0 ∈ Ê and
φ̂+
0
ψ̂ ψ̂ 00 ∈ Ĥ
Ê 0 = Ê+ ∪ Êε ∪ Ê−
Ĥε = ψ̂ ψ̂ 00 : ψ̂ ψ̂ 0 ∈ Ĥ and ψ̂ 0 ψ̂ 00 ∈ Ĥ
A
Ĥ+− = ψ̂ ψ̂ 000 : ψ̂ ψ̂ 0 ∈ Ê and ψ̂ 0 ψ̂ 00 ∈ Ĥ
φ̂+
A0
and ψ̂ 00 ψ̂ 000 ∈ Ê
φ̂−
0
Ĥ = Ĥε ∪ Ĥ+−
A
P̂0 = P̂ ∪ ψ̂ 0 : ψ̂ ψ̂ 0 ∪ {(e, ⊥, ⊥)}
g
(
tˆ(R) = λ ψ̂.
0
0 A
( T (φ̂ ) ∪ R(ψ̂ ) : ψ̂ ψ̂ ∈ Ê
[
)
∪ R(ψ̂ 0 ) : ψ̂ 0 ψ̂ ∈ Ĥ )
φ̂+
R = lfp(tˆ).
Fig. 9: Approximate pushdown garbage-collecting control-flow analysis.
1. The topmost frame on a stack for a given control state q. This is essential for return
transitions, as this frame should be popped from the stack and the store and the
environment of a caller should be updated respectively.
2. Whether a given control state q is reachable or not from the initial state q0 along
realizable sequences of stack actions. For example, a path from q0 to q along edges
labeled “push, pop, pop, push” is not realizable: the stack is empty after the first pop,
so the second pop cannot happen—let alone the subsequent push.
Knowing about a possible topmost frame on a stack and initial-state reachability is
enough for a classic pushdown reachability summarization to proceed one step further,
and we presented an efficient algorithm to compute those in Section 8. However, to deal
with the presence of an abstract GC in a conditional PDS, we add:
3. For a given control state q, what are the touched addresses of all possible frames that
could happen to be on the stack at the moment the CPDS is in the state q?
ZU064-05-FPR
paper-jfp
6 February 2018
3:38
PUSHDOWN FLOW ANALYSIS WITH ABSTRACT GARBAGE COLLECTION
41
The crucial addition to the algorithm is maintaining for each node q0 in the CRPDS a
~g
set of ε-predecessors, i.e., nodes q, such that q 7−
−→
→ q0 and [~g] = ε. In fact, only two out of
M
three kinds of transitions can cause a change to the set of ε-predecessors for a particular
node q: an addition of an ε-edge or a pop edge to the CRPDS.
One can notice a subtle mutual dependency between computation of ε-predecessors and
top frames during the construction of a CRPDS. Informally:
• A top frame for a state q can be pushed as a direct predecessor (e.g., q follows a
nested let-binding), or as a direct predecessor to an ε-predecessor (e.g., q is in tail
position and will return to a waiting let-binding).
ε
• When a new ε-edge q →
− q0 is added, all ε-predecessors of q become also ε-predecessors
of q0 . That is, ε-summary edges are transitive.
γ−
• When a γ− -pop-edge q −→ q0 is added, new ε-predecessors of a state q1 can be
obtained by checking if q0 is an ε-predecessor of q1 and examining all existing εpredecessors of q, such that γ+ is their possible top frame: this situation is similar to
the one depicted in the example above.
The third component—the touched addresses of all possible frames on the stack for a
state q—is straightforward to compute with ε-predecessors: starting from q, trace out only
the edges which are labeled ε (summary or otherwise) or γ+ . The frame for any action γ+
in this trace is a possible stack action. Since these sets grow monotonically, it is easy to
cache the results of the trace, and in fact, propagate incremental changes to these caches
when new ε-summary or γ+ nodes are introduced. This implementation strategy captures
the approximative approach to performing GC, as discussed in the previous section. Our
implementation directly reflects the optimizations discussed above.
14 Experimental Evaluation
A fair comparison between different families of analyses should compare both precision
and speed. We have implemented a version k-CFA for a subset of R5RS Scheme and
instrumented it with a possibility to optionally enable pushdown analysis, abstract garbage
collection or both. Our implementation source (in Scala) and benchmarks are available:
http://github.com/ilyasergey/reachability
In the experiments, we have focused on the version of k-CFA with a per-state store (i.e.,
without widening), as in the presence of single-threaded store, the effect of abstract GC
is neutralized due to merging. For non-widened versions of k-CFA, as expected, the fused
analysis does at least as well as the best of either analysis alone in terms of singleton flow
sets (a good metric for program optimizability) and better than both in some cases. Also
worthy of note is the dramatic reduction in the size of the abstract transition graph for the
fused analysis—even on top of the already large reductions achieved by abstract garbage
collection and pushdown flow analysis individually. The size of the abstract transition
graph is a good heuristic measure of the temporal reasoning ability of the analysis, e.g.,
its ability to support model-checking of safety and liveness properties (Might et al. 2007).
ZU064-05-FPR
paper-jfp
6 February 2018
42
3:38
J.I. Johnson, I. Sergey, C. Earl, M. Might, and D. Van Horn
Program #e #v k
mj09
19
8
eta
21 13
kcfa2
20 10
kcfa3
25 13
blur
40 20
loop2
41 14
sat
51 23
0
1
0
1
0
1
0
1
0
1
0
1
0
1
k-CFA
83
107
454
812
63
74
33
33
194
236
970
1935
272
327
> 32662 > 88548
4686
7606
123
149
149
163
> 10867 > 16040
3844
5547
> 28432 > 37391
k-CFA + GC
k-PDCFA
4
1
4
8
3
1
4
–
4
10
7
–
4
–
38
38
44
48
32
32
32
31
36
35
87
144
58
63
1761 4046
115
146
94
101
69
73
411
525
545
773
12828 16846
4
36
39
1
34
35
6
28
27
8
28
27
4
35
34
2
35
34
5
53
52
2
53
52
4
90
95
10 76
82
7
43
46
3 151 163
4 1137 1543
4 958 1314
4
1
8
8
4
2
5
2
10
10
7
3
4
5
k-PDCFA + GC
33
32
28
28
35
35
53
53
68
75
34
145
254
71
32
31
27
27
34
34
52
52
76
81
35
156
317
73
4
1
8
8
4
2
5
2
10
10
7
3
4
10
Table 1: Benchmark results for toy programs. The first three columns provide the name
of a benchmark, the number of expressions and variables in the program in the ANF,
respectively. For each of eight combinations of pushdown analysis, k ∈ {0, 1} and garbage
collection on or off, the first two columns in a group show the number of control states
and transitions/CRPDS edges computed during the analysis (for both less is better). The
third column presents the amount of singleton variables, i.e, how many variables have a
single lambda flow to them (more is better). Inequalities for some results of the plain kCFA denote the case when the analysis explored more than 105 configurations (i.e., control
states coupled with continuations) or did not finish within 30 minutes. For such cases we
do not report on singleton variables.
14.1 Plain k-CFA vs. pushdown k-CFA
In order to exercise both well-known and newly-presented instances of CESK-based CFAs,
we took a series of small benchmarks exhibiting archetypal control-flow patterns (see
Table 1). Most benchmarks are taken from the CFA literature: mj09 is a running example
from the work of Midtgaard and Jensen designed to exhibit a non-trivial return-flow behavior (Midtgaard 2007), eta and blur test common functional idioms, mixing closures and
eta-expansion, kcfa2 and kcfa3 are two worst-case examples extracted from the proof of
k-CFA’s EXPTIME hardness (Van Horn and Mairson 2008), loop2 is an example from
Might’s dissertation that was used to demonstrate the impact of abstract GC (Might 2007,
Section 13.3), sat is a brute-force SAT-solver with backtracking.
14.1.1 Comparing precision
In terms of precision, the fusion of pushdown analysis and abstract garbage collection
substantially cuts abstract transition graph sizes over one technique alone.
We also measure singleton flow sets as a heuristic metric for precision. Singleton flow
sets are a necessary precursor to optimizations such as flow-driven inlining, type-check
elimination and constant propagation. It is essential to notice that for the experiments in
ZU064-05-FPR
paper-jfp
6 February 2018
3:38
PUSHDOWN FLOW ANALYSIS WITH ABSTRACT GARBAGE COLLECTION
43
Table 1 our implementation computed the sets of values (i.e., closures) assigned to each
variable (as opposed to mere syntactic lambdas). This is why in some cases the results
computed by 0CFA appear to be better than those by 1CFA: the later one may examine
more states with different environments, which results in exploring more different values,
whereas the former one will just collapse all these values to a single lambda (Might et al.
2010). What is important is that for a fixed k the fused analysis prevails as the best-of- or
better-than-both-worlds.
Running on the benchmarks, we have re-validated hypotheses about the improvements to
precision granted by both pushdown analysis (Vardoulakis and Shivers 2010) and abstract
garbage collection (Might 2007). Table 1 contains our detailed results on the precision of
the analysis. In order to make the comparison fair, in the table we report on the numbers of
control states, which do not contain a stack component and are the nodes of the constructed
CRPDS. In the case of plain k-CFA, control states are coupled with stack pointers to obtain
configurations, whose resulting number is significantly bigger.
The SAT-solving benchmark showed a dramatic improvement with the addition of contextsensitivity. Evaluation of the results showed that context-sensitivity provided enough fuel
to eliminate most of the non-determinism from the analysis.
14.1.2 Comparing speed
In the original work on CFA2, Vardoulakis and Shivers present experimental results with
a remark that the running time of the analysis is proportional to the size of the reachable
states (Vardoulakis and Shivers 2010, Section 6). There is a similar correlation in our fused
analysis, but with higher variance due to the live address set computation GC performs.
Since most of the programs from our toy suite run for less than a second, we do not report
on the absolute time. Instead, the histogram on Figure 10 presents normalized relative
times of analyses’ executions. To our observation the pure machine-style k-CFA is always
significantly worse in terms of execution time than either with GC or push-down system,
so we excluded the plain, non-optimized k-CFA from the comparison.
Our earlier implementation of a garbage-collecting pushdown analysis (Earl et al. 2012)
did not fully exploit the opportunities for caching ε-predecessors, as described in Section 13. This led to significant inefficiencies of the garbage-collecting analyzer with respect
to the regular k-CFA, even though the former one observed a smaller amount of states and
in some cases found larger amounts of singleton variables. After this issue has been fixed,
it became clearly visible that in all cases the GC-optimized analyzer is strictly faster than
its non-optimized pushdown counterpart.
Although caching of ε-predecessors and ε-summary edges is relatively cheap, it is not
free, since maintaining the caches requires some routine machinery at each iteration of
the analyzer. This explains the loss in performance of the garbage-collecting pushdown
analysis with respect to the GC-optimized k-CFA.
As it follows from the plot, fused analysis is always faster than the non-garbage-collecting
pushdown analysis, and about a fifth of the time, it beats k-CFA with garbage collection in
terms of performance. When the fused analysis is slower than just a GC-optimized one, it is
generally not much worse than twice as slow as the next slowest analysis. Given the already
paper-jfp
6 February 2018
44
3:38
J.I. Johnson, I. Sergey, C. Earl, M. Might, and D. Van Horn
�����������������������������������
�����������������������������������
���
���
����
����
�������
�������
����������
����������
������������
������������
��
���
���
��
��
���
���
��
����
����
���
���
�����
�����
�����
�����
����
����
�����
�����
���
���
�����
�����
�����
�����
����
����
�����
�����
���
���
���
���
�����������������������������������
�����������������������������������
ZU064-05-FPR
����
����
�������
�������
����������
����������
������������
������������
��
���
���
��
��
���
���
��
����
����
���
���
Fig. 10: Analysis times relative to worst (= 1) in class; smaller is better. At the top is
the monovariant 0CFA class of analyses, at the bottom is the polyvariant 1CFA class of
analyses. (Non-GC k-CFA omitted.)
substantial reductions in analysis times provided by collection and pushdown analysis, the
amortized penalty is a small and acceptable price to pay for improvements to precision.
14.2 Analyzing real-world programs with garbage-collecting pushdown CFA
Even though our prototype implementation is just a proof of concept, we evaluated it not
on a suite of toy programs, tailored for particular functional programming patterns, but on a
ZU064-05-FPR
paper-jfp
6 February 2018
3:38
PUSHDOWN FLOW ANALYSIS WITH ABSTRACT GARBAGE COLLECTION
Program
#e
primtest 155
rsa
211
regex
344
scm2java 1135
#v !#v
44
93
150
460
16
36
44
63
k = 0, GC off
790 955 1400
1267 1507 2300
943 956 5400
376 375 1300
k = 0, GC on
113
355
578
376
k = 1, GC off
127 100 >43146 >54679 ∞
407 600
20746
28895 210
589 4500
1153
1179
8800
375 1300
376
375
1400
45
k = 1, GC on
442 562 1300
926 1166 2800
578 589 5000
376 375 1300
Table 2: Benchmark results of PDCFA on real-world programs. The first four columns
provide the name of a program, the number of expressions and variables in the program in
the ANF, and the number of singleton variables revealed by the analysis (same in all cases).
For each of four combinations of k ∈ {0, 1} and garbage collection on or off, the first two
columns in a group show the number of visited control states and edges, respectively, and
the third one shows absolute time of running the analysis (for both less is better). The
results of the analyses are presented in minutes (0 ) or seconds (00 ), where ∞ stands for an
analysis, which has been interrupted due to the an execution time greater than 30 minutes.
set of real-world programs. In order to set this experiment, we have chosen four programs,
dealing with numeric and symbolic computations:
• primtest – an implementation of the probabilistic Fermat and Solovay-Strassen
primality testing in Scheme for the purpose of large prime generation;
• rsa – an implementation of the RSA public-key cryptosystem;
• regex – an implementation of a regular expression matcher in Scheme using Brzozovski derivatives (Might et al. 2011; Owens et al. 2009);
• scm2java – scm2java is a Scheme to Java compiler;
We ran our benchmark suite on a 2.7 GHz Intel Core i7 OS X machine with 8 Gb
RAM. Unfortunately, k-CFA without global stores timed out on most of these examples
(i.e., did not finish within 30 minutes), so we had to exclude it from the comparison and
focus on the effect of a pushdown analyzer only. Table 2 presents the results of running the
benchmarks for k ∈ {0, 1} with a garbage collector on and off. Surprisingly, for each of the
six programs, those cases, which terminated within 30 minutes, found the same number
of global singleton variables.6 However, the numbers of observed states and runtimes are
indeed different in most of the cases except scm2java, for which all the four versions of
the analysis were precise enough to actually evaluate the program: happily, there was no
reuse of abstract addresses, which resulted in the absence of forking during the CRPDS
construction. In other words, the program scm2java, which used no scalar data but strings
being concatenated and was given a simple input, has been evaluated precisely, which is
confirmed by the number of visited control states and edges.
Time-wise, the results of the experiment demonstrate the general positive effect of the
abstract garbage collection in a pushdown setting, which might improve the analysis performance by more than two orders of magnitude.
6
Of course, the numbers of states explored for each program by different analyses were different,
and there were variations in function parameters cardinalities, which we do not report on here.
ZU064-05-FPR
paper-jfp
46
6 February 2018
3:38
J.I. Johnson, I. Sergey, C. Earl, M. Might, and D. Van Horn
15 Discussion: Applications
Pushdown control-flow analysis offers more precise control-flow analysis results than the
classical finite-state CFAs. Consequently, introspective pushdown control-flow analysis
improves flow-driven optimizations (e.g., constant propagation, global register allocation,
inlining (Shivers 1991)) by eliminating more of the false positives that block their application.
The more compelling applications of pushdown control-flow analysis are those which
are difficult to drive with classical control-flow analysis. Perhaps not surprisingly, the best
examples of such analyses are escape analysis and interprocedural dependence analysis.
Both of these analyses are limited by a static analyzer’s ability to reason about the stack,
the core competency of introspective pushdown control-flow analysis. (We leave an indepth formulation and study of these analyses to future work.)
15.1 Escape analysis
In escape analysis, the objective is to determine whether a heap-allocated object is safely
convertible into a stack-allocated object. In other words, the compiler is trying to figure out
whether the frame in which an object is allocated outlasts the object itself. In higher-order
languages, closures are candidates for escape analysis.
Determining whether all closures over a particular λ -term lam may be heap-allocated is
straightforward: find the control states in the compact rooted pushdown system in which
closures over lam are being created, then find all control states reachable from these states
over only ε-edge and push-edge transitions. Call this set of control states the “safe” set.
Now find all control states which are invoking a closure over lam. If any of these control
states lies outside of the safe set, then stack-allocation may not be safe; if, however, all
invocations lie within the safe set, then stack-allocation of the closure is safe.
15.2 Interprocedural dependence analysis
In interprocedural dependence analysis, the goal is to determine, for each λ -term, the set
of resources which it may read or write when it is called. Might and Prabhu (2009) showed
that if one has knowledge of the program stack, then one can uncover interprocedural
dependencies. We can adapt that technique to work with compact rooted pushdown systems. For each control state, find the set of reachable control states along only ε-edges and
pop-edges. The frames on the pop-edges determine the frames which could have been on
the stack when in the control state. The frames that are live on the stack determine the
procedures that are live on the stack. Every procedure that is live on the stack has a readdependence on any resource being read in the control state, while every procedure that is
live on the stack also has a write-dependence on any resource being written in the control
state. In control-flow terms, this translates to “if f calls g and g accesses a, then f also
accesses a.”
ZU064-05-FPR
paper-jfp
6 February 2018
3:38
PUSHDOWN FLOW ANALYSIS WITH ABSTRACT GARBAGE COLLECTION
47
16 Related Work
The Scheme workshop presentation of PDCFA (Earl et al. 2010) is not archival, nor were
there rigorous proofs of correctness. The complete development of pushdown analysis from
first principles stands as a new contribution, and it constitutes an alternative to CFA2. It
goes beyond work on CFA2 by specifying specific mechanisms for reducing the complexity
to polynomial time (O(n6 )) as well. Vardoulakis (2012) sketches an approach to regain
polynomial time in his dissertation, but does not give a precise bound. An immediate advantage of the complete development is its exposure of parameters for controlling polyvariance and context-sensitivity. An earlier version of this work appeared in ICFP 2012 (Earl
et al. 2012). We also provide a reference implementation of control-state reachability in
Haskell. We felt this was necessary to shine a light on the “dark corners” in the formalism,
and in fact, it helped expose both bugs and implicit design decisions that were reflected
in the revamped text of this work. The development of introspective pushdown systems is
also more complete and more rigorous. We expose the critical regularity constraint absent
from the ICFP 2012 work, and we specify the implementation of control-state reachability
and feasible paths for conditional pushdown systems in greater detail. More importantly,
this work uses additional techniques to improve the performance of the implementation
and discusses those changes.
Garbage-collecting pushdown control-flow analysis draws on work in higher-order controlflow analysis (Shivers 1991), abstract machines (Felleisen and Friedman 1987) and abstract
interpretation (Cousot and Cousot 1977).
Context-free analysis of higher-order programs The motivating work for our own is
Vardoulakis and Shivers recent discovery of CFA2. CFA2 is a table-driven summarization
algorithm that exploits the balanced nature of calls and returns to improve return-flow precision in a control-flow analysis. Though CFA2 exploits context-free languages, contextfree languages are not explicit in its formulation in the same way that pushdown systems
are explicit in our presentation of pushdown flow analysis. With respect to CFA2, our pushdown flow analysis is also polyvariant/context-sensitive (whereas CFA2 is monovariant/contextinsensitive), and it covers direct-style.
On the other hand, CFA2 distinguishes stack-allocated and store-allocated variable bindings, whereas our formulation of pushdown control-flow analysis does not: it allocates all
bindings in the store. If CFA2 determines a binding can be allocated on the stack, that
binding will enjoy added precision during the analysis and is not subject to merging like
store-allocated bindings. While we could incorporate such a feature in our formulation, it
is not necessary for achieving “pushdownness,” and in fact, it could be added to classical
finite-state CFAs as well.
CFA2 has a follow-up that sacrifices its complete abstraction with the machine that only
abstracts bindings in order to handle first-class control (Vardoulakis and Shivers 2011).
We do not have an analogous construction since loss of complete abstraction was an antigoal of this work. We leave an in-depth study of generalizations of CFA2’s method to
introspection, polyvariance and other control operators to future work.
ZU064-05-FPR
paper-jfp
48
6 February 2018
3:38
J.I. Johnson, I. Sergey, C. Earl, M. Might, and D. Van Horn
Calculation approach to abstract interpretation Midtgaard and Jensen (2009) systematically calculate 0CFA using the Cousot-style calculational approach to abstract interpretation (Cousot 1999) applied to an ANF λ -calculus. Like the present work, Midtgaard and
Jensen start with the CESK machine of Flanagan et al. (1993) and employ a reachablestates model.
The analysis is then constructed by composing well-known Galois connections to reveal a 0CFA incorporating reachability. The abstract semantics approximate the control
stack component of the machine by its top element. The authors remark monomorphism
materializes in two mappings: one “mapping all bindings to the same variable,” the other
“merging all calling contexts of the same function.” Essentially, the pushdown 0CFA of
Section 4 corresponds to Midtgaard and Jensen’s analysis when the latter mapping is
omitted and the stack component of the machine is not abstracted. However, not abstracting
the stack requires non-trivial mechanisms to compute the compaction of the pushdown
system.
CFL- and pushdown-reachability techniques This work also draws on CFL- and pushdownreachability analysis (Bouajjani et al. 1997; Kodumal and Aiken 2004; Reps 1998; Reps
et al. 2005). For instance, ε-closure graphs, or equivalent variants thereof, appear in many
context-free-language and pushdown reachability algorithms. For our analysis, we implicitly invoked these methods as subroutines. When we found these algorithms lacking (as
with their enumeration of control states), we developed rooted pushdown system compaction.
CFL-reachability techniques have also been used to compute classical finite-state abstraction CFAs (Melski and Reps 2000) and type-based polymorphic control-flow analysis (Rehof and Fähndrich 2001). These analyses should not be confused with pushdown
control-flow analysis, which is computing a fundamentally more precise kind of CFA.
Moreover, Rehof and Fähndrich’s method is cubic in the size of the typed program, but the
types may be exponential in the size of the program. Finally, our technique is not restricted
to typed programs.
Model-checking pushdown systems with checkpoints A pushdown system with checkpoints has designated finite automata for state/frame pairs. If in a given state/frame configuration, and the automaton accepts the current stack, then execution continues. This
model was first created in Esparza et al. (2003) and describes its applications to modelchecking programs that use Java’s AccessController class, and performing better dataflow analysis of Lisp programs with dynamic scope, though the specific applications are
not fully explored. The algorithm described in the paper is similar to ours, but not “on-thefly,” however, so such applications would be difficult to realize with their methods. The
algorithm discussed has multiple loops that enumerate all transitions within the pushdown
system considered. Again this is a non-starter for higher-order languages, since up-front
enumeration would conservatively suggest that any binding called would resolve to any
possible function. This strategy is a sure-fire way to destroy precision and performance.
Meet-over-all-paths for conditional weighted pushdown systems A conditional pushdown system is essentially a pushdown system in which every state/frame pair is a check-
ZU064-05-FPR
paper-jfp
6 February 2018
3:38
PUSHDOWN FLOW ANALYSIS WITH ABSTRACT GARBAGE COLLECTION
49
point. The two are easily interchangeable, but weighted conditional pushdown systems
assign weights to reduction rules from a bounded idempotent semiring in the same manner
as Reps et al. (2005). The work that introduces CWPDSs uses them for points-to analysis
for Java. They solve the meet-over-all-paths problem by an incrementally translating a
skeleton CFG into a WPDS and using WPDS++ (Lal and Reps 2006) to discover more
points-to information to fill in call/return edges. The translation involves a heavy encoding
and is not obviously correct. The killer for its use for GC is that it involves building
the product automaton of all the (minimized) condition automata for the system, and
interleaving the system states with the automaton’s states — there are exponentially many
such machines in our case, and even though the overall solution is incremental, this large
automaton is pre-built. It is not obvious how to incrementalize the whole construction, nor
is it obvious that the precision and performance are not negatively impacted by the repeated
invocation of the WPDS solver (as opposed to a work-set solution that only considers
recently changed states).
The approach to incremental solving using first-order tools is an interesting approach
that we had not considered. Perhaps first-order and higher-order methods are not too far
removed. It is possible that these frameworks could be extended to request transitions —
or even further, checkpoint machines — on demand in order to better support higher-order
languages. As we saw in this article, however, we needed access to internal data structures
to compute root sets of addresses, and the ability to update a cache of such sets in these
structures. The marriage could be rocky, but worth exploring in order to unite the two
communities and share technologies.
Model-checking higher-order recursion schemes There is terminology overlap with
work by Kobayashi (2009) on model-checking higher-order programs with higher-order
recursion schemes, which are a generalization of context-free grammars in which productions can take higher-order arguments, so that an order-0 scheme is a context-free grammar.
Kobyashi exploits a result by Ong (2006) which shows that model-checking these recursion schemes is decidable (but ELEMENTARY-complete) by transforming higher-order
programs into higher-order recursion schemes.
Given the generality of model-checking, Kobayashi’s technique may be considered an
alternate paradigm for the analysis of higher-order programs. For the case of order-0, both
Kobayashi’s technique and our own involve context-free languages, though ours is for
control-flow analysis and his is for model-checking with respect to a temporal logic. After
these surface similarities, the techniques diverge. In particular, higher-order recursions
schemes are limited to model-checking programs in the simply-typed lambda-calculus with
recursion.
17 Conclusion
Our motivation was to further probe the limits of decidability for pushdown flow analysis
of higher-order programs by enriching it with abstract garbage collection. We found that
abstract garbage collection broke the pushdown model, but not irreparably so. By casting
abstract garbage collection in terms of an introspective pushdown system and synthesizing
ZU064-05-FPR
paper-jfp
50
6 February 2018
3:38
J.I. Johnson, I. Sergey, C. Earl, M. Might, and D. Van Horn
a new control-state reachability algorithm, we have demonstrated the decidability of fusing
two powerful analytic techniques.
As a byproduct of our formulation, it was also easy to demonstrate how polyvariant/contextsensitive flow analyses generalize to a pushdown formulation, and we lifted the need to
transform to continuation-passing style in order to perform pushdown analysis.
Our empirical evaluation is highly encouraging: it shows that the fused analysis provides further large reductions in the size of the abstract transition graph—a key metric for
interprocedural control-flow precision. And, in terms of singleton flow sets—a heuristic
metric for optimizability—the fused analysis proves to be a “better-than-both-worlds”
combination.
Thus, we provide a sound, precise and polyvariant introspective pushdown analysis for
higher-order programs.
Acknowledgments
We thank the anonymous reviewers of ICFP 2012 and JFP for their detailed reviews, which
helped to improve the presentation and technical content of the paper. Tim Smith was
especially helpful with his knowledge of stack automata. This material is based on research
sponsored by DARPA under the programs Automated Program Analysis for Cybersecurity (FA8750-12-2-0106) and Clean-Slate Resilient Adaptive Hosts (CRASH). The U.S.
Government is authorized to reproduce and distribute reprints for Governmental purposes
notwithstanding any copyright notation thereon.
References
Bouajjani, A., J. Esparza, and O. Maler (1997). Reachability analysis of pushdown
automata: Application to Model-Checking. In Proceedings of the 8th International
Conference on Concurrency Theory, CONCUR ’97, pp. 135–150. Springer-Verlag.
Cousot, P. (1999). The calculational design of a generic abstract interpreter. In M. Broy and
R. Steinbrüggen (Eds.), Calculational System Design. NATO ASI Series F. IOS Press,
Amsterdam.
Cousot, P. and R. Cousot (1977). Abstract interpretation: A unified lattice model for static
analysis of programs by construction or approximation of fixpoints. In Conference
Record of the Fourth ACM Symposium on Principles of Programming Languages, pp.
238–252. ACM Press.
Earl, C., M. Might, and D. Van Horn (2010). Pushdown Control-Flow analysis of HigherOrder programs. In Workshop on Scheme and Functional Programming.
Earl, C., I. Sergey, M. Might, and D. Van Horn (2012). Introspective pushdown analysis
of higher-order programs. In Proceedings of the 17th ACM SIGPLAN International
Conference on Functional Programming (ICFP 2012), ICFP ’12, pp. 177–188. ACM.
Esparza, J., A. Kucera, and S. Schwoon (2003). Model checking LTL with regular
valuations for pushdown systems. Inf. Comput. 186(2), 355–376.
Felleisen, M. and D. P. Friedman (1987). A calculus for assignments in higher-order
languages. In POPL ’87: Proceedings of the 14th ACM SIGACT-SIGPLAN Symposium
on Principles of Programming Languages, pp. 314+. ACM.
ZU064-05-FPR
paper-jfp
6 February 2018
3:38
PUSHDOWN FLOW ANALYSIS WITH ABSTRACT GARBAGE COLLECTION
51
Flanagan, C., A. Sabry, B. F. Duba, and M. Felleisen (1993, June). The essence of
compiling with continuations. In PLDI ’93: Proceedings of the ACM SIGPLAN 1993
Conference on Programming Language Design and Implementation, pp. 237–247.
ACM.
Ginsburg, S., S. A. Greibach, and M. A. Harrison (1967). One-way stack automata. Journal
of the ACM 14(2), 389–418.
Johnson, J. I. and D. Van Horn (2013). Concrete semantics for pushdown analysis: The
essence of summarization. In HOPA 2013: Workshop on higher-order program analysis.
Kobayashi, N. (2009, January). Types and higher-order recursion schemes for verification
of higher-order programs. SIGPLAN Not. 44(1), 416–428.
Kodumal, J. and A. Aiken (2004, June). The set constraint/CFL reachability connection
in practice. In PLDI ’04: Proceedings of the ACM SIGPLAN 2004 Conference on
Programming Language Design and Implementation, pp. 207–218. ACM.
Lal, A. and T. W. Reps (2006). Improving pushdown system model checking. In T. Ball
and R. B. Jones (Eds.), CAV, Volume 4144 of Lecture Notes in Computer Science, pp.
343–357. Springer.
Li, X. and M. Ogawa (2010). Conditional weighted pushdown systems and applications.
In J. P. Gallagher and J. Voigtländer (Eds.), PEPM, pp. 141–150. ACM.
Melski, D. and T. W. Reps (2000, October). Interconvertibility of a class of set constraints
and context-free-language reachability. Theoretical Computer Science 248(1-2), 29–98.
Midtgaard, J. (2007). Transformation, Analysis, and Interpretation of Higher-Order
Procedural Programs. Ph. D. thesis, University of Aarhus.
Midtgaard, J. and T. P. Jensen (2009). Control-flow analysis of function calls and returns
by abstract interpretation. In ICFP ’09: Proceedings of the 14th ACM SIGPLAN
International Conference on Functional Programming, pp. 287–298. ACM.
Might, M. (2007, June). Environment Analysis of Higher-Order Languages. Ph. D. thesis,
Georgia Institute of Technology.
Might, M., B. Chambers, and O. Shivers (2007, January). Model checking via GammaCFA. In Verification, Model Checking, and Abstract Interpretation, pp. 59–73.
Might, M., D. Darais, and D. Spiewak (2011). Parsing with derivatives: a functional
pearl. In ICFP ’11: Proceeding of the 16th ACM SIGPLAN international conference
on Functional Programming, pp. 189–195. ACM.
Might, M. and P. Manolios (2009). A posteriori soundness for non-deterministic abstract
interpretations. In Proceedings of the 10th International Conference on Verification,
Model Checking, and Abstract Interpretation, VMCAI ’09, pp. 260–274. SpringerVerlag.
Might, M. and T. Prabhu (2009). Interprocedural dependence analysis of higher-order
programs via stack reachability. In Proceedings of the 2009 Workshop on Scheme and
Functional Programming.
Might, M. and O. Shivers (2006a). Environment analysis via Delta-CFA. In Conference
Record of the 33rd ACM SIGPLAN-SIGACT Symposium on Principles of Programming
Languages (POPL 2006), pp. 127–140. ACM.
Might, M. and O. Shivers (2006b). Improving flow analyses via Gamma-CFA: Abstract
garbage collection and counting.
In Proceedings of the 11th ACM SIGPLAN
International Conference on Functional Programming (ICFP 2006), pp. 13–25. ACM.
ZU064-05-FPR
paper-jfp
52
6 February 2018
3:38
J.I. Johnson, I. Sergey, C. Earl, M. Might, and D. Van Horn
Might, M., Y. Smaragdakis, and D. Van Horn (2010). Resolving and exploiting the kCFA paradox: Illuminating functional vs. object-oriented program analysis. In PLDI
’10: Proceedings of the 2010 ACM SIGPLAN Conference on Programming Language
Design and Implementation, PLDI ’10, pp. 305–315. ACM Press.
Ong, C. H. L. (2006). On Model-Checking trees generated by Higher-Order recursion
schemes. In 21st Annual IEEE Symposium on Logic in Computer Science (LICS’06),
LICS, pp. 81–90. IEEE.
Owens, S., J. Reppy, and A. Turon (2009). Regular-expression derivatives re-examined.
Journal of Functional Programming 19(02), 173–190.
Rehof, J. and M. Fähndrich (2001). Type-based flow analysis: From polymorphic
subtyping to CFL-reachability. In POPL ’01: Proceedings of the 28th ACM SIGPLANSIGACT Symposium on Principles of Programming Languages, pp. 54–66. ACM.
Reps, T. (1998, December). Program analysis via graph reachability. Information and
Software Technology 40(11-12), 701–726.
Reps, T., S. Schwoon, S. Jha, and D. Melski (2005, October). Weighted pushdown
systems and their application to interprocedural dataflow analysis. Science of Computer
Programming 58(1-2), 206–263.
Rounds, W. C. (1973). Complexity of recognition in intermediate level languages. In
Switching and Automata Theory, 1973. SWAT ’08. IEEE Conference Record of 14th
Annual Symposium on, pp. 145–158.
Shivers, O. G. (1991). Control-Flow Analysis of Higher-Order Languages. Ph. D. thesis,
Carnegie Mellon University.
Sipser, M. (2005, February). Introduction to the Theory of Computation (2 ed.). Cengage
Learning.
Van Horn, D. and H. G. Mairson (2008). Deciding kCFA is complete for EXPTIME.
In ICFP ’08: Proceeding of the 13th ACM SIGPLAN International Conference on
Functional Programming, pp. 275–282. ACM.
Van Horn, D. and M. Might (2012). Systematic abstraction of abstract machines. Journal
of Functional Programming 22(Special Issue 4-5), 705–746.
Vardoulakis, D. (2012). CFA2: Pushdown Flow Analysis for Higher-Order Languages. Ph.
D. thesis, Northeastern University.
Vardoulakis, D. and O. Shivers (2010). CFA2: A Context-Free approach to ControlFlow analysis. In A. D. Gordon (Ed.), Programming Languages and Systems, Volume
6012 of Lecture Notes in Computer Science, Chapter 30, pp. 570–589. Springer Berlin
Heidelberg.
Vardoulakis, D. and O. Shivers (2011). Pushdown flow analysis of first-class control.
In Proceedings of the 16th ACM SIGPLAN International Conference on Functional
Programming, ICFP ’11, pp. 69–80. ACM.
Wright, A. K. and S. Jagannathan (1998, January). Polymorphic splitting: An effective
polyvariant flow analysis. ACM Transactions on Programming Languages and
Systems 20(1), 166–207.
ZU064-05-FPR
paper-jfp
6 February 2018
3:38
PUSHDOWN FLOW ANALYSIS WITH ABSTRACT GARBAGE COLLECTION
53
18 Full Proofs
18.1 Pushdown reachability
Proof of Theorem 8.1. The space ICRPDS is further constrained than stated in the main
article:
S
{{s, s0 } : s s0 ∈ H} ⊆ S,
ICRPDS = ((S, E), H, (∆S, ∆E, ∆H)) :
∆S ∩ S = 0,
/ ∆E ∩ E = 0,
/ and ∆H ∩ H = 0/
For this section we assume
M = (Q, Γ, δ , q0 ) ∈ RPDS
G = ((S, E), H, (∆S, ∆E), ∆H) ∈ CRPDS where (S, E) ⊆ (Q, δ )
and q0 = s0
Let ORD be the class of ordinal numbers. We define a termination measure on the fixedpoint computation of F 0 ((Q, Γ, , )), d : CRPDS → ORD.
2 ·|Γ|
d((S, E), H, (∆S, ∆E, ∆H)) = (2|Q|
2
− |E|)ω + (2|Q| − |H|)
Lemma 18.1 (Termination)
Either G = F 0 (M)(G) or d(F 0 (M)(G)) ≺ d(G)
Proof
If both ∆E and ∆H are empty, there are no additions made to S, E or H, meaning G is a
fixed point. Otherwise, due to the non-overlap condition, one or both of E and H grow,
meaning the ordinal is smaller.
A corollary is that the fixed-point has empty ∆E and ∆H.
Lemma 18.2 (Key lemma for PDS reachability)
If inv(G) then inv(F 0 (M)(G))
Proof
All additional states and edges come from ∆Ei and ∆Hi for i ∈ [0..4], so by cases on the
sources of edges:
g
Case s s0 ∈ ∆E0 , s00 s000 ∈ ∆H0 .
By definition of sprout and path extension.
g
Case s s0 ∈ ∆E1 , s00 s000 ∈ ∆H1 .
φ̂+
If g ≡ φ̂− , then by definition of addPush there are q q0 ∈ ∆E, q0 s ∈ H, such
that (s, φ̂+ , s0 ) ∈ δ .
Let ~g be the witness of the invariant on q0 s given from definition of inv. Let κ̂ be
φ̂+~gφ̂−
arbitrary. We have [φ̂+~gφ̂− ] = ε. We also have (q, κ̂) 7−→∗ (s0 , κ̂). Root reachability
M
φ̂+
follows from path concatenation with the root path from (q, κ̂) 7−→ (q0 , φ̂ κ̂) from
M
inv.
The balanced path for s00 s000 comes from a similar push edge from ∆E and
concatenation with the path from the invariant on H.
ZU064-05-FPR
paper-jfp
6 February 2018
54
3:38
J.I. Johnson, I. Sergey, C. Earl, M. Might, and D. Van Horn
Case s00 s000 ∈ ∆H2 .
φ̂−
By definition of addPop, ∆E2 = 0/ and there are q s000 ∈ ∆E, q0 q ∈ H such that
φ̂+
s00 q0 ∈ E. Let ~g be the witness of the invariant on q0 q. Let κ̂ be arbitrary. We
φ̂+~gφ̂−
know by the invariant on E, (s00 , κ̂) 7−→∗ (s000 , κ̂) and [φ̂+~gφ̂− ] = ε.
M
g
s0
, s00
s000
Case s ∈ ∆E3 ∪ ∆E4
∈ ∆H3 ∪ ∆H4 .
Follows from definition of inv and path concatenation, following similar reasoning
as above cases.
We define “π is a subtrace of π 0 ,” π v π 0
~g ∗
π v (s, κ̂) 7−→ (s0 , κ̂ 0 ) (s0 , g0 , s00 ) ∈ δ
hi ∗
~g ∗
(s0 , κ̂ 0 ) 7−→ (s0 , κ̂ 0 ) v (s, κ̂) 7−→ (s0 , κ̂ 0 )
M
M
M
~g ∗
g0
M
M
π v (s, κ̂) 7−→ (s0 , κ̂ 0 ) 7−→ (s00 , [κ̂+0 g0 ])
∗
g~00
~g ∗
(s, κ̂) 7−→ (s0 , κ̂ 0 ) v (s000 , κ̂ 00 ) 7−→ (s0 , κ̂ 0 ) (s0 , g0 , s00 ) ∈ δ
M
M
∗
g0
g~00
0
0
00
0
0
(s , κ̂ ) 7−→ (s , [κ̂+ g ]) v (s000 , κ̂ 00 ) 7−→
M
M
~g ∗
(s, κ̂) 7−→
M
g0
(s0 , κ̂ 0 ) 7−→ (s00 , [κ̂+0 g0 ])
M
Theorem 8.1 is a corollary of the following theorem.
Theorem 18.1
lfp(F 0 (M)) = (C (M), E C G (M), (0,
/ 0),
/ 0)
/
Proof
(⊆): Directly from 18.2.
~g ∗
(⊇): Let π ≡ (s0 , hi) 7−→ (s, κ̂) be an arbitrary path in C (M) (the inclusion of root is not
M
a restriction due to the definition of CRPDSs). Let n ∈ Nats be such that lfp(F 0 (M)) =
F 0 (M)n . We show
• the same path through G,
g
g
• for each s ∈ S, s s0 ∈ E, s s0 ∈ H, there is an m < n such that s ∈ ∆Sm s s0 ∈
∆Em s s0 ∈ ∆Hm respectively, where F 0 (M)m = (Gm , Hm , (∆Sm , ∆Em , ∆Hm )), and
g~ε ∗
• all non-empty balanced subtraces have edges in H: ∀(sb , κ̂) 7−→ (sa , κ̂) v π.~
gε 6=
M
hi ∧ [~
gε ] = ε =⇒ sb sa ∈ H.
By induction on π,
Case Base: s0 .
Follows by definition of F 0 . No non-empty balanced subtrace.
~g0 ∗
g00
M
M
Case Induction step: (s0 , hi) 7−→ (s0 , κ̂) 7−→ (s, [κ̂+ g00 ]).
~g0 ∗
By IH, (s0 , hi) 7−→
G
(s0 , κ̂).
By cases on g00 :
ZU064-05-FPR
paper-jfp
6 February 2018
3:38
PUSHDOWN FLOW ANALYSIS WITH ABSTRACT GARBAGE COLLECTION
55
Case γ+ .
g00
Let m be the witness for s0 by the IH. By definition of F 0 , (s0 , κ̂) 7−→ (s, [κ̂+ g00 ])
M
is in ∆Em+1 and Em+2 (and thus s ∈ ∆Sm+1 and Sm+2 ). Thus the path is constructible through Gn . All balanced subtraces carry over from IH, since the last
push edge cannot end a balanced path.
Case ε.
The path is constructible the same as γ+ . Let m be the witness used in the path
g~ε ∗
construction. Let π 0 ≡ (sb , κ̂) 7−→ (se , κ̂) be an arbitrary non-empty balanced
M
subtrace. If se 6= s, then the IH handles it. Otherwise, g~ε = g~0ε ε. If sb = s0 , then
the ε-edge is added by sprout (so the witness number is m + 1). If not, then there
∗
g~0
ε
is a balanced subtrace (sb , κ̂) 7−→
(s0 , κ̂), thus sb s0 ∈ H. Let m0 be the witness
M
for sb s0 ∈ ∆Hm0 . Then sb s ∈ ∆max{m,m0 }+1 by definition of addEmpty.
Case γ− .
γ+
Since [~g] is defined, there is a push edge in the trace (call it su 7−→ sv ) with a
M
(possibly empty) balanced subtrace following to s0 . Thus by IH, there are some
γ+
m, m0 such that su sv ∈ Em , (if the subtrace is non-empty) sv s0 ∈ Hm0 If
γ−
m ≥ m0 by definition of addPush, s0 s ∈ ∆Em+1 . Otherwise, the edge is in Em0
γ−
and by definition of addEmpty, s0 s ∈ ∆Em0 +1 .
g~ε ∗
Let π 0 ≡ (sb , κ̂) 7−→ (se , κ̂) be an arbitrary non-empty balanced subtrace. If se 6=
M
∗
g~0
γ+
M
M
ε
(su , κ̂) 7−→
s, the IH handles it. Otherwise, g~ε = g~0ε γ+ g~00ε γ− and π 0 ≡ (sb , κ̂) 7−→
∗
g~0ε
γ−
M
M
(sv , γ κ̂) 7−→ (s0 , γ κ̂) 7−→ (s, κ̂). su s is added to ∆Hmax{m,m0 }+1 and thus sb
su is in Hmax{m,m0 }+3 .
18.2 RIPDS reachability
We use metafunction •++• : Cont × Cont → Cont to aid proofs:
ε++κ̂ = κ̂
φ : κ̂++κ̂ 0 = φ : (κ̂++κ̂ 0 )
split(ε) = [¢, ε]
split(φ : κ̂) = [¢κ̂, φ ]
Lemma 18.3 (Down spin)
For (q, ε, ε, ↓, ε, q) ∈ δ , (q, [κ̂B ++κ̂B0 , κ̂T ], w) 7−→∗ (q, [κ̂B , κ̂B0 ++κ̂T ], w)
Proof
By induction on κ̂B0 .
ZU064-05-FPR
paper-jfp
6 February 2018
56
3:38
J.I. Johnson, I. Sergey, C. Earl, M. Might, and D. Van Horn
Case Base: κ̂B = ε.
Reflexivity.
Case Induction step: κ̂B0 = κ̂φ .
By δ , (q, [κ̂B ++κ̂B0 , κ̂T ], w) 7−→ (q, [κ̂B ++κ̂, φ κ̂T ], w). By IH, (q, [κ̂B ++κ̂, φ κ̂T ], w)
7−→∗ (q, [κ̂B , κ̂++φ κ̂T ], w). This final configuration is the same as (q, [κ̂B , κ̂B0 ++κ̂T ], w).
Lemma 18.4 (gadget correctness)
For (δ , S) = gadget(s, K̂, g, s0 ), (s, split(κ̂), w) 7−→∗ (s0 , split([κ̂+ g]), w) iff κ̂ ∈ K̂ and [κ̂+ g]
δ
defined.
Proof
(⇒): By inversion on the rules for δ , the path must go through three stages: the downspin, the middle path, and the pop-off. By 18.3, (s, split(κ̂), w) 7−→ (qdown , [¢κ̂, $], w) 7−→∗
(qdown , [ε, ¢κ̂$], w). Then the (qdown , ¢, ε, ↑, ε, q0 ) rule must apply. We can construct an accepting path in the machine recognizing K̂ from the middle path via the following lemma:
κ̂ 0
κ̂ 0
δ
N
(q0 , [¢, κ̂$], w) 7−→∗ (q, [¢κ̂ 0 , κ̂ 00 $], w) implies (q0 , κ̂) 7−→∗ (q, κ̂ 00 ). Proof by induction.
Then (q, $, ε, ·, $− , qout ) must apply, and then (qout , ε, ε, g, s0 ) must apply, meaning that
[κ̂+ g] is defined.
(⇐): Since K̂ is regular, there must be a path in the chosen NFA N = (Q, Σ, δN , q0 , F)
from q0 to a final state q ∈ F, (q0 , κ̂) 7−→∗ (q, ε).
N
In the first stage, (s, split(κ̂), w) 7−→∗ (q0 , [¢, κ̂$], w).
The follows first by the (s, ε, ε, ·, $+ , qdown ) transition, then by 18.3 (qdown , split(κ̂$), w)
7−→∗ (qdown , [ε, ¢κ̂$], w), finally by the (qdown , ¢, ε, ↑, ε, q0 ) rule.
In the second stage we construct a path (q0 , [¢, κ̂$], w) 7−→∗ (q, [¢κ̂, $], w), from an accepting path in N: (q0 , κ̂) 7−→∗ (q, ε) where q ∈ F. The statement we can induct on to get this
κ̂ 0
κ̂ 0
is (q0 , κ̂) 7−→∗ (q, κ̂ 00 ) implies (q0 , [¢, κ̂$], w) 7−→∗ (q, [¢κ̂ 0 , κ̂ 00 $], w).
N
δ
κ̂ 0
Case Base: = ε, q = q0
Reflexivity.
, κ̂ 00
= κ̂.
κ̂ 000
φε
Case Induction step: κ̂ 0 = κ̂ 000 φε , (q0 , κ̂) 7−→∗ (q0 , κ̂ 0000 ) 7−→ (q, κ̂ 00 ).
N
7 →∗ (q0 , [¢κ̂ 000 , κ̂ 0000 $], w).
−
N
By IH, (q0 , [¢, κ̂$], w)
If φε = ε, then κ̂ 000 = κ̂ 0 , κ̂ 0000 = κ̂ 00
0
and we apply the (q , ε, ε, ·, ε, q) rule to get to (q, [¢κ̂ 0 , κ̂ 00 $], w). Otherwise, κ̂ 0 = κ̂ 000 φ
and we apply the (q0 , φ , ε, ↑, ε, q) rule to get to (q, [¢κ̂ 0 , κ̂ 00 $], w).
In the third and final stage, (qout , [¢κ̂, $], w) 7−→ (qout , split(κ̂), w) and since [κ̂+ g] is defined, we reach the final state by (qout , ε, ε, ·, g, s0 ).
Lemma 18.5 (Checking lemma)
If (q, a, a, ↑, ε, q) ∈ δ and (q, [κ̂B , κ̂T 0 ++κ̂T $], w) 7−→∗ (q, [κ̂B ++κ̂T 0 , κ̂T $], w0 ) (through
the one rule) then w = κ̂T 0 w0 .
ZU064-05-FPR
paper-jfp
6 February 2018
3:38
PUSHDOWN FLOW ANALYSIS WITH ABSTRACT GARBAGE COLLECTION
57
Proof
Simple induction.
Lemma 18.6 (Stack machine correctness)
For all M ∈ CPDS, G ∈ CCPDS, q ∈ G, if G v C C (M) then
−→
K̂,g
L (Stacks(G)(q)) = κ̂ : (q0 , hi) 7−→∗ (q, κ̂) .
G
Proof
(⊆): Let (sstart , [ε, ε], κ̂) 7−→∗ (sfinal , split(κ̂), ε) be an accepting path for κ̂ ∈ L (Stacks(G)(q)).
We inductively construct a corresponding path in G that realizes κ̂. We first see that the
given path is split into three phases: setup, gadgetry, checking. The first step must be
(sstart , ε, ε, ·, ¢+ , s0 ), which we call setup. The only final state must be preceded by scheck ,
sdown , and the final occurrence of s, which we call checking. Thus the middle phase is a
trace from s0 to s. This must be through gadgets, which are disjoint for each rule of the
IPDS, and thus each edge in G.
(sstart , [ε, ε], κ̂) 7−→ (s0 , [ε, ¢], κ̂) 7−→∗ (s, split(κ̂), κ̂) 7−→
(sdown , [¢κ̂, $], κ̂) 7−→∗ (sdown , [ε, ¢κ̂$], κ̂) 7−→
(scheck , [¢, κ̂$], κ̂) 7−→∗ (scheck , [¢κ̂, $], ε) 7−→ (sfinal , split(κ̂), ε)
We induct on the path through gadgets, s0 to s in the above path, invoking 18.4 at each step.
(⊇): Simple induction between setup and teardown, applying 18.4.
Proof of Theorem 12.1
Proof
The finiteness of the state space and monotonicity of F ensures the least fixed point exists.
lfp(F (M)) ⊆ C C (M) follows from 18.6 and the definition of F .
−→ ∗
K̂,g
To prove C C (M) ⊆ lfp(F (M)), suppose not. Then there must be a path (s0 , hi) 7−→
M
K̂ 0 ,g0
(s, κ̂) 7−→ (s0 , [κ̂+ g0 ])
M
where the final edge is the first edge not in lfp(F (M)).
By definition of C C , κ̂ ∈ K̂ 0 and (s, K̂ 0 , g, s0 ) ∈ δ . Since κ̂ is realizable at s in G, by
K̂ 0 ,g
definition of F and 18.6, (s, κ̂) 7−→ (s0 , [κ̂+ g]) contra the assumption. Thus C C (M) ⊆
G
lfp(F (M)) holds by contradiction.
ZU064-05-FPR
paper-jfp
58
6 February 2018
3:38
J.I. Johnson, I. Sergey, C. Earl, M. Might, and D. Van Horn
mon \
\Γ−
\ Γ = ICRPDS×
We first prove an invariant of fˆ : Exp → System
−→ SystemΓ , where System
\ × OPState)
\
P(OPState
IΓ (e) = ((e, ⊥, ⊥), hi)
IΓ0 (e) = (((e, ⊥, ⊥), 0),
/ hi)
hφ̂1 , . . . , φ̂n i+ = φ̂1+ . . . φ̂n+
\ Γ → Prop
invΓ : Exp → System
G
z }| {
o
[ n
g
Ω̂, Ω̂0 : Ω̂ Ω̂0 ∈ Ê )
invΓ (e)((P̂, Ê), Ĥ) = (P̂ =
g
∧ ∀(ψ̂, A) (ψ̂ 0 , A0 ) ∈ Ê.let K̂ = {κ̂ : StackRoot(κ̂) = A} in
∀κ̂ ∈ κ̂ ∈ K̂ : [κ̂+ g] defined .StackRoot([κ̂+ g]) = A0
K̂,g
∧ (ψ̂, κ̂) 7−→ (ψ̂ 0 , [κ̂+ g])
M
−→ ∗
−−→
K̂,g
∧ ∀(ψ̂, ) (ψ̂ , ) ∈ Ĥ.∃K̂, g.[~g] = ε ∧ (ψ̂, hi) 7−→ (ψ̂ 0 , hi)
0
M
0
where M = I \
PDS (e)
Lemma 18.7 ( fˆ invariant)
For all e, if invΓ (e)(G) then invΓ (e)( fˆe (G))
Proof
Same structure as in Lemma 18.2 without reasoning about worklists.
Proof of Theorem 12.2
Proof
0
Let M = I \
PDS (e), G = C C (M) and G0 = ((P̂, Ê), Ĥ) = lfp( fˆe ).
(G0 approximates G):
−→
K̂,g
We strengthen the statement to π ≡ IΓ (e) 7−→∗ (ψ̂, κ̂) implies
G
~g
• IΓ0 (e) 7−→0 ∗ ((ψ̂, StackRoot(κ̂)), κ̂).
G
−→ ∗
K̂,g
K̂ 0 ,g0
• for all (ψ̂, κ̂) 7−→ (ψ̂ 0 , [κ̂+~g]) 7−→ (ψ̂ 00 , [κ̂+~gg0 ]) v π, if [~gg0 ] = ε, then ∃κ̂ ∈ K̂ and
G
G
(ψ̂, StackRoot(κ̂)) (ψ̂ 00 , StackRoot([κ̂+~gg0 ])) ∈ Ĥ
By induction on π,
Case Base: IΓ (e).
By definition of fˆe , IΓ0 (e) = (Ω̂0 , hi), Ω̂0 ∈ P̂. First goal holds by definition of
StackRoot(hi) and reflexivity. Second goal vacuously true.
Case Induction step:
−−→
K̂ 0 ,g0
((e, ⊥, ⊥), hi) 7−→∗
G
K̂ 00 ,g00
(ψ̂ 0 , [~g0 ]) 7−→ (ψ̂, κ̂).
G
~g0
Let A = StackRoot([~g0 ]). By IH, IΓ0 (e) 7−→0 ∗ ((ψ̂ 0 , A), [~g0 ]).
G
ZU064-05-FPR
paper-jfp
6 February 2018
3:38
PUSHDOWN FLOW ANALYSIS WITH ABSTRACT GARBAGE COLLECTION
59
0
Let K̂root = {κ̂ : StackRoot(κ̂) = A} By definition of I \
PDS and the case assumpK̂ 00 ,g00
0
00
0
0
~
tion, (ψ̂ , K̂root , g , ψ̂) ∈ δ . By cases on (ψ̂ , [g ]) 7−→ (ψ̂, κ̂):
G
K̂ 00 ,φ̂+
Case (ψ̂ 0 , [~g0 ]) 7−→ (ψ̂, κ̂).
G
φ̂
+
By definition of fˆ, (ψ̂ 0 , A) (ψ̂, A ∪ T (φ̂+ )) ∈ G0 . By definition of StackRoot
and A, StackRoot([~g0 g00 ]) = StackRoot([~g]) = StackRoot(κ̂).
K̂ 00 ,ε
Case (ψ̂ 0 , [~g0 ]) 7−→ (ψ̂, κ̂).
G
Similar to previous case.
K̂ 00 ,φ̂−
Case (ψ̂ 0 , [~g0 ]) 7−→ (ψ̂, κ̂).
G
Since is [~g0 φ̂− ] is defined, there is an i such that gi = φ̂+ , which is witness
K̂ 000 ,φ̂+
to an edge in the trace with that action, ψ̂b 7−
−→
→ ψ̂e By definition of [ ], the
G
actions from ψ̂e to ψ̂ 0 cancel to ε, meaning by IH (ψ̂e , A) (ψ̂ 0 , A) ∈ H, and
φ̂
+
(ψ̂b , A0 ) (ψ̂e , A) ∈ E. Thus the pop edge is added by definition of fˆ0 . The new
balanced path (ψ̂b , A0 ) (ψ̂, A0 ) is also added, and extended paths get added
with propagation.
Approximation follows by composition with Theorem 12.1.
(G approximates G0 ): Directly from 18.7.
The approximate GC has a similar invariant, except the sets of addresses are with respect
to the tˆ computation.
\ Γ → Prop
invΓ̂ : Exp → System
G
z }| {
[
A,g
0
0
invΓ̂ (e)((P̂, Ê), Ĥ) = (P̂ =
ψ̂, ψ̂ : ψ̂ ψ̂ ∈ Ê )
A,g
g
∧ ∀ψ̂0 ψ̂1 ∈ Ê.∃(ψ̂0Γ , AΓ ) (ψ̂1Γ , A0Γ ) ∈ Ê.(∀i.ψ̂iΓ v ψ̂i ) ∧ AΓ ⊆ A
∧ ∀ψ̂0 ψ̂1 ∈ Ĥ.∃(ψ̂0Γ , AΓ ) (ψ̂1Γ , AΓ ) ∈ Ĥ 0 .∀i.ψ̂iΓ v ψ̂i
∧ ∀ψ̂ ∈ P̂.∃(ψ̂ Γ , A) ∈ P̂0 .ψ̂ Γ v ψ̂ ∧ lfp(tˆ)(ψ̂) ⊆ A
where ((P̂0 , Ê 0 ), Ĥ 0 ) = lfp( fˆe )
Lemma 18.8 (Approx GC invariant)
For all e, if invΓ̂ (e)(G) then invΓ̂ (e)( fˆe0 (G))
Proof
Straightforward case analysis.
Proof of Theorem 12.3
Proof
Induct on path in lfp( fˆe ) and apply 18.8.
ZU064-05-FPR
paper-jfp
60
6 February 2018
3:38
J.I. Johnson, I. Sergey, C. Earl, M. Might, and D. Van Horn
19 Haskell implementation of CRPDSs
Where it is critical to understanding the details of the analysis, we have transliterated the
formalism into Haskell. We make use of a two extensions in GHC:
-XTypeOperators -XTypeSynonymInstances
All code is in the context of the following header, and we’ll assume the standard instances
of type classes like Ord and Eq.
import Prelude hiding ((!!))
import Data.Map as Map hiding (map,foldr)
import Data.Set as Set hiding (map,foldr)
import Data.List as List hiding ((!!))
type P s = Set.Set s
type k :-> v = Map k v
(==>) :: a -> b -> (a,b)
(==>) x y = (x,y)
(//) :: Ord a => (a :-> b) -> [(a,b)] -> (a :-> b)
(//) f [(x,y)] = Map.insert x y f
set x = Set.singleton x
19.1 Transliteration of NFA formalism
We represent an NFA as a set of labeled forward edges, the inverse of those edges (for
convenience), a start state and an end state:
type NFA state char =
(NFAEdges state char,NFAEdges state char,state,state)
type NFAEdges state char = state :-> P(Maybe char,state)
19.2 ANF
data Exp
=
|
|
data AExp
=
|
data Lambda =
data Call
=
type Var
=
Ret AExp
App Call
Let1 Var Call Exp
Ref Var
Lam Lambda
Var :=> Exp
AExp :@ AExp
String
ZU064-05-FPR
paper-jfp
6 February 2018
3:38
PUSHDOWN FLOW ANALYSIS WITH ABSTRACT GARBAGE COLLECTION
-- Abstract
type AConf
type AEnv
type AStore
type AD
data AVal
type AKont
type AFrame
61
state-space:
= (Exp,AEnv,AStore,AKont)
= Var :-> AAddr
= AAddr :-> AD
= (AVal)
= AClo (Lambda, AEnv)
= [AFrame]
= (Var,Exp,AEnv)
data AAddr = ABind Var AContext
type AContext = [Call]
Abstract configuration space transliterated into Haskell. In the code, we defined abstract
addresses to be able to support k-CFA-style polyvariance.
Atomic expression evaluation implementation:
aeval :: (AExp,AEnv,AStore) -> AD
aeval (Ref v, ρ, σ ) = σ !!(ρ!v)
aeval (Lam l, ρ, σ ) = set $ AClo (l, ρ)
We encode the transition relation it as a function that returns lists of states:
astep :: AConf -> [AConf]
astep (App (f :@ ae), ρ, σ , κ) = [(e, ρ’’, σ ’, κ) |
AClo(v :=> e, ρ’) <- Set.toList $ aeval(f, ρ, σ ),
let a = aalloc(v, App (f :@ ae)),
let ρ’’ = ρ’ // [v ==> a],
let σ ’ = σ [a ==> aeval(ae, ρ, σ )] ]
astep (Let1 v call e, ρ, σ , κ) =
[(App call, ρ, σ , (v, e, ρ) : κ)]
astep (Ret ae, ρ, σ , (v, e, ρ’) : κ) = [(e, ρ’’, σ ’, κ)]
where a = aalloc(v, Ret ae)
ρ’’ = ρ’ // [v ==> a]
σ ’ = σ [a ==> aeval(ae, ρ, σ )]
19.3 Partial orders
We define a typeclass for lattices:
class Lattice a where
bot :: a
top :: a
(v) :: a -> a -> Bool
(t) :: a -> a -> a
(u) :: a -> a -> a
ZU064-05-FPR
paper-jfp
62
6 February 2018
3:38
J.I. Johnson, I. Sergey, C. Earl, M. Might, and D. Van Horn
And, we can lift instances to sets and maps:
instance (Ord s, Eq s) => Lattice (P s) where
bot = Set.empty
top = error "no representation of universal set"
x t y = x ‘Set.union‘ y
x u y = x ‘Set.intersection‘ y
x v y = x ‘Set.isSubsetOf‘ y
instance (Ord k, Lattice v) => Lattice (k :-> v) where
bot = Map.empty
top = error "no representation of top map"
f v g = Map.isSubmapOfBy (v) f g
f t g = Map.unionWih (t) f g
f u g = Map.intersectionWith (u) f g
(t) :: (Ord k, Lattice v) => (k :-> v) -> [(k,v)] -> (k :-> v)
f t [(k,v)] = Map.insertWith (t) k v f
(!!) :: (Ord k, Lattice v) => (k :-> v) -> k -> v
f !! k = Map.findWithDefault bot k f
19.4 Reachability
We can turn any data type to a stack-action alphabet:
data StackAct frame = Push { frame :: frame }
| Pop { frame :: frame }
| Unch
type CRPDS control frame = (Edges control frame, control)
type Edges control frame = control :-> (StackAct frame,control)
We split the encoding of δ into two functions for efficiency purposes:
type Delta control frame =
(TopDelta control frame, NopDelta control frame)
type TopDelta control frame =
control -> frame -> [(control,StackAct frame)]
type NopDelta control frame =
control -> [(control,StackAct frame)]
If we only want to know push and no-change transitions, we can find these with a NopDelta
function without providing the frame that is currently on top of the stack. If we want pop
transitions as well, we can find these with a TopDelta function, but of course, it must have
ZU064-05-FPR
paper-jfp
6 February 2018
3:38
PUSHDOWN FLOW ANALYSIS WITH ABSTRACT GARBAGE COLLECTION
63
access to the top of the stack. In practice, a TopDelta function would suffice, but there
are situations where only push and no-change transitions are needed, and having access to
NopDelta avoids extra computation.
At this point, we must clarify how to embed the abstract transition relation into a pushdown transition relation:
adelta :: TopDelta AControl AFrame
adelta (e, ρ, σ ) γ = [ ((e’, ρ’, σ ’), g) |
(e’, ρ’, σ ’, κ) <- astep (e, ρ, σ , [γ]),
let g = case κ of
[]
-> Pop γ
[γ1 , ] -> Push γ1
[ ]
-> Unch ]
adelta’ :: NopDelta AControl AFrame
adelta’ (e, ρ, σ ) = [ ((e’, ρ’, σ ’), g) |
(e’, ρ’, σ ’, κ) <- astep (e, ρ, σ , []),
let g = case κ of
[γ1] -> Push γ1
[ ]
-> Unch ]
The function crpds will invoke the fixed point solver:
crpds :: (Ord control, Ord frame) =>
(Delta control frame) ->
control ->
frame ->
CRPDS control frame
crpds (δ ,δ ’) q0 0 =
(summarize (δ ,δ ’) etg1 ecg1 [] dE dH, q0) where
etg1 = (Map.empty // [q0 ==> Set.empty],
Map.empty // [q0 ==> Set.empty])
ecg1 = (Map.empty // [q0 ==> set q0],
Map.empty // [q0 ==> set q0])
(dE,dH) = sprout (δ ,δ ’) q0
Figure 11 provides the code for summarize, which conducts the fixed point calculation,
the executable equivalent of Figure 6:
ZU064-05-FPR
paper-jfp
64
6 February 2018
3:38
J.I. Johnson, I. Sergey, C. Earl, M. Might, and D. Van Horn
summarize :: (Ord control, Ord frame) =>
(Delta control frame) ->
(ETG control frame) ->
(ECG control) ->
[control] ->
[Edge control frame] ->
[EpsEdge control] ->
(Edges control frame)
To expose the structure of the computation, we’ve added a few types:
-- A set of edges, encoded as a map:
type Edges control frame =
control :-> P (StackAct frame,control)
-- Epsilon edges:
type EpsEdge control = (control,control)
-- Explicit transition graph:
type ETG control frame =
(Edges control frame, Edges control frame)
-- Epsilon closure graph:
type ECG control =
(control :-> P(control), control :-> P(control))
An explicit transition graph is an explicit encoding of the reachable subset of the transition relation. The function summarize takes six parameters:
1. the pushdown transition function;
2. the current explicit transition graph;
3. the current ε-closure graph;
4. a work-list of states to add;
5. a work-list of explicit transition edges to add; and
6. a work-list of ε-closure transition edges to add.
The function summarize processes ε-closure edges first, then explicit transition edges and
then individual states. It must process ε-closure edges first to ensure that the ε-closure
graph is closed when considering the implications of other edges.
Sprouting
ZU064-05-FPR
paper-jfp
6 February 2018
3:38
PUSHDOWN FLOW ANALYSIS WITH ABSTRACT GARBAGE COLLECTION
summarize (δ ,δ ’) (fw,bw) (fe,be) [] [] [] = fw
summarize (δ ,δ ’) (fw,bw) (fe,be) (q:dS) [] []
| fe ‘contains‘ q = summarize (δ ,δ ’) (fw,bw) (fe,be) dS [] []
summarize (δ ,δ ’) (fw,bw) (fe,be) (q:dS) [] [] =
summarize (δ ,δ ’) (fw’,bw’) (fe’,be’) dS dE’ dH’ where
(dE’,dH’) = sprout (δ ,δ ’) q
fw’ = fw t [q ==> Set.empty]
bw’ = bw t [q ==> Set.empty]
fe’ = fe t [q ==> set q]
be’ = be t [q ==> set q]
summarize (δ ,δ ’) (fw,bw) (fe,be) dS ((q,g,q’):dE) []
| (q,g,q’) ‘isin’‘ fw = summarize (δ ,δ ’) (fw,bw) (fe,be) dS dE []
summarize (δ ,δ ’) (fw,bw) (fe,be) dS ((q,Push ,q’):dE) [] =
summarize (δ ,δ ’) (fw’,bw’) (fe’,be’) dS’ dE’’ dH’ where
(dE’,dH’) = addPush (fw,bw) (fe,be) (δ ,δ ’) (q,Push ,q’)
dE’’ = dE’ ++ dE’’
dS’ = q’:dS
fw’ = fw t [q ==> set (Push ,q’)]
bw’ = bw t [q’ ==> set (Push ,q) ]
fe’ = fe t [q ==> set q ]
be’ = fe t [q’ ==> set q’]
summarize (δ ,δ ’) (fw,bw) (fe,be) dS ((q,Pop ,q’):dE) [] =
summarize (δ ,δ ’) (fw’,bw’) (fe’,be’) dS’ dE’’ dH’ where
(dE’,dH’) = addPop (fw,bw) (fe,be) (δ ,δ ’) (q,Pop ,q’)
dE’’ = dE ++ dE’
dS’ = q’:dS
fw’ = fw t [q ==> set (Pop ,q’)]
bw’ = bw t [q’ ==> set (Pop ,q) ]
fe’ = fe t [q ==> set q ]
be’ = fe t [q’ ==> set q’]
summarize (δ ,δ ’) (fw,bw) (fe,be) dS ((q,Unch,q’):dE) [] =
summarize (δ ,δ ’) (fw’,bw’) (fe’,be’) dS’ dE [(q,q’)] where
dS’ = q’:dS
fw’ = fw t [q ==> set (Unch,q’)]
bw’ = bw t [q’ ==> set (Unch,q) ]
fe’ = fe t [q ==> set q ]
be’ = fe t [q’ ==> set q’]
summarize (δ ,δ ’) (fw,bw) (fe,be) dS dE ((q,q’):dH)
| (q,q’) ‘isin‘ fe = summarize (δ ,δ ’) (fw,bw) (fe,be) dS dE dH
summarize (δ ,δ ’) (fw,bw) (fe,be) dS dE ((q,q’):dH) =
summarize (δ ,δ ’) (fw,bw) (fe’,be’) dS dE’ dH’ where
(dE’,dH’) = addEmpty (fw,bw) (fe,be) (δ ,δ ’) (q,q’)
fe’ = fe t [q ==> set q ]
be’ = fe t [q’ ==> set q’]
Fig. 11: An implementation of pushdown control-state reachability.
65
ZU064-05-FPR
paper-jfp
66
6 February 2018
3:38
J.I. Johnson, I. Sergey, C. Earl, M. Might, and D. Van Horn
sprout :: (Ord control) =>
Delta control frame ->
control ->
([Edge control frame], [EpsEdge control])
sprout (δ ,δ ’) q = (dE, dH) where
edges = δ ’ q
dE = [ (q,g,q’) | (q’,g) <- edges, isPush g ]
dH = [ (q,q’)
| (q’,g) <- edges, isUnch g ]
Pushing
addPush :: (Ord control) =>
ETG control frame ->
ECG control ->
Delta control frame ->
Edge control frame ->
([Edge control frame], [EpsEdge control])
addPush (fw,bw) (fe,be) (δ ,δ ’) (s,Push γ,q) = (dE,dH) where
qset’ = Set.toList $ fe!q
dE = [ (q’,g,q’’) | q’ <- qset’, (q’’,g) <- δ q’ γ, isPop g ]
dH = [ (s,q’’)
| (q’,Pop ,q’’) <- dE ]
Popping
addPop :: (Ord control) =>
ETG control frame ->
ECG control ->
Delta control frame ->
Edge control frame ->
([Edge control frame], [EpsEdge control])
addPop (fw,bw) (fe,be) (δ ,δ ’) (s’’,Pop γ,q) = (dE,dH) where
sset’ = Set.toList $ be!s’’
dH = [ (s,q) | s’ <- sset’,
(g,s) <- Set.toList $ bw!s’, isPush g ]
dE = []
Clearly, we could eliminate the new edges parameter dE for the function addPop, but we
have retained it for stylistic symmetry.
Adding empty edges The function addEmpty has many cases to consider:
ZU064-05-FPR
paper-jfp
6 February 2018
3:38
PUSHDOWN FLOW ANALYSIS WITH ABSTRACT GARBAGE COLLECTION
67
addEmpty :: (Ord control) =>
ETG control frame ->
ECG control ->
Delta control frame ->
EpsEdge control ->
([Edge control frame], [EpsEdge control])
addEmpty (fw,bw) (fe,be) (δ ,δ ’) (s’’,s’’’) = (dE,dH) where
sset’
= Set.toList $ be!s’’
sset’’’’ = Set.toList $ fe!s’’’
dH’
= [ (s’,s’’’’) | s’ <- sset’, s’’’’ <- sset’’’’ ]
dH’’ = [ (s’,s’’’)
| s’ <- sset’ ]
dH’’’ = [ (s’’,s’’’’) | s’’’’ <- sset’’’’ ]
sEdges = [ (g,s) | s’ <- sset’, (g,s) <- Set.toList $ bw!s’ ]
dE = [ (s’’’’,g’,q) | s’’’’ <- sset’’’’,
(g,s) <- sEdges,
isPush g, let Push γ = g,
(q,g’) <- δ s’’’’ γ,
isPop g’ ]
dH’’’’ = [ (s,q) | ( ,s) <- sEdges, ( , ,q) <- dE ]
dH = dH’ ++ dH’’ ++ dH’’’ ++ dH’’’’
ZU064-05-FPR
paper-jfp
6 February 2018
3:38
| 6cs.PL |
Hybrid Beamforming with Spatial Modulation
in Multi-user Massive MIMO mmWave
Networks
Merve Yüzgeçcioglu and Eduard Jorswieck
arXiv:1709.04826v1 [cs.IT] 14 Sep 2017
Communications Theory, Communications Laboratory
Dresden University of Technology, D-01062 Dresden, Germany
Email: {merve.yuzgeccioglu, eduard.jorswieck}@tu-dresden.de
Abstract—The cost of radio frequency (RF) chains is
the biggest drawback of massive MIMO millimeter wave
networks. By employing spatial modulation (SM), it is
possible to implement lower number of RF chains than
transmit antennas but still achieve high spectral efficiency.
In this work, we propose a system model of the SM scheme
together with hybrid beamforming at the transmitter
and digital combining at the receiver. In the proposed
model, spatially-modulated bits are mapped onto indices
of antenna arrays. It is shown that the proposed model
achieves approximately 5dB gain over classical multi-user
SM scheme with only 8 transmit antennas at each antenna
array. This gain can be improved further by increasing
the number of transmit antennas at each array without
increasing the number of RF chains.
I. I NTRODUCTION
Recent technological developments also bring a challenge to wireless communications since the number
of network devices increases drastically. Even though
there are several works on achieving high data rates,
the bandwidth shortage of today’s commercial networks
also limits the capacity of the channel. In order to
overcome this bottleneck, millimeter wave (mmWave)
spectrum became tempting due to the availability of large
bandwidths. However, the signal in mmWave spectrum
experiences a severe propagation loss and the resulting
channel is poorly scattered. In spite of the unfavorable
characteristics of the channel, beamforming techniques
can be employed at massive multiple-input multipleoutput (mMIMO) networks to direct the beam with high
array gain. Thanks to the small wave length at high
frequencies, it is possible to pack a large amount of
antennas in small areas and asymptotically achieve the
capacity of the channel [1]. However, packing high number of antennas comes with a price: energy consumption.
It is not practical to implement dedicated RF chains
for each antenna at a transceiver due to the power and
space restrictions. Nevertheless, by employing hybrid
beamforming techniques, it is possible to benefit from
The work of Merve Yüzgeçcioglu has received partly funding
from the European Union’s Horizon 2020 research and innovation
programme under the Marie Sklodowzka-Curie grant agreement No
641985
978-1-5386-3531-5/17/$31.00 c 2017 European Union
the gain of a large amount of antennas while consuming
less energy by using lower number of RF chains than
antennas.
There are several works on optimal precoding in hybrid beamforming [2]–[8]. In [4], a method to construct a
near-optimal hybrid beamforming structure is presented
instead of an exhaustive search to maximize the spectral
efficiency. The work is extended to multi-user scenario in
[5] where users employ only analog combiner while the
base station (BS) has a hybrid structure and constructs
baseband precoder in order to mitigate the inter-userinterference. In [6], together with a two-stage hybrid
beamformer/combiner algorithm, the minimum number
of RF chains to realize the performance with a fullyconnected structure is given. Another approach is presented in [7] that exploits the singular vectors of the
channel to generate analog beamformer and combiner
which provides a lower complexity. Furthermore, an
asymptotic rate expression is given for the proposed
scheme.
Another popular research topic recently is SM that
enables transmitting additional bits to conventional modulated symbol without requiring extra power [9]–[11].
There have been extensive work on SM in different
aspects for MIMO networks. Furthermore, implementation of SM at multi-user MIMO networks is studied
at [12] and a precoding method is proposed in order
to cancel inter-user-interference. In this work, SM is
employed at sub-groups which are dedicated for each
user. Combination of SM and beamforming is introduced in [13]. In this work, analog beamforming is
employed to a generalized SM (GSM) scheme for singleuser scenario at Rician fading channels. A closer look
into the transmitter design of SM together with analog
beamforming is given in [14]. Furthermore, the extension
of SM with hybrid beamforming for a mmWave railway
communication system is studied in [15]. Here, the
receive antenna arrays (AA) at the front and end of
the train are assumed as virtual users served by two
different data streams transmitted from the BS. Analog
beamformer and combiner are employed at the transmitter and receiver, respectively. However, the beamformer
and combiner are not designed specifically for multi-
user spatial modulation system. Rather the design in [3]
is considered. As performance metric, an upper bound
rate for AWGN channel with Gaussian input is considered and maximum-likelihood detection is employed
for decoding. However, it is not clearly stated how the
inter-user-interference is eliminated and data decoding is
performed at each user.
In this paper, implementation of SM to a multi-user
mmWave network is introduced. In this scheme, BS
employs analog beamformer to direct the beam to the
intended user and digital precoder to cancel inter-userinterference while users employ the digital combiners.
The novelty of the paper is summarized as follows
• Analog beamforming is designed in order to maximize the achievable rate of the link. Two different codebooks are considered to choose analog
beamformer vector such that one provides optimal
beamformer while the other allows practical implementation. The quantization error is characterized
between these two codebooks.
• Digital precoder at the transmitter is designed in
order to eliminate inter-user-interference. Digital
combiner at the receiver is designed to successfully
reconstruct transmitted symbol when it is matched
with the correct analog beamformer.
• A low-complexity maximum-likelihood detector is
introduced.
• Analytical expression of the achievable rate of the
system is derived and the tightness of the expression
is shown.
The rest of the paper is organized as follows. In Sec. II,
the system model of the proposed scheme is described.
In Sec. III, design of the analog beamformer and digital
precoder for transmitter side and the digital combiner
for receiver side is explained in detail, and the quantization error between two codebooks is characterized.
In Sec. IV, analytical expression of achievable rate for
the proposed scheme is derived. The performance is
analyzed numerically in Sec. V and finally the paper
is concluded in Sec. VI.
II. S YSTEM M ODEL
The block diagram of the system is shown in Fig. 1.
There, the BS is equipped with NA uniform linear arrays
(ULA), each with NT transmit antennas connected to
NRF RF chains. We assume without loss of generality
that NT is fixed for all AAs and there are no common
transmit antennas between arrays. The BS serves K users
with NR receive antennas and it is assumed that NRF ≥
K. The number of AAs may vary independently from
the number of RF chains at the transmitter.
Data is transmitted to mobile stations (MS) in two
parts: first part is modulated according to spatial modulation principles by choosing one of the AAs which allows
to transmit log2 NA bits and the second part is modulated
according to a conventional M -ary modulation scheme
which allows to transmit log2 M bits. The index of
the active AA conveys information without consuming
additional power. According to incoming bits for each
bit stream, it is possible that one AA is used for more
than one user at a channel use. This is only possible
by the help of analog beamformer and allows to exploit SM for multi-user scenario without any limitation.
After determining the AA indices for every user, the
transmitter constructs the digital precoder and the analog
beamformer. The received signal at i-th user can be
expressed as follows
ri =
K
√ X
ρ
Haj ,i faj ,j paj ,j s + ni .
(1)
j=1
Herein, Haj ,i ∈ CNR ×NT is the L-path channel between
the aj -th AA and i-th user where aj is the selected AA
index to transmit data to j-th user. The channel follows
a geometry-based model shown below [16]
r
L
NT NR X
Haj ,i =
αaj ,il aRaj ,i (θl )aH
Taj ,i (φl ), (2)
L
l=1
q
Here, aRaj ,i (θl ) = N1R[1,ejπ sin θl,. . .,ejπ(NR −1) sin θl]T
q
1
jπ sin φl
,. . .,ejπ(NT −1) sin φl]T
and aTaj ,i (φl ) =
NT [1,e
are the receive and transmit AA responses of the lth path where l = 1, . . . , L. φl and θl are the angle
of departure (AoD) and angle of arrival (AoA) of the
path and drawn from the uniform distribution U(0, 2π].
Finally, αaj ,il , CN (0, 1), represents the channel gain
and the path loss of the path. Throughout the paper,
(.)T and (.)H denote transpose and hermitian operators,
respectively. It is assumed that the antennas are placed on
the array with half-wavelength separation. paj ,j ∈ C1×K
and faj ,j ∈ CNT ×1 are the digital precoder and analog
beamformer vectors for j-th user when aj -th AA array
is chosen for transmission. ρ represents the total transmit
power and with equal power allocation, the transmit
power of i-th user is ρi = ρ/K. The selection procedure
for precoder and beamformer is explained in detail in
Sec. III. s ∈ CK×1 is the data vector contains symbols
for K users and finally ni ∈ CNR ×1 is the noise vector
with elements distributed as CN (0, σ 2 ).
A. Receiver Structure
The receiver structure of the users is shown in Fig. 1.
In this system model, MS has NR receive antennas
connected to the same number of RF chains. The received signal is passed on to the combiner with the
help of RF chains and digital combining is performed
to successfully reconstruct the transmitted symbol. In
order to decode the received symbol, a low complexity maximum-likelihood (ML) detector is employed to
jointly estimate the AA index and M -ary modulated
symbol
2
[âi , ŝi ] = arg min
a,m
1 H
w ri − sm ,
β a,i
(3)
where wa,i ∈ CNR ×1 is the digital combiner from a-th
AA to i-th user, a = 1, . . . , A, β is the normalization
RF chain
Digital
Combiner
RF chain
W1
NR
log2 NA
log2 M
.
.
.
K
.
.
.
log2 NA
RF chain
Digital
Precoder
P
log2 M
.
.
.
NRF
.
.
.
.
.
.
NA
.
.
.
Analog
Beamformer
F
NA
Decoder
â1 , ŝ1
NA
Decoder
âK , ŝK
.
.
.
RF chain
RF chain
Digital
Combiner
RF chain
WK
NR
Fig. 1: Block diagram of hybrid beamforming with spatial modulation in multi-user downlink transmission
coefficient and sm is the m-th symbol from M -ary
constellation diagram. This detector structure allows the
user to decode without having Haj ,i , faj ,j and paj ,j
knowledge but only wa,i vectors and β. The construction
of wa,i is also explained in detail in Sec. III.
III. P RECODER , B EAMFORMER AND C OMBINER
D ESIGN
We consider two different approaches to design the
analog beamformer vector. The first method is choosing the vector among the transmit AA response vectors aT (φ). The second method is using a predefined
beamsteering codebook with quantized angles between
(0, 2π].
Let us define the codebook F = {fn ∈ CNT ×1 :
H
fn fn = 1, n = 1, . . . , N } and name the two different
codebooks as FA for transmit AA response vectors
collection and FB for the beamsteering codebook. For
both cases,
q the individual codewords are constructed
1
T
as f =
NT [1, exp (jπφ), . . . , exp (jπ(NT − 1)φ)]
where φ is the AoD of each path for codebook FA
and quantized angle 2πn
N where n = 0, . . . , N − 1 for
codebook FB .
Furthermore, the analog beamformer f is chosen in
order to maximize the signal-to-noise ratio (SNR) of the
link between the selected AA and the intended user.
For each user, the BS computes the RF beamformer
f1,i , . . . , fNA ,i as follows
fa,i = arg max ||Ha,i fn ||2 ,
∀fn ∈F
(4)
where fa,i is the optimum RF beamformer from the ath AA to i-th user. The resulting analog beamforming
matrix for all users is F = [fa1 ,1 , . . . , faK ,K ].
After determining the optimum RF beamformers for
each AA-user pair, the BS constructs the matrix Hi =
[H1,i f1,i , . . . , HNA ,i fNA ,i ]. Hi is used to calculate the
digital combiner for each user to successfully reconstruct
the transmitted symbol
Wi = w1,i . . . wNA ,i = (H†i )H ,
(5)
where (.)† denotes the pseudo inverse of a matrix. With
this design, it is possible to feedback only Hi matrices
to the users instead of having a training period of the
channel matrices for each AA-user pair. In this case, it is
sufficient for users to have Hi with dimension NA ×NR
instead of full channel matrices H1,i , . . . , HNA ,i with
dimension NA NT × NR .
After determining the optimum RF beamformer and
receive combiner, the transmitter generates the effective
channel in Eq. [6] order to calculate the digital precoder
to eliminate the inter-user-interference
H
wa1 ,1 Ha1 ,1 fa1 ,1 . . . waH1 ,1 HaK ,1 faK ,K
..
..
..
Hef f =
.
.
.
.
waHK ,K Ha1 ,K fa1 ,1 . . . waHK ,K HaK ,K faK ,K
(6)
Finally the precoding vector for each user is determined
as P = [pTa1 ,1 . . . pTaK ,K ]T = βH†ef f ,where pak ,k ∈
C1×K is the precoder vector of k-th user when ak -th AA
is chosen by spatially-modulated bits. The coefficient β
satisfies the average power constraint which is calculated
as
(s
)
K
β=E
,
(7)
tr(H†ef f (H†ef f )H )
where the expectation is over channel realizations. Since
the complex channel coefficients are the key parameters
for spatially-modulated bits, instantaneous power constraint is not applicable for such systems [12].
A. Characterization of the Quantization Error
For large scale AAs, the right and left singular vectors
of the channel are the transmit and receive AA response
vectors, respectively [17]. Let us define the optimal
analog beamformer for the channel given in Eq. [2] in
accordance with this statement as follows
fA = arg
max
∀fA,n ∈FA
||HfA,n ||2 ,
(8)
where N = L is the number of paths of the channel
and the total number of elements in the codebook.
H ∈ CNR ×NT is the channel matrix, fA,n ∈ CNT ×1
is the n-th element of the codebook FA . In order to
achieve the performance of the optimal beamformer,
infinite resolution phase shifters should be implemented
at the system. Since this is not possible for practical
systems, beamsteering codebooks with quantized angle
are commonly considered. However, selecting the analog
beamformer among beamsteering codebooks leads to a
quantization error.
Information theoretical methods can be used to find the
achievable rate of the proposed scheme. The mutual
information between transmitted and received symbols
is written as
UB
Exact
LB
2
bits/s/Hz
1.5
I(yi ; xi ) = h(yi ) − h(zi ).
1
Therein, xi is considered as the symbol that contains
the AA index and M -ary modulated symbol to ease
the understanding. zi is the noise term after receive
combining which follows the same distribution with ni .
0.5
0
-10
-5
0
5
10
15
SNR [dB]
Fig. 2: Rate approximation of a user for NA = 2, NT =
8, NR = 1 with BPSK modulation
Proposition 1. Assume that fA and fB are the solution
to Eq. [4] for codebooks FA and FB , respectively. Then,
the quantization error can be characterized as follows
dc,min ≤
max
∀fA ∈G(NT ,1)
d2c (fA , fB ) ≤ dc,max ,
(9)
H
}, dc,max =
where dc,min = 1 − tr{fA fAH fB,n fB,n
−
1
(13)
−
B
c NT −1 2 NT −1 and c is the coefficient of the metric
ball volume defined in [18].
Proof: Let us assume the codebook FB with B bits
resolution phase shifters where codebook size is N =
2B . Eq. [4] can be reformulated as follows
fB , arg
= arg
min
d2c (fA , fB,n )
min
H
1 − tr{fA fAH fB,n fB,n
},
∀fB,n ∈FB
∀fB,n ∈FB
(10)
where dc (., .) is the chordal distance. Eq. [10] is an
example of Grassmannian quantization on the Grassmann manifold G(NT , 1). The codebook FB can be
considered as a Grassmannian subspace sphere-packing
codebook since the elements of FB are distributed over
H
a sphere with radius ||fB,n
fB,n ||2 = 1, ∀n = 1, . . . , N .
Let us define the distance between the optimal beamformer and the selected beamformer from the quantized
codebook as dc (fA , fB ). By using the results in [18]–[20]
we reach the Eq. [9].
It is clearly seen from Eq. [9] that the
quantization error vanishes asymptotically such as
− 1
− B
limB→∞ c NT −1 2 NT −1 = 0.
IV. ACHIEVABLE R ATE
The received signal at i-th user after receive combining is
K
√ X
yi = waHi ,i ρ
Haj ,i faj ,j paj ,j s + waHi ,i ni .
(11)
j=1
Since the interference from the other users is eliminated
by employing transmit precoder, the resulting signal at
the receiver can be reformulated as follows
√
yi = ρi waHi ,i Hai ,i fai ,j si + waHi ,i ni .
(12)
Proposition 2. The received signal yi follows Gaussian
mixture distribution. There is no closed-form solution
for the entropy of a Gaussian mixture but using a tight
approximation the achievable rate can be bounded as
hl (yi ) − h(zi ) ≤ Ri ≤ hu (yi ) − h(zi ),
(14)
where hl (yi ) = (γ + αN,N 0 ) log2 e + log2 σ, hu (yi ) =
(γ + βN,N 0 ) log2 e + log2 σ and h(zi ) = log2 (πeσ 2 ).
Proof: We begin the proof by deriving the probability density function of the received signal that can be
written as follows
fYi(yi) =
NA
M X
X
fYi(yi |a = j,s = sm)P(a = j,s = sm). (15)
m=1 j=1
Since the selection of the AA indices and M -ary symbols
are independent events, joint probability in Eq. [15] can
be separated as P (a = j, s = sm ) = P (a = j)P (s =
1
sm ) = N1A M
. Furthermore, it is clearly seen that
fYi (yi |a = j, s = sm ) follows complex Gaussian distri√
H
bution with different mean values ρi wj,i
Hj,i fj,i sm according to the j and m indices. The resulting probability
density function follows Gaussian mixture distribution
fYi (yi ) =
M NA
1 XX
M NA m=1 j=1
(
) (16)
√
H
|yi − ρi wj,i
Hj,i fj,i sm |2
1
exp −
,
π
σ2
where Hj,i , fj,i and wj,i are the channel matrix, analog
beamformer and receive combiner for j, i AA-user pair,
respectively. Herein, sm denotes the m-th symbol from
M -ary constellation diagram and ρi is the transmit power
for i-th user. By using the tight approximation derived
for Gaussian mixtures in [21], entropy of yi can be
bounded as (γ + αN,N 0 ) log2 e + log2 σ ≤ h(yi ) ≤
(γ + βN,N 0 ) log2 e + log2 σ. Due to space constraints
the detailed explanation on the parameters γ, αN,N 0 and
βN,N 0 are not given here and can be found in [21].
The tightness of the introduced approximation is
shown in Fig. 2. In this figure, the lower (LB) and upper
(UB) bounds are calculated by using the approximation
and the exact rate is found Rby calculating the exact
∞
integral for entropy h(y) = − −∞ fY (y) log fY (y)dy.
100
100
FB - B = 6
FB - B = 7
10-1
FB - B = 8
10-1
FB - B = 9
BER
BER
10-2
-2
10
10-3
10-3
SM - NT = 4
HBFSM - NA
HBFSM - NA
HBFSM - NA
HBFSM - NA
10-4
10-4
10-5
0
5
10
15
20
25
30
SNR [dB]
0
5
= 4,
= 4,
= 4,
= 4,
10
NT
NT
NT
NT
= 8, FB
= 8, FA
= 64, FB
= 64, FA
15
20
25
30
SNR [dB]
Fig. 3: BER performance for different codebook sizes
(K = 2, NA = 4, NT = 8, NR = 1, 4 − QAM )
Fig. 4: Comparison of classical SM and HBF-SM (K =
2, NR = 1, B = 8, 4 − QAM )
V. N UMERICAL R ESULTS
floor when codebook FB is used, the error performance
is better than the classical SM scheme up to high SNR
regime. When the optimal codebook is used, HBF-SM
outperforms classical SM regardless the SNR value. It
is possible to have 5dB gain over classical SM scheme
with only 8 transmit antennas at an AA. If we consider
64 transmit antennas at each array, the gain will increase
up to 12dB. Although there is an intersection point
between HBFM-SM and classical SM, by increasing the
codebook size and transmit antenna numbers at an array,
the performance of HBF-SM can be further increased.
Note that, each AA is served by a single RF chain.
Hence, increasing the number of transmit antennas at an
array does not come with a high cost while significantly
increasing the performance. As it is seen from the
simulation results, the number of antennas at an array is
one of the key parameters that affects the performance.
The larger the number of transmit antennas the better the
beam can be directed to intended direction.
In this section, the performance of the hybrid beamforming with spatial modulation (HBF-SM) is analyzed
numerically. Firstly, the effect of the quantization error
on the uncoded bit error rate (BER) performance is
investigated for different codebook sizes. Then, it is
compared with the conventional SM scheme for multiuser scenario [12].
In Fig. 3, the BER performance of the system is
shown for the cases that the analog beamformer vector
is chosen among the transmit AA response vectors
(FA ) and the beamsteering codebook (FB ). It is seen
that the quantization error of the codebook FB leads
to an error floor for the low codebook sizes. Since
the digital combiner at the receiver side is generated
by using the analog beamforming vector also, when
the beamforming vector is erroneous, it cumulatively
increases the resulting error probability. As it is expected,
when the analog beamformer vector is chosen from
codebook FA , the performance is better and there is
no error floor in the system. Although the codebook
FA provides the optimal beamformer, there is no phase
shifter with infinite resolution. Hence, the codebook
FB with B-bit resolution would be used in practical
systems. When the phase shifters are changed from 6bits to 9-bits the performance is increased significantly.
As it is shown in Sec. III-A the performance of the
beamsteering codebook will eventually converge to the
optimal beamformer.
In Fig. 4, the BER performance of the HBF-SM is
compared with the conventional SM scheme in multiuser scenario. Average power constraint is employed for
both schemes. The considered SM scheme has NT = 4
transmit antennas. For HBF-SM scheme, NA = 4 antenna arrays are considered with NT = 8 and NT = 64
transmit antennas at each array, respectively. In SM
scheme, total log2 NT = 2 bits are transmitted at a
channel use by the selected transmit antenna index.
In HBFM-SM scheme, also log2 NA = 2 bits are
transmitted by the selected AA index. It is seen from
the simulation results, although HBF-SM has an error
VI. C ONCLUSION
In this work, we have introduced a system model
of HBF-SM scheme. Novel design of the beamformer
and digital combiners for the proposed model have been
derived. The effect of the codebook selection for analog
beamformer on performance has been shown and the
quantization error has been studied. Additionally, a tight
approximation for the rate has been introduced. Uncoded
BER performance of the system has been shown for
different parameters and also compared with a classical
SM implementation for multi-user networks. Simulation
results showed that even with low number of transmit
antennas at the AAs, performance of classical SM is
outperformed. For future work, the receiver design can
be improved by employing hybrid beamforming that also
allows reducing the number of RF chains. Since the
most important objective of the SM schemes is providing
higher spectral efficiency while consuming lower energy
compared to conventional schemes, optimizing the energy efficiency of the system is an important direction
for future research.
R EFERENCES
[1] T. L. Marzetta, “Noncooperative cellular wireless with unlimited
numbers of base station antennas,” IEEE Transactions on Wireless Communications, vol. 9, no. 11, pp. 3590–3600, Nov. 2010.
[2] Z. Pi, “Optimal transmitter beamforming with per-antenna power
constraints,” in IEEE Signal Processing for Communications
Symposium, Jun. 2012, pp. 3779–3784.
[3] T. Kim, J. Park, J.-Y. Seol, S. Jeong, J. Cho, and W. Roh, “Tens
of Gbps support with mmwave beamforming systems for next
generation communications,” in IEEE Global Communications
Conference, Dec. 2013, pp. 3685–3690.
[4] O. E. Ayach, S. Rajagopal, S. Abu-Surra, Z. Pi, and R. W. Heath
Jr., “Spatially sparse precoding in millimeter wave MIMO systems,” IEEE Transactions on Wireless Communications, vol. 13,
no. 3, pp. 1499–1513, Mar. 2014.
[5] A. Alkhateeb, G. Leus, and R. W. Heath Jr., “Limited feedback
hybrid precoding for multi-user millimeter wave systems,” IEEE
Transactions on Wireless Communications, vol. 14, no. 11, pp.
6481–6494, Nov. 2015.
[6] F. Sohrabi and W. Yu, “Hybrid digital and analog beamforming
design for large-scale MIMO systems,” in IEEE International
Conference on Acoustics, Speech and Signal Processing, Apr.
2015, pp. 2929–2933.
[7] S. Payami, M. Ghoraishi, and M. Dianati, “Hybrid beamforming
for large antenna arrays with phase shifter selection,” IEEE
Transactions on Wireless Communications, vol. 15, no. 11, pp.
7258–7271, Nov. 2016.
[8] T. E. Bogale, L. B. Le, A. Haghighat, and L. Vandendorpe, “On
the number of RF chains and phase shifters, and scheduling design with hybrid analog-digital beamforming,” IEEE Transactions
on Wireless Communications, vol. 15, no. 5, pp. 3311–3326, May
2015.
[9] R. Y. Mesleh, H. Haas, S. Sinanovic, C. Wook, and S. Yun,
“Spatial modulation,” IEEE Transactions on Vehicular Technology, vol. 57, no. 4, pp. 2228–2241, Jul. 2008.
[10] M. D. Renzo, H. Haas, A. Ghrayeb, S. Sugiura, and L. Hanzo,
“Spatial modulation for generalized MIMO: Challenges, opportunities and implementation,” IEEE in Proceedings, vol. 102, no. 1,
pp. 56–103, Jan. 2014.
[11] M.-C. Lee, W.-H. Chung, and T.-S. Lee, “Generalized precoder
design formulation and iterative algorithm for spatial modulation
in MIMO systems with CSIT,” IEEE Transactions on Communications, vol. 63, no. 4, pp. 1230–1244, Apr. 2015.
[12] S. Narayanan, M. J. Chaudhry, A. Stavridis, M. D. Renzo,
F. Graziosi, and H. Haas, “Multi-user spatial modulation MIMO,”
in IEEE Wireless Communications and Networking Conference,
Apr. 2014, pp. 671–676.
[13] N. Ishikawa, R. Rajashekar, S. Sugiura, and L. Hanzo,
“Generalized-spatial-modulation-based
reduced-RF-chain
millimeter-wave communications,” IEEE Transactions on
Vehicular Technology, vol. 66, no. 1, pp. 879–883, Jan. 2017.
[14] M.-C. Lee and W.-H. Chun, “Transmitter design for analog
beamforming aided spatial modulation in millimeter wave MIMO
systems,” in IEEE International Symposium on Personal, Indoor
and Mobile Radio Communications: Fundamentals and PHY,
Sep. 2016, pp. 1–6.
[15] Y. Cui, X. Fang, and L. Yan, “Hybrid spatial modulation beamforming for mmwave railway communication systems,” IEEE
Transactions on Vehicular Technology, vol. 65, no. 12, pp. 9597–
9606, Dec. 2016.
[16] A. A. M. Saleh and R. A. Valenzuela, “A statistical model for
indoor multipath propogation,” IEEE Journal on Selected Areas
in Communications, vol. SAC-5, no. 2, pp. 128–137, Feb. 1987.
[17] O. E. Ayach, R. W. Heath Jr., S. Abu-Surra, S. Rajagopal, and
Z. Pi, “The capacity optimality of beamsteering in large millimeter wave MIMO systems,” in IEEE International Workshop on
Signal Processing Advances in Wireless Communications, Jun.
2012, pp. 100 – 104.
[18] R. T. Krishnamachari and M. K. Varanasi, “On the geometry
and quantization of manifolds of positive semi-definite matrices,”
IEEE Transactions on Signal Processing, vol. 61, no. 18, pp.
4587–4599, Sep. 2013.
[19] W. Dai, Y. Liu, and B. Rider, “Quantization bounds on Grassmann manifolds and applications to MIMO communications,”
IEEE Transactions on Information Theory, vol. 54, no. 3, pp.
1108–1123, Mar. 2008.
[20] P. Cao, A. Zappone, and E. A. Jorswieck, “Grouping-based
interference alignment with IA-cell assignment in multi-cell
MIMO MAC under limited feedback,” IEEE Transactions on
Signal Processing, vol. 64, no. 5, pp. 1336–1351, Mar. 2016.
[21] K. Moshksar and A. K. Khandani, “Arbitrarily tight bounds on
differential entropy of Gaussian mixtures,” IEEE Transactions on
Information Theory, vol. 62, no. 6, pp. 3340–3354, Jun. 2016.
| 7cs.IT |
arXiv:1710.10069v1 [math.GR] 27 Oct 2017
THE MULTIPLICITIES OF ROOT NUMBER FUNCTIONS
STEFAN-CHRISTOPH VIRCHOW
Abstract. We consider the qth root number function for the symmetric
group. Our aim is to develop an asymptotic formula for the multiplicities
of the qth root number function as q tends to ∞. We use character theory,
number theory and combinatorics.
1. Introduction
Let q be a positive integer. We define the qth root number function rq : Sn → N0
via
rq (π) := #{σ ∈ Sn : σ q = π}.
Obviously, rq is a class function of the symmetric group Sn . For each irreducible
character χ of Sn let
mχ(q) := hrq , χi
be the multiplicity of χ in rq .
Scharf [7] proved that the qth root number functions rq are proper characters,
(q)
that is, the multiplicities mχ are non-negative integers. A good account of results
on root number functions can be found in [3, Chapter 6.2 and 6.3].
(q)
We now pay attention to the multiplicities mχ . Müller and Schlage-Puchta
(q)
established estimates for the multiplicities mχλ (cf. [5, Proposition 2 and 3]). In
addition, they showed the following: Let ∆ ∈ N be fixed and let q > 2 be an integer.
Given a partition µ ⊢ ∆, there exists some constant Cµq , depending only on µ and
q, such that for n sufficiently large and for a partition λ = (λ1 , . . . , λl ) ⊢ n with
(q)
λ\λ1 := (λ2 , . . . , λl ) = µ, we have mχλ = Cµq . In particular, we get
q
C(1)
= σ0 (q) − 1
1
(σ1 (q) + σ0 (q)2 − 3σ0 (q) + σ0′ (q))
2
1
= (σ1 (q) + σ0 (q)2 − 3σ0 (q) − σ0′ (q)) + 1,
2
q
C(2)
=
q
C(1,1)
where σ0′ (q) is the number of odd divisors of q.
Our aim is to generalize this result. We establish an asymptotic formula for the
(q)
multiplicities mχλ as q tends to ∞. More precisely, we claim the following:
log q
Theorem 1.1. Let q ∈ N be sufficiently large and let ∆ ∈ N with ∆ 6 log
2 . In
addition, let n > ∆q be an integer. Then, for partitions λ ⊢ n and µ ⊢ ∆ with
1
2
STEFAN-CHRISTOPH VIRCHOW
λ\λ1 = µ, we have
σ0 (q) + O(1)
1 σ1 (q) + O (σ0 (q))2
2
(q)
mχλ = χµ (1)
1 (q))
6 σ2 (q) + O(σ0 (q)σ
χµ (1) σ∆−1 (q) + O χµ (1) q ∆−2 ∆σ0 (q) + 2∆
∆!
∆!
if ∆ = 1
if ∆ = 2
if ∆ = 3
if ∆ > 4.
The O-constant is universal.
Remark 1.2. The error term in our asymptotic formula is essentially optimal.
The proof of our Theorem proceeds as follows: At first, we realize that for π ∈ Sn
the value χλ (π) is a polynomial in the number ci (π) of i-cycles of the permutation
π for i = 1, . . . , n. We use this result to establish a formula with main and error
(q)
term for mχλ , where the random variables ci appear again. Secondly, we summarize
identities and estimates for Stirling numbers of the first and second kind and then
we review bounds for the divisor function. Thirdly, we examine the distribution of
cycle in Sn and compute the mean of (ck1 )m1 · . . . · (ckj )mj . The formula for the
mean includes Stirling numbers of the second kind. Finally, we calculate the main
(q)
and the error term of mχλ obtained in the first step using the outcomes of step two
and three.
Some notation. A partition of n is a sequence λ = (λ1 , λ2 , . . . , λl ) of positive
integers such that λ1 > λ2 > . . . > λl and λ1 + λ2 + . . . + λl = n. We write
λ ⊢ n to indicate that λ is a partition of n. By λ\λ1 we mean the partition
Pl
λ\λ1 = (λ2 , λ3 , . . . , λl ). The weight |λ| of λ is |λ| = j=1 λj . For partitions λ, µ
we write µ ⊂ λ, if µj 6 λj for all j.
We denote by χλ the irreducible character of the symmetric group Sn corresponding
to the partition λ of n.
For a permutation π ∈ Sn and 1 6 i 6 n let ci (π) be the number of i-cycles of π.
Furthermore, for α ∈ R let
X
dα
σα (q) :=
d|q
be the divisor function. As usual, we denote by ζ(s) the Riemann zeta function
and we write (n,
k) forthe greatest common divisor of n and k.
We denote by nk and nk the Stirling numbers of the first and second kind, respectively. Finally, (x)n denotes the falling factorial.
2. Proof of the Theorem
In this section, we use character theory to derive a formula with main and error
(q)
term for the multiplicities mχλ . We apply this result to prove our Theorem.
Müller and Schlage-Puchta [5, Lemma 7] established the following.
Lemma 2.1. Let λ ⊢ n be a partition, µ = λ\λ1 , and let π ∈ Sn be a permutation.
Then
X
X
Y ci (π)
|µ̃|
χλ (π) =
(−1)
,
χµ,µ̃ (c)
ci
µ̃⊆µ
µ̃1 61
c⊆S|µ|−|µ̃|
i6|µ|
THE MULTIPLICITIES OF ROOT NUMBER FUNCTIONS
3
where c runs over all conjugacy classes of S|µ|−|µ̃| , χµ,µ̃ (c) denotes the number of
ways to obtain µ̃ from µ by removing rim hooks according to the cycle structure of
c, counted with the sign prescribed by the Murnaghan-Nakayama rule1, and ci is
the number of i-cycles of an element of c.
This result shows that χλ (π) is a polynomial in ci (π) for i = 1, . . . , |µ| with leading
term χµ (1)(|µ|!)−1 c1 (π)|µ| . We now observe:
Lemma 2.2. Let λ ⊢ n and µ ⊢ ∆ be partitions with µ = λ\λ1 , and let π ∈ Sn be
a permutation. Then we have
∆
X
c
(π)
+
.
.
.
+
c
(π)
c1 (π)
1
j+1
.
χλ (π) = χµ (1)
+ O χµ (1)
∆
−
j
∆
j=1
Proof. Applying Lemma 2.1, we obtain
X
c1 (π)
χλ (π) = χµ (1)
+
(−1)|µ̃|
∆
µ̃⊆µ
µ˜1 61
X
c⊆S∆−|µ̃|
Y ci (π)
χµ,µ̃ (c)
,
ci
i6∆
where c runs over all conjugacy classes of S∆−|µ̃| except the trivial class of S∆ .
Therefore, we realize the expected main term. We shall show that the second term
in the above formula can be absorbed into the error term.
At first, we observe that |χµ,µ̃ (c)| 6 χµ (1) for a conjugacy class c of S∆−|µ̃| . Secondly, let c be a conjugacy class of Sk with 1 6 k 6 ∆ and let ci be the number
of i-cycles of an element of c. Suppose that c1 + c2 + . . . + c∆ = ∆ − j for some
positive integer j. Then we have ci = 0 for all i > j + 2.
Therefore, it follows that the absolute value of the considered second term in the
preceding formula is bounded above by
∆
∆
Y ci (π)
X
X
X
c1 (π) + . . . + cj+1 (π)
.
6 χµ (1)
χµ (1)
∆−j
ci
j+1
j=1
j=1
i6j+1
(c1 ,...,cj+1 )∈N0
1c1 +...+(j+1)cj+1 6∆
c1 +...+cj+1 =∆−j
This yields our assertion.
Due to
mχ(q)
=
λ
1 X
χλ (π q ),
n!
π∈Sn
we obtain as immediate consequence of the preceding Lemma the following result.
Proposition 2.3. Let λ ⊢ n and µ ⊢ ∆ be partitions with µ = λ\λ1 , and let q ∈ N.
Then we get
∆ X
q
q
X c1 (π q )
X
χ
(1)
c
(π
)
+
.
.
.
+
c
(π
)
χ
(1)
µ
1
j+1
µ
.
mχ(q)
=
+O
λ
n!
n! j=1
∆
∆−j
π∈Sn
Now, we give the
1Cf. for instance [6, Theorem 4.10.2]
π∈Sn
4
STEFAN-CHRISTOPH VIRCHOW
(q)
Proof of Theorem 1.1. We stated a formula with main and error term for mχλ in
Proposition 2.3. In section 5, we will evaluate the main term: see Proposition 5.3.
In section 6, we will estimate the error term: cf. Lemma 6.2. Therefore, the proof of
our Theorem is completed. Moreover, the error term in our Theorem is essentially
optimal due to the Remark 5.4.
In the next two sections, we shall establish some auxiliary results.
3. Combinatorics and number theory
In this section we review some results about Stirling numbers of the first and
second kind as well as basic facts about the divisor function.
Definition 3.1.
Let n and k be positive integers. The Stirling numbers of the
second kind nk count the number of ways to partition a set of n labeled objects
into k nonempty unlabeled subsets.
Lemma 3.2. Let n and k be positive
n integers.
1) We have n2 = 2n−1 − 1 and n−1
= n2 .
2) In addition, we state the recurrence
k−1
X
n
k!
n
n
.
k!
=k −
(k
−
j)!
j
k
j=1
Proof. 1) follows from the definition. For the proof of 2) see [4, Theorem 7.2.6].
This recurrence yields an upper bound for nk . Next, we would like to represent
the ordinary powers xn by falling factorials (x)k := x(x − 1) · . . . · (x − k + 1).
Lemma 3.3. Let n be a positive integer. Then the identity
n
X
n
n
x =
(x)k
k
k=1
holds.
Proof. See for instance [1, Formula (6.10)].
Now, we pay attention to the Stirling numbers of the first kind.
Definition
3.4. Let n and k be positive integers. The Stirling numbers of
the
first
kind nk count the number of ways to arrange n objects into k cycles. So nk equals
the number of permutations of n elements with exactly k disjoint cycles.
Lemma 3.5. Let n and k be positive integers.
1) We have the recurrence
n
n−1
n−1
= (n − 1)
+
.
k
k
k−1
2) We obtain the estimate
(n − 1)!
n
n
6
.
k
k!
k−1
THE MULTIPLICITIES OF ROOT NUMBER FUNCTIONS
5
Proof. 1) Cf. [1, Formula (6.8)].
2) By induction over n: Obviously, the estimate is true for n 6 k. Applying the
recurrence 1) yields for n > k
k
n! n + 1
n
n
n!
n+1
+
6
.
6
k!
n k−2
k! k − 1
k−1
k
Stirling numbers of the first kind are (up to sign) the coefficients of ordinary powers
that yield the falling factorial (x)n . More precisely, we get
Lemma 3.6. Let n be a positive integer. Then the identity
n
X
n k
x
(x)n =
(−1)n−k
k
k=1
holds.
Proof. See for instance [1, Formula (6.13)].
Finally, we state results about the divisor function.
log q
Lemma 3.7. 1) Let ǫ > 0. Then we have σ0 (q) 6 (2 + ǫ) log log q for all q > q0 (ǫ).
2) σ1 (q) ≪ q log log q for all q > q0 .
3) Let k > 2. Then q k 6 σk (q) 6 ζ(2)q k .
Proof. For 1) and 2) see [2, Theorem 317 and Theorem 323].
3) The lower bound is obvious. The upper bound follows from the fact
X 1
.
σk (q) = q k
dk
d|q
4. Statistics of the symmetric group
Müller and Schlage-Puchta [5, Lemma 13] showed that, for π ∈ Sn chosen at
random, the distribution of ck (π) converges to a Poisson distribution with mean
1
→ ∞. In addition, they proved that the mean of (ck (·))m converges to
k as n
P
m m
−s
as n → ∞. We generalize this result and make it more explicit.
s=1 s k
Proposition 4.1. Let k1 , . . . , kl be distinct positive integers and let mj ∈ N for
j = 1, . . . , l. Then
!
mj
l
l
X
Y
mj
1 X Y
mj −s
.
kj
6
ckj (π)
s
n!
j=1 s=1
j=1
π∈Sn
If
Pl
j=1
kj mj 6 n is fulfilled, then we have an equality.
Proof. 1) Let k1 , . . . , kl be distinct positive integers and let sj ∈ N for j = 1, . . . , l.
We observe
Pl
0
l
if
X Y
ckj (π)
j=1 kj sj > n
−1
=
Ql
Pl
sj
n!
sj
if
j=1 kj sj !
j=1 kj sj 6 n.
j=1
π∈Sn
6
STEFAN-CHRISTOPH VIRCHOW
You can see this equality as follows.
Pl
Case 1: Let
j=1 kj sj > n. Then there exists no π ∈ Sn such that ckj (π) > sj
for all j = 1, . . . , l. Therefore, the considered sum is equal to 0.
Pl
Case 2: Let
j=1 kj sj 6 n. Then the left hand side of the equation is equal to
the number of tuples (τ1 , . . . , τl+1 ), which satisfy the following condition: There
exists distinct, disjoint cycles σij and a non-negative integer sl+1 such that τj =
σ1j · . . . · σsj j for all j = 1, . . . , l + 1, the cycles σij have length kj for all j = 1, . . . , l
Ql+1 Qsj
σij is the cycle decomposition for a permutation from Sn . Finally,
and j=1 i=1
the number of these tuples is equal to
n!
(n − k1 s1 )!
· ...
s2
− k1 s1 )! k2 s2 !(n − k1 s1 − k2 s2 )!
Pl−1
l
n − j=1 kj sj !
X
n−
kj sj !
...·
Pl
klsl sl ! n − j=1 kj sj !
j=1
k1s1 s1 !(n
Canceling yields our assertion.
2) It follows from 1) that
l
1 X Y
ckj (π) s =
j
n!
j=1
π∈Sn
(
0
Ql
−sj
j=1
kj
Pl
if
kj sj > n
Pj=1
l
if
j=1 kj sj 6 n.
3) Eventually, we compute the desired mean of a product of random variables ck (·).
Applying Lemma 3.3 and 2) yields
! l
!
l
l
Y mi
Y
X
Y
mj
(ckj )sj
E
=
ckj
E
si
i=1
j=1
j=1
(s1 ,...,sl )
16sj 6mj
6
X
(s1 ,...,sl )
16sj 6mj
=
l
Y
j=1
Obviously, we have an equality if
Pl
l
Y
mj
−s
kj j
sj
j=1
mj
X
mj −s
kj
s
s=1
j=1
!
.
kj mj 6 n is fulfilled.
Furthermore, Müller and Schlage-Puchta [5, Formula (33)] established the following
useful identity.
Lemma 4.2. Let d and q be positive integers and let π ∈ Sn . Then
X
cd (π q ) =
(k, q)ck (π).
k
k/(k,q)=d
5. The main term
We carry out the first step of the plan formulated in the proof at the end of
section 2: We compute the main term in Proposition 2.3. At first, we draw our
attention to the mean of a power of c1 (π q ).
THE MULTIPLICITIES OF ROOT NUMBER FUNCTIONS
Lemma 5.1. Let q ∈ N be sufficiently large and let δ ∈ N with
addition, let n > δq be an integer. Then we obtain
σ0 (q)
if
2
δ σ1 (q) + (σ0 (q))
if
1 X
c1 (π q ) =
n!
σ2 (q) + O(σ0 (q)σ1 (q))
if
π∈Sn
σδ−1 (q) + O q δ−2 δσ0 (q) + 2δ
if
7
δ 6
log q
log 2 .
In
δ=1
δ=2
δ=3
δ > 4.
The O-constant is universal.
Proof. 1) At first, we consider the case δ ∈ {1, 2}. We sketch the argument for
δ = 2 (the case δ = 1 is similar). Using Lemma 4.2 we get
2 X 1 X
X
1 X
1 X
2
k2
c1 (π q ) =
k1 k2
(ck (π)) +
ck1 (π)ck2 (π).
n!
n!
n!
π∈Sn
π∈Sn
k|q
(k1 ,k2 )
ki |q
k1 6=k2
π∈Sn
Since n > 2q, it follows with Proposition 4.1 that the considered mean is equal to
σ1 (q) + (σ0 (q))2 . This shows our claim for δ = 2.
2) We generalize this method for an arbitrary δ. Let n > δq. Applying Lemma 4.2
and Proposition 4.1 yields
δ
δ X
1 X
c1 (π q ) =
n!
π∈Sn
X
X
l=1 {M1 ,...,Ml } (k1 ,...,kl )
ki |q
ki 6=kj (i6=j)
j|
l |M
X
Y
|Mj | |Mj |−s
kj
,
s
j=1 s=1
where the second sum on the right is over all set partitions of {1, . . . δ} in exactly
l sets M1 , . . . , Ml .
3) We direct our attention to
Tm
m
m
XX
m m−s X m
σm−s (q).
k
=
:=
s
s
s=1
s=1
k|q
log q
Let q be sufficiently large and m 6 log
2 . Then
if m = 1
σ0 (q)
Tm 6 σ0 (q) + σ1 (q) if m = 2
m−1
3q
if m > 3.
In particular, we have T2 6 qσ0 (q).
You can see this estimate as follows: The cases m = 1 and m = 2 are obvious. Let
m > 3. It results from Lemma 3.2 and Lemma 3.7 for a constant C > 0
m−2
X m
2
m−1
Tm 6 Cm q log log q + ζ(2)q
q −s+1 .
s
s=1
8
STEFAN-CHRISTOPH VIRCHOW
Applying the estimate
m
6 sm (s!)−1 (see Lemma 3.2) we find that
s
m−2
X
s=1
m −s+1
q
6e−1
s
This proves our assertion.
4) Let q be sufficiently large and δ 6
m1 + . . . + ml = δ, we have
l
Y
j=1
log q
log 2 .
Then, for positive integers mi such that
l
Tmj 6 q δ−l max{3, σ0 (q)} .
This results immediately from 3).
5) Using the outcomes of step 2) and 4), we obtain
δ
δ
δ X X
X
δ
1 X
k δ−s + O
c1 (π q ) =
n!
s
s=1
π∈Sn
where
X
l
Y
l=2 {M1 ,...,Ml } j=1
k|q
= σδ−1 (q) + O F1 + F2 + F3 ,
T|Mj |
δ
X
δ
σδ−s (q),
s
s=2
X
T|M1 | T|M2 | ,
F2 :=
F1 :=
{M1 ,M2 }
F3 :=
δ
l
X
δ δ−l
max{3, σ0 (q)} .
q
l
l=3
The sum in F2 is over all set partitions of {1, . . . , δ} in exactly two sets M1 , M2 .
Therefore, we realize the expected main term. We shall show that the error term
is sufficiently small.
6) For the rest of the proof let 3 6 δ 6
F1 ≪
(
log q
log 2 .
σ1 (q)
2δ q δ−2
Lemma 3.7 and Lemma 3.2 yield
if δ = 3
if δ > 4
and
F3 ≪ 2δ q δ−2 ,
which is sufficiently small.
THE MULTIPLICITIES OF ROOT NUMBER FUNCTIONS
9
7) Finally, we examine F2 . The term F2 determines the order of the error term.
More precisely, we get
X
δ
δ
T|M1 | T|M2 |
F2 6
T1 Tδ−1 +
T2 Tδ−2 +
1
2
{M1 ,M2 }
|Mi |>3
if δ = 3
σ0 (q)σ1 (q)
≪ σ0 (q)q 2
if δ = 4
δ−2
δ
q
δσ0 (q) + 2
if δ > 5.
In the last estimate, we used the outcome of step 3). In addition, we applied
the inequality (σ1 (q))2 6 98 σ0 (q)q 2 for δ = 4 and the fact that δ 2 σ1 (q)q −1 6
δ 2 (log σ0 (q) + 1) ≪ max{2δ , δσ0 (q)} for the case δ > 5. Therefore, the proof is
completed.
Our next aim is to show that the result stated in Lemma 5.1 is essentially optimal.
log q
Lemma 5.2. Let q be prime and let δ 6 log
2 be a positive integer. In
n > δq be an integer. Then we get
σ0 (q)
δ
1 X
q
c1 (π ) = σ1 (q) + (σ0 (q))2
n!
π∈Sn
σ (q) + q δ−2 δ + 2δ−1 − 1 + O 3δ q δ−3
δ−1
addition, let
if δ = 1
if δ = 2
if δ > 3.
The O-constant is universal.
Proof. 1) For δ ∈ {1, 2} see Lemma 5.1. So let δ > 3. Since q is prime, Lemma 4.2
and Proposition 4.1 yield
δ
δ
1 X X δ
1 X
c1 (π q ) =
(c1 (π))δ−k (qcq (π))k
n!
n!
k
π∈Sn
π∈Sn k=0
= σδ−1 (q) + q δ−2 (δ + 2δ−1 − 1) + F1 + F2 + F3 ,
where
δ
X
δ δ−t
q ,
t
s
t=3
s=2
!
! k
δ−2 δ−k
X
X k
X δ − k
δ
k−t
,
F2 :=
q
t
s
k
t=1
s=1
k=1
X
δ−1
δ − 1 δ−1−t
δ
F3 :=
q
.
t
δ − 1 t=2
F1 :=
δ
X
δ
+
So we found the expected main term. We shall show that F1 , F2 and F3 can be
absorbed into the error term.
δ
q
δ
−1
2) Taking into account that 3 6 δ 6 log
, we obtain
log 2 and t 6 t (t!)
!
X
δ
δ
δ
δ 3−t
F1 ≪
+ q δ−3
+
q
≪ 3δ q δ−3 ,
2
3
t
t=4
10
STEFAN-CHRISTOPH VIRCHOW
and
!
δ−2 δ−k
δ−2
X
X
X δ − k
δ
δ
k−1
δ−3
F2 ≪
q
≪q
6 2δ q δ−3
s
k
k
s=1
k=1
as well as
k=1
F3 ≪ 2δ q δ−3 δ.
So we are done.
Proposition 5.3. Let q ∈ N be sufficiently large and let ∆ ∈ N with
addition, let n > ∆q be an integer. Then
σ0 (q)
1
2
q
1 X c1 (π )
2 σ1 (q) + O (σ0 (q))
= 1
σ2 (q) + O(σ0 (q)σ1 (q))
n!
∆
π∈Sn
6
1
1 ∆−2
∆σ0 (q) + 2∆
∆! σ∆−1 (q) + O ∆! q
∆6
log q
log 2 .
In
if ∆ = 1
if ∆ = 2
if ∆ = 3
if ∆ > 4.
The O-constant is universal.
Proof. 1) It follows from Lemma 3.6 that
1 X c1 (π q )
1 X
∆!
=
(c1 (π q ))∆ + O(F ),
n!
n!
∆
π∈Sn
π∈Sn
where
F :=
∆−1
X
δ=1
∆ 1 X
(c1 (π q ))δ .
δ n!
π∈Sn
The main term in the above formula is the mean we computed in Lemma 5.1.
Applying this Lemma, we get the expected main and error term. Therefore, it only
remains to be examined whether the error term F is sufficiently small.
2) Before we analyze the error term, we give two technical estimates.
q
Let ∆ 6 log
log 2 . Then Lemma 3.5 yields:
∆−3
i) For ∆ > 3 and δ ∈ {1, 2} we have ∆
.
δ ≪q
∆
ii) For ∆ > 4 and 1 6 δ 6 ∆ − 1 we get δ 6 ∆2 q ∆−δ−1 .
3) Now we estimate the error term F . For ∆ = 1 we obtain F = 0. For ∆ = 2
Lemma 5.1 yields, that F = σ0 (q). So let ∆ > 3. Applying Lemma 5.1, Lemma
3.7 and the upper bounds of step 2) we get
F ≪ q ∆−3 σ1 (q) +
∆−1
X
δ=3
which is sufficiently small.
∆2 q ∆−δ−1 σδ−1 (q) 6 q ∆−2
σ1 (q)
+ ∆3 ,
q
Remark 5.4. The error term in the preceding Proposition is essentially optimal:
Confer Lemma 5.2 and step 1) in the above proof.
THE MULTIPLICITIES OF ROOT NUMBER FUNCTIONS
11
6. The error term
We implement the second step of our plan: We compute the error term in Proposition 2.3.
Lemma 6.1. Let q ∈ N be sufficiently large and let 1 6 r 6 exp(q 1/3 ). In addition
log q
let δ ∈ N with δ 6 log
2 and n be a positive integer. Then
1 X
n!
π∈Sn
where Hr :=
Pr
σ0 (q)Hr
!δ
r
X
(σ0 (q))2 Hr2 + σ1 (q)Hr
cd (π q )
≪
q 2 Hr
d=1
δ−1 2
q
Hr
1
d=1 d .
if δ = 1
if δ = 2
if δ = 3
if δ > 4,
The O-constant is universal.
Proof. 1) Similar to step 2) in the proof of Lemma 5.1, we obtain
!δ
|Mj |
r
δ
l
X
X
Y X
X |Mj |
1 X X
cd (π q )
k −s
(k, q)|Mj |
6
,
n!
s
s=1
j=1
π∈Sn
d=1
l=1 {M1 ,...,Ml }
k
k6(k,q)r
where the second sum on the right is over all set partitions of {1, . . . δ} in exactly
l sets M1 , . . . , Ml .
2) We draw our attention to
X
(k, q)m
k
k6(k,q)r
r
m
m
X
m −s X X 1 m
k =
s
ds s
s=1
s=1
d=1
X
k
k=(k,q)d
m−s
k
d
r
m
X
1X m
σm−s (q)
6
d s=1 s
d=1
= Hr T m ,
where Tm is defined as in step 3) in the proof of Lemma 5.1.
3) The previous considerations in combination with Step 4) in the proof of Lemma
5.1 yield
!δ
r
1 X X
q
cd (π )
n!
π∈Sn
d=1
6 Hr Tδ + Hr2
X
{M1 ,M2 }
T|M1 | T|M2 | +
δ
X
l=3
Hrl
l
δ δ−l
q
max{3, σ0 (q)} .
l
Taking into account that Hr 6 log r + 1, our claim follows from step 3) and 7) in
the proof of Lemma 5.1 and from the estimates in Lemma 3.2 and 3.7.
12
STEFAN-CHRISTOPH VIRCHOW
q
Lemma 6.2. Let q ∈ N be sufficiently large and let ∆ ∈ N with ∆ 6 log
log 2 . In
addition, let n be a positive integer. Then
1
if ∆ = 1
∆ X
q
q
X
σ0 (q)
if ∆ = 2
1
c1 (π ) + . . . + cj+1 (π )
≪
n! j=1
∆−j
σ
(q)
if
∆=3
1
π∈Sn
1 q ∆−2 if ∆ > 4.
(∆−1)!
The O-constant is universal.
Proof. 1) The case ∆ = 1 is obvious. So let ∆ > 2. For 1 6 i 6 ∆ consider
i! X c1 (π q ) + . . . + c∆−i+1 (π q )
Q(i, ∆) : =
n!
i
π∈Sn
!δ
i
X
X ∆−i+1
X
i−δ i 1
q
=
(−1)
cd (π ) .
δ n!
δ=1
π∈Sn
The above equality is true due to Lemma 3.6. It
of Proposition 5.3 and with Lemma 6.1 that
σ0 (q)H∆
2
Q(i, ∆) ≪ σ1 (q)H∆−1
i−1 2
q H∆−i+1
d=1
follows with step 2) in the proof
if i = 1
if i = 2
if i > 3.
In the case i = 2 we also used the estimate (σ0 (q))2 6 2σ1 (q).
2) Finally, we look at
∆−1
∆
X Q(i, ∆)
1 X X c1 (π q ) + . . . + cj+1 (π q )
.
=1+
R∆ :=
n! j=1
i!
∆−j
i=1
π∈Sn
Due to the result of step 1), we find that
σ0 (q)
R∆ ≪ σ1 (q)
1 q ∆−2
(∆−1)!
So we are done.
if ∆ = 2
if ∆ = 3
if ∆ > 4.
References
[1] R. L. Graham, D. E. Knuth, O. Patashnik: Concrete Mathematics, Addison-Wesley, Reading, Massachusetts (1991).
[2] G. H. Hardy, E. M. Wright: An Introduction to the Theory of Numbers, Clarendon, Oxford
(1954).
[3] A. Kerber: Algebraic Combinatorics Via Finite Group Actions, BI-Wissenschaftsverlag,
Mannheim-Wien-Zürich (1991).
[4] V. H. Moll: Numbers and Functions, AMS, Providence, Rhode Island (2012).
[5] T. W. Müller, J.-C. Schlage-Puchta: Character theory of symmetric groups, subgroup
growth of Fuchsian groups, and random walks, Adv. Math. 213 (2007), 919-982.
[6] B. E. Sagan: The Symmetric Group, Springer, New York (2001).
[7] T. Scharf: Die Wurzelanzahlfunktion in symmetrischen Gruppen, J. Algebra 139 (1991),
446-457.
THE MULTIPLICITIES OF ROOT NUMBER FUNCTIONS
Author information
Stefan-Christoph Virchow, Institut für Mathematik, Universität Rostock
Ulmenstr. 69 Haus 3, 18057 Rostock, Germany
E-mail: stefan.virchow@uni-rostock.de
13
| 4math.GR |
Efficient, Safe, and Probably Approximately Complete Learning of Action Models
arXiv:1705.08961v1 [cs.AI] 24 May 2017
Roni Stern
Ben Gurion University of the Negev
Be’er Sheva, Israel
roni.stern@gmail.com
Abstract
In this paper we explore the theoretical boundaries
of planning in a setting where no model of the
agent’s actions is given. Instead of an action model,
a set of successfully executed plans are given and
the task is to generate a plan that is safe, i.e., guaranteed to achieve the goal without failing. To this
end, we show how to learn a conservative model
of the world in which actions are guaranteed to be
applicable. This conservative model is then given
to an off-the-shelf classical planner, resulting in a
plan that is guaranteed to achieve the goal. However, this reduction from a model-free planning to
a model-based planning is not complete: in some
cases a plan will not be found even when such exists. We analyze the relation between the number of
observed plans and the likelihood that our conservative approach will indeed fail to solve a solvable
problem. Our analysis show that the number of trajectories needed scales gracefully.
1
Introduction
In classical planning problems, a model of the acting agent
and its relationship to the relevant world is given in a formal planning description language, e.g., the classical STRIPS
model [Fikes and Nilsson, 1971] or PDDL [McDermott et al.,
1998]. Planning algorithms (planners) use this model to generate a plan for achieving a given goal condition. Creating a
planning domain model, however, is acknowledged as a notoriously hard knowledge engineering task. This has motivated
much work on learning such knowledge.
One such well-studied approach is to learn a domain
model by observing the agent’s interactions with the environment. The problems that arise in such approaches, however, are frequently intractable [Kearns and Valiant, 1994;
Daniely and Shalev-Shwartz, 2016]. An alternative approach
that is commonly used in reinforcement learning is to skip the
model-learning phase and directly learn how to act by observing the agent’s past actions and observations, and by guiding
the agent towards performing exploratory actions [Kearns et
al., 2002, inter alia]. In most prior work, the agent may fail to
execute a planned action. When this occurs, the agent replans,
possibly refining an underlying domain or action model it has
Brendan Juba
Washington University in St. Louis
St. Louis, MO, 63130 USA
bjuba@wustl.edu
learned. Thus, the agent learns from both positive and negative examples.
In this work we address a different setting, in which such
execution failures must be avoided. This setting is relevant
when execution failures are very costly or when the agent
has limited computation capabilities, and thus does not have
the capability to re-plan after the plan it has tried to execute
has failed. Consider, for example, a team of nano-bots deployed inside a human body for medical target identification
and drug delivery [Cavalcanti et al., 2007]. Re-planning is
likely not possible in such nano-bots, and, of course, failing
to cure the disease is undesirable. Thus, the planning task we
focus on in this paper is how to find a plan that is safe, i.e.,
is guaranteed to achieve the goal, in a setting where a domain
model is not available. We call this problem the safe modelfree planning problem.
Since performing actions that might fail is not allowed, exploration actions cannot be performed. The only source of
information available is a set of trajectories of previously executed plans. First, we show how to learn a set of actions that
the agent can use from the given trajectories. For every such
action a, we bound the set of predicates that are a’s preconditions and effects. This bounded action model is then used
to construct a classical planning problem such that a solution
to it is a solution to our model-free problem. This approach
to solve the model-free planning problem is sound and can
be very efficient, since current classical planners are very efficient. However, it is not complete, as the planning problem
that uses the learned action model might not be solvable even
if the underlying model-free planning problem is. Nonetheless, we prove that under some assumptions, the probability
of this occurring decreases quasi-linearly with the number of
observed trajectories.
This positive result comes in contrast to the hardness of
other tasks related to model learning. For example, learning
to predict labels computed by finite-state machines [Kearns
and Valiant, 1994] or even DNF formulas [Daniely and
Shalev-Shwartz, 2016] is believed to be computationally intractable. Thus, we cannot hope to predict the values that
fluents will take merely on the basis of the fact that these can
be computed by a finite-state machine or a DNF. Similarly,
even the problems of finding memoryless policies [Littman,
1994] or finite-state policies [Meuleau et al., 1999] in simple
environments is computationally intractable. Finally, in the
standard interactive learning model, simple examples (that
would be captured by a STRIPS environment, for example)
are known to force a learner to explore an exponential number of paths in any reasonable measure of the environment’s
complexity [Kakade, 2003, Section 8.6].
The key difference between the model learning we propose
and these hardness results is that we limit our attention to
the learning of STRIPS domain models in a PAC (“Probably
Approximately Correct”) sense: we do not aim to learn an
accurate action model, only one that is sufficient for finding a
safe plan in most cases. We introduce and desribe this PAClike setting in Section 4.
2
Problem Definition
The setting we address is a a STRIPS planning domain, represented using the SAS+ formalism [Bäckström and Nebel,
1995]. A planning domain in SAS+ is defined by the tuple
D = hX , Oi, where
• X = {x1 , . . . , xn } is a set of state variables, each associated with a finite domain Dom(xi ).
• O is a set of actions, where each action a ∈ O is defined
by a tuple hpre(a), eff(a)i, where pre(a) and eff(a) are
assignments of values to state variables, i.e., a set of assignments of the form xi = v where v ∈ Dom(xi ). We
refer to O and its associated sets of preconditions and
effects as the action model of the domain.
A state is also a set of assignments of the form xi = v
such that every variable in X is assigned a single value from
its corresponding domain. As a shorthand notation, if a state
s contains an assignment xi = v we will write s(xi ) = v.
A planning problem in SAS+ is defined by the tuple Π =
hD, sI , sG i, where sI is the initial state and sG is a partial
assignment of some state variables that defines the goal. A
state s is a goal state if sG ⊆ s. For an action a and a state
s we denote by apply(s, a) the state resulting from applying
a on state s. A solution to an SAS+ planning problem is
a plan, which is a sequence of actions a1 , . . . , am such that
sG ⊆ apply(. . . apply(s, a1 ), a2 ) . . . am ).
The key challenge we address in this paper is how to solve a
SAS+ planning problem Π = hD, sI , sG i without having the
action model of D. Instead, a set of trajectories are assumed
to be given.
Definition 1 (Trajectory) A
trajectory
T
=
hs1 , a1 , s2 , a2 , . . . , an−1 , sn i is an alternating sequence
of states (s1 , . . . , sn ) and actions (a1 , . . . , an ) that starts and
ends with a state.
A trajectory represents a successful execution of a sequence
of actions by the agent. A set of trajectories may be obtained,
for example, by monitoring the acting agent when it is controlled manually by a human operator. The states in the given
trajectories are assumed to be fully observable.
Finally, we can define the safe model-free planning problem, which is the focus of this work.
Definition 2 (Safe model-free planning) Let
Π
=
hhX , Oi , sI , sG i be a planning problem and let
T = {T 1 , . . . , T m } be a set of trajectories in the
planning domain hX , Oi. The input to a safe model-free
planning problem is the tuple hX , sI , sG , T i and the task is
to generate a plan π that is a solution to Π. We denote this
safe model-free planning problem as ΠT .
3
Conservative Planning
To solve the safe model-free planning problem, we propose
to learn a conservative action model, and then use it to find
sound plans.
Following prior work on learning action models [Wang,
1995; Wang, 1994; Walsh and Littman, 2008], we partition
every observed trajectory hs1 , a1 , s2 , . . . , sn+1 i ∈ T into a
set of action triplets, where each action triplet is of the form
hsi , ai , si+1 i. Let T (a) be all the action triplets for action a.
A state s and s0 are called pre- and post-state of a, respectively, if there is an action triplet hs, a, s0 i. Following Walsh
and Littman [2008] and Wang [1994; 1995], we observe that
from the set of trajectories T we can “bound” the set of predicates in an action’s preconditions and effects, as follows.
∅ ⊆ pre(a) ⊆
\
s
(1)
s0
(2)
hs,a,s0 i∈T (a)
[
hs,a,s0 i∈T (a)
s0 \ s ⊆ eff(a) ⊆
\
hs,a,s0 i∈T (a)
Equation 1 holds because a value assignment (xi = v) cannot be a precondition of a if it is not in every pre-state
of a, and thus only value assignments that exists in all the
pre-states
of a may be preconditions of a (hence pre(a) ⊆
T
0
hs,a,s i∈T (a) s). On the other hand, the fact that a state variable xi happened to have the same value v in all the pre-states
of the observed trajectories does not necessarily mean that
(xi = v) is a precondition of a. It may even be the case that
a has no preconditions at all, and thus the “lower bound” on
an action’s precondition is ∅.
Equations 2 holds because a value assignment (xi = v)
cannot be an effect of a if it is not in every post-state of
a, and thus only value assignments that exists in all the
post-states of a may be preconditions of a (hence eff(a) ⊆
T
0
hs,a,s0 i∈T (a) s ). On the other hand, every variable that has
a value in the post-state that isSdifferent from the pre-state
then it must be an effect (hence hs,a,s0 i∈T (a) s0 \ s). We denote the “upper bound” of the preconditions by preU
T (a) and
the “lower bound” of the effects by effL
(a).
T
3.1
Compilation to Classical Planning
Next, we use the bounds in Equations 1 and 2 to compile a
safe model-free planning problem ΠT = hX , sI , sG , T i to
a classical SAS+ problem F (ΠT ), such that a solution to
F (ΠT ) is a solution to ΠT , i.e., it is a solution for the underlying planning problem Π. F (ΠT ) is defined as follows.
It has exactly the same set of state variables (X ), start state
(sI ), and goal (sG ) as ΠT . The actions of F (ΠT ) is the set
of all the actions seen in an observed trajectory. We denote
this set of actions by A(T ). The preconditions of an action
a in F (ΠT ) are defined as the “upper bound” estimate given
in Equation 1 (preU
T (a)) and the effects of a in F (ΠT ) are
defined to be the ”lower bound” estimate given in Equation 2
(effL
T (a)).
Definition 3 (Safe) An action model O0 is safe with respect
to an action model O if for every state s and action a it holds
that (1) if a is applicable in s according to O0 then it is also
applicable in s according to O, and (2) applying a to s results
in exactly the same state when using either O0 or O.
Theorem 1 The action model in F (ΠT ) is safe with respect
to the action model of Π.
Proof: Let OΠ and OF (ΠT ) be the action models of Π and
F (ΠT ), respectively. Since a is applicable in s according to
OF (ΠT ) , it means that preU
T (a) ⊆ s and consequently a is
also applicable in s according to OΠ , since pre(a) ⊆ preU
T
(Equation 1).
Now, let s0Π be the state resulting from applying a on s
according to OΠ , and let v denote the value of a state variable xi in s0Π , i.e., v = s0Π (xi ). Since effL
T (a) ⊆ eff(a), then
according to OF (ΠT ) either (xi = v) is an effect of a or a
has no effect on xi . If the former is true then s0F (ΠT ) (xi ) =
v = s0Π (xi ). Otherwise, it means that in the observed trajectories, applying a never changed the value of xi , i.e., xi was
equal to v in both the pre-state and post-state. By definition,
this means that (xi = v) is a precondition of a according to
OF (ΠT ) , and thus s(xi ) = s0F (ΠT ) (xi ) = v = s0Π (xi ). Thus,
the effects of a on s will be the same in both action models,
and hence a s0Π = s0F (ΠT ) .
Corollary 1 (Soundness) Every solution to F (ΠT ) is also a
solution to ΠT
Corollary 1 is a direct result of Theorem 1, and its practical
implication is the following algorithm for solving any safe
model-free planning problem ΠT : compile it to a classical
planning problem F (ΠT ), run an off-the-shelf classical planner, and return the resulting plan. We refer to this algorithm
as the conservative model-free planner. The conservative
model-free planner is sound, but it is not complete. There can
be planning problems that have a solution but the observed
trajectories are not sufficient to induce a corresponding compiled planning problem that is solvable. As an extreme example, if we do not receive any observed trajectories, the compiled planning problem will not have any actions in its action
model and thus will not be able to solve anything. In the next
section we show that the required number of trajectories is
actually reasonable.
Figure 1 illustrates how to generate F (ΠT ) from observed
trajectories in a simple logistic-like domain with one truck,
one package, and three possible location A, B, and C. The
state variables are TruckAt, with domain A, B, and C, and
PackageAt with domain A, B, C, and T , where T represents that the package is on the truck. The possible actions
are MoveX,Y , PickupX , and UnloadX , for every X, Y ∈
{A, B, C}.
The three tables on the left-hand side, T1 , T2 , and T3 are
three observed trajectories, where the column T represents
the value of the state variable TruckAt and the column P represents the value of the state variable PackageAt. For example, T1 represents a trajectory where the truck starts at A,
moves to B, picks up the package, moves to C, and unloads
the package there. The tables on the right-hand side of Figure 1 show the action model learned from these trajectories.
For didactic reasons, we show the action model learned given
just T1 (F (Π{T1 } )), then the action model learned given T1
and T2 (F (Π{T1 ,T2 } )), and finally the action model learned
using all three trajectories.
As can be observed, given only T1 we do not have any
knowledge of many actions such as PickA , PickC , etc.
Also, the preconditions learned for the actions MoveA,B and
MoveB,C are too restrictive, requiring that the package is
at some location (while clearly a Move action only requires
knowing the truck’s location). However, given T1 and T2 ,
these redundant preconditions are removed, and thus task that
can be achieved with the actions MoveA,B , MoveB,C , PickB ,
and UnloadC will be found by our conservative model-free
planner.
4
Learning to Perform Safe Planning
In general, we cannot guarantee that any finite number of trajectories will suffice to obtain precisely the underlying action
model. This is because, for example, if some action never
appears in a trajectory, we may not know its effects; or, if
an action is only used in a limited variety of states, it may be
impossible to distinguish its preconditions. Consequently, we
cannot guarantee a complete solution to the model-free safe
planning problem. However, as the number of trajectories increases, we can hope to learn enough of the actions accurately
enough to be able to find plans for most goals in practice. This
gives raise to a statistical view of the model-free safe planning
task (Definition 2) that follows the usual statistical view of
learning, along the lines of Vapnik and Chervonenkis [1971]
and Valiant [1984].
Definition 4 (Safe Model-Free Learning-to-Plan) We suppose that there is an arbitrary, unknown (prior) probability
distribution D over triples of the form hsI , sG , T i, where sI
is a state, sG is a goal condition, and T is a trajectory that
starts in sI and ends in a state that satisfies sG , and all trajectories are applicable in a fixed planning domain D = hX , Oi.
In the safe model-free
learning-to-plan
D
E
Dtask, we are given
E
(1) (1)
(m) (m)
a set of triplets sI , sG , T (1) , . . . , sI , sG , T (m)
drawn independently from D, and a new SAS+ planning
problem Π = hD, sI , sG i such that the initial state and goal
condition (sI , G) are from some (sI , sG , T ) drawn from D.
The task is to either output a plan for Π or, with probability
at most , return that no plan was found.
Remarks on the task formulation
We stress that D is arbitrary, and thus the conditional distribution of trajectories given a start and goal state, D(T |sI , sG ),
can also be any arbitrary distribution. For example, D
could be the distribution of trajectories obtained by running
some (unknown, sophisticated) planning algorithm on input
(sI , sG ), or produced by hand by a human domain expert.
More generally, the adversarial choice of D in our model
subsumes a model in which a trajectory T (sI , sG ) is nondeterministically and adversarially chosen to be paired with the
start state sI and goal sG . (Indeed, the distributions used in
𝑻𝟏
A
𝑻𝟐
B
C
A
B
𝐅(𝚷 𝑻𝟏 )
𝑻𝟑
C
A
B
T
C
P
T
MoveA,B
s1
T
P
A
B
a1 MoveA,B
s2
B
B
a2 PickB
s3
B
s1
T
P
A
A
a1 MoveA,B
s2
B
A
a2 MoveB,C
T
s3
C
A
a3 MoveB,C
s4
C
T
A
B
C
s1
T
P
A
A
A
C
B
C
Eff B
a2 MoveA,B
Eff C
s3
B
T
a3 UnloadB
s4
B
B
A
B
C
T
PickB
T
P
P
Pre A
Pre A
Pre B
T
Eff B
Eff B
Eff
B
UnloadC
MoveB,C
Pre B
Pre B
Pre C
T
Eff C
Eff C
Eff
C
PickB
PickA
Pre B
B
Pre B
B
Pre A A
Eff
T
Eff
T
Eff
UnloadC
T
UnloadB
MoveA,B
MoveB,C
T
A
a4 UnloadC
s5
Pre A B
Pre B
s2
P
MoveA,B
MoveB,C
a1 PickA
𝐅(𝚷 𝑻𝟏 ,𝑻𝟐 ,𝑻𝟑 )
𝐅(𝚷 𝑻𝟏, 𝑻𝟐 )
UnloadC
T
PickB
Pre C
T
Pre C
T
Pre B B
Eff
C
Eff
C
Eff
T
C
Figure 1: An example of creating F (ΠT ) from observed trajectories in a simple logistics-like domain. The left-hand side of the
figure shows the three given trajectories T1 , T2 , and T3 . The right hand side of the figure shows our conservative action model,
learned using different subsets of the trajectories T1 , T2 , and T3 together.
such a case satisfy the stronger restriction that D(T |sI , sG )
produces a deterministic outcome T (sI , sG ), which does not
necessarily hold in our model.)
We also note that our conservative model-free planner does
not actually require knowledge of the goals sG associated
with the trajectories drawn from D. Thus, our approach actually solves a more demanding task that does not provide
the goals to the learning algorithm. But, such a distribution
over goals is nevertheless a central feature in our notion of
“approximate completeness,” and features prominently in the
analysis as we discuss next.
Analysis of learning
A key question is how our certainty that a plan can be generated for a new start and goal state grows with the number
of trajectories. Let log and ln denote log2 and loge , respectively, and let d denote the largest number of values for a state
variable.
Theorem 2 Using the conservative model-free planner, it is
sufficient to observe m ≥ (2 ln d)|A| (|X | + log 2|A|
δ ) trajectories to solve the safe model-free learning-to-plan problem
with probability 1 − δ.
Proof Outline. First, Lemma 1 shows that the set of actions
used by our conservative model-free planner (A(T )) is sufficient to solve a randomly drawn problem with high probability. Then, Lemma 2 shows that under certain conditions, the
preconditions we learned for these actions (preU
L ) are not too
conservative, i.e., they are adequate for finding a plan with
high probability. Finally, we prove that with high probabil-
ity these conditions over the action model we learned indeed
hold.
Lemma 1 Let A be the set of actions such that each action
a ∈ A appears in a trajectory sampled from D with prob
and let A(T ) be the set of every action
ability at least 2|A|
that appeared in a trace. The probability that all the actions
a ∈ A appear in A(T ) is at least 1 − δ/2.
Proof: By definition, the probability that an action a ∈ A
does not exist in a trajectory drawn from D is 1 − 2|A|
. Since
the observed trajectories T are drawn independently from D
)m ≤
we have that the probability that a ∈
/ A(T ) is (1 − 2|A|
·m
e− 2|A| , using the inequality 1−x ≤ e−x . Since we assume in
Theorem 2 that m ≥ (2 ln d)|A| (|X |+log 2|A|
δ ) which is larger
than (2 ln 2)|A| log 2|A|
,
we
have
that
the
probability
that a ∈
/
δ
A(T ) is at most
·m
e− 2|A| ≤ e− 2|A| ·
(2 ln 2)|A|
log
2|A|
δ
= e− ln
2|A|
δ
=
δ
2|A|
(3)
Hence, by a union bound over a ∈ A (noting |A | ≤ |A|),
we have that A ⊆ A(T ) with probability 1 − δ/2 as needed.
Stated informally, Lemma 1 says that with high probability
we will observe all the “useful” actions, i.e., the actions used
in many trajectories. However, we may have learned preconditions for these actions that are too conservative, preventing
the planner from finding a plan even if one exists. We next
define a property of action models that states that this does
not occur frequently.
Definition 5 (Adequate) We call an action model -adequate
if, with probability at most /2, we sample a trajectory T
from D such that T contains an action triplet hs, a, s0 i where
a ∈ A and s does not satisfy preU
T.
We say that an action model is -inadequate if it is not adequate. An equivalent way to define the notion of an adequate action model is that with probability at most /2 a
trajectory is sampled in which an action a ∈ A is invoked on
a state that does not satisfy the conservative preconditions of
a we learned from the given set of trajectories (T ).
Lemma 2 If the learned action model is -adequate and
A ⊆ A(T ), then with probability 1 − our conservative model-free planner will find a plan for a start-goal pair
(sI , sG ) sampled from D.
Proof:
Let T be the (unknown) trajectory sampled for
(sI , sG ). The probability that T uses an action that is not
in A is at most |A| · 2|A|
= /2. Thus, T contains only actions known to our planner with probability at least 1 − /2,
since we assumed that A ⊆ A(T ). Since the action model is
-adequate then with probability 1 − /2 the learned preconditions are satisfied on all of the states in T . Thus, by a union
bound, we find that with probability 1 − , our planner could
at least find T . Hence it will find a trajectory from sI to sG ,
as required.
Lemma 3 The action model used by our conservative modelfree planner is -adequate with probability at least 1 − δ/2.
Proof: Whether an action model is -adequate or not depends on the assignment of preconditions to actions. Since
there are |X | state variables each with at most d values, then
there are d|X | possible assignments of preconditions for an individual action and a total of d|X |·|A| possible preconditions
assignments for an action model . Let BAD be the subset
of these action model preconditions assignments that are not
-adequate. Clearly, BAD has size at most d|X |·|A| .
Consider a particular assignment of preconditions in an
-inadequate action model OBAD . Since OBAD is inadequate, it has a set of state-action pairs (s, a) associated
with it such that a ∈ A and a cannot be applied to s according to OBAD . The action model OBAD can only be
learned by our algorithm if none of these state-action pairs
were observed in the given trajectories T . On the other
hand, by the definition of inadequacy the probability of having a state-action pair from that list in a trajectory drawn
from D is at least /2. Thus, the probability that our algorithm will learn a particular preconditions assignment of
an -inadequate action model is at most (1 − /2)m . Since
2 ln d
2
m ≥ (2 ln d)|A| (|X | + log 2|A|
δ ) >
(|A| · |X | + log δ ),
then (1 − /2)m is smaller than
− 2 ·m
−(ln d·(|A|·|X |+log
2
δ ))
− ln d·|A|·|X |
− ln
2
δ
e
≤e
≤e
·e
(4)
which is at most 2·d|Xδ|·|A| . Thus, by a union bound over this
set of inadequate assignments BAD, the probability that any
inadequate assignment of preconditions could be output is at
most δ/2. Thus, with probability 1 − δ, the algorithm indeed produces an assignment of preconditions that is neither
unsafe for A(T ) nor inadequate for A , as needed.
S(Π)
¬S(Π)
Marginals
P(Π)
1−
0
µ(1 − )
¬P(Π)
1
1 − µ(1 − )
Priors
µ
1−µ
Table 1: The table shows the priors, conditional probabilities,
and marginals obtained with probability 1 − δ when using
our conservative model-free planner when it is given m ≥
(2 ln d)|A|
(|X | + log 2|A|
δ ) trajectories.
4.1
Unsolvable Instances
The implication of Theorem 2 is that by observing a number
of trajectories that is quasi-linear in the number of actions and
the number of state variables, we expect our safe model-free
planner to be complete with high probability, in the sense that
if a solution exists it will be found. But what if some of the
drawn problem instances are not solvable?
Corollary 2 If the probability of drawing a solvable startgoal pair from D is µ then it is sufficient to observe m ≥
(2 ln d)|A|
(|X | + log 2|A|
δ ) trajectories (of solvable instances)
to guarantee that with probability of at least 1−δ our conservative model-free planner will solve a start-goal pair drawn
from D with probability at least µ · (1 − )
Proof: Let S(Π) be true or false if a given planning problem Π is solvable or unsolvable, respectively, and let P(Π)
be true or false if our planner returns a solution to Π or not
respectively. We aim to bound P r(P(Π)). Since our planner
is sound, P(Π) → S(Π) and so
P r(P(Π)) = P r(P(Π)∧S(Π)) = P r(P(Π)|S(Π))·P r(S(Π))
According to Theorem 2, P r(P(Π)|S(Π)) ≥ 1 − and
P r(S(Π)) = µ by definition.
Table 1 shows the priors, conditional probabilities, and
marginals use by the proof of Corollary 2 The first row
shows the probabilities P r(P(Π)|S(Π)), P r(¬P(Π)|S(Π)),
abd P r(S(Π)); the second row shows the probabilities
P r(P(Π)|¬S(Π)), P r(¬P(Π)|¬S(Π)), and P r(¬S(Π)); and
the last row shows the marginal probabilities P r(P(Π)) and
P r(¬P(Π)).
Corollary 2 and Table 1 are valuable in that they provides
a relationship between µ, , δ, and m. Thus, we can increase
m to satisfy more demanding values of µ, , and δ and different types of error bounds. For example, consider an application that requires bounding, by some γ, the probability that
our planner outputs incorrectly that no plan exists. In other
words, an application that requires
P r(S(Π)|¬P(Π)) ≤ γ
Using Bayes’ rule and the values from Table 1, this means
that
γ · (1 − µ)
·µ
≤γ⇔≤
1 − (1 − ) · µ
µ · (1 + γ)
Plugging γ·(1−µ)
µ·(1+γ) into the sample complexity instead of in
Theorem 2 will give the required number of trajectories to obtain a bound of γ on the probability of incorrectly outputting
that a problem is not solvable.
4.2
Limited Planner Capabilities
The given trajectories T are presumably generated by some
planning entity. Since planning in general is a hard problem,
it may be the case that the planner that generated the given set
of trajectories has drawn a solvable problem from D but was
just not able to solve it due to memory or time constraints.
Learning from such a set of trajectories does not enable
bounding the probability of solving problems in general.
What can be obtained in such cases is to bound the solving
capabilities of our conservative model-free planner with respect to the capabilities of the planner that generated the observed trajectories. Thus, instead of having µ represent the
probability that an instance is solvable, we will have µ represent the probability that an instance is solvable by the original
planner. The rest of the analysis follows exactly the same as
in the previous section.
5
Related Work
Our work relates to several well-studied types of problems:
planning under uncertainty, reinforcement learning, and domain model learning.
Planning under uncertainty. In common models for planning under uncertainty, such as Markov Decision Problems
(MDP) and Partially Observable MDPs (POMDP), the uncertainty stems from the stochastic nature of the world or
from imperfect sensors that prevent full observability of the
agent’s state. Our setting is different in that our uncertainty
only stems from not knowing the agent’s action model.
Reinforcement learning. Reinforcement learning algorithms learn how to act by interacting with the environment.
Thus, they are designed for a trial-and-error approach to learn
the domain and/or how to plan in it. Our task is to generate a
plan that must work, so a trial-and-error approach is not sufficient.
Domain model learning. Most prior work on learning a
domain model in general or a STRIPS action model from observed trajectories, such as ARMS [Yang et al., 2007] and
LOCM [Cresswell et al., 2013], learn approximate models
that do not guarantee safety. Hence, such work generally
also involves some form of trial-and-error as well, iteratively
requesting more example trajectories or interacting directly
with the environment to refine the learned model [Mourão
et al., 2012; Wang, 1994; Wang, 1994; Walsh and Littman,
2008; Levine and DeJong, 2006; Jiménez et al., 2013]. In addition, most works learn from both positive and negative examples – observing successful and failed trajectories, while
we only require successful trajectories to be provided.
Another key difference is that unlike our work, most prior
works do not provide statistical guarantees on the soundness
of the plan generated with their learned model. An exception
to this is the work of Walsh and Littman [2008], that also
discussed the problem of learning STRIPS operators from
observed trajectories and provided theoretical bounds on the
sample complexity – the number of interactions that may fail
until the resulting planner is sound and complete. By contrast, we do not assume any planning and execution loop and
do not allow failed interactions. Hence, we aim for a planning
algorithm that is guaranteed to be sound, at the cost of completeness. This difference affects their approach to learning.
They attempted to follow an optimistic assumption about the
preconditions and effects of the learned actions, in an effort
to identify inaccuracies in their action model. By contrast,
we are forced to take a pessimistic approach, as we aim for a
successful execution of the plan rather than information gathering to improve the action model.
6
Conclusions
This paper deals with a planning problem in which the planner agent has no knowledge about its actions. Instead of an
action model, the planner is given a set of observed trajectories of successfully executed plans. In this setting we introduced the safe model-free planning problem, in which the
task is to find a plan that is guaranteed to reach the goal, i.e.,
there is no tolerance for execution failure. This type of problem is important in cases where failure is costly or in cases
where the agent has no capability to replan during execution.
We showed how to use the given set of trajectories to learn
about the agent’s actions, bounding the set of predicates in the
actions’ preconditions and effects. Then, we proposed a conservative approach to solve the safe model-free problem that
is based on a translation to a classical planning problem. This
solution is sound but is not complete, as it may fail to find
a solution even if one exists. However, we prove that under
some assumptions the likelihood of finding a solution with
this approach grows linearly with the number of predicates
and quasi-linearly with the number of actions.
Future directions for safe model-free planning include
studying how to address richer underlying planning models
including parametrized actions, conditional effects, stochastic action outcomes, and partial observability. While some
of these more complex action models can be compiled away
(e.g., a problem with conditional effects can be compiled to
a problem without conditional effects [Nebel, 2000]), the resulting problem can be significantly larger. A particularly interesting direction is how to learn there lifted action model,
i.e., what can be learned from a trajectory with an action a(x)
on the action model of a(y), where a is a parameterized action and x and y are different values for the same parameter.
Acknowledgements
B. Juba was partially supported by an AFOSR Young Investigator Award. R. Stern was partially supported by the Cyber
Security Research Center at BGU.
References
[Bäckström and Nebel, 1995] Christer Bäckström and Bernhard Nebel. Complexity results for SAS+ planning. Computational Intelligence, 11(4):625–655, 1995.
[Cavalcanti et al., 2007] Adriano Cavalcanti, Bijan Shirinzadeh, Robert A Freitas Jr, and Tad Hogg. Nanorobot
architecture for medical target identification. Nanotechnology, 19(1):015103, 2007.
[Cresswell et al., 2013] Stephen N. Cresswell, Thomas L.
McCluskey, and Margaret M. West. Acquiring planning
domain models using LOCM. The Knowledge Engineering Review, 28(02):195–213, 2013.
[Daniely and Shalev-Shwartz, 2016] Amit Daniely and Shai
Shalev-Shwartz. Complexity theoretic limtations on learning DNF’s. In Proceedings of the 29th Conference on
Computational Learning Theory, volume 49 of JMLR
Workshops and Conference Proceedings, pages 1–16.
2016.
[Fikes and Nilsson, 1971] Richard E Fikes and Nils J Nilsson. STRIPS: A new approach to the application of theorem proving to problem solving. Artificial intelligence,
2(3-4):189–208, 1971.
[Jiménez et al., 2013] Sergio Jiménez, Fernando Fernández,
and Daniel Borrajo. Integrating planning, execution, and
learning to improve plan execution. Computational Intelligence, 29(1):1–36, 2013.
[Kakade, 2003] Sham M. Kakade. On the sample complexity of reinforcement learning. PhD thesis, University of
London, 2003.
[Kearns and Valiant, 1994] Michael Kearns and Leslie
Valiant. Cryptographic limitations on learning boolean
formulae and finite automata. Journal of the ACM
(JACM), 41(1):67–95, 1994.
[Kearns et al., 2002] Michael Kearns, Yishay Mansour, and
Andrew Ng. A sparse sampling algorithm for near-optimal
planning in large Markov decision processes. Mach.
Learn., 49(2):193–208, 2002.
[Levine and DeJong, 2006] Geoffrey Levine and Gerald DeJong. Explanation-based acquisition of planning operators. In ICAPS, pages 152–161, 2006.
[Littman, 1994] Michael L. Littman. Memoryless policies:
Theoretical limitations and practical results. In From Animals to Animats 3: Proceedings of the Third International Conference on Simulation of Adaptive Behavior,
volume 3, page 238. MIT Press, 1994.
[McDermott et al., 1998] Drew McDermott, Malik Ghallab,
Adele Howe, Craig Knoblock, Ashwin Ram, Manuela
Veloso, Daniel Weld, and David Wilkins. PDDL-the planning domain definition language. Technical report, AIPS
’98 - The Planning Competition Committee, 1998.
[Meuleau et al., 1999] Nicolas Meuleau, Kee-Eung Kim,
Leslie Pack Kaelbling, and Anthony R. Cassandra. Solving POMDPs by searching the space of finite policies. In
Proceedings of the Fifteenth conference on Uncertainty in
artificial intelligence, pages 417–426. Morgan Kaufmann
Publishers Inc., 1999.
[Mourão et al., 2012] Kira Mourão, Luke S Zettlemoyer,
Ronald Petrick, and Mark Steedman. Learning STRIPS
operators from noisy and incomplete observations. In UAI,
pages 614–623, 2012.
[Nebel, 2000] Bernhard Nebel. On the compilability and
expressive power of propositional planning formalisms.
Journal of Artificial Intelligence Research, 12:271–315,
2000.
[Valiant, 1984] Leslie G. Valiant. A theory of the learnable.
Communications of the ACM, 18(11):1134–1142, 1984.
[Vapnik and Chervonenkis, 1971] V. Vapnik and A. Chervonenkis. On the uniform convergence of relative frequencies of events to their probabilities. Theory of Probability
and its Applications, 16(2):264–280, 1971.
[Walsh and Littman, 2008] Thomas J. Walsh and Michael L.
Littman. Efficient learning of action schemas and webservice descriptions. In the National Conference on Artificial Intelligence (AAAI), pages 714–719, 2008.
[Wang, 1994] Xuemei Wang. Learning planning operators
by observation and practice. In the International Conference on Artificial Intelligence Planning Systems (AIPS),
pages 335–340, 1994.
[Wang, 1995] Xuemei Wang. Learning by observation and
practice: An incremental approach for planning operator
acquisition. In ICML, pages 549–557, 1995.
[Yang et al., 2007] Qiang Yang, Kangheng Wu, and Yunfei
Jiang. Learning action models from plan examples using weighted MAX-SAT. Artificial Intelligence, 171(23):107–143, 2007.
| 2cs.AI |
arXiv:1709.07021v1 [math.GR] 20 Sep 2017
UNIQUELY LABELLED GEODESICS OF COXETER GROUPS
ELISABETH FINK AND KIRILL ZAINOULLINE
Abstract. Studying geodesics in Cayley graphs of groups has been a very
active area of research over the last decades. We introduce the notion of a
uniquely labelled geodesic, abbreviated with u.l.g. These will be studied first
in finite Coxeter groups of type An . Here we introduce a generating function,
and hence are able to precisely describe how many u.l.g.’s we have of a certain
length and with which label combination. These results generalize several
results about unique geodesics in Coxeter groups. In the second part of the
paper, we expand our investigation to infinite Coxeter groups described by
simply laced trees. We show that any u.l.g. of finite branching index has finite
length. We use the example of the group D̃6 to show the existence of infinite
u.l.g.’s in groups which do not have any infinite unique geodesics. We conclude
by exhibiting a detailed description of the geometry of such u.l.g.’s and their
relation to each other in the group D̃6 .
1. Introduction
Coxeter groups are historically very important and occur naturally as reflection
groups (see e.g. [Hu90]). Over the decades they have sparked immense interest from
various sides of mathematics and physics.
In particular, geodesics on their Cayley graphs, shortest connections between
points which represent reduced words, have been of special interest [St84], [BH93],
[He94], [Ed95], [St97], [El97], [EE10], [LP13], [HNW], [Ha17]. In [AC13] and [CK16]
the authors introduce a formal power series with coefficients the number of geodesics
for right-angled and even Coxeter groups based on trees. The paper [MT13] relates
geodesics and quasi-geodesics for Coxeter groups. Related to the formal power
series of geodesic growth is the growth series of a group introduced in [P90]. This
series has been studied for different types of Coxeter groups in for example [CD91],
[M03], [A04] and been generalized in [GN97].
In this paper we introduce the notion of a uniquely labelled geodesic. Instead
of limiting our investigation to the existence of geodesics and their uniqueness, we
reach out to geodesics which are unique with respect to their total label combination
seen along the path going to a fixed point in the graph. For example, in abelian
groups these consist only of powers of the generators, as any word with more than
one letter can be written in any other order of the generators and would still yield
the same element. On the other hand, any unique geodesic is also the unique
geodesic with that label combination reaching to the element it represents.
2010 Mathematics Subject Classification. 20F55, 20F65, 05E15.
Key words and phrases. Coxeter group, unique geodesic, generating function, reduced word.
1
2
E. FINK AND K. ZAINOULLINE
Geometrically speaking, it can be shown that a word in the generators of a group
can only be a u.l.g. if it is a connected path on the Coxeter diagram which is a graph
describing the group. Hence asking for a certain u.l.g. is the equivalent of the graph
theoretical problem of finding a connected path in a graph with visiting each vertex
a given number of times.
In the first part of the paper, we study the finite Coxeter groups of type A with
n generators. We introduce a generating function, a power series in n variables,
where each monomial represents a certain label combination. We then give a precise
formula for each coefficient depending on the monomial (Corollary 3.6). Based on
this, we give exact formulas for the number of non-zero coefficients (Corollary 3.7)
as well as the total number of u.l.g.’s in these groups (Theorem 3.8), both dependent
on n.
In the second part of the paper, we expand our study to infinite groups. We
introduce the notion of a branching index (Definition 5.1) which roughly speaking
describes the turning behaviour of a connected path in the Coxeter-Dynkin diagram.
We show that any u.l.g. with finite branching index has finite length (Theorem 5.9).
We then study the affine group D̃6 [BB, Appendix A1, Table 2]. In this group, we
exhibit an infinite periodic u.l.g. (Theorem 6.6). We show that there are indeed two
more different u.l.g.’s and relate these with each other geometrically (Theorem 6.8).
Acknowledgements. Both authors were partially supported by the NSERC Discovery Grant RGPIN-2015-04469.
2. Coxeter groups
Let W be a Coxeter group of rank n that is given by generators and relations
W = hs1 , . . . , sn | s2i = 1, (si sj )mij = 1i,
where mij ≥ 2 are the Coxeter exponents. Consider its Cayley graph C with respect
to the chosen generators s1 , . . ., sn : vertices of C correspond to elements w ∈ W
and two vertices w and w′ are connected by an edge and labelled by i iff
w′ = wsi and l(w′ ) = l(w) + 1,
where l : W → Z≥0 is the length function on W . Then the shortest path ρ(w)
connecting 1 and w ∈ W corresponds to a reduced expression for w = si1 . . . sil and
its length coincides with the length l of w; such a path will be called a geodesic.
We will use a slightly different labeling of the Cayley graph: instead of an integer
i ∈ {1, . . . , n} we put the standard vector {(0, . . . , 1, . . . , 0)} where 1 is at the ith
position. By a (total) label of a geodesic ρ(w) denoted by ρ
~(w) we call the sum of
labels (considered as vectors in Zn ) of all edges of ρ or, equivalently, it is an n-tuple
(i1 , . . . , in ), where ik is the number of generators sk used to express w.
We say ρ(w) is a uniquely labelled geodesic (u.l.g.) on the Cayley graph of W , if
there is only one geodesic connecting 1 and w with label ρ~(w). Observe that u.l.g.’s
correspond to elements w ∈ W that have a unique reduced expression for a given
label.
We define the generating function
UC (t1 , . . . , tn ) =
X
(i1 ,...,in )
ai1 ,...,in ti11 . . . tinn ,
UNIQUELY LABELLED GEODESICS OF COXETER GROUPS
3
where ai1 ,...,in is the number of u.l.g.’s with label (i1 , . . . , in ), i.e.,
~(w) = (i1 , . . . , in )}|.
ai1 ,...,in = |{w ∈ W | ρ(w) is a u.l.g. with ρ
2.1. Example. The Cayley graph of the symmetric group
S4 = hs1 , s2 , s3 | s2i = 1, (s1 s2 )3 = 1, (s2 s3 )3 = 1, [s1 , s3 ] = 1i
has the form
❧ 1 ❘❘❘❘❘
❘3❘❘
❧❧❧❧
1❧
❧
2
❘❘❘❘
❧
)
u❧❧❧❧
s2 ❘❘
s1 ❘❘
❧ s3 ◗◗◗◗◗
❧
♠
❘
❘
❧
❧
♠
❘
❘
❧
❧
♠
❘
❘
❧
❧
♠
❘
❘
2◗◗◗
♠2♠
❧❧❧❘❧❘3❘❘❘
❧❧❧❘❧❘3❘❘❘
◗◗(
u❧❧❧1
)
u❧❧❧1
)
v♠♠♠♠
s1 s3
s2 s1
s2 s3 ◗
s3 s2
s1 s2
◆◆◆◆
◗◗◗
❋❋
♠♠
♣♣
①
♠
♣
♠
♣
❋
①
◗
2
2 ◗
1◆◆◆
3
2
1
3
♠
❋
①
♣
◗
♠
♣
❋❋
◆'
◗(
v♠♠
w♣♣
①①
①
s3 s2 s1
s1 s2 s3
s2 s3 s2
3❋
❋❋ ①s1 s3 s2❋❋ ①①1
❃❃❖❖❖ s1 s2 s1
♦
❃❃ ❖❖❖❖
♦♦♦
①①❋❋
①①❋❋
❖❖❖
❃
♦♦♦
①① ❋" |①① ❋❋
♦
①
♦
s2 s1 s3
❖❖❖
♦
3
1❋
1❃
❋❋
①3
❃❃
❖2❖
♦♦♦ 1
①①
❋❋
2♦
❖❖❖
❃❃3
♦
①
2
♦
❋" w♦♦♦
❖' |①①
s1 s2 s3 s1
s3 s2 s1 s3
◗◗◗◗ s1 s2 s3 s2 ❘❘❘❘❧❧❧ s2 s1 s3 s2 ❘❘❘❘❧❧❧ s3 s2 s1 s2
♠
❧❘❘
♠♠♠
2◗◗◗
2
❧❧❧❘❘ 1❘❘
◗( u❧❧3 ❧❧ 1❘❘)
) v♠♠♠♠
u❧❧3
s3 s2 s1 s3 s2
s1 s2 s3 s1 s2
❘❘❘❘ s1 s3 s2 s3 s1
❧❧❧
3❧
2
1❘❘❘
❧
❘) u❧❧❧
s1 s2 s3 s1 s2 s1
which gives a polynomial generating function
UC (t1 , t2 , t3 ) = 1 + (t1 + t2 + t3 ) + (2t1 t2 + 2t2 t3 )
+ (t21 t2 + t1 t22 + 2t1 t2 t3 + t22 t3 + t2 t23 )
+ 2t1 t22 t3 + (t21 t22 t3 + t1 t22 t23 + 2t1 t32 t3 ).
2.2. Example. Suppose Γ is a finite graph where each two vertices are connected
by at most one edge. Let W be a right-angled Coxeter group associated to Γ,
i.e. adjacent vertices α and β of Γ correspond to generators sα and sβ in W with
(sα sβ )∞ = 1 (there are no relations between sα and sβ ), nonadjacent vertices
correspond to commuting generators and s2α = 1 for all α.
Since there are no relations between adjacent generators, u.l.g.’s in W are in
1-1 correspondence to (connected) paths on Γ. Hence, if we index vertices of Γ
(generators of W ) as α1 , . . . , αn , then the coefficient ai1 ,...,in of the generating
function UC counts
the number of connected paths in Γ that pass through the vertex
αj exactly ij times.
2.3. Example. Suppose W is an affine Weyl group Ã2 , i.e.
W = {s1 , s2 , s3 | s2i = 1, (si sj )3 = 1, i 6= j}.
Its Coxeter-Dynkin diagram Γ is a triangle
α1
④④
④④
④
④
④④
α3 ❈
❈❈
❈❈
❈❈
❈
α2
4
E. FINK AND K. ZAINOULLINE
where edges correspond to the braid relations (si sj )3 = 1 (there are no commuting
generators). Then the infinite word (s1 s2 s3 )∞ has the property that any finite
connected subword is a unique geodesic, in particular, it is a u.l.g. So the generating
function UC is a formal power series (not a polynomial) in Z[[t1 , t2 , t3 ]]. This has
been extensively studied in [LP13] and is closely related to the famous problem of
constructing infinite reduced words.
In the present paper we construct infinite u.l.g.’s (hence, infinite reduce words)
for some Coxeter groups whose Coxeter-Dynkin diagrams are simply-laced tree.
3. Coxeter groups of type A
Consider the case of a finite Coxeter group W of type A of rank n, i.e., mij = 2
if |i − j| > 1 and mij = 3 if |i − j| = 1 or, equivalently, the Coxeter-Dynkin diagram
of W is a chain:
◦2
◦3
...
◦n−1
◦n
◦1
We have the following observations that hold for any group of type A
3.1. Lemma. All non-zero monomials of UC are of the form
i
l+1
till tl+1
. . . timm ,
1 ≤ l ≤ m ≤ n,
where all the exponents il , il+1 , . . . , im are non-zero.
Proof. Assume that a geodesic (reduced word) ρ(w) contains no generator sk , where
2 ≤ k ≤ n − 1, but it contains generators from both subsets S = {s1 , . . . , sk−1 } and
T = {sk+1 , . . . , sn }. Then without loss of generality, ρ(w) must contain a subword
xy where x ∈ S, y ∈ T , i.e., w = u · xy · v. Since xy = yx, w can also be written
as w = u · yx · v. Both ways of writing w result in two different geodesics with the
same labels. Hence, ρ(w) can not be a u.l.g.
From the proof it follows
3.2. Corollary. In a u.l.g. ρ(w) with l(w) ≥ 2, any two adjacent generators w =
. . . si sj . . . must satisfy the condition |i − j| = 1.
3.3. Lemma. Suppose ρ(w) is a u.l.g. Then it can not contain subwords of the
form
sm sm−1 . . . sl+1 sl sl+1 . . . sm−1 sm sm−1 ,
sm−1 sm sm−1 . . . sl+1 sl sl+1 . . . sm−1 sm ,
sl sl+1 . . . sm−1 sm sm−1 . . . sl+1 sl sl+1 ,
sl+1 sl sl+1 . . . sm−1 sm sm−1 . . . sl+1 sl , where l + 2 ≤ m.
Proof. It is enough to prove it for the first word only (other words follow by symmetry). Suppose ρ(w) ∈ W contains such a subword. Then applying relations in
the Coxeter group we obtain
sm sm−1 . . . sl+1 sl sl+1 . . . sm−1 sm sm−1 =
sm sm−1 . . . sl+1 sl sl+1 . . . sm sm−1 sm =
sm sm−1 sm . . . sl+1 sl sl+1 . . . sm−1 sm =
sm−1 sm sm−1 . . . sl+1 sl sl+1 . . . sm−1 sm .
UNIQUELY LABELLED GEODESICS OF COXETER GROUPS
5
Since the first and the last subwords are different but have the same number of
occurrences of each generator, i.e., the same label, ρ(w) can not be a u.l.g.
Any word ρ(w) that does not contain subwords of the lemma, must have one of
the following forms (up to inversing the indices of generators sk 7→ sn+1−k ):
I. Suppose n ≥ 1 and l ≤ m. It decreases from m to l, that is ρ(w) = sm sm−1 . . . sl .
II. Suppose n ≥ 2 and l < i, j. First, it decreases from i till l and then it increases
till j.
si
sj
❇❇
⑤>
❇❇
⑤
❇ ⑤⑤⑤
sl
III. Suppose n ≥ 3 and l < i, j < m. First, it decreases from i till the absolute
minimum index l, then it increases till the absolute maximum index m and, finally,
decreases again till some index j. This can be depicted as follows:
si
✶✶
✶✶
✶✶
sm
✡D ✸✸✸
✡
✸✸
✡✡
✸
✡✡
✡
sj
✡
✡
✡✡
sl
3.4. Corollary. For n ≥ 3 the maximal length of a u.l.g. is 3n − 4.
Proof. The longest such word is of the form III (i = n − 1, l = 1, m = n, j = 2)
sn−1 sn−2 . . . s2 s1 s2 . . . sn−1 sn sn−1 . . . s2 ,
its length is 3n − 4 and it is a u.l.g. with label (1, 3, 3, . . . , 3, 3, 1).
The following theorem describes monomials of UC
3.5. Theorem. A non-zero monomial of UC has to be of the following type
I. tl tl+1 . . . tm for l ≤ m,
II. tl (tl+1 . . . ti )2 ti+1 . . . tm (and the inverse by tk 7→ tn+1−k ) for l < i ≤ m,
III. (a) tl (tl+1 . . . ti )2 (ti+1 . . . tj−1 )(tj . . . tm−1 )2 tm for l < i < j < m,
(b) tl (tl+1 . . . tj−1 )2 (tj . . . ti )3 (ti+1 . . . tm−1 )2 tm for l < j ≤ i < m.
Observe that Type III(a) for |i − j| = 1 overlaps with Type II for m = i + 1.
Proof. All words of forms I, II and III are u.l.g.’s that have labels and, hence,
monomials of the respective types I, II and III.
3.6. Corollary. We have the following formula for the generating function
X
aı ti11 . . . tinn ,
UC (t1 , . . . , tn ) =
ı=(i1 ,...,in )
6
E. FINK AND K. ZAINOULLINE
where the coefficients aı depend on the type of the label (monomial) ı = (i1 , . . . , in )
as follows
Type I with l = m or Type II with m = i
1
2
Type I with l < m or Type II with m > i + 1 or Type III(b),
aı =
4
Type III(a) with |i − j| > 1,
2(m − l) Type III(a) with |i − j| = 1 or Type II with m = i + 1.
Proof. We prove the last case only (previous cases follow similarly). Given a minimum index l and a maximum index m (m ≥ l + 3) as the initial index i of
a generator of the word we can choose any i ∈ {l + 1, . . . , m − 2} which gives
(m − 2) − (l + 1) + 1 = m − l − 2 different options. Inversing the indices gives the
same number of options. Hence, we obtain 2(m − l − 2) options for Type III(a).
As for Type II, a minimum index l and a maximum index m = i + 1 (m ≥ l + 2)
give two different words (up to an inverse), so we have exactly 4 options. Hence,
aı = 2(m − l − 2) + 4 = 2(m − l).
3.7. Corollary. There are exactly
in UC (incl. the constant term).
1 4
1 3
17 2
1
12 n − 6 n + 12 n − 3 n + 1
non-zero coefficients
Proof. We sum the number of respective coefficients for each type of a u.l.g.
In type I there are following cases for each pair (l, m)
(1) Case |m − l| > 1: the choices of m and l divide the list of n indices into
three parts 0 . . . 01 | 1 . . . 1 | 10 . . . 0 which gives n−1
options.
2
(2) Case |m − l| = 1: there are n − 1 options to choose (l, l + 1).
(3) Case m = l: we have exactly n options.
Hence, in total for type I, we obtain n−1
+ 2n − 1 options.
2
In type II we have the following cases
(1) Case m = i+1: this amounts
to a partition into three parts 0 . . . 01 | 2 . . . 2 |
10 . . . 0 which gives n−1
options.
2
(2) Case |m − i| > 1: this amounts to a partition into four
non-trivial parts
0 . . . 01 | 2 . . . 2 | 1 . . . 1 | 10 . . . 0, hence, giving us n−1
options. Since a
3
monomial is not symmetric, this number doubles to 2 · n−1
by reversing
3
the generators.
(3) Case m = i: if m < n, we split the list of indices
into three non-trivial parts
0 . . . 01 | 2 . . . 2 | 0 . . . 0 which leads to n−1
options; if m = n, then we
2
split it into two non-trivial
parts
and,
hence,
obtain
n − 1 options. So, in
total we get n−1
+
n
−
1
possibilities.
Since
a
monomial
is not symmetric,
2
this number doubles to (n − 1)(n − 2) + 2n − 2 by reversing the generators.
Finally, in type III we have
(1) Type III(a), case |j − i| > 1: this is the same number as in Type III(b)
with |i + 1 − (j − 1)| > 1. Both need 5 partitions of n
0 . . . 01|2 . . . 2|1 . . . 1|2 . . . 2|10 . . . 0,
0 . . . 01|2 . . . 2|3 . . . 3|2 . . . 2|10 . . . 0,
UNIQUELY LABELLED GEODESICS OF COXETER GROUPS
7
n−1
hence, in both cases individually we have n−1
5−1 options. This gives 2 ·
4
options in total.
(2) Type III(a), case |j − i| = 1: we have a partition into three parts 0 . . . 01 |
2 . . . 2 | 10 . . . 0. However, we have already considered coefficients with these
exponents in Type II, case (1).
(3) Type III(b), j = l + 1: wehave a partition into three parts 0 . . . 01 | 3 . . . 3 |
10 . . . 0 giving again n−1
options.
2
(4) Type III(b) with m = i + 1: we have a partition into four non-trivial
parts
0 . . . 01 | 2 . . . 2 | 3 . . . 3 | 10 . . . 0 for which we obtain again n−1
options.
3
Since a monomial is not symmetric, this number doubles to 2 · n−1
by
3
reversing the generators.
A complete list of different monomials with non-zero coefficients is given for
A3 , A4 , A5 in the Appendix 7.1.
3.8. Theorem. In a Coxeter group of type An there are exactly
if n = 2
6
UC (1, 1, . . . , 1) = 19
if n = 3
1 4 3 3 20 2 19
3 n − 2 n + 3 n − 2 n + 1 if n ≥ 4
uniquely labelled geodesics.
Proof. We combine the Corollaries 3.6 and 3.7. The geodesics counted in the proof
of 3.7 have the following coefficients and multiplicities:
Case of the proof
multiplicity
Type I, (1)
2
number of different geodesics of this type
n−1
Type I, (2)
2
n−1
Type I, (3)
1
n
Type II, (1)
4
n−1
2
Type II, (2)
2
Type II, (3)
1
Type III, (1)
4
Type III, (2)
2(m − l − 2)
Type III, (3)
2
Type III, (4)
2
2
2·
n−1
3
(n − 1)(n − 2) + 2n − 2
2 · n−1
4
1≤l<m≤n
2 · n−1
3
2 · n−1
3
The only non-obvious case is Type III, case (2), which only applies if n ≥ 4.
Here we obtain the sum:
n−2
n
X X
2(m − l − 2).
l=1 m=l+2
8
E. FINK AND K. ZAINOULLINE
Since only the case k = m − l − 2 > 0 matters and each k occurs (n − k)-times, the
sum transforms into
(*)
n−2
X n−l−2
X
l=1
2k = 31 n3 − 2n2 +
11
3 n
− 2.
k=1
2
In view of the table above, it gives 31 n4 − 34 n3 + 11
3 n − 5n for n < 4. If n ≥ 4, then
4
2
we add (*) and the result becomes n3 − n3 + 8n3 − n + 13 .
Using the description of u.l.g.’s we can recover a result of Hart [Ha17]
3.9. Corollary. In a symmetric group Sn+1 there are n2 + 1 elements with a unique
geodesic, i.e. a uniquely reduced expression.
Proof. A uniquely reduced expression corresponds to a word of type I. Hence, there
are at most n words of length 1, n − 1 (decreasing) words of length 2 which are of
the form si si−1 . In general, there are n − l words of length l, all of which must have
the form sj sj−1 . . . sj−l+1 . In total this gives n(n+1)
words. Adding an inverse for
2
n(n+1)
2
each word of length ≥ 2 gives
· 2 − n + 1 = n + 1 unique geodesics.
2
3.10. Remark. Observe that in type A the property of being a u.l.g. can be also
interpreted using the language of rhombic tilings of Elnitsky [El97].
Following [El97] we say that two geodesics ρ1 (w) and ρ2 (w) are C1 -equivalent if
ρ2 (w) is obtained from ρ1 (w) by applying a finite number of commuting relations,
i.e., by commuting subsequent generators si sj with |i−j| > 1 in the reduced expression for w. A function ρ(w) → ρ
~(w) (which assigns to a geodesic its label) factors
through C1 -equivalence, hence, if w has a u.l.g. ρ(w), then the C1 -equivalence class
of ρ(w) must contain only one element (the geodesic ρ(w) itself). The latter means
that
A rhombic tiling of the 2(n−1)-polygon corresponding to the equivalence class of a u.l.g. ρ(w) must have a unique ordering.
We say that a tile touches a border strongly if it touches it with 2 sides and the
border is on the left from the tile. Then a tiling has a unique ordering iff it satisfies
the following property:
Any border except the rightmost one has exactly one tile that
touches it strongly (i.e. with two sides).
4. Simply laced trees
We will now investigate the case when the Coxeter-Dynkin diagram Γ describing
the Coxeter group W is no longer a chain as in the type A case, but a finite graph
where any two vertices are connected by at most one edge, i.e., W has the Coxeter
exponents mij = 2 or 3 only. More precisely, a vertex α ∈ Π corresponds to a
generator sα of W . If two vertices α, β are adjacent (connected by an edge), then
the generators satisfy the braid relation sα sβ sα = sβ sα sβ , otherwise the generators
sα and sβ commute.
UNIQUELY LABELLED GEODESICS OF COXETER GROUPS
9
We index elements of Π from 1 to n = |Π|, i.e., Π = {α1 , . . . , αn }. Consider the
Cayley graph C of W with respect to the generators sα1 , . . . , sαn and the generating
function
X
aı ti11 . . . tinn
UC (t1 , . . . , tn ) =
ı=(i1 ,...,in )
which counts the number of u.l.g.’s in the Cayley graph of W . Observe that by
reindexing Π we reindex the variables ti and the label coordinates of ı. Sometimes
we will write the elements of Π as subscripts meaning the respective indices, i.e.,
tαj = tj and iαj = ij .
By the same arguments as in the type A-case, we obtain the following generalizations of Lemma 3.1, Corollary 3.2 and Lemma 3.3:
Q
4.1. Lemma. All non-zero monomials of UC are of the form α∈Π′ tiαα , where Π′
is a connected subset of vertices in Γ and iα > 0.
4.2. Corollary. In a u.l.g. ρ(w) with l(w) ≥ 2 any two adjacent generators w =
. . . sα sβ . . . correspond to adjacent (connected by an edge) vertices α, β in Γ.
In other words, a u.l.g. is necessarily a path (with possible returns) on the
Coxeter-Dynkin diagram Γ.
4.3. Lemma. A u.l.g. can not contain a subword of the following form
sα sβ · v · sβ sα sβ
and
sβ sα sβ · v · sβ sα ,
where v does not contain generators adjacent to sα .
By Corollary 4.2 we restrict to study paths on Γ. A path in Γ is called a simple
path or a path with no returns, if every vertex on it occurs exactly once. By a
turning vertex ν of a path ρ(w) we call a vertex corresponding to a generator sν
such that sµ sν sµ is a subword of w: we go from µ to ν and then back to µ. Let
(ν1 , . . . , νr ) be a list of subsequent (following the direction of the path) turning
vertices of ρ(w) (r = 0 corresponds to the empty list). Let ν0 and νr+1 denote the
starting and the ending vertex of the path ρ(w).
sν2
sνr
✞C ✽✽✽
✠D ✼✼✼
✞
✠
✼✼
✽✽
✞
✠
sν0
✼
✹✹
✽✽
✞✞
✠✠
✞
···
✼
✠
✹✹
✼✼
✞
✠
✞
✠
s
νr+1
✹✹ ✠✠
✼✼ ✞✞
✹ ✠✠
✼ ✞✞
sν1
sνr−1
We now restrict to the case when Γ is a tree.
4.4. Lemma. Let ρ(w) be a path corresponding to a reduced word in W . Let
(ν0 , . . . , νr+1 ) be the list of turning vertices (incl. the starting and the end point).
(1) Then for all 0 ≤ i ≤ r the subpath [νi , νi+1 ] passes through νi and νi+1
exactly one time. In particular, νi 6= νi+1 for all i.
(2) Moreover, if ρ(w) is a u.l.g., then for all 1 ≤ i ≤ r the subpath [νi−2 , νi+2 ]
passes through νi exactly one time (here to simplify the notation we set
ν−1 = ν0 and νr+2 = νr+1 ).
Proof. By definition, the subpath [νi , νi+1 ] passes through the intermediate turning
vertex νi . Hence, it is enough to show that all other vertices in the subpath are
10
E. FINK AND K. ZAINOULLINE
different from νi . Since Γ is a tree, all the subpaths [νi , νi+1 ] are simple, so (1)
follows.
Let ρ(w) be a u.l.g. Suppose the subpath [νi−2 , νi+2 ] contains a second copy of
νi . Then by (1) it has to be either in [νi−2 , νi−1 ) or in (νi+1 , νi+2 ]. Suppose it is in
(νi+1 , νi+2 ]. Since Γ is a tree, any path of length ≥ 2 that starts and ends at νi and
does not go through νi has to go through the same adjacent to νi vertex µi . Hence,
ρ(w) contains a subword sµi sνi sµi vsµi sνi of Lemma 4.3, a contradiction.
4.5. Corollary. A uniquely labelled geodesic ρ(w) has to be necessarily of the following form
Type I. A simple path, i.e. each generator in ρ(w) occurs exactly once.
Type II. A path with a single turn, i.e. r = 1.
Type III. A path with r ≥ 2 turns such that for all 1 ≤ i ≤ r the subpath [νi−2 , νi+2 ]
passes through the turning vertex νi exactly one time.
Observe that if Γ is a chain, then the types I, II and III become the respective
types of the A-case, hence, they are also provide sufficient conditions for being a
u.l.g. In general, there are paths of type III which are not u.l.g’s.
4.6. Example. Consider the Weyl group E6 with the Coxeter-Dynkin diagram
α2
α1
α3
α4
α5
α6
Consider a reduced word w = s3 s1 s3 s4 s2 s4 s5 s4 s3 s1 . It corresponds to a path of
type III with r = 3 and turning vertices (α3 , α1 , α2 , α5 , α1 ), however, it contains
a subword of Lemma 4.3, hence, it is not a u.l.g. So the condition that a reduced
word ρ(w) has type III is not sufficient for being a u.l.g.
4.7. Example. Consider the Weyl group D4 that is
W = {s0 , s1 , s2 , s3 | s2k = 1, [si , sj ] = 1 and (si s0 )3 = 1 for i, j > 0}.
Its Dynkin diagram Γ is (here αi correspond to si )
α2
α1
α0
α3
Consider the reduced word ρ(w) = s0 s1 s0 s2 s0 s3 s0 s1 s0 . It corresponds to the
path of type III
ρ(w) : α0 → α1 → α0 → α2 → α0 → α3 → α0 → α1 → α0
with r = 4 and turning vertices (α1 , α2 , α3 , α1 ).
The reduced expression graph taken modulo C1 -equivalence classes (two representatives of C1 -equivalence classes are connected by a directed edge labelled by i
if the first class is obtained from the second by applying s0 si s0 → si s0 si ) for w is
UNIQUELY LABELLED GEODESICS OF COXETER GROUPS
11
given by
/ 102301303
102301030
3
f◆◆◆
@ ◆◆◆
◆◆
◆◆
✁✁
✁
✁
1◆◆
1
3◆◆
✁
◆◆◆
◆◆◆
✁
&
✁
✁
/
o
/ 120230103
101203010
101230310
102013031
102030103
3
3
1
2
♣8
❤❤❤
✁✁ ♣♣♣8
♣
❤
❤
♣
❤
✁
❤
✁ ♣♣
♣♣
❤❤❤
2
1
✁✁ ♣♣1
♣♣3
❤❤❤❤1
✁✁♣♣♣
♣♣♣
s❤❤❤❤❤
010203010
101203101 o
102010301
120213031 ❱
1
k ❱❱❱❱
❂❂◆◆◆
◆◆◆
O
O
❱❱❱❱❱
❂❂ ◆◆
◆◆
❂❂ 1◆◆◆
1❱❱❱
2◆◆
1
3
❱❱❱❱
◆
◆◆&
❂❂ ◆◆&
❱❱
/
/ 201023031
010203101
012023101
120210301 o
201020301
3
2
2❂
1
8
O
❂❂
♣♣
♣♣
♣
♣
❂❂
♣
♣
♣
♣
❂❂
1
1
♣♣2
❂ ♣♣♣♣♣
x♣♣♣
/ 202102301
020102301
012023010
2
010230310
Since there is only one reduced expression with 5 generators s0 , the reduced word
ρ(w) is a u.l.g. for the label (5, 2, 1, 1).
Observe that the word w = (s0 s1 s0 s2 s0 s3 )2 s0 is not reduced. Indeed, we have
0102030102030 →(101)2(303)1(202)30 → 1023(101)230230 → 1023010230230 →
1023010203020 → 1023012023202 → 10230120302.
5. Uniquely labelled geodesics with finite branching index
Let Γ be a simply laced tree. By the valency of a vertex in Γ we denote the
number of vertices adjacent to it. A branching vertex is a vertex of valency at least
3. An end vertex is a vertex of valency 1. A branch of Γ is a maximal connected
subchain of Γ where all vertices have valency less or equal than 2.
5.1. Definition. Let ρ(w) be a u.l.g. and let (ν1 , . . . , νr ) be the list of its turning
vertices (without the starting and the end points) so that ρ(w) contains subwords
sµi sνi sµi , 1 ≤ i ≤ r. If the adjacent vertex µi is a branching vertex, then the νi is
called a short turning vertex, otherwise it is called a long turning vertex.
We define a branching index of ρ(w) with respect to the tree Γ as the number
(repetitions are possible) of short turning vertices that is
ı(ρ(w)) = |{i ∈ {1, . . . , r} | µi is a branching vertex of Γ}|.
If ρ(w) does not have turning vertices, we set ı(ρ(w)) = 0.
As an immediate consequence of Corollary 4.5 we obtain
5.2. Lemma. Suppose a u.l.g. visits a branch on the tree Γ, i.e., goes in and out
via the branching vertex γ attached to the branch. Then it has exactly one turning
vertex inside that branch. In other words, each visit of a branch corresponds to a
turning vertex in that branch.
5.3. Lemma. Suppose a u.l.g. ρ(w) visits the same branch more than once. Let
(γ1 , . . . , γs ), s > 1 be the corresponding list of turning vertices on that branch
(observe that it is a sublist of the list of all turning vertices of ρ(w)). Let di denote
the distance between γi and the branching vertex γ (observe that a short turning
vertex has distance 1 and a long one has distance > 1).
12
E. FINK AND K. ZAINOULLINE
Then any subpath [γi , γj ], i 6= j must contain a turning vertex γk of distance
dk = max(1, min(di , dj ) − 1).
Proof. Assume this is not the case. Then there are two long turning vertices a and
b of distances da > 1 and db > 1 such that all turning vertices between a and b are
of distance < (min(da , db ) − 1). Hence, the subpath [a, b] is a subword of a word of
Lemma 4.3, a contradiction.
5.4. Corollary. If γi is a long turning vertex, then either γi−1 or γi+1 has to be a
short turning vertex.
Consider now a u.l.g. ρ(w) of maximal length with trivial branching index, i.e.,
ı(ρ(w)) = 0.
5.5. Lemma. The u.l.g. ρ(w) contains all vertices of the tree.
Proof. Assume ρ(w) does not contain a vertex v of the tree but it does contain a
vertex x adjacent to v. Because of the tree structure, removing v will divide the tree
into two disconnected sets A and B and now ρ(w) can only use vertices in either of
the two sets. Since ρ(w) goes through x at least once it has the form w1 xw2 and
we can extend the path by replacing w1 xw2 with . . . w1 xvxw2 . . .. Hence ρ(w) is
not of maximal length, a contradiction.
5.6. Corollary. All turning vertices of ρ(w) have valency 1. In particular, every
end vertex can only be visited once.
5.7. Corollary. Let (ν1 , . . . , νr ) be turning vertices of ρ(w) and let d(νi , νi+1 ) denote
P
the distance between νi and νi+1 . Then the sum S = r−1
i=1 d(νi , νi+1 ) is maximal.
5.8. Proposition. Suppose the Coxeter-Dynkin diagram of W is a simply-laced tree
with n vertices. The maximal length of a u.l.g. with ı(ρ(w)) = 0 is bounded above
2
by n2 + 25 n − 7.
Proof. The maximal distance between end vertices is n − 1 (which is achieved in
the case of a chain). Let Ve denote the set of end vertices of the tree. Since |Ve | ≤ n
we obtain 12 (n − 1)n as a bound for the maximal value of S. However we need to
consider the beginning and the end of a word which can lie on a maximal subchain.
For this we use Corollary 3.4 and add 3(n − 1) − 4 and, hence, obtain the desired
bound.
5.9. Theorem. Let Γ be an arbitrary simply-laced tree with n vertices. Then a
u.l.g. with finite branching index ı(ρ(w)) = B has length at most n2 (B + 1) + n · B.
In other words, any u.l.g. in Γ with finite branching index has finite length.
Proof. Assume the branching index of ρ(w) is ı(ρ(w)) = B. We argue using the
total number of turning vertices V . If a vertex v appears in the list V twice, then
its occurances have to be separated by a short turning vertex on the same branch.
Consecutive long turning vertices have to be on different branches. Hence, a long
turning vertex x occurs at most B + 1 times in the list V . So the list V has length
at most B + (the number of short turning vertices)(B + 1) ≤ B + n · (B + 1).
Between each pair of turning vertices we visit at most n vertices. Hence, in total
we obtain the bound n · (B + n · (B + 1)).
UNIQUELY LABELLED GEODESICS OF COXETER GROUPS
1
2
3
1
4
2
3
1
4
a
ab
b a
a b
b a
a b
b a
a b
b a
a b
b a
1
13
3 2
2 4
4 1
1 3
3 2
2 4
4 1
1 3
3 2
a
31
b
2 3
Figure 1.
a
4 2
b
1 4
a
b
3 1
a
2 3
4 2
b
a
1 4
b
3 1
13
2
3
2 3
4
a
a
b
b
2 4
4
b
4 2
The transformations of w2 to the word with reduced middle part (1b23a4)3 .
6. Infinite uniquely labelled geodesics
In the present section we construct an example of an infinite reduced word so
that any finite connected subword is a u.l.g. We do it for a simply laced tree.
Observe that one such example is produced in Example 2.3 which is not a tree
case.
Consider the infinite group D̃6 with Coxeter-Dynkin diagram as follows:
2
1
3
a
b
4
We first give a word which is not reduced:
6.1. Lemma. The word (a1a2ab3b4b)n is not reduced.
Proof. The sequence b4ba1a2ab3b4ba1a2ab3b4b can be reduced to a sequence of
smaller length as
b4ba1a2a · 3b3 · 4ba1a2ab3b4b → b4ba1a2a3b4 · 3ba1a2ab3b4b →
b4ba1a2a3b4 · b3b3 · a1a2ab3b4b → b4ba1a2a3b4 · b3b · a1a2a · 3b3b4b →
b4ba1a2a3 · 4b4 · 3ba1a2a · b34b → b4b4 · a1a2a3 · b43b · a1a2ab34b → 4ba1a2a3b4a1a2ab34b.
Let W be a Coxeter group generated by S. In [S09] the authors rephrase a result
from [FZ07] as
6.2. Lemma. If S = I ∪ J where all elements in I respectively
Q J commute pairwise,
Q
and if W is infinite and irreducible, then the word ( i∈I si j∈J sj )n is reduced for
all n.
We define I = {a, 3, 4} and J = {b, 1, 2} and show that there exists an infinite
word which is ’close’ to the one that is reduced by Lemma 6.2.
We will then show that for w = a1ab3ba2ab4b the word wn is a u.l.g.
6.3. Lemma. The word w2 can be transformed into
1a3 · 1b2 · 3a4 · 2b1 · 4a3 · 1b2 · 3a4 · 2b4,
where the subword 1b2 · 3a4 · 2b1 · 4a3 · 1b2 · 3a4 is reduced.
These word reductions are depicted in Figure 1 and show a part of the Cayley
graph of D̃6 .
Proof. The word w2 = 1a1 · 3b3 · 2a2 · 4b4 · 1a1 · 3b3 · 2a2 · 4b4 can be written as
1a · 31b · 23 · a · 42 · b · 14 · a · 31 · b · 23 · a4 · 2 · b4 → 1a3 · 1b2 · 3a4 · 2b1 · 4a3 · 1b2 · 3a4 · 2b4.
The subword 1b2 · 3a4 · 2b1 · 4a3 · 1b2 · 3a4 is now reduced by Corollary 6.2.
14
E. FINK AND K. ZAINOULLINE
12p − 1
......
12p − 6
length 12n − 6
Figure 2.
The word wn , each block represents wp as in Figure 1 where the red line
represents the reduced part.
6.4. Corollary. The word wn has form
wn = 1a3(1b23a4)2n−12b4
and, hence, it has length at least 12n − 12.
Proof. Note that w2 · w reduces to
1a3 · 1b2 · 3a4 · 2b1 · 4a3 · 1b2 · 3a4 · 2b4 · a1ab3ba2ab4b →
1a3 · 1b23a42b14a31b23a4 · 2b4 · 1a13b32a24b4 →
1a3 · 1b23a42b14a31b23a4 · 2b1 · 4a3 · 1b2 · 3a4 · 2b4 → 1a3 · (1b23a4)5 2b4.
In general, we see that wn reduces to
1a3 · 1b2 · 3a4 · 1b2 · 3a4 · 1b2 · 3a4 · 2b4 → 1a3(1b23a4)2n−1 2b4.
The estimate of the length comes from having 6(2n − 1) + 6 letters. The part of
length 6(2n − 1) is reduced and we allow for cancellation of the prefix and suffix,
with 3 letters each, cancelling at most 12 letters.
6.5. Proposition. The word wn = (a1ab3ba2ab4b)n is reduced for all n ≥ 2.
Proof. We assume it is not, then there is a power p of w such that wp is not
p
p
reduced. The word wp has 12p letters. Denote by w−
and w+
the beginning and
p
end of the path in the Cayley graph labelled by w . We assume first, there is a
p
p
shorter connection between w−
and w+
in the Cayley graph with at most 12p − 1
n
n
the beginning and end
and w+
letters. Now assume n ≥ 13 · p and denote by w−
n
of the path in the Cayley graph labelled by w . This is depicted in Figure (2).
n
n
. Denote by s− and s+ the
with w+
There are n/p copies of wp connecting w−
two vertices connecting the reduced middle part (1b23a4)2n−1 of length 12n − 6 of
wn . Because each part wp of wn is connected by a path of length 12p − 1, we get
a connection from s− to s+ of length at most
n
p
· (12p − 1) + 6 = 12n −
n
p
+ 6 ≤ 12n − 13 + 6 = 12n − 7.
However, by Corollary 6.2 the subword (1b23a4)2n−1 between s− and s+ has length
12n − 6.
6.6. Theorem. The word wn = (a1ab3ba2ab4b)n is a u.l.g.
Proof. We observe that asymptotically there are each one third a’s, one third b’s
1
th every number 1, 2, 3 and 4. We first show that the only way to have a
and 12
u.l.g. with this distribution is a periodic word. Assume the infinite word is not
periodic. Then the maximal distance between two occurances of one of the letters
1, 2, 3, 4 must be bigger than 12, because it is 12 on average. Assume without loss
of generality that it is the letter 1 which has two occurances in the infinite word
of distance more than 12. Because of the symmetries of the Coxeter tree defining
UNIQUELY LABELLED GEODESICS OF COXETER GROUPS
15
D̃6 the same arguments work for 2, 3 and 4. This can only happen in one of the
following cases. (Recall that a u.l.g. has to be a connected path in the graph.)
We go through all other possiblities which do not yield the word (a1ab3ba2ab4b)n.
Most of them either contain either xyxy, xyx . . . xyx or xy . . . yxy with no x in the
dots. The first two are not reduced words, the latter not a u.l.g. A complete list of
cases can be found in the Appendix 7.2. We discuss only the cases which are not
immediate:
(5)
(17)
(23)
(27)
(28)
(31)
(33)
a1a2ab4b
a1ab4ba2ab3ba1
a1ab4b3ba2ab4ba1a
a1ab4b3bab4b3b
a1ab4ba2ab3b4bab3b
a1ab4ba2ab3b4ba1a
a1ab4ba2ab3b4ba1ab3b
(36)
(52)
(59)
(60)
(62)
(68)
a1ab4ba2ab3b4ba1a2ab3b
a1ab4ba2ab3b4ba1a2aba1a2a
a1ab4b3ba2ab4b3ba1a2ab4b
a1ab4b3ba2ab4b3ba1a2aba1a
a1ab4b3ba2ab4b3ba1a2abab
a1ab4b3ba2ab4b3ba1ab4b
(5) Isomorphic to (6) and follows because of the isomorphism 3 ↔ 4.
(17) a1ab4ba2ab3ba1 in this case the letter 1 occurs with distance 12.
(23) a1ab4b3ba2ab4ba1a this is the complex case: We easily verify that neither
b4b, b3b nor ab can be a prefix. Hence the only case to check is a2a1a . . . :
This can only occur after aba2a1a . . . , which can only occur after either
a1 · aba2a1a . . . or a2 · aba2a1a . . . .
The word a2aba2a1a = 2a2 · ba2a1a = 2ab · 2a2a · 1a is not reduced. Hence we need
to check a1aba2a · a1a . . . :
a1aba2a1ab4b3ba2ab4ba1a = a1ab · 2a2 · 1ab4b3ba2ab4ba1a
a1ab · 2a1 · 2a · b4b3ba2ab4ba1a = a1ab · 2a1 · a2a2 · b4b3ba2ab4ba1a
a1ab2a1a2ab4b3b · 2a2a · b4ba1a = 1a1 · b2a1a2ab4b3ba2b4ba1a
1a · b2 · 1a1a · 2ab4b3ba2b4ba1a = 1ab2 · a1 · 2ab4b3ba2b4ba1a.
The first word has length 23 the last word has length 21, hence it is not reduced
and cannot occur as part of a u.l.g.
(27)
a1ab4b3bab4b3b = a1ab4b3 · aba · 4b3b = a1a · 4b4 · 3aba4b3b
a1a4b3a · 4b4 · ab3b = a1a4b3a · b4b · ab3b
a14aba3 · b4b · ab3b = a14 · bab · 3b4bab3b.
Both the first and last term have 3 times a, six times b, once 1 and twice each 3
and 4. Hence it cannot be part of a u.l.g.
(28) This word is not reduced:
a1ab4ba2ab3b4bab3b = a1ab4ba2a · 3b3 · 4bab3b
a1ab4ba2a · 3b · 4 · 3b · ab3b = a1ab4ba2a · b3b3 · 4 · 3b · ab3b
a1ab4ba2a · 3b4 · b3b3 · ab3b = a1ab4ba2a · 3b4 · b3b · a · 3b3b
a1ab4ba2a · 3b4b3b · a · b3 = a1ab4ba2a3 · 4b4 · 3bab3
= a1a · b4b4 · a2a3b43bab3.
16
E. FINK AND K. ZAINOULLINE
(31) This is the inverse of 23. Hence it cannot be succeeded by anything valid.
(33) This word is not reduced:
a1ab4ba2ab3b4ba1ab3b = a1ab4ba2a · 3b3 · 4ba1ab3b
a1ab4ba2a3b · 4 · b3b3 · a1ab3b = a1ab4ba2a3b4 · b3b · a1a · 3b3b
a1ab4ba2a3 · 4b4 · 3ba1a · b3 = a1ab4b4 · a2a3 · b4 · 3ba1ab3
= a1a · 4b · a2a3b43ba1ab3.
(36) This word is not reduced:
a1ab4ba2ab3b4ba1a2ab3b = a1ab4ba2a · 3b3 · 4ba1a2ab3b
a1ab4ba2a · 3b4 · b3b3 · a1a2ab3b = a1ab4ba2a · 3 · 4b4 · 3ba1a2a · b3
a1a · b4b4 · a2a34b43ba1a2ab3 = a1a · 4b · a2a34b43ba1a2ab3.
(52) We study the end of the sequence a2ab3b4ba1a2aba1a2a and show it is not
reduced:
a2ab3b4ba1a2aba1a2a = a2ab3b4ba1 · 2a2 · ba1a2a
a2ab3b4ba · 21 · a2ba1a2a = a2ab3b4b · 2a2a · 1a2b · 1a1 · 2a
a2a · 2 · b3b4ba2 · 1a1 · 2b · 1a1 · 2a = 2a · b3b4ba2 · 1a · 2b · a1 · 2a.
(59) This word is not reduced:
a1ab4b3ba2ab4b3ba1a2ab4b = a1ab4b3ba2ab4 · 3b3 · a1a2ab4b
a1ab4b3ba2a · 3b3b · 4b3a1a2ab4b = a1ab4 · 3b · a2ab3 · 4b4 · 3a1a2ab4b
a1ab43ba2ab34b · 3a1a2a · 4b4b.
(60) This word is not reduced:
a1ab4b3ba2ab4b3ba1a2aba1a = a1ab4b3ba2ab4b3ba1 · 2a2 · ba1a
a1ab4b3ba2ab4b3b · 2a2a · 1 · a2 · ba1a = a1ab4b3b · a2a2 · b4b3ba2a1a2ba1a
a1ab4b3b · 2a · b4b3ba2 · 1a1 · 2ba1a = a1ab4b3b2ab4b3ba21a2b · 1a1a.
(62) The same as in 59, since the extra a1a2 in (59) has no effect on the technique.
(68) a1ab4b3ba2ab4b3ba1ab4b is not reduced by the same argument as in (59) as
well.
It is left to show that the above word does not transform into the one under the
isomorphisms 3 7→ 4 or 1 7→ 2.
Case 1: We look at 3 7→ 4 first:
w2 = a1ab3ba2ab4ba1ab3ba2ab4b = 1a1 · 3b3 · 2a2 · 4b4 · 1a1 · 3b3 · 2a2 · 4b4
= 1a31b23a42b14a31b23a42b4 = 1a32b13a41b23a42b13a42b4
= 1a32b31a14b42a23b31a24b4 = 1a23b3a1ab4ba2ab3b1a24b4.
Even though the last transformation of w2 contains the word w2 = a1ab4ba2ab3b,
the letter 1 occurs more often than in w2 . Further, the last line is not a u.l.g. which
implies that wn is a u.l.g.
UNIQUELY LABELLED GEODESICS OF COXETER GROUPS
a
1
ab
1 a
3
14
41
3
42
a
a
3
2
ab
2 a
3
b
23
42
13
ba
4
4
ba
13
ab
Figure 3.
ab
3
42
a
b
2
1
3 1 a 14
b
b
13
42
42
a
b
ab
1
ba
ba
3
ba
42
13
ab
4
42
a
ba
4
ab
a 23 b
42
13
b
1
2
ba
ab
ab
3
ba
2
ab
3 1 a 14
b
13
42
42 a
13
23
42
13
13
42
42
a
b
2
1
17
ba
3
ab
1
a
b
ba
4
4
b
b
3
1
1
b
ab
2
b
3
n
The geodesic wn on the top, the geodesic w2
on the bottom.
Case 2: We now check the isomorphism 1 7→ 2.
w2 = a1ab3ba2ab4b · a1ab3ba2ab4b = 1a1 · 3b3 · 2a2 · 4b4 · 1a1 · 3b3 · 2a2 · 4b4
= 1a31b23a42b14a31b23a42b4 = 1a3 · 2b1 · 3a4 · 1b2 · 3a4 · 2b1 · 3a42b4
= 1a3 · b41 · 2a2 · 3b3 · 1a1 · 4b4 · 2a2 · 3b4 = 1a3b41 · a2ab3ba1ab4b · 2a23b4.
6.7. Corollary. The words w2n = (a1ab4ba2ab3b)n and w3n = (a2ab3ba1ab4b)n are
u.l.g.’s for all n ≥ 1.
We finish with an observation regarding the two bi-infinite geodesic rays of wn
and w2n in the Cayley graph of W . We say two geodesics γ1 and γ2 are fellow
travelling if there exists a constant D such that every point on γ2 is at distance at
most D from a point on γ1 .
6.8. Theorem. The bi-infinite u.l.g.’s with labels
(1) wn = (a1ab3ba2ab4b)n
(2) w2n = (a1ab4ba2ab3b)n
(3) and w3n = (a2ab3ba1ab4b)n
are fellow travelling at distance at most 5 but at least 2 from each other on their
entire lengths.
Proof. We can reduce w2 = a1ab3ba2ab4b · a1ab3ba2ab4b as follows:
w2 = 1a1 · 3b3 · 2a2 · 4b4 · 1a1 · 3b3 · 2a2 · 4b4
= 1a31b23a42b14a31b23a42b4 = 1a32b13a41b23a42b13a42b4
= 1a32b31a14b42a23b31a24b4 = 1a23b3a1ab4ba2ab3b1a24b4.
We note that a cyclic permutation of the last line gives w22 and we applied 6
relations to get to the last line. However, as it can be verified in Figure 3, the
shortest distance actually remains uniformly bounded above by 5.
We note that this method is independent of how many copies of w we had in the
beginning. In a similar fashion it can be seen that we can obtain w3 = a2ab3ba1ab4b.
Hence we have a bundle of three bi-infinite geodesics which remain at distance at
most 5 from each other.
7. Appendix
7.1. Type A examples. We list the u.l.g.’s with different labels for the groups of
type An for n = 3, 4, 5 by enumerating all total labels which give a u.l.g. including
18
E. FINK AND K. ZAINOULLINE
the word of length 0, which gives the constant term in the generating function. It
can be verified that these numbers correspond to the formula in Corollary 3.7.
Type
Labels of u.l.g.’s
A3
100 110 120 122 010 011 012 221 001 111 021 131 121
210 000
A4
Type I: 1000 0100 0010 0001 1100 0110 0011 0111 1110
Total
number
15
33
1111 0000
Type II and III: 1200 0120 0012 1210 0121 1220 0122
0021 2210 0221 2221 1321 1231 1331 1310 0131 1211 1121
1222 1221 0210 2100
A5
Type I: 10000 01000 00100 00010 00001 11000 01100 00110
66
00011 11100 01110 00111 11110 01111 11111 00000
Type II: 12000 01200 00120 00012 21000 02100 00210 00021
01220 00122 22100 02210 00221 12220 01222 22210 02221
22221 12100 01210 00121 12110 01211 11210 01121 12111
12210 01221 12211 11221 12221 12200 12222 11121
Type III: 12121 12321 12310 13210 13100 01310 00131
01231 12331 13321 12231 13221 13310 01331 13331 01321
7.2. Infinite u.l.g.’s in the group D̃6 . We list all cases that have to be considered
in Theorem 6.6. Figure 4 depicts the first part of these words, and the corresponding
end-vertices of the trees labelled with T 2, T 3 and T 4 indicate that the tree continues
with Figure 5, 6 or 7.
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
(11)
(12)
(13)
(14)
(15)
(16)
(17)
(18)
(19)
(20)
a1a2a1a
a1a2a2
a1a2aba2a
a1a2abab
a1a2ab4b
a1a2ab3b3b
a1a2ab3b4ba2a
a1a2ab3b4b3b
a1a2ab3bab
a1a2ab3ba2a
a1aba2ab
a1aba2a2
a1ab3bab
a1ab3b4b
a1ab4bab
a1ab4ba2ab4b
a1ab4ba2ab3ba1
a1ab4b3b4b
a1ab4b3ba2aba
a1ab4b3ba2ab3b
(21)
(22)
(23)
(24)
(25)
(26)
(27)
(28)
(29)
(30)
(31)
(32)
(33)
(34)
(35)
(36)
(37)
(38)
(39)
(40)
a1ab4b3ba2ab4b3b3
a1ab4b3ba2ab4ba2a
a1ab4b3ba2ab4ba1a
a1ab4b3baba
a1ab4b3bab3b
a1ab4b3bab4ba
a1ab4b3bab4b3b
a1ab4ba2ab3b4bab3b
a1ab4ba2ab3b4bab4b
a1ab4ba2ab3b4baba
a1ab4ba2ab3b4ba1a
a1ab4ba2ab3b4ba2a
a1ab4ba2ab3b4ba1ab3b
a1ab4ba2ab3b4ba1ab4b
a1ab4ba2ab3b4ba1aba
a1ab4ba2ab3b4ba1a2ab3b
a1ab4ba2ab3b4ba1a2ab4b
a1ab4ba2ab3b4ba1a1
a1ab4ba2ab3b4b3b
a1ab4ba2ab3b4b4
UNIQUELY LABELLED GEODESICS OF COXETER GROUPS
(41)
(42)
(43)
(44)
(45)
(46)
(47)
(48)
(49)
(50)
(51)
(52)
(53)
(54)
a1ab4b4
a1ab4b3b3
a1ab4b3ba2a2
a1ab4b3bab4b4
a1ab4ba2a2
a1ab4ba2ab3b3
a1ab4b3ba2ab4b4
a1ab4b3ba2ab4bab
a1ab4ba2ab3b4ba1a2abab
a1ab4ba2ab3b4ba1a2aba2a
a1ab4ba2ab3b4ba1a2aba1ab
a1ab4ba2ab3b4ba1a2aba1a2a
a1ab4ba2ab3b4ba1a2aba1a1a
a1ab4b3ba2ab4b3b4b
(55)
(56)
(57)
(58)
(59)
(60)
(61)
(62)
(63)
(64)
(65)
(66)
(67)
(68)
19
a1ab4b3ba2ab4b3ba2a
a1ab4b3ba2ab4b3ba1aba
a1ab4b3ba2ab4b3ba1a1a
a1ab4b3ba2ab4b3ba1a2ab3b
a1ab4b3ba2ab4b3ba1a2ab4b
a1ab4b3ba2ab4b3ba1a2aba1a
a1ab4b3ba2ab4b3ba1a2aba2a
a1ab4b3ba2ab4b3ba1a2abab
a1ab4b3ba2ab4b3baba
a1ab4b3ba2ab4b3bab4b
a1ab4b3ba2ab4b3bab3b
a1ab4ba2aba
a1ab4b3ba2ab4b3ba1ab3b
a1ab4b3ba2ab4b3ba1ab4b
Checking the cases of the tree:
• The following cases have been shown in the proof of Theorem 6.6:
{(5), (17), (28), (31), (33), (36), (52), (59), (60), (62), (68)}.
• The following cases contain a sequence xyxy and are hence not reduced:
{2, 4, 6, 12, 21, 24, 30, 38, 40, 41, 42, 43, 44, 45, 46, 47, 49, 53, 62, 63}.
• The following cases contain xyx . . . xyx where the dots do not contain x or
y. These cases are not reduced:
{1, 3, 7, 8, 10, 16, 18, 20, 22, 25, 29, 32, 34, 39, 50, 54, 55, 57, 58, 61, 65, 67}.
• The following cases contain xy . . . yxy and are hence not a u.l.g.:
{9, 11, 13, 15, 19, 26, 35, 37, 48, 51, 56, 66}.
20
E. FINK AND K. ZAINOULLINE
1
a
b
2
3
a
4
b
4
a
T2
b
a
b
b
1
2
2
3
a
a
14
2
13
a
4
b
1
2
b
a
3
b
11
2
12
4
2
b
6
b
a
4
b
5
9
a
3
a
b
2
10
3
a
b
7
8
Figure 4. Tree 1
4
a
3
2
b
41
b
15
a
a
3
4
b
42
2
b
4
2
b
a
a
66
4
3
3
a
18
43
b
b
24
16
a
b
b
3
25
a
27
4
a
1
3
b
19
b
b
a
4
a
20
3
1
3
a
17
4
2
47
b
39
T3
a
40
a
T4
22
b
26
4
23
Figure 5. Tree 2
UNIQUELY LABELLED GEODESICS OF COXETER GROUPS
21
1
b
4
2
b
a
a
3
b
1
29
a
2
a
1
a
28
38
32
30
31
b
a
a
3
35
4
b
b
4
b
3
a
34
b
33
b
2
1
36
b
a
49
a
50
1
b
2
51
a
52
Figure 6. Tree 3
a
53
37
22
E. FINK AND K. ZAINOULLINE
3
b
a
4
b
1
21
2
b
54
3
4
a
a
a
63
55
b
1
b
65
a
64
3
4
2
57
a
56
b
b
b
67
68
a
4
3
b
b
b
1
2
59
62
58
a
a
60
61
Figure 7. Tree 4
References
[AC13] Y. Antolin, L. Ciobanu, Geodesic growth in right-angled and even Coxeter groups, European Journal of Combinatorics 34 (2013), 859–874.
[A04] A. Avasjö, Automata and growth functions of Coxeter groups, Licentiate Thesis Department of Mathematics, KTH, 2004.
[BB] A. Björner, F. Brenti, Combinatorics of Coxeter Groups, GTM 231, Springer, 2005.
[BH93] B. Brink and R. Howlett, A finiteness property and an automatic structure for Coxeter
groups, Math. Ann. 296, (1993), 179–190.
[CD91] R. Charney, M. Davis, Reciprocity of growth functions of Coxeter groups, Geometriae
Dedicata 39 (1991), Issue 3, 373–378.
[CK16] L. Ciobanu, A. Kolpakov, Geodesic growth of right-angled Coxeter groups based on trees,
J. Algebr. Comb. 44 (2016), 249–264.
[Ed95] P. Edelman, Lexicographically first reduced words, Discrete Math. 147 (1995), no.1-3,
95–106.
[El97] S. Elnitsky, Rhombic tilings of polygons and classes of reduced words in Coxeter groups,
J. Combin. Theory Ser. A 77 (1997), no. 2, 193221.
[EE10] H. Eriksson, K. Eriksson, Words with intervening neighbours in infinite Coxeter groups
are reduced, The Electronic Journal of Combinatorics 17 (2010), no 9.
[FZ07] S. Fomin, A. Zelevinsky, Cluster algebras. IV. Coefficients, Compositio Math. 143 (2007),
no. 1, 112–164.
[GN97] R. I. Grigorchuk, T. Nagnibeda, Complete growth functions of hyperbolic groups, Invent.
Math. 130 (1997), 159–188.
[Ha17] S. Hart, How many elements of a Coxeter group have a unique reduced expression?
Preprint arXiv:1701.01138 (2017).
UNIQUELY LABELLED GEODESICS OF COXETER GROUPS
23
[He94] P. Headley, Reduced expressions in infinite Coxeter groups, PhD Thesis, University of
Michigan, 1994.
[HNW] C. Hohlweg, P. Nadeau and N. Williams, Automata, reduced words, and Garside shadows
in Coxeter groups, Journal of Algebra 457 (2016), 331–456.
[Hu90] J. Humphreys, Reflection groups and Coxeter groups. Cambridge studies in Advanced
Math. 29, Cambridge Univ. Press, 1990.
[LP13] T. Lam, P. Pylyavskyy, Total positivity for loop groups II: Chevalley generators, Transformation Groups 18 (2013), no. 1, 179–231.
[M03] M. J. Mamaghani, Complete growth series of Coxeter Groups with more than three generators, Bulletin of the Iranian Mathematical Society Vol. 29 No. 1 (2003), 65–76.
[MT13] M. Mihalik, S. Tschantz, Geodesically tracking quasi-geodesic paths for Coxeter groups,
Bull. London Math. Soc. 45 (2013), 700–714.
[P90] L. Paris, Growth series of Coxeter groups, Group theory from a geometrical viewpoint
(Trieste, 1990), 302–310, World Sci. Publ., River Edge, NJ, 1991.
[S09] D. Speyer, Powers of Coxeter elements in infinite groups are reduced, Proceedings of the
American Math. Soc. 137 (2009), no. 4, 1295–1302.
[St84] R. Stanley, On the number of Reduced Decompositions of Elements of Coxeter groups,
Europ. J. Combinatorics 5 (1984), 359–372.
[St97] J. Stembridge, Some combinatorial aspects of reduced words in finite Coxeter groups,
Transactions of the American Math. Soc. 349 (1997), no. 4, 1285–1332.
(Elisabeth Fink) Department of Mathematics and Statistics, University of Ottawa,
585 King Edward Street, Ottawa, ON, K1N 6N5, Canada
E-mail address: efink@uottawa.ca
URL: http://mysite.science.uottawa.ca/efink/
(Kirill Zainoulline) Department of Mathematics and Statistics, University of Ottawa,
585 King Edward Street, Ottawa, ON, K1N 6N5, Canada
E-mail address: kirill@uottawa.ca
URL: http://mysite.science.uottawa.ca/kzaynull/
| 4math.GR |
arXiv:1705.04200v1 [math.GR] 11 May 2017
ON BENJAMINI–SCHRAMM LIMITS OF CONGRUENCE
SUBGROUPS
ARIE LEVIT
Abstract. A sequence of orbifolds corresponding to pairwise non-conjugate
congruence lattices in a higher rank semisimple group over zero characteristic
local fields is Benjamini–Schramm convergent to the universal cover.
1. Introduction
A semisimple analytic group G is defined as follows. Let I be a finite index set.
Assume that ki is a zero characteristic local field and Gi is a connected simplyconnected ki -isotropic almost ki -simple linear ki -group for every i ∈ I. Denote
Gi = Gi (ki ) so that in particular Gi is an almost
simple non-compact linear group
Q
admitting a ki -analytic structure. Let G = i∈I Gi .
Definition. A sequence of lattices (Γn )n∈N in G weakly central1 if for every compact
subset Q ⊂ G we have that
n→∞
Pr {gΓn ∈ G/Γn : gΓn g −1 ∩ Q ⊂ Z(G)} −−−−→ 1
This note is dedicated to establishing the following result.
Theorem 1. Assume that |I| ≥ 2. Then every sequence of pairwise non-conjugate
congruence lattices in G is weakly central.
Recall that every irreducible lattice in G is arithmetic whenever |I| ≥ 2. A
congruence lattice is a particular kind of an irreducible arithmetic lattice strictly
containing a congruence subgroup. See §3 for a precise definition of this notion.
In particular whenever lattices in G are known to satisfy the congruence subgroup
property a stronger formulation of Theorem 1 is possible.
We remark that if |I| = 1 and rank(G) ≥ 2 then every sequence of pairwise nonconjugate lattices is weakly central by [1, 12]. The recent works of Raimbault [20]
and Fraczyk [9] establish closely related results for congruence lattices in the rank
one groups SL2 (R) and SL2 (C). We also mention [1, §5] dealing with congruence
subgroups in a fixed uniform arithmetic lattice.
Convergence of Plancherel measures. Let ν G denote the Plancherel measure
b of G. For every uniform lattice Γ in G, the quasi-regular
on the unitary dual G
representation ρΓ of G in L2 (G/Γ) decomposes as a direct sum of irreducible repb appears in ρΓ with finite
resentations. Every irreducible representation π ∈ G
1Of course, this definition makes sense for any locally compact group. In the case that G is
totally disconnected a weakly central sequence is called a Farber sequence.
1
2
ARIE LEVIT
multiplicity m(π, Γ). The corresponding relative Plancherel measure is
X
1
m(π, Γ)δπ
νΓ =
vol(G/Γ)
b
π∈G
Combining Theorem 1 with [1, 1.2] and [12, 1.3] we obtain a generalization of
one of the main results of [1, 12] on convergence of relative Plancherel measures.
Corollary 2. Assume that |I| ≥ 2. Let Γn be any sequence of pairwise nonn→∞
conjugate uniformly discrete torsion-free congruence lattices in G. Then νΓn (E) −−−−→
b
ν G (E) for every relatively quasi-compact ν G -regular subset E ⊂ G.
The applications of Corollary 2 to limit multiplicities formulas and normalized
Betti numbers [1, 1.3,1.4] carry over to our setting as well.
Benjamini–Schramm convergence. The semisimple analytic group G is acting
by isometries on a contractible non-positively curved metric space X, as follows.
Let Xi be the symmetric space or Bruhat-Tits building associated
Q to Gi for every
i ∈ I, depending on whether ki is Archimedean or not. Take X = i∈I Xi equipped
with the product metric.
The following geometric notion is equivalent to saying that a sequence of lattices
is weakly central, as explained in [12, §3].
Definition. Let (Γn ) be a sequence of lattices in G. The orbifolds Γn \X Benjamini–
Schramm converge to X if for every radius 0 < R < ∞ the probability that an R-ball
with base point taken uniformly at random is contractible tends to one as n → ∞.
As an example, we provide a geometric application of Theorem 1 to arithmetic
orbifolds, relying on the congruence subgroup property [22].
Corollary 3. Let F be a number field with ring of integers OF . Assume that F
has r real embeddings and 2s complex embeddings with r + s ≥ 2. Consider the
irreducible arithmetic lattice
r
s
Y
Y
SL2 (OF ) ֒→
SL2 (R) ×
SL2 (C)
i=1
i=1
Let X = (H2 )r × (H3 )s be a product of two and three dimensional hyperbolic spaces.
Then the orbifolds corresponding to any sequence of distinct finite-index subgroups
of SL2 (OF ) are Benjamini-Schramm convergent to X.
On properties (T ) and (τ ). Benjamini-Schramm convergence of lattices was first
investigated in [1], where it was shown that any sequence of pairwise non-conjugate
irreducible lattices in a semisimple Lie group with high rank and property (T ) is
weakly central. General local fields were dealt with in [12]. These proofs rely on
property (T ), most crucially in order to invoke the Stuck–Zimmer theorem [14, 23].
Our approach is to make use of property (τ ) instead, avoiding the Stuck–Zimmer
theorem which is presently unknown in the absence of property (T ). More precisely,
we rely on property (τ ) with respect to congruence lattices. This is sometimes called
the Selberg property as it generalizes his famous theorem on congruence subgroups
of the modular group. It is crucial that Selberg’s property is, perhaps surprisingly,
uniform with respect to all the congruence lattices inside G — see Theorem A.
In addition, we rely on topological properties of the Chabauty space of semisimple analytic groups recently established by Gelander and the author [12], thereby
replacing yet another argument of [1] which previously required property (T ).
ON BENJAMINI–SCHRAMM LIMITS OF CONGRUENCE SUBGROUPS
3
Spectral gap and essentially free actions. Towards proving Theorem 1 we
study a Borel G-space obtained by taking a certain limit with respect to a sequence
of congruence lattices. Selberg’s property implies that this limiting G-space has
spectral gap — see §2 for a discussion of this notion. The following theorem allows
us to deduce that such an action is essentially free, provided that |I| ≥ 2.
Theorem 4. Let G be a product of at least two locally compact simple groups and
X a Borel G-space admitting an invariant probability measure µ. Assume that X
is properly ergodic, irreducible and has spectral gap. Then (X, µ) is essentially free.
We do not claim any originality for Theorem 4 — it is a formal corollary of the
well-known work of Bader–Shalom [2] and the fact that an action with spectral gap
is not weakly amenable [8]. We state it here merely as an observation and in the
hope that it may prove useful in other situations as well.
Notice that for Theorem 4 to be applicable in our situation the limiting Gspace has to be irreducible. Indeed, this follows from a key feature of Theorem A
establishing spectral gap with respect to each simple factor of G individually.
Acknowledgements. I would like to thank Tsachik Gelander for his guidance
and many useful discussions about the ideas of [1]. I would like to thank Nicolas
Bergeron for bringing to my attention the important fact that Clozel’s theorem [6]
holds uniformly with respect to all congruence lattices. I would like to thank Uri
Bader and Alex Lubotzky for several useful discussions and remarks.
2. Uniform spectral gap
Let G be a compactly generated locally compact group and (X, µ) a Borel Gspace with an invariant probability measure. Recall that the G-action on X has
spectral gap if the Koopman representation of G in the space L20 (X, µ) of functions
with zero integral does not almost admit invariant vectors.
Definition. A sequence (Xn , µn ) of Borel G-spaces with invariant probability measures has uniform spectral gap if the natural representation of G on ⊕n L20 (Xn , µn )
does not almost admit invariant vectors.
For example, if G has property (T ) then any family of ergodic G-invariant probability measures has uniform spectral gap. More generally, such uniformity is useful
when passing to weak-∗ limits of probability measures on a given compact G-space.
Proposition 5. Let X be a compact G-space and µn a sequence of invariant Borel
probability measures on X with uniform spectral gap. If µ is a weak-∗ limit of the
sequence µn then µ has spectral gap.
Let π and πn denote the Koopman representations on the Hilbert spaces L2 (X, µ)
and L2 (X, µn ) for n ∈ N. Let k·k and k·kn denote the norms on these spaces.
Proof of Proposition 5. Let ν be probability measure on G which is symmetric,
absolutely continuous with respect to the Haar measure and such that supp(ν ∗ ν)
is a generating set for G. Uniform spectral gap for the measures µn means that
kπn (ν)|L20 (X,µn ) kn < β
for some constant 0 < β < 1 and all n ∈ N. This fact is established in [3, G.4.2].
We claim that kπ(ν)|L20 (X,µ) k < β as well. In estimating the norm of a continuous
operator we may restrict attention to a dense subspace. Consider any non-zero
4
ARIE LEVIT
continuous function f ∈ C(X) ∩ L20 (X, µ). Note that π(ν)f ∈ C(X) ∩ L20 (X, µ) as
well. Denote
Z
an =
f dµn ,
f = fn′ + an 1X
so that fn′ ∈ C(X) ∩ L20 (X, µn ) for all n ∈ N.
R
Since µ is a weak-∗ limit of the measures µn we have that limn an = f dµ = 0.
To estimate the operator norm of π(ν) on the space L20 (X, µ) we calculate
kπ(ν)f k = limkπ(ν)f kn ≤ lim supkπ(ν)fn′ kn + lim|an | <
n
n
n
< β lim supkfn′ kn ≤ β limkf kn + lim|an | = βkf k
n
n
n
Therefore the G-space (X, µ) has spectral gap as well [3, G.4.2].
Assume that G splits as a direct product G = G1 × · · · × Gn of n factors. It is
natural to consider the restriction of the G-action to each factor Gi individually.
Definition. (X, µ) has strong spectral gap if the restricted action of each factor
has spectral gap. A sequence (Xn , µn ) of Borel probability G-spaces has strong
uniform spectral gap if these restricted actions have uniform spectral gap.
Recall that a G-action is irreducible if each factor Gi is acting ergodically. As
spectral gap clearly implies ergodicity, strong spectral gap implies irreducibility.
3. Congruence lattices and the Selberg property
Let G be semisimple analytic group, so that G is a direct product of Gi = Gi (ki )
where each ki is a local field and i ranges over a finite index set I.
Congruence lattices. Let F be an algebraic number field and H an absolutely
simple linear F -group. Let R ⊂ VF denote these infinite valuations such that H(Fv )
is non-compact for v ∈ R. Assume that there is a finite set of valuations S with
R ⊂ S ⊂ VF and a bijection ι : I → S so that ki ∼
= Fι(a)
Q and Gi is ki -isomorphic to
H for all i ∈ I. In particular we may identify G with v∈S H(Fv ).
The group H(F (S)) is an irreducible lattice in G. Given a non-zero ideal a in the
ring F (S) let H(a) denote the kernel of the natural map H(F (S)) → H(F (S)/a).
Definition. A congruence lattice is any lattice in G containing some H(a) as above.
The following is essentially a reformulation of the well-known Selberg’s property.
Theorem A (Selberg’s property). Let G be a semisimple analytic group. Then the
family of G-spaces G/Γ with normalized probability measures and Γ ranging over
the congruence lattices in G has strong uniform spectral gap.
We begin our discussion of Theorem A with a few preliminary remarks.
• The two G-representations L2 (X, µ) and L2 (X, αµ) are equivalent for every
α > 0, so that renormalizing a finite measure on a Borel G-space has no
effect on spectral gap.
• Similarly, the two G-representations L2 (G/Γ) and L2 (G/Γg ) are equivalent
for every g ∈ G.
• For a pair of lattices Γ, Γ′ with Γ ≤ Γ′ the G-representation L2 (G/Γ′ ) is
contained in L2 (G/Γ).
ON BENJAMINI–SCHRAMM LIMITS OF CONGRUENCE SUBGROUPS
5
In light of theseQremarks we may restrict our attention to the situation of the
lattice H(a) inside v∈S H(Fv ) while making sure that the resulting strong spectral
gap is independent of the field F , the group H and the ideal a.
On Selberg’s property. The existence of spectral gap for congruence subgroups
3
of SL2 (Z) regarded as lattices in SL2 (R) is essentially Selberg’s classical λ1 ≥ 16
2
theorem [21]. The Archimedean case where H is still SL2 , F is any number field
and S consists of infinite places is treated in [24]. The remaining case of H = SL2
and S an arbitrary set of places follows from the work of Gelbart–Jacquet [13].
The Burger–Sarnak method [5] allows to go beyond SL2 . This method was
extended by Clozel–Ullmo [7], in particular covering the p-adic case. We refer the
reader to the useful discussion on [16, §4.2].
Theorem B (Burger–Sarnak, Clozel–Ullmo). Let H1 be a semi-simple F -subgroup
of H. Then for every valuation v ∈ S the restriction of L2 (G/H(a)) to H1 (Fv ) is
weakly contained in
!
Y
M
′
2
H1 (Fv )/H1 (a )
L
a′ ⊳F (S)
v∈S
where the direct sum is taken over all non-zero ideals in F (S).
Clozel made the final contribution towards Selberg’s property by dealing with
arbitrary absolutely simple groups. In fact Theorem A is essentially equivalent to
[6, Thm. 3.1]. Clozel’s proof for a general F -group H depends on whether it is F isotropic or not. If rankF (H) ≥ 1 then H is known to admit SL2 as a F -subgroup
[17, I.1.6.3]. In that case one may rely on [13] and Theorem B.
The main effort of [6] is in dealing with the anisotropic case, as follows. If H is
F -anisotropic then it admits a F -subgroup H1 with rankFv (H1 ) = 1 and such that
H1 ∼
= SU(D, ∗)
= SL1 (D) or H1 ∼
where D is a division algebra of degree p2 over F or over a quadratic extension of F
in the first and second cases respectively, and p is a prime [6, 1.1]. A careful analysis
[6, §3.2] reveals that the case of SL1 (D) can be reduced to SL2 . Similarly, the case
of SU(D, ∗) is reduced either to SL2 , SU(3, Fv ) with v finite or SU(n, 1) with v
infinite. The parameter n is clearly bounded as H1 (Fv ) embeds in G. Clozel then
establishes spectral gap directly for congruence lattices in these last two families of
rank one groups.
Uniformity of spectral gap. Note that while [6, Theorem 3.1] is stated with
respect to a fixed algebraic number field F and group H, the resulting spectral gap
for the subgroup H1 (Fv ) is independent of any such choices.
To conclude the discussion, observe that there are only finitely many possibilities
for the group H1 (Fv ) and that the validity of Theorem A for these implies the same
for G. We need to take into account the fact that H1 (Fv ) regarded as a subgroup
of G depends on the chosen F -structure. The argument relies on the preliminary
remarks made above and on Lemma 6.
Lemma 6. Let H1 be an F -group and v ∈ S a valuation such that rankFv (H1 ) = 1.
Denote H1 = H1 (Fv ) and let Q ⊂ H1 be a compact subset. Then there is a compact
2Note that this case already suffices for the purpose of our Corollary 3.
6
ARIE LEVIT
subset Q ⊂ Gι(v) such that every Fv -homomorphism ϕ : H1 → Gι(v) satisfies
ϕ(Q1 )g ⊂ Q for some g ∈ Gι(v) .
Proof. Denote i = ι(v) to simplify notation. Let S be a maximal Fv -split torus
of Gi and denote S = S(Fv ). The group H1 admits SL2 as a F -subgroup by
the Jacobson–Morozov theorem. Let T denote the diagonal subgroup of SL2 and
T = T(Fv ). By the assumptions T is a maximal Fv -split torus of H1 .
Consider the following Cartan decompositions [19, 3.2-3]
SL2 (Fv ) = K0 T K0 ,
H1 = K 1 T K 1 ,
Gi = KSK
with the corresponding maximal compact subgroups K0 ≤ SL2 (Fv ) , K1 ≤ H1 and
K ≤ Gi . These three decompositions can be made compatible in the sense that
K0 = K ∩ SL2 (Fv ) , K1 = K ∩ H1
and T = S ∩ SL2 (Fv ) = S ∩ H1
up to conjugating by an element of Gi . This is possible in the Archimedean case
by a theorem of Mostow [18], and in the non-Archimedean case due to the strong
transitivity property of affine buildings [10].
Assume without loss of generality that Q1 = K1 T1 K1 for some compact subset
T1 ⊂ T . We claim that Q = KS1 K for some sufficiently large compact subset
S1 ⊂ S is as required. Indeed, consider the two Lie algebras h = L(H1 ) and
g = L(Gi ) and let dϕ : h → g denote the differential of ϕ. Therefore dϕ(sl2 ) is a
split sl2 -triplet of g in the sense of [4, VIII.§11.1]. In this situation the restriction
of dϕ to the Lie algebra L(T ) of the torus T is bounded [4, VIII.§11.3].
4. Weakly amenable actions and Theorem 4
Let G be a locally compact group and (X, µ) a Borel G-space with an invariant
probability measure. We use spectral gap to deduce essential freeness for such an
action and establish Theorem 4.
Recall that a G-space (X, µ) is weakly amenable if the orbital equivalence relation
generated by the action is amenable; see e.g. [25, Section 4.3] and [23] for details.
Theorem C (Stuck–Zimmer, Bader–Shalom). Let G be a direct product of at least
two simple groups. Assume that the G-space (X, µ) is properly ergodic, irreducible
and not weakly amenable. Then it is µ-essentially free.
Proof. See [23] for the classical case of semisimple Lie groups, [14] for semisimple
linear groups over local fields and [2] for general locally compact groups.
We conclude that Theorem 4 follows at once by combining Theorem C with the
following observation, due to Creutz [8].
Proposition D. Let G be a second countable locally compact group. If the G-space
(X, µ) is properly ergodic and weakly amenable then it has no spectral gap.
Roughly speaking, the proof of Proposition D is as follows. A weakly amenable
action is orbit equivalent to an action of either Z or R, depending on whether G is
countable or not. The action of these amenable groups admits more than a single
invariant mean on L∞ (X, µ). This fact is invariant under orbit equivalence, so that
the action of G admits more than a single invariant mean as well. Therefore the
action of G is not strongly ergodic and in particular there can be no spectral gap.
See [8, 7.3.1] for complete details and references concerning this argument.
It is interesting to note the similarity of the above theme with the role played
by Selberg’s theorem in the solution of the Banach-Ruziewicz problem [15].
ON BENJAMINI–SCHRAMM LIMITS OF CONGRUENCE SUBGROUPS
7
5. Invariant random subgroups and a proof of Theorem 1
We now prove Theorem 1 relying on strong uniform spectral gap for congruence
lattices. Invariant random subgroups are a main tool used in the proof3.
Associated to any second countable locally compact group G is the Chabauty
space of closed subgroups denoted Sub (G). This is a compact G-space with the
Chabauty topology and the conjugation action. An invariant random subgroup of
G is a G-invariant probability measure on Sub (G). Let IRS (G) denote the compact
convex space of all invariant random subgroups of G with the weak-∗ topology.
Associated to any normal subgroup N ⊳ G is the point mass δN ∈ IRS (G). More
interestingly, to any lattice Γ in G we associate µΓ ∈ IRS (G) obtained by pushing
forward the G-invariant probability on G/Γ to Sub (G) via the G-equivariant map
G/Γ → Sub (G) ,
gΓ 7→ gΓg −1
The following two propositions are to be compared with [1, 4.4] and [12, 1.1].
Proposition 7. Let µ ∈ IRS (G) be irreducible and essentially transitive. If µ is an
accumulation point of {µΓ } where Γ are lattices in G then µ = δM with M ≤ Z(G).
Proof. Since (Sub (G) , µ) is essentially transitive there is a closed subgroup H ≤ G
such that µ is supported on the conjugacy class of H and (Sub (G) , µ) is isomorphic
to the homogeneous G-space G/NG (H). The factors Gi of G are non-compact and
the fields ki have zero characteristic so that the density theorem of Borel applies
[17, II.4.4]. Combined with the irreducibility of the action it implies either that
NG (H) is an irreducible lattice or that NG (H) = G and µ is equal to δM for some
normal subgroup M ⊳ G.
We deal with these two possibilities separately, relying on results from [12]. If
NG (H) is a lattice in G then H must be an irreducible lattice as well by the normal
subgroup theorem of Margulis [17, IV]. Every irreducible lattice of G admits a
Chabauty open neighborhood in Sub (G) consisting of conjugates [12, 1.9], so that
the corresponding point µH is isolated in the space of extreme points of IRS (G).
A non-discrete normal subgroup of G does not belong to the closure of the
Chabauty subspace of discrete subgroups [12, 6.7]. Therefore the second case where
µ = δM is impossible unless M is central, as required.
Proposition 8. Every sequence of distinct µΓn ∈ IRS (G) associated to congruence
lattices Γn in G is weak-∗ convergent to δM ∈ IRS (G) for some M ≤ Z(G).
Proof. Let µ ∈ IRS (G) be any accumulation point of the sequence µΓn . We claim
that µ is equal to δM for some M ≤ Z(G), and in particular that µΓn is convergent.
Theorem A implies that the sequence µΓn of G-invariant Borel probability measures on the compact G-space Sub (G) has strong uniform spectral gap. By Proposition 5 the G-space (Sub (G) , µ) has strong spectral gap as well. In particular µ
is both irreducible and has spectral gap.
Making use of the argument on [23, p. 729] we may assume that G has trivial
center to begin with. Since the G-action on (Sub (G) , µ) is certainly not essentially
free, Theorem 4 implies that it must be essentially transitive. We conclude from
Proposition 7.
3See Gelander’s lecture notes [11] on the Chabauty topology, invariant random subgroups and
Benjamini-Schramm spaces.
8
ARIE LEVIT
The following three modes of convergence are all equivalent [12] — weak-∗ convergence of µΓn to a central subgroup in IRS (G), Benjamini–Schramm convergence
of Γn \X to X and the fact that the sequence Γn is weakly central.
The proof of Theorem 1 is now complete.
References
[1] M. Abert, N. Bergeron, I. Biringer, T. Gelander, N. Nikolov, J. Raimbault, and I. Samet.
On the growth of L2 -invariants for sequences of lattices in Lie groups. arXiv preprint
arXiv:1210.2961, 2012.
[2] U. Bader and Y. Shalom. Factor and normal subgroup theorems for lattices in products of
groups. Inventiones mathematicae, 163(2):415–454, 2006.
[3] B. Bekka, P. de La Harpe, and A. Valette. Kazhdan’s property (T), volume 11. Cambridge
university press, 2008.
[4] N. Bourbaki. Lie groups and Lie algebras: chapters 7-9, volume 7. Springer Science & Business Media, 2008.
[5] M. Burger and P. Sarnak. Ramanujan duals II. Inventiones mathematicae, 106(1):1–11, 1991.
[6] L. Clozel. Démonstration de la conjecture τ . Inventiones mathematicae, 151(2):297–328, 2003.
[7] L. Clozel and E. Ullmo. Équidistribution des points de Hecke. Contributions to automorphic
forms, geometry, and number theory, pages 193–254, 2004.
[8] D. Creutz. Stabilizers of actions of lattices in products of groups. Journal of Ergodic Theory
and Dynamical Systems, 2015. (to appear).
[9] M. Fraczyk. Strong limit multiplicity for arithmetic hyperbolic surfaces and 3-manifolds.
arXiv preprint arXiv:1612.05354, 2016.
[10] P. B. Garrett. Buildings and classical groups. CRC Press, 1997.
[11] T. Gelander. A lecture on invariant random subgroups. arXiv:1503.08402, 2015.
[12] T. Gelander and A. Levit. Invariant random subgroups over non-archimedean local fields.
preprint.
[13] S. Gelbart and H. Jacquet. A relation between automorphic representations of GL(2) and
GL(3). In Annales scientifiques de l’École Normale Supérieure, volume 11, pages 471–542,
1978.
[14] A. Levit. The Nevo-Zimmer intermediate factor theorem over local fields. Geometriae Dedicata, 186(1):149–171, 2017.
[15] A. Lubotzky. Discrete groups, expanding graphs and invariant measures. Springer Science &
Business Media, 2010.
[16] A. Lubotzky and A. Zuk. On property (τ ). Notices Amer. Math. Soc, 52(6):626–627, 2005.
[17] G. A. Margulis. Discrete subgroups of semisimple Lie groups, volume 17. Springer Science &
Business Media, 1991.
[18] G. D. Mostow. Fully reducible subgroups of algebraic groups. American Journal of Mathematics, 78(1):200–221, 1956.
[19] V. P. Platonov and A. S. Rapinchuk. Algebraic groups and number theory. Russian Mathematical Surveys, 47(2):133–161, 1992.
[20] J. Raimbault. On the convergence of arithmetic orbifolds. arXiv:1311.5375, 2013.
[21] A. Selberg. On the estimation of fourier coefficients of modular forms. In Proc. Sympos. Pure
Math, volume 8, pages 1–15, 1965.
[22] J.-P. Serre. Le probleme des groupes de congruence pour SL2 . Annals of Mathematics,
92(3):489–527, 1970.
[23] G. Stuck and R. J. Zimmer. Stabilizers for ergodic actions of higher rank semisimple groups.
Annals of Mathematics, pages 723–747, 1994.
[24] M.-F. Vignéras. Quelques remarques sur la conjecture λ1 ≥ 1/4, Seminar on number theory,
Paris 1981–82. Progr. Math, 38:321–343, 1983.
[25] R. J. Zimmer. Ergodic theory and semisimple groups, volume 81. Springer Science & Business
Media, 1984.
| 4math.GR |
On the hardness of losing weight
Andrei Krokhin1 and Dániel Marx2
arXiv:1711.03894v1 [cs.DS] 10 Nov 2017
1
2
Department of Computer Science, Durham University, Durham, DH1 3LE, UK
andrei.krokhin@durham.ac.uk
Department of Computer Science and Information Theory, Budapest University of
Technology and Economics, H-1521 Budapest, Hungary
dmarx@cs.bme.hu
Abstract. We study the complexity of local search for the Boolean constraint satisfaction problem (CSP), in the following form: given a CSP
instance, that is, a collection of constraints, and a solution to it, the
question is whether there is a better (lighter, i.e., having strictly less
Hamming weight) solution within a given distance from the initial solution. We classify the complexity, both classical and parameterized, of
such problems by a Schaefer-style dichotomy result, that is, with a restricted set of allowed types of constraints. Our results show that there
is a considerable amount of such problems that are NP-hard, but fixedparameter tractable when parameterized by the distance.
1
Introduction
Local search is one of the most widely used approaches to solving hard optimization problems. The basic idea of local search is that one tries to iteratively improve a current solution by searching for better solutions in its (k-)neighborhood
(i.e., within distance k from it). Any optimization algorithm can be followed by
a local search phase, thus the problem of finding a better solution locally is of
great practical interest. However, a brute force search of a k-neighborhood is not
feasible for large k, thus it is natural to study the complexity of searching the
k-neighborhood.
The constraint satisfaction problem (CSP) provides a framework in which it
is possible to express, in a natural way, many combinatorial problems encountered in artificial intelligence and computer science. A constraint satisfaction
problem is represented by a set of variables, a domain of values for each variable, and a set of constraints between variables. The basic aim in a constraint
satisfaction problem is then to find an assignment of values to the variables that
satisfies the constraints. Boolean CSP (when all variables have domain {0, 1}),
aka generalized satisfiability, is a natural generalization of Sat where constraints
are given by arbitrary relations, not necessarily by clauses. Local search methods
for Sat and CSP are very extensively studied (see, e.g., [5,9,10,11]).
Complexity classifications for various versions of CSP (and, in particular,
for versions of Boolean CSP) have recently attracted massive attention from
researchers, and one of the most popular directions here is to characterise restrictions on the type of constraints that lead to problems with lower complexity
in comparison with the general case (see [2,3]). Such classifications are sometimes
called Schaefer-style because the first classification of this type was obtained by
T.J. Schaefer in his seminal work [16]. A local-search related Schaefer-style classification for Boolean Max CSP was obtained in [1], in the context of local
search complexity classes such as PLS.
The hardness of searching the k-neighborhood (for any optimisation problem)
can be studied very naturally in the framework of parameterized complexity [6,8],
as suggested in [7]; such a study for the traveling salesman problem (TSP) was recently performed in [14]. Parameterized complexity studies hardness in finer detail than classical complexity. Consider, for example, two standard NP-complete
problems Minimum Vertex Cover and Maximum Clique. Both have the
natural parameter k: the size of the required vertex cover/clique. Both problems
can be solved in nO(k) time by complete enumeration. Notice that the degree of
the polynomial grows with k, so the algorithm becomes useless for large graphs,
even if k is as small as 10. However, Minimum Vertex Cover can be solved
in time O(2k · n2 ) [6,8]. In other words, for every fixed cover size there is a
polynomial-time (in this case, quadratic) algorithm solving the problem where
the degree of the polynomial is independent of the parameter. Problems with this
property are called fixed-parameter tractable. The notion of W[1]-hardness in
parameterized complexity is analogous to NP-completeness in classical complexity. Problems that are shown to be W[1]-hard, such as Maximum Clique [6,8],
are very unlikely to be fixed-parameter tractable. A Schaefer-style classification
of the basic Boolean CSP with respect to parameterized complexity (where the
parameter is the required Hamming weight of the solution) was obtained in [13].
In this paper, we give a Schaefer-style complexity classification for the following problem: given a collection of Boolean constraints, and a solution to it, the
question is whether there is a better (i.e., with smaller Hamming weight) solution
within a given (Hamming) distance k from the initial solution. We obtain classification results both for classical (Theorem 9) and for parameterized complexity
(Theorem 3). However, we would like to point out that it makes much more sense
to study this problem in the parameterized setting. Intuitively, if we are able to
decide in polynomial time whether there is a better solution within distance k,
then this seems to be almost as powerful as finding the best solution (although
there are technicalities such as whether there is a feasible solution at all). Our
classification confirms this intuition: searching the k-neighborhood is polynomialtime solvable only in cases where finding the optimum is also polynomial-time
solvable. On the other hand, there are cases (for example, Horn constraints or
1-in-3 Sat) where the problem of finding the optimum is NP-hard, but searching the k-neighborhood is fixed-parameter tractable. This suggests evidence that
parameterized complexity is the right setting for studying local search.
The paper is organized as follows. Section 2 reviews basic notions of parameterized complexity and Boolean CSP. Section 3 presents the classificiation with
respect to fixed-parameter tractability, while Section 4 deals with polynomialtime solvability. The proofs omitted from Section 4 can be found in the appendix.
2
2
Preliminaries
Boolean CSP. A formula φ is a pair (V, C) consisting of a set V of variables
and a set C of constraints. Each constraint ci ∈ C is a pair hsi , Ri i, where si =
(xi,1 , . . . , xi,ri ) is an ri -tuple of variables (the constraint scope) and Ri ⊆ {0, 1}ri
is an ri -ary Boolean relation (the constraint relation). A function f : V → {0, 1}
is a satisfying assignment of φ if (f (xi,1 ), . . . , f (xi,ri )) is in Ri for every ci ∈ C.
Let Γ be a set of Boolean relations. A formula is a Γ -formula if every constraint
relation Ri is in Γ . In this paper, Γ is always a finite set. The (Hamming) weight
w(f ) of assignment f is the number of variables x with f (x) = 1. The distance
dist(f1 , f2 ) of assignments f1 , f2 is the number of variables x with f1 (x) 6= f2 (x).
We recall various standard definitions concerning Boolean constraints (cf. [3]):
– R is 0-valid if (0, . . . , 0) ∈ R.
– R is 1-valid if (1, . . . , 1) ∈ R.
– R is Horn or weakly negative if it can be expressed as a conjunction of clauses
such that each clause contains at most one positive literal. It is known that R
is Horn if and only if it is min-closed: if (a1 , . . . , ar ) ∈ R and (b1 , . . . , br ) ∈ R,
then (min(a1 , b1 ), . . . , min(ar , br )) ∈ R.
– R is affine if it can be expressed as a conjunction of constraints of the form
x1 +x2 +· · ·+xt = b, where b ∈ {0, 1} and addition is modulo 2. The number
of tuples in an affine relation is always an integer power of 2.
– R is width-2 affine if it can be expressed as a conjunction of constraints of
the form x = y and x 6= y.
– R is IHS-B− (or implicative hitting set bounded) if it can be represented by
a conjunction of clauses of the form (x), (x → y) and (¬x1 ∨ . . . ¬xn ), n ≥ 1.
– The relation Rp-in-q (for 1 ≤ p ≤ q) has arity q and Rp-in-q (x1 , . . . , xq ) is true
if and only if exactly p of the variables x1 , . . . , xq have value 1.
The following definition is new in this paper. It plays a crucial role in characterizing the fixed-parameter tractable cases for local search.
Definition 1. Let R be a Boolean relation and (a1 , . . . , ar ) ∈ R. A set S ⊆
{1, . . . , r} is a flip set of (a1 , . . . , ar ) (with respect to R) if (b1 , . . . , br ) ∈ R
where bi = 1 − ai for i ∈ S and bi = ai for i 6∈ S. We say that R is flip separable
if whenever some (a1 , . . . , ar ) ∈ R has two flip sets S1 , S2 with S1 ⊂ S2 , then
S2 \ S1 is also a flip set for (a1 , . . . , ar ).
It is easy to see that R1-in-3 is flip separable: every flip set has size exactly 2,
hence S1 ⊂ S2 is not possible. Moreover, Rp-in-q is also flip separable for every
p ≤ q. Affine constraints are also flip separable: to see this, it is sufficient to
verify the definition only for the constraint x1 + · · · + xr = 0.
The basic problem in CSP is to decide if a formula has a satisfying assignment:
CSP(Γ )
Input: A Γ -formula φ.
Question: Does φ have a satisfying assignment?
3
Schaefer completely characterized the complexity of CSP(Γ ) for every finite
set Γ of Boolean relations [16]. In particular, every such problem is either in
PTIME or NP-complete, and there is a very clear description of the boundary
between the two cases.
Optimization versions of Boolean CSP were investigated in [3,4]. A straightforward way to obtain an optimization problem is to relax the requirement that
every constraint is satisfied, and ask for an assignment maximizing the number
of satisfied constraints. Another possibility is to ask for a solution with minimum/maximum weight. In this paper, we investigate the problem of minimizing
the weight. As we do not consider the approximability the problem, we define
here only the decision version:
Min-Ones(Γ )
Input: A Γ -formula φ and an integer W .
Question: Does φ have a satisfying assignment f with w(f ) ≤ W ?
The characterization of the approximability of finding a minimum weight
satisfying assignment for a Γ -formula can be found in [3]. Here we state only the
classification of polynomial-time solvable and NP-hard cases:
Theorem 2 ([3]). Let Γ be a finite set of Boolean relations. Min-Ones(Γ ) is
solvable in polynomial time if one the following holds, and NP-complete otherwise:
– Every R ∈ Γ is 0-valid.
– Every R ∈ Γ is Horn.
– Every R ∈ Γ is width-2 affine.
A Schaefer-style characterization of the approximability of finding two satisfying assignments to a formula with a largest distance between them was obtained in [4], motivated by the blocks world problem from KR, while a Schaeferstyle classification of the problem of deciding whether a given satisfying assignment to a given CSP instance is component-wise minimal was presented in [12],
motivated by the circumscription formalism from AI.
The main focus of the paper is the local search version of minimizing weight:
LS-CSP(Γ )
Input: A Γ -formula φ, a satisfying assignment f , and an integer k.
Question: Does φ have a satisfying assignment f ′ with w(f ′ ) < w(f )
and dist(f, f ′ ) ≤ k?
LS in the above problem stands for both “local search” and “lighter solution.”
Observe that the satisfying assignments of an (x ∨ y)-formula correspond to
the vertex covers of the graph where the variables are the vertices and the edges
are the constraints. Thus LS-CSP({x ∨ y}) is the problem of reducing the size
of a (given) vertex cover by including and excluding a total of at most k vertices.
4
As we shall see (Lemma 7), this problem is W[1]-hard, even for bipartite graphs.
This might be of independent interest.
Parameterized complexity. In a parmeterized problem, each instance contains an integer k called the parameter. A parameterized problem is fixed-parameter tractable (FPT) if can be solved by an algorithm with running time f (k)·nc ,
where n is the length of the input, f is an arbitrary (computable) function depending only on k, and c is a constant independent of k.
A large fraction of NP-complete problems is known to be FPT. On the other
hand, analogously to NP-completeness in classical complexity, the theory of
W[1]-hardness can be used to give strong evidence that certain problems are
unlikely to be fixed-parameter tractable. We omit the somewhat technical definition of the complexity class W[1], see [6,8] for details. Here it will be sufficient
to know that there are many problems, including Maximum Clique, that were
proved to be W[1]-hard. To prove that a parameterized problem is W[1]-hard,
we have to present a parameterized reduction from a known W[1]-hard problem.
A parameterized reduction from problem L1 to problem L2 is a function that
transforms a problem instance x of L1 with parameter k into a problem instance
x′ of L2 with parameter k ′ in such a way that
– x′ is a yes-instance of L2 if and only if x is a yes-instance of L1 ,
– k ′ can be bounded by a function of k, and
– the transformation can be computed in time f (k) · |x|c for some constant c
and function f (k).
It is easy to see that if there is a parameterized reduction from L1 to L2 , and
L2 is FPT, then it follows that L1 is FPT as well.
3
Characterizing fixed-parameter tractability
In this section, we completely characterize those finite sets Γ of Boolean relations
for which LS-CSP(Γ ) is fixed-parameter tractable.
Theorem 3. Let Γ be a finite set of Boolean relations. The problem LS-CSP(Γ )
is in FPT if every relation in Γ is Horn or every relation in Γ is flip separable.
In all other cases, LS-CSP(Γ ) is W[1]-hard.
First we handle the fixed-parameter tractable cases (Lemmas 4 and 6)
Lemma 4. If every R ∈ Γ is Horn, then LS-CSP(Γ ) is FPT.
Proof. Observe that if there is a solution f ′ , then we can assume that f ′ (x) ≤
f (x) for every variable x: by defining f ′′ (x) := min{f (x), f ′ (x)}, we get that f ′′
is also satisfying (as every R ∈ Γ is min-closed) and dist(f ′′ , f ) ≤ dist(f ′ , f ).
Thus we can restrict our search to solutions that can be obtained from f by
changing some 1’s to 0’s, but every 0 remains unchanged.
Since w(f ′ ) < w(f ), there is a variable x with f (x) = 1 and f ′ (x) = 0. For
every variable x with f (x) = 1, we try to find a solution f ′ with f ′ (x) = 0 using
5
a simple bounded-height search tree algorithm. For a particular x, we proceed as
follows. We start with initial assignment f . Change the value of x to 0. If there is
a constraint h(x1 , . . . , xr ), Ri that is not satisfied by the new assignment, then we
select one of the variables x1 , . . . , xr that has value 1, and change it to 0. Thus
at this point we branch into at most r directions. If the assignment is still not
satisfying, the we branch again on the variables of some unsatisfied constraint.
The branching factor of the resulting search tree is at most rmax , where rmax is
the maximum arity of the relations in Γ . By the observation above, if there is a
solution, then we find a solution on the first k levels of the search tree. Therefore,
k+1
we can stop the search on the k-th level, implying that we visit at most rmax
nodes of the search tree. The work to be done at each node is polynomial in the
k+1
size n of the input, hence the total running time is rmax
· nO(1) .
⊓
⊔
If every R ∈ Γ is not only Horn, but IHS-B− (which is a subset of Horn), then
the algorithm of Lemma 4 actually runs in polynomial time:
Corollary 5. If every R ∈ Γ is IHS-B−, then LS-CSP(Γ ) is in PTIME.
Proof. We can assume that every constraint is either (x), (x → y), or (x̄1 ∨ · · · ∨
x̄r ). If a constraint (x̄1 ∨ · · · ∨ x̄r ) is satisfied in the initial assignment f , then it
remains satisfied after changing some 1’s to 0. Observe that if a constraint (x) or
(x → y) is not satisfied, then at most one variable has the value 1. Thus there is
no branching involved in the algorithm of Lemma 4, making it a polynomial-time
algorithm.
⊓
⊔
For flip separable relations, we give a very similar branching algorithm. However, in this case the correctness of the algorithm requires a nontrivial argument.
Lemma 6. If every R ∈ Γ is flip separable, then LS-CSP(Γ ) is FPT.
Proof. Let (φ, f, k) be an instance of LS-CSP(Γ ). If w(f ′ ) < w(f ) for some
assignment f ′ , there is a variable x with f (x) = 1 and f ′ (x) = 0. For every
variable x with f (x) = 1, we try to find a solution f ′ with f ′ (x) = 0 using
a simple bounded-height search tree algorithm. For each such x, we proceed
as follows. We start with the initial assignment f and set the value of x to 0.
Iteratively do the following: (a) if there is a constraint in φ that is not satisfied
by the current assignment and such that the value of some variable in it has not
been flipped yet (on this branch), then we select one of such variables, and flip
its value; (b) if there is no such constraint, but the current assignment is not
satisfying then we move to the next branch; (c) if every constraint is satisfied,
then either we found a required solution or else we move to the next branch. If
a required solution is not found on the first k levels of the search tree then the
algorithm reports that there is no required solution.
Assume that (φ, f, k) is a yes-instance. We claim that if f ′ is a required
solution with minimal distance from f , then some branch of the algorithm finds
it. Let X be the set of variables on which f and f ′ differ, so |X| ≤ k. We now show
that on the first k levels of the search tree, the algorithm finds some satisfying
6
assignment f0 (possibly heavier than f ) that differs from f only on a subset
X0 ⊆ X of variables. To see this, assume that at some node of the search tree,
the current assignment differs from the initial assignment only on a subset of X;
we show that this remains true for at least one child of the node. If we branch
on the variables (x1 , . . . , xr ) of an unsatisfied constraint, then at least one of its
variables, say xi , has a value different from f ′ (as f ′ is a satisfying assignment).
It follows that xi ∈ X: otherwise the current value of xi is f (xi ) (since so far
we changed variables only in X) and f (xi ) = f ′ (xi ) (by the definition of X),
contradicting the fact that current value of xi is different from f (xi ). Thus if we
change variable xi , it remains true that only variables from X are changed. Since
|X| ≤ k, this branch of the algorithm has to find some satisfying assignment f0 .
If w(f0 ) < w(f ), then, by the choice of f ′ , we must have f0 = f ′ . Otherwise,
let X0 ⊆ X be the set of variables where f and f0 differ and let f ′′ be the
assignment that differs from f exactly on the variables X \X0 . From the fact that
every constraint is flip separable, it follows that f ′′ is a satisfying assignment.
We claim that w(f ′′ ) < w(f ). Indeed, if changing the values of the variables in X
decreases the weight and changing the values in X0 does not decrease the weight,
then the set X \ X0 has to decrease the weight. This contradicts the assumption
that f ′ is a solution whose distance from f is minimal: f ′′ is a solution with
distance |X \ X0 | < |X|. Thus it is sufficient to investigate only the first k levels
of the search tree. As in the proof of Lemma 4, the branching factor of the tree
k+1
is at most rmax , and the algorithm runs in time rmax
· nO(1) .
⊓
⊔
All the hardness proofs in this section are based on the following lemma:
Lemma 7. LS-CSP({x ∨ y}) is W[1]-hard.
Proof. The proof is by reduction from a variant of Maximum Clique: given a
graph G(V, E) with a distinguished vertex x and an integer t, we have to decide
whether G has a clique of size t that contains x. It is easy to see that this problem
is W[1]-hard. Furthermore, it can be assumed that t is odd. Let n be the number
of vertices of G and let m be the number of edges. We construct a formula φ on
m + n(t − 1)/2 − 1 variables and a satisfying assignment f such that G has a
clique of size t containing x if and only if φ has a satisfying assignment f ′ with
w(f ′ ) < w(f ) and distance at most k := t(t − 1) − 1 from f .
Let d = (t − 1)/2 (note that t is odd). The formula φ has d variables v1 ,
. . . , vd for each vertex v 6= x of G and a variable ue for each edge e of G. The
distinguished vertex x has only d − 1 variables x1 , . . . , xd−1 . If a vertex v is the
endpoint of an edge e, then for every 1 ≤ i ≤ d (or 1 ≤ i ≤ d − 1, if v = x), we
add the constraint ue ∨ vi . Thus each variable ue is in 2d − 1 or 2d constraints
(depending on whether x is the endpoint of e or not). Set f (ue ) = 1 for every
e ∈ E and f (vi ) = 0 for every v ∈ V , 1 ≤ i ≤ d. Clearly, f is a satisfying
assignment.
Assume that G has a clique K of size t that includes x. Set f ′ (vi ) = 1 for
every v ∈ K (1 ≤ i ≤ d) and set f ′ (ue ) = 0 for every edge e in K; let f ′ be the
same as f on every other variable. Observe that f ′ is also a satisfying assignment:
if a variable ue was changed to 0 and there is a constraint ue ∨ vi , then v ∈ K
7
and hence f ′ (vi ) = 1. We have w(f ′ ) < w(f ): dt − 1 variables were changed to
1 (note that x ∈ K) and t(t − 1)/2 = dk variables were changed to 0. Moreover,
the distance of f and f ′ is exactly dt − 1 + t(t − 1)/2 = t(t − 1) − 1 = k.
Assume now that f ′ satisfies the requirements. Let K be the set of those
vertices v in G for which f ′ (vi ) = 1 for every i. We claim that K is a clique of size
t in G. Observe that there are at least d|K| − 1 variables vi with f ′ (vi ) > f (vi )
and f ′ (ue ) < f (ue ) is possible only if both endpoints of e are in K, i.e., e
is in the set E(K) of edges in K. Thus w(f ′ ) < w(f ) implies d|K| − 1 <
|E(K)| ≤ |K|(|K| − 1)/2, which is only possible if |K| ≥ t. If |K| > t, then
f ′ (vi ) > f (vi ) for at least (t + 1)d − 1 variables, hence there must be at least
that many variables ue with f ′ (ue ) < f (ue ). Thus the distance of f and f ′ is at
least 2(t + 1)d − 2 > t(t − 1) − 1. Therefore, we can assume |K| = t. Now dt − 1 <
|E(K)| ≤ |K|(|K| − 1)/2 = t(t − 1)/2 is only possible if |E(K)| = t(t − 1)/2
(i.e., K is a clique) and it follows that there are exactly dt − 1 variables vi with
f ′ (vi ) > f (vi ) (i.e., x ∈ K).
⊓
⊔
Now we are ready to present the main hardness proof of the section:
Lemma 8. If Γ contains a relation R1 that is not Horn and a relation R2 that
is not flip separable, then LS-CSP(Γ ) is W[1]-hard.
Proof. The proof is by reduction from LS-CSP({x ∨ y}). Let (phi1 , f1 , k) be
an instance of LS-CSP({x ∨ y}), i.e., every constraint relation in formula φ1 =
(V, C) is (x ∨ y). Since R1 is not min-closed, we can assume (by permuting the
variables) that for some r1 , r2 ≥ 1, r3 , r4 ≥ 0, if we define
r1
r2
r3
r4
z }| { z }| { z }| { z }| {
R1′ (x, y, w0 , w1 ) = R1 (x, . . . , x, y, . . . , y, w0 , . . . , w0 , w1 , . . . , w1 ),
then (0, 1, 0, 1), (1, 0, 0, 1) ∈ R1′ , but (0, 0, 0, 1) 6∈ R1′ . Since R1′ is obtained from
R1 by identifying variables, we can use the relation R1′ when specifying instances
of LS-CSP(Γ ). We consider two cases:
Case 1: (1, 1, 0, 1) ∈ R1′ . In this case R1′ (x, y, 0, 1) = x ∨ y, hence it is
easy to simulate LS-CSP({x ∨ y}). The only difficulty is how to simulate the
constants 0 and 1. We do this as follows. Let us construct a formula φ2 that
has every variable of V and new variables q0j , q1j for every 1 ≤ j ≤ k + 1 (these
new variables will play the role of the constants). We define assignment f2 of
φ2 by setting f2 (x) = f1 (x) for x ∈ V and f2 (q0j ) = 0 and f2 (q1j ) = 1 for
1 ≤ j ≤ k + 1. For 1 ≤ a, b, c ≤ k + 1, we add constraint c1a,b,c = R1′ (q1a , q0b , q0b , q1c ),
it is clearly satisfied by assignment f2 . To simulate a constraint x ∨ y, we add
c2x,y,j = R1′ (x, y, q0j , q11 ) for every 1 ≤ j ≤ k + 1.
It is easy to see that if there is a solution f1′ for the original instance (φ1 , f1 , k),
then by setting f2′ (x) = f1′ (x) for every x ∈ V and f2′ (q0j ) = 0, f2′ (q1j ) = 1 for
every 1 ≤ j ≤ k + 1 gives a solution f2′ for the constructed instance (φ2 , f2 , k).
We claim the converse is also true: if f2′ is a solution for the instance (φ2 , f2 , k),
then the restriction of f2′ to V gives a solution for (φ1 , f1 , k). Since the distance
of f2 and f2′ is at most k, there are 1 ≤ b, c ≤ k + 1 with f2′ (q0b ) = 0 and
f2′ (q1c ) = 1. Because of the constraint c1a,b,c , we have that f2′ (q1a ) = 1 for every
8
1 ≤ a ≤ k + 1. It follows that f2′ restricted V is a satisfying assignment of
φ1 : for every constraint x ∨ y ∈ C, the constraint c2x,y,b prevents the possibility
f2′ (x) = f2′ (y) = 0. We have seen that f2′ (q0j ) ≥ f2 (q0j ) and f2′ (q1j ) ≥ f2 (q1j ) for
every 1 ≤ j ≤ k + 1. Now w(f2′ ) < w(f2 ) implies that the weight of f2′ on V has
to be less than the weight of f2 on V . Thus w(f1′ ) < w(f1 ).
Case 2: (1, 1, 0, 1) 6∈ R1′ , which means that R1′ (x, y, 0, 1) is x 6= y. In this case
we have to rely on the fact that R2 is not flip separable to simulate the constraint
x ∨ y. We construct formula φ2 and its satisfying assignment f2 as follows. Each
variable x is replaced by 3 variables x1 , x2 , x3 . We set f2 (x1 ) = f2 (x2 ) = f1 (x)
and f2 (x3 ) = 1 − f1 (x). Furthermore, for 1 ≤ j ≤ 3k + 1, we add the variables
q0j and q1j and set f2 (q0j ) = 0 and f2 (q1j ) = 1.
For every 1 ≤ a, b, c ≤ 3k + 1, we add the constraint c1a,b,c = R1′ (q1a , q0b , q0b , q1c ),
as in the previous case. For every x ∈ V , 1 ≤ j ≤ 3k + 1, and ℓ = 1, 2, we
add c2x,ℓ,j = R1′ (xℓ , x3 , q0j , q11 ), as we shall see, the role of these constraints is to
ensure f2′ (x1 ) = f2′ (x2 ) 6= f2′ (x3 ).
Since R2 is not flip separable, there is a tuple (s1 , . . . , sr ) ∈ R2 that has flip
sets S1 ⊂ S2 , but S2 \ S1 is not a flip set. For every constraint x ∨ y of φ1 , we
add 3k + 1 constraints to φ2 as follows. First, for 1 ≤ i ≤ r and 1 ≤ j ≤ 3k + 1,
we define variable vij as
x1 if i ∈ S1 and si = 0,
x3 if i ∈ S1 and si = 1,
y
if i ∈ S2 \ S1 and si = 1,
1
vij =
y3 if i ∈ S2 \ S1 and si = 0,
q11 if i 6∈ S2 and si = 1,
j
q0 if i 6∈ S2 and si = 0.
For every 1 ≤ j ≤ 3k + 1, we add the constraint c3x,y,j = R2 (v1j , . . . , vrj ). For
example, assume that (0, 1, 0, 1) ∈ R2 and this tuple has flip sets S1 = {1, 2}
and S2 = {1, 2, 3, 4}, but S2 \ S1 = {3, 4} is not a flip set. This means that
(0, 1, 0, 1), (1, 0, 1, 0), (1, 0, 0, 1) ∈ R2 and (0, 1, 1, 0) 6∈ R2 . In this case, constraint
c3x,y,j is R2 (x1 , x3 , y3 , y1 ). Assuming f (x1 ) 6= f (x3 ) and f (y1 ) 6= f (y3 ), any
combination of values on x1 and y1 satisfies the constraint, except if f (x1 ) =
f (y1 ) = 0. Thus the constraint effectively acts as a constraint x1 ∨ y1 .
Finally, we set the maximum allowed distance to k ′ := 3k. This completes
the description of the constructed instance (φ2 , f2 , k ′ ).
Assume first that f1′ is a solution for the instance (φ1 , f1 , k). Define f2′ (x1 ) =
′
f2 (x2 ) = f1′ (x) and f2′ (x3 ) = 1 − f1′ (x) for every x ∈ V , and define f2′ (q0j ) = 0,
f2′ (q1j ) = 1 for every 1 ≤ j ≤ 3k + 1. The fact w(f1′ ) < w(f1 ) implies w(f2′ ) <
w(f2 ). Furthermore, the distance of f2 and f2′ is exactly three times the distance
of f1 and f1′ , i.e., at most 3k. We claim that f2′ satisfies the constraints of φ2 .
This is easy to see for c1a,b,c and c2x,ℓ,j . For c3x,y,j , this can be seen as follows:
– If f2′ (x) = 0, f2′ (y) = 1, then because (s1 , . . . , sr ) ∈ R2 .
– If f2′ (x) = 1, f2′ (y) = 0, then because S2 is a flip set.
9
– If f2′ (x) = 1, f2′ (y) = 1, then because S1 is a flip set.
For the other direction, assume that f2′ is a solution for instance (φ2 , f2 , k ′ ).
Define f1′ (x) = f2′ (x1 ) for every x ∈ V ; we claim that f1′ is a solution for instance
(φ1 , f1 , k). Since the distance of f2 and f2′ is at most 3k, there are 1 ≤ b, c ≤ 3k+1
with f2′ (q0b ) = 0 and f2′ (q1c ) = 1. Because of the constraint c1a,b,c , we have that
f2′ (q1a ) = 1 for every 1 ≤ a ≤ 3k + 1. The constraints c2x,1,b and c2x,2,b ensure
that f2′ (x1 ) = f2′ (x2 ) = 1 − f2′ (x3 ) (since (0, 0, 0, 1) 6∈ R1′ and (1, 1, 0, 1) 6∈ R1′ ).
It follows that the distance of f1 and f1′ is at most k: f1 (x) 6= f1′ (x) implies
f2 (xℓ ) 6= f2′ (xℓ ) for ℓ = 1, 2, 3, hence this can hold for at most k different
x ∈ V . Moreover, w(f1′ ) < w(f1 ): this follows from the facts w(f2′ ) < w(f2 ) and
f2′ (q0j ) ≥ f2 (q0k ), f2′ (q1j ) ≥ f2 (q1k ) (1 ≤ j ≤ 3k + 1).
We claim that every constraint x ∨ y of φ1 is satisfied. Assume that f1′ (x) =
′
f1 (y) = f2′ (x1 ) = f2′ (y1 ) = 0. Now c3x,y,b is not satisfied: this follows from the
fact that S2 \ S1 is not a flip set for (s1 , . . . , sr ) (with respect to R2 ).
⊓
⊔
4
Characterizing polynomial-time solvability
In this section, we completely characterize those finite sets Γ of Boolean relations
for which LS-CSP(Γ ) is polynomial-time solvable.
Theorem 9. Let Γ be a finite set of Boolean relations. The problem LS-CSP(Γ )
is in PTIME if every relation in Γ is IHS-B− or every relation in Γ is width-2
affine. In all other cases, LS-CSP(Γ ) is NP-hard.
Proof. If every relation in Γ is IHS-B−, then Corollary 5 gives a polynomialtime algorithm. If every relation in Γ is width-2 affine then the following simple
algorithm solves LS-CSP(Γ ): for a given instance (φ, f, k), compute the graph
whose vertices are the variables in φ and two vertices are connected if there is
a constraint in φ imposed on them. If there is a connected component of this
graph which has at most k vertices and such that f assigns more 1’s in this
component than it does 0’s, then flipping the values in this component gives a
required lighter solution. If such a component does not exists, then there is no
lighter solution within distance k from f .
By Lemma 8, if Γ contains a relation that is not Horn and a relation that is
not flip separable then LS-CSP(Γ ) is NP-hard. (Note that the proof is actually a
polynomial-time reduction from an NP-hard problem.) Therefore, we can assume
that every relation in Γ is Horn or every relation in Γ is flip separable. We now
give the proof for the former case, while the proof for the latter case can be
found in Appendix.
Assume now that Γ is Horn, and there is a relation R ∈ Γ that is not IHS-B−.
We prove that LS-CSP({R}) is NP-hard. It is shown in the proof of Lemma 5.27
of [3] that then R is at least ternary and one can permute the coordinates in R
and then substitute 0 and 1 in R in such a way that that the ternary relation
R′ (x, y, z) = R(x, y, z, 0, . . . , 0, 1, . . . , 1) has the following properties:
1. R′ contains tuples (1, 1, 1), (0, 1, 0), (1, 0, 0), (0, 0, 0), and
10
2. R′ does not contain the tuple (1, 1, 0).
Note that if (0, 0, 1) ∈ R′ then R′ (x, x, y) is x → y. If (0, 0, 1) 6∈ R′ then,
since R (and hence R′ ) is Horn (i.e., min-closed), at least one of of the tuples
(1, 0, 1) and (0, 1, 1) is not in R′ . Then it is easy to check that at least one of
the relations R′ (x, y, x) and R′ (y, x, x) is x → y. Hence, we can use constraints
of the form x → y when specifying instances of LS-CSP({R′ }).
We reduce Minimum Dominating Set to LS-CSP({R′ }). Let G(V, E) be
a graph with n vertices and m edges where a dominating set of size at most t
has to be found. Let v1 , . . . , vn be the vertices of G. Let S = 3m. We construct
a formula with nS + 2m + 1 variables as follows:
– There is a special variable x.
– For every 1 ≤ i ≤ n, there are S variables xi,1 , . . . , xi,S . There is a constraint
xi,j → xi,j ′ for every 1 ≤ j, j ′ ≤ n.
– For every 1 ≤ i ≤ n, if vs1 , . . . , vsd are the neighbors of vi , then there
are d variables yi,1 , . . . , yi,d and the following constraints: xs1 ,1 → yi,1 ,
R′ (xs2 ,1 , yi,1 , yi,2 ), R′ (xs3 ,1 , yi,2 , yi,3 ), . . . , R′ (xsd ,1 , yi,d−1 , yi,d ), R′ (xi,1 , yi,d , x).
– For every variable z, there is a constraint x → z.
Observe that the number of variables of type yi,j is exactly 2m. Setting every
variable to 1 is a satisfying assignment. Set k := St + S − 1.
Assume that there is a satisfying assignment where the number of 0’s is at
most k (but positive). Variable x has to be 0, otherwise every other variable is
1. If xi,1 is 0, then xi,j is 0 for every 1 ≤ j ≤ S. Thus k < S(t + 1) implies that
there are at most t values of i such that xi,1 is 0. Let D consist of all vertices
vi such that xi,1 is 0. We claim that D is a dominating set. Suppose that some
vertex vi is not dominated. This means that if vs1 , . . . , vsd are the neighbors of
vi , then the variables xs1 ,1 , . . . , xsd ,1 , xi,1 all have the value 1. However, this
means that these variables force variables yi,1 , . . . , yi,d and variable x to value
1, a contradiction. Thus D is a dominating set of size at most t.
The reverse direction is also easy to see. Assume that G has a dominating set
D of size at most t. For every 1 ≤ i ≤ n and 1 ≤ j ≤ S, set variable xi,j to 1 if
and only vi is not contained in D. Set x to 0. It is easy to see that this assignment
can be extended to the variables yi,j to obtain a satisfying assignment: indeed,
if vs1 , . . . , vsd are the neighbors of vi and none of them is in D then vi ∈ D,
and we set yi,1 = . . . = yi,d = 1. Otherwise, if j is minimal such that vsj ∈ D,
we set yi,1 = . . . = yi,j−1 = 1 and yi,q = 0 for q ≥ j. This satisfying assignment
contains at most St + 2m + 1 ≤ k variables with value 0, as required.
Finally, we reduce LS-CSP({R′ }) to LS-CSP({R}) (and so to LS-CSP(Γ )).
Take an instance (φ, f, k) of LS-CSP({R′ }), let V be the variables of φ and
c1 , . . . , cp the constraints of φ. We build an instance φ′ of LS-CSP({R}) as
follows.
1. For each 1 ≤ i ≤ max(p, k + 1), introduce new variables xi0 , xi1 .
2. For each constraint ci = R′ (x, y, z) in formula φ, replace it by the constraint
R(x, y, z, xi0 , . . . , xi0 , xi1 , . . . , xi1 ).
11
3. For each ordered pair (i, j) where 1 ≤ i, j ≤ max(p, k+1), add the constraints
R(xi0 , xi0 , xj0 , xj0 , . . . , xj0 , xj1 , . . . , xj1 ) and R(xj1 , xj1 , xi1 , xj0 , . . . , xj0 , xj1 , . . . , xj1 ).
Finally, extend f so that, for all i, we have xi0 = 0 and xi1 = 1. It is clear that
the obtained mapping f ′ is a solution to the new instance. Note that, by the
choice of R′ , the tuple (1, 1, 0, 0, . . . , 0, 1, . . . , 1) does not belong to R. Hence, the
constraints added in step (3) above ensure that if a variable of the form xi0 or
xi1 in f ′ is flipped then, in order to get a solution to φ′ different from f ′ , one
must flip at least one of xi0 or xi1 for each 1 ≤ i ≤ max(p, k + 1). Consequently,
all solutions to φ′ that lie within distance k from f ′ must agree with f ′ on all
such variables. In other words, searching for such a solution, it makes sense to
flip only variables from V . Thus, clearly, the instances (φ, f, k) and (φ′ , f ′ , k) are
equivalent.
⊓
⊔
References
1. P. Chapdelaine and N. Creignou. The complexity of Boolean constraint satisfaction
local search problems. Annals of Mathematics and Artificial Intelligence, 43:51–63,
2005.
2. D. Cohen and P. Jeavons. The complexity of constraint languages. In F. Rossi,
P. van Beek, and T. Walsh, editors, Handbook of Constraint Programming, chapter 8. Elsevier, 2006.
3. N. Creignou, S. Khanna, and M. Sudan. Complexity Classifications of Boolean Constraint Satisfaction Problems, volume 7 of SIAM Monographs on Discrete Mathematics and Applications. 2001.
4. P. Crescenzi and G. Rossi. On the Hamming distance of constraint satisfaction
problems. Theoretical Computer Science, 288(1):85–100, 2002.
5. E. Dantsin, A. Goerdt, E. Hirsch, R. Kannan, J. Kleinberg, C. Papadimitriou,
2
P. Raghavan, and U. Schöning. A deterministic (2 − k+1
)n algorithm for k-SAT
based on local search. Theoretical Computer Science, 289:69–83, 2002.
6. R. Downey and M. Fellows. Parameterized Complexity. Springer, 1999.
7. M. Fellows. Parameterized complexity: new developments and research frontiers.
In Aspects of Complexity (Kaikura, 2000), volume 4 of de Gruyter Series in Logic
and Applications, pages 51–72. 2001.
8. J. Flüm and M. Grohe. Parameterized Complexity Theory. Springer, 2006.
9. J. Gu, P. Purdom, J. Franko, and B.W. Wah. Algorithms for the Satisfiability
Problem. Cambridge University Press, Cambridge, MA, 2000.
10. E. Hirsch. SAT local search algorithms: worst-case study. Journal of Automated
Reasoning, 24:127–143, 2000.
11. H. Hoos and E. Tsang. Local search methods. In F. Rossi, P. van Beek, and
T. Walsh, editors, Handbook of Constraint Programming, chapter 5. Elsevier, 2006.
12. L. Kirousis and Ph. Kolaitis. The complexity of minimal satisfiability problems.
Information and Computation, 187:20–39, 2003.
13. D. Marx. Parameterized complexity of constraint satisfaction problems. Computational Complexity, 14:153–183, 2005.
14. D. Marx. Searching the k-change neighborhood for TSP is W[1]-hard. Operations
Research Letters, 36:31–36, 2008.
15. C. Moore and J. M. Robson. Hard tiling problems with simple tiles. Discrete
Comput. Geom., 26(4):573–590, 2001.
12
16. T.J. Schaefer. The complexity of satisfiability problems. In STOC’78, pages 216–
226, 1978.
13
Appendix
The following proposition completes the proof of Theorem 9.
Proposition 10. If R is a flip separable relation that is not width-2 affine then
LS-CSP({R}) is NP-hard.
The above proposition will be proved through a sequence of lemmas, the
main lemmas being Lemma 16 and Lemma 19.
We need three auxiliary lemmas that will be used in the subsequent proofs.
Lemma 11. Let R is the set of solutions to a Γ -formula φ, and let R′ =
prJ (R) where J = {j | prj (R) = {0, 1}}. Then the problems LS-CSP(Γ ) and
LS-CSP(Γ ∪ {R′ }) are polynomial-time equivalent.
Proof. Note that, for any fixed j 6∈ J, each solution to φ takes the same value on
the corresponding variable. In every instance of CSP(Γ ∪ {R′ }), every constraint
of the form c = R′ (s) can be replaced by the constraints from φ where variables
from s keep their places, while all other variables are new and do not appear
elsewhere. This transformation (with the distance k unchanged) is a polynomialtime reduction from LS-CSP(Γ ∪ {R′ }) to LS-CSP(Γ )
⊓
⊔
Let C0 = {0} and C1 = {1}, and let Γ c = Γ ∪ {C0 , C1 }.
Lemma 12. If Γ contains at least one of the relations = or 6= then LS-CSP(Γ )
is polynomial-time equivalent to LS-CSP(Γ c ).
Proof. Assume that Γ contains the equality relation, the other case is very similar. We only need to provide a reduction from LS-CSP(Γ c ) to LS-CSP(Γ ).
Let (φ, f, k) be an instance of the former problem. For every variable x in φ
such that φ contains the constraint C0 (x) or C1 (x) (i.e., the value of variable
x is forced to be a constant 0 or 1), remove this constraint, introduce k new
variables yx,1 , . . . , yx,k , and add new constraints x = yx,i , (i = 1, . . . , k). Call the
obtained Γ -formula φ′ . Extend f to a solution f ′ of φ′ by setting f ′ (yx,i ) = f (x)
for all i, and consider the instance (φ′ , f ′ , k) of LS-CSP(Γ ). The equality constraints in φ′ ensure that every solution to φ′ whose distance is at most k from
f ′ must coincide with f ′ on all variables of the form yx,i . It is clear now that the
instances (φ, f, k) and (φ′ , f ′ , k) are equivalent.
⊓
⊔
Let R1 be an arbitrary n-ary relation (n ≥ 2) and let R2 be an (n + 1)-ary
relation such that (a1 , . . . , an , an+1 ) ∈ R2 ⇔ (a1 , . . . , an ) ∈ R1 and an 6= an+1 .
Lemma 13. The problems LS-CSP({R1 , 6=}) and LS-CSP({R2 , 6=}) are polynomial time equivalent.
Proof. It is clear that LS-CSP({R2 , 6=}) reduces to LS-CSP({R1 , 6=}), by simply replacing each constraint involving R2 by its definition via R1 and 6=. Let us
reduce LS-CSP({R1 , 6=}) to LS-CSP({R2 , 6=}). Let (φ, f, k) be an instance of
LS-CSP({R1 , 6=}) where φ is over a set V of variables. For each variable x ∈ V ,
14
introduce two new variables x′ , x′′ along with constraints x 6= x′ , x′ 6= x′′ . Replace every constraint of the form R1 (xi1 , . . . , xin ) in φ by R2 (xi1 , . . . , xin , x′in ),
and leave all other constraints in φ unchanged. Let φ′ be the obtained instance of
CSP({R2 , 6=}). Clearly, f has a unique extension to a solution f ′ to I ′ . It is clear
that the instance (φ′ , f ′ , 3k) of LS-CSP({R2 , 6=}) is equivalent to (φ, f, k). ⊓
⊔
We consider affine constraints first.
Lemma 14. If Γ contains the ternary relation defined by equation x+y +z = 1,
then LS-CSP(Γ ) is NP-hard.
Proof. First, we show that we can assume that the relation 6= is in Γ . Notice
that if R is the set of all solutions to the instance x + y + z = 1, z + z + z = 1.
Then pr{3} (R) = 1 while pr{1,2} (R) is the equality relation. By Lemma 11, we
may assume that the relation = is in Γ . Then, by Lemma 12, we can assume
that C0 , C1 ∈ Γ . By considering the instance C0 (z), x + y + z = 1 and using Lemma 11 again, we get that LS-CSP(Γ ) is polynomial-time equivalent to
LS-CSP(Γ ∪ {6=}).
The hardness proof is by reduction from CSP(R1-in-3 ), which is known to
be NP-hard even if every variable appears in exactly 3 constraints and each
constraint contains 3 distinct variables [15]. (This implies that the number of
variables equals the number of constraints and the weight of every solution is
exactly n/3, where n is the number of variables.) Given a CSP(R1-in-3 ) formula
φ with n variables x1 , . . . , xn , we construct a Γ -formula φ′ with variables
– xi,j for 1 ≤ i ≤ n, 1 ≤ j ≤ 2n,
– vt for every 1 ≤ t ≤ n,
– yj for every 0 ≤ j ≤ 2n2 .
For every 1 ≤ i ≤ n, 1 ≤ j, j ′ ≤ 2n, we add the constraint xi,j = xi,j ′ . For
every 1 ≤ j, j ′ ≤ 2n2 , we add the constraint yj = yj ′ . Assume that the t-th
constraint in φ is on variables xa , xb , xc . For 1 ≤ j ≤ 2n, we add the constraints
xa,j + xb,j + vt = 1 and xc,j + vt + y0 = 1. Finally, we add the constraint
y0 6= y1 . Define assignment f such that f (vt ) = 1 for 1 ≤ t ≤ n and f (yj ) = 1
for 1 ≤ i ≤ 2n2 , and every other variable is 0. The weight of f is 2n2 + n. Set
k := 2n2 + 1 + n + 2n · n/3. This completes the description of the reduction.
Assume that φ has a solution f0 . Define f ′ such that f ′ (xi,j ) = f0 (xi ) (1 ≤
i ≤ n, 1 ≤ j ≤ 2n), f ′ (y0 ) = 1, f ′ (yj ) = 0 (1 ≤ j ≤ 2n2 ). This assignment can
be extended to each vt in a unique way: if the t-th constraint in φ is on variables
xa , xb , xc , then exactly two of the variables xa,1 , xb,1 , xc,1 , y0 have value 1, hence
we can set vt accordingly. Thus we can obtain a satisfying assignment f ′ this
way. Observe that the weight of f0 is exactly n/3: each variable with value 1 in
f0 appears in exactly 3 constraints and each constraint contains exactly one such
variable. Thus the weight of f ′ is at most 2n · n/3 + 1 + n < 2n2 + n, strictly less
than the weight of f . Assignments f and f ′ differ on at most 2n2 +1+n+2n·n/3 =
k variables.
Assume now that φ′ has a satisfying assignment f ′ with w(f ′ ) < w(f ) and
dist(f, f ′ ) ≤ k. We claim that f ′ (y0 ) = 1 and f ′ (yi ) = 0 (1 ≤ i ≤ 2n2 ).
15
Otherwise, w(f ′ ) < w(f ) would imply that f ′ (vt ) = 0 for at least one 1 ≤ t ≤ n.
But this means that there is at least one 1 ≤ i ≤ n such that f ′ (xi,j ) = 1 for
every 1 ≤ j ≤ 2n. Thus the weight of f ′ is at least 2n2 + 2n > 2n2 + n, a
contradiction.
Define f0 (xi ) := f ′ (xi,1 ) for every 1 ≤ i ≤ n. The weight of f0 is at most n/3:
otherwise, dist(f, f ′ ) would be at least 2n(n/3+1)+2n2 +1 > k. Let xa , xb , xc be
the variables in the t-th constraint in φ. The facts f ′ (xa,1 )+ f ′ (xb,1 )+ f ′ (vt ) = 1,
f ′ (xc,1 ) + f ′ (vt ) + f ′ (y0 ) = 1 and f ′ (y0 ) = 1 imply that at least one of f ′ (xa,1 ),
f ′ (xb,1 ), f ′ (xc,1 ) is 1. Thus if we denote by X the set of those variables of φ that
have value 1 in f0 , then each constraint in φ contains at least one variable of X.
Moreover, it is not possible that a constraint contains more than one variables
of X. To see this, observe that each variable is contained in 3 constraints, thus
|X| ≤ n/3 implies that the variables of X appear in at most n constraints.
However, we have seen that each constraint contains at least one variable of X,
hence the variables of X appear in exactly n contraints. Equality is possible only
if any two variables of X appear in disjoint constraints, that is, no constraint
contains more than one variables from X. Therefore, f0 is a solution for the
instance φ.
⊓
⊔
Define the 6-ary affine relation R6 by the following system of equations:
x1 + x2 + x3 = 1, xi + xi+3 = 1(1 ≤ i ≤ 3). Let J be a subset of {1, . . . , 6}, and
let RJ denote the projection of R6 onto the set of indices J.
Lemma 15. LS-CSP({RJ }) is NP-hard for every J ⊆ {1, . . . , 6} containing at
least one number from each pair i, i + 3 where i = 1, 2, 3.
Proof. Note that the relation R6 does not change if the first equation in the
system is replaced by x4 + x2 + x3 = 0. We will assume that 1, 2, 3 ∈ J. All other
cases are very similar. As in the (beginning of the) proof of Lemma 14, we can
argue that LS-CSP({RJ }) is polynomial-time equivalent to LS-CSP({RJ , 6=})
Now the lemma follows from Lemmas 14 and 13.
⊓
⊔
Lemma 16. LS-CSP({R}) is NP-hard if R is affine, but not of width 2.
Proof. First, we show that it is enough to prove the result for Γ = {R, C0 , C1 },
that is, that LS-CSP(Γ ) and LS-CSP({R}) are polynomial-time equivalent. If
R is neither 0- nor 1-valid, then fix a tuple in R that contains both 0 and 1,
say (0, . . . , 0, 1, . . . , 1), and consider the relation R′ (x, y) = R(x, . . . , x, y, . . . , y),
where positions of the x’s correspond to the 0’s in the tuple. It is clear that either
R1 = {(0, 1)} in which case we are done, or R1 is the disequality relation, and
then we are done by Lemma 12. If R is 0-valid, but not 1-valid, then take a tuple
a ∈ R that contains a minimal positive number of 1’s, say (0, . . . , 0, 1, . . . , 1).
The fact that R is affine and 0-valid
P implies that it can be defined by a system
of linear equations of the form
xi = 0. It follows that the number of 1’s in a
is at least two. Consider the instance R(z, . . . , z, x, y, . . . , y), R(z, . . . , z) where
the z’s in the first constraint correspond to the 0’s in a. by the choice of a,
the projection on x, y of the set of solutions of this instance is the equality
16
relation, while the projection on z is C0 . Now we can use Lemmas 11 and 12.
The argument is very similar if R is 1-valid, but not 0-valid. If R is both 0- and
1-valid then take any tuple, say (0, . . . , 0, 1, . . . , 1) 6∈ R and consider the relation
R′ (x, y) = R(x, . . . , x, y, . . . , y), where positions of the x’s correspond to 0’s in
the tuple. Since R′ is affine, it can only be the equality relation, and we are done
again.
Now let R′ be a minimium arity relation which can be obtained from R
by substituting constants and identifying variables (note that such a relation
is also affine) and which is not of width-2. We claim that R′ , possibly after
permuting variables, is a relation of the form RJ , as in Lemma 15. By Lemma 11,
establishing this claim would finish the proof of the present lemma.
Note that by the minimality of (the arity of) R′ , none of the projections of
′
R onto a single coordinate can be one-element, and none of the projections of
R′ onto a pair of coordinates can be the equality relation. Thus every binary
projection is either the disequality relation or {0, 1}2 (note that the binary projection cannot contain 3 tuples, since R′ is affine). Furthermore, if two different
binary projections of R′ are disequality relations then the corresponding pairs
of coordinates are disjoint. Let R′′ be a largest arity projection of R′ such that
every binary projection of R′′ is {0, 1}2. Note that, to prove the lemma, it is
sufficient to show that R′′ is ternary and can be described by a single equation
x1 + x2 + x3 = a, a ∈ {0, 1}. It is easy to see that then a relation of the form
RJ can obtained from R′ by permuting coordinates: every variable of R not in
{x1 , x2 , x3 } forms a disequality relation with one of x1 , x2 , x3 .
Consider the relations R0′′ = R′′ (x1 , . . . , xn , 0) and R1′′ = R′′ (x1 , . . . , xn , 1).
By the choice of R′ , both of these relations are of width-2 affine, that is each
of them can be expressed by a system of equations of the form xi + xj = a,
a ∈ {0, 1}. Note that it is impossible that, for a pair of coordinates i, j, exactly
one of the projections of R0′′ and R1′′ onto i, j is the full relation {0, 1}2 . Indeed,
this would imply that the size of the projection of R′′ onto {i, j} would not be
a power of 2. Therefore, if, for a pair of indices i, j, one of the projections of R0′′
and R1′′ onto i, j is the equality relation then the other must be the disequality
relation (to ensure that the projection of R′′ onto {i, j} is {0, 1}2). It follows
that R′′ can be described by the following system of equations: for each pair
i, j such that pri,j (R0′′ ) is the equality relation the system contains the equation
xi + xj + xn+1 = 0, and for each pair i, j such that pri,j (R0′′ ) is the disequality
relation the system contains the equation xi + xj + xn+1 = 1, and there are no
other equations in the system.
Note that if some xi with 1 ≤ i ≤ n participates in at least two equations in
this system (which also involve xj and xj ′ ) then the projection of R′′ onto j, j ′
would not be {0, 1}2, which is a contradiction. Hence, the only variable that may
appear in more than one equation is xn+1 . Assume that the system contains at
least two equations, say, x1 + x2 + xn+1 = a and x3 + x4 + xn+1 = b. Then, by
identifying x4 with xn+1 in R′ , we would be able to obtain a relation which is
affine but not width-2 (because of the equation x1 + x2 + xn+1 = a), and has
arity smaller than the arity of R′ , which is a contradiction.
17
⊓
⊔
The lemma is proved.
It remains to consider the case when Γ contains a relation that is flip separable, but not affine. Again, we first consider one particular relation.
Lemma 17. LS-CSP({R1-in-3 }) is NP-hard.
Proof. First, we show that we can assume that the relation 6= is available. Consider the instance R1-in-3 (x, y, z0 ), R1-in-3 (z0 , z0 , z1 ). It is easy to see that the
projection on x, y of the set of all solutions of this instance is 6=, and Lemma 11
implies that it is now sufficient to prove that LS-CSP({R1-in-3, 6=}) is NP-hard.
The proof is by reduction from LS-CSP({x ∨ y}). Let (φ1 , f1 , k) be an instance of LS-CSP({x∨y}) with n variables and m constraints. Set S := 10n2 m2 .
We construct a formula φ2 with the following variables:
– For each variable v of φ1 with f (v) = 1, there is a variable x0v and S − 2m
variables x1v,1 , . . . , x1v,S−2m .
– For each variable v of φ1 with f (v) = 0, there is a variable x0v and S variables
x1v,1 , . . . , x1v,S .
– For each constraint c of φ1 , there is a variable yc .
The constraints of φ2 are as follows:
– For every variable v of φ1 with f (v) = 1, there is a constraint x0v 6= x1v,i
(1 ≤ i ≤ S − 2m).
– For every variable v of φ1 with f (v) = 0, there is a constraint x0v 6= x1v,i
(1 ≤ i ≤ S).
– For every constraint c = (u ∨ v) of φ1 , there is a constraint R1-in-3 (x0u , x0v , yc ).
We observe that the satisfying assignments of φ2 are in one-to-one correspondence with the satisfying assignments of φ1 . First, since x0v 6= x1v,i for every i,
we have that x1v,i1 = x1v,i2 for every i1 , i2 . Given a satisfying assignment f of
φ2 , we can obtain an assignment of φ1 by setting f (v) = f (x1v,1 ). Observe that
this gives a satisfying assignment of φ1 : if c = (u ∨ v) is a constraint of φ1 , then
f (u) = f (v) = 0 would imply f (x0u ) = f (x0v ) = 1, which would mean that the
constraint R1-in-3 (x0u , x0v , yc ) is not satisfied. On the other hand, if f satisfies φ1 ,
then setting f (x1v,1 ) = f (v) has a unique satisfying extension to φ2 .
Let f2 be the assignment of φ2 corresponding to assignment f1 of φ1 . Set
k ′ := k(S + 1) + m. We claim that (φ2 , f2 , k ′ ) is a yes-instance if and only if
(φ1 , f1 , k) is a yes-instance.
Let f1′ be an arbitrary satisfying assignment of φ1 and let f2′ be the corresponding assignment of φ2 . Let B (resp., B ′ ) contain variable v of φ1 if and only
if f1 (v) = 1 (resp., f1′ (v) = 1). Assignments f2 and f2′ differ on every variable
x0v , x1v,i with v ∈ |B \ B ′ | ∪ |B ′ \ B|, and possibly on some of the variables yc .
Counting the number of 0’s and 1’s on these variables, it follows that the weight
of f2′ is at least
w(f2 ) + |B ′ \ B|(S − 1) − |B \ B ′ |(S − 2m − 1) − m
18
(1)
and at most
w(f2 ) + |B ′ \ B|(S − 1) − |B \ B ′ |(S − 2m − 1) + m,
(2)
where the ±m term accounts for the variables of the form yc .
To prove that (φ2 , f2 , k ′ ) is a yes-instance if and only if (φ1 , f1 , k) is a yesinstance, first we assume that (φ1 , f1 , k) has a solution f1′ ; let f2′ be the corresponding assignment of φ2 . Define B and B ′ as above. It is easy to see that the
distance of f2 and f2′ is at most |B \ B ′ |(S − 2m + 1) + |B ′ \ B|(S + 1) + m ≤
k(S + 1) + m = k ′ . From |B ′ | < |B| we have |B ′ \ B| < |B \ B ′ |. Thus by (2),
the weight of f2′ is at most
w(f2 ) + |B ′ \ B|(S − 1) − |B \ B ′ |(S − 2m − 1) + m
≤ w(f2 ) + |B ′ \ B|(S − 1) − (|B ′ \ B| + 1)(S − 2m − 1) + m
= w(f2 ) + |B ′ \ B|2m + 3m + 1 − S < w(f2 ),
since S is sufficiently large.
For the second direction, suppose that there is a solution f2′ for instance
(φ2 , f2 , k ′ ). Let f1′ be the corresponding satisfying assignment, and define B and
B ′ as before. First we show that the distance of f1′ and f1 is at most k, i.e.,
|B ′ \ B| + |B \ B ′ | ≤ k. Assignments f2 and f2′ differ on every vertex of x1v,i for
v ∈ |B ′ \B|+|B \B ′ | and 1 ≤ i ≤ S −2m. Thus |B ′ \B|+|B \B ′ | > k would imply
that the distance of f2 and f2′ is at least (k+1)(S −2m) = k(S +1)+S −k−2m >
k(S +1)+m = k ′ , since S is sufficiently large. Let us show that |B ′ \B| < |B \B ′ |.
Suppose first that |B \ B ′ | = 0. By (1), the weight of f2′ is at least
w(f2 ) + |B ′ \ B|(S − 1) − |B \ B ′ |(S − 2m − 1) − m ≥ w(f2 ) + (S − 1) − m ≥ w(f2 ),
contradicting the assumption w(f2′ ) < w(f2 ). Suppose now that |B ′ \ B| ≥
|B \ B ′ | ≥ 1. Now the weight of f2′ is at least
w(f2 ) + |B ′ \ B|(S − 1) − |B \ B ′ |(S − 2m − 1) − m
≥ w(f2 )+|B\B ′ |(S−1)−|B\B ′ |(S−2m−1)−m = w(f2 )+|B\B ′ |2m−m > w(f2 ),
⊓
⊔
again a contradiction.
Consider the 6-ary relation T6 defined as follows: (a1 , . . . , a6 ) ∈ T6 if and only
if (a1 , a2 , a3 ) ∈ R1-in-3 and ai 6= ai+3 (1 ≤ i ≤ 3). Let J be a subset of {1, . . . , 6}
and let TJ denote the projection of T6 onto the set of indices J. Similarly to
Lemma 15, we have:
Lemma 18. LS-CSP({TJ }) is NP-hard if J contains at least one number from
each pair i, i + 3 where i = 1, 2, 3.
Lemma 19. LS-CSP({R}) is NP-hard if R is flip separable but not affine.
19
Proof. Assume first that R is 0-valid. Then we can assume that the relation C0 is
available. Indeed, if R is not 1-valid then R(x, . . . , x) is this relation. If R is also 1valid then we can get C0 and C1 as in the proof of Lemma 16. Note that a relation
obtained from a flip separable relation by substituting constants and identifying
coordinates is again flip separable. It is shown in the proof of Lemma 5.30 of [3]
that, since R is not affine, it is possible to permute coordinates in R in such a
way that the relation R′ (x, y, z) defined as R(0, . . . , 0, x, . . . , x, y, . . . , y, z, . . . , z)
satisfies the following properties:
– The tuples (0, 0, 0), (0, 1, 1), (1, 0, 1) belong to R′ , and
– the tuple (1, 1, 0) does not belong to R′ .
Since R′ is flip separable, it is easy to see that if one of the tuples from R1-in-3
belongs to R′ then all of them do. However, in this case (1, 0, 0) ∈ R′ has flip sets
{1} and {1, 2}, and therefore {2} should also be a flip set for this tuple, which
is impossible because (1, 1, 0) 6∈ R′ . It follows that R′ ∩ R1-in-3 = ∅. If the tuple
(1, 1, 1) was in R′ then it would have flip sets {1, 2, 3} and {1}, and hence {2, 3}
as well, which is impossible. It follows that R′ = {(0, 0, 0), (0, 1, 1), (1, 0, 1)}. It
is easy to see that R′ is obtained from T{2,3,4} (as in Lemma 18) by a cyclic
permutation of coordinates. The lemma now follows from Lemmas 11 and 18.
If R is 1-valid then a similar reasoning shows that one can obtain the relation {(1, 1, 1), (0, 0, 1), (0, 1, 0)} which is the relation T{3,4,5} with permuted
coordinates.
Assume now that R is neither 0- nor 1-valid. As in the proof of Lemma 16,
one can show that both C0 and C1 can be assumed to be available. It is shown
in the proof of Lemma 5.30 of [3] that if R is Horn then there is an instance
of CSP({R, C0 , C1 }) over variables {x, y, z, u0, u1 } which contains constraints
C0 (u0 ) and C1 (u1 ), and such that the projection on x, y, z of the set of all
solutions to the instance contains (0, 0, 0), (0, 1, 1), (1, 0, 1), but does not contain
(1, 1, 0). This set must be a Horn relation. On the other hand, we showed above
that this set must coincide with {(0, 0, 0), (0, 1, 1), (1, 0, 1)}, which is not Horn,
a contradiction. If R is not Horn, then it is easy to show (see the proof of the
Lemma 5.26 of [3]) that there is an instance of CSP({R, C0 , C1 }) over variables
{x, y, u0 , u1 } which contains constraints C0 (u0 ) and C1 (u1 ), and such that the
projection on x, y is not {0, 1}2 and contains the relation 6=. Since this set must
be a flip separable relation, it is equal to 6=. Again, it is shown in the proof
of Lemma 5.30 of [3] that there exists an instance of CSP({R, 6=, C0 , C1 }) over
variables {x, y, z, x′ , y ′ , z ′ , u0 , u1 } which contains constraints x 6= x′ , y 6= y ′ ,
z 6= z ′ , C0 (u0 ) and C1 (u1 ), and such that the projection on x, y, z of the set R′′
of all solutions to the instance contains (0, 0, 0), (0, 1, 1), (1, 0, 1), but does not
contain (1, 1, 0). Since R′′ is flip separable, it is easy to show (as above) that,
in fact, the projection of R′′ onto x, y, z, x′ , y ′ , z ′ is equal to T6 (with permuted
coordiantes). Now the result follows from by Lemmas 11 and 18.
⊓
⊔
20
| 8cs.DS |
MAXDIM OF SOME SIMPLE GROUPS
arXiv:1712.04553v1 [math.GR] 12 Dec 2017
LIU TIANYUE AND R. KEITH DENNIS
1. Introduction
There are various generalizations of the concept of dimension for vector spaces
in finite group theory. In this paper we are primarily concerned with three of them,
namely m(G), M axDim(G) and i(G).
Definition 1.1. A list\
of subgroups
\ {H1 , ..., Hn } are said to be in general position
if for all i ∈ {1, ..., n},
Hj
Hj .
j6=i
j
(i) M axDim(G) of a finite group G is defined as the maximal size of a list of
G’s maximal subgroups in general position.
(ii) m(G) is defined as the maximal size of an irredundant generating sequence
in G
(iii) i(G) is defined as the maximal size of an irredundant sequence in G
The basic relation between those group invariants is captured by the following
theorem.
Theorem 1.2. For a finite group G, m(G) ≤ M axDim(G) ≤ i(G)
M axDim, m and i can also be defined using the subgroup lattice. For a set [n]
with n elements, the power set P ([n]) has a partial order induced by inclusion. We
shall call a poset T a tower with dimension n if it is isomorphic to P ([n]), and call
the pre-images of (n − i)-element sets in P [n] the i-th level of the tower. In this
context, we can see that the following proposition holds.
Proposition 1.
(i) i(G) is the dimension of the largest tower embeddable in
the subgroup lattice of G.
(ii) M axDim(G) is the dimension of the largest tower embeddable in the subgroup lattice of G, so that all first level elements of the tower are maximal
subgroups. That is, for a first level element S, A ≥ S ⇒ A = S or A = G.
(iii) m(G) is the dimension of the largest tower embeddable in the subgroup
lattice of G, so that the join of all (n − 1)-th level elements of the tower is
G, and all (n − 1)-th level elements are the joins of the n-th level element
and cyclic.
Before we proceed to the results on the M axDim of simple groups, we shall
state some useful lemma here.
Lemma 1.3. m(G) ≤ max{i(M )|M is maximal subgroup of G}+1
i(G) = max{m(G), max{i(M )|M is maximal subgroup of G}}
Date: July 2017.
1
2
LIU TIANYUE AND R. KEITH DENNIS
2. The Suzuki Groups
In this section we compute m and MaxDim for some Suzuki groups of Lie type.
The main result was almost implied by a proof in the thesis of Whiston, although
he did not state the claim in his paper.
Proposition 2. If q is a prime power of 2, then Suz(q) has M axDim = m = 3.
Proof. To prove the proposition, we need to use the following theorem of Suzuki
[6]:
Theorem 2.1. Suz(q) acts 2-transitively on q 2 + 1 points. The maximal subgroups
of Suz(q) fall in one of the conjugacy classes listed below:
(i)
(ii)
(iii)
(iv)
(v)
Stabilizers of a point, Eq .Eq : Cq−1
Stabilizers of a two point subset, D2(q−1)
Cq+√2q+1 : C4
Cq−√2q+1 : C4
Suz(q0 ) when q = q0r , r is a prime number and q0 is not 2
The general strategy behind the proof of Prop. 1 is to show on one hand, since
Suzuki groups do not satisfy the B-properties, m ≥ 3; on the other hand, the
structure of the maximal subgroups does not allow M axDim to exceed 3. We
shall prove the upper bound for MaxDim by inspecting the intersections of those
maximal subgroups.
Proposition 3. q is an odd power of 2, A, B are maximal subgroups of Suz(q).
The intersection I := A ∩ B is:
(1)
(2)
(3)
(4)
(5)
(6)
Both A, B are of type (iii) or (iv), then I is a subgroup of C4
A is of type (ii) and B of type (iii) or (iv), then I is a subgroup of C2
A is of type (i) and B of type (iii) or (iv), then I is a subgroup of C4
Both A, B are of type (ii), then I is a subgroup of C2
Both A, B are of type (i), then I is isomorphic to Cq−1
A is of type (i) and B of type (ii), then I is isomorphic to either Cq−1 or
a subgroup of C2
To prove claim (1) ∼ (4) we need the following lemma:
Proposition 4. If F := Cn oB, G := Cm oB 0 are distinct maximal in simple group
S, gcd(|B|, m) = gcd(|B 0 |, n) = 1, then the intersection I = F ∩ G is isomorphic to
a common subgroup of H and H 0 , where H ∼
= B and H 0 ∼
= B0.
Proof. Assume there is a nontrivial x ∈ Cn lying inside the intersection I, then
x ∈ Cm (and vice versa) because any other element in G would have order dividing
|B 0 |, while as an element of Cn , the order of x cannot divide anything relatively
prime to n.
On one hand, if we assume Cn is not contained in the intersection, then there
must be a nontrivial element z ∈ Cn \I. Since G is maximal in S, S =< G, {z} >.
But z and x lie in the same cyclic group Cn , so z < x > z −1 =< x >. Moreover, by
the cyclicity of Cm , < x > is normal in G. That is to say, g < x > g −1 =< x > for
any finite word g using z and elements in G. Thus < x > must be a proper normal
subgroup of S, which contradicts S being a simple group. If Cm is not contained
in the intersection, we yield the same contradiction.
MAXDIM OF SOME SIMPLE GROUPS
3
On the other hand, if Cn , Cm ⊆ I then the normalizer of K := Cn ∩ Cm must
contain both F and G, so S = NS (K), also contradicting the simplicity of S. Thus
the intersection I must intersect trivially with Cn or Cm .
Write elements in F as ordered pairs. Since I intersects Cn trivially, for any
y ∈ B there is at most 1 element (x, y) ∈ I, so the homomorphism (x, y) 7→ y is
injective. Similarly, the homomorphism (x0 , y 0 ) ∈ G 7→ y 0 ∈ B 0 is also injective. I
is isomorphic to a common subgroup of B and B 0 .
√
√
Thus, since both q−1 and q± 2q+1 are odd,
√ we have that gcd(q± 2q+1, 4) = 1
implies claim (1), gcd(q − 1, 4) = gcd(q ± 2q + 1, 2) = 1 implies claim (2), and
gcd(q − 1, 2) = 1√implies claim (4). Claim (3) follows from Whiston’s result, namely
since gcd(4(q ± 2q + 1), q 2 (q − 1)) = 4, the order of the intersection must divide 4
[6]. However, a Klein four group cannot be embedded into a subgroup of type (iii)
or (iv), so the intersection must be a subgroup of C4 .
The intersection of two distinct point stabilizers, say Stab(a) and Stab(b), is
equivalent to the stabilizer of a in the pair-stabilizer Stab({a, b}). The orbit of
a in Stab({a, b}) is {a, b}, so by the Orbit-Stabilizer Theorem, the intersection of
Stab(a) and Stab(b) has index 2 in dihedral group Stab({a, b}) ∼
= D2(q−1) . Thus
the intersection is cyclic of order q − 1, and claim (5) is proven.
If I is the intersection of Stab(a) and Stab({u, v}), then there are two cases. If
a ∈ {u, v} then I ∼
/ {u, v}, then we need the
= Stab(u) ∩ Stab(v) ∼
= Cq−1 . If a ∈
following lemma of Huppert and Blackburn [1]:
Proposition 5. Any nontrivial element of Suz(q) has at most 2 fixed points.
That is to say, the intersection of any three point stabilizer must be trivial. If
a ∈
/ {u, v}, then since Stab(u) ∩ Stab(v) ∩ Stab(a) has index at most 2 in I :=
Stab(a) ∩ Stab({u, v}), I must be a subgroup of C2 and claim (6) holds.
Since any subgroup of C4 or C2 has i ≤ 1, a length 4 list lof maximal subgroups in
general position must include no subgroup of type (iii) or (iv), at most 1 subgroup
of type (ii), and at most 3 point stabilizers. Since 3-fold intersection of point
stabilizers is trivial, they cannot coexist with another dihedral group in the list, so
there cannot be a such a length 4 list l. That is to say, 3 = m = M axDim = 3.
3. The Suzuki Tower
The Suzuki tower is a family of five simple groups: P SL2 (7) < P SU3 (3) < J2 <
G2 (4) < Suz. Here J2 is the Hall-Janko group, G2 (4) is an exceptional group of
Lie type, and Suz is the sporadic Suzuki group. Each of them act transitively on
a distance transitive graph, denoted by Γ1 to Γ5 respectively. Moreover, for i =
1, 2, 3, 4, Γi are the local subgraphs of Γi+1 . That is to say, for any vertex v ∈ Γi+1 ,
the subgraph induced by v’s neighbors is Γi . When the groups in the Suzuki tower
acts as automorphisms of these graphs, the vertex stabilizers are isomorphic to
the group one level beneath it in the tower [7]. Using these information and some
knowledge on the structure of Γ1 , we can give an inductive argument to bound the
M axDim of these groups from below.
Proposition 6. The M axDim of P SU3 (3), J2 , G2 (4), and Suz are greater or
equal to 4, 5, 6 and 7 respectively.
Proof. As mentioned by Brouwer, Γ1 is the co-Heawood graph on 14 vertices, which
is the point-line non incidence graph of the Fano plane. In other words, it is the
4
LIU TIANYUE AND R. KEITH DENNIS
Figure 1. Heawood Graph
distance-3 graph of the Heawood graph in Figure 1 [3]. P SL2 (7) acts on it with
rank 4, and the stabilizer of a vertex v3 (e.g. the one labeled with a square in Figure
1) is S4 , which has vertex orbit size 1 + 4 + 6 + 3. The 6 orbit corresponds to the
vertices at distance 2 with v3 . Stabilizing another point in v3 ’s neighborhood (those
vertices labeled with circles in Figure 1) gives S3 , and stabilizing a point in the 6
orbit yields the Klein 4 group C2 × C2 . Stabilizing 2 points in v3 ’s neighborhood
would result in C2 , because the one vertex stabilizer S4 acts on the four neighbors
as the permutation group S4 . Notice that v3 ’s neighbors are at distance 2 with
each other in the co-Heawood graph.
Now to bound MaxDim of a Suzuki tower group at level i, i > 1, we can pick an
arbitrary vertex vi+2 first. Then we keep picking vertices in the common neighborhood of previously picked vertices, until the common neighborhood becomes the
co-Heawood graph. Since each time we add a new point the common neighborhood
induces a graph one level beneath the previous graph in the tower, we should have
collected i − 1 vertices so far. Pick a vertex in the co-Heawood graph and denote
it v3 . then pick two more vertices in v3 ’s neighborhood and denote them by v2 , v1 .
I claim that the point stabilizers of {v1 , ..., vi+2 } in Γi are maximal subgroups in
general position. The intersection of all those point stabilizers is the same as the
point-wise stabilizer of v1 , v2 , and v3 in the co-Heawood graph, so it must be C2 .
Meanwhile, the subgraph of Γi induced by {v1 , ..., vi+2 } is a complete i + 2 graph
with one edge removed, and its subgraphs induced on any i+1 vertices are complete
i + 1 graphs with or without an edge removed. If the subgraph on i + 1 vertices is
a complete graph, then the intersection of corresponding stabilizers is isomorphic
to S3 , i.e. the stabilizer of a vertex and one of its neighbor in the co-Heawood
graph. Otherwise, the instersection of point stabilizers must be C2 × C2 , which is
the point-wise stabilizer of two vertices at distance 2 in the co-Heawood graph. In
either case, the i + 1-fold intersections contains the overall intersection properly.
thus by definition the i + 2 point stabilizer are in general position.
Remark 3.1. We actually can prove a stronger result here, namely the lower bounds
for M axDim are the lower bonds form as well. Let {v1 , ..., vi+2 } be picked as
MAXDIM OF SOME SIMPLE GROUPS
5
mentioned above, and denote the two neighbors of v3 not in the list by a and b. Then
there are two elements σ, τ in the point-wise stabilizer Stab({v3 , ..., vi+2 }) ∼
= S4 so
that σ : a 7→ b, b 7→ v1 , v1 7→ a, v2 7→ v2 and τ : a 7→ v2 , v2 7→ a, b 7→ b, v1 7→ v1 . σ
and τ generates an S4 , which is exactly the point-wise
stabilizer
\
\ Stab({v3 , ..., vi+2 }).
Thus if we pick any α3 , ..., αi+2 so that αj ∈
Stab(vk )\ Stab(vk ), then since
k6=j
k
for all j ≥ 2, Stab({vj , ..., vi+2 }) is maximal in Stab({vj+1 , ..., vi+2 }), we know
that l := {σ, τ, α3 , ..., αi+2 } generates the entire group. Since sequence l certifies a
list of maximal subgroups in general position, it must be irredundant, so we know
m ≥ i + 2 for the group at level i. In particular, since the intersection of the point
stabilizers is C2 , applying a lemma of Keith Dennis tells us P SU3 (3) and J2 fails
the replacement property.
Remark 3.2. The lower bounds of M axDim are strict for P SU3 (3) and J2 , which
has been checked computationally.
4. The Mathieu Groups
In this section we give a computer-aided proof for the M axDim of Mathieu
groups.
Proposition 7. The M axDim of M11 , M12 , M22 , M23 , M24 are 5, 6, 6, 6, and 7
respectively.
Proof. M axDim for M11 and M12 has already been calculated by Brooks [2].
M22 , M23 , M24 can be defined as automorphism groups of Steiner systems. A
Steiner system (a, b, v) is a set S on v points with a collection of subsets B called
”blocks”, so that any block in B contains exactly b points and any a points are
contained in exactly one block [7]. In this section We shall utilize the axioms of
Steiner systems to prove our claim on M axDim.
We know that M24 is the automorphism group of the unique Steiner (5,8,24)
system, and is 5-transitive on the 24 points. M23 is the stabilizer of a point in
M24 . As will be shown later, we can find a list of 7 point stabilizers in M24
that are in general position. If {Stab(a1 ), ..., Stab(a7 )} are in general position in
M24 , then {Stab(a1 ) ∩ Stab(a2 ), ..., Stab(a1 ) ∩ Stab(a7 )} are in general position
in Stab(a1 ) ∼
= M23 . Since M23 is 4-transitive on 23 points, point stabilizers are
maximal in M23 . Thus {Stab(a1 ) ∩ Stab(a2 ), ..., Stab(a1 ) ∩ Stab(a7 )} is a list of
maximal subgroups in general position, so M axDim(M23 ) ≥ 6. Applying the same
argument to M22 the point stabilizer in M23 , we know that M axDim(M22 ) ≥ 5.
However, as suggested by Sophie Le’s computational results, for M22 there is a
way to construct a longer list of maximal subgroups in general position. Pick 2
points {a, b} in Steiner (5, 8, 24) system, the 22 points leftover together with the
blocks that contain {a, b} form a Steiner (3, 6, 22) system. Pick 3 points {α1 , α2 , α3 }
in the 22 points, a block O1 of S(5, 8, 24) that contains {a, b, α1 , α2 } , another block
O2 that contains {a, b, α3 } and intersect the first block at only 2 points {a, b}, and
a third block O3 that contains {a, α1 , α2 , α3 } and intersect both B1 and B2 at 4point sets. It turns out that M := Stab(a) ∩ Stab(b) ∼
= M22 and M ∩ Stab(α1 ), M ∩
Stab(α2 ), M ∩ Stab(α3 ), M ∩ Stab(B1 ), M ∩ Stab(B2 ), M ∩ Stab(B3 ) are maximal
in M and are in general position.
Sophie Le has checked computationally that M axDim(M22 ) = M axDim(M23 ) =
6.
6
LIU TIANYUE AND R. KEITH DENNIS
In a Steiner (5,8,24) system, a block is termed as an ”octad”. The stabilizer of
an octad in M24 is C24 : A8 , where A8 permutes the 8 points in the octad and each
involution in C24 moves all 16 points outside the octad and stabilizes the points
inside the octad [5]. Thus we can use point stabilizers again to give a lower bound
of M axDim. Pick an octad in the Steiner system, then pick 6 points a1 , ..., a6
inside the block and one point a7 outside the block. The stabilizers of them are
then in general position. We shall prove this by showing that on one hand, the
point-wise stabilizer of the 7 points is trivial, and on the other hand, point-wise
stabilizers of any 6 points are either C3 or C24 .
Since any five points determine a unique octad, a sixth point can either sit inside
or outside the octad. Let the stabilizer of that octad be denoted O ∼
= E : A, where
E∼
= A8 is alternating on 8 points.
= E16 is elementary abelian of order 16 and A ∼
Let πA : O → A8 be the projection onto the permutation group of points in the
octad, ιE : E16 → O be the injection from the normal subgroup, ιA : A8 → O be
the splitting homomorphism, and the point-wise stabilizer of 6 points be denoted
as I. If the sixth point sits inside the octad, then since any involution in the normal
∼
subgroup E stabilizes the points in the octad, ι−1
E (I) = E16 . Since the nontrivial
permutation on two points is odd, πA (I) must be trivial, thus I ∼
= E16 . If the
sixth point sits outside the octad, then all involutions in E cannot be contained
in I, and there is a permutation p in A8 that shuffles the three points left over in
the octad. Since O is a split extension, ιA (p) is an order 3 element in I. Since
πA (I) ∼
= C3 , ι−1
E (I) is trivial, I must be isomorphic to C3 . In either case, six-fold
intersections of point stabilizers are non-trivial. However, if R is the intersection of
∼
∼
all 7 point stabilizers, then ι−1
E (R) = πA (R) = 1, so R must be trivial. Thus the
point stabilizers are in general position and M axDim(M24 ) ≥ 7.
To obtain an upper bound, we checked computationally that all maximal subgroups of M24 have i ≤ 6 except the dyad groups M22 : 2, the octad groups C24 : A8
and the trio groups C26 : (P SL3 (2) × S3 ), all of which have i = 7. Thus i(M24 ) ≤ 8
and if there is a length 8 list l of maximal subgroups in general position, then only
dyad, octad, and trio subgroups may occur in that list.
The dyad subgroup is the stabilizer of a two point set (i.e. a dyad). Since M24 is
5-transitive, the intersection of two dyad groups falls into two conjugacy classes: if
the dyads intersect, then the intersection is isomorphic to P SL3 (4), and if the dyads
are disjoint, then the intersection is isomorphic to C25 : A5 : C2 : C2 . Both of them
have i ≤ 5, so they cannot be in a length 8 list of Maximal subgroups in general
position. That is, in a length 8 list l, there can be at most 1 dyad stabilizer. In the
following paragraphs, the high transitivity of M24 will be used extensively to show
that beside the potential dyad stabilizer, we can use at most 2 octad stabilizers and
4 trio stabilizers, thus l cannot exist.
Lemma 4.1. Had there been a length 8 list l of maximal subgroups in general
position, then octad stabilizers can appear at most twice in l.
Proof. Two distinct octads in the Steiner (5,8,24) system S(5, 8, 24) may be disjoint
or intersect at a set of size 2 or 4. Any pair of octads that intersect at a 2set are congruent to one another pair under the action of M24 . To see this, we
can denote the 2-point intersections in each pair {a, b} and {a0 , b0 } respectively.
The octads involved in each pair shall be denoted O1 , O2 and O10 , O20 . Then using
the 5-transitivity of M24 we can find an element σ1 that sends a to a0 and b to
b0 . The point-wise stabilizer of a0 and b0 is M22 , a group of automorphisms of
MAXDIM OF SOME SIMPLE GROUPS
7
the Steiner (3,6,22) system induced by fixing a0 , b0 in S(5, 8, 24). As Brouwer
mentioned, if we construct a M22 graph by connecting the 6-point blocks a Steiner
(3,6,22) system when they are disjoint, then M22 would act edge-transitively on
the graph [3]. That is to say, since the 6-point blocks are constructed by removing
{a0 , b0 } from the 77 octads in S(5, 8, 24) containing {a0 , b0 }, there is an element σ2
in the point-wise stabilizer of a0 and b0 that maps σ1 (O1 ) to O10 and σ1 (O2 ) to O20 .
Thus σ1 σ2 maps one pair of octad to the other pair, and it implies that Stab(O1 )
(resp.Stab(O2 )) can be mapped to Stab(O10 ) (resp.Stab(O20 )) under conjugation by
σ1 σ2 . As a consequence, we only need to check one pair of octad stabilizers to know
the intersection of any pair (Stab(O1 ), Stab(O2 )) where |O1 ∩ O2 |=2. It turns out
that the intersection is isomorphic to S6 , which has i = 5. Thus in l two octad
stabilizers must have their octads either disjoint or intersecting at a 4-point set.
Moreover, if an octad is split into a 4+2+2 partition by the blocks of a trio
(i.e. a partition of the 24 points into 3 octads), then the intersection of the octad
stabilizer with the trio stabilizer is (C2 × C2 × A4 ) : 2, which has i = 4, so they
cannot coexist in l. As in the previous case, by using transitivity we only need to
check one example in order to cover the entire case. Indeed, every octad-trio pair
with this pattern of intersection are congruent to each other under the action of
M24 , so the intersections of their stabilizers are all isomorphic. This claimed can
be inferred from the 2+1+1 transitivity of the trio group [5]. To prove the claim,
denote the octad-trio pairs as (O, T ) (resp. O0 , T 0 ) and the 3 octad blocks in T as
O1 , O2 , O3 (resp. O10 through O30 ), where |O1 ∩ O| = 4. Let {a1 , a2 } (resp. a01 , a02 )
be the two points in O ∩ O2 (resp. O ∩ O20 ) and a3 (resp. a03 ) be in O ∩ O1 (resp.
O ∩ O10 ), a4 (resp. a04 ) be in O ∩ O3 (resp. O ∩ O30 ). Since all trio groups are
conjugate to each other, there is a σ1 that maps trio T to T 0 . Since a trio group
acts as S3 on its three octads, there is an element σ2 that maps octad σ1 (O1 ) to O10
and σ1 (O2 ) to O20 . Since the trio group is 2+1+1 transitive after stabilizing each
octad in the trio individually, there is a σ3 that takes σ2 σ1 ({ai }) to {a0i }. Now we
have σ3 σ2 σ1 (O ∩ O2 ) = O0 ∩ O20 = {a01 , a02 }. If σ3 σ2 σ1 (O ∩ O3 ) = O0 ∩ O30 , then since
(O0 ∩ (O20 ∪ O30 )) ∪ {a03 } determines a unique octad, we have σ3 σ2 σ1 (O) = O0 . If
not, then I := σ3 σ2 σ1 (O ∩ O3 ) ∩ (O0 ∩ O30 ) must have only 1 element to ensure that
|O0 ∩ σ3 σ2 σ1 (O)| is less than 5. But in that case, the symmetric difference octad
O0 O10 would intersect σ3 σ2 σ1 (O) in a 6 point set (σ3 σ2 σ1 (O)\{a3 }) ∪ {a1 , a2 , a4 },
contradiction. Thus σ3 σ2 σ1 (O) = O0 and σ3 σ2 σ1 maps the first octad-trio pair to
the second.
If three octad stabilizers are in general position then their octads cannot be
mutually disjoint, for by stabilizing two disjoint octads we automatically stabilize
the third one. Thus there must be a pair of octad stabilizers intersecting at a tetrad
(i.e. 4-point set). They will be denoted O1 and O2 . In S(5, 8, 24), stabilizing a
tetrad implies stabilizing a sextet (i.e. a partition of the 24 points into 6 tetrads).
The tetrads in the sextet would be denoted as T et1 , ..., T et6 , where T et1 := O1 ∩O2 ,
T et2 := O1 − O2 , and T et3 := O2 − O1 . Adding a third octad stabilizer into the list
either preserves the tetrads in the sextet or refines them. However, the third octad
O3 must not intersect any tetrad in odd number of points. If |O3 ∩ T eti | = 1 for
some i, then since whenever i 6= j, T eti ∪ T etj is octad and the intersection of any
two distinct octads cannot have size other than 0, 2, or 4, |O3 ∩ T etj | = 1 or 3 for
all i 6= j. Because of the pigeon hole principal, we can then consider only the case
where |O3 ∩ T eti | = 3 for some i without loss of generality. If |O3 ∩ T eti | = 3 for
8
LIU TIANYUE AND R. KEITH DENNIS
some i, then |O3 ∩ T etj | = 1 for all i 6= j. Since we need |O3 ∩ O1 | = |O3 ∩ O2 | = 4,
i must be 1. But then we have stabilized 2 points inside and 1 point outside O1 ,
namely we have stabilized T et1 −O3 , T et2 ∩O3 , and T et3 ∩O3 . Thus the intersection
of such three octad stabilizers must be isomorphic to a subgroup of A6 , and by the
flatness of alternating groups it has i ≤ 4 [4]. If we want the list l to have length
8, any three elements of it must have an intersection with i ≥ 5, so |O3 ∩ T eti | = 2
or 4 for all i.
If the third octad O3 splits one of the tetrads into two dyads, then it must
split T et1 , ..., T et3 as well as one more tetrad T et4 into dyads. This is because
the 8 points in O3 must be split into 4 dyads, so it must intersect with O1 ∪
O2 . Since it must not intersect O1 or O2 in only 2 points, its intersection with
T et1 , ..., T et3 must all have size 2. Moreover, the sextet induced by |O3 ∩O1 | consists
of tetrads that their intersection with T et1 , ..., T et6 could only have size 0 or 2. Let
the tetrads in this sextet be denoted as T et01 , ..., T et06 , where T et01 := O1 ∩ O3 ,
T et02 := O1 − O3 , and T et03 := O3 − O1 . The tetrad containing T et3 − O3 shall
be denoted as T et04 . We know T et04 = T et3 − O3 ∪ T et4 − O3 , because otherwise
since there is an i that T et0i ⊂ T et5 ∪ T et6 , (T et04 ∪ T et0i ) ∩ (T et5 ∪ T et6 ) would
have cardinality 6, which is impossible. Therefore, T et5 ∪ T et6 = T et05 ∪ T et06 . Now
we shall show that |Stab(O1 ) ∩ Stab(O2 ) ∩ Stab(O3 )| ≤ 16 using those information.
Pick a point p in T et5 ∩ T et05 and another point q in T et4 ∩ T et04 and denote
Stab(p) ∩ Stab(q) ∩ Stab(O1 ) ∩ Stab(O2 ) ∩ Stab(O3 ) by I. Under the action of
Stab(O1 )∩Stab(O2 )∩Stab(O3 ), q can only be map to the other point in T et4 ∩T et04
and p can only be mapped to points in the octad T et5 ∪ T et6 , so I must have
index less than 16 in Stab(O1 ) ∩ Stab(O2 ) ∩ Stab(O3 ). But I must be trivial, for by
stabilizing p we have stabilized all 12 dyads that are the intersections of T eti ∩T et0j ,
therefore every nontrivial element in I must be involutions. Since we have stabilized
an octad T et5 ∪ T et6 , I must act as an even permutation on the octad. If there is a
nontrivial element σ in I, then it must permute the two points in T et5 − T et05 or the
two points in T et05 − T et5 . In fact, they must permute both of them simultaneously
to ensure that σ is an even permutation on both T et4 ∪ T et5 and T et4 ∪ T et6 . As
a result σ must stabilize the two points in T et6 ∩ T et06 and T et3 ∩ T et03 and move
the points in T et4 ∩ T et03 . But an involution in M24 moves either 16 or 24 points
[5], so σ must move every point in O1 , thus its action on T et1 ∪ T et4 is an odd
permutation, contradiction. Since i(I) is less than the number of prime factors in
I, i(I) ≤ 4, so we must avoid this pattern as well.
As a consequence, if three octad stabilizers are in general position, then their
pairwise intersections must have cardinality 4, so O3 is the union of two tetrads
in {T et1 , ..., T et6 }. We can add at most one more octad stabilizer into the list,
for four octad stabilizers in general position would have stabilized all six tetrads
individually. Meanwhile, we can add at most one trio stabilizer into the list. This
is because on one hand, if the octads inside the trio are all unions of tetrads in
{T et1 , ..., T et6 }, then T et5 , T et6 must be in different octads, so adding that trio
stabilizer stabilizes all the 6 tetrads, and no more such trio stabilizers can be added.
On the other hand, we can consider the case where one of the three octads in
trio Õ1 , Õ2 , Õ3 , say Õ1 , intersect with a tetrad at a set of cardinality 2, or 3.
Intersecting with a tetrad at one point implies intersecting with another one at a
set of cardinality 3, so without loss of generality we can omit the discussion of that
case. If |Õ1 ∩ T eti | = 3, then to ensure the intersection of Õ1 with O1 , O2 , O3 all
MAXDIM OF SOME SIMPLE GROUPS
9
have cardinality 0 or 4, we must have i = 1 and T et1 ⊂ O3 . But if we denote the
octad containing T et1 − Õ1 as Õ2 , then Õ2 ∩ T et2 and Õ2 ∩ T et3 must all have
cardinality 3 to ensure that |Õ2 ∩ O1 | = |Õ2 ∩ O2 | = 4, so |Õ2 ∩ (T et2 ∪ T et3 )| = 6,
contradiction. If |Õ1 ∩ T eti | = 2, then again using the same pigeon hole argument
we can infer that |Õ1 ∩ T eti | = |Õ2 ∩ T eti | = 2 for i = 1, 2, 3, 4, where T et4 ⊂ O3 .
But then an arbitrary element σ in the intersection of Stab(O1 ), Stab(O2 ) and the
trio stabilizer would not be able to move an element in O3 outside, so the three
octad stabilizers and the trio stabilizer are not in general position. Therefore, the
only trio stabilizers that can be added in consist of octads that are unions of tetrads
in {T et1 , ..., T et6 }. Thus if there are three octad stabilizers in the list l, then we can
add at most 1 dyad stabilizer, 1 octad stabilizer and 1 trio stabilizer into the list
(Actually, at most 1 octad stabilizer or 1 trio stabilizer along with a potential dyad
stabilizer), which means the list has size less than 8. It then follows immediately
that we can use at most two octad stabilizers in l.
Lemma 4.2. Had there been a length 8 list l of maximal subgroups in general
position, then trio stabilizers can appear at most 4 times in l.
Proof. Let us pick a trio {Õ1 , Õ2 , Õ3 } in the Steiner system. A second trio {Q̃1 , Q̃2 , Q̃3 }
may intersect with the first trio in one of the following four ways:
(i) All of Q̃1 , Q̃2 , Q̃3 are split into 4+2+2 partitions.
(ii) Q̃1 , Q̃2 are split into 4+2+2 partitions, and Q̃3 is split into two tetrads.
(iii) All of Q̃1 , Q̃2 , Q̃3 are split into tetrads.
(iv) Q̃1 overlaps with Õ1 , and Q̃2 , Q̃3 are split into tetrads.
In this section we shall denote a pair of trios as T rio1 = {Õ1 , Õ2 , Õ3 }, T rio2 =
{Q̃1 , Q̃2 , Q̃3 } and in the case we are discussing more than one pairs, the second
pair as T rio01 = {Õ10 , Õ20 , Õ30 }, T rio02 = {Q̃01 , Q̃02 , Q̃03 }. The intersection of the two
trio stabilizers shall be denoted as I(T rio1 , T rio2 ).
Let the type (i) pairs be considered the first. Assume without loss of generality
that |Õi ∩ Q̃i | = |Õi0 ∩ Q̃0i | = 4 for i = 1, 2, 3. Since Õi ∩ Q̃i can only be mapped
to Õj ∩ Q̃j for i, j = 1, 2, 3, I(T rio1 , T rio2 ) acts as a subgroup of S3 on {Õi ∩
Q̃i |i = 1, 2, 3}. (Õi ∩ Q̃i ) ∪ (Õj ∩ Q̃j ) cannot be octads in order to ensure that
(Q̃i − Õi ) ∪ (Õj ∩ Q̃j ) ∩ Õj has order less than 5, so if we find the stabilizers
Stab(Sex1 ) (resp.Stab(Sex2 )) of the sextet containing Õ1 ∩ Q̃1 (resp.Õ2 ∩ Q̃2 ), then
Stab(Sex1 ) ∩ Stab(Sex2 ) ∩ I(T rio1 , T rio2 ), which is the stabilizer of each Õi ∩ Q̃i ,
must have index less than 6 in I(T rio1 , T rio2 ). Let the tetrads in Sex1 be denoted
as T et1 , ..., T et6 . We know that |T eti ∩ (Õ2 ∩ Q̃2 )| can only be 0, 1, or 2 to ensure
|((Q̃1 − Õ1 ) ∪ T eti ) ∩ Õ2 |=2 or 4. But if there is no i that |T eti ∩ (Õ2 ∩ Q̃2 )| = 1,
then there must be two tetrads so that T eti ∪ T etj = (Õ2 ∩ Q̃2 ) ∪ (Õ3 ∩ Q̃3 ) in
order to ensure that |((Õ1 − Q̃1 ) ∪ T eti ) ∩ Q̃2 | = |((Õ1 − Q̃1 ) ∪ T etj ) ∩ Q̃2 | = 4.
In that case, (Õ2 ∩ Q̃2 ) ∪ (Õ3 ∩ Q̃3 ) forms an octad, which is impossible. Thus
both Õ2 ∩ Q̃2 and Õ3 ∩ Q̃3 are split into blocks of size 2+1+1 by tetrads in Sex1 .
In particular, there is a tetrad T et4 that intersect both Õ2 ∩ Q̃2 and Õ3 ∩ Q̃3
at 2-sets and two tetrads T et5 , T et6 that intersect both Õ2 ∩ Q̃2 and Õ3 ∩ Q̃3
at a single point. Pick one point p ∈ T et4 ∩ Õ2 ∩ Q̃2 and another point q ∈
T et5 ∩ Õ2 ∩ Q̃2 , then Stab(p) ∩ Stab(q) ∩ Stab(Sex1 ) ∩ Stab(Sex2 ) ∩ I(T rio1 , T rio2 )
must be trivial, for by doing so we have stabilized an octad Õ2 , 6 points inside
({p}, {q}, Õ2 ∩ T et4 − {p}, Õ2 ∩ Q̃2 ∩ T et6 , Õ2 ∩ T et5 − {q}, Õ2 ∩ T et6 − Q̃2 ) and
10
LIU TIANYUE AND R. KEITH DENNIS
one point outside (say, Õ3 ∩ Q̃2 ∩ T et5 ) the octad. Since p can only possibly be
mapped to the other point in Õ2 ∩T et4 −{p} and q can only be mapped to the point
in Õ2 ∩ Q̃2 ∩ T et6 by elements in Stab(Sex1 ) ∩ Stab(Sex2 ) ∩ I(T rio1 , T rio2 ), the
trivial group must have index less than or equal to 24 in I(T rio1 , T rio2 ). But then
I(T rio1 , T rio2 ) cannot have more than 4 prime factors, so i(I(T rio1 , T rio2 )) ≤ 4,
such a pair of trios stabilizers cannot appear in a list of length 8.
A pair of trio stabilizers of type (ii) cannot appear either. In fact, if {T rio1 , T rio2 }
and {T rio01 , T rio02 } are such trio pairs, then we can show that I(T rio1 , T rio2 )
is conjugate to I(T rio01 , T rio02 ). Assume without loss of generality that Õ1 , Q̃3
(resp.Õ10 , Q̃3 ’) are split into two tetrads. DenoteÕ1 ∩ Q̃1 , Õ1 − Q̃1 , Q̃1 − Õ1 , Q̃2 − Õ1
as T et1 , ..., T et4 . and their counterparts in the other pair of trios as T et01 , ..., T et04 .
Those tetrads must be in the same sextet. Moreover, the two tetrads left over in
T et1 , ..., T et6 must split both Õ2 ∩ Q̃3 and Õ3 ∩ Q̃3 in to dyads. Pick p1 , ..., p4 so
that p1 ∈ Q̃1 − Õ1 = T et3 , p2 , p3 ∈ Q̃3 ∩ Õ2 ∩ T et5 and p4 ∈ Q̃3 ∩ Õ2 ∩ T et6 .
Pick p01 , ..., p04 in a similar way for the other pair of trios. Since all sextet stabilizers
are conjugate to each other and a sextet stabilizer is transitive on the 6 tetrads,
there is in element σ ∈ M24 that maps T eti to T et0i . Since as Conway mentioned
the stabilizer of every tetrad in a sextet stabilizer is still 2+1+1+0+0+0 transitive
[5], there must be a τ that fixes each tetrad and maps pi to p0i . Now we have
τ σ(Q̃i ) = Q̃0i and τ σ(Õ1 ) = Õ10 . If τ σ(Õ2 ) 6= Õ20 then τ σ(Õ2 ) ∩ Õ20 = {p1 , ..., p4 },
therefore |τ σ(Õ2 ) ∩ Õ30 ∩ Q̃03 | = 1. But if we consider the sextet induced by Õ30 ∩ Q̃03
then we know there is an octad consisting of 4 points in Õ10 and the 4 points in
Õ30 ∩ Q̃03 . That octad would then intersect τ σ(Õ2 ) at a single point, which is not
possible. Thus τ σ(Õ2 ) = Õ20 and τ σ maps the first pair of trios to the second pair.
Checking an instance, we know that in this case I(T rio1 , T rio2 ) must be isomorphic to D8 × D8 , which has i = 4, and we can therefore eliminate all trio stabilizer
pairs of this type.
For type (iii) pairs, we know that {Õ1 , Õ2 , Õ3 , Q̃1 , Q̃2 , Q̃3 } are all unions of
tetrads in a particular sextet. A third trio may either consist of unions of tetrads
in that particular sextet, or contain an octad intersecting one of the tetrads at a
2-point set. Otherwise, if there is an octad intersecting a tetrad in a 1-point set or
3 point set, then that octad must be partitioned into blocks of size 3+1+1+1+1+1,
thus it must intersect one of the octads in T rio1 or T rio2 in a 2-point set, therefore
we can no longer obtain a length 8 list in the end. But if the third trio contains an
octad intersecting one of the tetrads at a 2-point set, say T et1 := Õ1 ∩ Q̃1 , then it
must intersect T et2 := Õ1 − Q̃1 , T et3 := Q̃1 − Õ1 at 2-point sets. Let Õ2 be the
octad containing T et3 and Q̃2 be the octad containing T et2 . Then the octad in
the third trio must intersect both T et4 := Q̃2 − T et2 , T et5 := Õ2 − T et3 at 2-point
sets, which is impossible-we only have 8 points in an octad. Thus the third trio
must consist of unions of tetrads in the sextet induced by T eti . That is to say,
the intersection of all trio stabilizers in the current list must contain the stabilizer
of each individual tetrads in {T eti }, which is proven by Conway to be isomorphic
to C26 : C3 . If we take Õi as vertices and Q̃i as edges, then we can construct
a triangle by attaching an edge to two vertices if their corresponding structures
satisfies Q̃i ⊂ Õj ∪ Õk . To say that the action of I(T rio1 , T rio2 ) on the 6 tetrads
preserves the octads is just saying that it acts as automorphisms of that triangle,
therefore we can see that I(T rio1 , T rio2 ) acts as s subgroup of D6 ∼
= S3 on those
tetrads, and it implies that the index of C26 : C3 in I(T rio1 , T rio2 ) has less than
MAXDIM OF SOME SIMPLE GROUPS
11
or equal to 2 prime factors. Thus if in our list of length 8 there are T rio1 , T rio2
of type (iii) then besides them we can add at most 2 more trio stabilizers into our
list.
Now we shall consider the last case that each pair of trio stabilizers in our list is of
type (iv). Let T rio1 , T rio2 be one of such a pair, Õ3 = Q̃3 , and the sextet induced
by Õ1 ∩ Q̃1 be denoted as {T eti }. If we introduce a third trio T rio3 = {R̃1 , R̃2 , R̃3 },
then there are two cases: in the first one, all octads in T rio3 are unions of tetrads in
{T eti }; in the second case, there is an octad in T rio3 that intersect one of the tetrads
at a 2-point set. Using the same argument as before we can ignore the case where
an octad in T rio3 intersect one of the tetrads at a 1-point or 3-point set. To finish
our proof we shall start with the first case. If Õ3 = Q̃3 = R̃3 then I(T rio1 , T rio2 )
would be contained in the intersection of three trio stabilizers. To see this we
can use T et1 := Õ1 ∩ Q̃1 , T et2 := Õ1 ∩ Q̃2 , T et3 := Õ2 ∩ Q̃1 , T et4 := Õ2 ∩ Q̃2 as
vertices and construct an edge-colored rectangle, where the edges correspond to the
octads containing a pair of tetrads in T et1 , ..., T et4 and they are colored according
to whether they belong to T rio1 or T rio2 . Elements in I(T rio1 , T rio2 ) are exactly
those elements in M2 3 that act as automorphisms of the rectangle, and Elements in
I(T rio1 , T rio2 , T rio3 ) are those elements preserving the diagonals as well. But an
automorphism of that colored triangle would automatically preserve the diagonals,
soI(T rio1 , T rio2 ) ⊂ I(T rio1 , T rio2 , T rio3 ), the trio stabilizers are not in general
position. Therefore the octad shared by T rio1 , T rio3 must be either Õ1 or Õ2 . On
the other hand, the octad shared by T rio2 , T rio3 must be either Q̃1 or Q̃2 , so T rio3
contains one of Õ1 , Õ2 and one of Q̃1 , Q̃2 at the same time, which is impossible.
Thus we cannot add in a T rio3 that has type (iv) intersections with both T rio1 and
T rio2 while keeping the stabilizers in general position without letting the octads in
T rio3 split the tetrads. But in the second case, if there is an octad R̃1 intersecting
one of the tetrads at a 2-point set, then that octad must split one of T et1 , .., T et4
into dyads. Since that octad is not equal to any octad in T rio1 , T rio2 , all of its
intersections with octads in T rio1 , T rio2 need to have 4 points, so that octad must
be split into 4 dyads by the 4 tetrads T et1 , .., T et4 . As a consequence R̃3 = Õ3 =
Q̃3 . If we want to introduce a fourth trio {P̃1 , P̃2 , P̃3 } that follows this pattern,
then P̃1 , P̃2 must split all the dyads outside P̃3 = Õ3 = Q̃3 = R̃3 , for otherwise
by observing the action of three-fold intersections on dyads we know that the 4fold intersection is the same as a 3-fold intersection. Using the same argument of
inspecting the action of 4-fold intersection on the intersection of those dyads again,
we can show that there is no hope to introduce a T rio5 with its stabilizer in general
position with the first 4 stabilizers in this case, therefore we can use at most 4 trio
stabilizers in the list.
Since in our list we can use at most 1 dyad stabilizer, 2 octad stabilizers and 4
trio stabilizers, the length of our list cannot be 8, so M axDim(M24 ) = 7.
5. The Higman-Sims Group and McLaughlin Group
In this section, we utilize the fact that M axDim(M22 ) = 6 to give a lower bound
for the M axDim of two sporadic simple groups, HS and M cL.
Proposition 8. M axDim(HS) ≥ 6, and M axDim(M cL) ≥ 6
Proof. Both the Higman-Sims group and the McLaughlin group can be defined in
terms of automorphisms of strongly regular graphs. As mentioned by Brouwer,
12
LIU TIANYUE AND R. KEITH DENNIS
the Higman-Sims graph has a 1+22+77 construction. Take the points and blocks
in S(3, 6, 22) as vertices and add in an extra vertex v. Connect v to all points
of S(3, 6, 22), connect a point to a block if the point is in the block, connect two
blocks if they are mutually disjoint, and keep points mutually unconnected, then
we get a strongly regular graph of parameter (100, 22, 0, 6) which is the HigmanSims graph [3]. A stabilizer M of vertices in HS is isomorphic toM22 . Each point
(resp. block) stabilizer in M22 then correspond to vertex stabilizers from the 22orbit (resp. 77-orbit) of M . It turns out that if we construct S(3, 6, 22) from
S(5, 8, 24) by stabilizing 2 points, then all stabilizers in M22 of those structures on
S(3, 6, 22) induced by octads containing only one of the 2 points correspond to the
intersections of M with the maximal subgroups in one of the conjugacy classes of
P SU3 (5) in HS[3]. Thus we can pick 6 maximal subgroups in general position in
M as described before and lift them to 5 vertex stabilizers and a P SU3 (5) in HS,
and the result list would be a length 6 list of maximal subgroups in general position
for HS.
This is not the only way to find 6 maximal subgroups in general position in HS.
In particular, if we pick 5 vertices in Higman-Sims graph that induces a pentagon
and another vertex not adjacent to any of them, then their point stabilizers are in
general position.
As mentioned by Brouwer again, the McLaughlin graph has a 22+77+176 construction. Pick 2 points {a, b} in Steiner (5, 8, 24) system, there are 22 points {p}
leftover, 77 octads {O} incident to both {a, b} and 176 octads {Q} incident to a
but not b. Take {p, O, Q} as vertices, connect p to O if p is not in O, connect
p to Q if p is in Q, connect O to O0 if their intersection is {a, b}, connect Q to
Q0 if their intersection has 2 points, and connect O to Q if their intersection has
size 4, then we get a strongly regular graph of parameter (275, 112, 30, 56) which is
the McLaughlin graph [3]. The stabilizer of a 22+77+176 partition is a maximal
subgroup M of M cL isomorphic to M22 . By the construction of the graph, we can
pick 3 vertices from the 22-orbit, 2 vertices from the 77 orbit, and 1 point from the
176 orbit such that the relation between their corresponding structures are exactly
the same as those corresponding to the length 6 general position sequence in M2 2
described above. Their stabilizers in M are therefore in general position. Since
each vertex stabilizer in M can be lifted to a maximal subgroup in M cL, we have
found a list of maximal subgroups in general position, so M axDim(M cL) ≥ 6.
Remark 5.1. The 176 octads {Q} incident to a but not b (resp. to b but not a) are
the points (resp. quadrics) in Higman geometry [3].
6. Remarks
Remark 6.1. By looking at intersections of point stabilizers, we can show that
M axDim(Spn (q)) ≥ n and M axDim(Un (q)) ≥ n. As Brouwer pointed out, the
symplectic groups and unitary groups act on strongly regular graphs. However,
this lower bound is usually not idealistic.
Every group with M axDim > m observed so far has i > M axDim as well. Is it
necessarily true for all groups with M axDim > m? In particular, does it contradict
the following conjecture of Shareshian?
MAXDIM OF SOME SIMPLE GROUPS
13
Conjecture 1. (Shareshian) Any open interval of a subgroup lattice of some finite
group is has the same homotopy as a wedge of spheres.
References
1. N. Blackburn B. Huppert, Finite groups, vol. III, Grundlehren der mathematischen Wissenschaften, no. ISBN: 978-3-642-67999-5, Springer-Verlag Berlin, 1982.
2. T. Brooks, Generating sets of mathieus groups, Senior Thesis for Bachelor at Cornell, May
2013.
3. A. E. Brouwer, A. M. Cohen, and A. Neumaier, Distance-regular graphs, 1 ed., vol. 18, Ergebnisse der Mathematik und ihrer Grenzgebiete, no. ISBN: 978-3-642-74343-6, Springer-Verlag
Berlin, July 1989.
4. P. J. Cameron and P. Cara, Independent generating sets and geometries for symmetric groups,
Journal of Algebra 258 (2002), no. 2, 641–650.
5. J. H. Conway and N. J. A. Sloane, Sphere packings, lattices and groups, 1 ed., vol. 290,
Grundlehren der mathematischen Wissenschaften, no. ISBN 978-1-4757-2018-1, SpringerVerlag New York Inc., 1988.
6. J. Whiston, The minimal generating sets of maximal size of selected groups, Ph.D. thesis,
Cambridge University, 2001.
7. R. A. Wilson, Finite simple groups, Graduate Texts in Mathematics, no. ISBN 978-1-84800987-5, Springer-Verlag London, 2009.
| 4math.GR |
arXiv:1310.6012v2 [q-bio.PE] 24 Nov 2015
Evolution of swarming behavior
is shaped by how predators attack
R ANDAL S. O LSON1,4,* , D AVID B. K NOESTER2,4,** , AND C HRISTOPH A DAMI2,3,4.***
1
Department of Computer Science & Engineering
Department of Microbiology & Molecular Genetics
3
Department of Physics and Astronomy
4
BEACON Center for the Study of Evolution in Action Michigan State University,
East Lansing, MI 48824, USA
*
rso@randaloson.com
**
dk@msu.edu
***
adami@msu.edu
2
November 25, 2015
Abstract
Animal grouping behaviors have been widely studied due to their implications for understanding
social intelligence, collective cognition, and potential applications in engineering, artificial intelligence,
and robotics. An important biological aspect of these studies is discerning which selection pressures
favor the evolution of grouping behavior. In the past decade, researchers have begun using evolutionary computation to study the evolutionary effects of these selection pressures in predator-prey models.
The selfish herd hypothesis states that concentrated groups arise because prey selfishly attempt to place
their conspecifics between themselves and the predator, thus causing an endless cycle of movement toward the center of the group. Using an evolutionary model of a predator-prey system, we show that
how predators attack is critical to the evolution of the selfish herd. Following this discovery, we show
that density-dependent predation provides an abstraction of Hamilton’s original formulation of “domains of danger.” Finally, we verify that density-dependent predation provides a sufficient selective
advantage for prey to evolve the selfish herd in response to predation by coevolving predators. Thus,
our work corroborates Hamilton’s selfish herd hypothesis in a digital evolutionary model, refines the
assumptions of the selfish herd hypothesis, and generalizes the domain of danger concept to densitydependent predation.
Keywords: group behavior, selfish herd theory, predator attack mode, density-dependent predation,
predator-prey coevolution, evolutionary algorithm, digital evolutionary model.
1
1
Introduction
Over the past century, researchers have devoted considerable effort into studying animal grouping behavior due to its important implications for social intelligence, collective cognition, and potential applications in engineering, artificial intelligence, and robotics [5]. Indeed, grouping behaviors are pervasive
across all forms of life. For example, European starlings (Sturnus vulgaris) are known to form murmurations of millions of birds which perform awe-inspiring displays of coordinated movement [12, 19].
Western honeybees (Apis mellifera) communicate the location of food and nest sites to other bees in their
group via a complex dance language [9]. Even relatively simple bacteria exhibit grouping behavior, such
as Escherichia coli forming biofilms which allow their group to survive in hostile environments [17].
Swarming is one example of grouping behavior, where animals coordinate their movement with conspecifics to maintain a cohesive group. Although swarm-like groups could arise by chance, e.g., Little
Egrets (Egretta garzetta) pursuing a common resource in water pools [28], typically swarms are maintained via behavioral mechanisms that ensure group cohesion [2]. As with many traits, swarming behavior entails a variety of fitness costs, such as increased risk of predation and the requisite sharing of
resources with the group [44]. With this fact in mind, significant effort has been dedicated to understanding the compensating benefits that grouping behavior provides [30]. Many such benefits of grouping behavior have been proposed, for example, swarming may improve mating success [68, 8], increase
foraging efficiency [47], or enable the group to solve problems that would be impossible to solve individually [5]. Furthermore, swarming behaviors are hypothesized to protect group members from predators
in several ways. For example, swarming can improve group vigilance [57, 27, 59, 46], reduce the chance
of being encountered by predators [59, 22], dilute an individual’s risk of being attacked [23, 58, 13, 18],
enable an active defense against predators [30], or reduce predator attack efficiency by confusing the
predator [24, 26, 29].
Unfortunately, many swarming animals take months or even years to produce offspring. These long
generation times make it extremely difficult to experimentally determine which of the aforementioned
benefits are sufficient to select for swarming behavior as an evolutionary response, let alone study the
behaviors as they evolve [26, 3]. In this paper, we use a digital model of predator-prey coevolution to
explore Hamilton’s selfish herd hypothesis [18]. Briefly, the selfish herd hypothesis states that prey in
groups under attack from a predator will seek to place other prey in between themselves and the predator, thus maximizing their chance of survival. As a consequence of this selfish behavior, individuals
continually move toward a central point in the group, which gives rise to the appearance of a cohesive
swarm. This paper expands on earlier work [41] by studying the long-term evolutionary effects of differing attack modes, exploring a new attack mode that directly selects against swarming behavior, and
providing an analysis of the control algorithms that evolved in the swarming prey.
2
Related Work
Hamilton’s original formulation of the selfish herd hypothesis introduced the concept of “domains of
danger” (DODs, Figure 1), which served as a method to visualize the likelihood of a prey inside a
group to be attacked by a predator [18]. Prey on the edges of the group would have larger DODs
than prey on the inside of the group; thus, prey on the edges of the group would be attacked more
frequently. Moreover, Hamilton proposed that prey on the edges of the group would seek to reduce
their DOD by moving inside the group, thus placing other group members between themselves and the
predator. Further work has expanded on this hypothesis by adding a limited predator attack range [25],
investigating the effects of prey vigilance [4], considering the initial spatial positioning of prey when the
group is attacked [38], exploring the role of prey body characteristics in shaping herd characteristics [31,
20], and even confirming Hamilton’s predictions in biological systems [48].
Additional studies have focused on the movement rules that prey in a selfish herd follow to minimize
2
Domain of Danger
Prey
Figure 1: Example “domains of danger” (DODs) from Hamilton’s selfish herd hypothesis. Each triangle represents a prey in the group, and the area around each triangle is its DOD. Prey on the inside
of the group have smaller DODs, which means they are less likely to be targeted when a predator attacks. As a consequence, “selfish” prey that move inside the group to minimize their DOD will have an
evolutionary advantage.
3
their DOD [62]. This line of work began by demonstrating that the simple movement rules proposed
by Hamilton reduce predation risk for prey inside the group [39], then opened some parameters of the
movement rules to evolution in an attempt to discover a more biologically plausible set of movement
rules [49, 65]. Importantly, these studies demonstrated that it is possible for selfish herd behavior to
evolve by natural selection on movement rules that rely on only local information for each agent, rather
than global information about the entire group. This paper builds on this work by studying the effects
of coevolving predators and predator attack mode (i.e., how predators select a prey in a group to attack)
on the evolution of the selfish herd.
That said, there are many other potential causes of swarming behavior that we do not address in
this study. For example, some studies have investigated the evolution of predator behavior in response
to prey density [56], the role of relative predator and prey speeds on the evolution of grouping behavior [66], elaborated upon the interaction between ecology and the evolution of grouping behavior [55, 64], and explored the role of group vigilance (i.e., the “many eyes” hypothesis) in the evolution
of grouping behavior [16, 43]. Two recent studies have explored the coevolution of predator and prey
behavior in the presence of the predator confusion effect [40, 32], and found that the predator confusion
effect is sufficient to select for the evolution of swarming behavior in the absence of any other group
benefits. Further, it has been shown that predators can adapt composite tactics to improve their efficacy
against swarms of prey, which greatly reduces the defensive benefit of swarming [7]. It is therefore necessary to keep in mind that while this study investigates the evolution of swarming behavior according
to Hamilton’s selfish herd hypothesis, in natural populations there are often many interacting benefits
and costs of swarming behavior that must be taken into account [30].
Of course, prior to this work there has been considerable research exploring the (co-)evolution of
animal behavior in agent-based models. Craig Reynolds’ work on the evolution of prey behavior in response to simulated predation was one of the earliest papers to demonstrate that predation can directly
select for aggregative behavior in prey populations [51]. Similarly, Karl Sims’ work in 1994 established a
new paradigm for the coevolution of agent behavior and morphology [54], which is still an object of intense study to this day. The work in this paper seeks to provide a stronger biological grounding to these
kinds of agent-based evolution experiments, in particular by implementing specific selection pressures
on the prey that are discussed in the animal behavior literature.
More broadly, in the past decade researchers have focused on the application of locally-interacting
swarming agents to optimization problems, called Particle Swarm Optimization (PSO) [45]. PSO applications range from feature selection for classifiers [67], to video processing [60], to open vehicle
routing [34]. A related technique within PSO seeks to combine PSO with coevolving “predator” and
“prey” solutions to avoid local minima [53], which has proven effective in other evolutionary computation domains as well [1]. Researchers have even sought to harness the collective problem solving
power of swarming agents to design robust autonomous robotic swarms [6]. Thus, elaborations on the
foundations of animal grouping behavior has the potential to improve our ability to solve engineering
problems.
3
Methods
To study the evolution of the selfish herd, we developed an agent-based model in which agents interact
in a continuous, toroidal virtual environment (736 × 736 virtual meters), shown in Figure 2. At the
beginning of each simulation, we place 250 agents in the environment at uniformly random locations.
These agents are treated as “virtual prey.” Each agent is controlled by a Markov Network (MN), which
is a probabilistic controller that makes movement decisions based on a combination of sensory input
(i.e., vision) and internal states (i.e., memory) [10]. We evolve the agent MNs with a genetic algorithm
(GA) [11, 14] under varying selection regimes, which will be described in more detail below.
During each simulation time step, all agents read information from their sensors and take action
4
Figure 2: A depiction of the simulation environment in which the agents interact. Black dots are prey
agents, the black triangle is a predator agent, and the lines around the predator agent indicate its field
of view. Agents wrap around the edges of the toroidal simulation environment.
5
PL PR
A
0 0
0 0
1
0
0
00
0
0
0
0 1
0 1
0
00 000
0
Markov
Network
0 0
0 0
0
0 0 0 00
1
L
retina
1
1 0
0 1
0
00 000
0
0
0
R
Figure 3: An illustration of the agents in the model. Light grey triangles are prey agents and the dark
grey triangles are predator agents. The agents have a 360◦ limited-distance retina (200 virtual meters) to
observe their surroundings and detect the presence of other agents. The current heading of the agent is
indicated by a bold arrow. Each agent has its own Markov Network, which decides where to move next
based off of a combination of sensory input and memory. The left and right actuators (labeled “L” and
“R”) enable the agents to move forward, left, and right in discrete steps.
(i.e., move) based on their effectors. In our first set of treatments, we simulate an ideal, disembodied
predator by periodically removing prey agents from the environment and marking them as consumed,
e.g., when they are on the outermost edges of the group. Subsequent treatments introduce an embodied,
coevolving predator agent which is controlled by its own MN. The data1 and source code2 from these
experiments is available online for further analysis. In the remainder of this section, we describe the
sensory-motor architecture of individual agents and present details related to the function and encoding
of MNs.
3.1
Agent Model
Figure 3 depicts the sensory-motor architecture of the agents used for this study. A prey agent can
sense predators and conspecifics with a limited-distance (200 virtual meters), pixelated retina covering
its entire 360◦ visual field. Its retina is split into 24 even slices, each covering an arc of 15◦ , which is
an abstraction of the broad, coarse visual systems often observed in grouping prey [35]. Regardless of
the number of agents present in a single retina slice, the prey agent only knows whether a conspecific
1 Data:
2 Code:
http://dx.doi.org/10.6084/m9.figshare.663680
https://github.com/adamilab/eos-selfish-herd
6
Table 1: Possible actions encoded by the agent’s output. Each output pair encodes a discrete action taken
by the agent. The agent’s MN changes the values stored in output states L and R to indicate the action
it has decided to take in the next simulation time step.
Output L Output R Encoded Action
0
0
Move forward
0
1
Turn right
1
0
Turn left
1
1
Stay still
or predator resides within that slice, but not how many. For example, in Figure 3, the fourth retina
slice to the right of the agent’s heading (labeled “A”) has both the predator and prey sensors activated
because there are two predator agents and a prey agent inside that slice. Once provided with its sensory
information, the prey agent chooses one of four discrete actions, as shown in Table 1. Prey agents turn
in 8◦ increments and move 1 virtual meter each time step.
In our coevolution experiments, the predator agents can detect only nearby prey agents using a
limited-distance (200 virtual meters), pixelated retina covering its frontal 180◦ that works just like the
prey agent’s retina (Figure 3). Similar to the prey agents, predators make decisions about how to move
next using their MN, as shown in Table 1, but move 3× faster than the prey agents and turn correspondingly slower (6◦ per simulation time step) due to their higher speed. This dramatically faster predator
movement speed is meant to represent predators that perform rapid attacks on groups of prey, such as a
Peregrine falcon dive bombing a swarm of Starlings. Finally, if a predator agent moves within 5 virtual
meters of a prey agent that is anywhere within its retina, the predator agent makes an attack attempt on
the prey agent. If the attack attempt is successful, we remove the prey agent from the simulation and
mark it as consumed.
3.2
Markov Networks
Each agent is controlled by its own Markov Network (MN), which is a probabilistic controller that makes
decisions about how the agent interacts with the environment and other agents within that environment.
Since a MN is responsible for the control decisions of its agent, it can be thought of as an artificial brain for
the agent it controls. Although we specifically use MNs as the artificial brain in these experiments, other
artificial brains such as Artificial Neural Networks, Genetic Programming, or many other evolvable
substrates that can produce agent-based behavior based on sensory inputs could also be used in these
experiments.
Every simulation time step, the MNs receive input via sensors (e.g., visual retina), perform a computation on inputs and any hidden states (i.e., memory), then place the result of the computation into
hidden or output states (e.g., actuators). We note that MN states are binary and only assume a value of
0 or 1. When we evolve MNs with a GA, mutations affect (1) which states the MN pays attention to as
input, (2) which states the MN outputs the result of its computation to, and (3) the internal logic that
converts the input into the corresponding output.
How Markov Networks Function
When we embed an agent into the simulation environment, we provide sensory inputs from its retina
into its MN every simulation step (labeled “retina” and “Markov Network”, respectively). Once we
provide a MN with its inputs, we activate it and allow it to store the result of the computation into
its hidden and output states for the next time step. MNs are networks of Markov Gates (MGs), which
perform the computation for the MN. In Figure 4, we see two example MGs, labeled “Gate 1” and “Gate
7
0
1
2
3
4
5
Gate 1
0
1
2
6
7
time t
7
time t+1
Gate 2
3
4
5
6
Figure 4: An example Markov Network (MN) with four input states (white circles labeled 0-3), two
hidden states (light grey circles labeled 4 and 5), two output states (dark grey circles labeled 6 and 7),
and two Markov Gates (MGs, white squares labeled “Gate 1” and “Gate 2”). The MN receives input
into the input states at time step t, then performs a computation with its MGs upon activation. Together,
these MGs use information about the environment, information from memory, and information about
the MN’s previous action to decide where to move next.
2.” At time t, Gate 1 receives sensory input from states 0 and 2 and retrieves state information (i.e.,
memory) from state 4. At time t + 1, Gate 1 then stores its output in hidden state 4 and output state 6.
Similarly, at time t Gate 2 receives sensory input from state 2 and retrieves state information in state 6,
then places its output into states 6 and 7 at time step t + 1. When MGs place their output into the same
state, the outputs are combined into a single output using the OR logic function. Thus, the MN uses
information from the environment and its memory to decide where to move in the next time step t + 1.
In a MN, states are updated by MGs, which function similarly to digital logic gates, e.g., AND & OR.
A digital logic gate, such as XOR, reads two binary states as input and outputs a single binary value
according to the XOR logic. Similarly, MGs output binary values based on their input, but do so with a
probabilistic logic table. Table 2 shows an example MG that could be used to control a prey agent that
avoids nearby predator agents. For example, if a predator is to the right of the prey’s heading (i.e., PL =
0 and PR = 1, corresponding to the second row of this table), then the outputs are move forward (MF)
with a 20% chance, turn right (TR) with a 5% chance, turn left (TL) with a 65% chance, and stay still (SS)
with a 10% chance. Thus, due to this probabilistic input-output mapping, the agent MNs are capable of
producing stochastic agent behavior.
The MGs in this model can receive input from a maximum of 4 states, and write into a maximum of
4 states, with a minimum of 1 input and 1 output state for each MG. Any state (input, output, or hidden)
in the MN can be used as an input or output for a MG. MNs can be composed of any number of MGs,
and the MGs are what define the internal logic of the MN. Thus, to evolve a MN, mutations change the
connections between states and MGs, and modify the probabilistic logic tables that describe each MG.
Mutations act directly on the genetic encoding of the MN, which is described next.
Genetic Encoding of Markov Networks
We use a circular string of bytes as a genome, which contains all the information necessary to describe a
MN. The genome is composed of genes, and each gene encodes a single MG. Therefore, a gene contains
the information about which states the MG reads input from, which states the MG writes its output to,
8
Start
N in Nout Input State IDs
Output State IDs
Probabilities
Gene 1
42 213 207 100 40 130
4
#
140 46
#
#
89
113 .......
255
Gene 2
42 213 134 97
20 238
#
#
254
#
#
71 248 101 .......
167
7
1
Figure 5: Example circular byte strings encoding the two Markov Gates (MGs) in Figure 4, denoted Gene
1 and Gene 2. The sequence (42, 213) represents the beginning of a new MG (white blocks). The next
two bytes encode the number of input and output states used by the MG (light grey blocks), and the
following eight bytes encode which states are used as input (medium grey blocks) and output (darker
grey blocks). The remaining bytes in the string encode the probabilities of the MG’s logic table (darkest
grey blocks).
and the probability table defining the logic of the MG. The start of a gene is indicated by a start codon,
which is represented by the sequence (42, 213) in the genome.
Figure 5 depicts an example genome. After the start codon, the next two bytes describe the number
of inputs (Nin ) and outputs (Nout ) used in this MG, where each N = 1 + (byte mod Nmax ). Here, Nmax
= 4. The following Nmax bytes specify which states the MG reads from by mapping to a state ID number
with the equation: (byte mod Nstates ), where Nstates is the total number of input, output, and hidden
states. Similarly, the next Nmax bytes encode which states the MG writes to with the same equation
as Nin . If too many inputs or outputs are specified, the remaining sites in that section of the gene are
ignored, designated by the # signs. The remaining 2Nin +Nout bytes of the gene define the probabilities in
the logic table.
The maximum number of states allowed and which states are used as inputs and outputs are specified as constants by the user. In these experiments, we provided 64 states for the MNs to work with:
24 sensory inputs, 2 outputs for the actuators, and 38 hidden states for optional internal computations.
Combined with these constants, the genome described above unambiguously defines a MN.
All evolutionary changes such as point mutations, duplications, deletions, or crossover are performed on the byte string genome, with probabilities as shown in Table 3. During a point mutation,
a random byte in the genome is replaced with a new byte drawn from a uniform random distribution. If
a duplication event occurs, two random positions are chosen in the genome and all bytes between those
points are duplicated into another part of the genome. Similarly, when a deletion event occurs, two random positions are chosen in the genome and all bytes between those points are deleted. Crossover for
MNs is not implemented in this experiment to allow for a succinct reconstruction of the line of descent
Table 2: An example MG that could be used to control a prey agent which avoids nearby predator
agents. “PL” and “PR” correspond to the predator sensors just to the left and right of the agent’s heading,
respectively, as shown in Figure 3. The columns labeled P(X) indicate the probability of the MG deciding
on action X given the corresponding input pair. MF = Move Forward; TR = Turn Right; TL = Turn Left;
SS = Stay Still.
PL PR P(MF) P(TR) P(TL) P(SS)
0
0
0.7
0.05
0.05
0.2
0
1
0.2
0.05
0.65
0.1
1
0
0.2
0.65
0.05
0.1
1
1
0.05
0.8
0.1
0.05
9
Table 3: Genetic algorithm and experiment settings.
GA Parameter
Value
Selection
Fitness proportionate
Population size
250
Per-gene mutation rate 1%
Gene duplication rate
5%
Gene deletion rate
2%
Crossover
None
Generations
40,000
Replicates
100
of the population [33] (described more below), which is a useful tool in evolutionary studies that we
harness in this paper.
4
Artificial Predation
In our first set of experiments, we observe the evolution of prey behavior in response to various forms
of artificial predation. This enables us to experimentally control the specific modes of predation and
observe their effect on the evolution of the selfish herd. We evolve the prey genomes with a GA with the
settings described in Table 3. We begin the evolutionary process by seeding the prey genome pool with
a set of randomly-generated ancestor genomes of length 5,000 with four random MGs. Following this,
we evaluate the relative fitness of each prey genome by translating the genome into its corresponding
MN, embodying each MN in a prey agent, and competing the prey agents in a simulation environment
for 1,000 simulation time steps. This evaluation period is akin to the agents’ lifespan, hence each agent
has a potential lifespan of 1,000 time steps. We assign each prey genome an individual fitness according
to how long its corresponding prey agent survived, following the equation:
Wprey = T
(1)
where T is the number of time steps the prey agent survived in the simulation environment. Thus,
individual prey genomes are rewarded for their agent surviving longer than other agents in the group.
Once all of the prey genomes are assigned fitness values, we perform fitness-proportionate selection on
the population of genomes via a Moran process [36], increment the generation counter, and repeat the
evaluation process on the new population of genomes until the final generation (40,000) is reached.
In all cases, we give the prey an initial 250 simulation time steps without predation to move around,
so that prey starting on the outside of the group have the chance to move toward the center of the group
if they wish to. Once the initial 250 simulation time steps elapse, we apply artificial predation every
4 simulation time steps to simulate an ideal predator attacking the group. Artificial predators succeed
with their attacks every time. We limit the artificial predator attack rate to one attack attempt every 4
simulation time steps, which is called the handling time. The handling time represents the time it takes
the simulated predator to consume and digest a prey after successful prey capture, or the time it takes
to refocus on another prey in the case of an unsuccessful attack attempt. We selected a handling time
of 4 because it reduces the herd of prey down to 25% of its original size by the end of the simulation,
therefore applying strong selection pressure for survivorship in the herd.
For each experiment, we characterize the grouping behavior by measuring the swarm density of the
entire prey population every generation [21]. We measure the swarm density as the mean number of
prey within 30 virtual meters of each other over a lifespan of 1,000 simulation time steps, which we have
experimentally shown to differentiate between swarming and non-swarming behavior in previous pub10
A)
B)
x
x
x
x
x
x
x
x
x
x x
x
C)
x
x
D)
x
xxx
x
x
x
x
Figure 6: An illustration of the four artificial predator attack modes. A) Random attacks, B) Random
walk attacks, C) Outside attacks, and D) High-density area attacks.
lished experiments [40]. Qualitatively, a swarm density of ≥ 15 indicates cohesive swarming behavior,
between 15 and 5 loosely grouping behavior, and ≤ 5 random, non-grouping behavior. Thus, swarm
density captures how cohesively the prey are swarming, or if the prey are even grouping at all.
In the following sections, we study the effect of four different attack modes on the evolution of
swarming behavior: uncorrelated random attacks (Figure 6A), correlated random attacks (random walk
attacks, Figure 6B), peripheral attacks (Figure 6C), and attacks that target the most dense area of the
swarm (Figure 6D).
4.1
Random Attacks
Our initial study sought to verify Hamilton’s selfish herd hypothesis by modeling evolving prey under
attack by predators that ambush prey from a random location in the simulation environment. If the
selfish herd hypothesis holds, we expect prey to minimize their “domain of danger” to the predators
by placing as many conspecifics as possible around them [18]. Similar to previous models studying
the selfish herd [65], a random attack proceeds by selecting a uniformly random location inside the
simulation space, then attacking the prey closest to that location, as shown in Figure 6A.
As seen in Figure 7, swarming behavior is weakly selected for when the predators make uniformly
random attacks on the prey3 (light grey triangles). Particularly, we found that prey took upwards of
5,000 generations to evolve cohesive swarming behavior when experiencing random attacks, compared
to less than 1,000 generations with the other attack modes. However, even random attacks selected
for more cohesive swarming behavior than no attacks at all, which resulted in completely dispersive
behavior (Figure 7, light grey stars).
This finding has important implications, namely that one of the original assumptions of the selfish
herd hypothesis—that the predator attack mode has no impact on the evolution of swarming behavior—
is not corroborated by this model. Following this discovery, we hypothesized that the directionality of the
predators’ attacks play a critical role in the evolution of the selfish herd. To test this hypothesis, we next
explore two different predator attack modes, each with their own distinct directionality of predation.
4.2
Random Walk Attacks
Our next experiment alters the mode of predation from a predator that attacks randomly selected locations to a predator that follows a random walk within the simulation environment. Shown in Figure 6B,
after each attack made by this predator, it is then moved to a random location within 50 virtual meters
3 Video:
Evolution of prey under Random Attack treatment: http://dx.doi.org/10.6084/m9.figshare.658857
11
50
Outside Attack
Random Walk Attack
Mean Nearby Prey
40
Random Attack
No Attacks
30
20
10
0
0
5000 10000 15000 20000 25000 30000 35000 40000
Generation
Figure 7: Mean swarm density over all replicates over evolutionary time, measured by the mean number of prey within 30 virtual meters of each other over a lifespan of 1,000 simulation time steps. Prey in
groups attacked randomly (light grey triangles) took much longer to evolve cohesive swarming behavior than prey in groups attacked by a predator that follows a random walk (dark grey circles) or always
from the outside of the group (black squares). When prey experience no attacks, they do not evolve
swarming behavior at all (light grey stars). Error bars indicate two standard errors over 100 replicates.
of its previous location. This models a predator that persistently feeds on a group of prey, rather than
ambushing.
Figure 7 shows that swarming evolved quickly when the prey were attacked by a predator following
a random walk4 (dark grey circles). Notably, even by generation 40,000, prey experiencing random walk
attacks formed significantly more cohesive swarms than prey experiencing random attacks. Thus, the
random walk predator attack mode appears to capture an important aspect of predation that selects for
swarming behavior.
4.3
Outside Attacks
In the last of our initial artificial predation experiments, we simulate a predator that always approaches
from the outside of the group and attacks the prey nearest to it, as in [63]. This predator attack mode
effectively has the predators consistently attacking prey on the outer edges of the group. As shown in
shown in Figure 6C, we simulate this predator attack mode by first choosing a random angle outside
of the group for the predator to approach from. Once an angle is chosen, we convert the angle into a
location on the edge of the visible simulation space and attack the prey nearest to that location.
As shown in Figure 7, this form of predation has the most significant impact on the evolution of the
selfish herd so far. When attacked by predators that consistently target prey on the edges of the group,
prey quickly evolve cohesive swarming behavior5 (black squares). Taken together, the results of these
4 Video:
5 Video:
Evolution of prey under Random Walk treatment: http://dx.doi.org/10.6084/m9.figshare.658856
Evolution of prey under Outside Attack treatment: http://dx.doi.org/10.6084/m9.figshare.658854
12
50
Outside Attack
Random Walk Attack
Random Attack
Mean Nearby Prey
40
30
20
10
0
0
5000 10000 15000 20000 25000 30000 35000 40000
Generation
Figure 8: Mean swarm density over all replicates over evolutionary time, measured by the mean number
of prey within 30 virtual meters of each other over a lifespan of 1,000 simulation time steps. Even when
experiencing density-dependent predation, prey in groups attacked randomly (light grey triangles) took
much longer to evolve swarming behavior than prey in groups attacked by a persistent artificial predator
(dark grey circles) or always from the outside of the group (black squares). Error bars indicate two
standard errors over 100 replicates.
artificial predation experiments demonstrate another discovery of this work: The more predators attack
prey on the outside of the group, the faster the selfish herd will evolve.
One translation of this finding is that in order for the selfish herd to evolve, prey must experience
a higher predation rate on the outside of the group than in the middle of the group. While this phenomenon can be explained by each prey having a “domain of danger” (DOD) influenced by its relative
position in the group [18, 25, 39], an alternative hypothesis is that of density-dependent predation.
4.4
Density-Dependent Predation
To study the impact of density-dependent predation on the evolution of the selfish herd, we impose a
constraint on the predator which reduces its attack efficiency when it attacks areas of the group with
high prey density. This reduced attack efficiency is meant to represent the increased predation rate that
prey on edges of the group are expected to endure [18, 25, 39], and such density-dependence can also
be thought of as a proxy for group defense. We compute the predator’s probability of capturing a prey
during a given attack (Pcapture ) with the following equation:
Pcapture =
1
Adensity
(2)
where Adensity is the number of prey within 30 virtual meters of the target prey, including the target
prey itself. For example, if the predator attacks a prey with 4 other prey nearby (Adensity = 5), it has a
20% chance of successfully capturing the prey. As a consequence of this mechanism, the prey experience
density-dependent predation.
13
Table 4: High-density area attack (HDAA) experiment treatments. The values listed for each treatment
are the handling times for the corresponding predator attack mode.
HDAA?
Outside Attack Frequency HDAA Frequency
No
10
N/A
Infrequent 10
250
Frequent
10
25
Figure 8 demonstrates the effect of density-dependent predation on the previous artificial predation
experiments. Just as before, when predators did not preferentially attack prey on the outside of the
group, as in the random attack experiment (light grey triangles), swarming behavior took much longer
to evolve. In contrast, when the predators followed a random walk (dark grey circles) or always attacked
from the outside of the group (black squares), the prey experiencing density-dependent predation again
quickly evolved swarming behavior. The most noticeable effect of density-dependent predation is on
the random attack treatment, where the swarm density measurement at generation 5,000 increased from
11.19±2.58 (mean ± two standard errors) to 17.61±2.72, indicating significantly stronger selection for
swarming.
4.5
High-Density Area Attacks
Thus far, we have explored attack modes that select for the evolution of swarming behavior. It is not surprising that there are also attack modes exhibited by natural predators that must select against swarming
behavior in their prey. For example, blue whales (Balaenoptera musculus) are known to dive into the densest areas in swarms of krill, consuming hundreds of thousands of krill in the middle of the swarm in a
single attack [15]. We call this kind of attack mode a high-density area attack. Such an attack clearly selects
against swarming behavior because it targets the prey that swarm the most. If krill swarms consistently
experience these high-density area attacks, then why do they still evolve swarming behavior?
It is important to note that krill swarms are also fed on by smaller species, such as crabeater seals
(Lobodon carcinophagus), that consistently attack the krill on the outside of the swarm [37]. Thus, krill
swarms are experiencing two forms of attack modes simultaneously: High-density area attacks from
whales and outside attacks from crabeater seals. Thus, it is possible that the selection pressure to swarm
from outside attacks (Figure 7) could outweigh the selection pressure to disperse from high-density area
attacks.
Shown in Figure 6D, we model high-density area attacks as an artificial attack that always targets
the prey at the most dense area of the swarm (i.e., highest Adensity ). We note that this attack mode
is the opposite of the density-dependent mechanism explored in the previous section, which favors
predators that target prey in the least dense area of the swarm. Once the target is selected, we execute
the attack by removing the target prey and all other prey within 30 virtual meters of the target prey.
Outside attacks are modeled as described above. To study the effect of high-density area attacks on
the evolution of swarming behavior, we allow the prey to evolve while experiencing both attack modes
simultaneously. We vary the relative handling times of both attacks (Table 4) to explore whether relative
attack frequency could explain why some swarming animals evolved swarming behavior despite the
fact that they experience high-density area attacks.
As shown in Figure 9, prey experiencing only outside attacks quickly evolve cohesive swarming
behavior (light grey triangles). However, when we introduce infrequent high-density area attacks (dark
grey circles), the selection pressure for prey to swarm is reduced. Finally, when we introduce frequent
high-density area attacks (black squares), the prey do not evolve swarming behavior at all. Thus, one
possible explanation for animals evolving swarming behavior despite experiencing high-density area
attacks is that the high-density area attacks are too infrequent relative to other attack types to exert a
14
Mean Nearby Prey
40
35
30
25
20
15
10
5
0
Outside Attack
Combined Attacks - Infrequent HDAA
Combined Attacks - Frequent HDAA
0
5000 10000 15000 20000 25000 30000 35000 40000
Generation
Figure 9: Mean swarm density over all replicates over evolutionary time, measured by the mean number of prey within 30 virtual meters of each other over a lifespan of 1,000 simulation time steps. Swarm
density was measured from evolved populations that were not experiencing predation during measurement, eliminating any possible effects of attack modes that kill more prey faster. Prey in groups attacked
only by outside attacks (light grey triangles) evolved cohesive swarming behavior. Increasing the relative frequency of high-density area attacks from infrequent (dark grey circles) to frequent (black squares)
caused the prey to evolve increasingly dispersive behavior. Error bars indicate two standard errors over
100 replicates.
strong enough selection pressure for prey to disperse.
In summary, the artificial predation experiments provided us with two important findings regarding
the evolution of the selfish herd: (1) attacks on prey on the periphery of the herd exert a strong selection
pressure for prey to swarm and (2) prey in less dense areas, such as those on the outside of the herd,
must experience a higher predation rate than in areas of dense prey, such as those in the center of the
herd.
5
Predator-Prey Coevolution
Building upon the artificial predation experiments, we implemented density-dependent predation in a
predator-prey coevolution experiment. Adding predators into the simulation environment enables us
to observe how embodied coevolving predators affect the evolution of the selfish herd.
For this experiment, we coevolve a population of 100 predator genomes with a population of 100
prey genomes using a GA with settings described in Table 3. Specifically, we evaluate each predator
genome against the entire prey genome population for 2,000 simulation time steps each generation.
During evaluation, we place 4 clonal predator agents inside a 512 × 512 virtual meters simulation environment with all 100 prey agents and allow the predator agents to make attack attempts on the prey
agents. The prey genome population size, simulation environment area, and total number of GA generations were decreased in this experiment due to computational limitations imposed by predator-prey
15
Mean Nearby Prey
40
35
30
25
20
15
10
5
0
Density Dependent Predation
No Density Dependent Predation
0 100 200 300 400 500 600 700 800 900 100011001200
Generation
Figure 10: Mean swarm density over all replicates over evolutionary time, measured by the mean number of prey within 30 virtual meters of each other over a lifespan of 2,000 simulation time steps. Prey in
groups experiencing density-dependent predation (black circles) evolved cohesive swarming behavior,
whereas prey in groups not experiencing density-dependent predation (light grey triangles) evolved
dispersive behavior. Error bars indicate two standard errors over 100 replicates.
coevolution. We assigned the prey individual fitness values as in the previous experiments, and evaluated predator fitness according to the following equation:
Wpredator =
tX
max
(S0 − At )
(3)
t=1
where t is the current simulation time step, tmax is the total number of simulation time steps (here, tmax =
2,000), S0 is the starting group size (here, S0 = 100), and At is the number of prey alive at update t. Thus,
predators are selected to consume more prey faster, and prey are selected to survive longer than other
prey in the group. Once all of the predator and prey genomes are assigned fitness values, we perform
fitness proportionate selection on the populations via a Moran process [36], increment the generation
counter, and repeat the evaluation process on the new populations until the final generation (1,200) is
reached.
To evaluate the coevolved predators and prey quantitatively, we obtained the line of descent (LOD)
for every replicate by tracing the ancestors of the most-fit prey MN in the final population until we
reached the randomly-generated ancestral MN with which the starting population was seeded (see [33]
for an introduction to the concept of a LOD in the context of digital evolution). We again characterized the prey grouping behavior by measuring the swarm density of the entire prey population every
generation.
Figure 10 depicts the prey behavior measurements for the coevolution experiments with densitydependent predation6 (black circles; mean swarm density at generation 1,200 ± two standard errors:
6 Video:
Prey from predator-prey coevolution treatment: http://dx.doi.org/10.6084/m9.figshare.658855
16
L
R
Number of prey MNs connected
Markov
Network
Figure 11: Number of sensory input connections from 100 evolved prey Markov Networks mapped
onto a prey agent. Only causal connections from the sensory inputs to the actuators are shown. The
arrow indicates the facing of the agent. The prey Markov Networks evolved a strong preference for
connecting to prey sensors in front and a slight preference for sensors behind the prey agent, but tended
to not connect to the sensors on the sides.
26.2±2.3) and without density-dependent predation (light grey triangles; 3.9±0.8). Without densitydependent predation, the prey evolved purely dispersive behavior as a mechanism to escape the predators, even after 10,000 generations of evolution (Supplementary Figure S1). In contrast, with densitydependent predation, the prey quickly evolved cohesive swarming behavior in response to attacks from
the predators within 400 generations. As expected, the coevolving predators adapted to the prey swarming behavior in the density-dependent treatments by focusing on prey on the edges of the swarm, where
the density of prey is lowest. As a caveat, density-dependent predation only selects for cohesive swarming behavior when the predators are faster than the prey (Supplementary Figure S2), which corroborates
earlier findings exploring the role of relative predator-prey speeds in the evolution of swarming behavior [66].
Here we see that density-dependent predation provides a sufficient selective advantage for prey to
evolve the selfish herd in response to predation by coevolving predators, despite the fact that swarming
prey experience an increased attack rate from the predators due to this behavior (see [40], Figures S3
& S4). Accordingly, these results uphold Hamilton’s hypothesis that grouping behavior could evolve
in animals purely due to selfish reasons, without the need for an explanation that involves the benefits
to the whole group [18]. Moreover, the discoveries in this work refine the selfish herd hypothesis by
clarifying the effect that different attack modes have on the evolution of the selfish herd.
17
6
Evolved Prey Markov Network Analysis
Now that we have evolved emergent swarming behavior in an agent-based model under several different treatments, we can analyze the resulting Markov Networks (MNs) to gain a deeper understanding of
the individual-based mechanisms underlying swarming behavior. For this analysis, we chose the mostabundant prey MN from each of the Outside Attack artificial predation experiment replicates, resulting
in 100 MNs that exhibit swarming behavior.
First, we analyze the structure of the 100 MNs by looking at the specific retina sensors that the MNs
evolved to connect to. Shown in Figure 11, the prey MNs show a strong bias for connecting to the preyspecific retina sensors in front of the prey, but not to the sides. Strangely, there appears to be a strong
preference to connect to the front-right sensor but not the front-left sensor, which is an artifact of the fact
that the front-right sensor is the only sensor that activates when other prey are directly in front of the
prey. Additionally, some of the prey MNs show a preference for connecting to the prey-specific retina
sensors behind the prey. From this analysis alone, we can deduce that the retina sensors that are most
conducive to swarming behavior are in front of the prey agent.
To understand how prey make movement decisions based on their sensory inputs, we map every possible input combination in the prey’s retina to the corresponding movement decision that the
prey made. This mapping is accomplished by generating all 224 possible input combinations—e.g.,
000000000000000000000000, 000000000000000000000001, 000000000000000000000010, etc.—and passing
them into the evolved MN as a simulated sensory input. Upon activating the MN, we receive an output
that corresponds to the action that the prey decided to make in respond to the simulated sensory input.
Due to the stochastic nature of MNs, the prey agents do not always make the same movement decision
when given the same input. Thus, we take the most-likely output from 1,000 repeats as the representative decision for a given sensory input combination. Effectively, this process produces a truth table that
maps every possible sensory input to its corresponding movement decision. An example truth table can
be seen in Table 5.
Table 5: An example truth table mapping every possible sensory input combination to the corresponding
most-likely movement decision from the evolved prey Markov Network.
Sensory input
Corresponding output
000000000000000000000000 00
000000000000000000000001 10
...
...
111111111111111111111111 00
Once we have the truth table of all 224 input-output mappings, we pass the truth table to the logic
minimization software espresso [52], which eliminates the inputs that have no effect on the outputs and
provides the minimal representative logic of the truth table. This process results in a truth table that is
reduced enough to make the evolved prey behavior comprehensible by humans. An example output
minimal logic table can be seen in Table 6.
Surprisingly, the individual-based mechanisms underlying the emergent swarming behavior are remarkably simple. Most of the prey MNs evolved to make their movement decisions based off of only
one prey sensor in front of the prey agent. If the prey sensor does not detect another prey agent, the
agent repeatedly turns in one direction until it detects another prey agent in that sensor. Once the
agent detects another prey agent in the sensor, it moves forward until the agent is no longer visible.
This mechanism alone proved sufficient to produce cohesive swarming behavior in the majority of our
experiments. Interestingly, this discovery corroborates the findings in earlier studies suggesting that
complex swarming behavior can emerge from simple movement rules when applied over a population
of locally-interacting agents [40, 61, 50].
18
Table 6: An example minimal logic table resulting from a Markov Network. Input 12 corresponds to a
frontal sensor, whereas Input 19 corresponds to a back-right sensor, where 1 indicates that the sensor
detects a prey agent and 0 means no prey are in that sensor. The output 01 translates into the agent
turning right and the output 00 translates into the agent moving forward. Thus, this example agent
moves forward if it sees anything in its frontal sensor. Otherwise, the agent turns right if it sees another
prey in its back-right sensor or if it sees nothing at all.
Input 12 Input 19 Corresponding output
0
0
01
1
0
00
0
1
01
1
1
00
In a small subset of the evolved prey MNs, we observe MNs that occasionally connect to one of the
prey sensors behind them. These MNs watch for a prey agent to appear in a single prey sensor behind
the agent and turn repeatedly in one direction until a prey agent is no longer visible in that sensor. Once
a prey agent is no longer visible in the back sensor, the MN moves forward or turns depending on the
state of the frontal sensor. We note that this mechanism only evolved in prey MNs that already exhibited
swarming behavior using one of the frontal sensors, which suggests that this mechanism does not play a
major role in swarming behavior. Instead, this mechanism seems to cause the prey agent to turn toward
the center of the swarm instead of swarming in a circle with the rest of the prey agents. This mechanism
can be thought of as a “selfish herd” mechanism that attempts to selfishly move the agent toward the
center of the swarm to avoid predation.
7
Conclusions and Future Work
The contributions of this work are as follows. First, we demonstrate Hamilton’s selfish herd hypothesis
in a digital evolutionary model and highlight that it is the attack mode of the predator that critically
determines the evolvability of swarming behavior. Second, we show that density-dependent predation
is sufficient for the selfish herd to evolve as long as the predators cannot consistently attack prey in the
center of the group. Finally, we show that density-dependent predation is sufficient to evolve grouping behavior in prey as a response to predation by coevolving predators. Consequently, future work
exploring the evolution of the selfish herd in animals should not only consider the behavior of the prey
in the group, but the attack mode of the predators as well. Following these experiments, we analyze
the evolved control algorithms of the swarming prey and identify simple, biologically-plausible agentbased algorithms that produce emergent swarming behavior, including a mechanism that produces
“selfish” behavior that drives the prey toward the center of the swarm.
Of course, the evolved prey behavior shown in the videos accompanying this paper may not closely
resemble the anti-predator behavior of many species of group-living prey that we observe in nature. We
provide these videos to demonstrate that grouping-like behavior has indeed evolved—and to confirm
that the swarm density count metric accurately captures when the prey evolve grouping-like behavior—
but we do not seek to claim that we have evolved a particular behavioral phenotype that would match
the grouping behavior we observe in nature. Presumably such grouping behavior has been selected for
by a variety of environmental factors that are not completely captured in this model, which would make
a fascinating venue of research in the future.
While this work shows one method by which the the evolution of grouping behavior can be studied, there remain many different hypotheses explaining the evolution of grouping behavior [30]. Our
future work in this area will focus on directly exploring these hypotheses in similar digital evolutionary
19
models, which has been detailed in [42].
8
Acknowledgments
We thank the three anonymous reviewers for their many insightful comments and suggestions. This
research has been supported in part by the National Science Foundation (NSF) BEACON Center under
Cooperative Agreement DBI-0939454, by NSF grant OCI-1122617, and by Michigan State University
through computational resources provided by the Institute for Cyber-Enabled Research. Any opinions,
findings, and conclusions or recommendations expressed in this material are those of the author(s) and
do not necessarily reflect the views of the NSF.
References
[1] Angeline, P. J. & Pollack, J. B. (1993). Competitive environments evolve better solutions for complex
tasks. In Proceedings of the 5th International Conference on Genetic Algorithms. San Francisco, CA, USA:
Morgan Kaufmann Publishers Inc., (pp. 264–270).
[2] Ballerini et al., M. (2008). Interaction ruling animal collective behavior depends on topological
rather than metric distance: Evidence from a field study. P. Nat. Acad. Sci. USA, 105, 1232–7.
[3] Beauchamp, G. (2004). Reduced flocking by birds on islands with relaxed predation. P. Roy. Soc. B,
271, 1039–42.
[4] Beauchamp, G. (2007). Vigilance in a selfish herd. Anim. Behav., 73, 445–451.
[5] Couzin, I. (2009). Collective cognition in animal groups. Trends Cogn. Sci., 13, 36–43.
[6] Şahin, E. (2004). Swarm robotics: From sources of inspiration to domains of application. In Proceedings of the International Conference on Swarm Robotics (SAB). (pp. 10–20).
[7] Demsar, J., Hemelrijk, C. K., Hildenbrandt, H., & Bajec, I. L. (2015). Simulating predator attacks on
schools: Evolving composite tactics. Ecological Modelling, 304, 22–33.
[8] Diabaté, A., Yaro, A., Dao, A., Diallo, M., Huestis, D., & Lehmann, T. (2011). Spatial distribution
and male mating success of Anopheles gambiae swarms. BMC Evol. Biol., 11, 184.
[9] Dyer, F. (2002). The biology of the dance language. Annu. Rev. of Entomol., 47, 917–49.
[10] Edlund, J., Chaumont, N., Hintze, A., Koch, C., Tononi, G., & Adami, C. (2011). Integrated information increases with fitness in the evolution of animats. PLoS Comput. Biol., 7, e1002236.
[11] Eiben, A. & Smith, J. (2003). Introduction to Evolutionary Computing. Springer-Verlag.
[12] Feare, C. (1984). The Starling. Oxford University Press, USA.
[13] Foster, W. & Treherne, J. (1981). Evidence for the dilution effect in the selfish herd from fish predation on a marine insect. Nature, 293, 466–467.
[14] Goldberg, D. (1989). Genetic Algorithms in Search, Optimization and Machine Learning. Boston, MA:
Addison-Wesley, 1st ed.
[15] Goldbogen, J., Calambokidis, J., Oleson, E., Potvin, J., Pyenson, N., Schorr, G., & Shadwick, R.
(2011). Mechanics, hydrodynamics and energetics of blue whale lunge feeding: Efficiency dependence on krill density. J. Exp. Biol., 214, 131–146.
20
[16] Haley, P., Olson, R., Dyer, F., & Adami, C. (2014). Exploring conditions that select for the evolution
of cooperative group foraging. In Proceedings of the International Conference on the Simulation and
Synthesis of Living Systems (ALIFE). Cambridge, MA, USA: MIT Press, (pp. 310–311).
[17] Hall-Stoodley, L., Costerton, J., & Stoodley, P. (2004). Bacterial biofilms: From the natural environment to infectious diseases. Nat. Rev. Micro., 2, 95–108.
[18] Hamilton, W. (1971). Geometry for the selfish herd. J. Theor. Bio., 31, 295–311.
[19] Hemelrijk, C. & Hildenbrandt, H. (2011). Some causes of the variable shape of flocks of birds. PLoS
ONE, 6, e22479.
[20] Hemelrijk, C. K. & Kunz, H. (2005). Density distribution and size sorting in fish schools: an
individual-based model. Behav. Ecol., 16, 178–187.
[21] Huepe, C. & Aldana, M. (2008). New tools for characterizing swarming systems: A comparison of
minimal models. Physica A, 387, 2809–2822.
[22] Inman, A. & Krebs, J. (1987). Predation and group living. Trends Ecol. Evol., 2, 31–32.
[23] Ioannou, C., Guttal, V., & Couzin, I. (2012). Predatory fish select for coordinated collective motion
in virtual prey. Science, 337, 1212–1215.
[24] Ioannou, C., Tosh, C., Neville, L., & Krause, J. (2008). The confusion effect–from neural networks
to reduced predation risk. Behav. Ecol., 19, 126–130.
[25] James, R., Bennett, P., & Krause, J. (2004). Geometry for mutualistic and selfish herds: The limited
domain of danger. J. Theoret. Biol., 228, 107–13.
[26] Jeschke, J. & Tollrian, R. (2007). Prey swarming: Which predators become confused and why?
Anim. Behav., 74, 387–393.
[27] Kenward, R. (1978). Hawks and doves: Factors affecting success and selection in Goshawk attacks
on Woodpigeons. J. Anim. Ecol., 47, 449–460.
[28] Kersten, M., Britton, R., Dugan, P., & Hafner, H. (1991). Flock feeding and food intake in Little
Egrets: The effects of prey distribution and behaviour. J. Anim. Ecol., 60, 241–252.
[29] Krakauer, D. (1995). Groups confuse predators by exploiting perceptual bottlenecks: A connectionist model of the confusion effect. Behav. Ecol. Sociobiol., 36, 421–429.
[30] Krause, J. & Ruxton, G. (2002). Living in groups. Oxford University Press, USA.
[31] Kunz, H. & Hemelrijk, C. K. (2003). Artificial fish schools: Collective effects of school size, body
size, and body form. Artif. Life, 9, 237–253.
[32] Kunz, H., Züblin, T., & Hemelrijk, C. (2006). On prey grouping and predator confusion in artificial
fish schools. In Proceedings of the International Conference on the Simulation and Synthesis of Living
Systems (ALIFE). (pp. 365–371).
[33] Lenski, R., Ofria, C., Pennock, R., & Adami, C. (2003). The evolutionary origin of complex features.
Nature, 423, 139–44.
[34] Marinakis, Y. & Marinaki, M. (2011). A honey bees mating optimization algorithm for the open
vehicle routing problem. In Proceedings of the Genetic and Evolutionary Computation Conference
(GECCO). (pp. 101–108).
21
[35] Martin, G. (1986). The eye of a passeriform bird, the European starling (Sturnus vulgaris): Eye
movement amplitude, visual fields and schematic optics. J. Comp. Physiol. A, 159, 545–557.
[36] Moran, P. (1962). The statistical processes of evolutionary theory. Clarendon Press.
[37] Mori, M. & Butterworth, D. (2006). A first step towards modelling the krill-predator dynamics of
the Antarctic ecosystem. CCAMLR Sci., 13, 217–277.
[38] Morrell, L., Ruxton, G., & James, R. (2010). Spatial positioning in the selfish herd. Behav. Ecol., 22,
16–22.
[39] Morton, T., Haefner, J., Nugala, V., Decino, R., & Mendes, L. (1994). The selfish herd revisited: Do
simple movement rules reduce relative predation risk? J. Theor. Bio., 167, 73–79.
[40] Olson, R., Hintze, A., Dyer, F., Knoester, D., & Adami, C. (2013). Predator confusion is sufficient to
evolve swarming behaviour. J. Roy. Soc. Interface, 10, 20130305.
[41] Olson, R., Knoester, D., & Adami, C. (2013). Critical interplay between density-dependent predation and evolution of the selfish herd. In Proceedings of the Genetic and Evolutionary Computation
Conference (GECCO). (pp. 247–254).
[42] Olson, R. S. (2015). Elucidating the evolutionary origins of collective animal behavior. Ph.D. thesis,
Michigan State University.
[43] Olson, R. S., Haley, P. B., Dyer, F. C., & Adami, C. (2015). Exploring the evolution of a trade-off
between vigilance and foraging in group-living organisms. Royal Society Open Science, 2.
[44] Parrish, J. (1999). Complexity, pattern, and evolutionary trade-offs in animal aggregation. Science,
284, 99.
[45] Poli, R. (2008). Analysis of the publications on the applications of particle swarm optimisation. J.
Artif. Evol. Appl., 2008, 4:1–4:10.
[46] Pulliam, H. (1973). On the advantages of flocking. J. Theor. Bio., 38, 419–22.
[47] Pulliam, R. & Caraco, T. (1984). Living in groups: Is there an optimal group size? In J. Krebs &
N. Davies (Eds.) Behavioral Ecology: An Evolutionary Approach. Oxford, UK: Blackwell Scientific, 2nd
ed., (pp. 122–147).
[48] Quinn, J. & Cresswell, W. (2006). Testing domains of danger in the selfish herd: Sparrowhawks
target widely spaced redshanks in flocks. P. Roy. Soc. B, 273, 2521–6.
[49] Reluga, T. & Viscido, S. (2005). Simulated evolution of selfish herd behavior. J. Theor. Bio., 234,
213–25.
[50] Reynolds, C. (1987). Flocks, herds and schools: A distributed behavioral model. In Proceedings of
the Conference on Computer Graphics and Interactive Techniques (SIGGRAPH). (pp. 25–34).
[51] Reynolds, C. W. (1993). An evolved, vision-based behavioral model of coordinated group motion.
In Proceedings of the Second International Conference on From Animals to Animats 2 : Simulation of
Adaptive Behavior: Simulation of Adaptive Behavior. Cambridge, MA, USA: MIT Press, (pp. 384–392).
[52] Rudell, R. (1986). Multiple-valued logic minimization for PLA synthesis. Tech. Rep. UCB/ERL
M86/65, EECS Department, University of California, Berkeley.
22
[53] Silva, A., Neves, A., & Costa, E. (2002). An empirical comparison of particle swarm and predator
prey optimisation. In Proceedings of the Irish Conference on Artificial Intelligence and Cognitive Science
(AICS). (pp. 103–110).
[54] Sims, K. (1994). Evolving 3d morphology and behavior by competition. Artif. Life, 1, 353–372.
[55] Spector, L., Klein, J., Perry, C., & Feinstein, M. (2003). Emergence of collective behavior in evolving
populations of flying agents. In Proceedings of the Genetic and Evolutionary Computation Conference
(GECCO). (pp. 61–73).
[56] Tosh, C. (2011). Which conditions promote negative density dependent selection on prey aggregations? J. Theor. Bio., 281, 24–30.
[57] Treherne, J. & Foster, W. (1981). Group transmission of predator avoidance behaviour in a marine
insect: The Trafalgar effect. Anim. Behav., 29, 911–917.
[58] Treherne, J. & Foster, W. (1982). Group size and anti-predator strategies in a marine insect. Anim.
Behav., 30, 536–542.
[59] Treisman, M. (1975). Predation and the evolution of gregariousness. I. Models for concealment and
evasion. Anim. Behav., 23, 779–800.
[60] Vellasques, E., Sabourin, R., & Granger, E. (2012). Gaussian mixture modeling for dynamic particle
swarm optimization of recurrent problems. In Proceedings of the Genetic and Evolutionary Computation
Conference (GECCO). (pp. 73–80).
[61] Vicsek, T., Czirók, A., Ben-Jacob, E., Cohen, I., & Shochet, O. (1995). Novel type of phase transition
in a system of self-driven particles. Phys. Rev. Lett., 75, 1226–1229.
[62] Viscido, S., Millery, M., & Wethey, D. (2002). The dilemma of the selfish herd: The search for a
realistic movement rule. J. Theor. Bio., 217, 183–194.
[63] Viscido, S. V., Miller, M., & Wethey, D. S. (2001). The response of a selfish herd to an attack from
outside the group perimeter. J. Theor. Bio., 208, 315–328.
[64] Ward, C., Gobet, F., & Kendall, G. (2001). Evolving collective behavior in an artificial ecology. J.
Artificial Life, 7, 191–209.
[65] Wood, A. & Ackland, G. (2007). Evolving the selfish herd: Emergence of distinct aggregating
strategies in an individual-based model. P. Roy. Soc. B, 274, 1637–42.
[66] Wood, A. J. (2010). Strategy selection under predation; evolutionary analysis of the emergence of
cohesive aggregations. J. Theor. Bio., 264, 1102–1110.
[67] Xue, B., Zhang, M., & Browne, W. (2012). Multi-objective particle swarm optimisation (PSO) for
feature selection. In Proceedings of the Genetic and Evolutionary Computation Conference (GECCO).
(pp. 81–88).
[68] Yuval, B., Wekesa, J., & Washino, R. (1993). Effect of body size on swarming behavior and mating
success of male Anopheles freeborni (Diptera: Culicidae). J. Insect Behav., 6, 333–342.
23
Supplementary Text
Olson et al.: Evolution of swarming behavior is shaped by how predators attack.
Predator-prey coevolution experiments for 10,000 generations
To ensure that the “No Density Dependent Predation” treatment never evolves swarming behavior,
we ran this treatment out for a full 10,000 generations, or over 8x the original number of generations.
Figure S1 depicts the result of these experiments, where the prey populations evolve dispersive behavior
by generation 1,000 and maintain the same behavior over the whole evolutionary time span.
40
No Density Dependent Predation
35
Mean Nearby Prey
30
25
20
15
10
5
0
0
1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
Generation
Figure S1: Mean swarm density over all replicates over evolutionary time, measured by the mean number of prey within 30 virtual meters of each other over a lifespan of 2,000 simulation time steps. Prey
under the “No Density Dependent Predation” treatment never evolves swarming behavior even after
10,000 generations of evolution. Error bars indicate two standard errors over 30 replicates.
24
Effect of relative predator-prey speed in predator-prey coevolution experiments
To explore the role of relative predator and prey speed in our evolutionary digital model, we reran the
predator-prey coevolution experiments with and without density-dependent predation and with varying relative predator and prey speeds. In the “Predator faster” treatment, the predators move 3x faster
than the prey, whereas in the “Predator slower” treatment, the predators move 0.5x the speed of the
prey. As shown in Figure S2, regardless of whether density-dependent predation is in effect, swarming
behavior only evolves in the prey when the predators are faster than the prey. In the treatments where
the predator is the same speed or slower, the prey are simply able to outrun the predators, and therefore
do not need to evolve swarming behavior as a defensive response to predation.
60
Density Dependent Predation - Predator faster
No Density Dependent Predation - Predator faster
Density Dependent Predation - Predator & prey same speed
No Density Dependent Predation - Predator & prey same speed
Density Dependent Predation - Predator slower
No Density Dependent Predation - Predator slower
Mean Nearby Prey
50
40
30
20
10
0
0
100 200 300 400 500 600 700 800 900 1000 1100 1200
Generation
Figure S2: Mean swarm density over all replicates over evolutionary time, measured by the mean number of prey within 30 virtual meters of each other over a lifespan of 2,000 simulation time steps. Prey only
evolve swarming behavior when the predator is faster and they are experiencing density-dependent
predation. Error bars indicate two standard errors over 30 replicates.
25
| 9cs.NE |
arXiv:1801.07230v1 [cs.CV] 22 Jan 2018
DiscrimNet: Semi-Supervised Action Recognition from Videos using Generative
Adversarial Networks
Unaiza Ahsan
Georgia Institute of Technology
Chen Sun
Google
Irfan Essa
Georgia Institute of Technology
uahsan3@gatech.edu
chensun@google.com
irfan@gatech.edu
Abstract
We propose an action recognition framework using Generative Adversarial Networks. Our model involves training a deep convolutional generative adversarial network
(DCGAN) using a large video activity dataset without label information. Then we use the trained discriminator
from the GAN model as an unsupervised pre-training step
and fine-tune the trained discriminator model on a labeled
dataset to recognize human activities. We determine good
network architectural and hyperparameter settings for using the discriminator from DCGAN as a trained model to
learn useful representations for action recognition. Our
semi-supervised framework using only appearance information achieves superior or comparable performance to
the current state-of-the-art semi-supervised action recognition methods on two challenging video activity datasets:
UCF101 and HMDB51.
1. Introduction
One of the biggest challenges in recognizing activities in
videos is obtaining large labeled video datasets. Annotating videos is largely both expensive and cumbersome due
to variations in viewpoint, scale and appearance within a
video. This suggests a need for semi-supervised approaches
to recognize actions in videos. One such approach is to use
deep networks to learn a feature representation of videos
without activity labels but with temporal order of frames
as a ‘weak supervision’ [1, 2]. This approach still requires
some supervision in terms of deciding sampling strategies
and related video encoding methods to input to neural networks (such as dynamic images [3]) and designing ‘good
questions’ of correct/incorrect orders as input to the deep
network.
Generative models such as the recently introduced Generative Adversarial Networks (GANs) [4] approximate high
dimensional probability distributions like those of natural
images using an adversarial process without requiring ex-
pensive labeling. To this end, our research question is: How
can we use abundant video data without labels to train a
generative model such as a GAN and use it to learn action
representation in videos with little to no supervision?
GANs are conventionally used to learn a data distribution
of images starting from random noise. Adversarial learning in GANs involves two networks: a discriminator network and a generator network. The discriminator network
is trained on two kinds of inputs – one consisting of samples
drawn from a high dimensional data source such as images
and the other consisting of random noise. Its goal is to distinguish between real and generated samples. The generator network uses the output of a discriminator to generate
‘better’ samples. This minimax game aims to converge to
a setting where the discriminator is unable to distinguish
between real and generated samples. We propose to use
the discriminator trained to only differentiate between a real
and generated sample for learning a feature representation
of actions in videos.
We use the GAN setup to train a discriminator network
and use the learned representation of discriminator as “initialized weight.” Then fine-tune that discriminator on labeled video dataset such as UCF101 [5]. Recent works have
done small experiments [6] but to our knowledge, nobody
has done an in-depth study and especially considered all the
architecture/hyperparameter settings that can yield a good
performance across datasets (we do well on HMDB51 too)
using only appearance information in the video. This unsupervised pre-training step avoids any manual feature engineering, video frame encoding, searching for the best video
frame sampling technique and results in an action recognition performance competitive to the state-of-the-art using
only appearance information.
Our key contributions and findings are:
• We propose a systematic semi-supervised approach to
learn action representations from videos using GANs.
• We perform a comprehensive study of best practices to
recognize actions from videos using the GAN training
process as a good initialization step for recognition.
Figure 1: Our approach to learn action representation from GANs
• We find that appearance-based unsupervised pretraining for video action recognition performs superior
or comparable to the state-of-the-art semi-supervised
multi-stream video action recognition approaches.
• Our unsupervised pre-training step does not require
weak supervision or computationally expensive steps
in the form of video frame encoding, video stabilization and search for best sampling strategies.
2. Related Work
To date, action recognition is one problem in Computer
Vision where deep Convolutional Neural Networks (CNNs)
have not outperformed hand-crafted features. Action recognition from videos has come a long way from holistic feature learning such as Motion Energy Image (MEI) and
Motion History Image (MHI) [7], space-time volumes [8]
and Action Banks [9] to local feature learning approaches
such as space-time interest points [10], HOG3D [11], histogram of optical flow [12] and tracking feature trajectories [13, 14, 15, 16].
The recent success of CNNs in image recognition has
enabled many researchers to treat a video as a set of RGB
images, perform image classification on the video frames
and aggregate the network predictions to achieve video level
classification [17]. Our approach is also inspired by local
appearance encoding methods for videos. 3D convolutional
networks capture spatio-temporal features via 3D convolutions in both spatial and temporal domains [18]. Various
fusion techniques are proposed to pool the temporal information to construct video descriptors [19, 20]. Recurrent
Neural Networks (RNNs) and Long Short Term Memory
(LSTM) networks have also been used to model videos for
action recognition [21, 22]. Using multiple networks to
model appearance and motion was first introduced by Simonyan and Zisserman [17]: the two-stream architecture,
where the spatial architecture is the standard VGG Net [23]
and the temporal stream network takes input stacked optical
flow fields. Wu et al. [24] added audio and LSTMs to the
network to improve video classification performance. We
do not experiment with multiple modalities in this paper as
we use only RGB frames as input to the model for our proof
of concept.
Generative models have been successfully used to avoid
manual supervision in labeling videos with the most common application being video frame prediction [25, 26, 27,
28, 29, 30, 27, 31, 32]. Since appearance changes are
smooth across videos, temporal consistency [33] and other
constraints [34] are useful to learn video representations.
Our work proposes a generative model as an unsupervised pre-training method for action recognition. While approaches that take temporal coherency into account such as
[1, 28, 35, 36] are similar to our work, they are different
in that enforcing temporal coherency still involves weak supervision [1] where they have to pre-select good samples
from a video. We do not do any weak supervision in our
approach but only use the generative adversarial training as
an unsupervised pre-training step to recognize actions.
Recently [2] train a network to predict the odd video out
of a set of videos where the “odd one out” is a video with
its frames in wrong temporal order. The key difference between our work and theirs is that we do not require any
weak supervision in terms of selecting the right video encoding method, sampling strategies or designing effective
odd-one-out questions to improve accuracy. Another recent
related approach is that of [37] where a network is trained
to sort a tuple of frames from videos. This sequence sorting
task forms the “unsupervised pretraining” step and the network is finetuned on labelled datasets. Our approach does
not use weak supervision (such as selecting the right tuple
via optical flow for example) for the unsupervised pretraining task and uses only appearance information in this work.
Generative Adversarial Networks [4] have been used for
semi-supervised feature learning particularly after the introduction of Deep Convolutional GANs (or DCGANs) [38].
Radford [38] et al. use the discriminator (pre-trained on
ImageNet) to compute features on CIFAR10 dataset [39]
for classification. Other works to use GANs for semisupervised learning [40, 41, 42, 43, 44] are all designed for
image recognition, not videos.
A recent work is [6] where the authors train GANs for
tiny video generation. They fine-tune their trained discriminator model on UCF101 and show promising results.
However, their model is significantly more complicated and
requires stabilized videos which involves SIFT [45] and
RANSAC [46] computation per video frame, something
that is not required by our method which achieves comparable accuracy after finetuning.
3. Approach
We briefly review the main principles behind GAN models and describe our methodology in detail to recognize actions by leveraging their unsupervised feature learning capability on videos.
3.1. Generative Adversarial Networks
GAN networks [4] exploit game theoretic approaches to
train two different networks; a generator and a discriminator. The generator represented by function G parameterized
by θ(G) starts with an input noise vector z that is sampled
from a normal distribution pnoise (z), up-samples this noise
ˆ The discriminator netdistribution and outputs an image I.
work is a CNN network (represented by function D) parameterized by θ(D) that takes as input an image (I (real
image) or Iˆ (generated or fake image)) and outputs a probability ∈ {0, 1} that whether the input image is from the
real distribution or generated distribution. Training GANs
involve a minimax game in which the generator attempts to
‘fool’ the discriminator into predicting a generated image
as real whereas the discriminator attempts to identify correctly which input images are fake. The discriminator cost
function is a cross entropy loss defined by:
J (D) (θ(D) , θ(G) ) = EI∼pdata(I) [logD(I)] + Ez∼pnoise(z) [log(1 − D(G(z)))]
(1)
The minimax objective from Equation 1 can be optimized using gradient-based methods since both discriminator and generator are functions (D and G) that are differentiable with respect to their inputs and parameters [47]. The
solution to this problem is a Nash equilibrium as both functions are trained to minimize their costs while maximizing
the other’s objective. GANs can be trained using Stochastic
Gradient Descent (SGD) with any optimizer of choice.
Figure 2: Results after 100 epochs of running DCGAN [38]
on UCF101 video frames. The images in the top three rows
are real while those on the bottom are generated by the
model
3.2. Training GANs with Video Frames
So far in the research community, GANs have been primarily used for sample generation. Thus, focus has been on
modifying generator parameters, network architecture and
loss functions in order to generate higher resolution images
with minimal artifacts. The discriminator network in all
variants of GANs is trained with binary cross entropy loss
(see Equation 1) [47]. Since our focus is not image generation but learning useful features to transfer to the task
of action recognition, we are motivated to train and use the
discriminator network in GANs for action recognition. The
discriminator network in a GAN learns a representation of
local appearance features thus modeling objects and scenes
in video frames as context. Lastly, it does so in an unsupervised manner i.e. we do not require explicit labels for
objects, scenes or actions to pre-train our action recognition
model.
Consider a set of videos V where V = {V1 , ..., Vn } and
n is the number of videos in the dataset. Each video consists
of a variable number of frames (sampled at the rate of one
frame per second). We use all the frames in the training set
of videos from two challenging video activity datasets without any label information to train the GAN model. Our approach is shown in Figure 1. We train GANs using a variety
of techniques proposed in prior research to generate images.
To compare with GANs pre-trained on an object recognition
dataset, we also train a GAN model on ImageNet [48] images. We use the same architecture as proposed in the DCGAN [38] paper since the authors have demonstrated the
transfer learning capability of DCGAN model on CIFAR10
dataset.
3.3. Unsupervised Pre-training
When dealing with small datasets, a CNN’s generalization performance decreases so that the test accuracy remains
small even while training accuracy may increase. This is
why a common practice is to initialize the weights of the
layers with ImageNet pre-trained CNN weights instead of
training from scratch. This is referred to as supervised pretraining since ImageNet labels have been used to determine
the initial weights.
Our approach is different in that we are trying to do unsupervised pre-training - determining starting weights for
a CNN model (discriminator) which is pre-trained without
label information using adversarial training. This unsupervised pre-training setup is compared with initializing the
weights in the discriminator network using other settings
and we show that the GAN-based initialization significantly
outperforms other initialization strategies on the test set of
UCF101.
3.4. Fine-tuning Discriminator Model
In this step of our approach we initialize the network
with the learned weights from adversarial training and finetune it on two video activity datasets. In the process of
fine-tuning, we are faced with numerous choices of network
architecture, learning rate schemes, optimization and data
augmentation. We explore in the space of these variations
and report all results on the test split 1 of UCF101 dataset.
4. Experiments
4.1. Datasets
UCF101 [5] is a benchmark action recgonition dataset
comprising 13320 YouTube videos of 101 action categories.
Actions include human-object interactions such as ‘apply
lipstick’, body motion such as ‘handstand walking’, humanhuman interactions, playing musical instruments and sports.
The dataset is small but challenging in that the videos
vary in viewpoint changes, illumination, camera motion and
Figure 3: Sample frames from the UCF101 dataset [5] with
action classes (from top to bottom): apply eye makeup, juggling balls and rowing
Figure 4: Sample frames from the HMDB51 dataset [49]
blur. The second dataset we experiment on is the HMDB51
dataset [49] which contains 6766 videos of 51 actions such
as chew, eat, laugh etc. Sample frames from both datasets
are shown in Figures 3 and 4.
4.2. Unsupervised Pre-training
This section describes three experiments to determine:
(a) Whether GANs can generate action images (b) Training
Protocol of GANs and (c) Data Augmentation steps
Can GANs Generate Action Images? Since we consider
a video as a set of RGB frames, we address the first question: Are GANs, traditionally used for generating faces,
objects and scenes capable of generating an image representing an action? This question is crucial to address because it determines the validity of using the trained GAN
discriminator as a CNN network and fine-tune it on a labelled video activity dataset. To answer this question, we
use all the videos from the train split 1 of UCF101 [5] and
sample 1 frame per second from each video. We train a DCGAN model with default parameters and after 100 epochs,
obtain results shown in Figure 2. From visual inspection
we can see that vanilla DCGAN is able to learn a coarse
representation of activities involving humans. The question now remains whether we can use the feature representation learned by GAN’s discriminator as an unsupervised
pre-training step to classify actions in labeled video action
recognition datasets.
Training Protocol of GANs: We use DCGAN’s public
implementation in torch and train three separate GAN models: One with UCF101 video frames, second with ImageNet
[48] images and third with a subset of Sports1M dataset
[50] frames. We train all three models for 100 epochs using
the architectural guidelines proposed in [38], namely, batch
normalization [51] in discriminator as well as the generator,
leaky Rectified Linear Units (leaky ReLU) [52] in all layers
of discriminator, strided convolutions in discriminator instead of pooling layers and fractional-strided convolutions
in the generator. There are no fully-connected (FC) layers
in the DCGAN architecture as the authors of [38] report no
loss in generator performance for not including FC layers.
Hence we also use the same architecture for training the
GAN model.
Data Augmentation: The main difference between our
GAN training and the DCGAN [38] approach is that DCGAN [38] performs data augmentation via taking 64 x 64
sized random crops of the image as well as scaling the images to range [-1,1]. This scaling is done for the tanh activation function in the generator. We change that protocol
and avoid random cropping. We only scale the frames of
videos to the range [-1,1] and scale the size to 64 x 64. The
reason why we avoid random cropping is because the action
frames from videos are much larger and contain much more
information than the original images used for training DCGAN (bedrooms, faces and the like). Taking random crops
from action frames will not result in a useful representation
because too much information will be lost. Thus, we only
scale the images to 64x64 as our aim is not just to generate
action images but to learn an effective action representation
for recognition.
4.3. Fine-tuning for Action Recognition
Here we describe the set of experiments conducted after
the GAN model has been trained. We use the pre-trained
discriminator network from our GAN model and fine-tune
it on the two labeled video action datasets: UCF101 [5]
and HMDB51 [49]. We begin by replacing the last spatial
convolutional layer (CONV5) with one that has the correct
number of outputs (equal to the number of action classes).
See Figure 5. This layer is initialized randomly and the
network is trained again with the previous layers initialized
with the pre-trained discriminator’s weights.
We perform a comprehensive experimental analysis of
architectural choices, hyperparameter settings and other
good practices and report the accuracy on the test set of
UCF101 dataset.
Does Source Data Distribution Matter? In this experiment, we determine whether the dataset we train GAN with
(which we refer to as the source dataset) determines performance on the target dataset (the labeled dataset on which
we fine-tune the discriminator network). To this end, we
train DCGAN on three large scale datasets: ImageNet [48]
images, UCF101 [5] video frames and frames of 10,000
videos from Sports1M [19] dataset. We use the same sampling strategy of 1 frame per second for both video datasets
and train all three GAN models separately for 100 epochs.
For each video Vi , there is a set of frames Fi where
Fi = V = {fn1 , fn2 , ..., fni } where ni is the number of
frames extracted for video Vi . Each video’s frames are
passed through the trained GAN’s discriminator and we extract CONV4’s activations as features on each frame. We
average frame-level features to obtain video-level features.
Source Dataset
ImageNet
UCF101
Sports1M
Destination Dataset (accuracy %)
UCF101
HMDB51
43.88
12.82
47.20
12.94
42.50
13.02
Table 1: Comparing the accuracy on target dataset with
three large scale datasets used to train GAN models
We train a linear SVM classifier [53] on top of these features using the train/test split1 provided by the dataset authors and obtain classification accuracy on the test set. We
use the same setting for training all three GAN models as
described in the training protocol earlier. Our results are
shown in Table 1.
As can be seen from Table 1 training a GAN with
UCF101 frames results in the best test accuracy on both
UCF101 and HMDB51. The difference between training
a GAN model with ImageNet and Sports1M frames and
training it with UCF101 frames is significant. Note that we
did not use all videos from the Sports1M dataset; we randomly selected 10,000 videos from the dataset, extracted 1
frame per second from each video and used those frames
to train the GAN model. For HMDB51 dataset the difference in test accuracy between using a GAN discriminator
pre-trained on UCF101 and other datasets is not very large.
But the superior performance of training a GAN model with
video action frames is clearly demonstrated by this experiment. The features learned by the discriminator network are
strong enough to transfer to other video datasets as well.
Choice of Architecture: There are several ways of
changing the architecture of the pre-trained discriminator
network for fine-tuning. Note that the discriminator is just
another CNN network with spatial convolutional layers and
no fully connected layers. For fine-tuning on the UCF101
dataset, we replace the last convolutional layer (CONV4)
with one that has the correct number of outputs, initialize
this layer randomly and train this network (fine-tune) for
160 epochs. This fine-tuning experiment is called ‘CONV4’
in Table 2. Network depth determines the model’s performance both in theory and practice [54]. Hence we add another convolutional layer (CONV5) and a fully connected
layer (FC), initialize them from scratch and retrain the network till convergence. We extract CONV4, CONV5 and
FC features from the finetuned network. We concatenate
CONV5 and CONV5 features and test the performance as
well as CONV4, CONV5 and FC features. We do not
freeze any layers before fine-tuning and keep a learning rate
of 0.001 to fine-tune the network. We empirically found
that freezing the earlier layers and finetuning only the last
layer(s) did not increase performance. We train a linear
SVM on top of the extracted features and compute results
Figure 5: Our network architecture: DCGAN discriminator architecture on the left and our added layers on the right
on UCF101’s test set. Our results are shown in Table 2. Our
network architecture is shown in Figure 5.
Our results in Table 2 show that with all other parameters kept the same, adding a convolutional layer and a fully
connected layer in the discriminator network architecture
results in only a slight improvement in performance. We
note that this is not a huge difference and this may seem
counterintuitive but the reason why this happens is that we
are initializing the added network layers randomly before
fine-tuning. Also, the dataset size of UCF101 frames is not
very large with 84,747 frames in the training set and 33,187
frames in the test set. This may lead to over fitting resulting in only a slight increase in performance on the test set
especially when the fully connected layer is added.
To reduce overfitting, we add dropout [55] after the additional convolutional and fully connected layers. We note the
performance with/without dropout by extracting CONV4
features from both networks (after finetuning) and training a
linear SVM. Adding dropout regularizes the network more
thus increasing the performance on test set of UCF101.
Fine-tuning vs Linear SVM: Once we fine-tune the discriminator model on the datasets, we have a choice of
whether to extract the CONV4’s activations and train a linear SVM on top of it or fine-tune the last layers with softmax classifier. We do both in our experiments and note that
the outcome is dependent on the dataset. We find that when
Architectural changes
CONV4
CONV4 + CONV5 + FC
CONV4 + CONV5
Test Accuracy (%)
48.35
49.30
50.12
Table 2: Effect of making the network deeper: Adding
more layers slightly improves action recognition
performance
we fine-tune the discriminator network on UCF101, the test
set accuracy using softmax is lower than extracting CONV4
features and training a linear SVM to recognize actions.
However when using HMDB51, the softmax classification
on the test set results in a higher accuracy than extracting
Layer 9 features and training a linear SVM classifier. This
result is shown in Table 3.
UCF101
HMDB51
Accuracy (%) on test set
CONV4 + linear SVM Softmax
48.35
41.40
14.40
21.04
Table 3: Comparing two ways of evaluating fine-tuned
network performance on UCF101 and HMDB51 test sets
From Table 3 it is apparent that for UCF101, feature embedding and training a linear SVM results in a better accuracy than softmax classification. The complete opposite
is true with HMDB51 dataset. We dig deeper to investigate why this happens. We find that the label distribution
of the dataset on which a deep network is being fine-tuned
on is the key to determine which method results in a better
test accuracy. The label distribution of UCF101 test set is
shown in Figure 6. This distribution is not balanced while
that of HMDB51 is completely balanced in terms of number
of videos per action category. Hence it appears that when
classes are unbalanced, since we have not used weighted
loss in the neural network, the linear SVM learns the features better hence resulting in an increased performance on
the test set. In the case of HMDB51, all classes are balanced
equally leading to the superior performance of the softmax
classifier over the feature embedding approach.
Unsupervised Pretraining vs Random Initialization
We validate the use of our unsupervised pre-training approach by comparing it with a network that is initialized
Method
STIP features [12]
DrLim [57]
TempCoh [31]
Obj. Patch [35]
Shuffle [1]
VideoGAN [6]
O3N [2]
OPN [37]
DiscrimNet (ours) CONV4 + linear SVM
DiscrimNet (ours) CONV5 + linear SVM
DiscrimNet (ours) (CONV4 + CONV5) + linear SVM
UCF101-split1 (%)
43.9
45.7
45.4
40.7
50.9
52.1
60.3
56.3
49.33
48.88
50.12
Table 5: Comparing our method to state-of-the-art
semi-supervised approaches on UCF101
Figure 6: Label distributions of UCF101 test set. The
HMDB51 dataset has uniform distribution of 30 videos per
action class
Initialization
Xavier + finetuning
DiscrimNet (ours) + finetuning
UCF101 (%)
33.10
49.30
Method
DrLim [57]
TempCoh [31]
Obj. Patch [35]
Shuffle [1]
O3N [2]
OPN [37]
DiscrimNet (ours) (fine-tuned)
HMDB51 (%)
11.6
20.4
Table 4: Validating the use of our unsupervised pretraining
approach vs training with random initialization
randomly. We initialize all the layers of the network using ‘xavier’ initialization. Proposed by [56], the authors
recommend initializing weights by drawing from a distribution with zero mean and variance given by: V ar(W ) =
2/(nin + nout ) where W is the distribution which the neuron is initialized with, nin is the number of neurons feeding
into the layer and nout is the number of output neurons from
this layer. We initialize all layers with this scheme and train
the network till convergence on UCF101. For HMDB51,
we train a network for 50 epochs with xavier initialized layers and compare that to our proposed discriminator initialized method at 50 epochs. The results are shown in Table 4
and clearly validate the use of our unsupervised pretraining
approach to initialize the network before finetuning. As a
reference, a supervised ImageNet pretrained network finetuned on UCF101 yields an accuracy of 67.1% and finetuned on HMDB51 yields an accuracy of 28.5% [1].
Table 6: Comparing our method to state-of-the-art
semi-supervised approaches on HMDB51
•
•
•
•
4.4. Comparison with the state-of-the-art
We compare our approach with several recent semisupervised baselines which recognize actions in videos.
The baselines are:
• STIP features: Handcrafted Space Time Interest Point
(STIP) features introduced by [10].
• DrLim [57]: This method uses temporal coherency by
minimizing the L2 distance metric between features of
neighboring frames in videos and enforcing a margin
δ between far apart frames.
• TempCoh [31]: Enforce temporal coherencFrom the
mid-1980s through 2015 the average number of acres
HMDB51 (%)
16.3
15.9
15.6
19.8
32.5
22.1
21.0
•
burned has grown from about 2 million acres a year to
around 8 millione by using L1 distance instead of L2.
Similar to DrLim [57].
Obj. Patch [35]: They extract similar object patches
using videos and learn a representation of objects by
tracking them through time. This model is used and
fine-tuned on UCF101 by [1].
Shuffle [1]: They use sequence verification as an unsupervised pre-training step for vidoes. The model is
then fine-tuned on UCF101.
VideoGAN [6]: They generate tiny videos using a two
stream GAN network. Their model is fine-tuned on
UCF101.
O3N [2]: They use odd-one-out networks to predict
the wrong temporal order from the right ones. Their
model is then fine-tuned on UCF101.
OPN [37]: They train a network to predict the order
of 4-tuple frames. Their model is then fine-tuned on
UCF101.
The results are shown in Table 5 and Table 6.
5. Discussion
Our comparison with several state-of-the-art semisupervised approaches to recognize actions in vidoes yields
important insights. Our results show competitive performance as compared to the state-of-the-art approaches in
semi-supervised learning given that:
• We only use appearance features and do not experiment with motion content of the video. This is especially intriguing given that our method outperforms
STIP features on this dataset. All methods in the results we compare to use temporal coherency as a signal
and do motion encoding.
• We do not do weak supervision in the form of temporal
consistency and do not design temporal order based
networks. The only supervision provided to the GAN
is the difference between a real image and noise.
• Our model outperforms several state-of-the-art approaches on HMDB51 given that no video from the
dataset was used in the unsupervised pre-training step
of this approach. This shows the domain adaptation capability of GAN discriminator networks and that they
are able to capture enough information to learn useful
representation of actions in video frames.
The methods that outperform our proposed approach are
either computationally expensive or require much more supervision in the form of selecting sampling strategies, video
encoding methods or in the case of O3N networks [2], designing effective odd-one-out questions for the network to
learn feature representations for action recognition.
6. Conclusion
We propose an unsupervised pre-training method using
GANs for action recognition in videos. Our method does
not require weak supervision in the form of temporal coherency, sampling selection or video encoding methods.
Purely on appearance information alone, our method performs either better than or comparable to the state-of-the-art
semi-supervised action recognition methods.
References
[1] Ishan Misra, C Lawrence Zitnick, and Martial Hebert. Shuffle and learn: unsupervised learning using temporal order
verification. In European Conference on Computer Vision,
pages 527–544. Springer, 2016.
[2] Basura Fernando, Hakan Bilen, Efstratios Gavves, and
Stephen Gould.
Self-supervised video representation
learning with odd-one-out networks.
arXiv preprint
arXiv:1611.06646, 2016.
[3] Hakan Bilen, Basura Fernando, Efstratios Gavves, Andrea
Vedaldi, and Stephen Gould. Dynamic image networks for
action recognition. In Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition, pages 3034–
3042, 2016.
[4] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing
Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and
Yoshua Bengio. Generative adversarial nets. In Z. Ghahramani, M. Welling, C. Cortes, N. D. Lawrence, and K. Q.
Weinberger, editors, Advances in Neural Information Processing Systems 27, pages 2672–2680. Curran Associates,
Inc., 2014.
[5] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah.
Ucf101: A dataset of 101 human actions classes from videos
in the wild. arXiv preprint arXiv:1212.0402, 2012.
[6] Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba.
Generating videos with scene dynamics. arXiv preprint
arXiv:1609.02612, 2016.
[7] Aaron F. Bobick and James W. Davis. The recognition of human movement using temporal templates. IEEE
Transactions on pattern analysis and machine intelligence,
23(3):257–267, 2001.
[8] Alper Yilmaz and Mubarak Shah. Actions sketch: A novel
action representation. In 2005 IEEE Computer Society
Conference on Computer Vision and Pattern Recognition
(CVPR’05), volume 1, pages 984–989. IEEE, 2005.
[9] Sreemanananth Sadanand and Jason J Corso. Action bank:
A high-level representation of activity in video. In Computer
Vision and Pattern Recognition (CVPR), 2012 IEEE Conference on, pages 1234–1241. IEEE, 2012.
[10] Ivan Laptev. On space-time interest points. International
Journal of Computer Vision, 64(2-3):107–123, 2005.
[11] Alexander Klaser, Marcin Marszałek, and Cordelia Schmid.
A spatio-temporal descriptor based on 3d-gradients. In
BMVC 2008-19th British Machine Vision Conference, pages
275–1. British Machine Vision Association, 2008.
[12] Ivan Laptev, Marcin Marszalek, Cordelia Schmid, and Benjamin Rozenfeld. Learning realistic human actions from
movies. In Computer Vision and Pattern Recognition, 2008.
CVPR 2008. IEEE Conference on, pages 1–8. IEEE, 2008.
[13] Ross Messing, Chris Pal, and Henry Kautz. Activity recognition using the velocity histories of tracked keypoints. In
2009 IEEE 12th international conference on computer vision, pages 104–111. IEEE, 2009.
[14] Pyry Matikainen, Martial Hebert, and Rahul Sukthankar.
Trajectons: Action recognition through the motion analysis
of tracked features. In Computer Vision Workshops (ICCV
Workshops), 2009 IEEE 12th International Conference on,
pages 514–521. IEEE, 2009.
[15] Yu-Gang Jiang, Qi Dai, Xiangyang Xue, Wei Liu, and
Chong-Wah Ngo. Trajectory-based modeling of human actions with motion reference points. In European Conference
on Computer Vision, pages 425–438. Springer, 2012.
[16] Heng Wang and Cordelia Schmid. Action recognition with
improved trajectories. In Proceedings of the IEEE International Conference on Computer Vision, pages 3551–3558,
2013.
[17] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In
Advances in Neural Information Processing Systems, pages
568–576, 2014.
[18] Shuiwang Ji, Wei Xu, Ming Yang, and Kai Yu. 3d convolutional neural networks for human action recognition. IEEE
transactions on pattern analysis and machine intelligence,
35(1):221–231, 2013.
[31] Hossein Mobahi, Ronan Collobert, and Jason Weston. Deep
learning from temporal coherence in video. In Proceedings
of the 26th Annual International Conference on Machine
Learning, pages 737–744. ACM, 2009.
[19] Andrej Karpathy, George Toderici, Sanketh Shetty, Thomas
Leung, Rahul Sukthankar, and Li Fei-Fei. Large-scale video
classification with convolutional neural networks. In Proceedings of the IEEE conference on Computer Vision and
Pattern Recognition, pages 1725–1732, 2014.
[32] Graham W Taylor, Rob Fergus, Yann LeCun, and Christoph
Bregler. Convolutional learning of spatio-temporal features.
In European conference on computer vision, pages 140–153.
Springer, 2010.
[20] Joe Yue-Hei Ng, Matthew Hausknecht, Sudheendra Vijayanarasimhan, Oriol Vinyals, Rajat Monga, and George
Toderici. Beyond short snippets: Deep networks for video
classification. In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, pages 4694–
4702, 2015.
[21] Moez Baccouche, Franck Mamalet, Christian Wolf,
Christophe Garcia, and Atilla Baskurt. Sequential deep
learning for human action recognition. In International
Workshop on Human Behavior Understanding, pages 29–39.
Springer, 2011.
[22] Jeffrey Donahue, Lisa Anne Hendricks, Sergio Guadarrama,
Marcus Rohrbach, Subhashini Venugopalan, Kate Saenko,
and Trevor Darrell. Long-term recurrent convolutional networks for visual recognition and description. In Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition, pages 2625–2634, 2015.
[23] Ken Chatfield, Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman.
Return of the devil in the details:
Delving deep into convolutional nets.
arXiv preprint
arXiv:1405.3531, 2014.
[24] Zuxuan Wu, Yu-Gang Jiang, Xi Wang, Hao Ye, Xiangyang
Xue, and Jun Wang. Fusing multi-stream deep networks for
video classification. arXiv preprint arXiv:1509.06086, 2015.
[25] Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to
sequence learning with neural networks. In Advances in neural information processing systems, pages 3104–3112, 2014.
[26] Yipin Zhou and Tamara L Berg. Temporal perception and
prediction in ego-centric video. In Proceedings of the IEEE
International Conference on Computer Vision, pages 4498–
4506, 2015.
[27] Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Anticipating the future by watching unlabeled video. arXiv
preprint arXiv:1504.08023, 2015.
[28] Ross Goroshin, Joan Bruna, Jonathan Tompson, David
Eigen, and Yann LeCun. Unsupervised learning of spatiotemporally coherent metrics. In Proceedings of the IEEE
International Conference on Computer Vision, pages 4086–
4093, 2015.
[29] Nitish Srivastava, Elman Mansimov, and Ruslan Salakhutdinov. Unsupervised learning of video representations using
lstms. CoRR, abs/1502.04681, 2, 2015.
[30] Michael Mathieu, Camille Couprie, and Yann LeCun. Deep
multi-scale video prediction beyond mean square error.
arXiv preprint arXiv:1511.05440, 2015.
[33] Zhang Zhang and Dacheng Tao. Slow feature analysis for
human action recognition. IEEE Transactions on Pattern
Analysis and Machine Intelligence, 34(3):436–450, 2012.
[34] Dinesh Jayaraman and Kristen Grauman. Slow and steady
feature analysis: higher order temporal coherence in video.
arXiv preprint arXiv:1506.04714, 2015.
[35] Xiaolong Wang and Abhinav Gupta. Unsupervised learning
of visual representations using videos. In Proceedings of the
IEEE International Conference on Computer Vision, pages
2794–2802, 2015.
[36] Xiaolong Wang, Ali Farhadi, and Abhinav Gupta. Actions˜
transformations. arXiv preprint arXiv:1512.00795, 2015.
[37] Hsin-Ying Lee, Jia-Bin Huang, Maneesh Singh, and MingHsuan Yang. Unsupervised representation learning by sorting sequences. In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, pages 667–676,
2017.
[38] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks.
arXiv preprint
arXiv:1511.06434, 2015.
[39] Alex Krizhevsky. Learning multiple layers of features from
tiny images. 2009.
[40] Xi Chen, Yan Duan, Rein Houthooft, John Schulman, Ilya
Sutskever, and Pieter Abbeel. Infogan: Interpretable representation learning by information maximizing generative
adversarial nets. arXiv preprint arXiv:1606.03657, 2016.
[41] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki
Cheung, Alec Radford, and Xi Chen. Improved techniques
for training gans. arXiv preprint arXiv:1606.03498, 2016.
[42] Jost Tobias Springenberg.
Unsupervised and semisupervised learning with categorical generative adversarial
networks. arXiv preprint arXiv:1511.06390, 2015.
[43] Jeff Donahue, Philipp Krähenbühl, and Trevor Darrell. Adversarial feature learning. arXiv preprint arXiv:1605.09782,
2016.
[44] Augustus Odena. Semi-supervised learning with generative adversarial networks. arXiv preprint arXiv:1606.01583,
2016.
[45] David G Lowe. Distinctive image features from scaleinvariant keypoints. International journal of computer vision, 60(2):91–110, 2004.
[46] Martin A Fischler and Robert C Bolles. Random sample
consensus: a paradigm for model fitting with applications to
image analysis and automated cartography. Communications
of the ACM, 24(6):381–395, 1981.
[47] Ian Goodfellow. Nips 2016 tutorial: Generative adversarial
networks. arXiv preprint arXiv:1701.00160, 2016.
[48] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li,
and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Computer Vision and Pattern Recognition,
2009. CVPR 2009. IEEE Conference on, pages 248–255.
IEEE, 2009.
[49] H. Kuehne, H. Jhuang, E. Garrote, T. Poggio, and T. Serre.
HMDB: a large video database for human motion recognition. In Proceedings of the International Conference on
Computer Vision (ICCV), 2011.
[50] Andrej Karpathy, George Toderici, Sanketh Shetty, Thomas
Leung, Rahul Sukthankar, and Li Fei-Fei. Large-scale video
classification with convolutional neural networks. In CVPR,
2014.
[51] Sergey Ioffe and Christian Szegedy. Batch normalization:
Accelerating deep network training by reducing internal covariate shift. arXiv preprint arXiv:1502.03167, 2015.
[52] Bing Xu, Naiyan Wang, Tianqi Chen, and Mu Li. Empirical
evaluation of rectified activations in convolutional network.
arXiv preprint arXiv:1505.00853, 2015.
[53] Rong-En Fan, Kai-Wei Chang, Cho-Jui Hsieh, Xiang-Rui
Wang, and Chih-Jen Lin. Liblinear: A library for large
linear classification. Journal of machine learning research,
9(Aug):1871–1874, 2008.
[54] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In
European Conference on Computer Vision, pages 646–661.
Springer, 2016.
[55] Nitish Srivastava, Geoffrey E Hinton, Alex Krizhevsky, Ilya
Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way
to prevent neural networks from overfitting. Journal of machine learning research, 15(1):1929–1958, 2014.
[56] Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks.
[57] Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant mapping. In Computer
vision and pattern recognition, 2006 IEEE computer society
conference on, volume 2, pages 1735–1742. IEEE, 2006.
| 1cs.CV |
Australia's long-term electricity demand forecasting
using deep neural networks
Homayoun Hamedmoghadam2*, Nima Joorabloo1* and Mahdi Jalili1#,
1 School of Engineering, RMIT University, Melbourne, Australia
2 Faculty of Engineering, Monash University, Melbourne, Australia
* These two authors contributed equally to this work.
# corresponding author: Mahdi.jalili@rmit.edu.au
Abstract
Accurate prediction of long-term electricity demand has a significant role in demand side
management and electricity network planning and operation. Demand over-estimation results in
over-investment in network assets, driving up the electricity prices, while demand underestimation may lead to under-investment resulting in unreliable and insecure electricity. In this
manuscript, we apply deep neural networks to predict Australia’s long-term electricity demand.
A stacked autoencoder is used in combination with multilayer perceptrons or cascade-forward
multilayer perceptrons to predict the nation-wide electricity consumption rates for 1-24 months
ahead of time. The experimental results show that the deep structures have better performance
than classical neural networks, especially for 12-month to 24-month prediction horizon.
Keywords: Electricity demand forecasting, prediction, optimization, deep neural networks,
deep learning.
1. Introduction
Power grids are one of the most critical infrastructures and have a major role in sustainable
development and economic growth. Smart grids are the future technologies in power grid
development, management, and control [1-3]. They have revolutionized the regime of existing
power grids, by employing advanced monitoring, communication and control technologies to
provide secure and reliable energy supply. Traditional operation of distribution systems in power
grids is a top-down oriented and unidirectional, where the operations are set by distribution
companies without any kind of engagement from the customers’ side. New technologies have
changed energy consumption, making it necessary to use effective energy management
strategies. For example, in residential areas with many rooftop photovoltaic cells installed,
during daytime there might be a reverse power flow from the original power system design [4].
Such a reverse flow can create problems such as voltage and frequency distortions if not properly
managed. Another example is electric vehicles with vehicle-to-grid capability that can benefit the
demand management by providing certain services for load shifting and powerline congestion
management [5]. Integration of information and communication infrastructures with power grids
enables considering the load as an additional degree of freedom in network management and
makes the grid “smart” by properly engaging the customers in the decision making process [6].
In order to have an effective demand side management strategy and also efficient network
planning and operation, reliable short- and long-term load forecasting is essential [7].
Distribution-level short-term load forecasting systems predict the load of substations, feeders and
individual customers [8]. Short-term load forecasting is often performed for half an hour to one
week ahead time intervals. Individual loads often have abrupt temporal variations due to changes
driven by building stock variation and customers’ behavioral changes [9-11]. This makes the
short-term load forecasting at an individual level a challenging task that many conventional
prediction models fail to provide accurate predictions. Often, peaks are overestimated to
accommodate generation margins, resulting in non-efficient use of resources [9]. Load patterns at
an individual level include high levels of uncertainty and prediction models should properly
consider such volatility and uncertainty. A solution to cancel out (or minimize) the uncertainty is
to perform the prediction task at an aggregated level, as aggregated loads have more regular
patterns than individual loads [8, 12-14].
Forecast terms from one month to one year are considered mid-term and those for the future
periods of more than one year are long-term. Both mid- and long-term load forecasts have
significant importance for network strategic planning and development of the grid such as
scheduling of maintenance activities, installation of new generation and transmission capacities,
and long-term demand side management [15, 16]. There are a number of socioeconomic factors
influencing the long-term electricity load of the network [17, 18]. For example, while
industrialization and rural electrification have a major role in long-term demand increase in
developing nations, increased use of electric appliances is one of the major factors behind
increased demand in developed nations [16]. Lee and Hong used fuzzy logic and introduced a
hybrid mid-term demand forecasting that could successfully predict the electricity demand for
periods of up to four-months ahead [19]. Smart grids have many probabilistic factors making the
demand forecasting a more challenging problem than before. Liu et al. proposed a probabilistic
model for long-term demand forecast and showed it such a framework can provide a reliable
prediction [20].
Network planning is a forward-looking process that needs short- and long-term load forecast at
the same time. A recent report carried out by the Australian Chief Scientists, known as Finkel
Report [21], pinpoints inaccurate demand forecasting as one of the main reasons behind some of
Australian electricity grid problems. Demand over-estimation may lead to wasteful and
inefficient over-investment in the assets, thus driving up the costs, which is passed onto
customers making the energy prices less affordable. On the other hand, demand under-estimation
may contribute to inadequate investment that can possibly result in energy reliability and security
issues.
This manuscript studies the potential of deep neural networks to predict Australia’s long-term
electricity demand. Deep learning techniques have been successfully applied to a number of
modelling and classification tasks [22-25]. Recently, deep learning has also been applied for
short-term demand forecasting. Dedinec et al. used deep belief networks that are made up from
multiple layers of Boltzmann machines for short-term demand forecasting [26]. They showed
that such a framework can obtain better results than traditional multilayer neural networks. Kong
et al. proposed long short-term memory recurrent neural networks for residential short-term
demand forecasting [27]. Ryu et al. used deep neural networks for load forecasting and demand
side management and showed their better performance than shallow neural networks [28]. Shi et
al. used pooling-based deep recurrent neural networks for short-term load forecasting of
individual households [29]. Their proposed framework outperformed traditional recurrent neural
networks. This manuscript is the first attempt to apply deep neural networks for long-term
demand forecast. Our results show that deep neural networks outperform shallow networks for
long-term forecasting.
2. Methods
2.1. Datasets
In this study, our problem is to predict the monthly amounts of country-wide electricity
consumption in Australia. We also account for socio-economic and environmental factors. We
use publicly available monthly record of electricity demand reported by the Australian Energy
Market Operator (AEMO), which is the regulatory body responsible for operating Australia’s
largest electricity and gas markets and power systems. AEMO’s operations include National
Electricity Market, the interconnected power grid for Australia’s eastern and south-eastern areas,
and Wholesale Electricity Market that operates in state of Western Australia.
The socio-economical inputs to our model are Gross Domestic Product (GDP) and population.
We use yearly Australia’s GDP and population provided as time series data in the World Bank’s
data bank. Environmental inputs of the model are monthly precipitation, average temperature,
average minimum temperature, and average maximum temperature. Precipitation time series is
the monthly average value in depth over space within Australia, provided by the World Bank.
Land surface temperature data used in the study is reported as anomalies relative to the average
of monthly values from January 1951 to December 1980 adopted from Berkeley Earth data
collection. We also use the amount of Carbon dioxide emissions during a year in the country as
an independent variable to model electricity demand. The data used here covers the time frame
from January 1999 to August 2013.
2.2. Modeling framework
In order to model the electricity demand as a dependent variable based on a number of
independent variables, we utilize deep and shallow artificial neural networks with different
architectures and settings. The primary input variables considered in this study are socioeconomic and environmental factors, assumed to be relevant and influential to the amount of
electricity consumed by residential and industrial. Independent variables, namely GDP,
population, CO2 emissions, precipitation, average temperature, average high temperature, and
average low temperature are fed to the model as inputs and the model is trained to output the
estimated monthly electricity demand. Each sample in the data is associated with a particular
month within the interval starting from January 1999 until August 2013 (176 months) which
makes the sample size equal to 176. The time series of the dataset can be shown as (Xi, yi), i = 1,
2, …, 176, where Xi is a vector of inputs and yi is the corresponding output.
A proportion of the available data is used for supervised learning, which is called the training
time window here. For each sample within the training time window, i.e. a set of the inputs and
the output associated with a particular month, the model is used to estimate the output with the
given input values. The error of estimated target value is used to adjust the weights of the links
connecting neurons of different layers. The neural network is expected to learn the complex
relation between the independent variables (inputs) and the dependent variable (output) through
the supervised learning procedure by adjusting its synaptic weights.
Through the experiments we perform a supervised learning process using samples within a
certain time window of length l, always starting with the first sample (first month of the data);
i.e. (Xi, yi), i = 1, 2, …, l. In order to test the trained model based on the training dataset, we feed
the neural network with the input vector of samples associated with 1 to 24 months ahead of the
last month within the training time window, i.e. Xi with i = l + 1, l + 2, …, l + 24. Then, the
estimated values by the neural network, i.e. ̂ , i = l+1, l+2, …, l+24, are compared against the
corresponding target values, i.e. yi, i = l + 1, l + 2, …, l + 24, and the estimation error is
calculated as a measure of performance (prediction accuracy) for the model.
2.3. Evaluation Criteria
In the following experiments, one or more of three widely-used accuracy measures are used,
namely, Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute
Percentage Error (MAPE) defined as below:
∑
|
√ ∑
∑
̂|
(
(1)
̂)
̂
(2)
(3)
where yi is the actual value of the output, ̂ is the forecasted (or estimated) value of the output,
and N is the number of test samples. Smaller error values correspond to better prediction
accuracy.
2.4. Neural network architectures
2.4.1. Conventional artificial neural networks
Artificial Neural Networks (ANNs) are powerful computational tools for modeling and
estimation that have been applied to many applications [30]. As conventional neural networks,
we consider two models: Multi-Layer Perceptron (MLP) and Cascade-Forward Multi-Layer
Perceptron (CFMLP). These two models are two well-known neural network architectures,
widely used for non-linear modeling. Both MLP and CFMLP consist of an input (output) layer
with number of neurons equal to the number of independent (dependent) variables. They can
incorporate any number of hidden layers each with one or more neurons, between their input and
output layers. MLP and CFMLP are feed-forward architectures meaning that information
propagates in only one direction from input layer towards output layer. In other words, there are
no cycles in their networked structures. In MLP neurons in each layer are connected only to
neurons of the immediate successive layer (see Fig. 1.a), whereas in CFMLPs neurons in each
layer are connected to neurons of all the layers to the front (Fig. 1.b), i.e. towards the output
layer.
ANNs have often many parameters that should be optimized based on the dataset in hand. The
free parameters of the model are learnt such as a loss function (e.g. prediction error) is
minimized. In order to enhance generalizability of the model, often regularization terms are also
considered in the optimization process. A number of optimization approaches, with different
computational and memory requirements, have been developed for this purpose. Here, we apply
a number of optimization (or training) algorithms and choose the best performing one. Using 13
training time window sizes, where smallest window can be described with l = 120 and the largest
window with l = 132, different training algorithms are applied to adjust the networks’ weights
and biases. Trained models are then used to predict the electricity demand for 24 months ahead
of their corresponding training time window, and the average prediction accuracy is used to
compare the effectiveness of different training algorithms.
The training algorithms we use are [30]: (i) Simple loss function of squared errors and
Levenberg-Marquardt optimization algorithm denoted as LM, (ii) LM with Bayesian
regularization term aiming at better generalization of the model denoted as LMbr, (iii) simple
loss function and Gradient Descent optimization algorithm denoted as GD, (iv) GD using
momentum (GDm) which allows the algorithm to escape the shallow local minima, (v) GD with
adaptive learning rate (GDa) which balances the stability of the solution and fast convergence of
the algorithm by changing the learning rate during the training process, (vi) GD with momentum
and adaptive learning rate denoted as GDma, (vii) Conjugate Gradient algorithm which
periodically switches the search direction with Powell-Beale reset rule denoted as CGpb, (viii)
conjugate gradient algorithm using Fletcher-Reeves update rule to determine the new search
direction denoted as CGfr, (ix) conjugate gradient with Polak-Ribiére updates denoted as CGpr,
(x) scaled conjugate gradient denoted as SCG, (xi) BFGS quasi-Newton method which is an
alternative to conjugate gradient methods with more computational complexity but converges
faster and here denoted as BFGS, (xii) one-step secant optimization method which is improved
BFGS algorithm to have less computational and storage requirements here denoted as OSS, and
(xiii) resilient backpropagation algorithm (RBP) aiming at solving the problem of small
magnitude gradient which leads to slow convergence.
2.4.1. Deep neural networks
Deep Neural Networks (DNNs) are a particular type of ANNs with multiple layers which apply
non-linear operations to data in order to automatically represent the complicated functions via
high-level abstraction. One approach to create deep architectures is to use Autoencoders, also
called Autoassociators, and wire them with conventional classifiers or regression models.
Autoencoders are a class of neural networks with the ability to encode the input data into a new
representation with unsupervised learning [25, 31]. An autoencoder is a hidden layer of neurons
trained to encode the raw input data into a new representation and decode them back to
reconstruct the original input with the minimum deformation possible (Fig. 2). A stacked
autoencoder is trained layer-wise in an unsupervised manner; each time a layer is trained to
reconstruct its inputs through unsupervised learning using samples of raw inputs or the
transformed raw inputs by previous layer(s). Then the front building block, which is a classical
neural network here, is trained with the features extracted by the last autoencoder as its inputs
and the monthly electricity consumption yi as its target output; this procedure is called pretraining. Pre-training can be followed by adjusting weights in the deep network utilizing
backpropagation. This second training phase is called fine-tuning for the deep neural network.
Here we use MLP and CFMLP and stack them with autoencoders to create DNN architectures
(Fig. 3), and evaluate them as models for electricity demand prediction. We use different shallow
architectures, namely, one-layer MLP, two-layer MLP, one-layer CFMLP, and two-layer
CFMLP each with the best performer settings found (i.e. the number of neurons in the layers and
the best training algorithm) and stack them with one or two layers of autoencoders. We vary the
number of neurons in each layer of autoencoders between 1 and 15. For each one of the four
shallow architectures and with one or two layers of stacked autoencoders, the experiments are
performed using training time windows with l = 120 to l = 132 and evaluation is done according
to average prediction accuracy calculated over the whole set of tests, i.e. 24 subsequent months
of each training time window in 10 independent runs. Also, experiments are repeated once with
only a pre-training phase for blocks of deep network, and once following with a fine-tuning
phase for the whole deep neural network after pre-training of the blocks using the most efficient
backpropagation algorithm found for the classical neural network stacked in the front of the deep
structure.
Fig. 1. Schematic view of two classical neural network architectures both with a single hidden layer. a) Multi-Layer Perceptron (MLP) and b)
Cascade-forward multi-layer perceptron (CFMLP) are two widely used feed forward neural network architectures.
Fig. 2. Schematic view of a single autoencoder layer with unsupervised learning process. The autoencoder is fed with a vector of k inputs and the
network weights are adjusted to encode the inputs to create a new representation with a vector of k' values, extracting k' features. The autoencoder
then decodes the new representation to estimate the inputs with minimum error in the output layer.
Fig. 3. Schematic view of a deep neural network with multiple layers of autoencoders stacked with a classical neural network, e.g. MLP.
Autoencoders extract features from the raw inputs and feed the regression model with the new representation of data.
3. Results
3.1. Conventional artificial neural networks
The first stage in learning an ANN is to find the optimal number of hidden layers and the number
of neurons in each layer. In this work, we use only 1 or 2 hidden layers for MLP and CFMLP. In
order to tune the model, we perform the learning on MLP and CFMLP architectures with 1 and 2
hidden layers and with different number of neurons in each layer. We then evaluate the model’s
performance through a set of tests to discover the best performing setting for each architecture.
Using the training time window with l = 120 (i.e. the last sample as (X120, y120)), first, the
learning process is performed on the models, and then the model with the optimized set of
parameters is used to predict the monthly electricity demand for the next consecutive 24 months
of the data. For each set of experiments, the training time window is first expanded for a number
of times, each time with one month, and then the prediction is performed again for 1 to 24
months ahead of the end of the training time window. Furthermore, for each time window, 10
independent runs of training-then-prediction are performed to make the results more reliable.
The prediction accuracy measures are calculated by averaging over all test errors calculated
during a set of experiments. A set of experiments includes a range of window sizes with 10
independent runs for each window size.
In order to find the best structure for the two classical ANN architectures, the learning time
window is expanded 13 times, so the first and last training time windows have l = 120 and l =
132, respectively. The learning time window expansion is performed each time by adding a
single sample corresponding to the first month. First, a set of experiments are performed using
MLP and CFMLP architectures with a single hidden layer. Prediction performance is compared
for different number of neurons in the hidden layer to find the best setting for the case of having
a single hidden layer (Fig. 4). The results show that when there is only one hidden layer, the
prediction performance worsens or no significant improvement is obtained by increasing the
number of neurons more than 5. Furthermore, MLP outperforms CFMLP by showing much less
prediction error.
Fig. 4. Average prediction error of MLP and CFMLP as classical artificial neural network architectures. The plots show the performance of MLP
and CFMLP both having a single hidden layer with different number of neurons. The performance criterion is MAPE calculated for predicting the
electricity demand of 1 to 24 months ahead of the learning time window. Neural networks are trained using Levenberg-Marquardt optimization
algorithm.
We also performed the experiments on the same learning time window sizes to assess the
classical architectures having two hidden layers with the number of neurons varying between 1
and 15 within each hidden layer. According to the results, MLP with more than two neurons in
the first hidden layer have relatively good prediction accuracy, and further increase in the
number of neurons in the first layer does not affect the performance significantly. However,
increasing the number of neurons in the first hidden layer of CFMLP reduces the prediction
accuracy. For both MLP and CFMLP architectures, performance usually declines as the number
of neurons in the second layer increases more than 2 for MLP and 9 CFMLP. According to the
results, the best performance for MLP and CFMLP with single hidden layer is achieved with 4
and 5 neurons respectively. When two hidden layers are used, the best architecture is (5, 2), i.e. 5
neurons in the first layer and 2 neurons in the second layer, for MLP and (2, 9) for CFMLP.
Utilizing different training approaches led to a wide range of prediction performances, according
to our experiments (Fig. 5). The results show that effectiveness of different training algorithms is
not the same for MLP and CFMLP. The MLP architecture with one hidden layer of 4 neurons
achieves the best prediction accuracy when trained with Levenberg-Marquardt algorithm
equipped with Bayesian regularization, while resilient backpropagation algorithm is the best
learning algorithm for MLP with two hidden layers. The best learning algorithm for CFMLP
with one (two) hidden layers is Levenberg-Marquardt algorithm with simple loss function
(conjugate gradient with Polak-Ribiére updates). These optimal learning algorithms are used for
DNN architectures.
Fig. 5. Average prediction error for (a) MLP and (b) CFMLP, achieved by applying different training algorithms. For each architecture, the bestperforming structure found through the previous set of experiments is used, i.e. single hidden layer with 4 and 5 neurons for MLP and CFMLP,
respectively and (5, 2) neurons and (2, 9) neurons for two-layer MLP and CFMLP, respectively.
3.2. Deep neural networks
In this section, we apply the proposed stacked autoencoder with MLP or CFMLP to the data. For
MLP with one layer, the best performance is achieved with one layer of stacked autoencoder
with 2 neurons. When there are two hidden layers in MLP, the best found structure consists of
one layer of stacked autoencoder with 8 neurons. DNNs using the CFMLP with one and two
layers have the best prediction accuracy with a single stacked autoencoder layer with 10 and 4
neurons, respectively. All architectures result in higher prediction accuracy when they go
through the fine-tuning phase.
We compare the four classical architectures (MLP and CFMLP with one and two hidden layers)
with their corresponding deep architectures. Note that the deep networks are built by adding a
block of one-layer stacked autoencoder to the classical models with the best-performing settings
(i.e. the number of neurons in the hidden layer and the optimal learning algorithm) determined
by experimental results. The learning in stacked autoencoder is unsupervised, while supervised
learning is used to tune parameters of MLP and CFMLP. In order to compare the classical model
with deep networks, we evaluate their prediction error when the demand is forecasted 1-24
months ahead.
Figure 6 shows the prediction error as a function of the length of look-ahead time. It is expected
that as prediction horizon is extended, the performance worsens, i.e. the prediction error
increases. The rate of decline in performance by increasing the prediction horizon is significantly
less in deep networks as compared to their classical counterparts. The results suggest that deep
neural networks are always more effective for long-term prediction, i.e. more than 1 year to 2
years ahead, rather than the prediction for up to 1 year ahead (mid-term prediction). Deep
networks with one layer of MLP or CFMLP have always better prediction performance than
classical MLP or CFMLP, where the outperformance of deep CFMLP over its classical
counterpart is more significant than that of MLP. Deep MLP with two hidden layers and deep
CFMLP with one layer show the best overall performance, while the deep MLP with two layers
is almost robust against increasing the prediction horizon. Overall prediction error is also
calculated for different structures by averaging over the prediction horizons (i.e. 24 values each
to corresponding to different prediction horizon), and the results are shown in Table 1. It
indicates that deep CFMLP with one hidden layer has the best performance in terms of all
evaluation criteria (MAPE, RMSE and MAE). These results also show that stacking a layer of
autoencoder almost always enhances the overall performance of a classical neural network.
Fig. 6. Classical neural networks (MLP and CFMLP) and their corresponding deep architectures. Deep network structures are built by stacking an
autoencoder to extract features from the raw inputs and feed the classical neural network with the new representation. The plots show the average
prediction error for 1 to 24 months ahead. The classical networks are a) MLP with one hidden layer, b) MLP with two hidden layers, c) CFMLP
with one hidden layer, and d) CFMLP with two hidden layers.
Table 1. Comparison between classical and deep neural networks in terms of three different evaluation criteria (MAPE, RMSE and MAE)
calculated by averaging over all experiments for training time windows with l = 133, 134, …, 152 and 10 independent runs for each window size.
24 tests (one for each prediction horizon 1-24 months) are performed for each run of each training window size; an average over these 24 tests is
obtained for each run. Error value showing the best performance is shown in bold.
MLP with one
hidden layer
Classical
Deep
MLP with two
hidden layers
Classical Deep
CFMLP with one
hidden layer
Classical Deep
CFMLP two
hidden layers
Classical Deep
MAPE
22.83
19.88
24.13
21.59
27.68
18.68
27.90
24.66
RMSE
MAE
0.166
0.145
0.177
0.192
0.213
0.138
0.209
0.182
0.142
0.123
0.149
0.149
0.170
0.115
0.172
0.150
4. Discussion and Conclusion
In this study, we aimed to predict the monthly electricity demand in Australia based on time
series of consumption rates as well as socio-economic and environmental factors. We
approached the modeling of electricity demand by employing artificial neural networks and
predicted the nation-wide electricity consumption for a range from one month (mid-term
forecast) to two years (long-term forecast) ahead. Such a long-term demand forecast has
significant role in effective electricity network planning and operation. Inaccurate demand
forecast has been identified as one of the major reasons behind recent problems with Australian
energy systems. This has contributed to reliability and security issues in the energy system, and
thus driving up the energy prices.
We collected the electricity consumption, socioeconomic and environmental data from various
resources and preprocessed them to make them ready for further processing and neural network
modeling. We then carried out a set of tests to find the best performing settings for two widelyused classical neural network architectures, namely, multi-layer perceptron and cascade-forward
multi-layer perceptron. Then, both tuned classical architectures with their best performing
structures and training approaches, were used with stacked autoencoders as building blocks to
create deep neural architectures. Such a deep neural network structure has been successfully used
in a number of modeling and classification applications.
We investigated the effect of the number of stacked autoencoder layers with the tuned classical
network architectures and found the optimal design settings for deep neural networks. As a final
step, the classical neural networks were compared with the corresponding deep architectures
built upon them. The comparison was carried out based on the prediction error of the models to
predict the electricity demand in a range from 1 to 24 months ahead of the training time window.
Experiments revealed the effectiveness of deep neural networks architectures for monthly
demand prediction. Stacked autoencoders, when trained in an unsupervised manner to reproduce
inputs through an encoding-decoding process, learns to produce an alternative representation of
input data. A single layer autoencoder is proved to have the tendency to learn first-order features
from the raw input data. Stacked autoencoder feeds the classical neural network block with the
new representation of inputs. Our investigations in this manuscript showed how using stacked
autoencoders can improve the performance of classical neural network models by capturing
latent useful features from the raw inputs. As a future work, similar deep neural network
structures will be applied to predict short-term electricity demand in varying granularity levels,
from individual household to substations, state and nation-wide levels.
Acknowledgments
Mahdi Jalili is supported by Australian Research Council through project No DP170102303. The
authors thank Dr. Peter Sokolowski (RMIT University, Melbourne, Australia) and Mr. Peter
McTaggart (Powercor, Melbourne, Australia) for stimulating discussions on the topic.
References
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
H. H.-C. Iu, C.-C. Chu, C. O. Nwankpa, and C. W. Wu, "Guest Editorial Complex
Network for Modern Smart Grid Applications (Part 1: Analysis and Application)," IEEE
Journal on Emerging and Selected Topics in Circuits and Systems, no. 7, p. 2, 2017.
X. Yu and Y. Xue, "Smart grids: a cyber-physical systems perspective," Proceeding of
the IEEE, vol. 104, no. 5, pp. 1058-1070, 2016.
X. Yu, C. Cecati, T. Dillon, and M. Simoes, "The new frontier of smart grids: An
industrial electronics perspective," IEEE Industrial Electronics Magazine, vol. 5, no. 3,
pp. 49-63, 2011.
E. Yao, P. Samadi, V. W. S. Wong, and R. Schober, "Residential demand side
management under high penetration of rooftop photovoltaic units," IEEE Transactions on
Smart Grids, vol. 7, no. 3, pp. 1597-1608, 2016.
M. A, López, S. de la Torre, S. Martín, and J. A. Aguado, "Demand-side management in
smart grid operation considering electric vehicles load shifting and vehicle-to-grid
support," International Journal of Electrical Power & Energy Systems, vol. 64, pp. 689698, 2015.
P. Palensky and D. Dietrich, "Demand side management: demand response, intelligent
energy systems, and smart loads," IEEE Transactions on Industrial Informatics, vol. 7,
no. 3, pp. 381-388, 2011.
D. Alahakoon and X. Yu, "Smart electricity meter data intelligence for future energy
systems: A survey," IEEE Transactions on Industrial Informatics, vol. 12, no. 1, pp. 425436, 2016.
X. Sun et al., "An efficient approach to short-term load forecasting at the distribution
level," IEEE Transactions on Power Systems, vol. 31, no. 4, pp. 2526-2537, 2016.
B. Stephen, X. Tang, P. R. Harvey, S. Galloway, and K. I. Jennett, "Incorporating
practice theory in sub-profile models for short term aggregated residential load
forecasting," IEEE Transactions on Smart Grids, vol. 8, no. 4, pp. 1591-1598, 2017.
J. Stephenson, B. Barton, G. Carrington, D. Gnoth, R. Lawson, and P. Thorsnes, "Energy
cultures: A framework for understanding energy behaviours," Energy Policy, vol. 38, no.
10, pp. 6120-6129, 2010.
H. Jiang, Y. Zhang, E. Muljadi, J. J. Zhang, and D. W. Gao, "A short-term and highresolution distribution system load forecasting approach using support vector regression
with hybrid parameters optimization," IEEE Transactions on Smart Grids, 2016.
A. Espasa and I. Mayo-Burgos, "Forecasting aggregates and disaggregates with common
features," International Journal of Forecasting, vol. 29, pp. 718-732, 2013.
M. Espinoza, J. Suykens, R. Belmans, and B. D. Moor, "Electric load forecasting," IEEE
Control Systems Magazine, vol. 27, no. 5, pp. 43-57, 2007.
S. Fan and Y. Wu, "Multiregion load forecasting for system with large geographical
area," IEEE Transactions on Industry Applications, vol. 45, no. 4, pp. 1452-1459, 2009.
A. H. Sanstad, S. McMenamin, A. Sukenik, G. L. Barbose, and C. A. Goldman,
"Modeling an aggressive energy-efficiency scenario in long-range load forecasting for
electric power transmission planning," Applied Energy, vol. 128, pp. 265-276, 2014.
F. J. Ardakani and M. M. Ardehali, "Long-term electrical energy consumption
forecasting for developing and developed economies based on different optimized models
and historical data types," Energy, vol. 65, pp. 452-461, 2014.
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
H. Hamedmoghadam, M. Jalili, H. Davari, and R. Maknoon, "Prediction of Iran's annual
electricity demand: artificial intelligence approaches," presented at the Innovations in
Information Technology, 2015.
J. Moral-Carcedo and J. Perez-García, "Integrating long-term economic scenarios into
peak load forecasting: An application to Spain," Energy 140, vol. 140, pp. 682-695, 2017.
W.-J. Lee and J. Hong, "A hybrid dynamic and fuzzy time series model for mid-term
power load forecasting," Electrical Power and Energy Systems, vol. 64, pp. 1057-1062,
2015.
B. Liu, J. Nowotarski, T. Hong, and R. Weron, "Probabilistic load forecasting via
quantile regression averaging on sister forecasts," IEEE Transactions on Smart Grids,
vol. 8, no. 2, pp. 730-737, 2017.
A. Finkel, K. Moses, C. Munro, T. Effeney, and M. O’Kane, "Independent review into
the future security of the national electricity market: blueprint for the future," 2017.
T. Young, D. Hazarika, S. Poria, and E. Cambria, "Recent trends in deep learning based
natural language processing," arXiv:1708.02709, 2017.
J. Schmidhuber, "Deep learning in neural networks: An overview," Neural Networks, vol.
61, pp. 85-117, 2015.
Y. LeCun, Y. Bengio, and G. Hinton, "Deep learning," Nature, vol. 521, pp. 436-444,
2015.
P. Vincent, H. Larochelle, I. Lajoie, Y. Bengio, and P.-A. Manzagol, "Stacked denoising
autoencoders: learning useful representations in a deep network with a local denoising
criterion," Journal of Machine Learning Research, vol. 11, p. 3371−3408, 2010.
A. Dedinec, S. Filiposka, A. Dedinec, and L. Kocarev, "Deep belief network based
electricity load forecasting: An analysis of Macedonian case," Energy, vol. 115, pp.
1688-1700, 2016.
W. Kong, Z. Y. Dong, Y. Jia, D. J. Hill, Y. Xu, and Y. Zhang, "Short-term residential
load forecasting based on LSTM recurrent neural network," IEEE Transactions on Smart
Grids, 2017.
S. Ryu, J. Noh, and H. Kim, "Deep neural network based demand side short term load
forecasting," Energies, vol. 10, p. 3, 2017.
H. Shi, M. Xu, and R. Li, "Deep learning for household load forecasting – A novel
pooling deep RNN," IEEE Transactions on Smart Grids, 2017.
S. Haykin, Neural networks: a comprehensive foundation, second ed. Prentice Hall,
1999.
Y. Bengio, "Learning deep architectures for AI," Foundations and trends® in Machine
Learning, vol. 2, no. 1, pp. 1-27, 2009.
| 9cs.NE |
A SPLINE-ASSISTED SEMIPARAMETRIC APPROACH TO
NONPARAMETRIC MEASUREMENT ERROR MODELS
arXiv:1804.00793v1 [math.ST] 3 Apr 2018
By Fei Jiang∗ and Yanyuan Ma†
University of Hong Kong∗ and Penn State University†
Nonparametric estimation of the probability density function of
a random variable measured with error is considered to be a difficult
problem, in the sense that depending on the measurement error property, the estimation rate can be as slow as the logarithm of the sample
size. Likewise, nonparametric estimation of the regression function
with errors in the covariate suffers the same possibly slow rate. The
traditional methods for both problems are based on deconvolution,
where the slow convergence rate is caused by the quick convergence
to zero of the Fourier transform of the measurement error density,
which, unfortunately, appears in the denominators during the construction of these methods. Using a completely different approach
of spline-assisted semiparametric methods, we are able to construct
nonparametric estimators of both density functions and regression
mean functions that achieve the same nonparametric convergence
rate as in the error free case. Other than requiring the error-prone
variable distribution to be compactly supported, our assumptions
are not stronger than in the classical deconvolution literatures. The
performance of these methods are demonstrated through some simulations and a data example.
1. Introduction. Density estimation is a familiar problem in the nonparametric estimation literature. Generally, we observe independent and
identically distributed (iid) variables X1 , . . . , Xn from a distribution with
probability density function (pdf) fX (x) and nonparametric estimators such
as kernel methods are available in the literature to estimate fX (x). Even
when the Xi ’s are not directly observed, nonparametric estimation of fX (x)
can still be carried out based on their surrogates. Specifically, assume that
instead of observing Xi , we observe Wi ≡ Xi + Ui , where Ui is a mean
zero random error independent of Xi and follows a distribution with pdf
fU (u). This problem was studied extensively in the literature and by large
the main stream approach is deconvolution (Carroll and Hall, 1988; Liu and
Taylor, 1989; Stefanski and Carroll, 1990; Zhang, 1990; Fan, 1991). However,
as established in these works, the deconvolution based estimator of fX (x)
may converge very slowly. For example, when the error distribution fU (u)
MSC 2010 subject classifications: Primary 60K35, 60K35; secondary 60K35
Keywords and phrases: errors in covariates, measurement error, semiparametrics, spline
1
2
is normal or other distributions that are “super smooth”, the deconvolution
estimator only converges to fX (x) at the rate of {log(n)}−k where k is a
positive constant. When the error distribution fU (u) is Laplace or other
“ordinary smooth” type, the convergence rate is n−k , where k is a positive
constant smaller than 0.2. Here “super smooth” and “ordinary smooth” are
characteristics that play an important role in determining the convergence
properties of the deconvolution estimators. Because these characteristics are
irrelevant in our approach, we skip the precise description and refer to Fan
(1991) for details.
Given that when Xi ’s are observed, nonparametric density estimation
usually performs better than these results, it prompts us to ask whether we
could further improve the estimation performance for the error-in-variable
case. However, the convergence rates described above are shown to be “optimal” in the literature as long as deconvolution is used. Thus, if better
convergence is to be achieved, the estimation has to go beyond the deconvolution framework. In other words, novel approaches totally different from
the existing literature will be required.
A related problem to nonparametric density estimation is nonparametric
regression. Likewise, when iid observations (Xi , Yi ), i = 1, . . . , n, are available, many nonparametric estimators such as kernel and spline based methods have been proposed to estimate the regression mean of Y conditional
on X. Here the assumption is that Yi = m(Xi ) + i , where for simplicity,
we assume i is independent of Xi and has mean zero with density f ().
When Xi is unavailable and instead only Wi described above is available,
we encounter the problem of nonparametric regression with measurement
error. The traditional way to estimate m(x) is also deconvolution (Fan and
Truong, 1993), resulting in the same possibly slow rate of convergence. Because the same reason that results in the slow convergence rate in the density
estimation is also responsible for the nonparametric regression estimation,
methods that do not rely on deconvolution have to be employed in order to
achieve improvement.
Our breakthrough came when it struck us that if the functions that we
want to estimate, fX (x) and m(x), had been parametric, we would have
much simpler problems. If these problems can be solved without using deconvolution, then we can approximate fX (x) and m(x) with B-spline, and
operationally we would be dealing with parametric models and deconvolution can be avoided too. Thus, we study both problems in this work with the
help of a spline representation. Using spline representation in measurement
error models is not entirely new, although it was mostly used in the Bayesian
framework (Berry, Carroll and Ruppert, 2002; Staudenmayer, Ruppert and
NEW NONPARAMETRIC MEASUREMENT ERROR MODEL ESTIMATOR
3
Buonaccorsi, 2008). While the density estimator is relatively easy to devise,
regression estimator turns out to be challenging and requires further semiparametric treatment. A second challenge in both problems is in establishing
the convergence rates of the resulting estimators. The common obstacle in
both estimators is the fact that it is a latent function that needs to be approximated with a spline representation, which requires unusual treatment
different from the typical handling of the spline approximation. In addition,
for the regression problem, we encounter further difficulties because not only
the estimator, but also the estimating equation that generates the estimator, do not have closed forms. Much effort is spent in clearing these hurdles, where we resort to extensive use of bilinear operators (Conway, 1990),
which are very different from typical regression spline asymptotic analysis
(Ma et al., 2015). The detailed proofs are respectively in the Appendix at
the end of the paper and in an online supplementary document.
The success of the B-spline-assisted procedures in achieving the better
rates of convergence in these problems is encouraging. Despite of the satisfactory result, it is not immediately clear what is the intuitive reason behind
the success. To rationalize the convergence rate of the B-spline based procedures, we can consider the parametric pdf and regression problems with
measurement error. In these simpler problems, root-n convergence of the
parameter estimation is indeed achievable. Thus, bridging a nonparametric problem to a parametric problem via B-splines, it is not unreasonable
to expect a classical nonparametric convergence rate for measurement error
problems as well. A critical assumption in such B-spline treatment is that the
possible range of the unobserved covariates are bounded, hence we can confine to a compact set in estimating the functions of interest. This restricted
the impact from the tails of the measurement error distribution, which is exactly what brings down the convergence rate of the deconvolution methods.
More explicitly, in the B-spline approximation, the compact set on which
we perform the estimation is built into the procedure at the very beginning
and we benefit from that throughout the procedure. However, in the deconvolution procedures, the Fourier and inverse Fourier transformation steps
do not take advantage of the compact set knowledge, and it automatically
estimates these functions on the whole interval, which is much harder to do.
In the following, we devise the estimation procedures for both the probability density function and the regression mean function in Section 2, and
summarize the theoretical properties of our estimators in Section 3. We
provide simulation studies to demonstrate the rate properties of the new
estimators in Section 4, and illustrate the methods in a data example in
Section 5. The paper is concluded with a discussion in Section 6.
4
2. B-spline-assisted estimation procedures.
2.1. Probability density function estimation. To set the notation, we use
fX (x) to denote a generic pdf function of the random variable X, and use
fX0 (x) to denote the true pdf that generates the data. We approximate
fX0 (x) on its support using B-splines (Masri and Redner, 2005). For simplicity, let the support be contained in [0, 1]. To ensure that the density
function is nonnegative and integrates to 1, we let the approximation be
fX (x, θ) ≡ R 1
0
exp{Br (x)T θ}
exp{Br (x)T θ}dx
,
where Br (x) is a vector of B-spline basis functions, and θ is the B-spline
coefficient vector to be estimated. Then
R1
exp{Br (x)T θ}fU (w − x)dx
fW (w, θ) ≡ 0 R 1
T
0 exp{Br (x) θ}dx
is an approximation to the pdf of W ≡ X + U , a surrogate of X. We then
perform a simple maximum likelihood estimator (MLE) to obtain θ, i.e. we
maximize
n
X
=
i=1
n
X
i=1
logfW (Wi , θ)
log
Z
0
1
T
exp{Br (x) θ}fU (Wi − x)dx − nlog
Z
1
exp{Br (x)T θ}dx
0
b and then reconstruct fX (x, θ)
b and use it as
with respect to θ to obtain θ,
b
b
the estimator for fX0 (x), i.e. fX0 (x) = fX (x, θ).
While the estimation procedure for fX0 (x) is extremely simple, it is not as
straightforward to establish the large sample properties of the estimator. In
Section 3, we will show that fbX0 (x) converges to fX0 (x) at a nonparametric
rate under mild conditions.
2.2. Regression function estimation. Unlike in the density estimation
case, the estimation procedure in the regression model is much more complex as we now describe. The key observation is that as soon as the nonparametric function m(x) is approximated with the B-spline representation, the
regression function itself without measurement error is a purely parametric
model (Wang and Yang, 2009), hence the idea behind Tsiatis and Ma (2004)
can be adapted. Specifically, we treat the B-spline coefficients as parameters of interest, treat the unspecified distribution of X, fX (x), as nuisance
NEW NONPARAMETRIC MEASUREMENT ERROR MODEL ESTIMATOR
5
parameter, and cast the problem as a semiparametric estimation problem.
We can then construct the efficient score function, which relies on fX0 (x).
A key observation of Tsiatis and Ma (2004) is that by replacing fX0 (x) with
an arbitrary working model, the consistency of the estimation can still be
retained. We now describe the estimation procedure in detail.
∗ (x) be a working pdf of X. Note that f ∗ (x) may not be the
First, let fX
X
∗ (x) is possibly misspecified. Following the practice
same as fX0 (x), that is, fX
of density estimation in Section 2.1, assume that the true density function
fX0 (x) has compact support contained in [0, 1]. Therefore we only need to
consider m(x) on [0, 1]. We approximate m(x) using spline representation
Br (x)T β. Define
Z 1
∂
∗
log
f {Y − Br (x)T β}fU (W − x)fX
(x)dµ(x)
∂β
0
R1 0
∗ (x)B (x)dµ(x)
f {Y − Br (x)T β}fU (W − x)fX
r
.
= − 0 R1
∗
T
0 f {Y − Br (x) β}fU (W − x)fX (x)dµ(x)
S∗β (W, Y, β) =
As the notation suggests, S∗β (W, Y, β) is the score function with respect to
∗ (x)
β calculated from the joint pdf of (W, Y ) under the working model fX
∗ (x),
and the spline approximation. Due to the possible misspecification of fX
∗
the mean of Sβ (W, Y, β) is not necessarily zero even if the mean function
P
is exactly Br (x)T β. Therefore simply solving ni=1 S∗β (Wi , Yi , β) = 0 may
generate an inconsistent estimator. The idea behind our estimator is to find
a function a(x, β) so that
(1)
E{S∗β (W, Y, β) | X}
"R 1
#
T β}f (W − x)f ∗ (x)dµ(x)
a(x,
β)f
{Y
−
B
(x)
r
U
X
= E 0 R1
|X ,
T β}f (W − x)f ∗ (x)dµ(x)
f
{Y
−
B
(x)
r
U
X
0
and then solve for β using the estimating equation
(2)
n
X
∗
Sβ (Wi , Yi , β)
i=1
−
R1
0
#
∗ (x)dµ(x)
a(x, β)f {Yi − Br (x)T β}fU (Wi − x)fX
= 0.
R1
T β}f (W − x)f ∗ (x)dµ(x)
f
{Y
−
B
(x)
i
r
i
U
X
0
This will guarantee a consistent estimator of β if the mean function is indeed
Br (x)T β, because our construction ensures the left hand side of (2) has mean
zero. Note that the right hand side of (1) is the conditional expectation of
6
a(X, β) calculated under the B-spline approximation and the posited model
∗ (x), hence we alternatively write it as E ∗ {a(X, β) | Y , W , β}.
fX
i
i
To solve for
a(x,
β),
we
discretize
the
integral
equation
(1).
In particular,
PL
∗
let fX (x) = j=1 cj I(x = xj ), where xj ’s are the points selected on [0, 1],
P
and cj ≥ 0, L
j=1 cj = 1. Then
S∗β (W, Y, β)
= −
PL
0
T
j=1 Br (xj )f {Y − Br (xj ) β}fU (W − xj )cj
.
PL
T
j=1 f {Y − Br (xj ) β}fU (W − xj )cj
Next, to write out the right hand side of (1) upon discretization, let A(β)
be an L × L matrix with its (i, j) entry
Z
f {y − Br (xj )T β}fU (w − xj )cj
Aij (β) =
PL
T
j=1 f {y − Br (xj ) β}fU (w − xj )cj
×f {y − Br (xi )T β}fU (w − xi )dµ(y)dµ(w).
Let ai = a(xi , β), a = (a1 , . . . , aL ). Further, define Hj (β) = {H1j (β), . . . , HLj (β)},
where
Z
f 0 {y − Br (xj )T β}fU (w − xj )cj
Hij (β) = − PL
T
j=1 cj f {y − Br (xj ) β}fU (w − xj )
×f {y − Br (xi )T β}fU (w − xi )dµ(y)dµ(w),
and let b(β) be a p × L matrix, with its ith column
bi (β) =
L
X
Hij (β)Br (xj ).
j=1
P
Then b(β) = L
j=1 Br (xj )Hj (β).
With
these
notations,
the integral equation (1) is equivalently written
PL
PL
as
j=1 Aij aj =
j=1 Hij (β)Br (xj ) for i = 1, . . . , L, or more concisely,
PL
T
aA (β) = j=1 Br (xj )Hj (β). Therefore
a(β) =
L
X
Br (xk )Hk (β){A−1 (β)}T ,
k=1
hence
aj (β) =
L
X
k=1
Br (xk )Hk (β){A−1 (β)}T ej ,
7
NEW NONPARAMETRIC MEASUREMENT ERROR MODEL ESTIMATOR
where ej is a length L vector with the jth component 1 and all others zero.
Thus, we have obtained a(X, β) on the discrete set x1 , . . . xL and can form
E ∗ {a(X, β) | Y, W, β}
PL
T
j=1 aj (β)f {Y − Br (xj ) β}fU (W − xj )cj
=
PL
T
j=1 f {Y − Br (xj ) β}fU (W − xj )cj
P
L
−1
T
T
X
Hk (β) L
j=1 {A (β)} ej f {Y − Br (xj ) β}fU (W − xj )cj
=
Br (xk )
PL
T
j=1 f {Y − Br (xj ) β}fU (W − xj )cj
k=1
=
L
X
Br (xk )Pk (W, Y, β),
k=1
where
Pk (W, Y, β) =
Hk (β)
PL
−1
T
T
j=1 {A (β)} ej f {Y − Br (xj ) β}fU (W
PL
T
j=1 f {Y − Br (xj ) β}fU (W − xj )cj
− xj )cj
.
We then obtain the estimator for β by solving the estimating equation (2)
with the corresponding a(x, β) plugged in. Note that in all the functions
that are explicitly written to depend on β, the dependence is always through
Br (·)T β.
b converges to m(x) at a nonparametric rate and we
We show that Br (x)T β
derive its estimation variance in Section 3 under mild conditions. As pointed
∗ (x),
out by an anonymous referee, instead of adopting a working model fX
we could in fact estimate fX0 (x) using the method developed in Section
∗ (x) to carry out the
2.1 to obtain fbX (x), and then use fbX (x) instead of fX
subsequent operations. We believe that this is the optimal thing to do and
will generate the most efficient estimator. However, rigorously establishing
the theoretical property of such practice is challenging so we leave this option
to future research.
3. Asymptotic results.
3.1. Results of probability density function estimation. We assume the
following regularity conditions.
(C1) The true density function fX0 (x) has compact support contained in
[0, 1], is bounded on its support and satisfies fX0 (x) ∈ C q ([0, 1]),
q ≥ 1. fU (u) is bounded above. The conditional pdf of X given W ,
fX|W (x|w), is bounded.
(C2) The spline order r ≥ q.
8
(C3) We define the knots t−r+1 = · · · = t0 = 0 < t1 < · · · < tN < 1 =
tN +1 = · · · = tN +r , where N is the number of interior knots and [0, 1]
is divided into N + 1 subintervals. Let dθ = N + r. N satisfies N → ∞,
N −1 n(logn)−1 → ∞, and N n−1/(2q) → ∞ when n → ∞.
(C4) Let hp be the distance between the (p + 1)th and pth interior knots
and let hb = max hp , hs = min hp . There exists a constant
r≤p≤N +r
r≤p≤N +r
chb , 0 < chb < ∞, such that
hb /hs < chb .
Therefore, hb = O(N −1 ), hs = O(N −1 ).
(C5) θ 0 is a spline coefficient such that supx∈[0,1] |Br (x)T θ 0 −log{fX0 (x)}| =
Op (hqb ). The existence of such θ 0 has been shown in De Boor (1978).
(C6) The expectation
"R 1
#!
R1
T
T
0 exp{Br (x) θ}fU (Wi − x)Br (x)dx
0 exp{Br (x) θ}Br (x)dx
E
−
R1
R1
T θ}f (W − x)dx
T
exp{B
(x)
r
i
U
0
0 exp{Br (x) θ}dx
is a smooth function of θ and has unique root for θ in the neighborhood
of θ 0 .
Condition (C1) contains some basic boundedness and smoothness conditions on various densities and is quite standard. The only requirement
that appears nonstandard is that fX0 (x) has compact support. This however is practically relevant in many situations as the range of the values
a random variable can be may very well be bounded. Condition (C2) is a
standard requirement to ensure that splines of sufficiently high order are
utilized. Condition (C3) requires suitable amount of spline basis to be used
according to the sample size, and Condition (C4) makes sure that the spline
knots are distributed sufficiently evenly. In summary, Conditions (C2), (C3)
and (C4) are standard requirements and together with Condition (C1), they
ensure Condition (C5). We list Condition (C5) instead of stating it in the
proof for convenience. Finally, Condition (C6) ensures that we are not in
the degenerate case where the expression in (C6) is constantly zero. Under
these conditions, we obtain the following properties of fbX0 (x).
b maxiProposition 1. Assume Conditions (C1) – (C6) to hold. Let θ
mize
n
X
logfW (Wi , θ)
i=1
=
n
X
i=1
log
Z
0
1
T
exp{Br (x) θ}fU (Wi − x)dx − nlog
Z
0
1
exp{Br (x)T θ}dx,
NEW NONPARAMETRIC MEASUREMENT ERROR MODEL ESTIMATOR
9
b − θ 0 = op (1) element-wise.
then θ
b maxiProposition 2. Assume Conditions (C1) – (C6) to hold. Let θ
mize
n
X
logfW (Wi , θ)
i=1
=
n
X
log
Z
0
i=1
1
T
exp{Br (x) θ}fU (Wi − x)dx − nlog
Z
1
exp{Br (x)T θ}dx.
0
b − θ 0 k2 = Op {(nhb )−1/2 }. Further, define
Then kθ
(
)
n Z 1
X
f
(x)f
(W
−
x)
i
X0
U
R00 = n−1
− fX0 (x) Br (x)dx.
R1
i=1 0
0 fX0 (x)fU (Wi − x)dx
Then
=
b − θ0
θ
"R 1
fX0 (x)fU (Wi − x)Br (x)Br (x)T dx
R1
0 fX0 (x)fU (Wi − x)dx
nR
o⊗2
1
Z 1
f
(x)f
(W
−
x)B
(x)dx
i
r
X0
U
0
−
fX0 (x)Br (x)Br (x)T dx
−
nR
o2
1
0
0 fX0 (x)fU (Wi − x)dx
Z
⊗2 #!−1
0
E
1
+
fX0 (x)Br (x)dx
0
R00 {1 + op (1)}.
Propositions 1 and 2 establishes the consistency and convergence rate
b to θ 0 defined in Condition (C5). The proof of Propositions
properties of θ
1 and 2 are in Supplement S.3 and S.4 respectively. We then utilize these
properties to analyze the bias, variance and convergence rate of fbX0 (x) in
Theorem 1, with its proof in Appendix A.1.
b be defined
Theorem 1. Assume Conditions (C1) – (C6) to hold. Let θ
in Proposition 2 and
b
exp{Br (x)T θ}
fbX0 (x) = R 1
.
Tb
0 exp{Br (x) θ}dx
Then supx∈[0,1] |log{fbX0 (x)} − log{fX0 (x)}| = Op {(nhb )−1/2 }. Further
q−1/2
bias{fbX0 (x)} ≡ E{fbX0 (x) − fX0 (x)} = Op (hb
),
10
and
h
i
nhb fbX0 (x) − fX0 (x) − bias{fbX0 (x)}
"R 1
p
fX0 (x)fU (Wi − x)Br (x)Br (x)T dx
∂
exp{Br (x)T θ 0 }
=
E 0
nhb T R 1
R1
∂θ 0 exp{Br (x)T θ 0 }dx
0 fX0 (x)fU (Wi − x)dx
nR
o⊗2
1
Z 1
0 fX0 (x)fU (Wi − x)Br (x)dx
−
−
fX0 (x)Br (x)Br (x)T dx
nR
o2
1
0
0 fX0 (x)fU (Wi − x)dx
Z
⊗2 #!−1
p
1
+
fX0 (x)Br (x)dx
0
×n−1
n Z
X
i=1
0
1
(
R1
0
fX0 (x)fU (Wi − x)
fX0 (x)fU (Wi − x)dx
)
− fX0 (x) Br (x)dx + op (1).
Theorem 1 shows that the B-spline MLE density estimator is root-nhb
consistent to fX0 (x), which is a standard nonparametric density estimation rate when no measurement error occurs, and is improved substantially
compared to the nonparametric deconvolution method.
3.2. Results of regression mean function estimation. To facilitate the description of the regularity conditions and the asymptotic results, we introduce some notations. Define
P
−1
T
T
H(x, β) L
j=1 {A (β)} ej f {Y − Br (xj ) β}fU (W − xj )cj
P (x, W, Y, β) =
,
R1
∗
T
0 f {Y − Br (x) β}fU (W − x)fX (x)dµ(x)
H(x, β) = {H1 (x, β), . . . , HL (x, β)},
Z
∗ (x)
f0 {y − Br (x)T β}fU (w − x)fX
Hi (x, β) = − R 1
∗
T
0 f {y − Br (x) β}fU (w − xj )fX (x)dµ(x)
×f {y − Br (xi )T β}fU (w − xi )dµ(y)dµ(w),
and write
∗
E {a(X, β) | Y, W, β} =
S∗β (W, Y, β)
=
Z
1
Br (x)P (x, W, Y, β)dµ(x),
0
Z
0
1
∗ (x)
Br (x)f0 {Y − Br (x)T β}fU (W − x)fX
dµ(x).
R1
∗ (x)
− 0 f {Y − Br (x)T β}fU (W − x)fX
We further define S∗β (Wi , Yi , m), E ∗ {a(X, m) | Yi , Wi , m}, P (x, W, Y, m),
Pk (W, Y, m) to be the resulting quantities when we replace all the appearance of Br (·)T β in S∗β (Wi , Yi , β), E ∗ {a(X, β) | Yi , Wi , β}, P (x, W, Y, β), and
NEW NONPARAMETRIC MEASUREMENT ERROR MODEL ESTIMATOR 11
Pk (W, Y, m) by m(·) respectively. Here a(X, m) is a function that satisfies
E[S∗β (Wi , Yi , m) − E ∗ {a(X, m)|Yi , Wi , m}|X, m] = 0,
where the last m is used to emphasize that the calculation of the outside
expectation depends on m. Note that these definitions do not conflict with
the notations used before. In fact, some generalize the previous notations.
We further define Sm (Y, W, m) to be a linear operator on Lp whose value at
s ∈ Lp is
Sm (Y, W, m)(s)
#
Z 1"
∗ (x)dµ(x)
f0 {Y − m(x)}fU (W − x)fX
=
−R 1
− P (x, Y, W, m) s(x)dx.
∗ (x)dµ(x)
f
{Y
−
m(x)}f
(W
−
x)f
0
U
X
0
The derivative of Sm (Y, W, m) is a bilinear operator on Lp × Lq with 1/p +
1/q = 1, 1 ≤ p, q ≤ ∞, whose value at s ∈ Lp , v ∈ Lq is
∂Sm (Y, W, m)
∂Sm (Y, W, m + tv)(s)
(s, v) =
∂m
∂t
.
t=0
Further, we define a bilinear operator
2
Sm
(Y, W, m)(s, v) = Sm (Y, W, m)(s)Sm (Y, W, m)(v).
From the above definition, we can see that
n
Also,
Pn
−1
E{Sm (Y, W, m)}(s) = 0,
i=1 Sm (Yi , Wi , m)(s)
− E{Sm (Yi , Wi , m)}(s) = op (1)kskp .
n
o
P
i ,Wi ,m)
i ,Wi ,m)
n−1 ni=1 ∂Sm (Y∂m
(s, v) − E ∂Sm (Y∂m
(s, v) = op (1)kskp kvkq ,
P
2 (Y , W , m)(s, v) − E{S 2 (Y , W , m)}(s, v) = o (1)ksk kvk ,
n−1 ni=1 Sm
i
i
i
p
p
q
m i
2 (Y , W , m)}(s, v) = E{S (Y, W, m)(s)S (Y, W, m)(v)} and
where E{Sm
i
i
m
m
E{∂Sm (Yi , Wi , m)/∂m}(s, v) = E[{∂Sm (Y, W, m)/∂m}(s, v)].
We now list the regularity conditions.
(D1) The true density functions fX0 (x), f () are bounded on their supports.
In addition, the support of fX0 (x) is compact.
(D2) E {∂Sm (Y, W, m)/∂m} is a bounded bilinear operator on L2 ×L2 , L1 ×
2 (Y , W , m)} is a bounded bilinear operator
L∞ , and L∞ × L1 . E{Sm
i
i
on L2 × L2 .
12
(D3) m(x) is bounded on [0, 1] and it satisfies m(x) ∈ C q ([0, 1]), q ≥ 1.
(D4) β 0 is a dβ dimensional spline coefficient such that supx∈[0,1] |Br (x)T β 0 −
m(x)| = Op (hqb ). The existence of such β 0 has been shown in De Boor
(1978).
(D5) The expectation
E[S∗β (Wi , Yi , β) − E ∗ {a(X, β)|Yi , Wi , β}]
has unique root for β in the neighborhood of β 0 . Its derivative with
respect to β is a smooth function of β in the neighborhood of β 0 , with
its singular values bounded and bounded away from zero. Denote the
unique zero β ∗ .
Conditions (D1) and (D2) contain bounded requirement on pdfs and operators and are not stringent. Condition (D1) further requires the compact
support of the distribution of X. This is similar to the density estimation
case and is crucial. Conditions (D3) and (D4) are regarding the mean function m(x) and its spline approximation which are quite standard. Condition
(D5) is a unique root requirement similar to (C6) and is used to exclude the
pathologic case where the estimating equation is constantly zero.
In the following, we establish the consistency of the parameter estimation
in Proposition 3 and then further analyze its convergence rate in Proposition
4. The results in these propositions are subsequently used to further establish
the asymptotic properties of the estimator of the regression mean function
m(x) in Theorem 2. The proofs of both the propositions and the theorem
are in Supplement S.5, S.6 and Appendix A.2 respectively.
Proposition 3.
b satisfy
Let β
n
X
i=1
Assume Conditions (C2) – (C4), (D1) – (D5) to hold.
b − E ∗ {a(X, β)
b | Yi , Wi , β}
b = 0.
S∗β (Wi , Yi , β)
b − β = op (1) element-wise.
Then β
0
Proposition 4.
b satisfy
Let β
n
X
i=1
Assume Conditions (C2) – (C4), (D3) – (D5) to hold.
b − E ∗ {a(X, β)
b | Yi , Wi , β}
b = 0.
S∗β (Wi , Yi , β)
NEW NONPARAMETRIC MEASUREMENT ERROR MODEL ESTIMATOR 13
b − β k2 = Op {(nh)−1/2 }. Further,
Then kβ
0
b −β =
β
0
(
E
∂[S∗β (Wi , Yi , β) − E ∗ {a(X, β) | Yi , Wi , β}]
∂β T
Br (·)T β=m(·)
!)−1
×T00 {1 + op (1)},
where
T00 = n
−1
n
X
i=1
S∗β (Wi , Yi , m) − E ∗ {a(X, m) | Yi , Wi , m}.
Theorem 2. Assume Conditions (C2) – (C4), (D1) – (D5) to hold. Let
b Then sup
b
− m(x)| = Op {(nhb )−1/2 }. Specifim(x)
b
= Br (x)T β.
x∈[0,1] |m(x)
q−1/2
cally, bias{m(x)}
b
= E{m(x)
b
− m(x)} = Op (hb
) and
p
nhb [m(x)
b
− m(x) − bias{m(x)}]
b
p
T
=
nh Br (x)
b(
!)−1
∂[S∗β (Wi , Yi , β) − E ∗ {a(X, β) | Yi , Wi , β}]
× − E
∂β T
Br (·)T β=m(·)
#
n
X
−1
×n
S∗β (Wi , Yi , m) − E ∗ {a(X, m) | Yi , Wi , m} + op (1).
i=1
∗ (x), m(x)
Theorem 2 shows that even with misspecified fX
b
converges to
m(x) with root-nhb rate, which is the standard nonparametric rate of the
regression mean function without measurement errors, and is considerably
faster than deconvolution estimators.
4. Simulation studies.
4.1. Performance of B-spline-assisted MLE. We conduct two simulation
studies to illustrate the finite sample performance of the proposed pdf and
regression mean estimators. To evaluate the B-spline MLE method for estimating the density functions, we generated data sets of sample sizes from
n = 500 to n = 2000, where X is generated from a beta distribution with
both shape parameters equal to 4. We then generated the measurement error
U from three models:
I(a): a normal distribution with mean 0, variance 0.25, denoted by
N (0, 0.25),
14
√
2, denoted by
I(b): a Laplace
distribution
with
mean
0
and
scale
0.5/
√
Lap(0, 0.5/ 2),
I(c): a uniform distribution on [−0.25, 0.25], denoted by Unif(−0.25, 0.25).
We used cubic splines to approximate the density functions, with number
of knots equal to the smallest integer larger than 1.7n1/5 . In the left panel
of Figure 1, we plotted
√ the averaged root-(nhb ) maximum absolute error
(MAE), calculated as nhb supx |fbX0 (x) − fX0 (x)|, versus the sample sizes
n. Following Theorem 1, the root-(nhb ) MAE has a constant order. This
translates to the curves in the plots that stabilize in a small range, which is
what we observe, especially when sample size grows to larger than 700.
We also compared the B-spline MLE method with the widely used deconvolution method for density estimation. In the upper row of Figure 2, we
plotted the average values of supx |fbX0 (x) − fX0 (x)| based on 200 simulations for both methods at different sample sizes. We adopted the two-stage
plug-in bandwidth selection method proposed in Delaigle and Gijbels (2002)
in implementing the deconvolution method. Much to our surprise, results in
the upper row of Figure 2 indicate that the B-spline MLE method outperforms the deconvolution method with rather significant gain in this case.
We suspect that this is because the measurement errors are quite large here
which caused difficulties for both methods but especially for the deconvolution method.
To further examine the performance of individual estimated pdf curves
from both methods, we also plotted the estimated density curves for sample
sizes 500, 1000 and 2000. Because the deconvolution method performs poorly
when the measurement errors are large (see upper row of Figure 2), we
reduced the error variances, and generated the three error distributions from
models
√
II(a): N (0, 0.0025), II(b): Lap(0, 0.05/ 2), II(c): Unif(−0.125, 0.125).
Under the reduced error variability, we performed the estimation and plotted the resulting density estimates and their 90% confidence bands. Figure
3 contains the results of the B-spline MLE and the deconvolution estimator
using the two-stage plug-in bandwidth (Delaigle and Gijbels, 2002), at sample size 500. The results show that although not as dramatic as the large
error case, the B-spline MLE is still closer to the true pdf with narrower
confidence band, hence is more precise than the deconvolution method. We
also provide the similar comparison results at sample sizes 1000 and 2000
in Figures S.1 and S.2 in the supplement. For a quantitative comparison,
we also computed the MAE between the true pdf curve and the estimated
pdf curve in the upper part of Table 1. These results show that the B-spline
NEW NONPARAMETRIC MEASUREMENT ERROR MODEL ESTIMATOR 15
MLE performs consistently better than deconvolution.
4.2. Performance of B-spline-assisted semiparametric estimator. We next
evaluated the finite sample performance of the B-spline semiparametric
mean regression method described in Section 2.2. We used sample sizes n
from 500 to 2000 and generated X from a beta distribution with both shape
parameters equal to 2. The true regression mean function is m(x) = sin(2πx)
and we generated the regression model errors from a normal distribution
with mean zero variance 0.25. We generated the measurement errors U from
the three different distributions described in model I (a)–(c). Cubic splines
were used to approximate the mean function, and the number of interior
knots were proportional to n1/5 . In the right √
panel of Figure 1, we plotted
the averaged root-(nhb ) MAE calculated via nhb supx |m(x)
b
− m(x)| as a
function of the sample size n. Similar to the density estimation experiment,
the curves stabilize as sample size increases, and is largely flattened after
n = 1000, indicating that supx |m(x)
b
− m(x)| has order (nhb )−1/2 .
We further compared the B-spline semiparametric method with the deconvolution method in the nonparametric mean regression model. In the lower
b
− m(x)| over 200 simurow of Figure 2, we plotted the average supx |m(x)
lations for both methods. Again, in this case, with moderate to significant
amount of noise, the B-spline semiparametric method greatly outperforms
the deconvolution method with smaller average error.
Similar to the pdf investigation, we reduce the measurement error variability and generated the errors from model II (a)–(c) to investigate the
mean function curve fitting. We plotted the mean function estimates and
the 90% confidence bands for the B-spline semiparametric and deconvolution methods in Figure 4 and Figures S.3 and S.4 in the supplement for
sample sizes 500, 1000 and 2000 respectively. These results indicate that the
B-spline semiparametric estimator indeed outperforms the deconvolution
method. Their performance difference in terms of averaged MAE between
the estimated and true mean curves are further provided in the lower half
of Table 1.
5. Data Example. Heavy fine particulate matter (PM2.5) air pollution has become a serious problem in China in recent years and its possible
effect on respiratory diseases has been a concern in public health. Starting from 2012, Beijing Environmental Protection Bureau (BEPB) has been
recording the daily PM2.5 levels in Beijing. Based on these data, Xu et al.
(2016) studied the effect of PM2.5 on asthma in year 2013. Specifically, they
explored the PM2.5 effect on the number of daily asthma emergency room
visits (ERV) in ten hospitals in Beijing, but found no significant effect. In
16
fact, the mean number of daily asthma ERVs even shows a decreasing trend
along the increase of measured PM2.5. This contradicts with the general
conclusion that PM2.5 has short term adverse effect on asthma (Fan et al.,
2016).
A potential reason of this inconsistency is the errors in the PM2.5 measurements which were not taken into account in the above analysis. In fact,
there are many debates on the accuracy of the PM2.5 reports, especially
in the early years such as 2013. For example, we compared the daily average PM2.5 reports in 2013 from 17 ambient air quality monitoring stations
and those reports from the “Mission China Beijing” website (Mission-China,
2016) maintained by the U.S. department of state, and show the two estimated pdfs of PM2.5 in Figure 5. It is clear that the estimated pdfs of PM2.5
from the two sources are very different, where the PM2.5 concentrations obtained from the BEPB yields a pdf estimate with the mode to the left of
that obtained from the “Mission China Beijing”, indicating a generally less
severe air pollution problem. This motivates us to consider the measurement
error issue in studying the effect of PM2.5 on the daily asthma ERV.
We restrict our analysis of PM2.5 in the range from 0 to 400. The largest
recorded PM2.5 value is 328. The data set we analyzed contains 337 observations. In the data available to us, the ith observation contains the number
of daily asthma ERVs, which we treat as response Yi , the average PM2.5
level over 17 ambient air quality monitoring stations from BEPB, which we
denote as Wi , and the PM2.5 level from “Mission China Beijing”, which
we write as W0i . To carry out the analysis, we let the true PM2.5 value
be Xi , and let Wki , Uki be the observed PM2.5 value and its corresponding
measurement error at the kth monitoring station, k = 1, . . . , 17.
We assume
P17
Uki ’s are independent
of each other and of Xi . Then Wi = k=1 Wki /17.
P
Writing U i = 17
U
k=1 ki /17, we have Wi = Xi + U i . Because our preliminary analysis result in Figure 5 suggests a possible discrepancy between
the measurements in Wi and in W0i , we allow a potential bias term b and
model W0i = b + Xi + U0i , where U0i is the measurement error of W0i .
We assume all the Uki , k = 0, . . . , 17 to have
with
P the same distribution
P
mean zero, and we estimate b by bb = n−1 ( ni=1 W0i − ni=1 Wi ), which
yields the value bb = 132. We further estimated the variance of U i based on
var(U ) = {var(W0i − Wi )}/18. This yields var(U
c ) = 0.008. Further, because
U i is the average of 17 iid Uki ’s, it is sensible to assume that U i has a normal
distribution.
Based on the above preliminary analysis, we proceed to estimate the pdf of
PM2.5, i.e. Xi , and the mean regression function of asthma ERVs conditional
on PM2.5, i.e. E(Yi | Xi ), using the B-spline-assisted MLE/semiparametric
NEW NONPARAMETRIC MEASUREMENT ERROR MODEL ESTIMATOR 17
methods in Sections 2.1 and 2.2. We further implemented 100 bootstraps to
estimate the asymptotic variances of the resulting estimators. We also compared the B-spline MLE and the B-spline semiparametric regression estimator with the deconvolution density and regression estimators. In implementing the B-spline approximation, we used two and three equally spaced knots
respectively, and in implementing the deconvolution methods, we used bandwidth 0.05. The number of knots is chosen based on the simulation studies in
Section 4. In terms of the bandwidth selection, because the crossvalidation
(Stefanski and Carroll, 1990) and the plug-in (Delaigle and Gijbels, 2002)
bandwidth selection algorithms tend to select small bandwidths that induce
large numerical errors, we used the smallest bandwidth that produced stable results in our analysis. The upper panel in Figure 6 shows the estimated
pdfs based on the B-spline MLE and the deconvolution method. Compared
with the kernel estimator in the same plots which ignores the measurement
errors, the B-spline MLE shows more difference than the deconvolution estimator. This is sensible given that the noise-to-signal ratio is quite large at
var(Ūi )/var(Xi ) = 0.2, hence measurement error is likely not ignorable. The
lower panel in Figure 6 provides the estimated mean of Yi as a function of
Xi . The B-spline semiparametric estimator shows a fluctuating pattern in
the range from 0 to 200, although the pattern is not significant. In the range
of PM2.5 concentration larger than 200, it shows clearly an increasing trend,
which agrees with the conclusion in Fan et al. (2016) that the exposure to
high PM2.5 has adverse effect on asthma onset. In contrast, the relation
from the deconvolution estimator is similar to that of the local linear regression estimator ignoring the measurement errors, and it is unable to detect
the increasing trend of the asthma ERVs as the PM2.5 level increases.
6. Discussion. We have developed a B-spline-assisted MLE for nonparametric pdf estimation and a B-spline-assisted semiparametric estimator
for nonparametric regression mean function, under the situation that the
covariates are measured with errors. The performance of both procedures
are superior to the widely used deconvolution methods, in terms of both
their theoretical convergence rate and their numerical performance.
A key difference between our procedures and the deconvolution procedures
is that we restrict our interest in estimating the functions on a compact set.
Given that practically most time, the possible range of a random variable
is indeed finite, hence the relevant information is indeed only functions in a
finite range, we are very curious if deconvolution methods can achieve the
same convergence property in such case, with possibly some modifications
on the existing procedures. To this end, Hall and Qiu (2005) provides some
18
relevant results, while we leave this general question as an open problem for
researchers who specialize in deconvolution methods.
Another curious issue is regarding kernel treatment to nonparametric
functions. In many nonparametric estimation problems, kernel methods and
B-spline methods generate comparable results. However, our procedures critically rely on the B-spline approximation to convert the unknown functions to “parametric” ones, hence we cannot readily replace the B-spline
method with kernel method. Whether or not kernel methods can be applied
in nonparametric density estimation and nonparametric regression estimation problems with measurement errors to achieve the same convergence rate
as B-spline based methods remains a challenging and interesting question.
The density and regression function estimation problems studied in this
work is the most basic ones. In practice, various complications may occur.
For example, the error distribution fU (u) may not be known and need to be
estimated parametrically or nonparametrically based on repeated measurements or other instruments. If one estimates fU (·) parametrically first then
insert it into our procedure, it will not have first order effect. However, if
fU (·) is estimated nonparametrically, it will in general affect both the bias
and variance of the resulting estimation of both the density and regression
functions. See Delaigle and Hall (2016) for deconvolution based estimation
incorporating unknown error distribution. The story may be even more interesting when U and X are correlated hence heteroscedastic measurement
errors occur. These are all very interesting problems and are worth careful
investigation.
Appendix.
A.1. Proof of Theorem 1. Note that
sup |log{fbX (x)} − log{fX0 (x)}|
x∈[0,1]
=
Tb
sup |Br (x) θ − log
x∈[0,1]
=
Z
0
1
b
exp{Br (x)T θ}dx
− log{fX0 (x)}|
b − Br (x)T θ 0 + Br (x)T θ 0 − log{fX0 (x)}
sup |Br (x)T θ
x∈[0,1]
−log
−log
Z
1
0
Z
0
1
b
exp{Br (x)T θ}dx
+ log
exp{Br (x)T θ 0 }dx|
Z
0
1
exp{Br (x)T θ 0 }dx
NEW NONPARAMETRIC MEASUREMENT ERROR MODEL ESTIMATOR 19
≤
b − θ 0 k2 } + sup |Br (x)T θ 0 − log{fX0 (x)}|
sup {kBr (x)k2 kθ
x∈[0,1]
+|log
x∈[0,1]
Z
1
exp{Br (x) θ}dx − log
0
+|log
Z
1
0
=
Z
1
0
exp{Br (x)T θ 0 }dx|
exp{Br (x)T θ 0 }dx|
b − θ 0 k2 } + sup |Br (x)T θ 0 − log{fX0 (x)}|
sup {kBr (x)k2 kθ
x∈[0,1]
x∈[0,1]
Z
Z 1
1
∂
T ?
b
log
exp{B
(x)
θ
}dx(
θ
−
θ
)|
+
|log
exp{Br (x)T θ 0 }dx|
r
0
T
∂θ
0
0
T
b
sup {kBr (x)k2 kθ − θ 0 k2 } + sup |Br (x) θ 0 − log{fX0 (x)}|
+|
≤
Tb
x∈[0,1]
Z
x∈[0,1]
1
∂
b − θ 0 )k2 + |log
+k T log
exp{Br (x)T θ ? }dxk2 k(θ
∂θ
0
= Op {(nhb )−1/2 + hqb },
Z
0
1
exp{Br (x)T θ 0 }dx|
b and θ 0 . The last equality holds
where θ ? is the point on the line connecting θ
R1
by Condition (C5), Proposition 2, and the fact that log 0 exp{Br (x)T θ 0 }dx
is a smooth function of Br (x)T θ 0 , which implies
Z 1
|log
exp{Br (x)T θ 0 }dx|
0
Z 1
Z 1
exp{Br (x)T θ 0 }dx − log
exp[log{fX0 (x)}]dx| = Op (hqb )
= |log
0
0
following Condition (C5). Further for any x,
"
#
T θ}
b
exp{B
(x)
r
bias{fbX (x)} = E R 1
− fX0 (x)
Tb
0 exp{Br (x) θ}dx
#
"
# "
b
exp{Br (x)T θ}
exp{Br (x)T θ 0 }
− R1
= E R1
Tb
exp{Br (x)T θ 0 }dx
0 exp{Br (x) θ}dx
"
# "0
#
exp{Br (x)T θ 0 }
exp[log{fX0 (x)}]
+ R1
− R1
T
0 exp{Br (x) θ 0 }dx
0 exp[log{fX0 (x)}]dx
"
#
!
∂
exp{Br (x)T θ 3 }
b − θ 0 ) + Op (hq ),
(A.1)
= E
(θ
R
b
∂θ T 01 exp{Br (x)T θ 3 }dx
b The third equality holds by
where θ 3 is on the line connecting θ 0 and θ.
Condition (C5). Consider the expectation in (A.1), recall that from (S.8),
b − θ 0 = −R1 (θ ∗ )−1 (R00 + R01 ),
(A.2)
θ
20
where R00 , R01 , R1 (θ) are defined in Proposition 2 and (S.10) and (S.9)
respectively. Further
!
#
"
∂
exp{Br (x)T θ 3 }
∗ −1
R1 (θ ) R00
E
R
∂θ T 01 exp{Br (x)T θ 3 }dx
"
#
∂
exp{Br (x)T θ 3 }
= E
R1 (θ ∗ )−1
T R1
T
∂θ
0 exp{Br (x) θ 3 }dx
)
!
Z 1(
fX0 (t)fU (Wi − t)
×
− fX0 (t) Br (t)dt
R1
0
0 fX0 (x)fU (Wi − x)dx
#
"
Z 1Z
∂
exp{Br (x)T θ 3 }
R1 (θ ∗ )−1
=
R
∂θ T 01 exp{Br (x)T θ 3 }dx
0
)
(
fX0 (t)fU (Wi − t)
× R1
− fX0 (t) fW 0 (Wi )Br (t)dWi dt
0 fX0 (x)fU (Wi − x)dx
#
"
Z 1Z
∂
exp{Br (x)T θ 3 }
=
R1 (θ ∗ )−1
T R1
T
∂θ
0
0 exp{Br (x) θ 3 }dx
!
fX0 (t)fU (Wi − t)
×R 1
fW 0 (Wi )Br (t) dWi dt
f
(x)f
(W
−
x)dx
i
X0
U
0
"
#
Z 1Z
∂
exp{Br (x)T θ 3 }
−
R1 (θ ∗ )−1 fX0 (t)fW 0 (Wi )Br (t)dWi dt
R
∂θ T 01 exp{Br (x)T θ 3 }dx
0
"
#
Z 1Z
exp{Br (x)T θ 3 }
∂
R1 (θ ∗ )−1 fX0 (t)fU (Wi − t)Br (t)dWi dt
=
R
∂θ T 01 exp{Br (x)T θ 3 }dx
0
"
#
Z 1Z
∂
exp{Br (x)T θ 3 }
−
R1 (θ ∗ )−1 fX0 (t)fU (Wi − t)Br (t)dWi dt
R
∂θ T 01 exp{Br (x)T θ 3 }dx
0
= 0.
(A.3)
In addition, we have
E
(A.4) ≤ E
∂
∂θ T
∂
∂θ T
"
R1
"
0
exp{Br (x)T θ 3 }
exp{Br (x)T θ 3 }dx
R1
0
exp{Br (x)T θ 3 }
#
exp{Br (x)T θ 3 }dx
R1 (θ ∗ )−1 R01
#
2
!
kR1 (θ ∗ )−1 k2 kR01 k2
!
.
NEW NONPARAMETRIC MEASUREMENT ERROR MODEL ESTIMATOR 21
First note that,
#
"
exp{Br (x)T θ 3 }
∂
(A.5)
R1
T
∂θ
0 exp{Br (x) θ 3 }dx
(d
θ
X
exp{Br (x)T θ 3 }Brk (x)
=
R1
T
k=1
0 exp{Br (x) θ 3 }dx
2
!2 1/2
R1
exp{Br (x)T θ 3 } 0 exp{Br (x)T θ 3 }Brk (x)dx
−
R1
[ 0 exp{Br (x)T θ 3 }dx]2
#2 1/2
"
dθ
T
X
exp{Br (x) θ 3 }Brk (x)
≤
R1
T
k=1
0 exp{Br (x) θ 3 }dx
!2 1/2
Z 1
dθ
X
T
T
exp{Br (x) θ 3 } exp{Br (t) θ 3 }
Brk (t)dt
+
R1
[ exp{Br (t)T θ 3 }dt]2
0
k=1
"
≤ R1
+
exp{Br (x)T θ 3 }
0
Z
0
exp{Br (x)T θ 3 }dx
1
0
"
#2
dθ
X
k=1
(x)T θ
1/2
2
Brk
(x)
(t)T θ
exp{Br
3 } exp{Br
R1
[ 0 exp{Br (t)T θ 3 }dt]2
#2
3}
!2
1/2
dθ
X
k=1
1/2
2
Brk (t)dt
dθ
X
exp{Br (x)T θ 3 }
2
= R1
(x)
Brk
T θ }dx
exp{B
(x)
r
3
k=1
0
1/2
!
dθ
Z 1 exp{B (x)T θ } exp{B (t)T θ } 2 X
r
3
r
3
∗
2
+
dt
Brk (t )
R1
0
[ exp{Br (t)T θ 3 }dt]2
k=1
0
= Op (1),
where t∗ is the point in (0, 1). In the above derivation, the first inequality
holds by the trianglar inequality, the second inequality holds by CauchySchwarz inequality, the second equality holds by the mean value theorem,
the last equality holds because
kBr (x)k22
=
dθ
X
k=1
2
Brk
(x)
≤ sup Brk (x)
k
dθ
X
k=1
Brk (x) = sup Brk (x) = Op (1).
k
22
Combining (A.4), (A.5), (S.12), (S.18), we have
!
#
"
exp{Br (x)T θ 3 }
∂
R1 (θ ∗ )−1 R01
E
R
∂θ T 01 exp{Br (x)T θ 3 }dx
(A.6)
1/2
q−1/2
q+1
= Op (h−1
b dθ hb ) = Op (hb
).
Combining (A.6), (A.2) and (A.3), we obtain
#
!
"
∂
exp{Br (x)T θ 3 }
b − θ 0 ) = Op (hq−1/2 ).
E
(θ
R
b
∂θ T 01 exp{Br (x)T θ 3 }dx
q−1/2
Plug this result to (A.1), we have bias{fbX (x)} = Op (hb
).
Hence,
h
i
p
nhb fbX (x) − fX0 (x) − bias{fbX (x)}
#
"
b
p
exp{Br (x)T θ}
− fX0 (x) − bias{fbX (x)}
=
nhb R 1
T θ}dx
b
exp{B
(x)
r
" 0
b
p
exp{Br (x)T θ 0 }
exp{Br (x)T θ}
=
− R1
nhb R 1
Tb
T
0 exp{Br (x) θ}dx
0 exp{Br (x) θ 0 }dx
#
exp{Br (x)T θ 0 }
+R 1
− fX0 (x) − bias{fbX (x)}
T θ }dx
exp{B
(x)
r
0
0
#
"
b
p
exp{Br (x)T θ 0 }
exp{Br (x)T θ}
− R1
+ Op (hqb n1/2 )
=
nhb R 1
T
T
b
0 exp{Br (x) θ}dx
0 exp{Br (x) θ 0 }dx
#
"R 1
"
p
fX0 (x)fU (Wi − x)Br (x)Br (x)T dx
exp{Br (x)T θ 0 }
∂
=
nhb T R 1
E 0
R1
T
∂θ
0 exp{Br (x) θ 0 }dx
0 fX0 (x)fU (Wi − x)dx
nR
o⊗2
1
Z 1
0 fX0 (x)fU (Wi − x)Br (x)dx
−
−
fX0 (x)Br (x)Br (x)T dx
nR
o2
1
0
0 fX0 (x)fU (Wi − x)dx
(
Z 1
⊗2 #!−1
n Z 1
X
fX0 (x)fU (Wi − x)
+
fX0 (x)Br (x)dx
n−1
R1
0
i=1 0
0 fX0 (x)fU (Wi − x)dx
−fX0 (x)} Br (x)dx + op (1).
The last equality hold by Proposition 2 and Condition (C3). This proves the
results.
NEW NONPARAMETRIC MEASUREMENT ERROR MODEL ESTIMATOR 23
A.2. Proof of Theorem 2.
sup |m(x)
b
− m(x)|
x∈[0,1]
=
b − Br (x)T β + Br (x)T β − m(x)|
sup |Br (x)T β
0
0
x∈[0,1]
≤
b − Br (x)T β | + sup |Br (x)T β − m(x)|
sup |Br (x)T β
0
0
x∈[0,1]
x∈[0,1]
−1/2
= Op {(nhb )
+
hqb },
by Proposition 2 and Condition (C5).
bias{m(x)}
b
(A.7)
= E{m(x)
b
− m(x)}
b − Br (x)T β } + {Br (x)T β − m(x)}
= E{Br (x)T β
0
0
b − Br (x)T β } + Op (hq ).
= E{Br (x)T β
0
b
The last equality holds by Condition (C5). By (S.21) and Proposition 2, we
can see that
b − Br (x)T β }|
|E{Br (x)T β
0
∗
∂[Sβ (Wi , Yi , β) − E ∗ {a(X, β) | Yi , Wi , β}]
= E Br (x)T E
∂β T
× (T00 + T01 ){1 + op (1)}}
∂[S∗β (Wi , Yi , β) − E ∗ {a(X, β) | Yi , Wi , β}]
T
= E Br (x) E
∂β T
×T01 {1 + op (1)}}
∂[S∗β (Wi , Yi , β) − E ∗ {a(X, β) | Yi , Wi , β}]
≤ E
E
∂β T
Br (·)T β=m(·)
!−1
Br (·)T β=m(·)
!−1
Br (·)T β=m(·)
!−1
T
×kBr (x) k2 kT01 k2 {1 + op (1)}
q+1/2
= Op (h−1
b hb
q−1/2
) = Op (hb
).
The second equality holds because E(T00 ) = 0. The fourth line holds by the
Pdβ
Pdβ
fact that k=1
{uk Brk (x)}2 = M k=1
Brk (x) = M for some finite constant
2
24
M and hence kBr (x)T k2 = O(1),
E
∂[S∗β (Wi , Yi , β) − E ∗ {a(X, β) | Yi , Wi , β}]
∂β
T
Br (·)T β=m(·)
1/2
q+1/2
by (S.28), and kT01 k2 ≤ dβ kT01 k∞ = Op (hb
result to (A.7), we have
q−1/2
Hence
p
=
p
=
p
bias{m(x)}
b
= Op (hb
!−1
2
= Op (h−1
b )
) by (S.25). Plug this
).
nhb [m(x)
b
− m(x) − bias{m(x)}]
b
nh Br (x)T
b(
!)−1
∂[S∗β (Wi , Yi , β) − E ∗ {a(X, β) | Yi , Wi , β}]
× − E
∂β T
Br (·)T β=m(·)
#
n
p
1X ∗
Sβ (Wi , Yi , m) − E ∗ {a(X, m) | Yi , Wi , m} − nhb bias{m(x)}
b
×
n
i=1
p
nh Br (x)T
=
b(
!)−1
∂[S∗β (Wi , Yi , β) − E ∗ {a(X, β) | Yi , Wi , β}]
× − E
∂β T
Br (·)T β=m(·)
#
n
X
∗
−1
∗
×n
Sβ (Wi , Yi , m) − E {a(X, m) | Yi , Wi , m} + Op (n1/2 hqb )
i=1
nh Br (x)T
b(
!)−1
∂[S∗β (Wi , Yi , β) − E ∗ {a(X, β) | Yi , Wi , β}]
× − E
∂β T
Br (·)T β=m(·)
#
n
X
−1
×n
S∗β (Wi , Yi , m) − E ∗ {a(X, m) | Yi , Wi , m} + op (1),
i=1
q−1/2
The second equality holds by the fact that bias{m(x)}
b
= Op (hb
last equality holds by Condition (C3). This proves the result.
). The
Acknowledgment. We thank Yang Li for providing the daily asthma
emergency room visits data and the referees for helpful comments.
NEW NONPARAMETRIC MEASUREMENT ERROR MODEL ESTIMATOR 25
normal error
laplace error
uniform error
7
6
5
4
Root nh maximum absolute error
3
10
2
5
Root nh maximum absolute error
15
8
normal error
laplace error
uniform error
500
1000
1500
2000
500
1000
1500
Sample size
2000
size
Fig 1. Performance of B-spline MLE pdf estimation (upper) and B-spline semiparametric
mean estimation (lower). Results based on 200 simulations.
1.0
0.40
0.35
0.25
0.30
Maximum absolute error
2.5
2.0
1.5
Maximum absolute error
1.0
0.8
0.6
0.20
0.4
Maximum absolute error
0.45
0.50
Fig 2. Comparison of pdf (upper) and mean (lower) estimators based on the B-spline MLE
or B-spline semiparametric estimator (solid) and the deconvolution (dashed) method, when
measurement errors are norm (left), Laplace (middle) and uniform (right) respectively.
Average maximum absolute error supx |fbX (x) − fX (x)| (upper) or supx |m
b X (x) − mX (x)|
(lower) is computed based on 200 simulations at sample sizes from 500 to 2000.
1500
2000
500
1000
1500
2000
500
1500
2000
0.6
Maximum absolute error
0.5
0.4
0.2
0.2
0.3
0.3
Maximum absolute error
2000
0.7
0.7
0.6
1.0
0.8
0.6
0.4
1500
Sample size
0.1
0.2
Maximum absolute error
1000
Sample size
0.8
Sample size
0.5
1000
0.4
500
500
1000
1500
Sample size
2000
500
1000
1500
Sample size
2000
500
1000
Sample size
26
4
3
f
2
1
0
0
1
f
2
3
4
Fig 3. B-spline MLE density estimation (left) and deconvolution estimation (right) from
200 simulations: The solid lines represent the true functions and the dash lines represent
the estimated functions and their 90% confidence bands. The first row to third row are the
results for model II (a)–(c) respectively. Sample size 500.
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.8
1.0
0.6
0.8
1.0
0.6
0.8
1.0
4
3
f
0
1
2
3
f
2
1
0
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
3
f
1
0
0
1
2
2
3
4
x
4
x
f
0.6
x
4
x
0.0
0.2
0.4
0.6
x
0.8
1.0
0.0
0.2
0.4
x
NEW NONPARAMETRIC MEASUREMENT ERROR MODEL ESTIMATOR 27
3
m
−2
−3
−2
−1
−1
0
0
m
1
1
2
2
3
Fig 4. B-spline semiparametric regression estimation (left) and deconvolution estimation
(right) from 200 simulations: The solid lines represent the true functions and the dash
lines represent the estimated functions and their 90% confidence bands. The first row to
third row are the results for model II (a)–(c), respectively. Sample size 500.
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
0.6
0.8
1.0
0.6
0.8
1.0
m
−2
−3
−2
−1
−1
0
0
m
1
1
2
2
3
x
3
x
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
m
0
0
−1
−1
−2
−2
−3
m
1
1
2
2
3
x
3
x
0.0
0.2
0.4
0.6
x
0.8
1.0
0.0
0.2
0.4
x
28
0.004
0.000
0.002
Density
0.006
Fig 5. The estimated pdf of PM2.5 without considering measurement error, based on data
from Beijing Environmental Protection Bureau (solid line) and “Mission China” website
(dash-line).
0
100
200
300
400
500
Observed PM2.5
Table 1
Performance comparison between B-spline MLE/semiparametric estimator and
deconvolution. Mean of the maximum absolute differences are reported. Results based on
average over 200 simulations.
model II(a)
model II(b)
model II(c)
model II(a)
model II(b)
model II(c)
pdf estimation: E{supx |fbX0 (x) − fX0 (x)|}
B-spline MLE
Deconvolution
n = 500 n = 1000 n = 2000 n = 500 n = 1000 n = 2000
0.204
0.160
0.109
0.238
0.212
0.160
0.213
0.153
0.110
0.230
0.197
0.158
0.222
0.184
0.124
0.315
0.242
0.230
mean estimation: E{supx |m(x)
b
− m(x)|}
B-spline semiparametric
Deconvolution
n = 500 n = 1000 n = 2000 n = 500 n = 1000 n = 2000
0.370
0.263
0.175
0.908
0.796
0.762
0.425
0.264
0.163
0.857
0.828
0.779
0.414
0.291
0.219
0.880
0.832
0.801
References.
Berry, S. M., Carroll, R. J. and Ruppert, D. (2002). Bayesian smoothing and regression splines for measurement error problems. Journal of the American Statistical
Association 97 160–169.
Carroll, R. J. and Hall, P. (1988). Optimal rates of convergence for deconvolving a
density. Journal of the American Statistical Association 83 1184-1186.
Conway, J. B. (1990). A Course in Functional Analysis. Springer-Verlag, New York.
De Boor, C. (1978). A practical guide to splines 27. Springer-Verlag, New York.
Delaigle, A. and Gijbels, I. (2002). Estimation of integrated squared density derivatives
from a contaminated sample. Journal of the Royal Statistical Society: Series B 64 869–
886.
NEW NONPARAMETRIC MEASUREMENT ERROR MODEL ESTIMATOR 29
Density
0
50
100
150
200
250
0.000 0.002 0.004 0.006 0.008 0.010 0.012 0.014
Density
0.000 0.002 0.004 0.006 0.008 0.010 0.012 0.014
Fig 6. The B-spline MLE (upper left) and deconvolution (upper right) pdf estimators
and their 95% confidence bands of PM2.5. The B-spline semiparametric (bottom left) and
deconvolution (bottom right) regression estimators and their 95% confidence bands. The
red line is the naive estimator ignoring measurement errors.
300
0
50
100
200
250
300
250
300
8
7
4
5
6
Mean asthma ERV
7
6
5
4
Mean asthma ERV
150
True PM2.5
8
True PM2.5
0
50
100
150
200
True PM 2.5
250
300
0
50
100
150
200
True PM 2.5
30
Delaigle, A. and Hall, P. (2016). Methodology for non-parametric deconvolution when
the error distribution is unknown. Journal of the Royal Statistical Society: Series B 78
231–252.
Fan, J. (1991). On the Optimal Rates of Convergence for Nonparametric Deconvolution
Problems. Annals of Statistics 19 1257-1272.
Fan, J. and Truong, Y. K. (1993). Nonparametric Regression with Errors in Variables.
Annals of Statistics 21 1900-1925.
Fan, J., Li, S., Fan, C., Bai, Z. and Yang, K. (2016). The impact of PM2.5 on asthma
emergency department visits: a systematic review and meta-analysis. Environmental
Science and Pollution Research 23 843–850.
Hall, P. and Qiu, P. (2005). Discrete-transform approach to deconvolution problems.
Biometrika 92 135–148.
Liu, M. C. and Taylor, R. L. (1989). A consistent nonparametric density estimator for
the deconvolution problem. Canadian Journal of Statistics 17 427-438.
Ma, S., Carroll, R. J., Liang, H. and Xu, S. (2015). Estimation and Inference in Generalized Additive Coefficient Models for Nonlinear Interactions with High-dimensional
Covariates. Annals of Statistics 43 2102–2131.
Masri, R. and Redner, R. A. (2005). Convergence rates for uniform B-spline density
estimators on bounded and semi-infinite domains. Nonparametric Statistics 17 555–
582.
Mission-China (2016). Accessed:09-30. http://www.stateair.net/web/historical/1/1.html.
Staudenmayer, J., Ruppert, D. and Buonaccorsi, J. P. (2008). Density Estimation in
the Presence of Heteroscedastic Measurement Error. Journal of the American Statistical
Association 103 726–736.
Stefanski, L. A. and Carroll, R. J. (1990). Deconvoluting kernel density estimators.
Statistics 21 169-184.
Tsiatis, A. A. and Ma, Y. (2004). Locally Efficient Semiparametric Estimators for Functional Measurement Error Models. Biometrika 91 835-848.
Wang, L. and Yang, L. (2009). Spline estimation of single-index models. Statistica Sinica
765–783.
Xu, Q., Li, X., Wang, S., Wang, C., Huang, F., Gao, Q., Wu, L., Tao, L., Guo, J.,
Wang, W. et al. (2016). Fine particulate air pollution and hospital emergency room
visits for respiratory disease in urban areas in Beijing, China, in 2013. PloS one 11
e0153099.
Zhang, C. H. (1990). Fourier methods for estimating mixing densities and distributions.
Annals of Statistics 18 806-830.
Department of Statistics
University of Hongkong
Pok Fu Lam, Hongkong
E-mail: feijiang@hku.hk
Department of Statistics
Penn State University
University Park, PA 16802
E-mail: yzm63@psu.edu
SUPPLEMENTARY DOCUMENT FOR “A SPLINE-ASSISTED
SEMIPARAMETRIC APPROACH TO NONPARAMETRIC
MEASUREMENT ERROR MODELS”
By Fei Jiang∗ and Yanyuan Ma†
University of Hong Kong∗ and Penn State University†
S.1. Definitions. Let H be the separable Hilbert space of square integrable
R1
functions on [0, 1], and let < φ1 , φ2 >≡ 0 φ1 (x)φ2 (x)dx denote the inner product
of the functions φ1 , φ2 in H. Let B ⊂ H be the space spanned by the B-spline bases.
Then the L2 norm k · k2 is the norm induced by the inner product. We denote the
functional sup norm as k · k∞ , and the Lp norm as k · kp .
In addition, for any random function g(x) with the ith observation gi (x), we
define Γ(g) to be a second moment based linear operator such that
Z 1
Γ(g)φ(x) ≡
E{gi (x)gi (s)}φ(s)ds
0
= E{< gi , φ > gi (x)}
while its empirical version can be written as
Γn (g)φ(x) ≡ n
−1
n
X
< gi , φ > gi (x).
i=1
Using these definitions, we can write
Z 1
Z
< Γ(g)Brk , Brl > = E
Brk (x)gi (x)dx
< Γn (g)Brk , Brl > = n
−1
0
n Z 1
X
i=1
0
1
0
Brk (x)gi (x)dx
Brl (x)gi (x)dx ,
Z
1
Brl (x)gi (x)dx.
0
b
Define C(g) as a dθ × dθ matrix with its (k, l) element < Γ(g)Brk , Brl >, and C(g)
as a dθ × dθ matrix matrix with its (k, l) element < Γn (g)Brk , Brl >.
S.2. Lemmas and their Proofs.
1
2
Lemma 1.
There is a constant Dr > 0 such that for each spline
and for each 1 ≤ p ≤ ∞
0
Dr kc kp ≤ k
dθ
X
k=1
Pdθ
k=1 ck Brk ,
ck Brk kp ≤ kc0 kp ,
where c0 = {ck {(tk − tk−r )/r}1/p , k = 1, . . . , dθ }T .
Proof: This is a direct consequence of Theorem 5.4.2 on page 145 in DeVore and
Lorentz (1993).
Lemma 2.
Let Γ(g) be a strictly positive definite operator on B. Let u be a
dθ -dimensional vector with kuk2 = 1. There exist positive constants D1 , D2 , D3 , D4
such that
D1 hb ≤ uT C(g)u ≤ D2 hb ,
−1 −1
T
−1
D2−1 h−1
b ≤ u C(g) u ≤ D1 hb ,
and
b
D3 hb ≤ uT C(g)u
≤ D4 hb ,
−1 −1
Tb
−1
D4−1 h−1
b ≤ u C(g) u ≤ D3 hb
in probability.
Proof: First note that by the Cauchy-Schwartz inequality, Lemma 1 and Condition
(A6), we have
E
"
Z
≤ E
uT
0
0
=
Z
2 #
Br (x)gi (x)dx
(d
θ
1 X
Z 1 (X
dθ
0
1
uk Brk (x)
k=1
)2
uk Brk (x)
k=1
≤ ku0 k22 O(1)
= O(hb ),
)2
dx
dxE
Z
1
0
Z
0
1
gi (x)2 dx
2
gi (x) dx
NEW NONPARAMETRIC MEASUREMENT ERROR MODEL ESTIMATOR
3
where u0 = {uk {(tk − tk−r )/r}1/2 , k = 1, . . . , dθ }T , whose L2 norm is of order
1/2
O(hb ). Thus we have uT C(β)u ≤ D2 hb for some positive constant D2 < ∞.
As shown in (28) in Cardot, Ferraty and Sarda (2003), since the eigenvalues of
the covariance operator Γ(g) are strictly positive definite on B, there is a positive
constant D such that
< Γ(g)φ, φ >≥ Dkφk2 , for φ ∈ B.
Note that BT
r u ∈ B, so we have
T
T
2
uT C(g)u =< Γ(g)BT
r u, Br u >≥ DkBr uk2 ≥ D1 kuk hb = D1 hb
for a positive constant D1 by Lemma 1 and Condition (A4). Therefore, D1 hb ≤
−1 −1
T −1
uT C(g)u ≤ D2 hb . And so D2−1 h−1
b ≤ u C (g)u ≤ D1 hb . Further, with Theo-
rem 1.19 in Chatelin (1983), we have
b
kC(g)
− C(g)k2 ≤ sup
dθ
X
1≤l≤dθ k=1
kΓn − Γk2 | < Brk , Brl > |
As shown in Cardot, Ferraty and Sarda (2003), Lemma 5.3 in Cardot, Ferraty and
Sarda (1999) implies kΓn − Γk2 = op (n(hb −1)/2 ). Further, by the property of BPθ
spline basis, we have when |k − l| > r, Brk Brl = 0. Therefore, sup1≤l≤dθ dk=1
|<
Brk , Brl > | = O(hb ), which implies
(S.1)
b
kC(g)
− C(g)k2 ≤ op (hb n(hb −1)/2 ).
Now because hb < 1, combine with the result that D1 hb ≤ uT C(g)u ≤ D2 hb , by the
triangular inequality we obtain
b
b
uT C(g)u
= uT C(g)u + uT {C(g)
− C(g)}u
b
≤ D2 hb + kuk2 k{C(g)
− C(g)}uk2
b
≤ D2 hb + k{C(g)
− C(g)}k2 kuk2
= D2 hb + op (hb )
and
b
b
uT C(g)u
= uT C(g)u + uT {C(g)
− C(g)}u
4
b
≥ D1 hb − kuk2 k{C(g)
− C(g)}uk2
b
≥ D1 hb − k{C(g)
− C(g)}k2 kuk2
= D1 hb + op (hb ).
b
Thus, D3 hb ≤ uT C(g)u
≤ D4 hb in probability for some positive constant D3 , D4 <
b −1 u ≤ D−1 h−1 in probability. This proves the result.
∞. And so D4−1 hb ≤ uT C(g)
3
b
Lemma 3.
Assume Ci (·) is a continuous random function of x ∈ [0, 1]. At each
x, |E{Ci (x)}| < ∞. kCi (·)k2 < ∞ a.s.. Then
|n
−1
n Z
X
1
Brk (x)Ci (t)dx| = Op (hb )
0
i=1
if E{Ci (x)} =
6 0 and
|n
n Z
X
−1
i=1
1
0
Brk (x)Ci (x)dx| = Op {
p
hb n−1 log(n)}
if E{Ci (x)} = 0.
Proof: By the Bernstein’s inequality in Bosq (1998), we have
n Z
X
1/n
i=1
"
= Op
n
X
i=1
0
1
Brk (x)Ci (x)dx − E
E 1/n
Z
Z
1
Brk (x)Ci (x)dx
0
1
Brk (x)Ci (x)dx
0
p
= Op { hb n−1 log(n)}.
2
logn
#1/2
The last equality holds from Lemma 1 by choosing gi = Ci . Now if E{Ci (x)} =
6 0,
then because Brk is positive in the interval (tk−r , tk ), and is 0 otherwise (page 88 in
De Boor (1978))
Z
E
0
1
Brk (x)Ci (x)dx
=
Z
1
Brk (x)E{Ci (x)}dx
Z 1
= |E{Ci (ξ)}
Brk (x)dx|
0
Z 1
= |E{Ci (ξ)}|
Brk (x)dx
0
0
NEW NONPARAMETRIC MEASUREMENT ERROR MODEL ESTIMATOR
5
≤ D9 (tk − tk−r )
(S.2)
= Op (hb ).
where ξ is a point in the interval [0, 1], D9 is a finite constant. The second equality holds by the assumption that Ci (·) is continuous function in x and the mean
value theorem. The inequality holds because the support of Brk is the interval
(tk−r , tk ) and |E{Ci (x)}| < ∞ for any x ∈ [0, 1]. Therefore, by Condition (A4)
that N −1 n(logn)−1 → ∞ we have
kn
−1
n Z
X
i=1
1
Brk (x)Ci (x)dxk2 = Op (hb )
0
for E{Ci (x)} =
6 0 and
kn
−1
n Z
X
i=1
1
Brk (x)Ci (x)dxk2 =
0
p
hb n−1 log(n)
for E{Ci (x)} = 0. This proves the results.
S.3. Proof of Proposition 1. First note that for any h(x), we have
(S.3)
R1
exp[log{fX0 (x)}]fU (Wi − x)h(x)dx
R1
0 exp[log{fX0 (x)}]fU (Wi − x)dx
!
R1
exp[log{f
(x)}]h(x)dx
X0
− 0R 1
0 exp[log{fX0 (x)}]dx
)
Z 1
Z (R 1
f
(x)f
(y
−
x)h(x)dx
i
X0
U
0
−
fX0 (x)h(x)dx
=
R1
0
0 fX0 (x)fU (yi − x)dx
Z 1
×
fX0 (x)fU (wi − x)dxdµ(yi )
0
Z 1Z
Z 1
=
fX0 (x)fU (wi − x)h(x)dµ(wi )dx −
fX0 (x)h(x)dx
E
0
0
(S.4)
0
= 0,
hence
"R 1
exp[log{fX0 (x)}]fU (Wi − x)Brk (x)dx
R1
0 exp[log{fX0 (x)}]fU (Wi − x)dx
#
R1
0 exp[log{fX0 (x)}]Brk (x)dx
−
=0
R1
exp[log{f
(x)}]dx
X0
0
E
0
6
for each Brk (x), k = 1, . . . , dθ . Therefore, by Condition (C5), we have
"R 1
exp{Br (x)T θ 0 }fU (Wi − x)Brk (x)dx
E 0 R1
T
0 exp{Br (x) θ 0 }fU (Wi − x)dx
#
R1
T
0 exp{Br (x) θ 0 }Brk (x)dx
= Op (hqb ),
−
R1
T
0 exp{Br (x) θ 0 }dx
which suggests that
Further, since
n
X
0
n
X
"R 1
b U (Wi − x)Brk (x)dx
exp{Br (x)T θ}f
R1
Tb
i=1
0 exp{Br (x) θ}fU (Wi − x)dx
#
R1
T θ}B
b rk (x)dx
exp{B
(x)
r
− 0 R1
= 0,
b
exp{Br (x)T θ}dx
n−1
(S.5)
"R 1
exp{Br (x)T θ 0 }fU (Wi − x)Brk (x)dx
R1
T
i=1
0 exp{Br (x) θ 0 }fU (Wi − x)dx
#
R1
T θ }B (x)dx
exp{B
(x)
r
0
rk
− 0 R1
= op (1).
T
0 exp{Br (x) θ 0 }dx
n
−1
0
0
we have
(S.6)
n
X
"R 1
exp{Br (x)T θ 0 }fU (Wi − x)Br (x)dx
R1
T
i=1
0 exp{Br (x) θ 0 }fU (Wi − x)dx
#
R1
T θ }B (x)dx
exp{B
(x)
r
0
r
− 0 R1
T
0 exp{Br (x) θ 0 }dx
"
R1
n
Tb
X
−1
0 exp{Br (x) θ}fU (Wi − x)Br (x)dx
−n
R1
Tb
i=1
0 exp{Br (x) θ}fU (Wi − x)dx
#
R1
Tb
0 exp{Br (x) θ}Br (x)dx
= op (1)
− R1
b
exp{Br (x)T θ}dx
n
−1
0
0
element-wise. By Condition (C6), as a function of θ,
"R 1
#!
R1
T
T
0 exp{Br (x) θ}fU (Wi − x)Br (x)dx
0 exp{Br (x) θ}Br (x)dx
E
−
R1
R1
T θ}f (W − x)dx
T
exp{B
(x)
r
i
U
0
0 exp{Br (x) θ}dx
has its derivative with respect to θ being a nonsingular matrix in the neighborhood
of its zero. Thus,
"R 1
#
R1
n
T
T
X
−1
0 exp{Br (x) θ}Brk (x)dx
0 exp{Br (x) θ}fU (Wi − x)Br (x)dx
n
−
,
R1
R1
T θ}f (W − x)dx
T θ}dx
exp{B
(x)
exp{B
(x)
r
i
r
U
i=1
0
0
NEW NONPARAMETRIC MEASUREMENT ERROR MODEL ESTIMATOR
7
also has nonsingular derivative with respect to θ and hence is an invertible function
of θ in the neighborhood. Therefore, from (S.5) and (S.6), by the continuous mapping
b − θ 0 = op (1) element-wise. This proves the results.
theorem, we have θ
S.4. Proof of Proposition 2. Proof: First note by using (S.3) at h(x) =
δ(x − x0 ) and then set x0 = x, we have
(S.7)
E
R1
0
= 0.
exp[log{fX0 (x)}]fU (Wi − x)
exp[log{fX0 (x)}]fU (Wi − x)dx
− R1
0
exp[log{fX0 (x)}]
exp[log{fX0 (x)}]dx
b is the maximum likelihood estimator, we have
Since θ
"
Z 1
n
X
∂
−1
b U (Wi − x)dx
n
log
exp{Br (x)T θ}f
0 =
∂θ
0
i=1
Z 1
Tb
−log
exp{Br (x) θ}dx
0
(S.8)
where for any θ,
b − θ 0 ) + op {kR1 (θ 0 )(θ
b − θ 0 )k2 },
= R0 + R1 (θ 0 )(θ
R1 (θ) =
(S.9)
"
Z 1
n
X
∂2
−1
n
log
exp{Br (x)T θ}fU (Wi − x)dx
∂θ∂θ T
0
i=1
Z 1
exp{Br (x)T θ}dx ,
−log
0
b and
and θ ∗ is on the line connecting θ 0 and θ,
n R1
T
X
−1
0 exp{Br (x) θ 0 }fU (Wi − x)Br (x)dx
R0 = n
R1
T
i=1
0 exp{Br (x) θ 0 }fU (Wi − x)dx
R1
exp{Br (x)T θ 0 }Br (x)dx
− 0 R1
T
0 exp{Br (x) θ 0 }dx
= R00 + R01 ,
where
R00 ≡ n
−1
n Z
X
1
R1
exp[log{fX0 (x)}]fU (Wi − x)
exp[log{fX0 (x)}]fU (Wi − x)dx
!
exp[log{fX0 (x)}]
−R 1
Br (x)dx
exp[log{f
(x)}]dx
X0
0
i=1
0
0
!
8
= n
−1
n Z
X
i=1
1
0
(
and
(S.10)
R01 ≡ n−1
n Z
X
R1
fX0 (x)fU (Wi − x)
0
1
fX0 (x)fU (Wi − x)dx
"
R1
)
− fX0 (x) Br (x)dx
exp{Br (x)T θ 0 }fU (Wi − x)
exp{Br (x)T θ 0 }fU (Wi − x)dx
#
exp{Br (x)T θ 0 }
Br (x)dx
−R 1
T θ }dx
exp{B
(x)
r
0
0
n Z 1
X
exp[log{fX0 (x)}]fU (Wi − x)
−1
−n
R1
i=1 0
0 exp[log{fX0 (x)}]fU (Wi − x)dx
!
exp[log{fX0 (x)}]
−R 1
Br (x)dx.
0 exp[log{fX0 (x)}]dx
Now
0
i=1
kR00 k∞ = sup n
−1
0
n Z
X
1
R1
exp[log{fX0 (x)}]fU (Wi − x)
exp[log{fX0 (x)}]fU (Wi − x)dx
!
exp[log{fX0 (x)}]
−R 1
Brk (x)dx
0 exp[log{fX0 (x)}]dx
p
= Op { hb n−1 log(n)}.
k
(S.11)
0
i=1
0
The last equality holds by Lemma 3 in the supplementary material with
Ci (x) =
R1
0
fX0 (x)fU (Wi − x)
fX0 (x)fU (Wi − x)dx
− R1
0
fX0 (x)
fX0 (x)dx
,
which satisfies E{Ci (x)} = 0 by (S.7), and kCi (x)k2 < ∞ by Condition (C1).
Further,
(S.12)
kR01 k∞
"
= sup
n−1
n Z
X
1
("
R1
exp{Br (x)T θ 0 }fU (Wi − x)
exp{Br (x)T θ 0 }fU (Wi − x)dx
#
exp[log{fX0 (x)}]fU (Wi − x)
exp{Br (x)T θ 0 }
−R 1
− R1
T
exp{fX0 (x)}fU (Wi − x)dx
0 exp{Br (x) θ 0 }dx
!) 0
#
exp[log{fX0 (x)}]
−R 1
Brk (x)dx
exp[log{f
(x)}]dx
X0
0
"
Z 1 ("
n
X
exp{Br (x)T θ 0 }fU (Wi − x)
q
−q
−1
= hb sup n
hb
R1
T
k
0
i=1
0 exp{Br (x) θ 0 }fU (Wi − x)dx
k
i=1
0
0
9
NEW NONPARAMETRIC MEASUREMENT ERROR MODEL ESTIMATOR
−R 1
0
exp{Br (x)T θ 0 }
exp{Br (x)T θ 0 }dx
#
exp[log{fX0 (x)}]
−R 1
exp[log{fX0 (x)}]dx
0
−
!)
R1
0
exp[log{fX0 (x)}]fU (Wi − x)
exp[log{fX0 (x)}]fU (Wi − x)dx
#
Brk (x)dx
= Op (hq+1
b ).
The last equality holds by using Lemma 3 in the supplementary material. with
=
Ci (x)
("
h−q
b
−
R1
0
R1
0
exp{Br (x)T θ 0 }fU (Wi − x)
exp{Br (x)T θ 0 }fU (Wi − x)dx
exp[log{fX0 (x)}]fU (Wi − x)
− R1
0
#
exp{Br (x)T θ 0 }dx
!)
exp[log{fX0 (x)}]
0
−R 1
exp[log{fX0 (x)}]fU (Wi − x)dx
exp{Br (x)T θ 0 }
exp[log{fX0 (x)}]dx
,
which satisfies kCi (x)k2 < ∞. Combining (S.11) and (S.12) we have R0 = R00 {1 +
op (1)} by Conditions (C3) and (C4). Further,
E(RT
00 R00 )
(
)2 d Z
2
θ
1
X
fX0 (ξ)fU (Wi − ξ)
≤ n−1 E sup R 1
− fX0 (ξ)
Brk (x)dx
ξ
f
(x)f
(W
−
x)dx
0
i
U
l=k
0 X0
= Op (n−1 hb ).
where ξ is a point in (0, 1). The last equality holds because
fX0 (ξ)fU (Wi − ξ)
sup R 1
ξ
fX0 (x)fU (Wi − x)dx
0
− fX0 (ξ)
is bounded by Condition (C1). Further,
dθ Z
X
k=1
0
1
Brk (x)dx
2
=
dθ
X
k=1
Dk (tk − tk−r )2
= sup |tk − tk−r |D
k
= Op (hb ),
for some finite constant Dk and D. The second equality holds because the support
of Brk is the interval (tk−r , tk ). The last equality holds because Dk is bounded above
since Brk (x) < 1 for any x. Therefore,
(S.13)
1/2
kR0 k2 = Op (n−1/2 hb ).
10
To analyze R1 (θ 0 ), we first have that for any u such that kuk2 = 1, from Lemma
1 in the supplementary material,
R1
T
T
T 0 Br (x)Br (x) exp{Br (x) θ 0 }fU (Wi − x)dx
u
u
R1
T θ }f (W − x)dx
exp{B
(x)
r
0
i
U
0
2
1/2
T
=
u exp{Br (x)T θ 0 /2}fU (Wi − x)
Br (x) R 1
[ 0 exp{Br (x)T θ 0 }fU (Wi − x)dx]1/2
1/2
u exp{Br (x)T θ 0 /2}fU (Wi − x)
= Op hb R 1
[ 0 exp{Br (x)T θ 0 }fU (Wi − x)dx]1/2
2
2
2
= Op (hb ),
and
R1
T
T
T 0 Br (x)Br (x) exp{Br (x) θ 0 }dx
u
u
R1
T
0 exp{Br (x) θ 0 }dx
=
T
u exp{Br (x)T θ 0 /2}
Br (x) R 1
[ 0 exp{Br (x)T θ 0 }dx]1/2 2
2
T
u exp{Br (x) θ 0 /2}
= O hb R 1
[ 0 exp{Br (x)T θ 0 }dx]1/2 2
= O(hb ).
Similarly
R
T θ }f (W − x)dx 1 BT (x) exp{B (x)T θ }dx
B
(x)
exp{B
(x)
0
r
r
r
0
i
U
r
u
uT 0
R01
R1
T
T
0 exp{Br (x) θ 0 }fU (Wi − x)dx 0 exp{Br (x) θ 0 }dx
"R 1
#2
T
T
0 u Br (x) exp{Br (x) θ 0 }fU (Wi − x)dx
≤
R1
T
0 exp{Br (x) θ 0 }fU (Wi − x)dx
"R 1
#
T (x)u exp{B (x)T θ }dx 2
B
r
0
+ 0 Rr1
,
T
0 exp{Br (x) θ 0 }dx
R1
while
"R 1
0
T
BT
r (x)u exp{Br (x) θ 0 }fU (Wi − x)dx
R1
T
0 exp{Br (x) θ 0 }fU (Wi − x)dx
#2
2
T
BT
r (x)u exp{Br (x) θ 0 }fU (Wi − x)
≤
R1
T
0 exp{Br (x) θ 0 }fU (Wi − x)dx
1
"
u exp{Br (x)T θ 0 }fU (Wi − x)
≤ O p hb R 1
T
0 exp{Br (x) θ 0 }dxfU (Wi − x)
≤
Op (h2b ),
2
1
#2
NEW NONPARAMETRIC MEASUREMENT ERROR MODEL ESTIMATOR
and
"R 1
0
T
BT
r (x)u exp{Br (x) θ 0 }dx
R1
T
0 exp{Br (x) θ 0 }dx
#2
2
T
BT
r (x)u exp{Br (x) θ 0 }
≤
R1
T
0 exp{Br (x) θ 0 }dx
1
"
T
u exp{Br (x) θ 0 }
≤ O hb R 1
T
0 exp{Br (x) θ 0 }dx
1
#2
≤ O(h2b ).
Combining these results, we have
∂ 2 fW (w, θ 0 )/∂θ 0 ∂θ T
0
u
fW (w, θ 0 )
"R 1
T
T
T
0 Br (x)Br (x) exp{Br (x) θ 0 }fU (w − x)dx
= u
R1
T
0 exp{Br (x) θ 0 }fU (w − x)dx
R1 T
R1
T
T
0 Br (x) exp{Br (x) θ 0 }fU (w − x)dx 0 Br (x) exp{Br (x) θ 0 }dx
−
R1
R1
T
T
0 exp{Br (x) θ 0 }fU (w − x)dx 0 exp{Br (x) θ 0 }dx
R
R1
1 T
T
T
0 Br (x) exp{Br (x) θ 0 }dx 0 Br (x) exp{Br (x) θ 0 }fU (w − x)dx
−
R1
R1
T
T
0 exp{Br (x) θ 0 }dx 0 exp{Br (x) θ 0 }fU (w − x)dx
R1
T
Br (x)BT
r (x) exp{Br (x) θ 0 }dx
− 0
R1
T
0 exp{Br (x) θ 0 }dx
#
R
R1
T θ }dx 1 BT (x) exp{B (x)T θ }dx
B
(x)
exp{B
(x)
r
0
r
r
0
r
0
+2 0
u
R1
{ 0 exp{Br (x)T θ 0 }dx}2
= Op (hb ),
uT
and
E
"
(
T∂
u
2 f (W , θ )/∂θ ∂θ T
i 0
0
W
0
"R 1
fW (Wi , θ 0 )
u
2 #
T
Br (x)BT
r (x) exp{Br (x) θ 0 }fU (Wi − x)dx
R1
T
0 exp{Br (x) θ 0 }fU (Wi − x)dx
R1
R
T θ }f (W − x)dx 1 BT (x) exp{B (x)T θ }dx
B
(x)
exp{B
(x)
r
r
0
i
r
0
U
r
− 0
R1
R01
T
T
0 exp{Br (x) θ 0 }fU (Wi − x)dx 0 exp{Br (x) θ 0 }dx
R1
R
T θ }dx 1 BT (x) exp{B (x)T θ }f (W − x)dx
B
(x)
exp{B
(x)
r
r
0
r
0 U
i
r
− 0
R1
R01
T
T
0 exp{Br (x) θ 0 }dx 0 exp{Br (x) θ 0 }fU (Wi − x)dx
R1
T
Br (x)BT
r (x) exp{Br (x) θ 0 }dx
− 0
R1
T
0 exp{Br (x) θ 0 }dx
# !2
R1
R1 T
T
T
Br (x) exp{Br (x) θ 0 }dx 0 Br (x) exp{Br (x) θ 0 }dx
+2 0
u
R1
{ exp{Br (x)T θ 0 }dx}2
= E
u
T
0
0
11
12
= Op (h2b ).
Now using Bernstein’s inequality, for = hb /log(n),
pr uT R1 (θ 0 )u − uT E{R1 (θ 0 )}u >
−n2 2 /2
≤ exp
2
nE(uT [∂ 2 fW (Wi , θ 0 )/∂θ 0 ∂θ T
0 /fW (Wi , θ 0 ) − E{R1 (θ 0 )}]u) + nhb /3
−n2 /2
≤ exp
2
E[{uT ∂ 2 fW (Wi , θ 0 )/∂θ 0 ∂θ T
0 /fW (Wi , θ 0 )u} ] + hb /3
−n2 /2
→0
≤ exp
2
hb + hb /3
when n → ∞. Thus,
|uT R1 (θ 0 )u − uT E{R1 (θ 0 )}u| = op (hb ).
(S.14)
Further,
(S.15)
E{R1 (θ 0 )}
"
2
#
∂ fW (Wi , θ 0 )/∂θ 0 ∂θ T
∂logfW (Wi , θ 0 ) ⊗2
0
= E
−E
fW (Wi , θ 0 )
∂θ 0
"
#
Z 2
∂ fW (w, θ 0 )/∂θ 0 ∂θ T
∂logfW (Wi , θ 0 ) ⊗2
0
=
fW 0 (w)dw − E
fW (w, θ 0 )
∂θ 0
"
#
∂logfW (Wi , θ 0 ) ⊗2
q−1
,
= Op (hb ) − E
∂θ 0
where fW 0 (w) is the true density of W . The last equality holds by Condition (C5)
so that
Z
∂ 2 fW (w, θ 0 )/∂θ 0 ∂θ T
0
fW 0 (w)dw
fW (w, θ 0 )
Z 2
∂ fW (w, θ 0 )/∂θ 0 ∂θ T
0
fW (w, θ 0 )dw
−
fW (w, θ 0 )
R
and the fact that fW (w, θ)dw = 1 so
Z
∂ 2 fW (w, θ 0 )/∂θ 0 ∂θ T
0
fW (w, θ 0 )
R
∂ 2 fW (w, θ 0 )dw
=
∂θ 0 ∂θ T
0
= 0.
2
= Op (hq−1
b )
fW (w, θ 0 )dw
NEW NONPARAMETRIC MEASUREMENT ERROR MODEL ESTIMATOR
Now
13
"
#
∂logfW (Wi , θ) ⊗2
E
∂θ
"R 1
T
0 exp{Br (x) θ}fU (Wi − x)Br (x)dx
= E
R1
T
0 exp{Br (x) θ}fU (Wi − x)dx
#⊗2
R1
T
exp{Br (x) θ}Br (x)dx
.
− 0 R1
T θ}dx
exp{B
(x)
r
0
(S.16)
First,
" R
#⊗2
R1
1
T
T
exp{Br (x) θ}Br (x)dx
exp{Br (x) θ}fU (Wi − x)Br (x)dx
6= 0
− 0 R1
E 0 R1
T
T θ}dx
exp{B
(x)
θ}f
(W
−
x)dx
exp{B
(x)
r
i
r
U
0
0
by Condition (C6). Also, it is easy to see that,
" R
#
R1
1
T θ}f (W − x)B (x)dx
T θ}B (x)dx ⊗2
exp{B
(x)
exp{B
(x)
r
i
r
r
r
U
E 0 R1
− 0 R1
T θ}f (W − x)dx
T
exp{B
(x)
r
i
U
0
0 exp{Br (x) θ}dx
is C(g) defined in Section S.1 in the supplementary material with
gi (x) = R 1
0
exp{Br (x)T θ}fU (Wi − x)
exp{Br (x)T θ}fU (Wi − x)dx
− R1
0
exp{Br (x)T θ}
exp{Br (x)T θ}dx
.
First note that Γ(g) is a strictly positive definite operator on B. We prove this
by contradiction. Note that Γ(g) is obviously positive definite. Assume Γ(g) is not
strictlyy positive definite, then it has a zero eigenvalue when operating on B, hence
the matix
"R 1
exp{Br (x)T θ}fU (Wi − x)Br (x)dx
R1
T
0 exp{Br (x) θ}fU (Wi − x)dx
#⊗2
R1
T
exp{Br (x) θ}Br (x)dx
− 0 R1
T θ}dx
exp{B
(x)
r
0
< Γ(g)Br , Br > = E
0
has a zero eigenvalue. By (S.15) and (S.16), this implies E{R1 (θ 0 )} is not be invertible, which contradicts with Condition (C6) that the score function has unique
solution around the truth. Hence by using Lemma 2 in the supplementary material,
for u such that kuk2 = 1, we have that there exist constants C > c > 0, such that
"R 1
T
T
0 exp{Br (x) θ}fU (Wi − x)Br (x)dx
(S.17)
chb ≤ u E
R1
T
0 exp{Br (x) θ}fU (Wi − x)dx
14
−
R1
0
exp{Br (x)T θ}Br (x)dx
R1
T
0 exp{Br (x) θ}dx
#⊗2
u ≤ Chb .
Combining the results in (S.14), (S.15), (S.16) and (S.17), there exist constants
0 < c1 < C1 , such that
−C1 hb ≤ uT R1 (θ 0 )u ≤ −c1 hb
in probability, and in turn
−1
−1 −1
T
−1
−c−1
1 hb ≤ u R1 (θ 0 ) u ≤ −C1 hb
(S.18)
b 0 k2 = Op (kR1 (θ 0 )−1 R0 k2 ) ≤ Op {kR1 (θ 0 )−1 k2 kR0 k2 } =
in probability. Therefore, red kθ−θ
b 0 )| ≤
Op {(nhb )−1/2 }. Therefore, for any vector a with kak2 = Op (1) we have |aT (θ−θ
b − θ 0 )k2 = Op {(nhb )−1/2 }.
kak2 k(θ
Further, expanding the second derivative, we can write
R1 (θ 0 )
R1
exp[log{fX0 (x)}]fU (Wi − x)Br (x)Br (x)T dx
R1
0 exp[log{fX0 (x)}]fU (Wi − x)dx
hR
i⊗2
1
exp[log{f
(x)}]f
(W
−
x)B
(x)dx
i
r
X0
U
0
−
hR
i2
1
exp[log{f
(x)}]f
(W
−
x)dx
i
X0
U
0
R1
exp[log{fX0 (x)}]Br (x)Br (x)T dx
− 0
R1
0 exp[log{fX0 (x)}]dx
hR
i⊗2
1
exp[log{f
(x)}]B
(x)dx
r
X0
0
+
{1 + op (1)}
hR
i2
1
exp[log{f
(x)}]dx
X0
0
"R 1
fX0 (x)fU (Wi − x)Br (x)Br (x)T dx
= E 0
R1
0 fX0 (x)fU (Wi − x)dx
nR
o⊗2
1
f
(x)f
(W
−
x)B
(x)dx
i
r
X0
U
0
−
nR
o2
1
f
(x)f
(W
−
x)dx
i
X0
U
0
Z 1
⊗2 #
Z 1
T
{1 + op (1)}.
−
fX0 (x)Br (x)Br (x) dx +
fX0 (x)Br (x)dx
0
= E
0
0
Hence we can write
b − θ0
θ
NEW NONPARAMETRIC MEASUREMENT ERROR MODEL ESTIMATOR
=
15
"R 1
fX0 (x)fU (Wi − x)Br (x)Br (x)T dx
R1
0 fX0 (x)fU (Wi − x)dx
nR
o⊗2
1
f
(x)f
(W
−
x)B
(x)dx
i
r
U
0 X0
−
nR
o2
1
f
(x)f
(W
−
x)dx
i
U
0 X0
Z 1
−
fX0 (x)Br (x)Br (x)T dx
0
Z
⊗2 #!−1
E
0
1
+
fX0 (x)Br (x)dx
0
R00 {1 + op (1)}.
This proves the results.
S.5. Proof of Proposition 3. First note that by the definition of a(X, m),
we have
E[S∗β (Wi , Yi , m) − E ∗ {a(X, m) | Yi , Wi , m} | X, m] = 0.
This leads to
(S.19)
E[S∗β (Wi , Yi , β 0 ) − E ∗ {a(X, β 0 ) | Yi , Wi , β 0 }] = op (1)
element-wise by Condition (D4). By Condition (D5), E[S∗β (Wi , Yi , β)−E ∗ {a(X, β) |
Yi , Wi , β}] is invertible near β ∗ and its inverse function is a one-to-one function with
bounded first derivative in the neighborhood of zero. Therefore kβ ∗ − β 0 k2 = op (1).
On the other hand, since
(S.20)
n−1
n
X
i=1
we obtain
b − E ∗ {a(X, β)
b | Yi , Wi , β}
b = 0,
S∗β (Wi , Yi , β)
b − E ∗ {a(X, β)
b | Yi , Wi , β}]
b = op (1).
E[S∗β (Wi , Yi , β)
b − β ∗ k2 = op (1).
Using the same argument regarding β 0 in (S.19), we obtain kβ
b − β k2 = op (1).
Thus, kβ
0
S.6. Proof of Proposition 4. By the Taylor expansion, we have
0 =
n
X
i=1
b − E ∗ {a(X, β)
b | Yi , Wi , β}
b
S∗β (Wi , Yi , β)
16
b − β ),
= T0 + T1 (β ∗ )(β
0
(S.21)
b , and for any β
where β ∗ is a point on the line connecting β 0 and β
n
X
∂[S∗β (Wi , Yi , β) − E ∗ {a(X, β) | Yi , Wi , β}]
T1 (β) = n−1
∂β T
i=1
,
and
T0 = n−1
n
X
i=1
S∗β (Wi , Yi , β 0 ) − E ∗ {a(X, β 0 ) | Yi , Wi , β 0 }
= T00 + T01 ,
where
T00 = n−1
n
X
S∗β (Wi , Yi , m) − E ∗ {a(X, m) | Yi , Wi , m},
i=1
and
T01 = n−1
"
n
X
i=1
− n
−1
S∗β (Wi , Yi , β 0 ) − E ∗ {a(X, β 0 ) | Yi , Wi , β 0 }
n
X
S∗β (Wi , Yi , m)
i=1
∗
#
− E {a(X, m) | Yi , Wi , m} .
By the definition of S∗β (Wi , Yi , m), we can write
kT00 k∞ = sup n−1
k
n Z
X
i=1
1
0
"
−R 1
0
∗ (x)
f0 {Yi − m(x)}fU (Wi − x)fX
∗ (x)dµ(x)
f {Yi − m(x)}fU (Wi − x)fX
−P (x, Wi , Yi , m) Brk (x)dµ(x) .
(S.22)
Now define
"
C1i (x) = − R 1
0
∗ (x)
f0 {Yi − m(x)}fU (Wi − x)fX
∗ (x)dµ(x)
f {Yi − m(x)}fU (Wi − x)fX
#
− P (x, Wi , Yi , m) ,
by the Bernstein’s inequality in Bosq (1998), we have
1/n
n Z
X
i=1
"
= Op
0
n
X
i=1
1
Brk (x)C1i (x)dx − E
E 1/n
Z
0
Z
1
Brk (x)C1i (x)dx
0
1
Brk (x)C1i (x)dx
2
#1/2
logn
NEW NONPARAMETRIC MEASUREMENT ERROR MODEL ESTIMATOR
= Op {
(S.23)
17
p
hb n−1 log(n)}.
The last equality holds from Lemma 2 in the supplementary material, by choosing
gi (x) = C1i (x) and setting u with uk = 1 and ul = 0, for l 6= k. Now
E[S∗β (Wi , Yi , m) − E ∗ {a(X, m) | Yi , Wi , m}]
#
!
Z 1"
∗ (x)
f0 {Y − m(x)}fU (W − x)fX
− P (x, W, Y, m) Br (x)dx
= E
−R 1
∗ (x)dx
f
{Y
−
m(x)}f
(W
−
x)f
0
U
X
0
= 0,
nR
1
which further implies E
0
o
Brk (x)C1i (x)dx = 0 for all k. Plug it into (S.23), and
combine with (S.22), we obtain
p
kT00 k∞ = Op { hb n−1 log(n)}.
(S.24)
Further,
kT01 k∞
(S.25)
= sup n−1
k
n Z
X
1
0
i=1
"
−R 1
0
"
∗ (x)
f0 {Yi − Br (x)T β 0 }fU (Wi − x)fX
∗ (x)dµ(x)
f {Yi − Br (x)T β 0 }fU (Wi − x)fX
−P (x, Wi , Yi , β 0 )] − − R 1
0
∗ (x)
f0 {Yi − m(x)}fU (Wi − x)fX
∗ (x)dµ(x)
f {Yi − m(x)}fU (Wi − x)fX
−P (x, Wi , Yi , m)]) Brk (x)dµ(x)
= sup n
−1
k
n
X
∂Sm (Yi , Wi , m)
i=1
= Op (1) sup
k
=
∂m
Op (hq+1
b ),
Z
0
1
Brk (x)dx
(Brk , BT
r β − m) {1 + op (1)}
(
)
sup |Br (x)T β − m(x)| {1 + op (1)}
x∈[0,1]
The second equality holds by the Taylor expansion with respect to Br (·)T β. The
third equality holds by Condition (D2). The last equality holds by Condition (D4)
and the fact that the support of Brk is the interval (tk−r , tk ). Combining (S.24) and
(S.25) we have T0 = T00 {1 + op (1)} by Conditions (C3) and (C4). Further,
E(TT
00 T00 )
18
dβ Z 1 "
X
∗ (x)
f 0 {Yi − m(x)}fU (Wi − x)fX
−R 1
= n−1 E
∗
l=1 0
0 f {Yi − m(x)}fU (Wi − x)fX (x)dµ(x)
o
−P (x, Wi , Yi , m)] Brl (x)dx)2
≤ n
=
−1
G1
dβ Z
X
l=1
−1
Op (n hb ),
1
Brl (x)dx
0
2
for some constant G1 . The second equality holds by Condition (D2) and the fact
that,
dβ Z
X
X
l=1
Brl (x)dx
0
l=1
dβ
=
1
2
D1l (tl − tl−r )2
≤ sup |tl − tl−r |D1
l
= Op (hb ),
for some finite constant D1l and D1 . The second equality holds because the support
of Brl is the interval (tl−r , tl ). The last equality holds because D1l are bounded
above since Brl (x) < 1 for any x. Therefore,
1/2
kT0 k2 = Op (n−1/2 hb ).
(S.26)
Further for arbitrary vector u with kuk2 = 1, we can write T1 (β ∗ ) as
dβ
dβ
n
∗)
X
X
X
∂S
(Y
,
W
,
m
m i
i
|uT T1 (β ∗ )u| = n−1
uk Brk ,
uk Brk
∂m∗
i=1
k=1
k=1
dβ
= {c + op (1)}k
X
k=1
uk Brk k22
for a positive constant c, where m∗ (·) = Br (·)T β ∗ . The last equality holds by ConP
dition (D2) and the fact that n−1 ni=1 ∂Sm (Yi , Wi , m∗ )/∂m∗ is a non-singular operator on B. To show the non-singularity, we prove by contradiction. Recall that
Sm (Y, W, m∗ )(s)
NEW NONPARAMETRIC MEASUREMENT ERROR MODEL ESTIMATOR
=
Z
0
and
1
"
∗ (x)dµ(x)
f0 {Y − m∗ (x)}fU (W − x)fX
−R 1
0
∗ (x)dµ(x)
f {Y − m∗ (x)}fU (W − x)fX
∗
19
#
− P (x, Y, W, m ) s(x)dx.
S∗β (Wi , Yi , β ∗ ) − E ∗ {a(X, β ∗ )|Yi , Wi , β ∗ }
"
#
Z 1
∗ (x)dµ(x)
f0 {Y − m∗ (x)}fU (W − x)fX
∗
Br (x) − R 1
=
− P (x, Y, W, m ) dx.
∗ (x)}f (W − x)f ∗ (x)dµ(x)
f
{Y
−
m
0
U
X
0
Suppose E{∂Sm (Yi , Wi , m∗ )/∂m∗ } is singular on B, then
)
(
∂S∗β (Wi , Yi , β ∗ ) − E ∗ {a(X, β ∗ )|Yi , Wi , β ∗ }
E
∂β ∗ T
#
(Z
"
1
∗ (x)dµ(x)
f0 {Y − m∗ (x)}fU (W − x)fX
∂
− P (x, Y, W, m∗ )
−R 1
= E
Br (x)
∗ (x)}f (W − x)f ∗ (x)dµ(x)
∂m∗ (x)
f
{Y
−
m
0
U
X
0
Br (x)T dx
is a singular matrix. However by Condition (D5) we know that the score function
E[S∗β (Wi , Yi , β) − E ∗ {a(X, β)|Yi , Wi , β}]
has a unique solution around β 0 , and hence
(
)
∂S∗β (Wi , Yi , β ∗ ) − E ∗ {a(X, β ∗ )|Yi , Wi , β ∗ }
E
∂β ∗ T
cannot be singular. Therefore, E{∂Sm (Yi , Wi , m∗ )/∂m} is indeed a non-singular operator on B.. Further,
P
n−1 ni=1 {∂Sm (Yi , Wi , m∗ )/∂m∗ } is a consistent estimator of E{∂Sm (Yi , Wi , m∗ )/∂m∗ },
and hence
P
n−1 ni=1 {∂Sm (Yi , Wi , m∗ )/∂m∗ } is also a non-singular operator on B for sufficiently large n. Further by using Lemme 1 in the supplementary material, we have
Pdβ
1/2
1/2
g3 hb = Dr ku0 k2 ≤ k k=1
uk Brk k2 ≤ ku0 k2 = G3 hb , where u0 = {uk {(tk −
tk−r )/r}1/2 , k = 1, . . . , dβ }T and g3 , G3 are finite constants. These together with
Lemma 2 imply
(S.27)
c9 hb ≤ uT T1 (β ∗ )u ≤ C9 hb ,
and
(S.28)
−1 −1
T
∗ −1
C9−1 h−1
b ≤ u T1 (β ) u ≤ c9 hb ,
20
in probability. Combining with (S.21), (S.26), we have
b − β k2 ≤ kT1 (β ∗ )−1 k2 kT0 k2 = Op {(nhb )−1/2 }.
kβ
0
By the consistency of Br (x)T β ∗ to m(x), we have
T1 (β ∗ )
= E
∂[S∗β (Wi , Yi , β) − E ∗ {a(X, β) | Yi , Wi , β}]
∂β T
Br (·)T β=m(·)
!
{1 + op (1)}.
Therefore, we can write
b −β
β
0
= −E
∂[S∗β (Wi , Yi , β) − E ∗ {a(X, β) | Yi , Wi , β}]
∂β T
Br (·)T β=m(·)
!−1
×T00 {1 + op (1)}.
This proves the results.
References.
Bosq, D. (1998). Nonparametric statistics for stochastic processes : estimation and prediction D.
Bosq. Lecture notes in statistics: 110. Springer, New York.
Cardot, H., Ferraty, F. and Sarda, P. (1999). Functional linear model. Statistics & Probability
Letters 45 11–22.
Cardot, H., Ferraty, F. and Sarda, P. (2003). Spline estimators for the functional linear model.
Statistica Sinica 13 571–592.
Chatelin, F. (1983). Spectral Approximation of Linear Operators. SIAM, New York.
De Boor, C. (1978). A practical guide to splines 27. Springer-Verlag, New York.
DeVore, R. A. and Lorentz, G. G. (1993). Constructive approximation 303. Springer Science
& Business Media, New York.
Department of Statistics
Department of Statistics
University of Hongkong
Penn State University
Pok Fu Lam, Hongkong
University Park, PA 16802
E-mail: feijiang@hku.hk
E-mail: yzm63@psu.edu
NEW NONPARAMETRIC MEASUREMENT ERROR MODEL ESTIMATOR
21
4
3
f
2
1
0
0
1
f
2
3
4
Fig S.1: B-spline MLE density estimation (left) and deconvolution estimation (right)
from 200 simulations: The solid lines represent the true functions and the dash lines
represent the estimated functions and their 90% confidence bands. The first row to
third row are the results for model II (a)–(c) respectively. Sample size 1000.
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.8
1.0
0.6
0.8
1.0
0.6
0.8
1.0
4
3
f
0
1
2
3
f
2
1
0
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
3
f
1
0
0
1
2
2
3
4
x
4
x
f
0.6
x
4
x
0.0
0.2
0.4
0.6
x
0.8
1.0
0.0
0.2
0.4
x
22
4
3
f
2
1
0
0
1
f
2
3
4
Fig S.2: B-spline MLE density estimation (left) and deconvolution estimation (right)
from 200 simulations: The solid lines represent the true functions and the dash lines
represent the estimated functions and their 90% confidence bands. The first row to
third row are the results for model II (a)–(c) respectively. Sample size 2000.
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.8
1.0
0.6
0.8
1.0
0.6
0.8
1.0
4
3
f
0
1
2
3
f
2
1
0
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
3
f
1
0
0
1
2
2
3
4
x
4
x
f
0.6
x
4
x
0.0
0.2
0.4
0.6
x
0.8
1.0
0.0
0.2
0.4
x
NEW NONPARAMETRIC MEASUREMENT ERROR MODEL ESTIMATOR
23
3
m
−2
−3
−2
−1
−1
0
0
m
1
1
2
2
3
Fig S.3: B-spline semiparametric regression estimation (left) and deconvolution estimation (right) from 200 simulations: The solid lines represent the true functions and
the dash lines represent the estimated functions and their 90% confidence bands.
The first row to third row are the results for model II (a)–(c), respectively. Sample
size 1000.
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
0.6
0.8
1.0
0.6
0.8
1.0
m
−2
−3
−2
−1
−1
0
0
m
1
1
2
2
3
x
3
x
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
m
0
0
−1
−1
−2
−2
−3
m
1
1
2
2
3
x
3
x
0.0
0.2
0.4
0.6
x
0.8
1.0
0.0
0.2
0.4
x
24
3
m
−2
−3
−2
−1
−1
0
0
m
1
1
2
2
3
Fig S.4: B-spline semiparametric regression estimation (left) and deconvolution estimation (right) from 200 simulations: The solid lines represent the true functions and
the dash lines represent the estimated functions and their 90% confidence bands.
The first row to third row are the results for model II (a)–(c), respectively. Sample
size 2000.
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
0.6
0.8
1.0
0.6
0.8
1.0
m
−2
−3
−2
−1
−1
0
0
m
1
1
2
2
3
x
3
x
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
m
0
0
−1
−1
−2
−2
−3
m
1
1
2
2
3
x
3
x
0.0
0.2
0.4
0.6
x
0.8
1.0
0.0
0.2
0.4
x
| 10math.ST |
Efficient Large-Scale Multi-Modal Classification
Douwe Kiela, Edouard Grave, Armand Joulin and Tomas Mikolov
arXiv:1802.02892v1 [cs.CL] 6 Feb 2018
Facebook AI Research
{dkiela,egrave,ajoulin,tmikolov}@fb.com
Abstract
While the incipient internet was largely text-based, the modern digital world is becoming increasingly multi-modal. Here,
we examine multi-modal classification where one modality
is discrete, e.g. text, and the other is continuous, e.g. visual
representations transferred from a convolutional neural network. In particular, we focus on scenarios where we have to
be able to classify large quantities of data quickly. We investigate various methods for performing multi-modal fusion and
analyze their trade-offs in terms of classification accuracy and
computational efficiency. Our findings indicate that the inclusion of continuous information improves performance over
text-only on a range of multi-modal classification tasks, even
with simple fusion methods. In addition, we experiment with
discretizing the continuous features in order to speed up and
simplify the fusion process even further. Our results show that
fusion with discretized features outperforms text-only classification, at a fraction of the computational cost of full multimodal fusion, with the additional benefit of improved interpretability.
Text classification is one of the core problems in machine
learning and natural language processing (Borko and Bernick 1963; Sebastiani 2002). It plays a crucial role in important tasks ranging from document retrieval and categorization to sentiment and topic classification (Deerwester et al.
1990; Joachims 1998; Pang and Lee 2008). However, while
the incipient Web was largely text-based, the recent decade
has seen a surge in multi-modal content: billions of images
and videos are posted and shared online every single day.
That is, text is either replaced as the dominant modality, as
is the case with Instagram posts or YouTube videos, or it is
augmented with non-textual content, as with most of today’s
web pages. This makes multi-modal classification an important problem.
Here, we examine the task of multi-modal classification
using neural networks. We are primarily interested in two
questions: what is the best way to combine (i.e., fuse) data
from different modalities, and how can we do so in the most
efficient manner? We examine various efficient multi-modal
fusion methods and investigate ways to speed up the fusion
process. In particular, we explore discretizing the continuous
features, which leads to much faster training and requires
Copyright c 2018, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
less storage, yet is still able to benefit from the inclusion
of multi-modal information. To the best of our knowledge,
this work constitutes the first attempt to examine the accuracy/speed trade-off in multi-modal classification; and the
first to directly show the value of discretized features in this
particular task.
If current trends continue, the Web will become increasingly multi-modal, making the question of multi-modal classification ever more pertinent. At the same time, as the Web
keeps growing, we have to be able to efficiently handle ever
larger quantities of data, making it important to focus on
machine learning methods that can be applied to large-scale
scenarios. This work aims to examine these two questions
together.
Our contributions are as follows. First, we compare various multi-modal fusion methods, examine their trade-offs,
and show that simpler models are often desirable. Second,
we experiment with discretizing continuous features in order to speed up and simplify the fusion process even further.
Third, we examine learned representations for discretized
features and show that they yield interpretability as a beneficial side effect. The work reported here constitutes a solid
and scalable baseline for other approaches to follow; our investigation of discretized features shows how multi-modal
classification does not necessarily imply a large performance
penalty and is feasible in large-scale scenarios.
Related work
Text classification. Neural network-based methods have
become increasingly popular for text classification (Socher
et al. 2011; Wang and Manning 2012). Recent work has used
neural networks for text classification either at a sentence
(Kim 2014; Hill, Cho, and Korhonen 2016) or full document
(Le and Mikolov 2014; Baker, Kiela, and Korhonen 2016;
Joulin et al. 2016) level. Many core NLP tasks are essentially text classification, from tweets (Sriram et al. 2010)
to reviews to spam. Even though there has been extensive
work on feature engineering for text classification (Chen et
al. 2009), modern approaches often make use of word embeddings (Mikolov et al. 2013) or sentence representations
(Kiros et al. 2015) learned from a large corpus in an unsupervised fashion.
Fusion strategies. Multi-modal fusion, or the integration
of input from various modalities, is an important topic in
the field of multimedia analysis (Wu et al. 2004; Atrey et al.
2010). The question of fusion has been explored in a variety
of tasks, from audio-visual speech recognition (Potamianos
et al. 2003) to multi-sensor management (Zhao et al. 2003)
and face recognition (Xiong and Svensson 2002). Much of
this research has focused on the combination of two or more
continuous modalities. Here, we are specifically interested in
the fusion of discrete textual input with another, continuous,
modality.
Multi-modal NLP. The usage of non-textual information
in natural language processing (Mooney 2008) has become
increasingly popular. On the one hand, there has been a
lot of interest in cross-modal applications, such as image annotation (Weston, Bengio, and Usunier 2011), image captioning (Bernardi et al. 2016), mapping images to
text or vice versa (Frome et al. 2013; Socher et al. 2013;
Lazaridou, Bruni, and Baroni 2014) and visual question
answering (Antol et al. 2015; Fukui et al. 2016). On the
other hand, multi-modal fusion has been extensively explored in the context of grounded representation learning for lexical semantics (Bruni, Tran, and Baroni 2014;
Kiela and Bottou 2014; Lazaridou, Pham, and Baroni 2015).
While much of this work has focused on vision (Baroni
2016), other perceptual modalities modalities (Lopopolo
and van Miltenburg 2015; Kiela and Clark 2015; Kiela, Bulat, and Clark 2015) have also been explored, as well as
robotics (Mei, Bansal, and Walter 2016), videos (Regneri et
al. 2013) and games (Branavan, Silver, and Barzilay 2012;
Narasimhan, Kulkarni, and Barzilay 2015). This work is
similar in spirit to (Bruni, Tran, and Baroni 2014), in that
we explore fusion techniques. However, similarly to (Lazaridou, Pham, and Baroni 2015), we learn how to integrate the
multi-modal inputs, and use transferred representations as in
(Kiela and Bottou 2014).
Multi-modal deep learning. Our work relates to previous work on integrating information from multiple modalities in neural networks (Ngiam et al. 2011; Srivastava
and Salakhutdinov 2012; Kiros, Salakhutdinov, and Zemel
2014). Here, we enhance a well-known neural network architecture for efficient text classification (Joulin et al. 2016)
with the ability to include continuous information, and explore methods for combining multi-modal features. The
works of (Arevalo et al. 2017) and (Fukui et al. 2016),
explore complex gating mechanisms and compact bilinear
pooling as multi-modal fusion methods. In order to obtain
visual representations, we transfer continuous features from
neural networks trained on other tasks (in this case ImageNet), as has been shown to work well for a wide variety
of tasks (Oquab et al. 2014; Razavian et al. 2014).
Evaluation
Surprisingly, there are not many large-scale multi-modal
classification datasets available. We evaluate on three
datasets that are large enough to examine accuracy/speed
trade-offs in a meaningful way. Two of our datasets
(Food101 and MM-IMDB) are medium-sized; while the
third dataset (FlickrTag) is very large by today’s standards.
The quantitative properties of the respective datasets are
shown in Table 1 and they are described in more detail in
what follows.
Food101
The UPMC Food101 dataset (Wang et al. 2015) contains
web pages with textual recipe descriptions for 101 food labels automatically retrieved online. Each page was matched
with a single image, where the images were obtained by
querying Google Image Search for the given category. Examples of food labels are Filet Mignon, Pad Thai, Breakfast
Burrito and Spaghetti Bolognese. The web pages were processed with html2text1 to obtain the raw text.
MM-IMDB
The recently introduced MM-IMDB dataset (Arevalo et al.
2017) contains movie plot outlines and movie posters. The
objective is to classify the movie by genre. This is a multilabel prediction problem, i.e., one movie can have multiple
genres. The dataset was specifically introduced to address
the lack of multi-modal classification datasets.
FlickrTag and FlickrTag-1
We use the FlickrTag dataset based on the massive
YFCC100M Flickr dataset of (Thomee et al. 2016) that
was used in (Joulin et al. 2016). The dataset consists of
Flickr photographs together (in most, but not all cases) with
short user-provided captions. The objective is to predict the
user-provided tags that belong to the photograph. This is a
very large-scale dataset, so we perform the multi-modal fusion operator and speed-versus-accuracy studies on a subset (specifically, the first shard, which corresponds to onetenth of the full dataset) for those studies, which we denote
FlickrTag-1. We show that the inclusion of discretized features yields classification accuracy improvements with respect to text on the whole dataset.
Approach
As a starting point, we take the highly efficient text classification approach of FastText (Joulin et al. 2016). To ensure
a fair comparison, we enhance that model with the capability to handle continuous or discretized features. Specifically, we use 2048-dimensional continuous features that
were obtained by transferring the pre-softmax layer of a
152-layer ResNet (He et al. 2016) trained on the ImageNet
classification task. In the case of the large-scale FlickrTag
datasets, we use ResNet-34 features (of 512 dimensions). It
has been shown that convolutional network features can be
transferred successfully to a variety of tasks (Razavian et al.
2014) and we take the same approach here. We explore a
variety of models and experiment with discretization.
The scenario of multi-modal classification certainly admits, or even invites, highly sophisticated models. In our
1
https://pypi.python.org/pypi/html2text
Dataset
Food101
MM-IMDB
FlickrTag
FlickrTag-1
#Train
#Words
#Valid
#Words
#Test
#Words
58,131
15,552
70,243,104
7,166,110
98,365,392
2,564,734
1,134,118,808
92,651,036
6,452
2608
656,687
48,048
10,893,597
425,863
10,100,945
682,663
21,519
7799
621,444
48,471
36,955,182
1,266,681
9,913,566
672,900
Table 1: Evaluation datasets with their quantitative properties.
case, however, we also have to take into account efficiency,
so we want to focus on models that are simple and efficient
enough to handle large-scale datasets, while obtaining improved performance over our baselines. We experiment with
a comprehensive set of models, listed below in increasing
order of complexity.
In all cases, given a set of N documents, the objective is
to minimize the negative log likelihood over the classes:
N
1 X
log(sof tmax(o(xn ), yn )),
−
N n=1
(1)
where o is the network’s output, xn is the multi-modal input
and yn is the label.
Baselines
Text The first baseline consists of FastText (Joulin et
al. 2016), a library for highly efficient word representation learning and sentence classification. FastText is trained
asynchronously on multiple CPUs using stochastic gradient descent and a learning rate that linearly decays with the
amount of words. It yields competitive performance with
more sophisticated text classification approaches, while being much more efficient. That is, we ignore the visual signal
altogether and only use textual information, i.e.,
o(xn ) = W U xtn ,
where W and U are weight matrices and xtn is the normalized bag of textual features representation.
Continuous The second baseline consists of training a
classifier only on top of the transferred ResNet features (He
et al. 2016). That is, we ignore the textual information and
only use the visual input, i.e.,
o(xn ) = W V xvn ,
where W and V are weight matrices and xvn consists of the
ResNet features, normalized to unit length.
Continuous Multi-Modal Models
Additive We combine the information from both modalities using component-wise addition, i.e.,
o(xn ) = W (U xtn + V xvn ).
Max-pooling We combine the information from both
modalities using the component-wise maximum, i.e.,
o(xn ) = W max(U xtn , V xvn ).
Gated We allow one modality to “gate” or “attend” over
the other modality, via a sigmoid non-linearity, i.e.,
o(xn ) = W (σ(U xtn ) ∗ V xvn ),
or alternatively,
o(xn ) = W (U xtn ∗ σ(V xvn )).
One can think of this approach as performing attention
from one modality over the other. It is a conceptually similar simplification of multi-modal gated units, introduced in
(Arevalo et al. 2017). The modality to be gated is a hyperparameter (see below).
Bilinear Finally, to fully capture any associations between
the two different modalities, we examine a bilinear model,
i.e.,
o(xn ) = W (U xtn ⊗ V xvn ).
This approach can be thought of as a simpler version of the
more complex multi-modal bilinear pooling introduced by
(Fukui et al. 2016). We also experiment with a method where
we introduce a gating non-linearity into the bilinear model,
which we call Bilinear-Gated.
Discretized Multi-Modal Models
A downside of continuous models is that they require an expensive matrix-vector multiplication V xvi and storing large
matrices of floating point numbers requires a lot of space.
While the ResNet features used in these experiments consist of a relatively small number of components, these can
easily run into the tens of thousands: consider e.g. combinations of SIFT and Fisher vectors used in state-of-the-art
computer vision applications (Perronnin and Larlus 2015).
Hence, we experiment with discretizing the continuous features, where we convert the continuous features to a discrete
sequence of tokens, which can be treated as if they are special tokens, which we normalize separately, and used in the
standard FastText setup. This is a simple, computationally
less intensive solution. Discretized features also obviously
require less storage.
In particular, we investigate product quantization (PQ)
(Jegou, Douze, and Schmid 2011), where we divide the continuous vector into subvectors of equal size, and then perform k-means clustering on each of the subvectors. For each
image, we subsequently determine the closest centroid for
each of its subwords, which is combined with the subvector
index in order to obtain a discretized vector. For example, a
Model
Food101
MM-IMDB
FlickrTag-1
Previous work
(Wang et al. 2015)
(Arevalo et al. 2017)-GMU
(Arevalo et al. 2017)-AVG
85.1
—
—
—
63.0
61.5
—
—
—
Baselines
FastText
Continuous
88.0 ± 0.1
56.7 ± 0.2
58.8 ± 0.1
49.3 ± 0.0
23.0 ± 0.0
12.4 ± 0.0
Continuous
Additive
Max-pooling
Gated
Bilinear
Bilinear-gated
90.4 ± 0.1
90.5 ± 0.1
90.1 ± 0.1
88.1 ± 0.3
90.8 ± 0.1
61.0 ± 0.0
62.2 ± 0.1
61.8 ± 0.1
61.5 ± 0.1
62.3 ± 0.2
26.8 ± 0.0
26.9 ± 0.0
27.7 ± 0.0
27.8 ± 0.0
28.6 ± 0.0
Discretized
PQ (n=4, k=256)
RSPQ (n=4, k=256, r=4)
89.5 ± 0.1
89.8 ± 0.0
60.5 ± 0.1
60.7 ± 0.1
25.6 ± 0.1
26.1 ± 0.1
Table 2: Accuracy (averaged over 5 runs) of continuous and discretized multi-modal models, compared to baselines.
100-dim continuous vector xvi may be divided into ten 10dimensional subvectors si . Let N (si ) denote the index of
the nearest centroid for si . The discretized representation of
v is then given as h(1, N (s1 )), (2, N (s2 )), .., (10, N (s10 ))i.
We include these tokens in the text and treat them as if they
were special tokens, in the standard fastText model, i.e.,
o(xn ) = W (U xtn + αU xdn ).
where xdn are the discretized features and α is a reweighting
hyperparameter. We normalize xtn and xdn independently. As
we can see, the discretized models are closely related to the
additive model, except that they use the same weight matrix
U with the discretized features used as “words” in the text.
While PQ is great for compressing information into a discretized sequence, it does impose hard boundaries on subvectors, which means that overlapping semantic content that
is shared between subvectors may be lost. Hence, we introduce a novel quantization method, called random sample
product quantization (RSPQ), in order to maintain (at least
some) overlapping semantic information. In RSPQ, the process is the same as in PQ, except we perform PQ over r
repetitions of random permutations of xvi . In both cases, we
treat the discretized features as if they are reweighted special
tokens included in the textual data and run standard fastText.
Model complexity
There are various trade-offs at stake between these models.
The additive, max-pooling and gated models are simplest
and result in a hidden layer of the same size as with the normal FastText. The computational complexity of the linear
classifier is thus O(HK), where K is the number of classes
and H is the size of U xt and V xv . The max-pooling and gating models are slightly more complicated than the additive
one, requiring an extra operation. For the bilinear model the
complexity amounts to O(H 2 K). Thus, the bilinear model
is by far the most expensive to compute. The additive model
has the benefit that it does not strictly require a continuous
input at all times.
Hyperparameters and training
In all experiments, the model is tuned on the validation set. We tried the following hyperparameters:
a learning rate in {0.1, 0.25, 0.5, 1.0, 2.0}, a number of epochs in {5, 10, 20}, a reweighting parameter
in {0.01, 0.02, 0.05, 0.1, 0.2, 0.5} and an embedding dimensionality of either 20 or 100. These hyperparameters
were sweeped using grid search and we used a softmax loss.
For other hyperparameters, such as the number of threads in
the parallel optimization and the minimum word count, we
fixed their values to standard values in FastText (4 threads,
minimum count of 1, respectively), since we found that
these did not impact classification accuracy. In the case
of the gated and bilinear-gated models, the modality used
to serve as a gate over the other modality is treated as a
hyperparameter as well.
Results
The results of the comparison may be found in Table 2. We
compare the continuous and discretized multi-modal models
against the text-only FastText model and to the continuous
features-only model. We also include results from (Wang et
al. 2015) on Food-101, where they used TF-IDF features
for text and a deep convolutional neural network features
for images, as well as results from (Arevalo et al. 2017)
for Gated Multimodal Units (GMU) and their AVG Probs
model. GMUs are a substantially more complicated model
architecture than any of our relatively simple fusion methods, so this study is a good test of their capability. We note
that in the case of Food101, our methods work considerably
better than previously reported results. For MM-IMDB, the
continuous multi-modal models perform very close to the
GMU model and outperform the AVG Probs method, while
being simpler and computationally more efficient.
We observe that multi-modal models always outperform
standard FastText and the continuous-only approach, disregarding the particular type of fusion. This shows that the
inclusion of multi-modal information (at least in these types
of classification tasks) always helps and that making use of
Model
Train time (FlickrTag-1)
FastText
0h01m
Additive
Max-pooling
Gated
Bilinear
Bilinear-Gated
0h39m
0h39m
0h40m
1h04m
1h06m
PQ
RSPQ
0h01m
0h02m
Table 3: Training time on FlickrTag-1
Model
FlickrTag
FastText
36.7
PQ (n=4, k=256)
RSPQ (n=4, k=256, r=4)
38.9
39.4
q 0 253
q 0 13
donuts 0.987
doughnuts 0.981
donut 0.980
doughnut 0.979
donuts? 0.939
crème 0.933
ramekins 0.928
brulee 0.925
brûlée 0.916
custards 0.916
q 1 253
oishii 0.905
shoga 0.885
tenkasu 0.884
octopus 0.883
aonori 0.881
Table 5: Examples of nearest neighbors for quantized features in Food101.
one minute. If we scale up to the full FlickrTag dataset, Table 4 shows that the discretized models substantially outperform standard FastText. An increase of 2.7% in accuracy,
as seen from FastText to RSPQ, represents having an additional 16778 test set documents correctly classified using
that model, which is a non-negligible amount.
Interpretability
Table 4: Performance on full FlickrTag.
multi-modal information, where available, will lead to increased performance. FastText outperforms the continuousonly method on all datasets, which indicates that text plays a
big role in these tasks, and that it is relatively more important
than the visual information.
If we examine the continuous multi-modal models, we see
that the bilinear-gated model is the clear winner: it outperforms all other methods on all three tasks. It is however also
the most complicated model, and as a result is less efficient.
We found that placing the gating non-linearity on the text
led to the best performance on Food101 and MM-IMDB,
while placing it on the visual modality led to the best performance on FlickrTag. It is interesting to observe that the more
complicated gated model, as well as the non-gated bilinear
model, do not necessarily outperform the simpler additive
and max-pooling models. In fact, the performance of these
much simpler models is not too far removed from the best
scores. The take-away message appears to be: if you care
more about accuracy, use the bilinear method with gating; if
you care more about speed, use a simple model like the additive or max-pooling one, which have the additional potential
benefit that they do not necessarily require the presence of
continuous information if none is available.
Speed
We can draw inspiration from the fact that the additive model
performed reasonably well: if speed is essential—if necessary at the expense of some accuracy—the discretized models are an obvious choice to further simplify and speed up the
model. Even though they outperform standard FastText by a
large margin, as shown in Table 2, they only come with a
minor performance impact. Table 3 shows the training times
for the various models on the FlickrTag-1 dataset: while the
bilinear models take around one hour to train (recall that
this constitutes only the first shard of the full dataset); the
discretized methods, similarly to FastText, only take around
An interesting side effect of the discretized multi-modal
methods is that they allow us to examine the nearest neighbors of the quantized features: if a particular feature corresponds to something that looks like a donut, for example, then its embedding should be close to words related to
donut. Indeed, as Table 5 shows, we can find clearly identifiable clusters, e.g. for donuts, crème brûlée and certain
types of Japanese food. Interpretability is an important but
often overlooked aspect of classification models: we show
that a simple and efficient method, that outperforms textonly methods by a large margin, yields the additional benefit that it allows for the interpretation of the visual features
that a classifier picks up on—something that is difficult to
achieve with standard convolutional features.
Conclusion & Outlook
The internet is becoming increasingly multi-modal, which
makes the task of multi-modal classification ever more pertinent. In order to be able to handle large quantities of data, we
need efficient models for large-scale multi-modal classification. In this work, we examined these two questions together.
First, we compared various multi-modal fusion methods and
found a bilinear-gated model to achieve the highest accuracy, while the simpler additive and max-pooling models
yielded reasonably high accuracy at higher speed. Second,
we showed that the model can be speeded up even further
by introducing discretized multi-modal features. Lastly, we
showed that this method yields the additional benefit of interpretability, where we can examine what the multi-modal
model picks up on when making its classification decision.
We hope that this work can serve as a useful baseline for
further work in multi-modal classification.
Acknowledgments
We thank the reviewers for their comments and our colleagues at FAIR for their feedback and support.
References
[Antol et al. 2015] Antol, S.; Agrawal, A.; Lu, J.; Mitchell,
M.; Batra, D.; Lawrence Zitnick, C.; and Parikh, D. 2015.
VQA: Visual question answering. In Proceedings of the
IEEE International Conference on Computer Vision, 2425–
2433.
[Arevalo et al. 2017] Arevalo, J.; Solorio, T.; Montes-y
Gómez, M.; and González, F. A. 2017. Gated multimodal
units for information fusion.
[Atrey et al. 2010] Atrey, P. K.; Hossain, M. A.; El Saddik, A.; and Kankanhalli, M. S. 2010. Multimodal fusion for multimedia analysis: a survey. Multimedia systems
16(6):345–379.
[Baker, Kiela, and Korhonen 2016] Baker, S.; Kiela, D.; and
Korhonen, A. 2016. Robust text classification for sparsely
labelled data using multi-level embeddings. In Proceedings
of COLING, 2333–2343.
[Baroni 2016] Baroni, M. 2016. Grounding distributional
semantics in the visual world. Language and Linguistics
Compass 10(1):3–13.
[Bernardi et al. 2016] Bernardi, R.; Cakici, R.; Elliott, D.;
Erdem, A.; Erdem, E.; Ikizler-Cinbis, N.; Keller, F.; Muscat,
A.; and Plank, B. 2016. Automatic description generation
from images: A survey of models, datasets, and evaluation
measures. J. Artif. Intell. Res.(JAIR) 55:409–442.
[Borko and Bernick 1963] Borko, H., and Bernick, M. 1963.
Automatic document classification. Journal of the ACM
10(2):151–162.
[Branavan, Silver, and Barzilay 2012] Branavan, S.; Silver,
D.; and Barzilay, R. 2012. Learning to win by reading
manuals in a monte-carlo framework. Journal of Artificial
Intelligence Research 43:661–704.
[Bruni, Tran, and Baroni 2014] Bruni, E.; Tran, N.-K.; and
Baroni, M. 2014. Multimodal distributional semantics. J.
Artif. Intell. Res.(JAIR) 49(2014):1–47.
[Chen et al. 2009] Chen, J.; Huang, H.; Tian, S.; and Qu, Y.
2009. Feature selection for text classification with naı̈ve
bayes. Expert Systems with Applications 36(3):5432–5435.
[Deerwester et al. 1990] Deerwester, S.; Dumais, S. T.; Furnas, G. W.; Landauer, T. K.; and Harshman, R. 1990. Indexing by latent semantic analysis. Journal of the American
society for information science 41(6):391.
[Frome et al. 2013] Frome, A.; Corrado, G. S.; Shlens, J.;
Bengio, S.; Dean, J.; Mikolov, T.; et al. 2013. Devise: A deep
visual-semantic embedding model. In Advances in neural
information processing systems, 2121–2129.
[Fukui et al. 2016] Fukui, A.; Park, D. H.; Yang, D.;
Rohrbach, A.; Darrell, T.; and Rohrbach, M. 2016. Multimodal compact bilinear pooling for visual question answering and visual grounding. CoRR abs/1606.01847.
[He et al. 2016] He, K.; Zhang, X.; Ren, S.; and Sun, J. 2016.
Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 770–778.
[Hill, Cho, and Korhonen 2016] Hill, F.; Cho, K.; and Korhonen, A.
2016.
Learning distributed representations of sentences from unlabelled data. arXiv preprint
arXiv:1602.03483.
[Jegou, Douze, and Schmid 2011] Jegou, H.; Douze, M.; and
Schmid, C. 2011. Product quantization for nearest neighbor
search. IEEE Transactions on Pattern Analysis and Machine
Intelligence 33(1):117–128.
[Joachims 1998] Joachims, T. 1998. Text categorization with
support vector machines: Learning with many relevant features. In European Conference on Machine Learning, 137–
142.
[Joulin et al. 2016] Joulin, A.; Grave, E.; Bojanowski, P.; and
Mikolov, T. 2016. Bag of tricks for efficient text classification. CoRR abs/1607.01759.
[Kiela and Bottou 2014] Kiela, D., and Bottou, L. 2014.
Learning image embeddings using convolutional neural networks for improved multi-modal semantics. In EMNLP, 36–
45.
[Kiela and Clark 2015] Kiela, D., and Clark, S. 2015. Multiand cross-modal semantics beyond vision: Grounding in auditory perception. In EMNLP, 2461–2470.
[Kiela, Bulat, and Clark 2015] Kiela, D.; Bulat, L.; and
Clark, S. 2015. Grounding semantics in olfactory perception. In ACL (2), 231–236.
[Kim 2014] Kim, Y. 2014. Convolutional neural networks
for sentence classification. In Proceedings of EMNLP.
[Kiros et al. 2015] Kiros, R.; Zhu, Y.; Salakhutdinov, R.;
Zemel, R. S.; Torralba, A.; Urtasun, R.; and Fidler, S. 2015.
Skip-thought vectors. In Proceedings of NIPS.
[Kiros, Salakhutdinov, and Zemel 2014] Kiros,
R.;
Salakhutdinov, R.; and Zemel, R. S. 2014. Multimodal neural language models. In Proceedings of ICML,
volume 14, 595–603.
[Lazaridou, Bruni, and Baroni 2014] Lazaridou, A.; Bruni,
E.; and Baroni, M. 2014. Is this a wampimuk? crossmodal mapping between distributional semantics and the visual world. In ACL (1), 1403–1414.
[Lazaridou, Pham, and Baroni 2015] Lazaridou, A.; Pham,
N. T.; and Baroni, M. 2015. Combining language and vision with a multimodal skip-gram model. arXiv preprint
arXiv:1501.02598.
[Le and Mikolov 2014] Le, Q. V., and Mikolov, T. 2014.
Distributed representations of sentences and documents. In
ICML, volume 14, 1188–1196.
[Lopopolo and van Miltenburg 2015] Lopopolo, A., and van
Miltenburg, E. 2015. Sound-based distributional models.
IWCS 2015 70.
[Mei, Bansal, and Walter 2016] Mei, H.; Bansal, M.; and
Walter, M. R. 2016. Listen, attend, and walk: Neural mapping of navigational instructions to action sequences. In Proceedings of AAAI.
[Mikolov et al. 2013] Mikolov, T.; Sutskever, I.; Chen, K.;
Corrado, G. S.; and Dean, J. 2013. Distributed representations of words and phrases and their compositionality. In
Advances in neural information processing systems, 3111–
3119.
[Mooney 2008] Mooney, R. J. 2008. Learning to connect
language and perception. In Proceedings of AAAI.
[Narasimhan, Kulkarni, and Barzilay 2015] Narasimhan, K.;
Kulkarni, T.; and Barzilay, R. 2015. Language understanding for text-based games using deep reinforcement learning.
In Proceedings of EMNLP.
[Ngiam et al. 2011] Ngiam, J.; Khosla, A.; Kim, M.; Nam,
J.; Lee, H.; and Ng, A. Y. 2011. Multimodal deep learning.
In Proceedings of ICML, 689–696.
[Oquab et al. 2014] Oquab, M.; Bottou, L.; Laptev, I.; and
Sivic, J. 2014. Learning and transferring mid-level image representations using convolutional neural networks. In
Proceedings of the IEEE conference on computer vision and
pattern recognition, 1717–1724.
[Pang and Lee 2008] Pang, B., and Lee, L. 2008. Opinion
mining and sentiment analysis. Foundations and trends in
information retrieval 2(1-2):1–135.
[Perronnin and Larlus 2015] Perronnin, F., and Larlus, D.
2015. Fisher vectors meet neural networks: A hybrid classification architecture. In Proceedings of the IEEE conference
on computer vision and pattern recognition, 3743–3752.
[Potamianos et al. 2003] Potamianos, G.; Neti, C.; Gravier,
G.; Garg, A.; and Senior, A. W. 2003. Recent advances in the
automatic recognition of audiovisual speech. Proceedings of
the IEEE 91(9):1306–1326.
[Razavian et al. 2014] Razavian, A. S.; Azizpour, H.; Sullivan, J.; and Carlsson, S. 2014. Cnn features off-the-shelf:
an astounding baseline for recognition. In Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition Workshops, 806–813.
[Regneri et al. 2013] Regneri, M.; Rohrbach, M.; Wetzel, D.;
Thater, S.; Schiele, B.; and Pinkal, M. 2013. Grounding action descriptions in videos. Transactions of the Association
for Computational Linguistics 1:25–36.
[Sebastiani 2002] Sebastiani, F. 2002. Machine learning
in automated text categorization. ACM Computing Surveys
34(1):1–47.
[Socher et al. 2011] Socher, R.; Pennington, J.; Huang,
E. H.; Ng, A. Y.; and Manning, C. D. 2011. Semisupervised recursive autoencoders for predicting sentiment
distributions. In Proceedings of the conference on empirical
methods in natural language processing, 151–161. Association for Computational Linguistics.
[Socher et al. 2013] Socher, R.; Ganjoo, M.; Manning, C. D.;
and Ng, A. 2013. Zero-shot learning through cross-modal
transfer. In Advances in neural information processing systems, 935–943.
[Sriram et al. 2010] Sriram, B.; Fuhry, D.; Demir, E.; Ferhatosmanoglu, H.; and Demirbas, M. 2010. Short text classification in twitter to improve information filtering. In Proceedings of the 33rd International ACM SIGIR Conference
on Research and Development in Information Retrieval, SIGIR ’10, 841–842. New York, NY, USA: ACM.
[Srivastava and Salakhutdinov 2012] Srivastava, N., and
Salakhutdinov, R. R. 2012. Multimodal learning with deep
boltzmann machines. In Advances in neural information
processing systems, 2222–2230.
[Thomee et al. 2016] Thomee, B.; Shamma, D.; Friedland,
G.; Elizalde, B.; Ni, K.; Poland, D.; Borth, D.; and Li, L.
2016. YFCC100M: The New Data in Multimedia Research.
Communications of the ACM 59(2):64–73.
[Wang and Manning 2012] Wang, S., and Manning, C. D.
2012. Baselines and bigrams: Simple, good sentiment and
topic classification. In Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics: Short
Papers-Volume 2, 90–94. Association for Computational
Linguistics.
[Wang et al. 2015] Wang, X.; Kumar, D.; Thome, N.; Cord,
M.; and Precioso, F. 2015. Recipe recognition with
large multimodal food dataset. In Workshop CEA of the
IEEE International Conference on Multimedia & Exposition
(ICME), 1–6. IEEE.
[Weston, Bengio, and Usunier 2011] Weston, J.; Bengio, S.;
and Usunier, N. 2011. Wsabie: Scaling up to large vocabulary image annotation. In IJCAI, volume 11, 2764–2770.
[Wu et al. 2004] Wu, Y.; Chang, E. Y.; Chang, K. C.-C.; and
Smith, J. R. 2004. Optimal multimodal fusion for multimedia data analysis. In Proceedings of the 12th annual ACM
international conference on Multimedia, 572–579. ACM.
[Xiong and Svensson 2002] Xiong, N., and Svensson, P.
2002. Multi-sensor management for information fusion: issues and approaches. Information fusion 3(2):163–186.
[Zhao et al. 2003] Zhao, W.; Chellappa, R.; Phillips, P. J.;
and Rosenfeld, A. 2003. Face recognition: A literature survey. ACM computing surveys (CSUR) 35(4):399–458.
| 1cs.CV |
Disentangled Variational Auto-Encoder for Semi-supervised Learning
Yang Li∗1 , Quan Pan1 , Suhang Wang3 , Haiyun Peng2 , Tao Yang 1 , and Erik Cambria2
1
arXiv:1709.05047v1 [cs.LG] 15 Sep 2017
2
School of Automation, Northwestern Polytechnical University
School of Computer Science and Engineering, Nanyang Technological University
3
Department of Computer Science and Engineering, Arizona State University
Abstract
for semi-supervised learning, the semi-supervised VAEs are typically composed of three main components: an encoder network qφ (z|x, y), a decoder pθ (x|y, z) and a classifier qφ (y|x).
In the application, the encoder, decoder and classifier can be
implemented using various models, e.g., MLP or CNN networks [19, 34]. Though the classifier plays a vital role in achieving the semi-supervised goal, it introduces extra parameters of
itself to learn. With the limited labeled data, it may not be an
optimal choice to introduce more parameters to VAE for semisupervised learning because it may memorize the limited data
with large quantities of parameters, namely overfiting.
In this paper, we develop a novel approach for semi-supervised
VAE without classifier. Specifically, we propose a new model
called SDVAE, which encodes the input data into disentangled
representation and non-interpretable representation, then the category information is directly utilized to regularize the disentangled representation via equation constraint. To further enhance
the feature learning ability of the proposed VAE, we incorporate
reinforcement learning to relieve the lack of data. The dynamic
framework is capable of dealing with both image and text data
with its corresponding encoder and decoder networks. Extensive Therefore, in this paper, we investigate if we can directly incorexperiments on image and text datasets demonstrate the effective- porate the limited label information to VAE without introducing
ness of the proposed framework.
a classifier so as to achieve the goal of semi-supervised learning and at the same time to reduce the number of parameters to
be learned. In particular, we investigate the following two challenges: (1) Without introducing classifier, how do we incorpo1 Introduction
rate the label information to VAE for semi-supervised learning?
and (2) How can we effectively use the label information for representation learning of VAE? In an attempt to solve these two
The abundant data generated online every day has greatly ad- challenges, we propose a novel semi-supervised learning model
vanced machine learning, data mining and computer vision com- named Semi-supervised Disentangled Variational Auto-Encoder
munities. However, manual labeling of large dataset is very time (SDVAE). SDVAE adopts the VAE with KKT conditions as it
and labor consuming. Sometimes it even requires domain knowl- has better representation learning ability than VAE. Unlike exedge. All the above results the majority of the data with limited isting semi-supervised VAEs that utilize classifiers, SDVAE enlabels. Therefore, semi-supervised learning, which utilizes both codes the input data into disentangled representation and nonlabeled and unlabeled data for model training, is attracting in- interpretable representation, and the category information is dicreasing attention [4, 22, 14, 33]. Existing semi-supervised mod- rectly utilized to regularize the disentangled representation as an
els can be generally categorized into those categories, i.e., dis- equation constraint. As the labeled data is limited, the labeled
criminative model, generative model, graph-based model and the information may not affects the model much. To remedy this,
combined model with those categories [29, 5, 9, 23, 31].
we further change the equation constraints into the reinforcement
Among various semi-supervised models proposed, the semi- learning format, which helps the objective gain the category insupervised generative models based on variational auto-encoder formation heuristics. The inverse auto-regression (IAF) is also
have shown strong performance in image classification [14, 19] applied to improve the latent variable learning. The proposed
and text classification [33]. The effectiveness of VAE for semi- framework is flexible in which it can deal with both image and
supervised learning comes from its efficiency in posterior dis- text data by choosing corresponding encoder and decoder nettribution estimation and its powerful ability in feature extract- works. The main contributions of the paper are:
ing from text data [2] and image data [14, 19]. To adapt VAE
• Propose a novel semi-supervised framework which directly
exploits the label information to regularize disentangled rep-
∗ liyangnpu@mail.nwpu.edu.cn
1
the RCE, namely maxθ,φ Eqφ (z|x) log pθ (x|z) [6]. Then the objective function is changed with the inequation constraints:
resentation with reinforcement learning;
• Extract the disentangle variable for classification and the
non-interpretable variable for the reconstruction from the
data directly; and
max Eqφ (zi |x) log pθ (x|z)
θ
(4)
subject to KL(qφ (z|x)||pθ (z) < )
• Conduct extensive experiments on image and text datasets Using KKT conditions [1], Eq.(4) can be rewritten as follows:
to demonstrate the effectiveness of the proposed SDVAE.
L̂(θ, φ; x, λ) = Eqφ (z|x) log pθ (x|z)
(5)
− λ(KL(qφ (z|x)||pθ (z)) − )
2
Preliminaries
where λ > 0 is the Lagrangian multiplier, which is used to penalize the deviation of the constraint KL(qφ (z|x)||pθ (z)) ≤ .
Given that λ > 0 and > 0, we have
In this section, we introduce preliminaries that will be useful to
understand our model.
L̂(θ, φ; x, λ) ≥ Eqφ (z|x) log pθ (x|z) − λKL(qφ (z|x)||pθ (z))
(6)
If λ = 1, then Eq.(6) reduces to the original variational autoencoder problems that proposed by Kingma [16]. However, if
2.1 Variational Auto-Encoder
0 < λ < 1, then L̂(θ, φ, λ; x) > L(θ, φ, λ; x), which is closer
to the target log pθ (x). This is just the mathematical description
Variational Auto-Encoders (VAEs) have emerged as one of the
of the fact that the more information in the latent variable z, the
most popular deep generative models. One key step of VAE is to
tighter of the lower bound is. Through the KKT condition, a loose
evaluate pθ (x), which can be interpreted as
constraint over the decoder is introduced. Empirical results show
log pθ (x) = KL(qφ (z|x)||pθ (z|x)) + L(θ, φ; x)
(1) that VAE with KKT condition performs better than original VAE.
Thus, in this paper, we use VAE with KKT condition as our basic
where KL(Q||P ) is Kullback-Leibler divergence between two model.
distributions Q and P and L(θ, φ; x) is the evidence lower bound
(ELBO). It is defined as
L(θ, φ; x) = Eqφ (z|x) (− log qφ (z|x) + log pθ (x, z))
2.3
(2)
When there is label information y in the observed data, it is easy
to extend Eq.(6) to include label information as follows [14].
The term qφ (z|x) is to extract latent feature from the observed data x and it is called encoder generally. By minimizing KL divergence, we try to find qφ (z|x) that can approximate the true posterior distribution pθ (z|x). Because
L(θ, φ; x) is non-negative and log p(x) is fixed, then minimizing KL(qφ (z|x)||pθ (z|x) is equivalent to maximizing L(θ, φ; x).
We can rewrite L(θ, φ; x) as
L̂(θ, φ; x, y, λ) ≥ Eqφ (z|x,y) log pθ (x|z, y)
+ λ(log p(y) + log p(z) − log qφ (z|x, y))
(7)
To achieve the semi-supervised learning, [14] introduce a classifier qφ (y|x) to Eq.(7), which results in
X
U (θ, φ, y; x, λ) =
qφ (y|x)(L̂(θ, φ; x, y, λ)) + H(qφ (y|x))
L(θ, φ; x) = Eqφ (z|x) log pθ (x|z) − KL(qφ (z|x)||pθ (z)) (3)
y
where the first term in the RHS of Eq.(3) is the reconstruction
error (RCE), and the second term in the RHS is the KL divergence
between the prior and the posterior (KLD). Those two values play
different roles during the approximation. We will introduce them
in details in the next section.
2.2
Semi-supervised VAE
(8)
Apart from the Eq.(7) and Eq.(8), the classification loss over the
label information Ep(x,y) (log qφ (y|x)) is added into the objective
function when facing with the labeled data. However, in this paper, the discriminative information is added from scratch and an
equation constrained VAE is proposed, in order to highlight the
contribution of labeled data.
VAE with KKT Conditions
In practice, we find that the RCE is usually the main error, while 3 The Proposed Framework
the term of KLD is regarded as the regularization to enforce
(qφ (z|x)) to be close to pθ (z|x), which is relatively small. If
we constrain the KL divergence term into a small component to In this section, we introduce the details of the proposed framegain a tighter lower bound, the goal is transformed to maximize work. Instead of using a classifier to incorporate the label infor2
3.2
mation, we seek to directly use label information to regularize the
latent representation so as to reduce the number of parameters.
SDVAE-I
The first way we consider is the cross entropy between y and v,
i.e.,
U = y log qφ (v|x)
(11)
3.1
Disentangled Representation
where y is the observed label information, qφ (·) is encoder for
the disentangle variable v. This is a popular loss function for
supervised learning and doesn’t introduce any new parameters.
Therefore, we choose this as the loss function for regularizing the
disentangled variable v. We name this method Semi-supervised
Disentangled VAE (SDVAE-I). By adding this loss function to
Eq.(10), the objective function of SDVAE-I is given as:
In order to incorporate the label information to the latent representation, we assume that the latent representation can be divided
into two parts, i.e., the disentangle variable and non-interpretable
variable. The disentangle variable captures the categorical information, which can be used for prediction task. Therefore, we
can use label information to constrain disentangled variable. The
(12)
L̂(θ, φ; x, λ, u) ≥ RE(z,v) − λ(KLz + KLv ) + µU
non-interpretable variable can be vectors comprised of any dimensions that combine other uncertain information from the data. µ is the weight parameter. When there is no labeled data, the
For the simplicity of notation, we use v to denote the disentan- equation condition will be U = 0.
gled variable and z to denote the non-interpretable representation. With v and z, the encoder can be rewritten as qφ (v, z|x).
We further assume that the disentangled variable and the non3.3 SDVAE-II
interpretable variable are independent condition on x, i.e.,
The drawback of the SDVAE-I is obviously, because the training
results depend on the number of the labeled data heavily. However, for semi-supervised learning, there is usually a small size
of the labeled data available. Thus, it is hard for the disentangle
It is a reasonable assumption, because given x, the categorical variable to capture the category information. To remedy this, we
information is only dependent on x and thus v, which captures the got inspired by the idea in [33]. The equation constraints can be
categorical information, is independent of z given x. This means expressed by the reinforcement learning, in which ELBO can be
there are seldom information about the category information in seen as the reward of the equation constraint. The disentangle
z, which is validated in the experiment part.
variable v acts as the agent which decides the output category
Now qφ (z|x) is the encoder for the non-interpretable representa- information. This can be seen in Eq.(14). Finally, a constant
tion, and qφ (v|x) is the encoder for the disentangle representa- number c is added to act as the bias. The parameter µ is changed
into as followed:
tion. Based on those assumptions, Eq.(7) is written as:
qφ (v, z|x) = qφ (v|x)qφ (z|x)
(9)
µ = RE(z,v) − (KLz + KLv ) + c
L̂(θ, φ; x, λ) ≥ Eqφ (z|x),qφ (v|x) log pθ (x|z, v)
+ λ(log p(v) + log p(z) − log qφ (z|x) − log qφ (v|x))
(13)
The partial update of this part is as followed:
(10)
∆φ = v(L + c)∇φ log qφ (v|x)
= RE(z,v) − λ(KLz + KLv )
(14)
Where L denotes the RE(z,v) − (KLz + KLv ).
where RE(z,v) = Eqφ (z|x),qφ (v|x) log pθ (x|z, v), which represents the reconstruction error given the variable (z, v). KLz and
KLv denote the KL(qφ (z|x)||p(z)) and KL(qφ (v|x)||p(v)) respectively. From the above equation, we can see that the categorical information is extracted from the data, i.e., captured in
disentangled variable v. Now if we have partial labels given, we
can directly use the label information to regularize v.
However, those terms only take effect over the labeled data. To
make up for this drawback, another term of log-likelihood expectation on disentangle variable v is added as the information entropy in Eq.(12), and this will be calculated both in labeled data
and the unlabeled data. It helps to reduce the large variance of the
disentangle information. Then the objective function in Eq.(12)
is changed into Eq.(15).
With v capturing the categorical information, there are many
ways to regularize v. Inspired by the work of [6], we add equation
L̂(θ, φ; x, λ, c) ≥ RE(z,v) − λ(KLz + KLv )
(15)
constraints on v over the ELBO, where the equation constraint is
+ y(β1 L + c) log qφ (v|x) + β2 H(qφ (v|x))
to enforce the disentangled representation v to be close to the label information y. In this work, we consider two ways to add the where y is the label information, β1 and β2 are the coefficient
constraint over the ELBO as discussed below.
parameters, and we name this model SDVAE-II.
3
3.4
With Inverse Autoregressive Flow
the same time are from both the labeled data and the unlabeled
data. Furthermore, we assume that these two variables are independent. However, it is not the same in the previous works, they
only extract the latent variable z from the data. When there is
no label information, label variable y infers from the x with the
shared parameters from qφ (z|x) or infers from z directly.
Because the two different latent variables are extracted from
the data directly, to make the posterior inference more flexible
and enhance the ability in disentangle representation in highdimension space, the inverse autoregressive flow (IAF) [35] is
applied in SDVAE-I and SDVAE-II. The chain is initialized with
the output µ0 and δ0 from the encoder. Together with the random sample ε ∼ N (0, I), the latent variable z is calculated as
z0 = µ0 + δ0 ⊗ ε. The way to update IAF chain is the same as
that in the LSTM shown in Eq.(16).
Then, based on different assumptions, there are differences with
the previous works in mathematics. The ELBO with two independent latent variable inferences is written as Eq.(10), and it is
different from that in Eq.(7) who only has one latent variable z
inference. Furthermore, if we ignore the assumption difference,
zt = δt ⊗ zt−1 + µt
(16) when facing with the labeled data in previous works, their objective function is a special case in Eq.(15) when β1 = β2 = 0.
where (δt , µt ) are the outputs of the auto-regression neural networks, whose input is the last latent variable zt−1 , and t is the When the label is missing, previous works apply the marginal
flow length.
posterior inference over the label information which is shown in
3.5
Eq.(8). In this paper, it is the inference for both latent variable
inference over the z and v, and this is shown in Eq.(17).
Training of SDVAE
U (x) = RE(z,v) − λ(KLz + KLv ) + β2 H(qφ (v|x))
(17)
The models can be trained end-to-end using mini-batch with the
ADAM optimizer [13]. The training algorithm is summarized in
Algorithm 1. In Line 1, we initialize the parameters. From Line
4 Experimental Results
3 to Line 5, we sample a mini-batch to encode the input data as
z and v. From Line 6 to Line 10, we apply IAF. We then update
the parameters from Line 11 to Line 13.
In this section, we conduct experiments to validate the effectiveness of the proposed framework. Specifically, we want to anAlgorithm 1 Training algorithm of the proposed models.
swer the following questions: (1) Is the disentangled representa1: Initialize the parameters , φ, θ
tion able to capture the categorical information? (2) Is the non2: repeat
interpretable variable helpful for the data reconstruction? (3) Is
3:
x ← Sample a miniBatch from the datapoints
the proposed framework effective for semi-supervised learning?
4:
← Random sample from the noise distribution
To answer the above questions, we conduct experiments on im5:
z, v ← qφ (z, v|x, )
age and text datasets, respectively.
6:
if IAF then
7:
for t < T do
8:
ẑ ← iaf (z, θ)
4.1 Experiments on Image Datasets
9:
end for
10:
end if
4.1.1 Datasets Description
11:
x̂ ← pθ (x|ẑ, v)
12:
g ← 5θ,φ L̂(θ, φ; x, λ, c)
Calculate the gradients of Eq.(15) for SDVAE-II, and Eq.(12) for For image domain, we choose two widely used benchSDVAE-I.
mark datasets for evaluating the effectiveness of SDVAE, i.e.,
13:
(θ, φ) ←Update with gradients g
MNIST [18] and SVHN [21]. In the MNIST, there are 55,000
14: until model convergence
data samples in the train set and 10,000 data samples in the test
set. In the SVHN, there are 73,257 data samples in the train set,
and 26,032 data samples in the test set. Both datasets contain 10
categories.
3.6 Discussion
The differences between the previous work [14, 33] and our work
4.1.2
will be discussed in this section.
Model Structure
Firstly, the assumptions are different. In this work, we assume For the image data, the encoder is a deep net composed of two
that the disentangle variable v and non-interpretable variable z at convolutional layers followed by two fully connected layers. The
4
Table 1: The classification errors on the MNIST data with part of labeled data
Models
NN
CNN
TSVM
CAE
MTC
AtlasRBF
Semi-VAE(M1)+TSVM
Semi-VAE(M2)
Semi-VAE(M1+M2)
SDVAE-I
SDVAE-I&IAF
SDVAE-II
SDVAE-II&IAF
([14])
600
11.44
7.68
6.16
6.3
5.13
5.72(±0.05)
4.94(±0.13)
2.59(±0.05)
2.75(±0.11)
2.74(±0.06)
2.49(±0.10)
1.97(±0.14)
ECVAE-I
convolutional layers are used to extract features from
the images
while the fully connected layers are used to convert the features to
the non-interpretable variable and the disentangle variable. The
decoder is a network composed of two fully connected layers to
map the latent features back to images. Dropout [26] is applied
to both the encoder and decoder networks.
4.1.3
1000
10.07
6.45
5.38
4.77
3.64
3.68(±0.12)
4.24(±0.07)
3.60(±0.56)
2.40(±0.02)
2.42(±0.08)
2.24(±0.08)
1.96(±0.09)
1.29(±0.11)
3000
6.04
3.35
3.45
3.22
2.57
3.49(±0.04)
3.92(±0.63)
2.18(±0.04)
1.70(±0.09)
1.33(±0.09)
1.58(±0.09)
1.00(±0.05)
SDVAE-I
ECVAE-I&IAF
SDVAE-I&IAF
SDVAE-II
ECVAE-II&IAF
5555555555 5
SDVAE-II&IAF
3333 3
00
0
888888888888
000000000 00
88888888888888888
6
6666 66666 00000
666 2 2222222 2
666666666666666666 666
88888888888888
6
0
6 2 2 22
666 6666
000 0
8
666 666666666 666666
0 00 0
6
6
8
0000000 0 00000 6 222222 2 222222222222222
6
8888888
6
8
4
666666666 66666 6
6666666666 6666 66 6
00 00
8 88 8 8
2
0
2
000 00 0000
0
2
2
0
2
6
6
6
6
2
4
2
2
0
0
6
0
2
2 22
8
0 0
6 66 66666666
4 6 6666666 6666666
4 4
888888 0000000 000
2 2222 2 2 2222
4444444 4444444444444444
6 666
000 00 0
6
8
4 444 44
2222 2 2 22 2
0
2 2222 222 2 2 22
44444444444444444444444 444
4
4
2
2
4
2
2
9
4
2
2
4444 444
4
4 44 444 44
11111111111
00000 00000
2
9
0 00000000000000 0000 0
4
55 8
1 611 1 111111
222222222222222222
1 111 1111
8 88888 88 8888 888
44 4 444
0 0 000 00
2 2222222 2222222221 2
00000000000 00000000000000
1111 1111 1
88 888888 8 888 88 888 5 555
4444 444444444444 44 4
33 3333 33333
1
2
7
2
2
2
2
2
7
3
0
8 8 8888 88888888 88 8
0
55
1 11 111 1
0 0000
32222 222222222222222
444 444 5 3 3 3 33333333 3
8
88888 8888
555 5 55
911 1 111111111111
2 2 27222
3 3 333 33 3333 33 3
8 55555555
8
3 33333333 3333333333 53 555 55 55 5555555
55 555 555
00
30 2 7 7777
3 33 3
55 555 5 5
9999 9
7
5
9
7
7
5
5
7
3
3
7 777 7
44444
9999999 999 9
99 9
3
33
7777777777777 77
444 44 4
4
4 4 4 999999999999999999999999999 9
555555555 55555 5
44 4 4 44
077777777777777
99999494 99 9 99999 9 9 99
55 5 5 5 555
3
11111
9999 9 9 9
4
9
9
55555555555555 555555555555555
39 7777777777777 9797
9
9 9999 9
1111111111111 99
44 4 9 99 99 9999999 9 99 9
7
7
5 55
7
7
1
7
7
7
77
9 9 9 99 9 9 9
11 1 111 1 1 5
55555 55 5
77 27
1 1
999949 9999999 9999 9 9999
7
99 9
1 11111111111111
77 777 7777 797 7 7
3
11111 111 1 1
3
3
3
3
3
3
777777 777 77777 777777777 7
1
3
3
1 111111 1
3 3333333 333 3
9
5
1 111 11 1
77777 77777777 77777777777777777
33333 33333333333 33333
11111111111 11
7 7 7 77 77 7 7 77
333 3333 333333333 3333333
1 111111 1
33 3
Disentangle Representation
ECVAE-II
35 55 5 555 55
5 3 55 5555555555555555
5 55555555555
9
5 55555555
3
3
3 3
5 565
33333333333333333 3593
6
6
33333 3333333 33333333
8
8
8888 8
333 3333 33
9
88888888888 8888
8
3 333333 333333
8
8
8
8
8
8
8
8
8
8
8
8
88 888 88 888 89
3 33333 3
5 66 6
9
8
8
8
8
8
8
6
4
66 6
8888888888888 8
3
9
444 444
666 666666 6
8 88888888
8 8
66666 6 66 6 6
44 44 44
8
44 44444444 444
6 6666 66 6
77
444444444444444444 4444 6 666666666 666 66666666
7
7
9
7
6 66666666666
444444444 444
77777777777 77
4
9 77777 7777 77777777777
1
6
111
9
7777777777777777777777
1
1
7
1
1
1 11
77777777 77 7
3
111111111 11111111111111
7
0000
777 772
111111111111
00 000000000000000 00009
8 32
0 00000 0 000 000000
111111
0000006 000000
1 111111111
22 22
00 0000000 00
2
1 11111111 1
02000000
2222 22222222 22 2 2
1 11
2222222222 2222222222222222222
2222 222222 2 22 2
22 4
2 222222 2222
22
99999 99999
99 9999
99999999999999999
9999999999999999999
The first experiment is to explore how the disentangle variable v
and non-interpretable variable z perform in the image reconstruction. The experiment is conducted on the MNIST dataset. In the
training data, we randomly select 3000 data samples as labeled
data and the remainings are unlabeled. The dimension of the disentangle variable v is 10 which is same as the category number,
and the dimension of z is 50.
We first train the model to learn the parameters. Then we use the
trained model to learn latent representation on the test data. After
learning the representations, we mask v and z in turn to see how
they affect the reconstruction of the input image. Two sample
results are shown in Fig.1. We also use t-SNE [28] to visualize v
of the testing data. The results from those four models (SDVAEI, SDVAE-I&IAF SDVAE-II and SDVAE-II&IAF) are shown in
Fig.2.
33333333333333
33333333333 333
33 33333333333 33333333 3
3 3333333
93
3
6666 666666666666
6 6666666666666666666666
6 6666666666 64
6666 66 66666
6 66
6
22 2
2 222 222222222 222
2222222222222222222 2222222 2
22222222 2222222 22222222
2 22222222222 24
2
2222
88888 8 8
88888888 8888888
88888888888888888888888
88
888888888888888
9 99
9 999999999
99 9999 9999
99 9999999999999999999999
9 999999 99999
11 11 11
11111111111111111111111111119
1111111111111 11111 1111111
1111 1111111111111
1
0
0000000000000000
00 00000000000000000
00000000000000000000
0
00 00 08
00
4 4444444 444 4
4 44 4 444
8 444444444444444444444444444
44444 4
4
7
77777777777777777777777
7777777 77777777
777777777777777777777777
777
7 777777777777
5 5
5555 55 5 55555
555 55555555555555555555 555555555
55555555555555
55555
Figure 2: The t-SNE distribution of the latent variable v from
proposed models, and different categories are in different colors
with number.
v mainly captures the categorical information, and it has little
influence over the reconstruction task. More specifically, from
Fig.2, we can see that images of the same class are clustered
Mask z
Mask v
together, implying that the disentangled representation captures
the categorical information. In addition, we find that cluster
SDVAE-I gives the worst visualization as clusters have intersections, while SDVAE-I&IAF and SDVAE-II&IAF give better
Figure 1: The first row in left figure and the right figure are the visualization, which suggests that SDVAE-I&IAF and SDVAEreconstruction images with the variable z and variable v masked II&IAF are better at capturing the categorical information.
respectively, and the images in the second row in both figures are
From Fig.1, we can see that when v is masked, z still reconstructs
the test images original.
the input image well, indicating that z is appropriate for reconstruction. To explore how variable z takes effect in the image
From Fig.1 and Fig.2, we can see that the disentangle variable reconstruction, we range a certain dimension of z from -2 to 2
5
4.2 Experiments on Text Dataset
on the specific labeled image, and the selected results are shown
in Fig.3. From the image, we can see that z can control differ3
4
3
0
4.2.1
Dataset Description
To test the model on text data, the IMDB data [20] is used. This
dataset contains 25,000 train samples and 25,000 test samples in
two categories.
Figure 3: The reconstruction images by varying z in a certain
dimension.
4.2.2 Model Structure
ent properties in image reconstruction with different dimensions,
such as italic controlling, bold controlling, transform control, and
In the application of the text data, the encoder is also the conthe style controlling, etc. These can be seen from images in Fig.3
volutional neural networks, but different from the case in image
left to right.
data, there are two convolutional neural networks which are referring from [12] parallelized together. One is extracting the feature
at the word level, and the other is extracting the feature at the
4.1.4 Semi-Supervised Learning
character level. As to the decoder, we applied the conditioned
LSTM [32], which is given as follows:
Furthermore, we conduct experiments to test the proposed models in semi-supervised learning on MNIST. We randomly select
ft = σ(Wf [z; v] + Uf ht−1 + bf )
x points from the training set as labeled data, where x is varit = σ(Wi [z; v] + Ui ht−1 + bi )
ied as {600, 1000, 3000}. The rest training data are used as unot = σ(Wo [z; v] + Uo ht−1 + bo )
labeled data. We compare with state-of-the-art supervised and
(18)
semi-supervised classification algorithms, which are used in [14].
It = Wc [z; v] + Uc ht−1 + bc
The experiments are conducted 10 times and the average accuct = ft ⊗ ct−1 + it ⊗ σ(It )
racy with standard deviation are showed in Table 1. Note that
ht = ot ⊗ relu(ct )
the performances of the compared methods are from [14] too.
From this table, we can see the proposed model SDVAE-II&IAF
performs best in classification and makes the least classification The conditional LSTM is same as the vanilla LSTM except for
errors (in black bold format) with small part of the labeled data. the current variable, which is replaced by the concatenation of
Although SDVAE-I performs not as good as other proposed mod- the latent variable z and v. The techniques of dropout [26] and
batch normalized [10] are both utilized in the encoder and deels, it still can achieve state-of-the-art results.
coder networks.
To further validate the observation, we also conduct the semisupervised learning over the SVHN, another popularly used
dataset. SVHN has 73,257 training samples and 26032 test samples. Among the training data, we randomly select 1000 data 4.2.3 Disentangle Representation
samples as labeled data and the rest as unlabeled data. The results
are shown in Table 2. Similarly, we can observe that SDVAE- We randomly select 20k samples from the training set as the laII&IAF gives the best performance.
beled data, and others are unlabeled during the training. Similarly, we use the t-SNE to visualize the disentangle variable
v ∈ N 2 and the non-interpretable variable z ∈ N 50 from the
proposed model on the test data and unlabeled data. Results are
showed in Fig.4,
Table 2: The results on the SVHN data
Method
KNN
TSVM
Semi-VAE(M1)+KNN
Semi-VAE(M1)+TSVM
Semi-VAE(M1+M2)
SDVAE-I
SDVAE-I&IAF
SDVAE-II
SDVAE-II&IAF
([14])
Test error rate
77.93% (±0.08)
66.55% (±0.10)
65.63% (±0.15)
54.33% (±0.11)
36.02% (±0.10)
47.32% (±0.13)
46.92% (±0.12)
44.16% (±0.14)
34.25% (±0.13)
From the left figure in Fig.4, we can see that the disentangle
representation v can clearly separate the positive and negative
samples while non-interpretable representation cannot, i.e., data
points from two clusters are interleaved with each other. This
suggests that the disentangle representation captures categorical
information well, and there is seldom categorical information in
non-interpretable variable.
6
Non-interpretable
variable z
1 01 0
1001
01
1
11 0
0 1
0
1
0 01
11 0
11001 1
11
1 0 010 010 1
0 10 101000 1 0
11111011
0 11
1
1 0 1 0 10 1
00 10
00 1
01 1
1100 01
1001
10
01
0
0 1 111 0 1000101 10
11010100
1 10100
11 1 01 0
011101
0101 00
0
0
1
0
101
0 1 1101011 1100010111100 10101
1
1
1111
000 111100
000
1011 10
0 110 10 0
01 01 01001001
0
11 1 1 1 01 1
0 00
0010100 110011101
1
011
001 1 11001110
100 100
01111
1 1 0 1 1 0 101111
11 001
101
01100
1
11 1 10 1001 01101001
0 10 1 0
10110111
10 0
011 1 00101101 0
1001 1011 00011
111
0 110000
1
000001011000
0 10 0 110110101 00
000 00101111 011
1
011
101 00 000 011 00000 1
1 10
00
10 11
000110 10100110
0 0 1110111010 1
1 101 1111010
1
1 10 10111
11
00001 100000
0
1
0
0
0
0
1
0
0
0
1
0
0
1
0
0
0
0
0
0100 0 1 11011 0010 1 0 1101 0 10001 0
1 11
10 00 11
100
00
0 11 0
10 000 10 001
01 1 011 1
101111 1 0
01
0 100 0
0
0 01 100
111011101010
1100
1 01 1
0 0111
0110
1 1
110 1110
10 001 111 10 01011000
0 0
00001110 00 0 0101
10 1
0 1
1 0010 0 01
1 0010 00 1
00
011 0
0 01 11 1 00 0 1011
0
0 0 1 11 100111000 101 101
0 1011010111 1 101 10 101
0 1 1 100 1
1
0
1
1
0
0
00 1
10 0 1 000100101
1 1 00
1 0 0 0 1011 01 1 1 1 11100 0
11 100011
1 0
11 1 1 01 010 1001 1 1100 0 1001 0
101 0
1 101010 0
0 1
000 10 00 0101 110 0
0 1
10 0
1
011
0 1
0 1 010 1 101 00 1 1 1 01
1
1
1
1
1
0
0
1 0 00
0
1
11
0
1
1 01 1 0
0
1
0 0
0
1
4.3
vd
Disentangle variable
0 00
00
0 000 00 0 00
0
00 0
1
1 111111
00000 000000 0
1111 1 11
0 000
000 0 000
0
11 1 1 11
11 1111
1
00 0 0
0
111111111
111 111 11 111
00000000
0000 00 000
1
0
0
111111 1
0 0000 0
111 111
0
0
0
1
0
1
1
1
000000 0
0
1 1 1
0000
0
0000
000
0
0
1 1 11 111
111
10
00
000 00
0
1
0 0 0000 0 00
1 11 1 1 11 1111
000 000 000000
00
1 1 1111 11111
0 0 00
0 0 00
00
1111 1 11 1 11
0
0
1
0
0
0
0
0
1
1
1
1
0
0
1
00 0
0
1
0 0000
11
11
11
1 1 111111 1
00 0 000000 00
1 11
1
000 0
111
111
0
1
1
1
11 1111 1
00 0 0
00
1
11
1 111 111
1111
00
1111111
1 1111 1
1
11
0
1111 1 1 11 1 111 1
00
11 11
00
11
1 11
00
0
0
1
00
1 11
0
1 11111
111
1 1111
0
000 00 0
1
0 0
111 1111
1 11 11
0
11
00 000 000
000000
0 00
1111
1
111 11111 1111
0
0
0
0
0
11 1 11111 11
0
1
1
1
0
1
0
1
0 00 00 000 0 000 0
1 1 1111 1 1111 1 111 11
111
0
0 0 0 00
1
1
1
0
0
1
1
1
0 0 0000
11 1 1 1 1
0 00
00
1 1
1 11
00000 0000 00
00000
00
11
00 0 0 0
1111 1 1 1
00 0
00
0 0 0000
1
0
11 11
0
0
1
1
1
0 0 0
00 00 000 0 00
111
1 1 11 1 11
0 0
0
00 0
00 000 00
11
1 1 1 111111
0000
0
000000 0
111
1111111
1
1111
0
000 0 00
11
0
0
1
0
0
1
0
0
1
1
0
1 11
00 00
1 1 11 1
0
11 111
11
1111
00 000000
11 1 111111
0 0000 00
0
0
0 000
1111 11 1
0 0000
11111
00 0
0
0
0
0000
0000
1111
000
00 0000000
111 11 1
11 11 1111
00000
11 111
00 000
00
11
11
000 0
11
1
0 00
Parameters Analysis
There are several important parameters need to be tuned for the
model, i.e., λ, β1 , β2 and the length of IAF. In this section, we
conduct experiments to analyze the sensitiveness of the model to
the parameters.
4.3.1
Effects of λ and the IAF Length
(a) Unlabeled Data
Non-interpretable
variable z
01
0 11
01 1
1 11 1 0 0
0 0 0 0 101 1
000
0 10 10000 1 10 011 00 00 11 11 11 1
1 1
1 0
1
1
0 11
10
0 001 00101 10 0 0 0
00001 00 1 01 1 0
0000 0 11101 11 0 0 0 01
01 1 01 00 0 00 1 0
0
1
0
0
11 1 10 00
1
1
0
1110 1 10
11
01
1
0 0 010 10 0 111 0 110 1 10 1
0 1 0011110 1110 010 1011011010001 111 0 001001 01 10 111
0 0 0001
1
0
0
0
1
1
1
1
0
1
0
0 0
1
11 0
1
01
1111 00 1 00 100101 00111 10100
00 0 101 1 00 0
01
10 001 0
1
1 1
01
01111 10
1101 01 0001 1 001 0
1
010101110
0
0 10000 0
00 100100
0
1
0
0 1 1
1
1
1
1
0
1
1
1
0 0
1 00000 0000
10
11110 00 0010
11
101 1 00 0
10
0 0
1
1 000 0 1 001 0
1 0 0010 1100 11111 1010000 1111
0000 010001
0 1 0
0
11 0 0
1 0 0 0 000 1 0
11
1 10010 001 1101
1
10
0010 1 11111
1
00 0 110100
1000 1101001
10 00 11111 01001 11 0 0
10 1 00
1 1 11
0
1
0
0
1
0
0
1
1
1 01 111 0001 1 0 1 0111 101001 1 001 1
0 0 00 0 1 10 11001
0
0
0 0
1
01
0 1 11101001
01 000 1001 10 0100
0 0011 01 101
00 00 0010 1
10 001 1
0
0 001010110
00 10
110010 1011 0010 1 1 0
0011 0
1110 100
011 10100101100 01 0 00100 011 011 1 1
0 00
1 000 10 010 1
0
1
0
1
0 0 0
0
0 0 0110 1
0
1
0
1
0 01 010 010110 01 1101 10
10 0 11 0
1
00 1 110 110 00 11 0 0001
001
0
101 0
1
0 0 00111 0 0010 0100 11 01101
1 0111 10000 111 00 010 110 1 0
10 0
11
1 0
0 1 11
0
11 111 00000 0 000 1101 00 11 0 0
1 00
10
1
1
0 1 00
0 00
1
0
1
1
1
0 0 01 0 1 1 01 1 0 11
1 1
1 0
0 0 1 0
000 1 1
01011 01 1 1 1
1
0
1
000
1
0
0
10 0 000 0 0 1
0
1
Disentangle0 0 variable vd
We firstly evaluate λ and the length of the IAF chain which are
proposed in the works of β-VAE [6] and IAF [15]. These experiments are conducted over the MNIST training dataset.
00
1
1
000
010
11
1
1
1
0 1101 00
010 10001
11 1
0 0
0 10
100
10
110
1
1
1
00
1010 0 00001
0100
11111 101111
1
1
0
00
0 0 01
1000
1
1
1
000000
01011
0 000 0
01 1 10101
11
11
1
1100 00 00000 0
010
001
1
11111
000
1
000
1
0
1
1
1
0
101 1 1
111
0 00
11
1 00
1
00000
1
1
00 1
1 10
00
1111
1111 111111
11 1
10010
00 0 00000
1
1 0110
1111
1
1
1
0
0 000
0
1
1
0
1
1
1
1
11111 111
0
10 1 0 01 0000 00 00001
0 011
1
0
0000 0 00
11 11111
110 0 0 1 0 1
0
1
1
0
1
0
1
0
0
000 000 00 0000
11 1 1 1
1 1
10
101 00000
00 00000
11 1111
0
10 010
00
000000
000 10000
0
01111111
10
1 1 11
0000
0 000000
10
0 10
1111 1
11 1
0 00000
00000000
000000 00000000
0000
00
1111 0 00
1 0101
1 00
0 000
1
0
11 1111
0
000 000 00
0
1
1
0
1
0
1
1
1
0
0
1
1
1
0
1
0
0
0
1111111
0 0 000 0 0000 0 000
1 1 00
0
11 1 11
1
0
0
11111
1
11 0 0 00101 000 0000 00
0 11111
1
11111
111011
01 1 1
0
0
0
0
111
111 111
0
0
001
1111
0
0000 0
1
11
1 1
0
0 100
0 0
00
1
11
1
0
101111111 1 0 0
11
1
00
11111011011
00
00
10
1 1 000
010 0100
01100
001000
11 1111
11110
0
11 01
00
1 00
0
00 0001
1111
1
0
00 000
10 0
1
00
000
10
11
1 10
10
0
11
0 00 00
0
00
11010
0 0000
1 000
0
11
1101
110
000000
111
001
0
1
1
0
0
1
0
0
1
11
0
1
00
0
111
0
11
10
1
1
11
11
The objective function in the proper λ finding is depicted in
Eq.(6). Results with different λ values are shown in Fig.5(a)
From the results, we can see that it is better for λ to have a small
value, which not only leads to a rich information in the latent
variable but also gets a better reconstruction error. But as de(b) Test Data
scribed before, the large value of KL-divergence is also the cause
Figure 4: The left figure is the t-SNE distribution of the non- of overfitting or the underfitting for the model. However, in the
interpretable variable z, the right figure is the t-SNE distribution case of λ = 0.1, there is a low reconstruction error, which is the
of the disentangle variable v correspondingly. Different cate- sign of the good performance.
gories are in different colors with number.
Then the model structure about the IAF chain is built according
to the Eq.(16), and the results with different length are shown in
4.2.4 Semi-Supervised Learning
the right figure in Fig.5(b). From the figure, we can see that it is
not good to set the chain too long if it is a long IAF. The RCEs are
We further conduct semi-supervised classification on the text not so good together with the KLDs, and the latent variable isvery
dataset using the representation learned from previous experi- unstable. On the contrary, there is a stable increase about the
ments and fine tuning the model. Similarly, we compare with KL divergence, and a stable decrease reconstruction error when
state-of-the-art semi-supervised learning algorithms. The aver- the length of the IAF chain is set to 1. This means that under
age test error rate is reported in Table 3. From the results, we can the good reconstruction, the latent variable captures more useful
see that: (i) SDVAE-II&IAF outperforms the compared meth- information. This is also validated in the results of the SDVAEods, which implies the effectiveness of the proposed framework I&IAF and SDVAE-II&IAF. Thus, in the experiments about the
for semi-supervised learning; and (ii) As we add reinforcement IAF, its length is set to 1 by default.
Powered by TCPDF (www.tcpdf.org)
Powered by TCPDF (www.tcpdf.org)
learning and IAF, the performance increases, which suggests the
two components contribute to the model.
4.3.2
Table 3: The results on the IMDB data
Method
LSTM ([4])
Full+Unlabeled+BoW ([20])
WRRBM+BoW ([20])
NBSVM-bi ([30])
seq2-bown-CNN ([11])
Paragraph Vectors ([17])
LM-LSTM ([4])
SA-LSTM ([4])
SSVAE-II&LM ([33])
SDVAE-I
SDVAE-I&IAF
SDVAE-II
SDVAE-II&IAF
Effects of β1 and β2
To decide the parameter β1 and β2 that in SDVAE-II, we made
the grid search both on the text data and the image data. For the
image data, the experiment is conducted on the SVHN dataset
with 1000 labeled samples. Experimental result with β1 ranges
from 0.1 to 1000, and β2 ranges from 0.01 to 100 are shown in
Fig.6(a). For the text data, the experiment is conducted on the
IMDB data with 20,000 labeled samples. Experimental result
with β1 and β2 range from 0.1 to 1000 are shown in Fig.6(b).
Test error rate
13.50%
11.11%
10.77%
8.78%
7.67%
7.42%
7.64%
7.24%
7.23%
12.56%
11.60%
7.37%
7.18%
From the Fig.6(a), we can see that, an acceptable range for β1 in
the image data is [0.1:100] and [0.01:10] for the β2 . Especially,
when β1 = 0.1 and β2 = 1, it is achieving the best result.
For the text data, the results in the Fig.6(b) show that the accuracy
7
λ=0.1
λ= 0 . 5
λ= 1 . 0
λ= 2 . 0
1 5 0
1 4 0
λ=
λ=
λ=
λ=
0 .1
0 .5
1 .0
2 .0
5 0
4 0
1 3 0
K L D
R C E
3 0
1 2 0
2 0
1 1 0
(a) The Image data
1 0
1 0 0
9 0
Figure 6: The grid search results for the proper β1 and β2 finding.
0
0
3 0
6 0
9 0
1 2 0
1 5 0
E p o c h
semi-supervised VAE all use a parametric classifier, which increases the burden to learn more parameters given the limited
labeled data. The proposed framework incorporates the label information directly into the disentangled representation and thus
avoids the parametric classifier.
(a) Validate λ
t=1
t=1
1 5 0
1 3 0
1 8
t=1 0
t=2 0
t=3 0
t=1 0
t=2 0
t=3 0
1 4 0
1 6
1 4
1 2
1 1 0
1 0
1 0 0
8
9 0
6
8 0
4
7 0
2
6 0
Variants of VAE Because of the great potential of VAE in image and text mining, various models based on VAE are proposed
to further improve its performance [16, 6, 7, 15]. For example,
[6] apply the KKT condition in the VAE, which gave a tighter
lower bound. Similarly, [3] introduce importance weighting to
VAE, which also tries to give a tighter bound. [24] consider the
stein based sampling to minimize the KL divergence. [7] rewrite
the evidence lower bound objective by decomposition, and give
a clear explanation for each term. To extend the flexible of the
posterior inference, IAF is introduced [15] which improves the
VAE a lot.
K L D
R C E
1 2 0
0
0
2 0
4 0
6 0
8 0
1 0 0
1 2 0
1 4 0
E p o c h
(b) Validate Length of IAF
Figure 5: The left y-axis in each figure is the reconstruction error
(BCE) which is axis of the solid lines, and the right y-axis is the
KL divergence (KLD) which is axis of the dash lines.
6
is not sensitive to β2 . However, when β1 is small, the result will
be more precise. In conclusion, it is better to set β1 to 0.1 and β2
can be set randomly.
5
(b) The Text data
Conclusions
In this work, we propose models that extract the disentangle variable v and the non-interpretable variable z from data at the same
time. The disentangle variable is designed to capture the category information and thus relieves the use of classifiers in semisupervised learning. The non-interpretable variable is designed to
reconstruct the data. Experiments show that it could even reflect
certain textual features, such as italic, bold, transform and style
in the hand writing digital data during the reconstruction. These
two variables cooperate well and each performs its own functions
in the SDVAE. The IAF improves the model effectively on the
basis of SDVAE-I and SDVAE-II. Especially in which, SDVAEII&IAF achieves the state-of-the-art results both in image data
and the text data in the semi-supervised learning tasks.
Related Works
Semi-supervised VAE Semi-supervised learning is attracting increasing attention, and lots of works are proposed [33, 8, 22, 31,
23, 14, 27, 5]. Those works can be divided into the discriminative models [29, 4], the generative models [33, 8, 22], graph based
models [27], and the combined model with those [5]. Because of
the effectiveness of deep generative models in capturing the data
distribution, semi-supervised models based on deep generative
models such as generative adversarial network [25] and variational auto-encoder (VAE) [14] become popular. Semi-VAE [14]
incorporates the learned latent variable into the classifier and improves the performance greatly. SSVAE [33] extends Semi-VAE
for sequence data, and also demonstrates its effectiveness in the
semi-supervised learning on the text data. The aforementioned
References
[1] Dimitri P Bertsekas. Nonlinear programming. Athena scientific Belmont, 1999.
8
[2] Samuel R Bowman, Luke Vilnis, Oriol Vinyals, and et al. [16] Diederik P Kingma and Max Welling. Auto-encoding variGenerating sentences from a continuous space. arXiv
ational bayes. arXiv preprint arXiv:1312.6114, 2013.
preprint arXiv:1511.06349, 2015.
[17] Quoc Le and Tomas Mikolov. Distributed representations
[3] Yuri Burda, Roger Grosse, and Ruslan Salakhutdiof sentences and documents. In Proceedings of the ICML,
nov. Importance weighted autoencoders. arXiv preprint
pages 1188–1196, 2014.
arXiv:1509.00519, 2015.
[18] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick
[4] Andrew M Dai and Quoc V Le. Semi-supervised sequence
Haffner. Gradient-based learning applied to document
learning. In Proceedings of the NIPS, pages 3079–3087,
recognition. Proceedings of the IEEE, 86:2278–2324, 1998.
2015.
[19] Lars Maaløe, Casper Kaae Sønderby, Søren Kaae Sønderby,
[5] Jingrui He, Jaime G Carbonell, and Yan Liu. Graph-based
and Ole Winther. Auxiliary deep generative models. arXiv
semi-supervised learning as a generative model. In Propreprint arXiv:1602.05473, 2016.
ceedings of the IJCAI, volume 7, pages 2492–2497, 2007.
[20] Andrew L Maas, Raymond E Daly, Peter T Pham, and et al.
[6] Irina Higgins, Loic Matthey, Arka Pal, and et al. beta-vae:
Learning word vectors for sentiment analysis. In ProceedLearning basic visual concepts with a constrained variaings of the ACL: Human Language Technologies-Volume 1,
tional framework. In Proceedings of the ICLR, 2017.
pages 142–150, 2011.
[7] Matthew D Hoffman and Matthew J Johnson. Elbo surgery: [21] Yuval Netzer, Tao Wang, Adam Coates, and et al. Reading
yet another way to carve up the variational evidence lower
digits in natural images with unsupervised feature learning.
bound. In Proceedings of the NIPS, Workshop in Advances
In Proceedings of the NIPS, Workshop on deep learning and
in Approximate Bayesian Inference, 2016.
unsupervised feature learning, volume 2011, page 5, 2011.
[8] Zhiting Hu, Zichao Yang, Xiaodan Liang, and et al. Toward
[22] Augustus Odena. Semi-supervised learning with generative
controlled generation of text. In Proceedings of the ICML,
adversarial networks. arXiv preprint arXiv:1606.01583,
pages 1587–1596, 2017.
2016.
[9] Gao Huang, Shiji Song, Jatinder ND Gupta, and Cheng
[23] Yong Peng, Bao-Liang Lu, and Suhang Wang. Enhanced
Wu. Semi-supervised and unsupervised extreme learning
low-rank representation via sparse manifold adaption for
machines. IEEE transactions on cybernetics, 44(12):2405–
semi-supervised learning. Neural Networks, 65:1–17, 2015.
2417, 2014.
[10] Sergey Ioffe and Christian Szegedy. Batch normalization: [24] Yunchen Pu, Zhe Gan, Ricardo Henao, and et al. Stein
variational autoencoder. arXiv preprint arXiv:1704.05155,
Accelerating deep network training by reducing internal co2017.
variate shift. In Proceedings of the ICML, pages 448–456,
2015.
[25] Jost Tobias Springenberg.
Unsupervised and semisupervised
learning
with
categorical
generative adversarial
[11] Rie Johnson and Tong Zhang. Effective use of word order
networks.
arXiv
preprint
arXiv:1511.06390,
2015.
for text categorization with convolutional neural networks.
arXiv preprint arXiv:1412.1058, 2014.
[12]
[13]
[14]
[15]
[26] Nitish Srivastava, Geoffrey E Hinton, Alex Krizhevsky, and
et al. Dropout: a simple way to prevent neural networks
Yoon Kim. Convolutional neural networks for sentence
from overfitting. Journal of Machine Learning Research,
classification. arXiv preprint arXiv:1408.5882, 2014.
15(1):1929–1958, 2014.
Diederik Kingma and Jimmy Ba. Adam: A method for
stochastic optimization. arXiv preprint arXiv:1412.6980, [27] Amarnag Subramanya and Partha Pratim Talukdar. Graphbased semi-supervised learning. Synthesis Lectures on
2014.
Artificial Intelligence and Machine Learning, 8(4):1–125,
Diederik P Kingma, Shakir Mohamed, Danilo Jimenez
2014.
Rezende, and Max Welling. Semi-supervised learning with
deep generative models. In Proceedings of the NIPS, pages [28] Laurens Van der Maaten and Geoffrey Hinton. Visualizing
data using t-sne. Journal of Machine Learning Research,
3581–3589, 2014.
9:2579–2605, 2008.
Diederik P Kingma, Tim Salimans, Rafal Jozefowicz, and
et al. Improved variational inference with inverse autore- [29] V Vapnik and A Sterin. On structural risk minimization or
gressive flow. In Proceedings of the NIPS, pages 4743–
overall risk in a problem of pattern recognition. Automation
4751, 2016.
and Remote Control, 10(3):1495–1503, 1977.
9
[30] Sida Wang and Christopher D Manning. Baselines and bigrams: Simple, good sentiment and topic classification. In
Proceedings of the ACL: Short Papers-Volume 2, pages 90–
94, 2012.
[31] Suhang Wang, Jiliang Tang, Charu Aggarwal, and Huan
Liu. Linked document embedding for classification. In Proceedings of the 25th ACM International on Conference on
Information and Knowledge Management, pages 115–124.
ACM, 2016.
[32] Tsung-Hsien Wen, Milica Gasic, Nikola Mrksic, and et al.
Semantically conditioned lstm-based natural language generation for spoken dialogue systems.
arXiv preprint
arXiv:1508.01745, 2015.
[33] Weidi Xu, Haoze Sun, Chao Deng, and Ying Tan. Variational autoencoder for semi-supervised text classification.
In Proceedings of the AAAI, pages 3358–3364, 2017.
[34] Xinchen Yan, Jimei Yang, Kihyuk Sohn, and Honglak Lee.
Attribute2image: Conditional image generation from visual
attributes. In Proceedings of the ECCV, pages 776–791,
2016.
[35] Zichao Yang, Zhiting Hu, Ruslan Salakhutdinov, and Taylor
Berg-Kirkpatrick. Improved variational autoencoders for
text modeling using dilated convolutions. arXiv preprint
arXiv:1702.08139, 2017.
10
| 2cs.AI |
Applying the Z-transform for the stati analysis of
oating-point numeri al lters
arXiv:0706.0252v1 [cs.PL] 2 Jun 2007
David Monniaux
April 3, 2018
1 Introdu tion
The stati
analysis of
ontrol/ ommand programs, with a view to proving the
absen e of runtime errors, has re ently pi ked up steam, with the in eption
of analyzers
it is nowadays possible to build
to realisti
(sour e
sound
and
industrial situations. A stati
ode or obje t
stati
analyzers s aling up
analyzer takes as input a program
ode) and outputs a series of fa ts, warnings or other
indi ations obtained by automati
A stati
pre ise
apable of s aling up to real industrial programs. In parti ular,
analysis of that program.
analyzer is said to be sound if all the fa ts that it derives from
a program (say, variable
x
is always positive) are always true, regardless
semanti s
of how and on whi h inputs the program is run. Sound stati
based on a
exe utions.
analyzers are
, that is, a mathemati al de nition of possible program
It is well-known that any method for program veri ation
annot be
at the same time sound (all results produ ed are truthful), automati
human intervention),
omplete (true results
minating (always produ es a result)
(no
an always be proved) and ter-
1 unless one supposes that program
false
memory is nite and thus that the system is available to model- he king
alarms
te hniques. As a result, sound stati
ality.
analyzers are bound to produ e
pre ise
sometimes; that is, warnings about issues that
One thus wants analyzers that are
annot happen in re-
, that is, model reality so
losely that they seldom produ e false alarms but also, one wants analyzers that are e ient, taking only reasonable amounts of time and memory
to perform an analysis.
One
ru ial
lass of errors for
over ows say, when
using oating-point
ontrol/ ommand systems is arithmeti
onverting some value to an integer in programs
omputations.
Su h errors have already proved to be
extremely dangerous, having for instan e
1
The formal version of this result is a
aused the explosion of the Ariane 5
lassi
theorem.
1
of re ursion theory, known as Ri e's
on its maiden ight [16℄. In order to prove the absen e of su h errors, stati
analyzers su h as Astrée
2 [1, 2℄ have to bound all oating-point variables in
the program. It is impossible to do so using simple interval arithmeti ; in
order to bound the output of a numeri al lter, one has to make the analyzer
understand the stability
onditions of the numeri al pro essing implemented
in the appli ation to be analyzed.
In
urrent
ontrol/ ommand designs, it is
utable is obtained by
by automati
ompiling C
ommonpla e that the exe-
ode, or assembly
ode, itself obtained
translation from a high-level spe i ation. This high-level spe -
i ation is typi ally given in a high-level language su h as Simulink
3 Lustre
TM ,4 These languages, in their simplest form, onsider programs
[4℄ or S ade
to be the software
ounterpart of a network of ele troni
integrators, rate limiters...)
onne ted by wires; this is a tually how several
of these languages represents programs graphi ally. Several
grouped into a
ir uits ( lters,
ir uits
an be
ompound lter.
One advantage of these high-level languages is that their semanti s is
onsiderably
and
of
leaner than those of low-level languages su h as C. The lter
ompound lter
onstru tions provide natural boundaries for blo ks
omputations that belong together and probably have some interesting
ompositional
modular
and identi able properties. It is thus interesting to be able to analyze these
languages in a
and
fashion; that is, the analysis of
some blo k ( ompound lter) is done independently of that of the rest of the
ode, and the result of that analysis may be plugged in when analyzing
larger programs.
ear
This paper deals with the
lters.
ompositional and modular analysis of
By this, we mean lters that would be linear had they been
implemented over the real eld. Of
ourse, in reality, these lters are imple-
mented over oating-point numbers and none of the
relationships hold.
point
lin-
lassi al mathemati al
We nevertheless provide sound semanti s for oating-
omputations and sound analysis for su h lters.
1.1 Digital ltering
Control/ ommand programs in embedded appli ations often make use of
linear lters (for instan e, low-pass, high-pass, et .). The design prin iples
of these lters over the real numbers are well known; standard basi
designs
(Butterworth, Chebyshev, et .) and standard assembly te hniques (parallel,
serial) are taught in the
2
3
urri ulum of signal pro essing engineers.
http://www.astree.ens.fr
Simulink
TM
is a tool for modelling dynami
e.g. networks of numeri
lters. The
systems and
ontrol appli ations, using
ontrol part may then be
ompiled to hardware or
software.
http://www.mathworks. om/produ ts/simulink/
4
Ample
S ade is a
ommer ial produ t based on LUSTRE.
http://www.esterel-te hnologies. om/produ ts/s ade-suite/
2
literature has been devoted to the design of digital lters implementing some
desirable response, for implementation in sili on or in software, in xed-point
and in oating-point.[12℄
However, dis rete-time lters are often dis ussed assuming
on real numbers. There is still some
omputations
onsiderable literature on the impli a-
tions of xed-point or oating-point numbers, but the vast majority of the
work has fo used on usual
ideal
ase or average
that worst- ase bounds on
ase bounds it is even argued
lters on real numbers are too pessimisti
and not relevant for lter design [12, 11.3℄. The study of the quantization
and roundo noise generated by xed-point or oating-point implementations has mostly been done from a sto hasti
average
point of view, in order to prove
ase properties.
For our analysis purposes, we need sound worst- ase bounds, and pra tial means for obtaining them with reasonable
omputational resour es. For
these reasons, the point of view of the designers of stati
from that of the lter designers.
A favorite tool of lter designers is the
whi h the overall ideal (i.e.
analyzers is dierent
Z-transform
[12,
hapter 3℄, with
implemented over the real numbers) transfer
fun tion of a lter is summarized in a rational fun tion with real
whose poles and zeroes determine the frequen y response.
we shall show how we
oe ients,
In this paper,
an use this transform to automati ally summarize
networks of linear lters; how this transform allows us to
ompute pre ise
bounds on the out ome of the lter, and to stati ally summarize
omplex
lters; and how to deal with roundo errors arising from oating-point
om-
putations.
1.2 Contributions of the arti le
This arti le gives a sound
abstra t semanti s
for linear numeri al lters im-
plemented in oating-point or xed-point arithmeti s, given as the sum of a
linear part (using the Z-transform) and a nonlinear part (given using a ne
bounds); this latter part
omes from the roundo noise (and, possibly, some
optional losses of linear pre ision done for the sake of the speed of the analysis). (Se t. 4 for the ideal, linear part, 7 for the nonlinear part).
In many o
real
(Se t. 2.3) of
the
asions, the
omputed bounds are obtained from the norms
ertain power series. In Se t. 5, we give ee tive methods on
numbers for bounding su h norms.
In Se t.8 we explain how to
implement some of these methods e iently and soundly using integer and
oating-point arithmeti s. In Se t. 9 we study a few
ases.
As with other numeri al domains su h as those developed for Astrée, we
pro eed as follows:
the exa t oating-point
on rete semanti s is overap-
proximated by a mathemati ally simple semanti s on real numbers, whi h
is itself overapproximated by proved bounds, whi h are themselves further
overapproximated by an exe utable semanti s (implemented partly in exa t
3
×α0
+
+
normal
input
delay
×α1
delay
+
×β1
+
delay
feedback input
delay
×α2
×β2
feedback sub−filter
compositional blocks
Figure 1: De omposition of the TF2 lter
+ β1 Sn−1 + β2 Sn−2
hained by serial
tary gates are
Sn = α0 En + α1 En−1 + α2 En−2
into elementary blo ks.
omposition. Inside ea h
The
ompositional blo ks are
ompositional on the left, elemen-
omposed in parallel. On the right hand side, a feedba k loop
is used.
arithmeti s, partly using some variant of interval oating-point
tions). This ensures the soundness of the ee tive
omputa-
omputations.
This paper is an extended version of [18℄.
1.3 Introdu tion to linear lters and Z-transforms
Let us
onsider the following pie e of C
ode, whi h we will use as a running
example ( alled TF2):
Y = A0*I + A1*Ibuf[1℄ + A2*Ibuf[2℄;
O = Y + B1*Obuf[1℄ + B2*Obuf[2℄;
Ibuf[2℄=Ibuf[1℄; Ibuf[1℄=I;
Obuf[2℄=Obuf[1℄; Obuf[1℄=O;
All variables are assumed to be real numbers (we shall explain in later
se tions how to deal with xed- and oating-point values with full generality
and soundness). The program takes
time
onstant
the
oe ients. This pie e of
I as an input and outputs O; A0 et
. are
ode is wrapped inside a (rea tive) loop;
is the number of iterations of that loop. Equivalently, this lter
an
be represented by the blo k diagram in Fig. 1.
onstants and in (resp. yn , on ) the
n. Then, assuming ok = 0 for k < 0, we an
develop the re urren e: on = yn + b1 .on−1 + b2 .on−2 = yn + b1 .(yn−1 + b1 .on−2
+b2 .on−3 )+b2 .(yn−2 +b1 .on−3 +b2 .on−4 ) = yn +b1 .yn−1 +(b2 +b21 b0 ).yn−2 +. . .
where . . . depends solely on yk with k < n − 2. More generally: there
P
exist oe ients c0 , c1 . . . su h that for all n, on =
k=0 ck yn−k . These
Let us note
value of
I
a0 et .
Y, O)
(resp.
the values of the
at time
4
oe ients solely depend on the
for
bk ;
we shall see later some general formulas
omputing them.
But, itself,
oe ients
c′n
yn = a0 .in + a1 .in−1 + a2 .in−2 . It follows that
P there exist
ak and the bk ) su h that on = k=0 c′k in−k .
(depending on the
onvolution kernel
We again nd a similar shape of formula, known as a
The
c′k
sequen e is
alled a
, mapping
MI
Let us now suppose that we know a bound
|in | ≤ MI ;
inequality,
we wish to derive a bound
P
′
k=0 |ck |.MI .
′
onvolution kernel c .
|On | ≤
norm of the
MO
onvolution produ t
i
to
on the input: for all
ompositionally
omputes
n,
on the output. By the triangle
The quantity
P
′
k=0 |ck | is
alled the
What our method does is as follows: from the des ription of a
linear lter, it
.
o.
l1-
omplex
ompa t, nite representations of
onvolution kernels mapping the inputs to the outputs of the sub-blo ks of
the lter, and a
a
urately
omputes the norms of these kernels (or rather,
lose upper bound thereof ).
As a result, one
an obtain bounds on any
variable in the system from a bound on the input.
2 Linear lters
In this se tion, we give a rough outline of what we designate by linear lters
and how their basi
properties allow them to be analyzed.
2.1 Notion of lters
ausality
We deal with numeri al lters that take as inputs and output some (unbounded) dis rete streams of oating-point numbers, with
the output of the lter at time
(times 0 to
t),
t
but not on the future inputs.
mented with a state, and the output at time
time
t
; that is,
depends on the past and present inputs
5 In pra ti e, they are imple-
t
is a fun tion of the input at
and the internal state, whi h is updated.
Su h lters are typi ally
implemented as one pie e of a syn hronous rea tive loop [2, 4℄:
while(true) {
...
(state, output) = filter(state, input);
}
5
There exist non- ausal numeri al ltering te hniques One striking example is Mat-
lab's
filtfilt
fun tion, whi h runs the same
ausal lter in one dire tion, then in the
reverse-time dire tion over the same signal; the overall lter has zero phase shift at all
frequen ies, a very desirable
hara teristi
in some appli ations. Unfortunately, as seen
on this example, non- ausal lters require buering the signal and thus are not usable for
real-time appli ations. They are outside the s ope of this paper.
5
2.2 Linear lters
We are parti ular interested in lters of the following form (or
thereof ): if
(sk )
and
(ek )
the lter, there exist real
for all time
t, s t
ompounds
are respe tively the input and output streams of
oe ients
(the output at time
st =
n
X
α0 , α1 , . . . αn and β1 ,
t) is de ned as:
αk et−k +
k=0
m
X
. . . βm su h that
βk st−k
(1)
k=1
or, to make apparent the state variables,
0
1
st−m+1
..
..
.. .
.
. =
0 ···
st
βm · · ·
If the
β
..
.
0
β2
0
···
st−m
.
.. ..
. . +
0 ···
1
st−1
β1
αn · · ·
0
et−n
..
. .
0
en
α0
.
.
.
(2)
nite impulsional response
in nite impulsional response
are all null, the lter has ne essarily
(FIR) while in the opposite
ase, it may have
(IIR). The reason for this terminology is the study of the rea tion of the
system to a unit impulse (e0
FIR lter,
n + 1 time
= 1
and
∀k > 0 ek = 0).
In the
ase of a
units after the end of the impulse, the output be omes
permanently null. In the
ase of an IIR lter, the output (when
stable
omputed
ideally in the real numbers) never be omes permanently null, but rather
follows some exponential de ay if the lter is
lter may be unstable.
. A badly designed IIR
Furthermore, it is possible to design lters that
should be stable, assuming the use of real numbers in
omputation, but that
exhibit gross numeri al distortions due to the use of oating-point numbers
in the implementation.
Linear lters are generally noted using their
Z-transform
α0 + α1 z + · · · + αn z n
1 − β1 z − · · · − βm z m
The reasons for this notation will be made
all the ideal
6
(3)
lear in Se t. 4.5. In parti ular,
ompound linear lters expressible with elementary elements
su h as produ ts by
onstants, delays, et .
an be summarized by their Z-
transform (Se t. 4); that is, they are equivalent to a lter whose output is
a linear
ombination of the last
will also be
n
inputs and
m
outputs. The Z-transform
entral in the semanti s of oating-point and xed-point lters
(Se t. 7).
6
An alternate notation [12℄ repla es all o
urren es of
z
by
z −1 .
In su h a formalism,
onditions su h as the poles must have a module greater than 1 are repla ed by the
equivalent for the inverse, e.g. the poles must have a module stri tly less than 1. We
hose polynomials in
z
be ause they allow using normal power series instead of Laurent
series.
6
To summarize some salient points of the following se tions, FIR lters
α's
given by
and de ay
are very easy to deal with for our purposes, while the stability
onditions of IIR lters are determined by the study of the above
zero
rational fun tion and espe ially the module of the zeroes of the
1 − β1 z − · · · −
βm z m polynomial (z0 is a
of
Q
if
Q(z) =
Q(z0 ) = 0).
Those
roots are the inverses of the eigenvalues of the transition matrix. Spe i ally,
the lter is stable if all the zeroes have module greater than 1.
2.3 Bounding the response of the lter
The output streams of a linear lter, as an element of
RN , are linear fun
tions
of the inputs and the initial values of the state variables (internal state
variables).
onvolu-
More pre isely, we shall see later that, negle ting the oating-point errors
tion produ t Q ⋆ E
of the input
(qn )n∈N
a sequen e
E
by some
of reals su h that for any
lter is FIR if this
S
onvolution kernel
P Q
and assuming zero in the initial state variables, the output
is the
: there exists
n
k=0 qk en−k . The
n, sn =
onvolution kernel is null ex ept for the rst few values,
and IIR otherwise. If the initial state values r1 , . . . , rn are nonzero, then
S = Q0 ⋆ E + r1 Q1 + rn Qn where the Qi are onvolution kernels.
Let E : (ek )n∈N be a sequen e of real or omplex numbers. We all ∞
of E , if nite, and note kEk∞ the quantity supk∈N |ek |. Be ause of the
L -
norm
isomorphism between sequen es and formal power series, we shall likewise
note
k
P
k
ak z k k∞ = supk |ak |.
We are interested in bounding the response
of the lter with respe t to the in nite norm: i.e. we want to
f
fun tion
the
has
su h that
kSk∞ ≤ f (kEk∞ ).
Said otherwise, if for all the past of
omputation sin e the last reset of the lter,
|s|
has been always less than
f (|M |)
If we do not have initialization
onstru t a
|e|
was less than
M,
then
sin e the last reset.
onditions nor oating-point errors,
f
will
be linear, otherwise it will be a ne. Let us pla e ourselves for now in the
former
ase: we are trying to nd a number
in nite norm
For any linear fun tion
of
f,
f
g
su h that
kSk∞ ≤ g.kEk∞.
mapping sequen es to sequen es, we
noted,
kf k∞
all
subordinate
supkxk∞ =1 kf (x)k∞ , assuming
g = kE 7→ Q ⋆ Ek∞ . If this quantity is
the quantity
is is nite. We are thus interested in
nite, the lter is stable; if it is not, it is unstable: it is possible to feed an
input sequen e to the lter, nitely bounded, whi h we result in arbitrarily
high outputs at some point in time.
L1-norm
For a sequen e (or formal series)
A, we note kAk1 =
, if nite. Then we have the following
[12, 11.3℄:
Lemma 1.
Proof.
kE 7→ Q ⋆ Ek∞ = kQk1
We shall rst prove that
sequen es
Q
and
E , kQ ⋆ Ek∞
P∞
k=0 |ak |,
alled its
ru ial and well-known result
.
kE 7→ Q ⋆ Ek∞ ≤ kQk1 ; that is, for any
≤ kQk1 .kEk∞ . Let us note C = Q ⋆ E .
7
P
cn = nk=0 qk en−k ,
kek∞ .kQk1 .
therefore
Pn
|cn | ≤
k=0 |qk ||en−k |
≤ kek∞ .
Pn
k=0 |qk |
≤
< kOk1 . Re all that kQk∞ =
P∞We shall then show equality. Let M P
N
|q
|
.
Then
there
exists
N
su
h
that
k=0 |qk | ≥ M . Choose ek = 1
k=0 k
if k ≤ N and qn−k ≥ 0, ek = −1 otherwise. Clearly, kEk∞ = 1, and
Pn
Pn
cn =
k=0 ek qn−k =
k=0 |qn−k | ≥ M , therefore kQ ⋆ Ek∞ ≥ M and
kE 7→ Q ⋆ Ek∞ ≥ M . Sin e this is valid for any M < kQk1 , then the
kE 7→ Q ⋆ Ek∞ = kQk1 equality holds.
Note that most of the dis ussion on numeri al lters found in the signal
pro essing literated is based on the L2-norm
is adapted to energy
kxk2 =
2 1/2 (whi h
k=0 |xk |
P∞
onsiderations) for instan e, for estimating the fre-
quen y spe trum of the rounding noise. We shall never use this norm in this
arti le.
3 Convolution kernels as formal power series
In the pre eding se tion, we said that the output of the ideal lter is just the
formal power series
rational fun tions
onvolution of the input with some (possibly in nite) kernel. In this se tion,
we show how
onvolution, and basi
given as
are a good framework for des ribing this
fa ts about the kernels of the lters we are interested,
.
3.1 Formal power series
We shall rst re all a few de nitions and fa ts about formal power series.
K[[X]] over a eld K = R or C is the ve tor
N
spa e of ountably in nite sequen es K where the produ t of two sequen es
A : (aP
k )k∈N and B : (bk )k∈N is de ned as A.B : (ck )k∈N by, for all n ∈ N,
cn = nk=0 ak bn−k ( onvolution). Remark that for any algebra operation
(addition, subtra tion, multipli ation) and any N , we obtain the same results for the oe ients cn for n ≤ N as if A and B were the oe ients of
polynomials and we were omputing the oe ient cn , the n-th degree oef7
ient of the polynomial A.B . For this reason, we shall from now on note
P∞
k
A(z) =
k=0 ak z by analogy with the polynomials. Note that for most
The algebra formal power series
of this arti le, we are interested in
formal
possible interpretation as holomorphi
all if the
onvergen e radius of the
the rare o
the needed
asions when we need
7
One
with the
P∞
A
k=0 ak z
ak
are null ex ept for a nite number,
is a polynomial.
K[[X]] as the proje tive limit of the K[X]/X n quotient
K[X]/X n+1 → K[X]/X n morphisms in the ategory of rings.
an therefore see
anoni al
k series is null); we shall note
onvergen e properties (and we shall prove
onvergen es). If all the
the formal series
power series and not with their
fun tions (i.e. it is not a problem at
8
rings
Wherever we have a
lently
onvolution
onsider a produ t
A.B
(ak ) ⋆ (bk )
inverse
of sequen es, we
an equiva-
of formal series.
We shall often wish to take the
of a power series, and the quotient
P
k
k bk b su h that b0 is
(n) as follows: A(0) = A, A(n+1) =
not null. We de ne a sequen e of series A
(n)
(n)
(n)
k
qn = an /b0 . Note that for all
n ∈ N, k < n Ak = 0
A − qn ∗ z B where
P
P
n
n
(n+1) +
k
k
n
and A = A
k=0 qk z B ; thus for all n, A ≡
k=0 qk z B (mod X ),
n
whi h may equivalently written as A ≡ Q.B (mod X ). Therefore, A =
A/B
of two series.
Q.B ,
whi h explains why
This is possible for any series
A very important
Another important
Q
an be
S
and
E
quotient
ase for the rest of the paper is
A by B . P
k
1/(1 − z) = ∞
k=0 z .
of
onstatation is that this quotient formula applied to
S = E.
where
alled the
α0 + α1 z + · · · + αn z n
1 − β1 z − · · · − βm z m
(4)
are expressed as formal power series is equivalent to running
the IIR lter de ned by the above rational fun tion with
S
E
the inputs and
the output.
3.2 Stability ondition
We manipulate
onvolution kernels expressed as rational fun tions where the
oe ient of degree
0
of the denominator is
1.
We shall identify a rational
fun tion with the asso iated formal power series. Using
omplex analysis, we
shall now prove the following lemma, giving the stability
to lter designers:
kQk1 < ∞
unit dis .
Lemma 2.
|z| ≤ 1
ondition familiar
if and only if all the poles of Q are outside of the
That is: a lter is stable in ideal real arithmeti s if and only if all its
poles have module greater than 1.
Proof.
Consider the poles of the rational fun tion
|z| ≤ 1
unit dis , then the radius of
the meromorphi
fun tion
Q
around 0 has a radius of
greater than 1. This implies that the series
and thus that
kQk1
onverges absolutely within the
|z| ≤ 1
If none are in the
onvergen e stri tly
z=1
kQk1 < ∞ then the series
onverges absolutely for
is nite. On the other hand, if
that dis .
Q.
onvergen e of the power series of
unit dis
and no pole
an be within
4 Compositional semanti s: real eld
exa t
Now, we have a se ond look at the basi
to give a pre ise and
the real numbers.
ompositional
semanti s of linear lters, in order
semanti s of
ompound lters on
We show that any linear lter with one input and one
output is equivalent (on the real numbers) to a lter as de ned in 2.2.
9
4.1 De nition
A lter or lter element has
• ni
inputs
I1 ,
• nr
reset state values
...,
In i
( olle tively, ve tor
I ),
ea h of whi h is a
of real numbers;
r1 ,
...,
rnr
R),
( olle tively, ve tor
stream
whi h are the
initial values of the state of the internal state variables of the lter
(inside delay operators) at the beginning of the
• no
If
M
be the
output streams
O1 ,
...,
Ono
omputation;
( olle tively, ve tor
O).
is a matrix (resp. ve tor) of rational fun tions, or series, let
oordinate-wise appli ation of the norm
k·kx
Nx (M )
to ea h rational fun tion,
or series, thereby providing a ve tor (resp. matrix) of nonnegative reals. We
note
mi,j
the element in
R(z)
We note by
M
at line
i
and
olumn
j.
R and by Q[z](z) the
P (z)/(1 − zQ(z)) where P and Q are
the eld of rational fun tions over
ring of rational fun tions of the form
polynomials (that is, the ring of rational fun tions su h that the
term of the denominator is not null).
8 When
L1-norm of the asso iated power series.
When
•
•
omputed upon the real eld, a lter
a matrix
response
F
is
we note
onstant
kF k1
h that
ing) linear response of output stream
F (I, R)
i
di,j
linear
hara terizes the (de ay-
with respe t to reset value
the ve tor of output streams of lter
F
j.
over the reals, on
R.
Then we have
∀I ∈ (RN )ni ∀R ∈ Rnr F (I, R) = T F .I + D F .R
(5)
the ve tor of input streams
I
the
hara terized by:
T F ∈ Mno ,ni (Q[z](z) ) su h that ti,j hara terizes the
of output stream i with respe t to input stream j ;
D F ∈ Mno ,nr (Q[z](z) ) su
a matrix
We note
F ∈ Q[z](z) ,
and the ve tor of reset values
When the number of inputs and outputs is one, and initial values are
assumed to be zero, the
hara terization of the lter is mu h simpler
all matri es and ve tors are s alars (reals, formal power series or rational
fun tions), and
TD
framework on this
is null. We re ommend that the reader instantiates our
ase for better initial understanding.
4.2 Basi arithmeti blo ks
Plus node implemented in oating point type
8
(z)
T = 1 1 , D = [];
This last ring is the lo alization of the ring
generated by
z,
thus the notation.
10
R[z]
f : ni = no = 1,
of real polynomials at the prime ideal
m
n
F
n
unit delay
Figure 2: A feedba k lter
S ale by
k
f : T = k , D = [];
n
ks): T = z , D = 0;
node implemented in oating point type
n lo k ti
T = z ,D= 1 ;
Delay without initializer (delay for
Unit delay with initializer :
4.3 Composition
Parallel
omposition
T
0
T = 1
0 T2
omposition through lter
T = T2 .T1 , D = T2 .D1 D2 .
Serial
,
D1 0
D=
0 D2
;
1, then 2:
4.4 Feedba k loops
Let us
onsider a lter
onsisting of a lter
and feedba k loops running the
delays.
(Fig. 2) We split
TO ∈ Mn,n (Q[z](z) )
TF
n
F
m+n inputs and n outputs
n inputs through unit
es TI ∈ Mn,m (Q[z](z) ) and
with
outputs to the last
into sub-matri
representing respe tively the responses to the global
inputs and to the feedba k loop. The system then veri es the linear equation
over the ve tors of formal power series:
(Idn −
zTOF )O
D F .R.
O = TIF .I + zTOF .P + D.R, and
thus
+
9
F
By Cor. 4, Idn − zTO is invertible in Mn,n (Q[z](z) ), thus T = (Idn −
F
−1
F
F
−1
F
zTO ) .TI and D = (Idn −zTO ) .D . Se tion 8.2 explains how to perform
su h
=
TIF .I
omputations in pra ti e.
4.5 Examples
A se ond order IIR linear lter is expressed by
α2 .delay2 (E) + β1 .delay1 (S) + β2 .delay2 (S).
(α0 + α1 z + α2 z 2 )E + (β1 z + β2 z 2 )S . This
S = (α0 + α1 z + α2 z 2 )(1 − β1 z − β2 z 2 )−1 .E .
9
S = α0 .E + α1 .delay2 (E) +
S =
This yields an equation
equation is easily solved into
This result is not surprising, be ause the system, by
solutions.
11
onstru tion, must admit
ausal
+
E
TF2
TF2
N1
×−k
Figure 3:
A
S
N2
delay (2)
ompound lter
onsisting of two se ond order lters and a
feedba k loop. Ea h TF2 node is a se ond-order lter whose transfer fun tion
is of the form
(α0 + α1 z + α2 z 2 )(1 − β1 z − β2 z 2 )−1 .
In Fig. 3, we rst analyze the two internal se ond order IIR lters separately and obtain
α0 + α1 z + α2 z 2
1 − β1 z − β2 z 2
a0 + a1 z + a 2 z 2
Q2 =
1 − b1 z − b 2 z 2
Q1 =
(6)
(7)
(8)
The we analyze the feedba k loop and obtain for the whole lter a rational
fun tion with a 6th degree dominator:
S=
Q1 .Q2
.E
1 + kz 2 .Q1 .Q2
where Q1 and Q2 are the transfer fun
α1 z + α2 z 2 )(1 − β1 z − β2 z 2 )−1 ), whi h
(9)
tion of the TF2 lters (form
we
(α0 +
omputed earlier.
5 Bounding the 1-norm of series expansions of rational fun tions
5.1 Inverses of produ ts of a ne forms
Let
ξi be
omplex numbers of module stri tly greater than 1. Let
formal power series
The
q (n) ,
n-th
the
degree
Qn
i=1 Qi
where the
oe ient of
oe ient of
zn
qi
is
Qi (z)
(n+1)
−ξi
Q(z) be the
(z − ξi )−1 .
are the power series
, by the easy expansion:
1
−1/ξ
=
z−ξ
1 − z/ξ
in the
Q
(10)
power series, is obtained by su
essive
onvolution produ ts; it is
q (n) =
X
P
∀i,ki ∈N∧ i ki =n
12
Y
(ki )
qi
(11)
We
an therefore bound its module:
q (n) ≤
X
P
∀i,ki ∈N∧ i ki =n
Y
(ki )
qi
(12)
The right hand side of the pre eding inequality is just the
Qn
(n)
(n)
= qi
i=1 Q̇i where q̇i
1
oe ient of the
|ξi |−z series. Sin e |ξ1 | >
q̇ (n) of the series
|ξi |(n+1) is the
=
1,
the
at
z = 1,
that is,
1
|ξi |−1 . We
n-th
order
onvergen e radius of
this last series is stri tly greater than 1; furthermore, all its
nonnegative; therefore, the sum of its
oe ient
oe ients are
oe ients is the value of the fun tion
an therefore give an upper bound:
1
(z − ξ1 ) · · · (z − ξn )
1
≤
1
(|ξ1 | − 1) · · · (|ξn | − 1)
(13)
5.2 Rough and less rough approximation in the general ase
P (z)/Q(z) be a rational fun tion, with P (z) a polynomial of degree
m Q(z) a moni polynomial of degree n. Let zeroes(Q) be the multiset of
zeroes of Q (multiple zeroes are ounted with their multipli ity). P (z) =
P
P
k
k |pk |.kQk1 . Therefore
k pk z Q(z), thus kP k1 ≤
Let
P
Q
This is, however, a very
1
≤Q
kP k1
ξ∈zeroes(Q) (|ξ| − 1)
oarse approximation. Intuitively, the mass of the
onvolution kernel expressed by the
P/Q
series lies in its initial terms. Still,
with the above formula, we totally negle t the
the
(14)
an ellations that happen in
omputation of this initial part of the kernel; i.e. instead of
|a−b|, we bound it by |a|+|b|. The solution is to split kP/Qk1
≥N
and kP/Qk1 . We shall elaborate on this in Se t. 5.5.
into
onsidering
kP/Qk<N
1
5.3 Se ond degree denominators with omplex poles
A
ommon
ase for ltering appli ations is when the denominator is a se ond
degree polynomial
are two
follows:
where
onjugate
Q
of negative dis riminant. In this
omplex numbers
ξ
and
ξ̄
λ = P (ξ)/(ξ − ξ̄).
We shall for now leave
oe ients
ak = −
λ
ξ k+1
13
ak
+
Q
and the de omposition is as
λ
λ̄
P (z)
= P0 (z) +
+
Q(z)
z − ξ z − ξ̄
We are interested in the
ase, the roots of
P0
(15)
out.
of this series:
λ̄
ξ̄ k+1
(16)
Let us write
ak = −
λ = |λ|eiα
and
ξ = |ξ|eiβ ;
then
|λ| iα −i(k+1)β
−iα i(k+1)β
e
.e
+
e
.e
|ξ|k+1
= −2
|λ|
cos (α − (k + 1)β)
|ξ|k+1
(17)
1/|ξ| mod|λ|/|ξ|. ThereIf Q is moni
To summarize, the sequen e is a de reasing exponential of rate
ulated by a sine wave and multiplied by a
onstant fa tor
omputing |λ| and |ξ| will be of prime importan e.
Q(z) = z 2 + z1 x + z0 , then |ξ|2 = ξ ξ¯ = c0 . In the ase of a rational
fore,
of the form
then
with
α0 + α1 z + α2 z 2
P (z)
=
Q(z)
1 − β1 z − β2 z 2
|ξ| = |β2 |−1/2
and
omplex numbers,
λ = P (ξ)/(ξ − ξ̄).
|λ|2 = λλ̄ =
The numerator is a symmetri
fore
are polynomials in the
P,
and
ξ̄ ,
(19)
Q,
Q ; its
roots of
oe ients of
and thereoe ients
therefore the whole polynomial
oe ients of
P
and
Q.
The denominator
α2 2 + β2 −α1 2 − α0 α1 β1 + α0 2 β2 + α2 α1 β1 + α0 β1 2 + 2 β2
−(β1 2 + 4 β2 )
we
an
Q.
We are now interested in bounding
≤ 1,
ξ
polynomial in
be expressed as a polynomial in the
|λ| =
ompute
P (ξ)(ξ̄ − ξ) + P (ξ̄)(ξ − ξ̄)
(ξ − ξ̄)2
oe ients of
is just the dis riminant of
(18)
Should we prefer not to
an be expressed as a polynomial in the
2
fun tion
|ak |.
If we just use
(20)
| cos (α − (k + 1)β) |
ome ba k to the earlier bounds obtained by totally separating the
series arising from the two poles.
We shall now obtain a better bound using the following
any real
| cos θ| =
using the
√
p
(1 + cos(2θ)) /2 ≤ 2−1/2 (1 + cos(2θ)/2)
√
inequality
1 + x ≤ 1 + x/2. Therefore
cos2 θ =
on avity
|ak | ≤
√
Now, we are interested
r<1
onstatation: for
θ,
∞
X
k=0
|λ|
(1 + cos(2(α − (k + 1)β))/2)
|ξ|k+1
P∞
in bounding
k=N [ak |. For any a and b,
(21)
2
cos(a + kb)r k =
cos a − r cos(a − b)
1 − 2r cos b + r 2
14
(22)
and
0≤
(23)
Let us now see the quality of su h bounds
S1 =
∞
X
k=0
S2 =
S1 ≤ S2 , S1 ≤ S3 :
| cos(a + kb)|r k
∞
X
(24)
1
1−r
rk =
k=0
∞
(25)
1 X
(1 + 2 cos(2(a + kb)))r k
S3 = √
2 k=0
1
1 cos(2a) − r cos(2(a − b))
1
+ .
S3 = √
1 − 2r cos b + r 2
2 1−r 2
(26)
(27)
Note that S3 is not ne essarily better than S2 (for a = 0 and b = 0, S3 /S2 =
√
3/(2 2) ≃ 1.06). However, some moderate gains may be obtained ; for
instan e, for r = 0.7, a = 0 and b = 0.3, S1 ≃ 2.60, S2 ≃ 3.33 and S3 ≃ 2.80.
For pra ti al purposes, the bound obtained using S2 is very su ient and
easy to
ompute. We thus opt for this one.
5.4 Finer bounds using partial fra tion de omposition
Q is their
P prime
with Q the fra tion P/Q admits
P
a
as P/Q = P0 +
i Pi /Qi , where P0 is the Eu lidean
quotient of P by Q and the degree of Pi is stri tly less than that of Qi .
Using the fundamental theorem of algebra, it follows that if the ξi are
the distin t roots of Q and mi their multipli ity, then there exist λi,j ∈ C
It is well known that if
Qi
are pairwise prime polynomials, and
partial de omposition
produ t, then for any polynomial
su h that
P/Q = P0 +
mi
XX
i
Sin e
Q
j=1
λi,j
(z − ξi )j
(28)
ξi and
j , λi′ ,j = λ̄i′ , j .
is a real polynomial, its roots are either real, either pairs of
onjugate
ξi′ = ξ̄ , with
the same multipli ity, and also for all
However, while theoreti ally sound, this result is numeri ally deli ate
when there are multiple roots, or dierent roots very
1.3℄ For instan e, let us
degree polynomial
Q,
lose to ea h other.[13,
onsider a rst-degree polynomial
and a se ond-
then
λ1
λ2
P (z)
=
+
Q(z)
z − ξ1 z − ξ2
and we obtain
P
λ1 = P (ξ1 )/(ξ2 − ξ1 )
(and
(29)
λ2 = P (ξ2 )/(ξ1 − ξ2 )). Both
ξ1 − ξ2 . While it is
numbers will get very large, in inverse proportion of
quite improbable that we should analyze lters where two separate poles
have been intentionally be pla ed very
lose together, it is possible that we
analyze lters with multiple poles (for instan e, the
15
omposition of a lter
with itself ), and, with numeri al
lose poles and thus a dramati
We still
an pro eed with a
instead of fa toring
Q
omputations, we would have two extremely
radius r de omposition
numeri al instability.
into a produ t of
z − ξi
of
P/Q
[13, Def 1.3℄:
fa tors, we fa tor it into
a produ t of Qi su h that for any i, and any roots ξ1 and ξ2 of Qi ,
|ξ1 − ξ2 | < 2r . The same referen e des ribes algorithms for performing
then
su h
de ompositions. We obtain a de omposition of the form
X
P
= P0 +
Pi /Qi
Q
(30)
i
where the roots of ea h
the degree of
Pi .
Qi
P
Q
whi h we
We
are
lose together, the degree of
Pi
is less than
From this we obtain the bound
1
≤ kP0 k1 +
X
i
kPi k.
1
Qi
(31)
1
an bound using the inequalities given in the pre eding subse tions.
an, as before, improve on this bound by splitting the series between an
initial sequen e and a tail.
5.5 Development of rational fun tions and normed bounds
P (z)/Q(z) ∈ Q[z](z) be a rational fun tion representing a power series
by its development (un )n∈N around 0. We wish to bound kuk1 , whi h we
shall note kP/Qk1 . As we said before, most of the mass of the development
of P/Q lies in its initial terms, whereas the tail of the series is negligible
(but must be a ounted for for reasons of soundness). We thus split P/Q
into an initial development of N terms and a tail, and use
Let
+ kP/Qk≥N
kP/Qk1 = kP/Qk<N
1
1
kP/Qk1
is
omputed by
velopment of
P/Q.
performing su h a
omputing expli itly the
N
(32)
rst terms of the de-
We shall see in Se t. 8.3 the di ulties involved in
omputation soundly using interval arithmeti s.
Let dQ be the degree of Q. The development D of P/Q yields an equation
P (z) = D(z).Q(z) + R(z).z N . We have P (z)/Q(z) = D(z) + R(z)/Q(z).z N
and thus
kP/Qk≥N
= kR/Qk1 ≤ kRk∞ .k1/Qk1
1
(33)
The pre eding sub-se tions give a variety of methods for bounding
using the zeroes of
Q(z);
bounds on the absolute values of the zeroes of
expli it
ompute
omputation of
D
k1/Qk1
Se tion 5.2 gives a rough method based on lower
R
Q(z). kRk∞
is bounded by
using interval arithmeti s; as we shall see, we
until the sign of the terms is unknown that is, when the norm
of the developed signal is on the same order of magnitude as the numeri al
16
error on it, whi h happens, experimentally, when the terms are small in
absolute values.
Therefore,
approximation used
k1/Qk1
kRk∞
is small, and thus the roughness of the
does not matter mu h in pra ti e.
6 Pre ision properties of xed- or oating-point operations
In this se tion, we shall re all a few fa ts on the errors introdu ed by xedand oating-point arithmeti s. They will be su ient for all our reasonings,
without need for further knowledge about numeri al arithmeti s.
Most types of numeri al arithmeti s, in luding the widely used IEEE-754
oating-point arithmeti , implemented in hardware in all
urrent mi ro om-
puters, de ne the result of elementary operations as follows: if
f
is the ideal
operation (addition, subtra tion, multipli ation, division et .) over the real
numbers and
where
that
r
is a
f˜ is
roundo
the
fun tion. The roundo fun tion
and is very
•
•
•
hooses a
an be exa tly represented in the used xed- or oating-point data type,
lose to
x; spe
i ally, most systems, in luding all IEEE-754 sys-
tems, provide the following roundo fun tions:
•
f˜ = r ◦ f
value r(x)
orresponding oating-point operation, then
round to
of
x
in the dire tion
+∞: r(x) is the representable real nearest to x in the dire
tion
−∞: r(x) is the representable real nearest to x in the dire
tion
+∞;
round to
of
is the representable real nearest to
0;
round to
of
0: r(x)
10
−∞;
round to nearest (generally, the default mode):
sentable real nearest to
r(x)
is the repre-
x.
not-a-number
In this des ription, we leave out the possible generation of spe ial values
su h as in nities (+∞ and
−∞)
ating unde ned results su h as
10
and
0/0.
(NaN), the latter indi-
We assume as a pre ondition to the
On Intel x86 systems, the des ription of the exa t properties of the oating-point
arithmeti s is
ompli ated by the fa t that, by default, with most operating systems and
languages, the 80287- ompatible oating-point unit performs
ing 80-bit long double pre ision numbers, even when the
omputations internally us-
ompiled program suggests the
use of standard 64-bit double pre ision IEEE numbers. Note that su h usage of supplemental pre ision for intermediate
The nal result of the
omputations is allowed by the C standard, for example.
omputation may therefore depend on the register s heduling and
optimizations performed by the
ompiler. Sin e we reason by
are always sound (albeit pessimisti ) in the fa e of su h
piler and the system do.
17
maximal errors, our bounds
ompli ations, whatever the
om-
numeri al lters that we analyze that they are not fed in nities or NaNs indeed, in some DSP (digital signal pro essor) implementations, the hardware
is in apable of generating or using su h values, and in many other implementations the system is
on gured so that the generation or usage of in nities
any
issues an ex eption resulting in bringing the system into a failure mode. Our
framework provides
quantity
x inside
onstru tive methods for bounding
the lters as
kxk∞ ≤ c0 +
Pn
k=1 ck .kek k∞
the input streams of the system; it is quite easy to
does not over ow (kxk
su ient
< M );
one
ek
are
he k that the system
an even easily provide some very wide
onditions on the input (kek k∞
not in lude su h
oating-point
where the
P
≤ (M − c0 )/( nk=1 ck )).
We will
onditions in our des ription, for the sake of simpli ity.
x
and the oating-point result x̃ is bounded, in absolute value, by max(εrel |x|, εabs )
11
(the least positive oating-point number)
where εabs is the
−1074
−324
≃ 4.94 · 10
and
and εrel is the
in urred (εabs = 2
−53
−16
εrel = 2
≃ 1.11 · 10
for IEEE double pre ision operations, for the
For any arithmeti
operation, the dis repan y between the ideal result
absolute error
relative error
worst
ase with respe t to rounding modes). We a tually take the
oarser
inequality
|x − x̃| ≤ εrel |x| + εabs
(34)
See [9℄ for more details on oating-point numbers and [17℄ for more about
the a ne bound on the error.
In the
ase of xed-point arithmeti s, we have
εrel = 0
and
εabs = δ (δ
is
the smallest positive xed-point number) if the rounding mode is unknown
(round to
+∞, −∞
et .) and
δ/2
is it is the rounding mode is known to be
round-to-nearest.
7 Compositional semanti s: xed- and oating-point
In this se tion, we give and a
ompositional
the oating-point numbers.
abstra t
semanti s of lters on
7.1 Constraint on the errors
Our abstra t semanti s
•
•
11
hara terizes a xed- or oating-point lter
the exa t semanti s of the asso iated lter
an abstra tion of the dis repan y
ideal and oating-point lters.
The absolute error results from the
F
ondition:
an ex eption), under ow is generally a benign
ondition.
18
between the
a number
rounded to 0. Contrary to over ow (whi h generates in nities, or is
relying on relative error bounds if one wants to be sound.
by:
over the real numbers
∆(I) = F̃ (I) − F (I)
under ow
F̃
lose to 0 is
on gured to issue
However, it pre ludes merely
We transform
a
F̃ (I)
into the sum of a term that we
urately using algebra and
∆(I)
that we
an bound very
omplex analysis, and a nondeterministi
annot analyze a
urately and soundly without
di ulties, but for whi h bounds are available:
a ne almost linear
onsiderable
assuming for the sake of
simpli ity a single input and a single output and no initialization
we obtain an
to
,
onstraint on the
F .
kIk∞ : k∆(I)k∞ ≤ εFrel kIk∞ + εabs
input
k∆(I)k∞
onditions,
with respe t
In short: sin e the lter is linear, the
magnitude of the error is (almost) linear.
We generalize this idea to the
abstra t semanti s
and
ase of multiple inputs and outputs. The
F
∆ is given by matri es εrel
,T ∈ Mno ,ni (R+ )
no
F
ve tor εabs ∈ R+ su h that
hara terizing
F
εrel
,D ∈ Mno ,nr (R+ )
and a
F
kF (I, R) − F̃ (I, R)k∞ ≤ εFrel,T .N∞ (I) + εrel
,D .N∞ (R) + εabs .
where
F̃ (I, R)
is the output on the stream
numbers on input streams
I
omputed upon the
and initial values
I.
(35)
oating-point
7.2 Basi arithmeti blo ks
Plus node implemented in oating point type
f : ni = no = 1, T = 1 1 ,
i
h
f
f
D = 0, εrel,T = εfrel εrel
, εrel,D = 0, εabs = εabs ;
k node implemented in oating point
f
f
, εrel,D = 0; εabs = εabs ;
εrel,T = |k|.εrel
S ale by
Delay without initializer (delay for
n
f : T = k , D = 0,
type
lo k ti ks):
εrel,T = 0, εrel,D = 0, εabs = 0
Unit delay with initializer :
εabs = 0
T = z n , D = 0,
T = z , D = 1 , εrel,T = 0, εrel,D = 0,
omposition blo k matri es and ve tors:
Parallel
εrel,T
ε1
= rel,T
0
0
2
εrel
,T
,
εrel,D
ε1
0
= rel,D 2
0
εrel,D
,
εabs
ε1
= 2abs
εabs
.
7.3 Serial omposition
The serial
omposition of two lters is more involved.
the ideal linear transfer fun tions of both lters, and
fun tions implemented over oating-point numbers.
We have
for
G).
We
F̃
F and G be
G̃ the transfer
Let
and
mutatis mutandis
∀I N∞ (F (I) − F̃ (I)) ≤ εFrel .N∞ (I) + εFabs (
are interested in ε = N∞ (F (I) − F̃ (I)): that
is, a ve tor of
positive numbers indexed by the outputs of the system su h that on every
oordinate
k, the dieren
e
δ
between output
19
k
omputed over the reals and
the oating-point numbers over the same input
extend
≤
to real ve tors
I
veri es
oordinate-wise.
kδk∞ ≤ εk .
We
The following is easier to understand when ea h lter has a single input
and a single output; then, all ve tors and matri es are s alars (either in
or
R
Nx (v) is simply kvkx .
R of (re)initialization values is split between RF (those onF ) and RG (those on erning G). We split the overall output error
Q[z](z) ,
and
The ve tor
erning
of the system between the part that was introdu ed by the rst lter (and
then ampli ed or attenuated by the se ond lter) and the part that was
introdu ed by the se ond lter, and use the triangle inequality:
N∞ ((G ◦ F )(I, R) − (G̃ ◦ F̃ )(I, R))
≤ N∞ (G ◦ F (I) − G ◦ F̃ (I)) + N∞ (G ◦ F̃ (I) − G̃ ◦ F̃ (I))
G
G
G
G
≤ N1 (G).(F (I) − F̃ (I)) + εrel
,T .N∞ (F̃ (I)) + εrel,D .N∞ (R ) + εabs
G
G
G
≤ N1 (G).(F (I) − F̃ (I)) + εG
rel,T .(N∞ (F̃ (I)) + N∞ (F̃ (I) − F (I))) + εrel,D .N∞ (R ) + εabs
G
G
G
G
≤ (N1 (G) + εG
rel,T ).N∞ (F̃ (I) − F (I)) + εrel,D .N∞ (R ) + εrel .N∞ (F (I)) + εabs
F
F
F
F
≤ (N1 (G) + εG
rel,T ).(εrel,T .N∞ (I) + εrel,D .N∞ (R ) + εabs )
G
G
G
G
+ εrel
,T .N1 (F ).N∞ (I) + εrel,D .N∞ (R ) + εabs
i
h
F
G
G
≤ (N1 (G) + εrel
,T ).εrel,T + εrel,T .N1 (F ) .N∞ (I)
h
i
h
i
F
F
G
G
+ (N1 (G) + εG
rel ).εrel,D .N∞ (R ) + εrel,D .N∞ (R )
i
h
F
G
+ (N1 (G) + εG
rel ).εabs + εabs
Thus
G◦F
εrel
,D
(36)
G◦F = (N (G) + εG ).εF + εG .N (F ),
εrel
1
1
rel
rel
rel
,T
G
F
G◦F = (N (G)+ εG ).εF + εG .
ε
).ε
,
and
εabs
= (N1 (G) + εG
1
rel
rel,D
rel,D
rel
abs
abs
7.4 Feedba k loops
Let us
all
o(n)
the ve tor of outputs of the lter at step
linear fun tion of the
n.
It is, ideally, a
urrent input, the pre eding inputs, and the pre eding
On = L(I≤n , O<N ). Let us all L̃ the asso iated oating-point
tion and Õ the oating-point output of the lter. Let us all ∆ = Õ − O .
outputs.
fun
∆n = L̃(I≤n , Õ<N ) − L(I≤n , O<N )
= L̃(I≤n , Õ<N ) − L(I≤n , Õ<N )) + L(I≤n , Õ<N ) − L(I≤n , O<N )
= L̃(I≤n , Õ<N ) − L(I≤n , Õ<N )) + L(0, ∆<N )
(37)
Cn = L̃(I≤n , Õ<N ) − L(I≤n , Õ<N )) be the sequen e of ve tors of error
F
reations at ea h iteration. Then ∆ veri es the equation ∆ = C + zTO .∆.
F
−1
As before, this means ∆ = (Idn − zTO )
.C and thus that N∞ (∆≤n ) ≤
F
−1
.N∞ (C≤n ).
N1 (Idn − zTO )
F
F
F
Let us split εrel,T ∈ Mn,n+m (R+ ) into εrel,I ∈ Mn,m (R+ ) and εrel,O ∈
Let
20
Mn,n (R+ ).
Then
F
F
F
F
N∞ (C≤n ) ≤ εrel
,I .N∞ (I≤N ) + εrel,O .N∞ (Õ<N ) + εrel,D .N∞ (R) + εabs
F
F
≤ εFrel,I .N∞ (I≤N ) + εrel
−O )
,O .N∞ (O<N ) + εrel,O .N∞ (Õ
| N {z <N}
∆<N
F
F
+ εrel,D .N∞ (R) + εabs
But then, noting
A = N1 (Idn − zTOF )−1
(38)
,
F
N∞ (∆≤n ) ≤ A.(εFrel,I .N∞ (I≤N ) + εrel
,O .N1 (T ).N∞ (I≤N )
F
F
F
+ εrel
,O .N∞ (∆<N ) + εrel,D .N∞ (R) + εabs )
Let
F
K1 = A.εrel
,O ∈ Mn,n (R+ )
(39)
and
F
F
K2 (ι, ρ) = A. εFrel,I + εFrel,O .N1 (T )).ι + εrel
,D .ρ + εabs
(40)
N∞ (∆≤n ) ≤ K1 .N∞ (∆<n )+K2 (N∞ (I), N∞ (R)). This means that the
sequen e un = N∞ (∆<n ) veri es u0 = 0 and un+1 ≤ K1 .un +K2 (N∞ (I), N∞ (R)).
This implies that for all n, un is less than the least xed point L of v 7→
K1 .v + K2 (N∞ (I), N∞ (R)).
Re all that the spe tral radius of a matrix M of real numbers is the
greatest absolute values of its eigenvalues. If K1 is ontra ting (spe tral
radius less than 1), then v 7→ K1 .v + K2 (N∞ (I), N∞ (R)) has a unique
xed point, by Bana h's xed point theorem; and 1 − K1 is invertible.
−1 K (N (I), N (R)). Let ε
This xed point is v = (1 − K1 )
rel,T = (1 −
2
∞
∞
Then
F
F
K1 )−1 .A. εrel
,I + εrel,O .N1 (T )
F .
K1 )−1 .A.εabs
,
εrel,D = (1 − K1 )−1 .εFrel,D ,
εabs = (1 −
and
N∞ (∆) ≤ εrel,T .N∞ (I) + εrel,D .N∞ (R) + εabs .
K1 = A.εFrel,O ∈ Mn,n (R+ ) where A is the matrix of norms
N1 (Idn − zTOF )−1 ; K1 bounds the amount of oating-point impre ision
that feeds ba k into the system. A is the ampli ation bounding matrix
Then
Re all that
of the lter
onsisting merely of the feedba k loop of the original lter; if
the original lter is stable and well-designed, the
F
moderate. εrel,O measures the
oe ients of
A
reation of impre ision in one iteration of the
internal lter; if the lter is numeri ally well-designed, then its
are very small. On real-world examples,
of
should be
10−15 .
K1
oe ients
was on the order of magnitude
This suggests an ee tive method for bounding from above the various
quantities of the form
(if
y
(1−K1 )−1 .y
that we listed, where
is a matrix, then split it into its
y
is a
olumn ve tor
olumn ve tors).
d∞ = (1 − K1 )−1 .y =
21
∞
X
k=0
K1k .y
(41)
is the unique xpoint of
φ = x 7→ K1 .x + y ,
whi h is monotoni
tra ting. Consider the matrix norm subordinate to
kK1 k = sup
i
This gives a rough bound on
kd∞ k∞ ≤
and
on-
on ve tors:
k1i,j
(42)
j
d∞ :
∞
X
k=0
kK1 kk .kyk∞ =
Pn
dn = (x 7→ K1 .x + y)n (y) =
Let
X
k · k∞
n
k=0 K1 .y .
kd∞ − dn k∞ ≤
kyk∞
.
1 − kK1 k
(43)
d∞ − dn = K1n+1 .d∞ ,
thus
K1n+1
.kyk∞ .
1 − kK1 k
(44)
d∞ :
.kyk∞ .V1
Therefore, the following is an upper bound on
B = dn + −
where
V1
K1n+1
kK1 k − 1
is a ve tor of ones of the same dimension as
(45)
y.
This
may be ee tively performed in oating-point arithmeti
omputation
in order to yield
a sound upper bound by
omputing Eqn. 42 and 45 in round-to-+∞ mode
(x
Remark that we
of
7→ −1/x is monotoni ).
the resulting B̃ by he
king that
an dire tly prove the soundness
K1 .B̃ + y
is less than
B̃
(this
he k-
ing phase, though unne essary assuming a sound implementation, may be
heaply performed for the sake of se urity; while it is possible that the result
should be
and
orre t and the
he k fails, this seems very unlikely in pra ti e,
an be worked around by
hoosing a slightly larger
B̃ ).
7.5 Trading some a ura y for omputation speed; nonlinear
elements
We have split the behavior of the lter into the sum of the
onvolution of
the input signal by the power development of a rational fun tion, representing the exa t behavior, and some error term.
fun tions exa tly over
pensively large.
Q[z](z) ,
then the rational
If we
ompute the rational
oe ients might grow ex-
It seems silly to use high pre ision for the
oe ients of
a system parameterized by oating-point numbers and implemented with
oating-point errors. Indeed, we may redu e the pre ision of the
oe ients
of the rational fun tion at the expense of adding to the margin of error.
An ideal lter of Z-transform the rational fun tion
P (z) =
Pdp
k
k=0 pk z and
Q(z) =
Pdq
is equivalent to a lter with ideal input Z-transform
Z-transform
Q
(Fig 4).
P (z)/(1−Q(z)) where
k
k=1 qk z with non initialization
P
ondition
and ideal feedba k
Su h a lter may be soundly approximated by a
22
P
+
delay
Q
Figure 4: An ideal lter equivalent to a lter of Z-transform
♯
P (z)/(1−Q(z)).
♯
F ♯ with TIF = P ♯ , TOF = Q♯ , εrel,I = kP ♯ − P k1 ,
♯
εrel,I = kQ − Qk1 , εabs = 0, whi h we know how to solve from Se t. 7.4.
♯
More generally: a lter F may be approximated by a lter F with
♯
♯
♯
F
G
F
G
G
F
F
F
transfer fun tion T
= T , εrel,T = εrel,T + εrel,T , εrel,D = εrel,D + εrel
,D ,
non-ideal feedba k lter
♯
F = εF where G is the feedba k lter with internal lter H
εabs
abs
H
H
H
♯
♯
P ♯ , TOH = Q♯ , εrel
,I = kP − P k1 , εrel,I = kQ − Qk1 , εabs = 0.
Note that this gives a generi
ements o
given
TIH =
method for approximating non-linear el-
uring in lters, provided that it is possible to split them into a
linear part and a nonlinear part, the output of whi h
an be bounded by an
a ne fun tion of bounds on the absolute value of the inputs.
8 Numeri al onsiderations
We have so far given many mathemati al formulas that are exa t in the
real
eld. In this se tion, we explain how to obtain sound abstra tions for these
formulas using oating-point arithmeti s.
8.1 Interval arithmeti s
upward rounded rounded-
IEEE oating-point arithmeti s [9℄ and good extended pre ision libraries
to-+∞
downward rounded rounded-to-˜−∞
su h as MPFR [7℄ provide fun tions
omputing
(or
f (x1 , . . . , xn )
f˜+ are the asso iated
oating-point downward and upward operations, then f (x1 , . . . , xn ) is guar−
+
anteed to be in the interval [f˜ (x1 , . . . , xn ), f˜ (x1 , . . . , xn )], whi h will guar) and
(or
is the exa t operation on real numbers and
antee the
soundness˜
f˜− (x1 , . . . , xn )
better bounds
will guarantee
of our approa h.
f−
) results: that is, if
and
Furthermore, for many operations,
+
and f (x1 , . . . , xn ) are guaranteed to be optimal; that is, no
optimality
an be provided within the desired oating-point format; this
of
ertain of our elementary operations.
23
8.2 Approximate algebrai
omputations
In many o
P asions, we ideally would like to ompute on real polynomiPP
= nk=1 pk z k but instead we ompute on oating-point polynomials
n
k
P̃ =
k=1 [lk , hk ]z abstra ting the set γ(P̃ ) of polynomials P su h that
∀k pk ∈ [lk , hk ]. In pra ti e, it will often be ne essary that 0 ∈
/ [lk , hk ]
als
in order to avoid un ertainties on the degree of the polynomial.
All the
usual polynomial operations (addition, multipli ation by a s alar, subtra tion, multipli ation) may be abstra ted using interval arithmeti s. We also
in lude a test
ontains0 (P̃ ) whether the null polynomial is in
this stru ture an
abstra t ring
Given a abstra t ring
R,
.
we
onstru t the
abstra t eld
γ(P̃ ).
We
all
of fra tions over
p1 /q1 +p2 /q2 = (p1 q2 +p2 q1 )/(q1 q2 ),
k.(p/q) = (k.p)/q , (p1 /q1 ).(p2 /q2 ) = (p1 .p2 )/(q1 .q2 ), (p1 /q1 )/(p2 /q2 ) = (p1 .q2 )
/(q1 .p2 ), ontains0 (p/q) = ontains0 (p). We an make a simple attempt at
that ring using the following operations:
redu ing the fra tions by
he king that there are no trivial
an ellations
abstra t ring of matri es
between the numerator and denominator in produ ts and quotients.
K,
Given an abstra t ring
we
onstru t the
over that ring with the usual operations: if
if
M = A.B , mi,j =
P
k
ai,k .bk,j .
If
K
is an abstra t eld, we
implement Gaussian elimination in order to
matrix
A
and a matrix
su h that
B.
M = A + B , mi,j = ai,j + bi,j ;
ompute
A−1 .B
an also
given a square
When we look for a pivot, we sele t elements
e
ontains0 (e) is false.
Unfortunately,
omputations on su h approximate stru tures may yield
unfavorable results. In parti ular, the absen e of simpli ation between the
numerator and denominator may yield fra tions
have some
P̃ (z)/Q̃(z)
where
P̃
and
Q̃
ommon zeroes. The spurious poles that are introdu ed not be
that mu h of a problem if we use partial fra tion de omposition (Se t. 5.4),
for they will yield very small
will make the
oe ients in the de omposition; however, they
omputations more
of Se t. 5.1, the results may be
A solution is to perform all
over
Q[z](z) .
Then,
omplex. If using the simple tail bounds
onsiderably worse.
omputations on rational fun tions exa tly
an ellation between a numerator and a denominator
be performed exa tly by division by their greatest
an
ommon divisor, whi h is
obtained from Eu lide's algorithm over the Eu lidean division of polynomials. No spurious poles may be introdu ed. However, on large lter networks,
exa t
omputations may produ e ex eedingly large integer numerators and
denominators.
It is then possible to apply the approximation s heme of
Se t. 7.5 in order to trade speed for potential pre ision. This is the solution
that we implemented in our system: exa t
omputations on rational num-
bers and safe approximations to limit the length of the numbers involved in
the
omputations.
24
8.3 Computation of developments
kP/Qk1 of a series quotient of two polynomials,
N initial terms of development, whi h we ompute
When bounding the norm
we split the series into its
expli itly, and a tail whose norm we bound. The rst idea is to
N
ompute the
rst terms of the series by quotienting the series, as explained in Se t. 3.1
or, equivalently, by running the lter for
1, 0, 0, . . . .
N
iterations on the Dira
input
In order to provide a sound result, one would work using interval
arithmeti s over oating-point numbers. However, as already noted by Feret,
after some number of iterations the sign of the terms be omes unknown and
then the magnitude of the terms in rease fast; it is therefore indi ated to
ompute the development until the rst term of unknown sign is rea hed,
N a ordingly (one may still also enfor e a maximal number of
Nmax ). In order to be able to develop the quotient further with
and assign
iterations
good pre ision, one
an use a library of extended-pre ision oating-point
omputations with sele table rounding dire tion, su h as the MPFR library
now part of GNU MP [7℄.
8.4 Bounding the roots
In order to bound
interval
kP/Qk1 ,
where
P
Q
and
may possibly be given using
oe ients, we have to bound the roots of
Q.
More formally, we
P̃ (z) =
k su h that 0 ∈
/
[l
,
h
]
,
nd
a
family
(
ξ̃
,
ρ
)
(
[l
;
h
]z
n n
k k 1≤k≤n ξk ∈ C
k=1 k k
with ℜξk and ℑξk oating-point numbers, ρk ∈ R+ a oating-point number)
Pn
su h that for any polynomial P =
k=1 pk su h that ∀k pk ∈ [lk , hk ], then,
up to a permutation, the n roots (ξk )1≤k≤n of P are su h that ξk ∈ D(ξ̃k , ρk )
where D(z, r) is the losed dis of enter z and radius r .
Often, what we need is a tually bounds on the |ξk |; this an easily be
have to solve the following problem: given an interval polynomial
Pn
obtained from the pre eding bounds using interval arithmeti
on plus, minus,
multiply and square root.
Our
oe ients are intervals
errors of oating-point
hk − lk
[lk , hk ]
in order to a
omputations. As a
ommodate possible
onsequen e, it is expe ted that
are small. This suggests to us a two-step method for obtaining the
desired bounds:
approximations x
1. Use an e ient and, in pra ti e, very a
j to the roots of
nomial).
Pn
urate algorithm to obtain
lk +hk k
k=1
2 z (the midpoint poly-
2. From those approximations, obtain bounds on the radius of the error
ommitted.
There exist a variety of methods and implementations to perform the rst
point. We used
gsl_poly_ omplex_solve of the GNU S
whi h is based on an eigenvalue de omposition of the
25
ienti
Library [8℄,
ompanion matrix.
For the se ond step, Rump des ribes a variety of bounding methods [22℄
whi h take a polynomial and approximate roots as an input and output
error radii; these methods may be performed using interval arithmeti s. We
ξj
implemented the simplest and roughest one:
xj − p j
and radius
|pj |
is in a
losed dis
of
enter
where
pj =
nP (xj )
Q
,
pn k6=j xj − xk
(46)
whi h is easily implemented using interval arithmeti s (P be omes
P̃
et .).
9 Implementation and ase studies
We implemented the algorithms des ribed above in a simple Obje tive Caml
[15℄ program: lters are represented by a re ord of all their
hara teristi s
(transfer matri es, bounding matri es); fun tions (in the OCaml) sense
stru t lter re ords, or perform
The formal
omposition operations.
omputations on fra tions are performed over
using GNU MP's
mpq
on-
type [7℄. We initially
Q, implemented
onsidered using MPFR [10℄, an
extended pre ision library with sound rounding modes, for interval
omputa-
tions; instead, we simply use the IEEE-754 rounding modes of the hardware
oating-point unit, whi h is mu h faster.
9.1 Composition of TF2 lters
Let us re all the example of Se t. 4.5. It is a
omposition of two TF2 lters
with a feedba k loop around it. The serial
omposition of the lter in Fig. 3
and another TF2 lter, all with realisti
oe ients, is analyzed in about
0.10 s on a re ent PC; the analyzer nds that
εrel ≃
10−12 and
εabs ≃
10−305 .
kSk ≤ gkEk
with
g ≃ 2,
with
The power series developments of rational fun tions (Se t. 8.3) are done
up to around the 27th order.
9.2 Complex nonlinear iterated lter
We now
onsider a nonlinear, iterated lter due to Roozbehani et al. [21℄[5℄.
We rst analyze separately
filter1() (2nd-order linear lter) and filter2()
(2nd-order a ne lter). So as to simplify matters, we do not give the transfer fun tions using matri es, matri es inverses et .
but as the solution of
z.
We obtain that system
a system of linear equations over polynomials in
very simply from the program: whenever we see an assignment
turn it into an equation
x=e
x := e,
we
(we assume without loss of generalities that
variables are only assigned on e in a single iteration step), where
original expression where a variable
v
e
is the
that has not yet been assigned in the
26
urrent iteration is repla ed by
value of
iv + z.v , iv
standing for the initialization
v.
void filter1 () {
stati float E[2℄, S[2℄;
if (INIT1) {
S[0℄ = X; P = X;
E[0℄ = X; E[1℄=0; S[1℄=0;
} else {
P =0.5*X-0.7*E[0℄ +0.4*E[1℄ p = 0.5e − 0.7(ie0 + z.e0 )
+1.5*S[0℄-S[1℄*0.7;
+0.4(ie1 + z.e1 )
}
}
E[1℄ = E[0℄;
E[0℄ = X;
S[1℄ = S[0℄;
S[0℄ = P;
X=P/6+S[1℄/5;
+1.5(is0 + z.s0 ) − 0.7(is1 + z.s1 )
e1 = ie0 + z.e0
e0 = e
s1 = is1 + z.e1
s0 = p
x = p/6 + s1 /5
We all e the input value for X. We solve the system and obtain x =
Q.e + Qie0 .ie0 + Qie1 .ie1 + Qis0 .is0 + Qis1 .is1 . The ommon denominator of
2
the Q fra tions is 10 − 15z + 7z , whi h has omplex onjugate roots z su h
that |z| ≃ 1.2. ie1 = is1 = 0 and ie0 = is0 = ι (the last value for input
e su h that INIT1 is true), thus kxk∞ ≤ kQk1 .kek∞ + kQie0 + Qis0 k∞ .kιk.
With a pre ondition kek∞ ≤ 400, this yields kxk∞ < 339. If we take the
oarser inequality kxk∞ ≤ kQk1 .kek∞ + (kQie k∞ + kQis k∞ ).kιk we get
0
0
kxk∞ < 528. Roozbehani et al. nd a bound ≃ 531.
void filter2 () {
stati float E2[2℄, S2[2℄;
if (INIT2) {
S2[0℄ =0.5*X; P = X;
E2[0℄ = 0.8*X; E2[1℄=0; S2[1℄=0;
} else {
P =0.3*X-E2[0℄*0.2+E2[1℄*1.4 p = 0.3e − 0.2(ie0 + z.e0 )
+S2[0℄*0.5-S2[1℄*1.7;
+1.4(ie1 + z.e1 )
}
}
E2[1℄ = 0.5*E2[0℄;
E2[0℄ = 2*X;
S2[1℄ = S2[0℄+10;
S2[0℄ = P/2+S2[1℄/3;
X=P/8+S2[1℄/10;
+0.5(is0 + z.s0 ) + 1.7(is1 + z.s1 )
e1 = 0.5(ie0 + z.e0 )
e0 = 2e
s1 = is0 + z.s0 + τ
s0 = p/2 + s1 /3
x = p/8 + s1 /10
τ = 10/(1 − z) and
x = Q.e + Qie0 .ie0 + Qie1 .ie1 + Qis0 .is0 + Qis1 .is1 + Qc . The ommon
We pro eed similarily (with the introdu tion of
obtain
27
60 + 35z + 51z 2 , with omplex onjugate roots z
su h that |z| ≃ 1.08. Then kxk∞ ≤ kQk1 .kek∞ + k0.8Qie + 0.5Qis k∞ .kιk +
0
0
kQc k∞ . This yields kxk∞ ≤ 1105.
denominator of the
Q
is
The two linear lters are
ombined into an iterated nonlinear lter.
filter1() (resp. filter2()) is run with
(resp. [−800, 800]). We repla e the all
X ∈ [−339, 339] (resp. X ∈ [−1105, 1105]).
a pre- ondition of
X ∈ [−400, 400]
to the lter by its post ondition
void main () {
X = 0;
INIT1 = TRUE; INIT2=TRUE;
while (TRUE) {
X = 0.98 * X + 85;
if (abs(X)<= 400) {
filter1 ();
X=X+100;
INIT1=FALSE;
} else
if (abs(X)<=800) {
filter2();
X=X-50;
INIT2=FALSE;
}
}}
The program then
an be abstra ted into:
while (TRUE) {
X = 0.98 * X + 85;
maybe hoose X in [−1155, 1055];
}
We obtain X ∈ [−1155, 4250.02] by
running Astrée with a large number
of narrowing iterations, whereas Astrée
pre isely and
(x
annot bound
= 0.98x+ 85
X.
In this
has for unique solution
algebrai ally, but in more
annot analyze the original program
ase, the exa t solution
x = 4250)
ase.
4560.
Note that the non-abstra ted program
for
omputed
omplex lters this would not have been the
Roozbehani et al. have a bound of
X ∈ [0, 209].
[−1155, 4250]
ould have been
onverges to a value
≃ 205,
with
However, this very simple program illustrates our methodology
ompositional analysis: nding the optimal solution is possible here be-
ause the program is simple, but would not be possible in pra ti e if we had
added more nonlinear behavior and nondeterministi
rea tive
inputs, as in real-life
ode; whereas by analyzing pre isely ea h linear lter and plugging
the results ba k into a generi
analyzer, we get reasonable results.
28
10 Related works
In the eld of digital signal pro essing, some sizable literature has been devoted to the study of the ee ts of xed-point and oating-point errors on
numeri al lters. In the area of xed-point
omputation, bounds on the sizes
of the various operands are of paramount importan e: operands that leave
the pres ribed range will undergo saturation and the output signal will be
distorted. For these reasons, operands are s aled so as not to produ e digital
saturation; yet, the s ale fa tor should be made large enough that rounding
errors are very small
ompared to the typi al magnitude of the signal. While
the fa t that the l1-norm of the
onvolution kernel is what matters for judg-
ing over ow, it is argued that this norm is overly pessimisti [12, 11.3℄
[11, eq 13℄, not to mention the di ulties in estimating it. In pra ti e, lter
designers have preferred
riteria that indi ate no saturation for most om-
monpla e inputs, ex luding pathologi al inputs. Our vision is dierent: our
results must be sound in all
ir umstan es, even pathologi al inputs.
The impa t of xed- and oating-point errors in digital lters was
las-
si ally studied from by modeling the errors as random sour es of known
distribution, independent of ea h other and with no temporal
(i.e.
orrelations between su
in reality, false: the
random; the
orrelation
essive values) [3, 20℄. These assumptions are,
omputational pro ess is fully deterministi , and not
omputations are generally interdependent (all
omputations
inside a lter depend on the past of the input variables); and there are temporal
orrelations. However,
ir uit designers are
on erned with the spe tral
distribution of output noise [11℄, and optimization of hardware or software
implementations with respe t to this noise, and these tools are adequate
for this. On the other hand, we merely aim at providing sound bounds for
the outputs of the system, but the bounds that we provide must be sound
without any extra and unfounded suppositions.
J. Feret has proposed an abstra t domain for analyzing programs
prising digital linear lters [6℄.
se ond degree lters. In
works, in a
om-
He provides ee tive bounds for rst and
omparison, we
onsider more
omplex lter net-
ompositional fashion; but we analyze spe i ations, and not C
ode (whi h is usually
loss of stru ture).
ompiled from those spe i ations, with
onsiderable
Another dieren e is that we do not perform abstra t
iterations. Feret's method
urrently
onsiders only se ond-order lters (i.e.
TF2), though it may be possible to adapt it to higher-order lters.
se ond-order lters, the bounds
in this paper are very
On
omputed by Feret's method and the method
lose (sin e both are based on a development of the
onvolution kernel, though they use dierent methods of tail estimation).
Lamb et al. [14℄ have proposed ee tive methods, based on linear algebra,
for
omputing equivalent lters for DSP optimization. They do not
ompute
bounds, nor do they study oating-point errors.
Roozbehani et al. [21℄ nd program invariants by Lagrangian relaxation
29
and semide nite programming, with quadrati
invariants. In order to make
problems tra table, they too apply a blo kwise abstra tion. The
lass of pro-
grams that they may analyze dire tly is potentially larger, but the results
are less pre ise than our method on some linear lters. They do not handle oating-point impre isions (though this
an perhaps be added to their
framework).
One possible appli ation of our method would be to integrate it as a
pre-analysis pass of a tool su h as Astrée [5℄. Astrée
omputes bounds on all
oating-point variables inside the analyzed program, in order to prove the
absen e of errors su h as over ow.
reasonably a
by-wire
In order to do so, it needs to
ompute
urate bounds on the behavior of linear lters. A typi al y-
ontroller
ontains dozens of TF2 lters, some of whi h may be
integrated into more
omplex feedba k loops; in some
of the lters may yield too
ases, separate analysis
oarse bounds.
11 Con lusions and future works
We have proposed ee tive methods for providing sound bounds on the outome of
omplex linear lters from their ow-diagram spe i ations, as found
in many appli ations. Computation times are modest; furthermore, the nature of the results of the analysis may be used for modular analyses the
analysis results of a sub- lter
the sub- lter
an be stored and never be re omputed until
hanges.
The usefulness of these methods is twofold. First, they
implemented in the graphi al user interfa e for designing
then be able to
a
ould be dire tly
ir uits. Users may
ompute gains or to
he k the stability of lters, taking into
ount oating-point errors (whi h
onventional Z-transform te hniques do
not
onsider). Se ond, they
stati
an be used as a way to automati ally obtain
analysis transformers or transfer fun tions: a stati
analysis tool
su h as Astrée may dete t that some program sequen e implements su h or
su h
omplex linear lter, and apply some invariant relation
omputed using
the te hniques in that paper.
In future works, we will examine the
positional, modular analysis.
non-linear lters
•
ase of non-linear lters and
The analysis of a
an be done in two ways or a
ombination of linear and
ombination thereof:
the overall behavior of a nonlinear lter may be
input-output relationship su h as
om-
onstrained by some
kOk∞ ≤ (1 + ǫ)kIk∞
rate limiter), and this input-output relationship
(example of a
an be integrated into
the abstra t semanti s as in Part 7;
•
the overall behavior of a linear lter
this bound information
an be pre isely bounded, and
an be fed into an analysis of a larger nonlinear
30
lter, su h as one based on stati ally
omputed relationships between
intervals [19℄
Referen es
[1℄ B. Blan het, P. Cousot, R. Cousot, J. Feret, L. Mauborgne, A. Miné,
D. Monniaux, and X. Rival. Design and implementation of a spe ial-
The Essen e of Computation: Complexity, Analysis, Trans-
purpose stati
formation
software. In
program analyzer for safety- riti al real-time embedded
, number 2566 in Le ture Notes in Computer S ien e, pages
85108. Springer Verlag, 2002.
[2℄ B. Blan het, P. Cousot, R. Cousot, J. Feret, L. Mauborgne, A. Miné,
PLDI
D. Monniaux, and X. Rival. A stati
software. In
analyzer for large safety- riti al
, pages 196207. ACM, 2003.
[3℄ Bru e W. Bomar, L. Montgomery Smith, and Roy D. Joseph. Round-
IEEE Trans. on Cir uits and Systems
o noise analysis of state-spa e digital lters implemented on oating-
II
point digital signal pro essors.
, 44(11):952955, 1997.
POPL '87:
Pro eedings of the 14th ACM SIGACT-SIGPLAN symposium on Priniples of programming languages
[4℄ Paul Caspi, Daniel Pilaud, Ni olas Halbwa hs, and John A. Plai e. LUSTRE: a de larative language for real-time programming. In
, pages 178188. ACM Press, 1987.
[5℄ Patri k Cousot, Radhia Cousot, Jér me Feret, Laurent Mauborgne, An-
ESOP
toine Miné, David Monniaux, and Xavier Rival. The ASTRÉE analyzer.
In
, number 3444 in Le ture Notes in Computer S ien e, pages
2130, 2005.
[6℄ Jér me Feret. Stati
analysis of digital lters. In
ESOP '04
, number
2986 in Le ture Notes in Computer S ien e. Springer-Verlag, 2004.
library
[7℄ Free Software Foundation.
, 2004.
GMP GNU multiple pre ision arithmeti
GSL GNU s ienti library
Standard for Binary Floating-Point Arithmeti
The MPFR Library
[8℄ Free Software Foundation.
[9℄ IEEE.
[10℄ INRIA et al.
, 2004.
. standard 754.
.
The Bell System Te hni al J.
[11℄ Leland B. Ja kson. On the intera tion of roundo noise and dynami
range in digital lters.
February 1970.
, 49(2):159184,
31
[12℄ Leland B. Ja kson.
Digital Filters and Signal Pro essing
. Kluwer, 1989.
Newton iteration for fa torization in
1998.
C[z].
C(z) and simultaneous
J. Complexity
[13℄ Peter Kirrinnis. Partial fra tion de ompostion in
, 14(3):378444,
PLDI '03
[14℄ Andrew A. Lamb, William Thies, and Saman Amarasinghe.
analysis and optimization of stream programs. In
25. ACM, 2003.
guide
[15℄ Xavier Leroy.
. INRIA.
Linear
, pages 12
The Obje tive Caml system, do umentation and user's
[16℄ Ja ques-Louis Lions et al.
Ariane 501: Flight 501 failure.
Te hni al
report, ESA / CNES, 1996. Available on WWW.
[17℄ A. Miné.
ESOP'04
LNCS
Relational abstra t domains for the dete tion of oating-
point run-time errors. In
Springer, 2004.
[18℄ David Monniaux.
meri al lters. In
, volume 2986 of
, pages 317.
Computer-aided veri ation: CAV '05
Compositional analysis of oating-point linear nu, number 3576
in Le ture Notes in Computer S ien e, pages 199212. Springer Verlag,
2005.
Stati analysis symposium (SAS)
[19℄ David Monniaux.
Optimal abstra tion on real-valued programs.
, 2007. To appear.
Trans. on Signal Pro essing
[20℄ Bhaskar D. Rao.
Floating point arithmeti
and digital lters.
, 40(1):8595, January 1992.
[21℄ M. Roozbehani, E. Feron, and A. Megretski.
and
omputation for software veri ation.
In
LNCS, page 606. Springer, 2005.
HSCC
In
IEEE
Modeling, optimization
, number 3414 in
J. of Computational and Applied Math.
[22℄ Siegfried M. Rump. Ten methods to bound multiple roots of polynomials.
, 156(2):403432, 2003.
M , let us note minori,j (M ) the determinant of the matrix
i and olumn j from M . We re all that for any
dimension n
n
X
(−1)n−1 mi,j .minor1,j (M )
det(M ) =
(47)
For any matrix
obtained by removing line
matrix
M
of
j=1
and that the determinant is
n-linear.
Re all that for any matrix
M
of in-
vertible determinant,
M −1 = det(M )−1 .
32
minori,j (M )
t
(48)
Id
Proof.
det(
If
, then there exists B ∈ Q[z] su h that
A ∈ Mn,n (Q[z](z) )
−
zA)
=
1 − zB
n
Lemma 3.
.
Proof by indu tion on
onsider
n.
(z)
The
ase
n = 1
is trivial.
Now let us
n > 1.
det(Idn − zA)
= (1 − za1,1 )minor1,1 (Idn − zA) +
= minor1,1 (Idn − zA) + z
n
X
j=2
n
X
j=1
(−1)n za1,j minor1,j (Idn − zA)
(−1)n za1,j minor1,j (Idn − zA)
(49)
The result follows by the appli ation of the indu tion hypothesis, and the
fa t that
BQ[z](z)
is a ring and thus the determinant of any matrix over that
ring is itself in the ring.
.
Corollary 4.
Mn,n (Q[z](z) )
Proof.
If A ∈ M
, then Id
n,n (Q[z](z) )
n
− zA)
has an inverse in
det(Idn −zA) is of the form 1−zP (z)/Q(z),
where P and Q are polynomials su h that the onstant oe ient of Q is
1, therefore (det(Idn − zA))−1 = Q(z)/(Q(z) − zP (z) is in Q[z](z) . All the
minori,j (Idn − zA) are elements of Q[z](z) , the result follows by applying
By the pre eding lemma,
Equ. 11.
33
| 6cs.PL |
1
Bio-Inspired Local Information-Based Control for
Probabilistic Swarm Distribution Guidance
arXiv:1711.06869v1 [cs.MA] 18 Nov 2017
Inmo Jang, Hyo-Sang Shin, and Antonios Tsourdos
Abstract—This paper addresses a task allocation problem for
a large-scale robotic swarm, namely swarm distribution guidance
problem. Unlike most of the existing frameworks handling this
problem, the proposed framework suggests utilising local information available to generate its time-varying stochastic policies.
As each agent requires only local consistency on information
with neighbouring agents, rather than the global consistency, the
proposed framework offers various advantages, e.g., a shorter
timescale for using new information and potential to incorporate
an asynchronous decision-making process. We perform theoretical analysis on the properties of the proposed framework. From
the analysis, it is proved that the framework can guarantee the
convergence to the desired density distribution even using local
information while maintaining advantages of global-informationbased approaches. The design requirements for these advantages
are explicitly listed in this paper. This paper also provides specific
examples of how to implement the framework developed. The
results of numerical experiments confirm the effectiveness and
comparability of the proposed framework, compared with the
global-information-based framework.
Index Terms—Swarm robotics, Distributed robot systems,
Networked robots, Markov chains.
I. I NTRODUCTION
HIS paper addresses a task allocation problem for a largescale multiple-robot system, called a robotic swarm.
Robotic swarms have attracted lots of attention because they
are regarded as promising solutions to handle complicated
missions that other systems may not be able to manage [1], [2].
Agents in a swarm are assumed to be homogeneous because
the swarm is usually realised through mass production [3]. In
this context, the task allocation problem can be reduced to a
problem of how to distribute a swarm of agents into given
tasks (or bins), satisfying the desired population fraction (or
swarm density) for each task. This problem is known as the
swarm distribution guidance problem [4]–[6].
For a large number of agents, probabilistic approaches based
on Markov chains [4]–[13] or differential equations [14]–[18]
have been widely utilised. Since these approaches focus not on
individual agents but instead on the ensemble dynamics, they
are also called Eulerian [11]–[13] or macroscopic frameworks
[18], [19]. In these approaches, swarm densities for each bin
are represented as system states, and a state-transition matrix
describes stochastic decision policies, i.e., the probabilities
that agents in a bin switch to another. Individual agents in
the swarm make decisions based on these policies, but in a
random, independent, and memoryless manner.
T
Inmo Jang, Hyo-Sang Shin, and Antonios Tsourdos are with the Centre for
Autonomous and Cyber-Physical Systems, Cranfield University, MK43 0AL,
United Kingdom (e-mail: inmo.jang@cranfield.ac.uk; h.shin@cranfield.ac.uk;
a.tsourdos@cranfield.ac.uk).
Initially, open-loop-type frameworks have been proposed
[4]–[7], [16]–[18]. Agents under these frameworks are controlled by time-invariant stochastic decision policies. The policies, which make a swarm converge to a desired distribution,
are pre-determined by a central controller and broadcasted
to each agent before executing the mission. Communication
between agents is hardly required during the mission, so
that it can reduce communication complexity under these
frameworks. However, the agents only have to follow the given
policies without incorporating any feedbacks, and thus there
still remain some agents who unnecessarily and continuously
switch bins even after the swarm reaches the desired distribution. This gives rise to a trade-off between convergence rate
and long-term system efficiency [17].
There have been also some other works, called closed-looptype frameworks [9]–[15]. This type of frameworks allows
agents to adaptively construct their own stochastic decision
policies at the expense of sensing the concurrent swarm
status through interactions with other agents. Based on such
information, agents can synthesise time-inhomogeneous transition matrices to achieve certain objectives and requirements:
for example, maximising convergence rates [10], minimising
travelling costs [13], and temporarily adjusting given policies
when bins are more overpopulated or underpopulated than
certain levels [14], [15]. In particular, Bandyopadhyay et. al.
[13] recently proposed a closed-loop-type algorithm that exhibits faster convergence as well as less undesirable transition
behaviours, compared with an open-loop-type algorithm. This
algorithm is expected to mitigate the trade-off raised in openloop-type frameworks.
To the best of our knowledge, most of the existing closedloop-type algorithms are based on Global Information Consistency Assumption (GICA) [20]. GICA implies that necessary
information is required to be consistently known by entire
agents. We refer to such information as global, because
achieving information consistency needs agents to somehow
interact with all the others through a multi-hop fashion and
thus it “happens on a global communication timescale” [20].
This paper proposes a framework that requires Local Information Consistency Assumption (LICA) [20]. Unlike GICAbased algorithms, the proposed framework require only local
consistency on information with neighbouring agents, not
the global consistency. LICA can provide various alternative
advantages to the proposed framework, compared with GICA.
Firstly, it “provides a much shorter timescale for using new
information because agents are not required to ensure that this
information has propagated to the entire team before using it”
[20]. Secondly, LICA enables a foundation on which an asyn-
2
chronous decentralised decision-making process can be developed. Note that the timescales for achieving the information
consistency between the agents can be different depending on
their local circumstances. Considering any possibly-extrinsic
heterogeneity of agents (e.g., different sensing frequency due
to local communication delays), an asynchronous algorithm is
regarded as more realistic in coordinating a robotic swarm,
so increasing its system efficiency [21]–[23]. Finally, LICA
makes the proposed approach additionally robust against dynamical changes in bins and those in agents. Given that
inclusions or exclusions of bins are perceived by neighbouring agents, the proposed approach works well even without
requiring other far-away agents to know the changes.
The LICA-based framework developed in this paper utilises
local information as its feedback gains, which is motivated
from the recent GICA-based work in [13]. This framework
is inspired by the mechanism of decision-making in a fish
swarm, in which each of them adjusts its individual behaviour
based on those of neighbours [24]–[27]. Similarly, each agent
in the framework developed uses its local status, i.e. the current
density of its associated bin relative to those of its neighbour
bins, to generate its time-varying stochastic decision policies.
The agent is not required to know any global information,
and hence the aforementioned advantages of LICA can be
exploited.
We prove that, even using local information, the proposed
framework asymptotically converges to a desired swarm distribution and it retains the advantages of existing closedloop-type approaches. This paper explicitly presents the design requirements for a time-inhomogeneous Markov chain
to achieve these desired features. It is thus expected that
the user can utilise the requirements in designing their own
algorithm. In addition, three specific examples are provided
to demonstrate how to implement the proposed framework:
1) minimising travelling cost; 2) maximising convergence rate
under upper flux bounds; and 3) generation of quorum-based
policies (similar to [14], [15]).
The rest of this paper are organised as follows. Section II introduces the desired features of a swarm distribution guidance
framework along with relevant definitions and notations. Section III proposes our framework with its design requirements,
the biological inspiration, and an analysis regarding whether
the desired features are satisfied. We provide examples of how
to implement the framework for specific problems in Section
IV, and an asynchronous implementation in Section V. The
results of numerical experiments are shown in Section VI,
followed by concluding remarks in Section VII.
Notations
∅, 0, I and 1 denote the empty set, the zero matrix of
appropriate sizes, the identity matrix of appropriate sizes, and
a row vector with all elements are equal to one, respectively.
v ∈ Pn is a stochastic (row) vector such that v ≥ 0 and
v ·1> = 1. v[i] indicates the i-th element of vector v. Prob(E)
denotes the probability that event E will happen.
TABLE I
N OMENCLATURE
Symbol
Description
Bi
A
nA
k
nk [i]
ajk
xjk
Mkj
Θ
µ?k
Ak
Ck
Nk (i)
The i-th bin amongst a set of nbin bins (Definition 1);
A set of agents (Definition 1);
The number of total agents at time instant k (Definition 1);
The number of agents at the i-th bin (Eqn. (4));
The j-th agent’s state indicator vector (Definition 2);
Stochastic state vector of the j-th agent (Definition 4);
Stochastic decision policy of the j-th agent (Definition 4);
Desired swarm distribution (Definition 5);
Current (global) swarm distribution (Definition 3);
Physical motion constraint matrix (Definition 6);
Communicational connectivity matrix (Definition 7);
A set of (communicationally-connected) neighbour bins of
the i-th bin (Definition 7);
A set of agents in Nk (i);
Current local swarm density at the i-th bin (Eqn. (4));
Estimate of µ̄?k [i] by the j-th agent;
Locally-desired swarm density at the i-th bin (Eqn. (5));
Primary guidance matrix (Eqn. (10));
Secondary guidance matrix (Eqn. (10));
Primary local-feedback gain (e.g., Eqn. (6));
Secondary local-feedback gain (Eqn. (9));
ANk (i)
µ̄?k [i]
µ̄jk [i]
Θ̄[i]
Pkj
Skj
ξ̄kj [i]
Gjk [i]
II. P RELIMINARIES
A. Definitions
This section presents necessary definitions and assumptions
for our proposed framework, which will be shown in Section
III. Since most of them are embraced from the recent existing
literature [10], [13], we here briefly provide their essential
meanings.
Definition 1 (Agents and Bins). A set of agents A are
supposed to be distributed over a prescribed region in a state
space B. The entire space is partitioned into nbin disjoint bins
bin
(subspaces) such that B = ∪ni=1
Bi and Bi ∩ Bj = ∅, ∀i 6= j.
We also regard B = {B1 , ..., Bnbin } as the set of all the bins.
Each bin Bi represents a predefined range of an agent’s state,
e.g., position. The number of the entire agents is time-varying,
and its value at time instant k is denoted by nA
k = |A|. Note
that we do not assume that the agents keep track of nA
k.
Definition 2 (Agent’s state). Let ajk ∈ {0, 1}nbin be the state
indicator vector of agent j ∈ A at time instant k. If the agent’s
state belongs to bin Bi , then ajk [i] = 1, otherwise 0.
Definition 3 (Current (global) swarm distribution). The current (global) swarm distribution µ?k ∈ Pnbin is a row-stochastic
vector such that each element µ?k [i] is the population fraction
(swarm density) of A in bin Bi at time instant k:
µ?k :=
1 X j
ak .
|A|
(1)
∀j∈A
Definition 4 (Agent’s stochastic state and decision policy).
Agent j’s stochastic state is a row-stochastic vector xjk ∈
3
Pnbin in which each element xjk [i] gives the probability that
the agent’s state belongs to bin Bi at time instant k:
xjk [i] := Prob(ajk [i] = 1).
(2)
The probability that agent j in bin Bi at time instant k will
transition to bin Bl before the next time instant is called its
stochastic decision policy, denoted as:
Mkj [i, l] := Prob(ajk+1 [l] = 1|ajk [i] = 1).
(3)
Note that Mkj ∈ Pnbin ×nbin is a row-stochastic matrix such
that Mkj ≥ 0 and Mkj · 1> = 1> , and will be referred as
Markov matrix.
Definition 5 (Desired swarm distribution). The desired swarm
distribution Θ ∈ Pnbin is a row-stochastic vector such that
each element Θ[i] indicates the desired swarm density for bin
Bi .
Assumption 1. For ease of description for this paper, we
assume that Θ[i] > 0, ∀i ∈ {1, ..., nbin }. Obviously, in practice, there may exist some bins whose desired swarm densities
are zero. These bins can be accommodated by adopting any
subroutines ensuring that all agents eventually move to and
remain in any of the positive-desired-density bins, for example,
an escaping algorithm in [13, Section III.C].
Assumption 2 (The number of agents [6], [10], [13], [15],
[17]). It is assumed that nA
k nbin so that the time evolution of the swarm distribution is governed by the stochastic
decision policy in Equation (3). Although the finite cardinality
of the agents normally cause a residual convergence error, a
lower bound on nA
k that probabilistically guarantees a desired
convergence error is analysed in [13, Theorem 6] by exploiting
Chebyshev’s equality. Note that this theorem is generally
appliable and thus is also valid for our work.
Definition 6 (Physical motion constraint [6], [10], [13]).
Motion constraints of agents are denoted by the matrix Ak ∈
{0, 1}nbin ×nbin , where Ak [i, l] = 1 if agents in bin Bi at
time instant k are allowed to transition to bin Bl by the next
time instant; Ak [i, l] = 0, otherwise. It is assumed that Ak
is symmetric and irreducible (i.e., strongly-connected); and
Ak [i, i] = 1 for all agents, bins, and time instants.
Definition 7 (Communicationally-connected). Bins Bi and Bl
are said to be communicationally-connected, if there exists at
least one agent in bin Bi who can directly communicate with
some agents in bin Bl , and vice versa. This communicational
connectivity over all the bins at time instant k is defined by the
matrix Ck ∈ {0, 1}nbin ×nbin , where Ck [i, l] = 1 indicates that
bins Bi and Bl are communicationally-connected. Note that
Ck is symmetric and all its diagonal entries are set to be one.
For each bin Bi , we define the set of its (communicationallyconnected) neighbour bins as Nk (i) = {∀Bl ∈ B | Ck [i, l] =
1}. The set of agents in any of bins in Nk (i) is denoted by
ANk (i) = {∀j ∈ A | ajk [l] = 1, ∀l : Bl ∈ Nk (i)}.
Assumption 3 (Communicational connectivity over bins). The
physical motion constraint of a robot is, in general, more
stringent than its communicational constraint. From this, it can
be assumed that if the transition of agents between bin Bi and
Bl is allowed within a unit time instant, then the both bins
are communicationally-connected, i.e., if Ak [i, l] = 1 then
Ck [i, l] = 1. Note that we set Ck [i, l] = 0 if Ak [i, l] = 0.
This implies that the matrix Ck is irreducible, as is Ak .
The communication network over the agents is assumed to
be strongly-connected [10], [13]. Using distributed consensus
algorithms [10], [11], [28], each agent can access necessary
local information in its neighbour bins.
Assumption 4 (Pre-known Information [13]). The desired
swarm distribution Θ, the motion constraint matrix Ak (also
Ck ), and other pre-determined values such as variables regarding objective functions and user-design parameters (which will
be introduced later) are known by all the agents before they
begin a mission.
Assumption 5 (Agent’s capability [10], [13]). Each agent can
determine the bin to which it belongs, and know the locations
of neighbour bins so that it can navigate toward any of these
bins. The agent is capable of collision avoidance behaviours
against other agents or obstacles.
B. Problem Statement
The objective of the swarm distribution guidance problem
considered in this paper is to distribute a set of agents A over
a set of bins B by the Markov matrix Mkj in a manner that
holds the following desired features:
Desired Feature 1. The swarm distribution µ?k asymptotically
converges to the desired swarm distribution Θ as time instant
k goes to infinity.
Desired Feature 2. Transitions of the agents between the bins
are controlled in a way that Mkj becomes close to I as µ?k
converges to Θ. This implies that the agents are settled down
after Θ is achieved, and thus unnecessary transitions can be
reduced. Moreover, the agents identify and compensate any
partial loss or failure of the swarm distribution.
Desired Feature 3. For each agent in bin Bi , the information required for generating time-varying stochastic decision policies is not global information (e.g., µ?k ) but locally
available information within ANk (i) . Thereby, the resultant
time-inhomogeneous Markov process is based on LICA, and
has benefits such as a shorter timescale for obtaining new
information (than GICA), the potential for an asynchronous
process, etc.
Remark 1. One of our main contributions is to provide
Desired Feature 3 as well as to retain Desired Features 1 and 2
by additionally adopting Assumption 3, which can be elicited
from other assumptions in the existing literature.
III. A C LOSED - LOOP - TYPE F RAMEWORK USING L OCAL
I NFORMATION
This section proposes a LICA-based framework for the
swarm distribution guidance problem. The framework is different from the recent closed-loop-type algorithms in [10], [13]
in the sense that they utilise the global information (e.g., the
4
current swarm distribution in Equation (1)) for constructing a
time-inhomogeneous Markov matrix, whereas ours uses the
local information in Equation (7). We present, in spite of
using such relatively insufficient information, how the desired
features described in the previous section can be achieved
in the proposed framework. Before that, we introduce the
biological idea, which is about decision-making mechanisms
of a fish swarm, that inspires this framework to particularly
attain Desired Feature 3. In addition, we explicitly provide
the design requirements for a Markov matrix in order for
prospective users to easily incorporate their own specific
objectives into this framework.
1
2
3
4
…
10
11
12
13
14
…
20
21
22
23
24
…
30
31
32
33
34
…
40
...
…
…
…
…
…
91
92
93
94
… 100
Fig. 1. Examples of how to calculate µ̄?k [i]: for bin B23 , µ̄?k [23] =
nk [23]/(nk [13] + nk [22] + nk [23] + nk [24] + nk [33]). In the proposed
framework, agents in the bin only need to obtain the local information
from other agents in its neighbour bins (shaded). Note that each square
indicates each bin, and the red arrow between two bins Bi and Bl means
that Ak [i, l] = 1.
A. The Biological Inspiration
For a swarm of fishes, it has commonly been assumed
that their crowdedness limits their perception ranges over
other members, and their cardinality restricts the capacity for
individual recognition [25]. How fishes end up with collective
behaviours is different from the ways of other social species
such as bees and ants, which are known to use recruitment
signals for the guidance of the entire swarm [29], [30]. Thus,
in biology domain, a question naturally has arisen about
the mechanism of fishes’ decision-making in an environment
where local information is only available and information
transfer between members does not explicitly happen [24]–
[27], [31], [32].
It has been experimentally shown that fishes’ swimming
activities vary depending on their perceivable neighbours.
According to [31], fishes have the tendency to maintain their
statuses (e.g., position, speed, and heading angle) relative to
those of other nearby fishes, which results in their organised
formation structures. In addition, it is presented in [32] that
spatial density of fishes has influences on both the minimum
distances between them and the primary orientation of the fish
school.
Based on this knowledge, the works in [24]–[27] suggest
individual-based models to further understand the collective
behavioural mechanisms of fishes: for example, their repelling,
attracting, and orientating behaviours [24], [26]; how the density of informed fishes affects the elongation of the formation
structure [25]; and group-size choices [27]. The common and
fundamental characteristic of these models is that every agent
maintains or adjusts its personal status with consideration of
those of other individuals within its limited perception range.
As inspired by the understanding of fishes, we believe
that there must be an enhanced swarm distribution guidance
approach in which each agent only needs to keep its relative
status by using local information available from its nearby
neighbours. In this approach, a global information is not
necessary to be known by agents, and thereby the corresponding requirement of extensive information sharing over all the
agents can be alleviated.
B. Fundamental Idea of the Proposed Approach
Suppose that each agent in bin Bi is required to keep its
local status µ̄?k [i], which we referred to as the current local
swarm density at bin Bi , at the value of the corresponding
locally-desired swarm density Θ̄[i]. They are respectively
defined as follows:
nk [i]
µ̄?k [i] := P
,
(4)
∀l:Bl ∈Nk (i) nk [l]
where nk [i] is the number of agents such that ajk [i] = 1; and
Θ̄[i] := P
Θ[i]
∀l:Bl ∈Nk (i)
Θ[l]
.
(5)
We use the term µ̄jk [i] as an estimate of µ̄?k [i] by agent j, which
can be obtained through a distributed information consensus
algorithm [10], [11], [28].
The fundamental idea of the proposed approach is to make
each agent j in bin Bi :
(i) only need to estimate the difference of µ̄?k [i] and
Θ̄[i], which are both locally-available information within
Nk (i); and
(ii) more reluctant to deviate from the current bin as the
difference becomes smaller (i.e., Mkj [i, i] → 1 and
Mkj [i, l] → 0 ∀l as µ̄jk [i] → Θ̄[i]).
Our proposed framework utilises the difference between
µ̄jk [i] and Θ̄[i] as a local-information-based feedback gain,
denoted by ξ¯kj [i], which is a scalar in (0, 1] that monotonically
decreases as µ̄jk [i] converges to Θ̄[i]. For instance, this paper
uses
|Θ̄[i]−µ̄jk [i]| α
if |Θ̄[i] − µ̄jk [i]| ≤ Θ̄[i]
( Θ̄[i] )
|Θ̄[i]−µ̄jk [i]| α
(6)
ξ¯kj [i] := ξ
if
(
) < ξ
Θ̄[i]
1
otherwise
where α > 0 and ξ > 0 are design parameters. We call
this gain primary local-feedback gain because it is utilised to
control the primary guidance matrix Pkj (shown in the next
subsection).
Remark 2. Equation (4) is equivalent to the i-th element of
the following vector:
X
1
µ̄k (i) =
ajk .
(7)
|ANk (i) |
∀j∈ANk (i)
µ̄?k [l]
Namely,
= µ̄k (i)[l] if l = i. Here, we intentionally
introduce Equation (7) for ease of comparison with the information required for feedback gains in the existing literature
5
(e.g., Equation (1)). From this, it is implied that, in order for
each agent in bin Bi to estimate µ̄k (i)[i] (i.e., the current
local swarm density µ̄?k [i]), the set of other agents whose
information is necessary is restricted within ANk (i) . That is,
each agent needs to have neither a large perception radius
nor an extensive information consensus process over the entire
agents.
This subsection presents our closed-loop-type framework
based on locally-available information feedbacks. The basic
form of the stochastic decision policy for agent j in bin Bi is
such that
(
(1 − ωkj [i])Pkj [i, l] + ωkj [i]Skj [i, l] if l = i
j
Mk [i, l] :=
(1 − ωkj [i])Pkj [i, l]
∀l 6= i.
(8)
Here, ωkj [i] ∈ [0, 1) is the weighting factor to have different
weights on the agent’s primary decision policy Pkj [i, l] ∈ P
and secondary decision policy Skj [i, l] ∈ P. It is defined as
(9)
where τ j is a design parameter; and Gjk [i] ∈ [0, 1] is secondary
local-feedback gain, which is based on the difference between
µ̄jk [i] and Θ̄[i]. Note that ωkj [i] is mainly affected by Gjk [i],
while diminishing as time instant k goes to infinity.
Equation (8) can be represented in matrix form as
Mkj = (I − Wkj )Pkj + Wkj Skj ,
Pkj
Requirement 4. Pkj is irreducible such that
Pkj [i, l] > 0
if Ck [i, l] = 1.
Pkj [i, l]
otherwise.
=0
(R4)
Note that Ck is already assumed to be irreducible in Assumption 3.
C. A LICA-based Closed-loop-type Framework
ωkj [i] := exp(−τ j k) · Gjk [i]
With consideration of (R1), this requirement can be fulfilled
by Θ[i]Pkj [i, l] = Θ[l]Pkj [l, i], ∀i. A Markov process satisfying
this property is said to be reversible.
(10)
Skj
where
∈ Pnbin ×nbin and
∈ Pnbin ×nbin are rowstochastic matrices, called primary guidance matrix and secondary guidance matrix, respectively. Wkj ∈ Rnbin ×nbin is a
diagonal matrix such that diag(Wkj ) = (ωkj [1], ..., ωkj [nbin ]).
The stochastic state vector of agent j is governed by the
Markov process:
xjk+1 = xjk Mkj .
(11)
For now, we claim that, in order for this Markov system to
achieve Desired Features 1-3, Pkj must satisfy the following
requirements.
Requirement 1. Pkj is a matrix with row sums equal to one,
i.e.,
n
bin
X
Pkj [i, l] = 1, ∀i.
(R1)
Requirement 5. Pkj becomes close to I as µ̄jk converges to
Θ̄, i.e.,
Pkj [i, i] → 1 as µ̄jk [i] → Θ̄[i] (or ξ¯kj [i] → 0), ∀i.
Pkj ,
Skj ,
(R5)
ξ¯kj [i]
Depending on the objectives of a user,
and
Gjk [i] can be designed differently under given specific constraints. As long as Pkj holds (R1)-(R5) for all time instant
k and all agent j ∈ A, the aforementioned desired features
are achieved. Note that (R1)-(R4) are associated with Desired
Feature 1, whereas (R5) is with Desired Feature 2. The detailed
analysis will be described in the next subsection.
Every agent executes the following algorithm at every time
instant. The detail regarding Line 6-8 will be presented in
Section IV, which shows examples of how to implement this
framework.
Algorithm 1 Decision making of agent j at time instant k
// Obtain the local information
1: Identify the current bin Bi ;
2: Identify neighbour bins Nk (i) (and Ck [i, l] ∀l);
3: Compute Θ̄[i] using (5);
j
4: Obtain µ̄k [i];
// Generate the stochastic decision policy
j
5: Compute ξ¯k [i] (using (6));
j
6: Compute Pk [i, l] ∀l;
j
7: Compute Sk [i, l] ∀l;
j
8: Compute Gk [i];
j
9: Compute ωk [i] using (9);
j
10: Compute Mk [i, l] ∀l using (8);
// Individually behave based on the policy
11: Generate a random number z ∈ unif[0, 1];
12: Select bin Bq such that
Pq−1 j
Pq
j
l=1 Mk [i, l];
l=1 Mk [i, l] ≤ z <
13: Move to the selected bin;
l=1
Pkj
In fact,
needs to be row-stochastic, for which it should
further hold that Pkj [i, l] ≥ 0, ∀i, l. Note that this constraint is
implied by (R4), which will be introduced later.
Requirement 2. All diagonal elements are positive, i.e.,
Pkj [i, i] > 0, ∀i.
(R2)
Requirement 3. The stationary distribution of Pkj is the
desired swarm distribution Θ, i.e.,
n
bin
X
i=1
Θ[i]Pkj [i, l] = Θ[l], ∀l.
(R3)
D. Analysis
We first show that the Markov process in Equation (11)
holds Desired Feature 1 under the assumption that Pkj satisfies the requirements (R1)-(R4) for each time instant. The
stochastic state of agent j at time instant k ≥ k0 , governed by
the Markov process from an arbitrary initial state xjk0 , can be
written as:
j
xjk = xjk0 Ukj0 ,k := xjk0 Mkj0 Mkj0 +1 · · · Mk−1
.
(12)
For ease of analysis, we assume that every agent j knows any
necessary information correctly, i.e., µ̄jk [i] = µ̄?k [i].
6
Theorem 1. Provided that the requirements (R1)-(R4) are satisfied for all time instants k ≥ k0 , it holds that limk→∞ xjk =
Θ pointwise for all agents, irrespective of the initial condition.
Proof. This claim can be proved by following similar steps in
proving [13, Theorem 4]. The claim is true if limk→∞ xjk =
xjk0 · limk→∞ Ukj0 ,k = xjk0 · 1> Θ = Θ. In order for that,
the matrix product Ukj0 ,k should (i) be strongly ergodic and
(ii) have Θ as its unique limit vector, i.e., limk→∞ Ukj0 ,k =
1> Θ. We will show that the two conditions are valid under
the assumption that (R1)-(R4) are satisfied.
Lemma 5 in Appendix describes the characteristics of Mkj
and Ukj0 ,k , which will be used for the rest of this proof. From
this lemma, (a) Ukj0 ,k is primitive (thus, regular); (b) there
exists a positive lower bound γ for Mkj , ∀k; and (c) Mkj is
asymptotically homogeneous. Then, from [33, Theorem 4.15,
p.150] it follows that Ukj0 ,k is strongly ergodic, which fulfils
the condition (i).
Let ek ∈ Pnbin be the unique stationary distribution vector
corresponding to Mkj (i.e., ek Mkj = ek ). Due to the prior
condition (b) and the fact that (d) Mkj is irreducible for
∀k ≥ k0 , it follows from [33, Theorem 4.12, p.149] that
the asymptotical homogeneity of Mkj with respect to Θ (i.e.,
limk→∞ ΘMkj = Θ) is equivalent to limk→∞ ek = e and
Θ = e, where e is a limit vector. According to [33, Corollary,
p.150], under the prior conditions (b) and (d), if Ukj0 ,k is
strongly ergodic with its unique limit vector v, then v = e.
Hence, it turns out that the unique limit vector of Ukj0 ,k is Θ
(i.e, limk→∞ Ukj0 ,k = 1> Θ). Thereby, the condition (ii) is also
fulfilled.
1
2
1
2
4
3
4
3
(a) Tree-type
(b) Strongly-connected
Fig. 2. Examples of simple bin topologies to help Lemma 1 & 2: (a) treetype; (b) strongly-connected. The red line in (b) indicates a newly-added route
between bin B1 and B4 based on the topology in (a).
Proof. The matrix B ∈ Rnbin ×nbin can be linearly decomposed into ne of the same-sized matrices B(i,j) , where ne is
the number of edges in the underlying graph of Ck . Here,
B(i,j) ∈ Rnbin ×nbin is a matrix such that B(i,j) [i, i] =
−Θ[j]/Θ[i] and B(i,j) [j, j] = −Θ[i]/Θ[j]; B(i,j) [i, j] =
B(i,j) [j, i] = 1; and all the other entries are zero. For example,
consider that four bins are given and connected as shown in
Figure 2(a). Clearly, B = B(1,2) + B(2,3) + B(2,4) , where
Θ[2]
1
0
0
− Θ[1]
1
− Θ[1]+Θ[3]+Θ[4]
1
1
Θ[2]
,
B=
1
− Θ[2]
0
0
Θ[3]
Θ[2]
0
1
0
− Θ[4]
− Θ[2]
Θ[1]
B(2,3)
0
0
=
0
0
B(2,4)
0
0
=
0
0
0
Θ[1]
− Θ[2]
0
0
B(1,2) = 1
0
0
Theorem 1 implies that the stochastic state of any agent
eventually converges to the desired swarm distribution, regardless of Skj , Gjk [i] and (R5). In other words, even if (R5) is not
satisfied, the Markov system can converge to Θ. However, the
system induces unnecessary transitions of agents even after
being close enough to the desired swarm distribution, which
means that Desired Feature 2 does not hold.
For now, we present that Desired Feature 2 can be obtained
by (R5) and Theorem 2, which will be described later. Suppose
that, for every bin Bi , µ̄?k [i] converges to and eventually
reaches Θ̄[i] at some time instant k. The following shows
that at this moment it also holds that µ?k reaches Θ. From
Equations (4)-(5) and the supposition
of µ̄?k [i] = Θ̄[i] ∀i, it
P
follows that 1/Θ̄[i] · nk [i] = ∀j:Bj ∈Nk (i) nk [j] ∀i. This can
be rearranged as:
0
0 0
,
0 0
0 0
1
0
− Θ[3]
Θ[2]
0
1
1
0
− Θ[2]
Θ[3]
0
0
0
0
0
0
Θ[4]
− Θ[2]
0
1
0
0
,
0
0
0
1
0
− Θ[2]
Θ[4]
.
where X ∈ Rnbin ×nbin is a diagonal matrix such that
diag(X) = (1/Θ̄[1], 1/Θ̄[2], ..., 1/Θ̄[nbin ]); Ck is the communicational connectivity matrix (in Definition 7); and nk ∈
Rnbin is a row vector such that the i-th element indicates nk [i],
i.e., the number of agents in bin Bi at time instant k.
It is trivial that the rank of every B(i,j) is one, and the matrix
has only one linearly independent column vector, denoted by
v(i,j) . Without loss of generality, we consider v(i,j) ∈ Rnbin
1
as a column vector such that the i-th entry is − Θ[i]
, the j-th
1
entry is Θ[j] , and the others are zero: for an instance, v(1,2) =
1
1
[− Θ[1]
, Θ[2]
, 0, 0]> .
It is obvious that v(i,j) and v(k,l) are linearly independent
when the bin pairs {i, j} and {k, l} are different. This implies
that the number of linearly independent column vectors of B
is the same as that of edges in the topology. Hence, for a treetype topology of nbin bins, since there exist nbin − 1 edges,
the rank of the corresponding matrix B is nbin − 1.
Lemma 1. Given nbin bins communicationally-connected as
a tree-type topology, the rank of its corresponding matrix B
in Equation (13) is nbin − 1.
Lemma 2. Given a strongly-connected topology of bins, the
rank of its corresponding matrix B is not affected by adding
a new edge that directly connects any two existing bins.
nk · B := nk · (Ck − X) = 0
(13)
7
Proof. We will show that this claim is valid even when a
tree-type topology is given, as it is a sufficient condition for
strong-connectivity. Given the tree-type topology in Figure
2(a), suppose that bin B1 and B4 are newly connected. Then,
the new topology becomes as shown in Figure 2(b), and it has
new corresponding matrix Bnew , where Bnew = B + B(1,4) .
As explained in the proof of Lemma 1, the rank of B(1,4) is one
and it has only linearly independent vector v(1,4) . However,
this vector can be produced as a linear combination of the
existing v vectors of B (i.e., v(1,4) = v(1,2) + v(2,4) ). Thus,
the rank of Bnew retains that of B. Without loss of generality,
this implies that the rank of B of a given strongly-connected
topology is not affected by adding a new edge that directly
connects any two existing bins.
Thanks to Lemma 1 and 2, we end up with the following
corollary and theorem:
Corollary 1. Given nbin bins that are communicationally
strongly-connected, the rank of its corresponding B is nbin −1.
Theorem 2. Given nbin bins that are communicationally
strongly-connected, convergence of µ̄?k to Θ̄ is equivalent to
convergence of µ?k to Θ.
Proof. From Equation (5), it can be said that Θ·B = 0. When
µ̄?k [i] is assumed to converge to Θ̄[i] at some time instant k for
every bin Bi , Equation (13) is valid (i.e., nk · B = 0). Since
the nullity of B is one, due to Corollary 1, there is only one
linearly-independent row-vector a ∈ Rnbin such that a·B = 0.
Hence, it is obvious that nk = · Θ, where is an arbitrary
scalar value. This implies that µ?k [i] = nk [i]/nA
k = Θ[i], ∀i :
Bi ∈ B. Therefore, convergence of µ?k to Θ is equivalent to
convergence of µ̄?k to Θ̄.
From this theorem and (R5), Desired Feature 2 finally holds.
Corollary 2. If Pkj satisfies (R5), it can be said from Theorem
2 that Pkj becomes I as µ?k converges to Θ. And this is also
the case for the Markov process Mkj , which satisfies Desired
Feature 2.
In order for each agent j in bin Bi to generate the timevarying stochastic decision policy Mkj [i, l] in Equation (8), the
agent only needs to obtain its local information within ANk (i) .
Therefore, Desired Feature 3 is also achieved.
Remark 3 (Robustness against dynamic changes of agents
and those of bins). The proposed framework is robust with
against dynamic changes in the number of agents and bins.
Similarly to what is claimed in [13, Remark 8], as each agent
behaves based on its current bin location and local information
in a memoryless manner, Desired Features 1-3 in the proposed
framework won’t be affected by inclusion or exclusion of
agents in a swarm. Furthermore, provided changes on bins
are perceived by at least nearby agents in the corresponding
neighbour bins, robustness against those changes can be hold
in the proposed framework. This is because agents in bin
Bi utilise only local information such as Θ̄[i] and µ̄jk [i],
and are not required to know any information from other
far-away bins. Moreover, the proposed framework does not
need to recalculate Θ, reflecting such changes on bins, so
P
that
∀i Θ[i] = 1 because computing Θ̄[i] in (5) includes
normalisation of Θ.
IV. I MPLEMENTATION E XAMPLES
A. Example I: Minimising Travelling Expenses
This section provides examples on implementations of the
framework proposed. In particular, this subsection addresses
a problem of minimising travelling expenses of agents during
convergence to a desired swarm distribution.
This problem can be defined as: given a cost matrix
Ek ∈ Rnbin ×nbin in which each element Ek [i, l] represents
the travelling expense of an agent from bin Bi to Bl , find Pkj
such that
n
bin n
bin
X
X
min
Ek [i, l]Pkj [i, l]
(P1)
i=1 l=1
subject to (R1)-(R5) and
M Θ[l]f (ξ¯kj [i], ξ¯kj [l])f (Ek [i, l]) ≤ Pkj [i, l]
if Ck [i, l] = 1, ∀i 6= l
(14)
where M ∈ (0, 1] is a design parameter. f (ξ¯kj [i], ξ¯kj [l]) ∈ (0, 1]
is set by
f (ξ¯kj [i], ξ¯kj [l]) = max(ξ¯kj [i], ξ¯kj [l])
(15)
so that the value monotonically increases with regard to
increase of either ξ¯kj [i] or ξ¯kj [l] and diminishes as ξ¯kj [i] and
ξ¯kj [l] simultaneously reduces. This value controls the lower
bound of Pkj [i, l] in Equation (14). Θ[l] enables agents in bin
Bi to be distributed over its neighbour bins in proportion to the
desired swarm distribution. f (Ek [i, l]) ∈ (0, 1] is a scalar that
monotonically decreases as Ek [i, l] increases (see Equation
(29) for instances), encouraging agents in bin Bi to avoid
spending higher transition expenses. Note that we assume that
Ek is symmetric; Ek [i, l] > 0 if Ak [i, l] = 1; and its diagonal
entries are zero.
Corollary 3. The optimal matrix Pkj of the problem (P1) is
given by: ∀i, l ∈ {1, ..., nbin } and i 6= l,
(
M Θ[l]f (ξ¯kj [i], ξ¯kj [l])f (Ek [i, l]) if Ck [i, l] = 1
j
Pk [i, l] =
0
otherwise
(16)
and ∀i = l,
X j
Pkj [i, i] = 1 −
Pk [i, l].
(17)
∀l6=i
Proof. We can prove this by following the proof of [13,
Corollary 1]. Suppose that the problem is only subject to (R4)
and (14), without (R1)-(R3) and (R5). Then, the off-diagonal
elements of an optimal matrix should be their corresponding
lower bounds in (14) if Ck [i, l] = 1. The diagonal elements of
the matrix do not affect the objective function due to the fact
that Ek [i, i] = 0, ∀i. Accordingly, the matrix Pkj that holds
(16) and (17) is also an optimal matrix for the simplified
problem.
Let us now consider (R1)-(R3) and (R5). Since M ,
fP(ξ¯kj [i], ξ¯kj [l]) and f (Ek [i, l]) are upper-bounded by 1 and
j
∀l6=i Θ[l] < 1, Pk [i, i] in (17) is always positive for all i,
which fulfils (R2). It is also obvious that (R1) is satisfied by
8
Equation (17). From Equation (16), it holds that Θ[i]Pkj [i, l] =
Θ[l]Pkj [l, i], complying with (R3). Since (R1)-(R4) are satisfied, the Markov process is converging to a desired distribution
due to Theorem 1. Noting that f (ξ¯kj [i], ξ¯kj [l]) diminishes as
ξ¯kj gets close to zero (i.e., µ̄jk → Θ̄), (R5) is also fulfilled
by Equation (17). Hence, Pkj is the optimal solution for the
problem (P1).
For reducing unnecessary transitions of agents during this
process, it is favourable that agents in bin Bi such that µ̄jk [i] ≤
Θ̄[i] (i.e., underpopulated) do not deviate. To this end, we set
Skj = I and Gjk [i] as follows [13]:
Gjk [i] :=
exp(β(Θ̄[i] − µ̄jk [i]))
exp(β|Θ̄[i] − µ̄jk [i]|)
.
(18)
The gain value is depicted in Figure 3(a) with regard to β.
Remark
of P
Convergence Rate). Due to the fact
P 4 (Increase
j
that
P
[i,
l]
≤
∀l6=i k
∀l:Bl ∈Nk (i)\Bi Θ[l] from Equation
(16), the total outflux of agents from bin Bi becomes smaller as
the bin has fewer connections with other bins. This eventually
makes the convergence rate of the Markov process slower.
Adding an additional variable into Pkj [i, l] in (16) does not
affect the obtainment of Desired Features 1-3 as long as Pkj
satisfies (R1)-(R5). Thus, in order to enhance the convergence
rate under the requirements, one can add
Θ := min{ P
1
∀s:Bs ∈Nk (i)\Bi
Θ[s]
,P
1
∀s:Bs ∈Nk (l)\Bl
}
Θ[s]
(19)
into Pkj [i, l], as follows:
¯j
¯j
Θ M Θ[l]f (ξk [i], ξk [l])f (Ek [i, l])
j
Pk [i, l] =
if Ck [i, l] = 1
0
otherwise,
(20)
which can be substituted for Equation (16).
Algorithm 2 Subroutine of Algorithm 1 (Line 6–8) for P1
Compute Pkj [i, l] ∀l using (16) (or (20)) and (17);
j
j
2: Set Sk [i, i] = 1 and Sk [i, l] = 0, ∀l 6= i
j
3: Compute Gk [i] using (18);
1:
for mitigating the trade-off between convergence rate and longterm system efficiency.
For the sake of imposing upper flux bounds during the entire
process, we consider the following one-way flux constraint:
nk [i]Pkj [i, l] ≤ c(i,l) ,
∀i, ∀l 6= i.
(21)
This means that the number of agents moving from bin Bi
to Bl is upper-bounded by c(i,l) . The bound value is assumed
to be very small with consideration of mission environments
such as the number of agents, the number of bins, and their
topology. Otherwise, all the agents can be distributed over
the bins very soon so that the upper flux bounds become
meaningless, and thus the corresponding problem can be
trivial.
Regarding the convergence rate of a Markov chain, there
are respective analytical methods depending on whether it
is time-homogeneous or time-inhomogeneous. For a timehomogeneous Markov chain, if the matrix is irreducible, the
second largest eigenvalue of the matrix is used as an index
indicating its asymptotic convergence rate [34, p.389]. In
contrast, for a time-inhomogeneous Markov chain, coefficients
of ergodicity can be utilised as a substitute for the second
largest eigenvalue, which is not useful for this case [35].
Particularly, this paper uses the following proper coefficient
of ergodicity, amongst others:
Definition 8. (Coefficient of Ergodicity [33, pp. 136–137]).
Given a stochastic matrix M ∈ Pn×n , a (proper) coefficient
of ergodicity 0 ≤ τ (M) ≤ 1 can be defined as:
τ (M) := max max |M[i, s] − M[l, s]|.
∀s
∀i,∀l
(22)
A coefficient of ergodicity is said to be proper if τ (M) = 0
if and only if M = 1> · v, where v ∈ Pn is a row-stochastic
vector.
The convergence rate of a time-inhomogeneous Markov
chain Mk ∈ Pn×n , ∀k > 1 can be maximised by minimising
τ (Mk ) at each time instant k,Qthanks to [33, Theorem 4.8,
r
p.137]: τ (M1 M2 · · · Mr ) ≤ k=1 τ (Mk ). Hence, the objective of the specific problem considered in this subsection
can be defined as: find Pkj such that
min τ (Pkj )
(23)
subject to (R1)-(R5) and (21).
B. Example II: Maximising Convergence Rate within Upper
Flux Bounds
This subsection presents an example in which the specific
objective is to maximise the convergence rate under upper
bounds regarding transitions of agents between bins, denoted
by upper flux bounds. The bounds can be interpreted as safety
constraints in terms of collision avoidance and congestion:
higher congestions may induce higher collisions amongst
agents, which may bring unfavourable effects on system
performance. A similar problem is addressed by an open-looptype algorithm in [17], where transitions of agents are limited
only at a desired swarm distribution. This restriction is not for
considering the aforementioned safety constraints, but rather
Remark 5 (Advantages of the coefficient of ergodicity in (22)).
Other proper coefficients in [33, p. 137] such as
X
τ1 (M) = 1 − min
min (M[i, s], M[l, s])
i,l
∀s
or
τ2 (M) = 1 −
X
∀s
min (M[i, s]) .
∀i
may have the trivial case such that τ1 (Pkj ) = 1 (or τ2 (Pkj ) = 1)
for some time instant k, when they are applied to this problem.
This is because, given a strongly-connected topology Ck , there
may exist a pair of bins Bi and Bl such that Pkj [i, s] = 0
or Pkj [l, s] = 0, ∀s. To avoid this trivial case, the work in
9
[13] instead utilises τ1 ((Pkj )dCk ) as the proper coefficient of
ergodicity, where dCk denotes the diameter of the underlying
graph of Ck . However, this implies that agents in bin Bi are
required to additionally access the information from other bins
beside Nk (i), causing additional communicational costs. The
coefficient of ergodicity in (22) does not suffer this issue. Note
that τ (M) ≤ τ1 (M) ≤ τ2 (M) [33, p. 137].
Finding the optimal solution for the problem (23) is another
challenging issue, which can be called fastest mixing Markov
chain problem. Since the purpose of this section is to show an
example of how to implement our proposed framework, we
heuristically address this problem at this moment.
Suppose that matrix Pkj satisfying (R1)-(R5) is given,
and the topology of bins is not fully-connected. Since the
matrix is non-negative and there exists at least one zerovalue entry in each column, the coefficient of ergodicity
can be said as τ (Pkj ) = max∀i,∀s (Pkj [i, s]). Assuming that
max∀l6=i Pkj [i, l] ≤ 1/|Nk (i)|, which is generally true due
to the smallness of c(i,l) , it turns out that each diagonal
element of Pkj is the largest value in each row. Thus, we can
say that τ (Pkj ) = max∀i Pkj [i, i]. TheP
objective function of
this problem can be said as maxmin∀i ∀l6=i Pkj [i, l] because
minimising the maximum diagonal element of a stochastic
matrix is equivalent to maximising the minimum row-sum of
its off-diagonal elements.
We turn now to the constraints (R1)-(R5) and (21). In order
to comply with (R3), we initially set Pkj [i, l] = Θ[l]Qjk [i, l],
where Qjk is a symmetric matrix that we will design now.
The constraint (21), (R4), and the symmetricity of Qk are
integrated into the following constraint: ∀i, ∀l 6= i,
c(l,i)
c(i,l)
,
) ≥ Qjk [i, l] > 0 if Ck [i, l] = 1
min(
nk [i]Θ[l] nk [l]Θ[i]
Qjk [i, l] = 0 otherwise.
(24)
For (R2) and (R5), we set the diagonal entries of Pkj as
Pkj [i, i] ≥ 1 − ξ¯kj [i],
∀i.
This
Pnbincan bej rewritten, with consideration of (R1) (i.e.,
l=1 Θ[l]Qk [i, l] = 1, ∀i), as
X
Θ[l]Qjk [i, l] ≤ ξ¯kj [i], ∀i.
(25)
∀l6=i
Then, the reduced problem can be defined as: find Qjk such
that
X
maxmin∀i
Θ[l]Qjk [i, l]
(P2)
∀l6=i
subject to (24) and (25).
The algorithm for this problem is shown in Algorithm 3.
If we neglect (25), an optimal solution can be obtained by
making Qjk [i, l] equal to its upper bound of (24) (Line 2).
However, this solution may not hold (25). Thus, we lower the
entries of Qjk to satisfy (25), while keeping them symmetric
and as higher as possible (Line 3–9). In details, Line 3 (or
Line 6) ensures the constraint (25) for each bin Bi in a way
that, if this is not the case, obtains the necessary lowering
factor ¯0Q [i] (or ¯Q [i]). In order to keep Qjk as higher as
possible, we temporarily take 0Q [i, l] as the maximum value
of {¯
0Q [i], ¯0Q [l]} (Line 4). After curtailing Qjk [i, l] by applying
0
Q [i, l], we obtain the corresponding lowering factor again
(Line 5–6). The minimum value is taken for both maintaining
Qjk symmetric and satisfying (25) (Line 7). Then, the corresponding stochastic decision policy is generated based on the
resultant Qjk (Line 9–10). Note that we set Gjk [i] = 0 for all
time instants, all bins, and all agents, so Mkj = Pkj .
Algorithm 3 Subroutine of Algorithm 1 (Line 6) for P1
// Initialise Pkj
j
1: Pk [i, l] = 0, ∀l ∈ {1, 2, ..., nbin };
// Compute Qjk satisfying (24)
c(l,i)
c(i,l)
j
2: Qk [i, l] = min( n [i]Θ[l] , n [l]Θ[i] ), ∀Bl ∈ Nk (i) \ {Bi };
k
k
j
// Lower Qk to satisfy (25)
j
ξ̄k
[i]
, 1);
j
∀l6=i Θ[l]Qk [i,l]
0
0
0
4: Q [i, l] = max(¯
Q [i], ¯Q [l]), ∀Bl ∈ Nk (i) \ {Bi };
j
j
0
5: Qk [i, l] := Q [i, l]Qk [i, l], ∀Bl ∈ Nk (i) \ {Bi };
j
ξ̄k
[i]
, 1);
6: ¯Q [i] = min( P
j
∀l6=i Θ[l]Qk [i,l]
3:
¯0Q [i] = min( P
Q [i, l] = min(¯
Q [i], ¯Q [l]), ∀Bl ∈ Nk (i) \ {Bi };
Qjk [i, l] := Q [i, l]Qjk [i, l], ∀Bl ∈ Nk (i) \ {Bi };
// Compute Pkj
j
j
9: Pk [i, l] = Θ[l]Qk [i, l], ∀Bl ∈ Nk (i) \ {Bi };
P
j
j
10: Pk [i, i] = 1 −
∀l6=i Pk [i, l];
7:
8:
C. Example III: Local-information-based Quorum Model
This subsection shows that the proposed framework is
able to incorporate a quorum model, which is introduced in
[14], [15]. In this model, if a bin is overpopulated above
a certain level of predefined threshold called quorum, the
probabilities that agents in the bin move to neighbour bins
are temporarily increased, rather than following given Pkj .
This feature eventually brings an advantage to the convergence
performance of the swarm.
To this end, we set the secondary guidance matrix Skj as
follows: ∀i, l ∈ {1, ..., nbin } and ∀j ∈ A,
(
1/|Nk (i)|
if Ck [i, l] = 1
j
Sk [i, l] :=
(26)
0
otherwise.
This matrix makes agents in a bin equally disseminated over
its neighbour bins. In addition, the secondary feedback gain
Gjk [i] is defined as
!−1
µ̄jk [i]
j
Gk [i] := 1 + exp γ(qi −
)
,
(27)
Θ̄[i]
where γ > 0 is a design parameter, and qi > 1 is the quorum
for bin Bi . The value of the gain is shown in Figure 3(b),
varying depending on γ and qi . As µ̄jk [i]/Θ̄[i] becomes higher
than the quorum, Gjk [i] gets close to 1 (i.e., Skj [i, l] becomes
more dominant than Pkj [i, l]). The steepness of the function at
the quorum value is regulated by γ.
The existing quorum models in [14], [15] require each agent
to know µ?k [i], which implies that the total number of agents
10
1
1
=1
=5
= 10
= 100
0.9
0.8
0.9
q i = 2,
=1
q i = 2,
= 10
0.8
q i = 2,
= 100
0.7
0.7
0.6
0.6
0.5
0.5
0.4
0.4
0.3
0.3
0.2
0.2
0.1
q i = 1.5,
= 10
0.1
0
-2
0
-1.5
-1
(underpopulated)
-0.5
0
0.5
1
1.5
2
(overpopulated)
0
0.5
1
(underpopulated)
(a) P1
1.5
2
2.5
3
3.5
4
(overpopulated)
(b) the Quorum Model
Gjk [i]
Fig. 3. The secondary feedback gains
depending on the associated
design parameters: (1) for P1 (Eqn. (18)); (2) for the quorum model (Eqn.
(27))
nA
k should be tracked in real time. It could be possible that
some agents in a swarm unexpectedly become faulted by
internal or external effects during a mission, which hinders
for other alive agents from keeping track of nA
k in a timely
manner. On the contrary, this requirement is not the case for
the quorum model in this subsection, and it works by using
the local information available from ANk (i) .
Algorithm 4 Subroutine of Algorithm 1 (Line 7-8) for the
quorum-based method
Compute Skj [i, l] ∀l using (26);
j
2: Compute Gk [i] using (27);
1:
V. A SYNCHRONOUS I MPLEMENTATION
A synchronous process induces extra time delays and interagent communications to make entire agents, who may have
different timescales for obtaining new information and make
decisions, remain in sync. Such unnecessary waiting time and
communications may cause unfavourable effects on mission
performance or even may not be realisable in practice [23].
In the previous sections, it was assumed that a swarm of
agents act synchronously at every time instant. Here we show
that the proposed framework allows agents to operate in an
asynchronous manner, assuming that the union of underlying
graphs of the corresponding Markov matrices across some time
intervals is frequently and infinitely strongly-connected.
Suppose that an algorithm to compute Pkj that satisfies (R1)(R5) in a synchronous environment is given (e.g., Algorithm 2
or 3). We propose an asynchronous implementation, as shown
in Algorithm 5, which substitutes for Line 6 in Algorithm
Algorithm 5 Asynchronous Construction of Pkj [i, l] (Substitute for Line 6 of Algorithm 1)
1:
2:
3:
4:
5:
6:
7:
+
if Bi ∈ R+
k & isnonempty(Rk \ {Bi }) then
j
Compute Pk [i, l] as usual, ∀Bl ∈ R+
k \ {Bi };
−
Pkj [i, l] = 0, ∀B
∈
R
;
l
k
P
Pkj [i, i] = 1 − ∀l6=i Pkj [i, l];
else
Pkj [i, i] = 1; Pkj [i, l] = 0, ∀l 6= i;
end if
1. We refer to a set of bins where agents are ready to use
their respective local information (e.g., µ̄jk [i]) as R+
k , and a
set of the other bins as R−
k . It is assumed that each agent
j in bin Bi ∈ R+
k also knows the local information of its
neighbour bin Bl ∈ Nk (i) if Bl ∈ R+
k . As shown in Line 2, the
agent follows an existing procedure as long as all information
required to generate Pkj [i, l] is available (e.g., µ̄jk [i] and µ̄jk [l]
for Algorithm 2, and ¯0Q [i] and ¯0Q [l] for Algorithm 3). On the
contrary, if any local information of its neighbour bin Bl ∈
Nk (i) is not available, the probability to transition to the bin
is set as zero (Line 3). In the meantime, each agent for whom
necessary local information is not ready does not deviate but
remains at the bin it belonged to. Equivalently, it can be said
that Pkj [i, i] = 1 and Pkj [i, l] = 0, ∀l 6= i (Line 6).
Hereafter, for the sake of differentiation from the original
Pkj generated in a synchronous environment, let us refer to the
matrix resulted by Algorithm 5 as asynchronous primary guidance matrix, denoted by P̄kj . Accordingly, the asynchronous
Markov matrix can be defined as:
M̄kj := (I − Wkj )P̄kj + Wkj Skj .
Here, we show that this asynchronous Markov process also
converges to the desired swarm distribution.
Lemma 3. The matrix P̄k , for every time instant k, satisfies
the following properties: (1) row-stochastic; (2) all diagonal
elementsP
are positive, and all other elements are non-negative;
nbin
and (3) i=1
Θ[i]P̄kj [i, l] = Θ[l], ∀l.
Proof. The matrix P̄kj is row-stochastic because of Line 4 and
6 in Algorithm 5. Furthermore, given that Pkj satisfies (R2),
the property (2) is valid for P̄kj because P̄kj [i, i] ≥ Pkj [i, i] for
∀i.
Let us now
to the property (3). For ∀Bi ∈ R−
k , it is
Pnturn
bin
trivial that l=1 Θ[l]P̄kj [l, i] = Θ[i] because of Line 6. For
j
∀Bi ∈ R+
k , it turns out from Algorithm 5 that (i) P̄k [i, l] =
j
j
j
−
P̄k [l, i] = 0 for ∀Bl ∈ Rk ; (ii) P̄k [i, l] =PPk [i, l] for ∀Bl ∈
j
j
R+
Pkj [i, l].
k \ {Bi } and (iii) P̄k [i, i] = Pk [i, i] +
∀l:Bl ∈R−
k
We apply the findings into the following equation:
n
bin
X
Θ[l]P̄kj [l, i] =
X
Θ[l]P̄kj [l, i]
∀l:Bl ∈R−
k
l=1
+
X
Θ[l]P̄kj [l, i] + Θ[i]P̄kj [i, i].
∀l:Bl ∈R+
k \{Bi }
(28)
The first term of the right hand side becomes zero because of
j
j
(i). Due to (ii) and the fact that Θ[i]P
P k [i, l] = Θ[l]Pjk [l, i]
∀l, the second term becomes Θ[i] ∀l:Bl ∈R+ \{Bi } Pk [i, l].
P k
The last term becomes Θ[i]Pkj [i, i] + Θ[i] ∀l:Bl ∈R− Pkj [i, l]
k
because of (iii). Putting
Equation (28)
P all of them together,
is equivalent to Θ[i] ∀l:Bl ∈R+ \{Bi } Pkj [i, l] + Θ[i]Pkj [i, i] +
k
P
Pnbin j
Θ[i] ∀l:Bl ∈R− Pkj [i, l] = Θ[i] l=1
Pk [i, l] = Θ[i].
k
Lemma 4. If the union of a set of underlying graphs of
{P̄k1 , P̄k1 +1 , ..., P̄k2 −1 } is strongly-connected, then the matrix
product P̄k1 ,k2 := P̄k1 P̄k1 +1 · · · P̄k2 −1 is irreducible.
11
Proof. Thanks to Lemma 3 and 4, the matrix product P̄ki ,ki+1
for each time interval [ki , ki+1 ) satisfies (R1)-(R4). Therefore,
one can prove this theorem by similarly following the proof
of Theorem 1.
VI. N UMERICAL E XPERIMENTS
A. Effects of Primary Local-feedback Gain ξ¯kj [i]
Depending on the shape of primary feedback gain ξ¯kj [i], the
performance of the proposed framework changes, especially
with respect to convergence rate, fraction of transitioning
agents, and residual convergence error. Let us first investigate
the effect of changes in the feedback gain using Algorithm 2
with Equation (20).
We consider a scenario where a set of 2, 000 agents are
supposed to be distributed over an arena consisting of 10 × 10
bins, as depicted in Figure 1. There are vertical and horizontal
paths between adjacent bins. Note that the agents are allowed
to move at most 3 paths away within a unit time instant. All
the agents start from a bin, which reflects the fact that they
are generally deployed from a base station at the beginning
of a mission. The desired swarm distribution Θ is uniformrandomly generated at each scenario. The agents are assumed
to estimate necessary information correctly, e.g. µ̄jk [i] = µ̄?k [i].
For the rigorous validation, the performance of the proposed
algorithm will be compared with that of the GICA-based
algorithm [13]. To this end, f (Ek [i, l]) is set to be the same
as the corresponding coefficient in [13, Corollary 1]:
f (Ek [i, l]) := 1 −
Ek [i, l]
Ek,max + E
(29)
where Ek,max is the maximum element of the travelling
expense matrix Ek , and E is a user-design parameter. Ek [i, l]
is defined as a linear function based on the distance between
bin Bi and Bl :
Ek [i, l] := E1 · ∆s(i,l) + E0
25
Fraction of Transitioning Agents (%)
Primary feedback gain with varying
0.8
0.7
0.6
0.5
0.4
= 0.2
= 0.4
= 0.6
= 0.8
= 1.0
= 1.2
0.3
0.2
0.1
0
= 0.2
= 0.4
= 0.6
= 0.8
= 1.0
= 1.2
0.2
0.4
0.6
Primary feedback gain when
15
10
5
0.8
1
0
500
1000
1500
=1
2000
2500
3000
3500
4000
Time Instant
(a) Primary Local-feedback Gain
(b) Fraction
Agents
1
of
Transitioning
0.08
α
α
α
α
α
α
0.9
0.8
0.7
= 0.2
= 0.4
= 0.6
= 0.8
= 1.0
= 1.2
α
α
α
α
α
α
0.07
0.6
0.5
0.4
0.3
0.2
0.06
= 0.2
= 0.4
= 0.6
= 0.8
= 1.0
= 1.2
0.05
0.04
0.03
0.02
0.01
0.1
0
20
0
0
Convergence Error ( D H(Θ, µ *k) )
Theorem 3. Suppose that there exists an infinite sequence of non-overlapping time intervals [ki , ki+1 ), i =
0, 1, 2, ..., such that the union of underlying graphs
of {P̄ki , P̄ki +1 , ..., P̄ki+1 −1 } in each interval is stronglyconnected. Let the stochastic state of agent j at time instant k ≥ k0 , governed by the corresponding Markov process from an arbitrary state xjk0 , be xjk = xjk0 Ūkj0 ,k :=
j
xjk0 M̄kj0 M̄kj0 +1 · · · M̄k−1
. Then, it holds that limk→∞ xjk = Θ
pointwise for all agents, irrespective of the initial condition.
1
0.9
Convergence Error ( D H(Θ, µ *k) )
Proof. Since the union of a set of underlying graphs of
{P̄k1 , P̄k1 +1
P̄k2 −1 } is strongly-connected, the underlying
P,k...,
2 −1
graph of
k=k1 P̄k is also strongly-connected. Noting that
every P̄k , ∀k ∈ {k1 , k1 + 1, ..., k2 − 1} is a nonnegative
nbin × nbin matrix and its diagonal elements are positive (by
Lemma 3), it follows from [36, Lemma 2] that P̄k1 ,k2 ≥
Pk2 −1
γ k=k
P̄k , where γ > 0. This implies that the underlying
1
graph of P̄k1 ,k2 is strongly-connected, and thus the matrix
P̄k1 ,k2 is irreducible.
0
500
1000
1500
2000
2500
3000
3500
4000
0
3500
3600
3700
Time Instant
(c) Convergence Error
3800
3900
4000
Time Instant
(d) Convergence Error (zoomed-in)
Fig. 4. Performance comparison depending on the primary local-feedback
gain ξ̄kj [i] with different setting of α: (a) the value of ξ̄kj [i]; (b) the fraction
of transitioning agents; (b) the convergence performance; (d) the convergence
performance (zoomed-in for time instant between 2000 and 4000)
are assumed to follow any shortest route when they transition
between two bins. The design parameters are set as follows:
E1 = 1 and E0 = 0.5 in (30); E = 0.1 in (29); ξ = 10−9
in (6); M = 1 in (20); β = 1.8 × 105 in (18); and τ j = 10−6
in (9).
As a performance index for the closeness between the current swarm distribution µ?k and Θ, we use Hellinger Distance,
i.e.,
v
unbin
2
q
uX p
1
?
Θ[i] − µ?k [i] ,
DH (Θ, µk ) := √ t
2 i=1
Hellinger Distance is known as a “concept of measuring
similarity between two distributions” [37] and is utilised as
a feedback gain in the existing work [13].
More importantly, to examine the effects of the shape of
ξ¯kj [i], we set α in (6) as 0.2, 0.4, 0.6, 0.8, 1 and 1.2.
Figure 4 reveals that the convergence rate can be traded off
against the fraction of transitioning agents and the residual
convergence error. As ξ¯kj [i] becomes more concave, i.e. the
value of α decreases, the summation of off-diagonal entries
of Pkj becomes higher, leading to more transitioning agents,
but a faster convergence rate. At the same time, a higher value
of ξ¯kj [i] even at a low value of |Θ̄[i]−µjk [i]| gives rise to unnecessarily higher off-diagonal entries of Pkj . Hence, the swarm
tends to be prevented from converging to the desired swarm
distribution properly, resulting in higher residual convergence
error.
B. Comparison with a GICA-based Method
(30)
where ∆s(i,l) is the minimum required number of paths from
Bi to Bl ; E1 and E0 are user-design parameters. The agents
Let us now compare the LICA-based method for (P1) with
the GICA-based method in [13]. The scenario considered is the
same as the one in the previous subsection except for α = 0.6.
12
1
25
LICA
GICA
4
LICA
GICA
0.8
0.7
0.6
0.5
0.4
0.3
0.2
LICA
GICA
20
Cumulative Travel Expenses
Fraction of Transitioning Agents (%)
Convergence Error (DH ( , *k ))
0.9
10
2.5
15
10
5
2
1.5
1
0.5
0.1
0
0
0
50
100
150
200
250
300
350
400
0
0
50
100
Time Instant
150
200
250
300
350
Time Instant
(a) Convergence Error
(b) Fraction of Transitioning Agents
400
0
50
100
150
200
250
300
350
400
Time Instant
(c) Cumulative Travel Expenses
Ratio of Cumulative Travel Expenses until DH = 0.1
Ratio of Required Time Instants for Convergence (LICA/GICA)
Fig. 5. Performance comparison between the proposed method (LICA) with the existing method (GICA): (a) the error between the current swarm status and
the desired status, shown as Hellinger Distance; (b) the fraction of agents who transitions between any two bins; (c) the cumulative travel expenses of all the
agents from the beginning.
1.4
1.3
1.2
1.1
1
0.9
0.8
method and those by the compared method are 1.72 × 104 and
1.96 × 104 , respectively, and their ratio is 0.878. This is also
confirmed by the statistical result in Figure 6(b). A possible
explanation is that when some of the bins do not meet their
desired swarm densities, the entire agents in the GICA-based
method would obtain higher feedback gains, which might lead
to unnecessary transitions. On the contrary, this is not the case
in the LICA-based method since agents are only affected by
their neighbour bins.
0.92
0.9
0.88
0.86
0.84
0.82
0.8
0.78
0.76
0.74
0.7
0.3
0.28 0.26 0.24 0.22
0.2
0.18 0.16 0.14 0.12
Convergence Error (DH ( , *k ))
(a) Converged Time Instants
0.1
1
LICA/GICA
(b) Cumulative
Expenses
Travel
Fig. 6. Performance comparison (Monte-Carlo experiments) between the
proposed method (LICA) and the existing method (GICA): (a) the required
time instants to converge to DH (Θ, µ?k ) ∈ {0.30, 0.28, ..., 0.12, 0.10} (i.e.,
convergence rate); (2) the ratio of the cumulative travel expenses by LICA to
those by GICA until converging to DH (Θ, µ?k ) = 1.
Note that Θ in Remark 4 can control convergence rate, but is
not discussed in [13]. For the fair comparison, Θ is applied
to both the methods.
We conduct 100 runs of Monte Carlo experiments. Figure
5 presents the results of one representative scenario and the
statistical results of the Monte Carlo experiments are shown
in Figure 6. According to Figure 5(a), the convergence rate
of the proposed method is slower at the initial phase, but
similar to that of the the GICA-based method as reaching
DH (Θ, µ?k ) = 0.10. This is confirmed by the statistical results
in Figure 6(a), where the ratio of the required time instants
for converging to DH (Θ, µ?k ) ∈ {0.30, 0.28, ..., 0.12, 0.10} in
the LICA-based method to those of the GICA-based method
is presented. At this point, it is worth noting that these convergence rate results are presented in respect to time instants
of each Markov process. As the LICA-based framework may
have a much shorter timescale, its convergence performance in
practice could be better than that of the GICA-based method.
Figure 5(c) shows that the cumulative travel expenses are
smaller in the proposed method. The expenses by the proposed
C. Robustness in Asynchronous Environments
This subsection investigates the effects of asynchronous
environments in the proposed LICA-based method for (P1)
and compares them with those in the GICA-based method
in [13]. Hence, a realistic scenario where an asynchronous
process is required is considered: in the scenario, it is assumed
that agents in some bins cannot communicate for some reason
(such bins are called blocked) and thus other agents in normal
bins have to perform their own process without waiting them.
The proportion of blocked bins to the entire bins is set to
be different values, i.e. 0%, 10%, 20% and 30%. At each time
instant, the corresponding proportion of bins are randomly
selected as blocked bins. For the proposed framework, the
asynchronous implementation in Section V is built upon
Algorithm 2. In the GICA-based method, for the comparison
purpose only, it is assumed that agents in normal bins obtain
µjk = µ?k without interacting with agents in the blocked bins.
The rest of scenario setting are the same as those in Section
VI-B.
Figure 7 illustrates the performance of each method: convergence rate, fraction of transitioning agents, and cumulative travel expenses. As the proportion of the blocked bins
increases, the GICA-based method tends to have faster convergence speed, whereas it loses Desired Feature 2 and thus
increases cumulative travelling expenses (as shown in Figure
7(a), 7(b), and 7(c), respectively). On the contrary, the LICAbased method shows a graceful degradation in terms of Desired
Feature 2 (as shown in Figure 7(b)). A possible explanation
for these results could be that higher feedback gains due to the
communication disconnection induce faster convergence performance in each method than the normal situation. This effect
13
30
Convergence Error (DH ( , *k ))
0.8
0.7
0.6
Fraction of Transitioning Agents (%)
LICA (Blocked 0%)
LICA (Blocked 10%)
LICA (Blocked 20%)
LICA (Blocked 30%)
GICA (Blocked 0%)
GICA (Blocked 10%)
GICA (Blocked 20%)
GICA (Blocked 30%)
0.9
0.5
0.4
0.3
0.2
10
14
LICA (Blocked 0%)
LICA (Blocked 10%)
LICA (Blocked 20%)
LICA (Blocked 30%)
GICA (Blocked 0%)
GICA (Blocked 10%)
GICA (Blocked 20%)
GICA (Blocked 30%)
25
20
4
LICA (Blocked 0%)
LICA (Blocked 10%)
LICA (Blocked 20%)
LICA (Blocked 30%)
GICA (Blocked 0%)
GICA (Blocked 10%)
GICA (Blocked 20%)
GICA (Blocked 30%)
12
Cumulative Travel Expenses
1
15
10
5
10
8
6
4
2
0.1
0
0
0
50
100
150
200
250
300
350
400
450
0
0
50
100
Time Instant
150
200
250
300
350
400
450
0
50
100
Time Instant
(a) Convergence Error
150
200
250
300
350
400
450
Time Instant
(b) Fraction of Transitioning Agents
(c) Cumulative Travel Expenses
80
Swarm Distribution (D H ( , *k ))
Fractions of Transitioning Agents (%)
P2
P2 + Quorum
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
P2
P2 + Quorum
70
60
50
40
30
20
10
0
0
500
1000
1500
2000
2500
3000
3500
4000
Time Instant
(a) Convergence Error
0
500
1000
1500
2000
2500
3000
3500
Time Instant
(b) Fraction of Transitioning Agents
4000
40
400
Actual (Case 1)
Probabilistic (Case 1)
Actual (Case 2)
Probabilistic (Case 2)
35
350
30
300
25
250
20
200
15
150
10
100
5
50
0
0
500
1000
1500
2000
2500
3000
3500
0
4000
Max (# Transitions of Agents for Each Path) - Case 2
1
Max (# Transitions of Agents for Each Path) - Case 1
Fig. 7. Performance comparison in communication-disconnected situations: (a) the error between the current swarm status and the desired status, shown as
Hellinger Distance; (b) the fraction of agents who transitions between any two bins; (c) the cumulative travel expenses of all the agents from the beginning.
Time Instant
(c) Maximum Number of Transitioning Agents via
Each Path (for P2)
Fig. 8. Comparison results between the method for P2 and the quorum-based model: (a) the error between the current swarm status and the desired status,
shown as Hellinger Distance, at each time instant; (b) the fraction of agents who transition between any two bins at each time instant. (c) The maximum
number of transitioning agents via each path in the method for P2 (Case 1: |A| = 10, 000 and c(i,l) = 20, ∀i 6= l; Case 2: |A| = 100, 000 and c(i,l) = 200,
∀i 6= l)
is dominant for the GICA-based method because it affects
the entire agents, who use global information. However, in
the LICA-based framework, the communication disconnection
only locally influences so that its effectiveness is relatively
modest.
D. Demonstration of Example II and III
This subsection demonstrates the LICA-based method for
(P2) (i.e., Algorithm 3) and the quorum model (i.e., Algorithm
4). For the former, we consider a scenario where 10, 000 agents
and an arena consisting of 10 × 10 bins are given. The arena
is as depicted in Figure 1, where the agents are allowed to
move only one path away within a unit time instant. For each
one-way path, the upper flux bound per time instant is set as
20 agents (i.e., c(i,l) = 20, ∀i 6= l). All the agents start from a
bin, and the desired swarm distribution is uniform-randomly
generated.
For the latter, we build the quorum model upon the LICAbased method for (P2). This can be a good strategy for a user
who wants to achieve not only faster convergence rate but
also lower unnecessary transitions after equilibrium, which are
regulated by the upper flux bounds. Thus, in the same scenario
described above, we will demonstrate the combined algorithm
that computes Skj and Gjk by Algorithm 4 and Pkj by Algorithm
3. We set qi = 1.3 and γ = 30 for (27); α = 1 and ξ = 10−9
for (6); and τ j = 10−6 for (9).
Figure 8(a) and 8(b) presents that the both approaches make
the swarm converge to the desired swarm distribution. It is
observed that the number of transitioning agents in the method
for (P2) are restricted because of the upper flux bound during
the entire process. Meanwhile, the quorum-based method very
quickly disseminates the agents, who are initially at one bin,
over other bins, and thus the fraction of transitioning agents
is very high at the initial phase. After that, the population
fraction by the quorum-based method drops and remains as
low as that by the method for (P2).
Figure 8(c) presents the maximum value amongst the number of transitioning agents via each (one-way) path. The
red line indicates the actual result by the method for (P2),
while the green line indicates the corresponding probabilistic
value (i.e., max∀i max∀l6=i nk [i]Pkj [i, l]). It is shown that the
stochastic decision policies reflect the given upper bound,
meanwhile this bound is often violated in practice due to the
finite cardinality of the agents. However, the result in the same
scenario with setting |A| = 100, 000 and c(i,l) = 200, ∀i 6= l
(denoted by Case 2), depicted by the blue and magenta lines
in Figure 8(c), suggests that such violation can be mitigated
14
as the number of given agents increases.
VII. C ONCLUSION
This paper poposed a LICA-based closed-loop-type framework for probabilistic swarm distribution guidance. Since the
feedback gains can be generated based on local information,
agents have shorter and different timescales for using new
information, and can incorporate an asynchronous decisionmaking process. Even using local information, the proposed
framework converges to a desired density distribution, while
maintaining scalability, robustness, and long-term system efficiency. Furthermore, the numerical experiments have showed
that the proposed framework is suitable for a realistic environment where communication between agents is partially and
temporarily disconnected. This paper has explicitly presented
the design requirements for the Markov matrix to hold all these
advantages, and has provided specific problem examples of
how to implement this framework.
Future works include optimisation of ξ¯kj [i], which can mitigate the trade-off between convergence rate and residual error.
In addition, it is expected that the communication cost required
for the proposed framework can be reduced by incorporating
a vision-based local density estimation [38].
ACKNOWLEDGEMENT
The authors gratefully acknowledge that this research was
supported by International Joint Research Programme with
Chungnam National University (No. EFA3004Z). Thanks to
Sangjun Bae for supportive discussions.
A PPENDIX
A. Regarding the Convergence Analysis in Theorem 1
Definition 9 (Irreducible). A matrix is reducible if and only
if its associated digraph is not strongly connected. A matrix
that is not reducible is irreducible.
Definition 10 (Primitive). A primitive matrix is a square
nonnegative matrix A such that for every i, j there exists k > 0
such that Ak [i, j] > 0.
Definition 11 (Regular). A regular matrix is a stochastic
matrix such that all the entries of some power of the matrix
are positive.
Definition 12. [33, pp.92, 149] [13] (Asymptotic Homogeneity) “A sequence of stochastic matrices Mk ∈ Mn×n , k ≥ k0 ,
is said to be asymptotically homogeneous (with respect to
d) if there exists a row-stochastic vector d ∈ Pn such that
limk→∞ dMk = d.”
Definition 13. [33, pp.92, 149] [13] (Strong Ergodicity)
The matrix product Uk0 ,k := Mk0 Mk0 +1 · · · Mk−1 , formed
from a sequence of stochastic matrices Mr ∈ Pn×n , r ≥
k0 , is said to be strongly ergodic if for each i, l, r we
get limr→∞ Uk0 ,r [i, l] = v[l], where v ∈ Pn is a rowstochastic vector. Here, v is called its unique limit vector (i.e.,
limr→∞ Uk0 ,r = 1> v).
Lemma 5. Given the requirements (R1)-(R4) are satisfied, Mkj
in Equation (11) has the following properties:
1) row-stochastic;
2) irreducible;
3) all diagonal elements are positive, and all other elements
are non-negative;
4) there is a positive lower bound γ such that 0 < γ ≤
j
+
min+
refers to the minimum
i,l Mk [i, l] (Note that min
of the positive elements);
5) asymptotically homogeneous with respect to Θ.
In addition, Ukj0 ,k in Equation (12) has the following
properties:
6) irreducible;
7) all diagonal elements are positive, and all other elements
are non-negative;
8) primitive [39, Lemma 8.5.4, p.541]
Proof. This lemma can be proved by similarly following the
mathematical development for [13, Theorem 4]. Mkj is rowstochastic because Mkj · 1> = (I − Wkj )Pkj · 1> + Wkj Skj · 1> =
(I − Wkj ) · 1> + Wkj · 1> = 1> . Pkj is irreducible and
ωkj [i] is always less than 1, thus Mkj is also irreducible
(i.e., Mkj [i, l] > 0 if Pkj [i, l] > 0). The property 3) is true
because diag(I − Wkj ) > 0, Wkj ≥ 0, and Pkj is also a nonnegative matrix such that its diagonal elements are positive.
The property 4) is implied by either the property 2) or 3). From
the definition of Wkj , it follows that limk→∞ Wkj = 0 (because
of exp(−τ j k)), and thereby limk→∞ Mkj = limk→∞ Pkj .
Hence, limk→∞ ΘMkj = limk→∞ ΘPkj = Θ, and the property
5) is valid.
Let us now turn to Ukj0 ,k . It is irreducible due to the
fact that if Mrj [i, l] > 0 for some r ∈ {k0 , ..., k − 1} and
∀i, l ∈ {1, ..., nbin }, then the corresponding element Ukj0 ,k [i, l]
is greater or equal to the product of positive diagonal elements
and Mrj [i, l] and the property 2). The property 7) is true
because Ukj0 ,k is a product of nonnegative matrices where all
diagonal entries are positive. It follows from [39, Lemma 8.5.4,
p.541] that Ukj0 ,k is primitive: “if a square matrix is irreducible,
nonnegative and all its main diagonal entries are positive, then
the matrix is primitive”.
R EFERENCES
[1] M. Brambilla, E. Ferrante, M. Birattari, and M. Dorigo, “Swarm
Robotics: a Review from the Swarm Engineering Perspective,” Swarm
Intelligence, vol. 7, no. 1, pp. 1–41, 2013.
[2] A. Kolling, P. Walker, N. Chakraborty, K. Sycara, and M. Lewis,
“Human Interaction With Robot Swarms: A Survey,” IEEE Transactions
on Human-Machine Systems, vol. 46, no. 1, pp. 9–26, 2016.
[3] E. Sahin, “Swarm Robotics: From Sources of Inspiration to Domains of
Application,” in Swarm Robotics. Berlin: Springer, 2005, pp. 10–20.
[4] B. Acikmese and D. S. Bayard, “A Markov Chain Approach to Probabilistic Swarm Guidance,” in American Control Conference, Montreal,
Canada, 2012, pp. 6300–6307.
[5] ——, “Probabilistic Swarm Guidance for Collaborative Autonomous
Agents,” in American Control Conference, Portland, OR, USA, 2014,
pp. 477–482.
[6] ——, “Markov Chain Approach to Probabilistic Guidance for Swarms
of Autonomous Agents,” Asian Journal of Control, vol. 17, no. 4, pp.
1105–1124, 2015.
[7] I. Chattopadhyay and A. Ray, “Supervised Self-Organization of Homogeneous Swarms Using Ergodic Projections of Markov Chains,” IEEE
Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics),
vol. 39, no. 6, pp. 1505–1515, 2009.
15
[8] N. Demir and B. Acikmese, “Probabilistic Density Control for Swarm
of Decentralized ON-OFF Agents with Safety Constraints,” in American
Control Conference, Chicago, IL, USA, 2015, pp. 5238–5244.
[9] R. Luo, N. Chakraborty, and K. Sycara, “Supervisory Control for CostEffective Redistribution of Robotic Swarms,” in IEEE International
Conference on Systems, Man, and Cybernetics (SMC), San Diego, CA,
USA, 2014, pp. 596–601.
[10] N. Demir, U. Eren, and B. Açkmee, “Decentralized Probabilistic Density
Control of Autonomous Swarms with Safety Constraints,” Autonomous
Robots, vol. 39, no. 4, pp. 537–554, 2015.
[11] S. Bandyopadhyay, S.-J. Chung, and F. Y. Hadaegh, “Inhomogeneous
Markov Chain Approach To Probablistic Swarm Guidance Algorithms,”
in 5th Int. Conf. on Spacecraft Formation Flying Missions and Technologies (SFFMT), Munich, Germany, 2013.
[12] D. Morgan, G. P. Subramanian, S. Bandyopadhyay, S.-J. Chung, and
F. Y. Hadaegh, “Probabilistic Guidance of Distributed Systems using
Sequential Convex Programming,” in IEEE/RSJ International Conference on Intelligent Robots and Systems, Chicago, IL, USA, 2014, pp.
3850–3857.
[13] S. Bandyopadhyay, S.-J. Chung, and F. Y. Hadaegh, “Probabilistic and
Distributed Control of a Large-Scale Swarm of Autonomous Agents,”
IEEE Transactions on Robotics, vol. 33, no. 5, pp. 1103–1123, 2017.
[14] A. Halasz, M. A. Hsieh, S. Berman, and V. Kumar, “Dynamic Redistribution of a Swarm of Robots among Multiple Sites,” in IEEE/RSJ
International Conference on Intelligent Robots and Systems, San Diego,
CA, USA, 2007, pp. 2320–2325.
[15] M. A. Hsieh, A. Halasz, S. Berman, and V. Kumar, “Biologically
Inspired Redistribution of a Swarm of Robots among Multiple Sites,”
Swarm Intelligence, vol. 2, no. 2-4, pp. 121–141, 2008.
[16] S. Berman, A. Halasz, M. Ani Hsieh, and V. Kumar, “Navigation-based
Optimization of Stochastic Strategies for Allocating a Robot Swarm
among Multiple Sites,” in 47th IEEE Conference on Decision and
Control, Cancun, Mexico, 2008, pp. 4376–4381.
[17] S. Berman, A. Halasz, M. A. Hsieh, and V. Kumar, “Optimized
Stochastic Policies for Task Allocation in Swarms of Robots,” IEEE
Transactions on Robotics, vol. 25, no. 4, pp. 927–937, 2009.
[18] T. W. Mather and M. A. Hsieh, “Macroscopic Modeling of Stochastic
Deployment Policies with Time Delays for Robot Ensembles,” The
International Journal of Robotics Research, vol. 30, no. 5, pp. 590–
600, 2011.
[19] K. Lerman, A. Martinoli, and A. Galstyan, “A Review of Probabilistic
Macroscopic Models for Swarm Robotic Systems,” in Swarm Robotics.
Berlin: Springer, 2005, pp. 143–152.
[20] L. B. Johnson, H.-L. Choi, and J. P. How, “The Role of Information Assumptions in Decentralized Task Allocation: A Tutorial,” IEEE Control
Systems, vol. 36, no. 4, pp. 45–58, 2016.
[21] S. Xue, C. Sun, J. Zeng, Y. Jin, and R. Cheng, “Effect of Communication
Modes to Swarm Robotic Search,” The Open Electrical & Electronic
Engineering Journal, vol. 8, no. 1, pp. 240–244, 2014.
[22] B. I. Koh, A. D. George, R. T. Haftka, and B. J. Fregly, “Parallel
Asynchronous Particle Swarm Optimization,” International Journal for
Numerical Methods in Engineering, vol. 67, no. 4, pp. 578–595, 2006.
[23] L. Johnson, S. Ponda, H.-L. Choi, and J. How, “Asynchronous
Decentralized Task Allocation for Dynamic Environments,” in Infotech@Aerospace 2011, St.Louis, MO, USA, 2011.
[24] I. D. Couzin, J. Krause, R. James, G. D. Ruxton, and N. R. Franks,
“Collective Memory and Spatial Sorting in Animal Groups.” Journal of
theoretical biology, vol. 218, no. 1, pp. 1–11, 2002.
[25] I. D. Couzin, J. Krause, N. R. Franks, and S. A. Levin, “Effective
Leadership and Decision-making in Animal Groups on the Move,”
Nature, vol. 433, no. 7025, pp. 513–516, 2005.
[26] J. Gautrais, C. Jost, and G. Theraulaz, “Key Behavioural Factors in a
Self-Organised Fish School Model,” Annales Zoologici Fennici, vol. 45,
no. 5, pp. 415–428, 2008.
[27] D. J. Hoare, I. D. Couzin, J. G. J. Godin, and J. Krause, “Contextdependent Group Size Choice in Fish,” Animal Behaviour, vol. 67, no. 1,
pp. 155–164, 2004.
[28] S. Bandyopadhyay and S.-J. Chung, “Distributed Estimation using
Bayesian Consensus Filtering,” American Control Conference, pp. 634–
641, 2014.
[29] T. D. Seeley, The Wisdom of the Hive. Cambridge, Massachusetts:
Havard University Press, 1995.
[30] L. Keller, M. J. B. Krieger, and J.-B. Billeter, “Ant-like Task Allocation
and Recruitment in Cooperative Robots,” Nature, vol. 406, no. 6799,
pp. 992–995, 2000.
[31] B. L. Partridge, “The Structure and Function of Fish Schools,” Scientific
American, vol. 246, no. 6, pp. 114–123, 1982.
[32] C. Becco, N. Vandewalle, J. Delcourt, and P. Poncin, “Experimental
Evidences of a Structural and Dynamical Transition in Fish School,”
Physica A: Statistical Mechanics and its Applications, vol. 367, pp. 487–
493, 2006.
[33] E. Seneta, Non-negative Matrices and Markov Chains, ser. Springer
Series in Statistics. New York, NY: Springer New York, 1981.
[34] Y. Bestaoui Sebbane, Planning and Decision Making for Aerial Robots,
ser. Intelligent Systems, Control and Automation: Science and Engineering. Cham: Springer International Publishing, 2014, vol. 71.
[35] I. C. F. Ipsen and T. M. Selee, “Ergodicity Coefficients Defined by Vector
Norms,” SIAM Journal on Matrix Analysis and Applications, vol. 32,
no. 1, pp. 153–200, jan 2011.
[36] A. Jadbabaie, Jie Lin, and A. Morse, “Coordination of Groups of Mobile
Autonomous Agents using Nearest Neighbor Rules,” IEEE Transactions
on Automatic Control, vol. 48, no. 6, pp. 988–1001, 2003.
[37] J. Chung, P. Kannappan, C. Ng, and P. Sahoo, “Measures of Distance
between Probability Distributions,” Journal of Mathematical Analysis
and Applications, vol. 138, no. 1, pp. 280–292, 1989.
[38] S. A. M. Saleh, S. A. Suandi, and H. Ibrahim, “Recent Survey on Crowd
Density Estimation and Counting for Visual Surveillance,” Engineering
Applications of Artificial Intelligence, vol. 41, pp. 103–114, 2015.
[39] R. A. Horn and C. R. Johnson, Matrix Analysis, 2nd ed. Cambridge:
Cambridge University Press, 2012.
| 10math.ST |
arXiv:1706.02410v1 [math.ST] 7 Jun 2017
A SHARP MULTIPLIER INEQUALITY WITH
APPLICATIONS TO HEAVY-TAILED REGRESSION
PROBLEMS
QIYANG HAN AND JON A. WELLNER
Abstract. We develop a sharp multiplier
Pinequality used to study the
size of the multiplier empirical process ( n
i=1 ξi f (Xi ))f ∈F , where ξi ’s
and F are multipliers and an indexing
function
class respectively. We
Pn
show that in general the size of E
i=1 ξi f (Xi ) F is determined jointly
by the growth rate of the corresponding empirical process, and the size
of the maxima of the multipliers.
The new inequality sheds light on the long-standing open question
concerning the rate behavior of classical empirical risk minimization procedures in the regression setting with heavy-tailed errors. In particular,
the convergence rate of the least squares estimator in a general nonparametric regression model, with usual entropy conditions on the function classes, is shown to be no worse than the maximum of the optimal
Gaussian regression rate and the noise level of the errors. Furthermore,
this rate cannot be improved in general. We also give an application to
the sparse linear regression model where both design and errors can be
heavy-tailed.
1. Introduction
1.1. The questions and motivation. In this paper we are interested in
the multiplier empirical process
(1.1)
n
X
ξi f (Xi )
i=1
indexed by a class F of measurable functions f : X → R. Here the real
valued multipliers ξ1 , . . . , ξn are independent of X -valued random variables
X1 , . . . , Xn . We do not impose any apriori structural assumptions on ξi ’s,
e.g. ξi can be arbitrarily correlated with each other. Neither do we impose
any structural assumptions on F.
The first question we address is the following.
Date: June 9, 2017.
2010 Mathematics Subject Classification. 60E15, 62G05.
Key words and phrases. multiplier empirical process, multiplier inequality, empirical
risk minimization, nonparametric regression, sparse linear regression, heavy-tailed errors.
Supported in part by NSF Grant DMS-1104832, DMS-1566514 and NI-AID grant R01
AI029168.
1
2
Q. HAN AND J. A. WELLNER
Question 1. Under what moment conditions on the ξi ’s can we assert that,
the multiplier empirical process (1.1) has roughly the same size as the empirical process counterpart for (nearly) all function classes F in a nonasymptotic manner?
Question 1 has been well understood in the asymptotic regime where
ξ1 , . . . , ξn are i.i.d. and F and X are fixed: the central limit theorem (CLT)
for the multiplier empirical process holds if and only if a CLT holds for the
empirical process, given the moment condition kξ1 k2,1 < ∞ (see Section 1.3
for notation), see [36] or Section 2.9 of [68].
Much less work has addressed non-asymptotic aspects concerning (1.1).
In deriving CLTs for the multiplier empirical process, a multiplier inequality relating the modulus of continuity of the multiplier empirical process
with that of the original empirical process is used, see [26, 27], or Lemma
2.9.1 of [68]. This inequality typically suffers from a huge loss in terms of
rates of convergence in the finite sample setting, see Section 5.1. In a recent
work of [49], (1.1) is studied using a multiscale truncated analogue of Talagrand’s γ2 functional only assuming weak moments on the ξi ’s. However,
the bound obtained in [49] does not take into account the Poisson (small
sample) behavior of the empirical process and hence reveals some difficulties in handling general function classes whose element-wise L2 size is much
smaller than its L∞ size, see Section 5.2.
The motivation for studying Question 1 arises from applications in statistical theory. In the empirical risk minimization (ERM) literature, one of
the long-standing fundamental questions is the following.
Question 2. What is the convergence rate of the nonparametric least squares
estimator(LSE) over a general regression function class with the presence
of heavy-tailed errors?
The nonparametric LSE is of fundamental importance in both statistical
theory and application, and serves as the building block for a large class
of more complicated penalized/aggregated estimators. By now, calculation
of rates of convergence for the LSE under light-tailed errors has been wellunderstood: we only need to evaluate the size of a localized Rademacher
or Gaussian process, e.g. [6, 7, 8, 30, 31, 43, 63, 65, 68]. This corresponds
to taking ξ1 , . . . , ξn to be i.i.d. Rademacher or Gaussian random variables
in (1.1). A precise understanding of Question 1 for general multipliers ξi ’s
would naturally lead to improved understanding towards Question 2.
Under heavy-tailed assumptions, one line of research [34, 45, 46, 47] asserts rate-optimality of related ERM procedures using least squares criteria, under a key structural assumption on the function class–namely Lsubgaussian or a small-ball condition. This shows that there are function
classes over which the LSE performs well under heavy-tailed errors. This
approach presents, to some degree, a best-case analysis in the sense that
the assumptions can fail for many other commonly used function classes
including (i) VC-subgraph classes, (ii) shape-restricted classes, (iii) smooth
MULTIPLIER INEQUALITIES
3
classes; see Section 5.2. Another line of research [3, 10, 11, 14, 15, 20, 28,
29, 38, 39, 40, 50] proposes alternative procedures with guaranteed optimal
rates of convergence in heavy-tailed problems (and some with subgaussian
performance), showing that there are good estimators other than the LSE
for general function classes under heavy-tailed errors. To date, it seems
that neither of the approaches gives a definite answer to the fundamental
Question 2.
1.2. Our contribution. 1 We attempt to make progress on the difficulties
2 by developing
mentioned above. We start addressing Question 1 in SectionP
a sharp multiplier inequality characterizing the size of E k ni=1 ξi f (Xi )kF .
Very informally,
E
n
X
ξi f (Xi )
i=1
F
≈ max ‘growth rate’ of k 7→ E
k
X
i=1
εi f (Xi )
F
, E max |ξi | .
1≤i≤n
In particular, the size of the multiplier empirical process is determined not
only by the size of the corresponding empirical process, but also by the
moments of the multipliers. Technically, we establish upper bounds in Theorems 1 and 2 for structureless multipliers ξi ’s and indexing function classes
F. A matching lower bound, established in Theorem 3, shows that the
moment barrier revealed by the multipliers in the size of the multiplier empirical process, is unimprovable without additional structural assumptions
on F.
As indicated before, a clear understanding of the size of the multiplier empirical process is crucial in addressing Question 2. Using our new multiplier
inequality, we show (see Theorem 4) that if the errors have finite Lp,1 (p ≥ 1)
moments, and the underlying function class is uniformly bounded with an
entropy of order ε−α for some α ∈ (0, 2), the rate of convergence for the LSE
in L2 metric is no worse than
1
1
1
OP n− 2+α ∨ n− 2 + 2p .
Furthermore, this rate is in general unimprovable (see Theorem 5). This
rate offers the first quantitative answer to the fundamental Question 2 for
a general regression function class, and shows a clear tradeoff between the
complexity of the function class and the heaviness of the tail of the errors.
We note that the uniform boundedness assumption of the function class is
more of a cliché of the localization property of the estimator : all that we
need is stochastic boundedness of the LSE.
1For notation see Section 1.3.
4
Q. HAN AND J. A. WELLNER
For least squares estimation in non-parametric regression models, we have
restricted our attention to Donsker classes. Such a situation changes drastically in the sparse linear regression model, where the empirical process part
of the Lasso problem is non-Donsker in the sense that its supremum over
the function class in question becomes unbounded as the problem dimension
grows. We suggest a new proof method that combines our new multiplier
inequality with Gaussian approximation techniques to quantify the growth
of the empirical process when both the design matrix and the errors can
be heavy-tailed. The new technique allows us to prove rate-optimality of
the Lasso estimator under a wide range of distributional assumptions on the
design and the errors.
The approach we adopt here, per se, can be viewed as a ‘worst-case’
analysis for the size of the multiplier empirical process, and the rate behavior
of the least squares estimator in the presence of heavy-tailed errors. Such
a view point naturally complements the ‘best-case’ analysis conducted in
[34, 45, 46, 47, 48]. We hope future contributions will go further in the
different directions suggested by both approaches.
1.3. Notation. For a real-valued random variable ξ and 1 ≤ p < ∞, let
1/p
kξkp := E |ξ|p
denote the ordinary p-norm. The Lp,1 norm for a random
variable ξ is defined by
Z ∞
P(|ξ| > t)1/p dt.
kξkp,1 :=
0
It is well known that Lp+ε ⊂ Lp,1 ⊂ Lp holds for any underlying probability
measure, and hence a finite Lp,1 condition requires slightly more than a p-th
moment, but no more than any p + ε moment, see Chapter 10 of [37].
For a real-valued measurable function f defined on (X , A, P ), kf kLp (P ) ≡
kf kP,p ≡ P |f |p )1/p denotes the usual Lp -norm under P , and kf k∞ ≡
supx∈X |f (x)|. f is said to be P -centered if P f = 0. Lp (g, B) denotes the
Lp (P )-ball centered at g with radius B. For simplicity we write Lp (B) ≡
Lp (0, B). To avoid unnecessary measurability digressions, we will assume
that F is countable throughout the article. As usual, for any φ : F → R,
we write kφ(f )kF for supf ∈F |φ(f )|.
Throughout the article ε1 , . . . , εn will be i.i.d. Rademacher random variables independent of all other random variables. Cx will denote a generic
constant that depends only on x, whose numeric value may change from
line to line unless otherwise specified. a .x b and a &x b mean a ≤ Cx b
and a ≥ Cx b respectively, and a ≍x b means a .x b and a &x b [a . b
means a ≤ Cb for some absolute constant C]. For two real numbers a, b,
a ∨ b ≡ max{a, b} and a ∧ b ≡ min{a, b}.
1.4. Organization. Section 2 is devoted to the new multiplier inequality,
while Sections 3-4 are devoted to applications of the multiplier inequalities
to nonparametric and sparse linear regression models, respectively. We make
MULTIPLIER INEQUALITIES
5
some further remarks and comparisons in Section 5. Most detailed proofs
are relegated to Sections 6-7.
2. The multiplier inequality
2.1. Upper bound. We first state the assumptions.
Assumption A. Suppose that ξ1 , . . . , ξn are independent of the random variables X1 , . . . , Xn , and either of the following conditions holds:
(A1) X1 , . . . , Xn are i.i.d. with law P on (X , A), and F is P -centered.
(A2) X1 , . . . , Xn are permutation invariant, and ξ1 , . . . , ξn are independent mean-zero random variables.
Theorem 1. Grant Assumption A. Let F1 ⊃ · · · ⊃ Fn be a non-increasing
sequence of function classes. Assume further that there exist non-decreasing
concave functions {ψn } : R≥0 → R≥0 with ψn (0) = 0 such that
k
X
E
(2.1)
εi f (Xi )
i=1
Fk
holds for all 1 ≤ k ≤ n. Then
E
(2.2)
n
X
ξi f (Xi )
i=1
where Γξn (t) ≡
1
n
Fn
Pn
≤4
|ξi | > t .
i=1 P
Z
∞
0
≤ ψn (k)
ψn nΓξn (t) dt,
The formulation of Theorem 1 requires non-increasing {Fk }. This is suitable for studying the size of the multiplier empirical process indexed by
function classes with a shrinking L2 size. It is also possible to use Theorem
1 for the case F1 = · · · = Fn ; see Section 4 for an application to the sparse
linear regression model.
The key ingredient in the proof of Theorem 1 is the following, which may
be of independent interest.
Proposition 1. Grant Assumption A. For any function class F,
" n
n
k
X
X
X
(2.3)
ξi f (Xi )
≤E
E
εi f (Xi )
( η(k) − η(k+1) )E
i=1
F
k=1
i=1
F
#
where η(1) ≥ · · · ≥ η(n) ≥ η(n+1) ≡ 0 are the reversed order statistics for
(i) (under (A1)) {2 |ξi |}ni=1 , (ii) (under (A2)) {|ξi − ξi′ |}ni=1 with {ξi′ } being
an independent copy of {ξi }.
In the simplest case where ξi ’s are uniformly bounded, (2.3) implies that
the multipliers can be ‘replaced’ by Rademachers. This is the content of
the classical contraction principle (see Section 3.1.4 of [24]). In the general
case for unbounded multipliers, (2.3) suggests that the size of the empirical
Pk
process E
needs to be adjusted with some measure for the
i=1 εi f (Xi )
F
6
Q. HAN AND J. A. WELLNER
heaviness of the tails of the ξi ’s. Hence Proposition 1 can also be viewed
as an interpolation inequality from the contraction principle for bounded
multipliers to unbounded multipliers. We note that in the simple case for
i.i.d. multipliers, an inequality in the spirit of (2.3) has implicitly been used
in [27], see also page 177-178 of [68].
A canonical example in applications is the following.
Theorem 2. Consider the same assumptions as in Theorem 1. Assume for
simplicity that ξi ’s have the same marginal distributions. Suppose that for
some γ ≥ 1, and some constant κ0 > 0,
E
(2.4)
k
X
i=1
εi f (Xi )
Fk
≤ κ0 · k1/γ
holds for all 1 ≤ k ≤ n. Then for any p ≥ 1 such that kξ1 kp,1 < ∞,
E
n
X
i=1
ξi f (Xi )
Fn
≤ 4κ0 · nmax{1/γ,1/p} kξ1 kmin{γ,p},1 .
Proof. First consider
ψn (t) ≡ κ0 t1/γ in Theorem
Pn γ ≤ p. In this case, letting
1/γ
1, we see that E k i=1 ξi f (Xi )kFn ≤ 4κ0 · n kξ1 kγ,1 . On the other hand,
P
if γ > p, we can take ψn (t) ≡ κ0 t1/p to conclude that E k ni=1 ξi f (Xi )kFn ≤
4κ0 · n1/p kξ1 kp,1 . Note that γ ≥ 1 ensures the concavity of ψn .
Theorem 2 says that the upper bound for the multiplier empirical process
has two components: one part comes from the growth rate of the empirical process; another part comes from the moment barrier of the multipliers ξi ’s. Note that the growth rate n1/p corresponds to the worst size of
E max1≤i≤n |ξi | under a p-th moment condition on ξi ’s.
Remark 1. Choosing γ ≥ 2 in Theorem 2 corresponds to the bounded
Donsker regime2 for the empirical process. In this case we only need kξ1 k2,1 <
∞ to ensure the multiplier empirical process to be also bounded Donsker.
This moment condition is generally unimprovable in view of [35]. On the
other hand, such a choice of γ can fail due to: (i) failure of integrability
of the envelope functions of the classes {Fk }, or (ii) failure of the classes
{Fk } to be bounded Donsker. (i) is related to the classical MarcinkiewiczZygmund strong laws of large numbers and the generalizations of those to
empirical measures, see [2, 41, 44]. For (ii), some examples in this regard
can be found in [57], see also Proposition 17.3.7 of [55].
Remark 2. An important assumption we make in Theorems 1 and 2 is the
independence of the multipliers (ξ1 , . . . , ξn ) and the samples (X1 , . . . , Xn ).
We conjecture that the independence assumption cannot be removed based
on a result of [1]: there is no universal multiplier moment for the CLT.
2F is said to be bounded Donsker if sup
n∈N E supf ∈F
√1
n
Pn
i=1
εi f (Xi ) < ∞.
MULTIPLIER INEQUALITIES
7
Remark 3. Although Theorems 1 and 2 are formulated in a non-asymptotic
fashion, they can also be used to examine asymptotic equicontinuity of the
multiplier empirical process; see Section 5.1 for more details.
Remark 4. One particular case for application of Theorems 1 and 2 is the
following. Let δ1 ≥ . . . ≥ δn ≥ 0 be a sequence of non-increasing nonnegative real numbers, and F be an arbitrary function class. Let Fk ≡
F(δk ) ≡ {f ∈ F : P f 2 < δk2 } be the ‘local’ set of F with L2 -radius at most
δk . An important choice in statistical applications for δk is given by
(2.5)
E
k
X
. kδk2 .
εi f (Xi )
i=1
F (δk )
As will be seen in Section 3, the above choice {δk } corresponds to the rate
of convergence for the LSE in a general nonparametric regression model. In
this case Theorems 1 and 2 yield that
(2.6)
E
n
X
. nδn2
ξi f (Xi )
i=1
given sufficient moments of the ξi ’s.
F (δn )
Theorems 1 and 2 only concern the first moment of the suprema of the
multiplier empirical process. For higher moments, we may use the following
Hoffmann-Jørgensen-Talagrand type inequality relating the q-th moment
estimate with the first moment estimate in some independent cases.
Lemma 1 (Proposition 3.1 of [23]). Suppose X1 , . . . , Xn are i.i.d. with law
P and ξ1 , . . . , ξn are i.i.d. with kξ1 k2 < ∞. Let F be a P -centered, countable
class of class of functions such that supf ∈F P f 2 ≤ σ 2 . Then for any q ≥ 1,
!q
q
n
n
X
X
q
ξi f (Xi )
ξi f (Xi ) ≤ K
E sup
E sup
f ∈F i=1
f ∈F i=1
+q
q/2
√
q
q
q
q
( n kξ1 k2 σ) + q E max |ξi | sup |f (Xi )| .
Here K > 0 is a universal constant.
1≤i≤n
f ∈F
2.2. Lower bound. Theorems 1 and 2 do not require any structural assumptions on the function class F. [49] showed that for a ‘L-subgaussian’
class, a 2 + ε moment on i.i.d. ξi ’s suffices to conclude that the multiplier
empirical process behaves like the canonical Gaussian process (see Section
5.2 for more details). One may therefore wonder if the moment barrier for
the multipliers in Theorem 2 is due to an artifact of the proof. Below in
Theorem 3 we show that this barrier is intrinsic for general classes F.
Theorem 3. Let X = [0, 1]d and P be a probability measure on X with
Lebesgue density bounded away from 0 and ∞. Let ξ1 , . . . , ξn be i.i.d. random variables such that E max1≤i≤n |ξi | ≥ κ0 n1/p for some p > 1 and some
8
Q. HAN AND J. A. WELLNER
p
complexity of function class dominates=n1/γ
2
moment barrier dominates=n1/p
2
α=
2
γ−1
Figure 1. Tradeoff between the complexity of the function
class and the moment of the multipliers in the size of the
multiplier empirical process. The critical curve (purple): p =
1 + 2/α.
constant κ0 independent of ξ1 . Then for any γ > 2, there exists a sequence
of function classes F1 ⊃ · · · ⊃ Fn all defined on X such that for n large
enough,
E
k
X
εi f (Xi )
i=1
Fk
≤ κ1 · k1/γ ,
holds for all 1 ≤ k ≤ n, and that
E
n
X
i=1
ξi f (Xi )
Fn
max{1/γ,1/p}
≥ κ−1
.
1 n
Here κ1 is a constant depending on κ0 , γ and P .
Remark 5. The condition on ξi ’s will be satisfied, e.g. if the ξi ’s are i.i.d.
with the tail condition P(|ξi | > t) ≥ κ′0 /(1 + tp ) for t > 0.
Combined with Theorem 2, it is seen that the growth rate nmax{1/γ,1/p}
of the multiplier empirical process cannot be improved in general. This
suggests an interesting phase transition phenomenon (cf. Figure 1): if the
complexity of the function class dominates the effect of the tail of the multipliers, then the multiplier empirical process essentially behaves as the empirical process counterpart; otherwise the size of the maxima of the multipliers
governs the growth of the multiplier empirical process in general.
3. Nonparametric regression: least squares estimation
In this section, we consider the nonparametric regression model:
(3.1)
Yi = f0 (Xi ) + ξi ,
i = 1, . . . , n
MULTIPLIER INEQUALITIES
9
where f0 ∈ F, X1 , . . . , Xn are i.i.d. random variables from P , and ξ1 , . . . , ξn
are mean-zero errors. Let fˆn be any least squares estimator (LSE) over F:
n
2
1X
Yi − f (Xi ) .
fˆn ∈ arg min
f ∈F n
i=1
We assume without loss of generality that fˆn is well-defined.
As discussed in the Introduction, there seems to be no satisfactory solution addressing the question, as to when the LSE performs ‘well’ or ‘badly’
under substantially weaker moment assumptions on the errors for a general
function class. Our theorem below tries to shed some light on this problem
in the realm where the function class F satisfies entropy conditions.
Let (F, k·k) be a subset of the normed space of real functions f : X →
R. Let N (ε, F, k·k) be the ε-covering number, and N[ ] (ε, F, k·k) be the
ε-bracketing number; see page 83 of [68] for more details.
Assumption B. Suppose that F ⊂ L∞ (1) satisfies either of the following
conditions: there exists some α ∈ (0, 2) such that for any ε ∈ (0, 1),
(F1) supQ log N ε kF kL2 (Q) , F, L2 (Q) . ε−α .
(F2) log N[ ] (ε, F, L2 (P )) . ε−α .
Here the supremum in (F1) is taken over all finitely discrete probability
measures Q, and F denotes the envelope function for F.
Theorem 4. Suppose that ξ1 , . . . , ξn are mean-zero errors independent of
X1 , . . . , Xn with the same marginal distributions, and kξ1 kp,1 < ∞ for some
p ≥ 1. Further suppose that F is a P -centered function class such that
F − f0 ⊂ L∞ (1) satisfies Assumption B with some α ∈ (0, 2). Then
1
−1+ 1
= OP n− 2+α ∨ n 2 2p .
fˆn − f0
(3.2)
L2 (P )
If ξi ’s are i.i.d. F need not be P -centered. Furthermore, if ξi ’s are i.i.d.
and p ≥ 2, then (3.2) holds in expectation.
The proof relies on the following reduction scheme.
Proposition 2. Suppose that ξ1 , . . . , ξn are mean-zero random variables
independent of X1 , . . . , Xn , and F − f0 ⊂ L∞ (1). Further assume that
n
(3.3)
E
sup
f ∈F :kf −f0 kL
2 (P )
≤δ
and
1 X
√
ξi (f − f0 )(Xi ) . φn (δ),
n
i=1
n
(3.4)
E
sup
f ∈F :kf −f0 kL
2 (P )
≤δ
1 X
√
εi (f − f0 )(Xi ) . φn (δ).
n
i=1
=
hold for some φn such that δ 7→ φn (δ)/δ is non-increasing. Then fˆn − f0
L2 (P )
√ 2
OP (δn ) holds for any δn such that φn (δn ) ≤ nδn . Furthermore, if ξ1 , . . . , ξn
10
Q. HAN AND J. A. WELLNER
are i.i.d. mean-zero with kξ1 kp < ∞ for some p ≥ 2, then E fˆn − f0
1
1
√
O(δn2 ) for any δn ≥ n− 2 + 2p such that φn (δn ) ≤ nδn2 .
2
L2 (P )
=
The remaining task in the proof of Theorem 4 is a calculation of the
modulus of continuity of the (multiplier) empirical process involved in (3.3)
and (3.4) using Theorem 1 and local maximal inequalities for the empirical
process (see Proposition 5).
Remark 6. Some remarks on the assumptions on F.
(1) Assumption B requires F to be a Donsker class. Although the proof
applies to non-Donsker function classes with α ≥ 2, the first term in
(3.2) becomes sub-optimal in general, see [8].
(2) F is assumed to be P -centered to allow errors with arbitrary dependence structure. It is known from [69] (see Theorem 1, page 638)
that for a centered function class, the minimax risk of estimating
a regression function under arbitrary errors with second moments
uniformly bounded, is no worse than that for i.i.d. Gaussian errors.
If the errors are i.i.d., then F need not be P -centered.
(3) The uniform boundedness assumption on F includes many classical
examples, cf. Section 9.3 of [65]. In fact, this is more of a localiza= OP (1).
tion property for the estimator : all that we need is fˆn
∞
Proving stochastic boundedness of the LSE usually serves as the first
step in deriving a rate of convergence and this may require casespecific techniques. For instance, for isotonic regression on [0, 1],
stochastic boundedness follows easily from the min-max representation of [53] if kf0 k∞ < ∞.
At this point, (3.2) only serves as an upper bound for the convergence
1
rates of the LSE. Since the rate n− 2+α corresponds to the optimal rate in
the Gaussian regression case [70], it is natural to conjecture that this rate
−1+ 1
cannot be improved. On the other hand, the ‘noise’ rate n 2 2p is due to
the reduction scheme Proposition 2, which relates the convergence rate of
the LSE to the size of the multiplier empirical process involved. It seems
natural to wonder if this noise rate is a proof artifact due to the deficiency
in Proposition 2.
Theorem 5. Let X = [0, 1]d and P be a probability measure on X with
Lebesgue density bounded away from 0 and ∞. Suppose that ξ1 , . . . , ξn are
i.i.d. mean-zero random variables with kξ1 kp,1 < ∞ for some p ≥ 2.
(1) For each α ∈ (0, 2), there exists a function class F and some f0 ∈ F
such that F − f0 satisfies Assumption B, and that for p ≥ 1 + 2/α,
there exists some least squares estimator fn∗ over F satisfying
1
E kfn∗ − f0 kL2 (P ) ≥ ρ · n− 2+α .
Here ρ > 0 is a (small) constant independent of n.
MULTIPLIER INEQUALITIES
11
p
1
Gaussian rate =n− 2+α
2
1
1
noise rate=n− 2 + 2p
2
α
Figure 2. Tradeoff between the complexity of the function
class and the noise level of the errors in the convergence rates
for the LSE. The critical curve (purple): p = 1 + 2/α.
(2) There exists a function class F ≡ Fn contained in a fixed uniformly
bounded VC-subgraph function
√ class, and some f0 ∈ F such that
the following holds: suppose log n ≤ p ≤ (log n)1−δ for some δ ∈
(0, 1/2). Then there exists some law for the error ξ1 with kξ1 kp,1 .
log n, such that for n large enough, there exists some least squares
estimator fn∗ over Fn satisfying
1
1
E kfn∗ − f0 kL2 (P ) ≥ ρ′ · n− 2 + 2p (log n)−2 .
Here ρ′ > 0 is a (small) constant independent of n.
The first claim of the above theorem justifies the heuristic conjecture
that the convergence rate for the LSE with heavy-tailed errors, should be
no better than the optimal rate in the Gaussian regression setting. Although
here we give an existence statement, the proof is constructive: in fact we
use (essentially) the Hölder class. Other function classes are also possible if
we can handle the Poisson (small-sample) domain of the empirical process
indexed by these classes. The second claim asserts that there is a ‘simple’
problem (=VC-subgraph class) for which the noise level dominates the risk
for the least squares estimator.
Remark 7. It is easy to derive from Proposition 2 that, considering the
size of kξ1 kp,1 , the rate for LSE in (2) of the above theorem is no worse
1/2
1
1
1
1
than kξ1 kp,1 n− 2 + 2p . n− 2 + 2p (log n)1/2 . Hence we miss at most a polylogarithmic factor in the lower
√ bound in (2) of Theorem 5. We suspect that
the moment condition p ≥ log n is a proof artifact. It would be interesting
to further investigate if this condition can be removed.
1
−1+
1
Theorem 5 implies that the rate OP (n− 2+α ∨ n 2 2p ) obtained in Theorem 4 cannot be improved in general. This translates the tradeoff in the
size of the multiplier empirical process between the complexity of the function class and the heaviness of the tail of the errors (multipliers) into the
convergence rate of the LSE (cf. Figure 2):
12
Q. HAN AND J. A. WELLNER
(1) If p ≥ 1 + 2/α, then fˆn − f0
L2 (P )
1
= OP n− 2+α . In this case, the
noise level is ‘small’ compared with the complexity of the function
class so that the LSE achieves the optimal rate as in the case for
i.i.d. Gaussian errors (see [70]).
1 1
− +
= OP n 2 2p . In this case,
(2) If p < 1 + 2/α, then fˆn − f0
L2 (P )
the noise is so heavy-tailed that the worst-case rate of convergence
of the LSE is governed by this noise barrier.
If p → ∞, i.e. the errors become increasingly light-tailed, then the optimal
rate in the Gaussian regression case (3.2) will eventually dominate. On the
other hand, if p = 2, i.e. there are only (essentially) second moments for the
errors, then (3.2) will be OP (n−1/4 ), which is the boundary rate for Donsker
and non-Donsker classes, cf. Figure 2.
4. Sparse linear regression: Lasso revisited
In this section we consider the sparse linear regression model:
(4.1)
Y = Xθ0 + ξ
where X ∈ Rn×d is a (random) design matrix and ξ = (ξ1 , . . . , ξn ) is a meanzero noise vector independent of X. When the true signal θ0 is sparse, one
popular estimator is the Lasso [60]:
1
2
kY − Xθk2 + λ kθk1 .
(4.2)
θ̂(λ) ≡ arg min
θ∈Rd n
The lasso estimator has been thoroughly studied in a vast amount of literature; we only refer the readers to the monograph [12] for a comprehensive
overview.
Our main interest here will be the following question: under what moment
conditions on the distributions of X and ξ can the lasso estimator enjoy
the optimal rate of convergence? In particular, neither X nor ξ need be
light tailed apriori, and ξ1 , . . . , ξn need not be independent from each other.
Previous works guaranteeing rate-optimality of the Lasso estimator typically
assume that both X and ξ are sub-gaussian, see [12, 51, 64], or at least X
is sub-gaussian, see [34]. The problem is challenging as we move out of
the sub-gaussian realm since, as we will see, the level of unboundedness of
the underlying empirical process depends on both the tail of X and ξ and
the parameters (d, n). This requires a different analysis compared with the
Donsker indexing function classes in Section 3.
Before stating the result, we need some notion of the compatibility condition: For any L > 0 and S ⊂ {1, . . . , d}, define
p
1
φ(L, S) = |S| min √ kXθS − XθS c k2 : kθS k1 = 1, kθS c k1 ≤ L .
n
Here for any θ = (θi ) ∈ Rd , θS ≡ (θi 1i∈S ) and θS c ≡ (θi 1i∈S
/ ). Let B0 (s) be
the set of s-sparse vectors in Rd , i.e. θ ∈ B0 (s) if and only if |{i : θi 6= 0}| ≤ s.
MULTIPLIER INEQUALITIES
13
Further let Σ = EΣ̂ where Σ̂ = X ⊤ X/n is the sample covariance matrix,
and σ d = σmin (Σ) and σ̄d = σmax (Σ) be the smallest and largest singular
value of the population covariance matrix, respectively. d = dn and s = sn
can either stay bounded or blow up to infinity in asymptotic statements.
Theorem 6. Let X be a design matrix with i.i.d. mean-zero rows, and
0 < lim inf σ d ≤ lim sup σ̄d < ∞. Suppose that
min φ(3, S) ≥ c0
(4.3)
|S|≤s
holds for some c0 > 0 with probability tending to 1 as n → ∞, and that for
some 1/4 ≤ α ≤ 1/2,
log d · M4 (X) ∨ log2 d
(4.4)
lim sup
< ∞,
n2−4α
n→∞
p
where M4 (X) ≡ E max1≤j≤d |X1j |4 . Then for θ̂ L ≡ θ̂(2L kξn k1/α,1 log d/n),
(4.5)
16L2 kξn k21/α,1 s log d
2
1
L
·
lim lim sup sup Pθ0
= 0.
X(θ̂ − θ0 ) >
L→∞ n→∞ θ0 ∈B0 (s)
n
n
2
c20
R∞
P
Here kξn k1/α,1 ≡ 0 Γξn (t)α dt, where Γξn (t) = n1 ni=1 P(|ξi | > t) is defined in Theorem 1.
p
The rate s log d/n in the above theorem is well-known to be (nearly)
minimax optimal for prediction in the sparse linear regression model. The
quantity kξn k1/α,1 should be thought as the ‘noise level’ of the regression
problem. For instance, if ξi ’s are i.i.d, and α = 1/2, then kξn k1/α,1 = kξ1 k2,1 .
Although in Theorem 6 we only consider prediction error, estimation error
L
θ̂ − θ0 can be obtained using completely similar arguments by noting
1
that Lemma 2 below also holds for estimation error.
As an illustration of the scope of Theorem 6, we consider some different
scaling regimes for the parameter space (d, n, s). For simplicity of discussion
we assume that the errors ξ1 , . . . , ξn have the same marginal distributions
and the design matrix X has i.i.d. entries such that X1,1 has a Lebesgue
density bounded away from ∞ and kX1,1 k2,P = 1.
Example 1. Consider the scaling regime d/n → λ ∈ (0, 1). We claim that
kX11 kP,4+ε ∨ kξk4,1 < ∞ for some ε > 0 guarantees the validity of (4.5).
First, (4.3) holds under the finite fourth moment condition, see [5]. Second,
(4.4) holds under the assumed moment conditions. Note that a fourth moment condition on X11 is necessary: if kX11 kP,4 = ∞, then lim sup σ̄d = ∞
a.s., see [4]. This appears to be a new result; [39] considered a different
‘tournament’ Lasso estimator with best tradeoff between confidence statement and convergence rate under heavy-tailed designs and errors.
Example 2. If the entries of X have sub-exponential tails, then Theorem
E of [33] showed that (4.3) holds with probability tending to 1 as n → ∞
14
Q. HAN AND J. A. WELLNER
under (s log d ∨ log3 d)/n → 0. Condition (4.4) is satisfied if kξk2+ε < ∞ up
to a mild dimension constraint d ≤ exp(ncε ) for some constant cε > 0. This
seems to be a new result; the recent result of [56] considered the similar tail
condition on X along with a sub-exponential tail for the errors ξi ’s, while
their rates are not optimal.
Example 3. If the entries of X have sub-gaussian tails, then (4.3) holds
with probability tending to 1 under the scaling s log d/n → 0 (see [52, 54]).
Condition (4.4) is satisfied if kξk2+ε < ∞, which recovers a recent result of
[34] in the case where X and ξ are independent, up to a mild dimension
′
constraint d ≤ exp(ncε ) for some constant c′ε > 0.
Now we prove Theorem 6. The following reduction is well-known.
Lemma 2 (Theorem 6.1 of [12]). On the event
(
)
r
n
2X
log d
EL ≡ max
ξi Xij ≤ L
,
(4.6)
1≤j≤d n
n
i=1
p
with tuning parameter λ ≡ 2L log d/n, it holds that
1
X(θ̂ L − θ0 )
n
2
2
≤
s0 log d
16L2
·
2
φ (3, S0 )
n
where S0 = {i : (θ0 )i 6= 0} and s0 = |S0 |.
The difficulty involved here is that both X and ξ can be heavy tailed.
By Theorem 1, to account for the effect of ξi ’s, we only need to track the
Pk
size of E max1≤j≤d
i=1 εi Xij at each scale k ≤ n. The empirical process
√
(normalized by k) is unbounded as d → ∞ and hence requires a different
analysis from that given in Section 3.
Lemma 3. Let X1 , . . . , Xn be i.i.d. random vectors in Rd with covariance
matrix Σ. If supd σmax (Σ) < ∞, then for all k, d ∈ N,
E max
1≤j≤d
k
X
i=1
1/4
+ (k log d)1/2 .
εi Xij . k log3 d · M4 (X) ∨ log2 d
The proof of the lemma is inspired by the recent work [17] who considered
Gaussian approximation of the maxima of high-dimensional random vectors
by exploiting second moment information for the Xi ’s. We modified their
method by taking into account the third moment information of Xi ’s induced
by the symmetric Rademacher εi ’s; such a modification proves useful in
identifying certain sharp moment conditions considered in the examples (in
particular Example 1). See Section 7.1 for a detailed proof.
Proof of Theorem 6. By Lemma 2 and the assumption on the compatibility condition (4.3), we see that with the choice for tuning parameter λ ≡
MULTIPLIER INEQUALITIES
2L kξn k1/α,1
15
p
log d/n, the left side of (4.5) can be bounded by
!
2 kξ k2
16L
2
n
s
log
d
1
1/α,1
+ o(1)
·
P θ0
X(θ̂ L − θ0 ) >
n
φ2 (3, S0 )
n
2
r
n
log d
2X
ξi Xij > L kξn k1/α,1
+ o(1).
≤ P max
1≤j≤d n
n
(4.7)
i=1
By Lemma 3, we can apply Theorem 1 with F1 = · · · = Fn ≡ {πj : Rd →
R, j = 1, . . . , d} where πj (x) = xj for any x = (xl )dl=1 ∈ Rd , and
p
1/4
ψn (k) ≡ C kα log3 d · M4 ∨ log2 d
+ k1/2 log d
for some 1/4 ≤ α ≤ 1/2 to conclude that
E max
1≤j≤d
n
X
ξi Xij . n
i=1
α
1/4
log d · M4 ∨ log d
3
2
+n
p
1/2
log d
Z
Z
∞
0
∞
0
Γξn (t)α dt
Γξn (t)1/2 dt
p
1/4
. nα log3 d · M4 ∨ log2 d
+ n1/2 log d kξn k1/α,1 .
By Markov’s inequality, (4.7) can be further bounded (up to constants) by
1/4
1 log d · (M4 ∨ log2 d)
+ o(1).
∨1
L
n2−4α
The claim of Theorem 6 therefore follows from the assumption (4.4).
5. Further remarks
5.1. Comparison of Theorem 1 with the multiplier inequality in
[68]. In this section we compare the multiplier inequality in Theorem 1
with the one in Section 2.9 of [68], which originates from [26, 27]: for i.i.d.
mean-zero ξi ’s and i.i.d. Xi ’s, and for any 1 ≤ n0 ≤ n,
n
(5.1)
1 X
ξi f (Xi )
E √
n
i=1
F
. (n0 − 1)E kf (X1 )kF
E max1≤i≤n |ξi |
√
n
k
1 X
εi f (Xi )
+ kξ1 k2,1 max E √
n0 ≤k≤n
k i=1
.
F
5.1.1. Asymptotic setting. The primary application of (5.1) rests in studying
asymptotic equicontinuity of the multiplier empirical process in the following
sense. Suppose that F is Donsker, then by theP
integrability of the empirical
n
3
−1/2
process (see Lemma 2.3.11 of [68]) , E n
i=1 εi f (Xi ) Fδ → 0 as n →
∞ followed by δ → 0. Now apply (5.1) via n → ∞, n0 → ∞ followed by
3Here F ≡ {f − g : f, g ∈ F, kf − gk
δ
L2 (P ) ≤ δ}.
16
Q. HAN AND J. A. WELLNER
P
δ → 0 we see that E n−1/2 ni=1 ξi f (Xi ) F → 0 as n → ∞ followed by
δ P
δ → 0 if kξ1 k2,1 < ∞. This shows that √1n ni=1 ξi f (Xi ) f ∈F satisfies a
CLT in ℓ∞ (F) if F is Donkser and the ξi ’s are i.i.d. with kξ1 k2,1 < ∞. The
moment condition is known to be unimprovable in general [35].
Our new multiplier inequality, Theorem 1, can also be used to study
asymptotic equicontinuity of the multiplier empirical process with the help
of the following lemma.
Lemma 4. Fix 0 ≤ γ ≤ 1 and a concave function ϕ : R≥0 → R≥0 such that
ϕ(x) → ∞ as x → ∞. Let {an } ⊂ R≥0 be such that an → 0 as n → ∞, and
ψ : R≥0 → R≥0 be the least concave majorant of {(n, an ϕ(n))}∞
n=0 . Then
ψ(t)/ϕ(t) → 0 as t → ∞.
P
Take any sequence δn → 0 and let an ≡ E n−1/2 ni=1 εi f (Xi ) F . By
δn
Lemma 4, the least concave majorant function ψ : R≥0 → R≥0 of the map
n 7→ an n1/2 (n ≥ 0) satisfies ψ(t)/t1/2 → 0 as t → ∞. Now an application of
Theorem 1 and the dominated convergence theorem shows that
Z ∞
n
ψ(nP(|ξ1 | > t)) p
1 X
p
· P(|ξ1 | > t) dt → 0
ξi f (Xi )
≤4
E √
n
nP(|ξ1 | > t)
0
i=1
as n → ∞.
Fδn
5.1.2. Non-asymptotic setting. The major drawback of (5.1) is that it is
not sharp in a non-asymptotic setting. For an illustration, let ξ1 , . . . , ξn be
i.i.d. multipliers with kξ1 kp,1 < ∞(p ≥ 3), Xi ’s be i.i.d. samples uniformly
distributed on [0, 1], and F be a uniformly bounded function class on [0, 1]
satisfying Assumption B with α ≡ 1. For example, F can be the class
of uniformly Lipschitz functions on [0, 1], or uniformly bounded monotonic
functions on [0, 1]. We apply (5.1) with F(n−1/3 ) (note that n−1/3 is the
usual local radius for 1-smooth problems) and local maximal inequalities for
the empirical process (Proposition 5) to see that
(5.2)
n
1 X
− 71 12 − p1
−1/6
−1/2+1/p
.
. inf (n0 − 1)n
+ n0
≍n
ξi f (Xi )
E √
1≤n0 ≤n
n
−1/3
i=1
F (n
)
On the other hand, Theorem 2 gives the sharp rate (recall p ≥ 3):
n
1 X
E √
ξi f (Xi )
n
i=1
. n−1/6 .
F (n−1/3 )
It is easily seen that the bound (5.2) calculated from (5.1) is far from optimal
even if p is very large.
5.2. Connections with Mendelson’s work.
MULTIPLIER INEQUALITIES
17
5.2.1. On the ‘L-subgaussian’ and ‘small-ball’ conditions. In a series of papers by Mendelson and his coauthors [32, 34, 38, 39, 45, 46], ERM procedures
from various perspectives are studied under ‘L-subgaussian’, ‘small-ball’ conditions and their variants.
Definition 1. A function class F
(1) is ‘L-subgaussian’ with respect to P if and only if for any f, g ∈
F, kf − gkψ2 (P ) ≤ L kf − gkL2 (P ) where k·kψ2 (P ) denotes the Orlicz
2
norm under P with ψ2 (x) = ex − 1.
(2) satisfies a small-ball condition with respect to P if and only if there
exists some constant κ > 0, ε ∈ (0, 1] such that for every f, g ∈
F ∪ {0},
P |f (X) − g(X)| ≥ κ kf − gkL2 (P ) ≥ ε.
These notions are especially well-suited for function classes with a certain moment equivalence property. In particular, F being sub-gaussian is
√
equivalent to kf kLp (P ) . p kf kL2 (P ) for all p ≥ 2. If there exists some
p > 2 such that the above inequality is true for all f ∈ F, then F satisfies
a small-ball condition4. A canonical example for such function classes is
given by F ≡ {fa (x) = x⊤ a : a ∈ Rp }, and P an isotropic and sub-gaussian
measure on Rp (see Definition 1.2 of [34]). On the other hand, we will see
below that such notions represent, in a certain sense, ‘best-case’ function
classes in statistical applications.
Proposition 3. The following function classes do not satisfy a small-ball
condition (and hence are not sub-gaussian) under the uniform law P on
[0, 1].
(1) The class of indicator functions on [0, 1]:
F = {1[a,b] : 0 ≤ a ≤ b ≤ 1}.
(2) The class of uniformly bounded monotonic/convex functions on [0, 1].
(3) The class of uniformly Lipschitz functions/C α functions with uniformly bounded derivatives on [0, 1].
These examples show that the best-case ‘small-ball’ or ‘sub-gaussian’ condition does not cover (i) VC-subgraph classes (ii) shape-restricted classes
(iii) smooth classes, and hence the theories developed in a series of work of
[34, 38, 39, 45, 46, 47] cannot be directly applied for these frequently used
function classes.
5.2.2. Connections with [49]. [49] studied the multiplier empirical process
(1.1) via a multiscale truncated analogue of Talagrand’s γ2 functional. We
briefly describe the machinery. For a random variable Z, and p ≥ 1, define
kZk(p) := sup1≤q≤p
kZkLq
√ .
q
This norm measures in a sense the ‘sub-gaussian’
4Hence the moment equivalence condition in [38, 39] is stronger than a small-ball
condition.
18
Q. HAN AND J. A. WELLNER
behavior of Z up to the p-th moment. The key quantity used in [49] is the
following: for a function class F, a given u ≥ 1 and s0 ≥ 0,
X
2s/2 kf − πs f k(u2 2s ) + 2s0 /2 sup kπs0 f k(u2 2s0 )
Λ̃s0 ,u (F) ≡ inf sup
f ∈F
f ∈F s≥s
0
where the infimum is taken over all admissible5 sequences {Fs }s≥0 and πs f
is the nearest point in Fs of f with respect to the (u2 2s ) norm. s0 is the
‘starting level’ of the chain. A standard generic chaining argument shows
that for any function class F, with probability at least 1−c0 exp(−u2 2s0 /c0 ),
sup |Gn f | ≤ c1 uΛ̃s0 ,u (F)
(5.3)
f ∈F
√
holds for some absolute constants c0 , c1 ≥ 1. Here Gn ≡ n(Pn − P ).
One main conclusion in [49] is to obtain a bound analogous to (5.3) for the
multiplier empirical process with i.i.d. multipliers ξi ’s admitting 2 + ε moments, cf. Theorem 1.9 in [49]. Such a result is easy to see for L-subgaussian
classes: if F is subgaussian, then Λ̃s0 ,u (F) . γ2 (F, ψ2 ) . γ2 (F, L2 ) 6. Hence
E supf ∈F |Gn f | . γ2 (F, L2 ). Applying the classical multiplier inequality
(5.1) with n0 = 1 we see that for i.i.d. mean-zero multipliers ξ1 , . . . , ξn
independent of samples X1 , . . . , Xn ,
n
1 X
ξi f (Xi ) . kξ1 k2,1 max E sup |Gk (f )|
E sup √
1≤k≤n f ∈F
n
f ∈F
i=1
. kξ1 k2,1 γ2 (F, L2 ) . kξ1 k2,1 E sup |Gf |
f ∈F
where {Gf }f ∈F is the centered Gaussian process indexed by F with covariance structure endowed by the L2 inner product. Here the last inequality
follows from Majorizing Measure Theorem [59]. This recovers Corollary
1.10 of [49] (with a slightly better moment condition on ξi ’s at the cost of a
weaker probability statement).
A basic question for Mendelson’s approach described above, is the sharpness of (5.3).
Proposition 4. Fix any ε > 0, and u ≥ 1, s0 ≥ 0. There exists an integer
n0 ∈ N such that for all n ≥ n0 , we can find a function class F such that
sup |Gn f | ≤ ε · c1 uΛ̃s0 ,u (F)
(5.4)
f ∈F
with probability at least 1 − (c0 + 1) exp(−u2 2s0 /c0 ). The constants c0 , c1 are
taken from (5.3) and its probability estimate.
Compared with (5.4), we see that (5.3) can be arbitrarily loose for some
function class if n is large enough. The major problem here is that the bound
(5.3) is independent of n and hence does not take into account of the Poisson
5A sequence {F } ⊂ F is admissible if |F | = 1 and |F | ≤ 22s .
s
0
s
6For definition of γ (F, ·), see [59].
2
MULTIPLIER INEQUALITIES
19
(small-sample) behavior of the empirical process. In fact, as n → ∞, the
size of supf ∈F |Gn f | →d supf ∈F |GP f |7 is governed by the Gaussian (largesample) behavior of the empirical process, i.e. the L2 size supf ∈F kf kL2 (P ) ;
while for n = 1 (or small), the quantity supf ∈F |G1 f | is governed by the
Poisson behavior of the empirical process, i.e. the L∞ size of the function
class. Such a discrepancy is actually the key to our construction in the proof
of Proposition 4. In fact any function class with this discrepancy can be
used for the construction. In this sense Proposition 4 shows some difficulties
of the approach of [49] in dealing with function classes whose element-wise
L2 size is much smaller than its L∞ size.
6. Proofs for the main results
6.1. Preliminaries. Let
J(δ, F, L2 ) ≡
(6.1)
Z
δ
sup
0
Q
q
1 + log N (ε kF kQ,2 , F, L2 (Q)) dε
denote the uniform entropy integral, where the supremum is taken over all
discrete probability measures, and
Z δq
(6.2)
1 + log N[ ] (ε, F, k·k) dε
J[ ] (δ, F, k·k) ≡
0
denote the bracketing entropy integral. The following local maximal inequalities for the empirical process play a key role throughout the proof.
Proposition 5. Suppose that F ⊂ L∞ (1), and X1 , . . . , Xn ’s are i.i.d. random variables with law P . Then with F(δ) ≡ {f ∈ F : P f 2 < δ2 },
(1) If the uniform entropy integral (6.1) converges, then
n
X
√
J(δ, F, L2 )
εi f (Xi )
. nJ(δ, F, L2 ) 1 + √ 2
kF kP,2 .
(6.3) E
nδ kF kP,2
i=1
F (δ)
(2) If the bracketing entropy integral (6.2) converges, then
n
X
J[ ] (δ, F, L2 (P ))
√
√ 2
.
. nJ[ ] (δ, F, L2 (P )) 1 +
εi f (Xi )
(6.4) E
nδ
i=1
F (δ)
Proof. (6.3) follows from [67]; see also Section 3 of [22]. (6.4) follows from
Lemma 3.4.2 of [68].
We will primarily work with F ≡ 1 in the above inequalities. A twosided estimate for the empirical process will be important for proving lower
bounds in Theorems 3 and 5. The following definition is taken from [22],
page 1167.
7G
P
is the canonical P -Brownian bridge process.
20
Q. HAN AND J. A. WELLNER
Definition 2. A function class F is α-full (0 < α < 2) if and only if there
exists some constant K1 , K2 > 1 such that both
log N ε kF kL2 (Pn ) , F, L2 (Pn ) ≤ K1 ε−α ,
a.s.
for all ε > 0, n ∈ N, and
log N σ kF kL2 (P ) /K2 , F, L2 (P ) ≥ K2−1 σ −α
hold. Here σ 2 ≡ supf ∈F P f 2 , F denotes the envelope function for F, and
Pn is the empirical measure for i.i.d. samples X1 , . . . , Xn with law P .
The following lemma, giving a sharp two-sized control for the empirical
process under the α-full assumption, is proved in Theorem 3.4 of [22].
Lemma 5. Suppose that F ⊂ L∞ (1) is α-full with σ 2 ≡ supf ∈F P f 2 . If
α/2
√
kF kL (P )
2
&α 1, then there exists some constant
nσ 2 &α 1 and nσ
σ
K > 0 depending only on α, K1 , K2 such that
n
X
kF kL2 (P ) α/2
−1 √
εi f (Xi )
≤E
K
nσ
σ
i=1
√
F
≤ K nσ
kF kL2 (P )
σ
α/2
.
Note that the right side of the inequality can also be derived from (6.3)
(taking supremum over all finitely discrete probability measures only serves
to get rid of the random entropy induced by L2 (Pn ) norm therein).
The following lemma guarantees the existence of a particular type of α-full
class that serves as the basis of the construction in the proof of Theorems 3
and 5.
Lemma 6. Let X , P be as in Theorem 3. Then for each α > 0, there
exists some function class F defined on X which is α-full and contains G ≡
{1[a,b]×[0,1]d−1 : 0 ≤ a ≤ b ≤ 1}.
Proof. Without loss of generality we assume P is uniform on X ≡ [0, 1]d .
1/α
Take F1 = C1 ([0, 1]) to be a 1/α-Hölder class on [0, 1] (see Section 2.7 of
[68]), and define F ≡ {f (x1 , . . . , xd ) := f1 (x1 ) : f1 ∈ F1 }. Let F̃ ≡ F ∪ G.
For any discrete probability measure Q on X = [0, 1]d ,
N (ε, F̃ , L2 (Q)) ≤ N (ε, F, L2 (Q)) + N (ε, G, L2 (Q))
≤ N (ε, F, L∞ ([0, 1]d )) + sup N (ε, G, L2 (Q)),
Q
where the last inequality follows from the fact that any ε-cover of F in L∞
metric on [0, 1]d induces an ε-cover on the function class F under any L2 (Q)
on X . Now by Theorem 2.7.1 of [68] and the fact that G is a bounded
VC-subgraph function class (see Section 2.6 of [68]), we have the following
entropy estimate:
(6.5)
sup log N (ε, F̃ , L2 (Q)) . ε−α .
Q
MULTIPLIER INEQUALITIES
21
where the supremum is taken over all discrete probability measures supported on X . On the other hand, note that
1/α
N (ε, F ∩ L2 (σ; P ), L2 (P )) = N (ε, C1
([0, 1]) ∩ L2 (σ), L2 ([0, 1])).
For ε ≡ cσ for some small c > 0, the latter is bounded from below by
a constant multiple of exp(c′ σ −α ) for another constant c′ > 0, due to the
classical work of [13, 61] in the context of more general Besov spaces. The
1/α
1/α
connection here is C1 ([0, 1]) = B∞,∞ (1) (in the usual notation for Besov
space, see Proposition 4.3.23 of [24]). See also [62], page 103-106 for an
explicit construction for a (local) minimax lower bound in L2 metric for the
Hölder class (which is essentially the same problem), where a set of testing
functions {fi : i ≤ M } is constructed such that M ≥ 2m/8 , kfj − fk kL2 &
m−1/α and kfj kL2 . m−1/α . Hence we see that
log N (cσ, F̃ ∩ L2 (σ; P ), L2 (P )) & σ −α .
(6.6)
The claim follows by combining (6.5) and (6.6).
6.2. Proof of Theorem 1.
Proof of Proposition 1. We will omit F in the proof. We first consider the
condition (A1). Note that for (X1′ , . . . , Xn′ ) being an independent copy of
(X1 , . . . , Xn ), we have
E
n
X
ξi f (Xi ) = Eξ,X
n
X
i=1
i=1
ξi f (Xi ) −
EX ′ f (Xi′ )
≤E
n
X
i=1
ξi f (Xi ) − f (Xi′ ) .
Here in the first equality we used the centeredness assumption on the function class F in (A1). Now conditional
on ξ, for fixed ε1 , . . . , εn , the map
Pn
′
(X1 , . . . , Xn , X1′ , . . . , Xn′ ) 7→
ξ
ε
is a permutation
i=1 i i f (Xi ) − f (Xi )
of the original map (without εi ’s). Since (X1 , . . . , Xn , X1′ , . . . , Xn′ ) is the
coordinate projection of a product measure, it follows by taking expectation
over ε1 , . . . , εn that
(6.7)
n
n
X
X
EX,X ′
ξi εi f (Xi ) − f (Xi′ ) .
ξi f (Xi ) − f (Xi′ ) = Eε,X,X ′
i=1
i=1
This entails that
(6.8)
n
n
n
X
X
X
|ξi | εi f (Xi )
|ξi | sgn(ξi )εi f (Xi ) = 2E
ξi f (Xi ) ≤ 2Eξ,ε,X
E
i=1
i=1
i=1
where the equality follows since the random vector (sgn(ξ1 )ε1 , . . . , sgn(ξn )εn )
has the same distribution as that of (ε1 , . . . , εn ) and is independent of
ξ1 , . . . , ξn . We will simply write |ξi | without the absolute value in the sequel
22
Q. HAN AND J. A. WELLNER
for notational convenience. Let π be a permutation over {1, . . . , n} such
that ξi = ξ(π(i)) . Then the right hand side of (6.8) equals
(6.9)
n
n
X
X
(by relabelling)
ξ(i) επ−1 (i) f (Xπ−1 (i) )
ξ(π(i)) εi f (Xi ) = E
E
i=1
i=1
n
X
=E
ξ(i) εi f (Xi )
i=1
P
(by invariance of (PX ⊗ Pε )n ).
Now write ξ(i) = k≥i (ξ(k) − ξ(k+1) ) where ξ(n+1) ≡ 0. The above display
can be rewritten as
(6.10)
n
n
k
n X
X
X
X
εi f (Xi ) .
(ξ(k) − ξ(k+1) )
(ξ(k) − ξ(k+1) )εi f (Xi ) = E
E
i=1 k=i
i=1
k=1
The claim under (A1) follows by combining (6.8)-(6.10). For (A2), let ξi′ ’s
be an independent copy of ξi ’s. Then the analogy of (6.8) becomes
E
n
X
ξi f (Xi ) = E
i=1
=E
n
n
X
X
(ξi − ξi′ )f (Xi )
(ξi − Eξi′ )f (Xi ) ≤ E
i=1
n
X
i=1
εi ξi − ξi′ f (Xi ) = E
i=1
n
X
i=1
εi |ηi | f (Xi )
where ηi ≡ ξi − ξi′ . The claim for (A2) follows by repeating the arguments
in (6.9) and (6.10).
Proof of Theorem 1. First consider (A1). Using Proposition 1 we see that,
n
n
k
X
X
X
.
ξi f (Xi )
≤ 2E ( ξ(k) − ξ(k+1) )E
εi f (Xi )
(6.11) E
i=1
i=1
k=1
Fn
Fn
By monotonicity of the function classes {Fk }, for any 1 ≤ k ≤ n,
(6.12)
E
k
X
εi f (Xi )
i=1
Fn
≤E
k
X
i=1
εi f (Xi )
Fk
≤ ψn (k).
Collecting (6.11)-(6.12), we see that
" n
#
n
n Z |ξ(k) |
X
X
X
≤ 2E
ξi f (Xi )
( ξ(k) − ξ(k+1) )ψn (k) = 2E
ψn (k) dt
E
i=1
k=1
k=1 |ξ(k+1) |
Fn
Z ∞
Z ∞
ψn nΓξn (t) dt
ψn (|{i : |ξi | ≥ t}|) dt ≤ 2
≤ 2E
0
0
where the last inequality follows from Fubini’s theorem and Jensen’s inequality, completing the proof for the upper bound for (A1). For (A2), mimicing
MULTIPLIER INEQUALITIES
the above proof, we have
Z ∞
n
X
≤
ψn
ξi f (Xi )
E
i=1
0
Fn
≤
Z
∞
=
Z
∞
ψn
0
ψn
0
n
X
P(
i=1
n
X
ξi − ξi′
!
≥ t)
23
dt
!
P(|ξi | ≥ t/2) + P( ξi′ ≥ t/2)
i=1
n
X
2
i=1
!
P(|ξi | ≥ t/2)
dt = 2
Z
∞
0
dt
ψn 2nΓξn (t) dt.
The proof of the claim for (A2) is complete by noting that ψn (2x) ≤ 2ψn (x)
due to the concavity of ψn and ψn (0) = 0.
6.3. Proof of Theorem 3. We first outline the proof strategy. The proof
contains two main steps:
(1) First we show that the size of the multiplier empirical process always
majorizes the corresponding empirical process (Lemma 7), whose
size from below can be obtained, if entropy control both from above
and below is possible. We construct some function class F̃ with this
property for each level γ > 2, yielding the lower bound n1/γ .
(2) The second step aims at a lower bound n1/p . The ideaPis to add to F̃
some ‘least favorable functions’ G such that E supf ∈G | ni=1 ξi f (Xi )| ≈
E max1≤i≤cn |ξi | & n1/p for some constant c > 0 independent of n.
In order to maintain the upper entropy estimate in the first step, G
needs to be chosen ‘small’ enough.
Lemma 7. Suppose that ξ1 , . . . , ξn are i.i.d. random variables independent
of i.i.d. X1 , . . . , Xn . Then
kξ1 k1 E
n
X
i=1
εi f (Xi )
F
≤ 2E
n
X
i=1
ξi f (Xi )
.
F
Proof. Let X1′ , . . . , Xn′ be independent copies of X1 , . . . , Xn . Then the left
hand side of the above display can be bounded by
(6.13)
n
n
X
X
′
εi |ξi | (f (Xi ) − f (Xi′ )) .
εi (f (Xi ) − f (Xi )) ≤ E
kξ1 k1 E
i=1
i=1
Using arguments similar to (6.8), the |ξi |’s in the above display can be
replaced by ξi ’s. The Rademacher εi ’s can be deleted using arguments in
(6.7), thereby completing the proof by further using triangle inequality.
Lemma 8. Let X1 , . . . , Xn be i.i.d. random variables distributed on [0, 1]d
with a probability law P admitting a Lebesgue density bounded away from
∞. Let {Ii }ni=1 be a partition of [0, 1]d such that Ii ∩ Ij = ∅ for i 6= j and
24
Q. HAN AND J. A. WELLNER
∪ni=1 Ii = [0, 1]d , and L−1 n−1 ≤ |Ii | ≤ Ln−1 for some absolute value L > 0.
Then there exists some τ ≡ τL,P ∈ (0, 1) such that for n large enough,
P (X1 , . . . , Xn lie in at most τ n intervals among {Ii }ni=1 ) ≤ 0.5n .
Proof. Note that the event in question equals
∪|I|=τ n {X1 , . . . , Xn ∈ ∪i∈I Ii } .
With K ≡ supx∈[0,1]d |(dP/dλ)(x)|, the above display implies that the probability in question can be bounded by
in
h e τ
n
n
· τ LK .
τ n · Ln−1 · K ≤ exp (τ n log(e/τ )) · (τ LK)n =
τ
τn
τ
Since τe · τ → 0 as τ → 0, it suffices to choose τ > 0 small enough.
Now we are in position to prove Theorem 3.
Proof of Theorem 3. As in the proof outline, the proof will proceed in two
steps. The first step aims at establishing a lower bound for the multiplier
empirical process on the order of n1/γ .
Let α = 2/(γ − 1), and F̃ be an α-full class on X in Lemma 6. Further
let δk = k−1/(2+α) and F̃k ≡ F̃ (δk ) = {f ∈ F̃ : P f 2 < δk2 }. Then it follows
from Lemma 5 that there exists some constant K > 0,
k
X
K −1 kα/(2+α) ≤ E
εi f (Xi )
i=1
F̃k
≤ Kkα/(2+α) .
P
Lemma 7 now guarantees that E k ni=1 ξi f (Xi )kF̃n can be bounded from
below by a constant multiple of nα/(2+α) = n1/γ where the constant depends
on kξ1 k1 . This completes the first step of the proof.
In the second step, we aim at establishing a lower bound of order n1/p . To
this end, let {Ij }nj=1 be a partition of X such that L−1 n−1 ≤ |Ij | ≤ Ln−1 .
On the other hand, let fj ≡ 1Ij ∈ F̃n for 1 ≤ j ≤ n (increase δn by constant
factors if necessary), and En denote the event that X1 , . . . , Xn lie in N ≥ τ n
sets among {Ij }nj=1 . Then Lemma 8 entails that P(En ) ≥ 1 − 0.5n ≥ 1/2 for
n large enough. Furthermore, let Ij ≡ {i : Xi ∈ Ij } and pick any Xι(j) ∈ Ij .
Note that Ij ’s are disjoint, and hence conditionally on X we have
E max |ξj | ≤ E max ξι(j)
1≤j≤τ n
(by i.i.d. assumption on ξi ’s)
1≤j≤N
≤ E max ξι(j) + E
1≤j≤N
≤ E max
1≤j≤N
X
i∈Ij
ξi
X
ξi
(Ij ’s are disjoint and Eξi = 0)
i∈Ij \ι(j)
(by Jensen’s inequality).
MULTIPLIER INEQUALITIES
25
Then
n
X
E
ξi f (Xi )
i=1
F̃n
≥ E max
1≤j≤n
n
X
ξi fj (Xi )
i=1
= EX Eξ max
≥ EX Eξ max |ξj | 1En ≥
1≤j≤τ n
1≤j≤N
1
Eξ max |ξj |
2 1≤j≤τ n
X
i∈Ij
ξi 1En
for n large enough. Now the second step follows from the assumption, and
hence completing the proof.
6.4. Proof of Theorem 4. We first prove Proposition 2.
Proof of Proposition 2. Let
Mn f =
n
n
i=1
i=1
1X
2X
(f − f0 )(Xi )ξi −
(f − f0 )2 (Xi ), M f = −P (f − f0 )2 .
n
n
Here we used the fact that Eξi = 0 and the independence assumption between {ξi } and {Xi }. Then it is easy to see that
n
2X
(f − f0 )(Xi )ξi + (Pn − P )(f − f0 )2 .
|Mn f − Mn f0 − (M f − M f0 )| ≤
n
i=1
The first claim follows by standard symmetrization and contraction principle
for the empirical process indexed by a uniformly bounded function class,
followed by an application of Theorem 3.2.5 of [68].
Now assume that ξ1 , . . . , ξn are i.i.d. mean-zero errors with kξ1 kp < ∞
for some p ≥ 2. Fix t ≥ 1. For j ∈ N, let Fj ≡ {f ∈ F : 2j−1 tδn <
kf − f0 kL2 (P ) ≤ 2j tδn }. Then by a standard peeling argument, we have
!
X
≥ tδn ≤
P fˆn − f0
P sup (Mn (f ) − Mn (f0 )) ≥ 0 .
L2 (P )
j≥1
f ∈Fj
Each probability term in the above display can be further bounded by
!
P
sup (Mn (f ) − Mn (f0 ) − (M f − M f0 )) ≥ 22j−2 t2 δn2
f ∈Fj
≤ P
n
X
√
1
√
ξi f (Xi ) ≥ 22j−4 t2 nδn2
n
sup
f ∈F −f0 :kf kL
2 (P )
+ P
≤2j tδn
i=1
sup
f ∈F −f0 :kf kL
2 (P )
≤2j tδn
n
X
√
1
√
εi f 2 (Xi ) ≥ 22j−3 t2 nδn2 .
n
i=1
26
Q. HAN AND J. A. WELLNER
By contraction principle and moment inequality for the empirical process
(Lemma 1), we have
2
2
E
sup
f ∈F −f0 :
kf kL (P ) ≤2j tδn
2
n
1 X
√
ξi f (Xi )
∨ E
n
sup
f ∈F −f0 :
kf kL (P ) ≤2j tδn
i=1
2
n
1 X
√
εi f 2 (Xi )
n
i=1
2
. φn (2j tδn ) + (1 ∨ kξ1 k2 )2 22j t2 δn2 + (1 ∨ kξ1 kp )2 n−1+2/p .
In the above calculation we used the fact that E max1≤i≤n |ξi |2 ≤ kξ1 k2p n2/p
under kξ1 kp < ∞. By Chebyshev’s inequality,
"
#
X φn (2j tδn ) 2
1
1
ˆ
√
∨ 2j 2 2 ∨ 4j 4 2−2/p 4 .
≥ tδn ≤ Cξ
P fn − f0
2 t nδn 2 t n
22j t2 nδn2
L2 (P )
δn
j≥1
−1+
1
Under the assumption that δn ≥ n 2 2p , and noting that φn (2j tδn ) ≤
2j tφn (δn ) by the assumption that δ 7→ φn (δ)/δ is non-increasing, the right
side of the above display can be further bounded up to a constant by
X φn (δn ) 2
1
1
√ 2
+ 2 . 2
j
2 t nδn
t
t
j≥1
for t ≥ 1. The expectation bound follows by integrating the tail estimate.
The following lemma calculates an upper bound for the multiplier empirical process at the target rate in Theorem 4.
Lemma 9. Suppose that Assumption A holds with i.i.d. X1 , . . . , Xn ’s with
law P , and F ⊂ L∞ (1) satisfies Assumption B with α ∈ (0, 2). Further
assume for simplicity that ξi ’s have the same marginal distributions with
1
−1+ 1
kξ1 kp,1 < ∞. Then with δn ≡ n− 2+α ∨ n 2 2p , we have
E
n
X
sup
2
P f 2 ≤ρ2 δn
i=1
1−α/2
≤ K̄α (ρ
ξi f (Xi ) ∨ E
∨ρ
−α
sup
n
X
2
P f 2 ≤ρ2 δn
i=1
εi f (Xi )
( α
n 2+α 1 ∨ kξ1 k1+2/α,1 ,
)
1
n p 1 ∨ kξ1 kp,1 ,
p ≥ 1 + 2/α,
1 ≤ p < 1 + 2/α.
1
Proof. If p ≥ 1 + 2/α, then δn = n− 2+α . By local maximal inequalities for
empirical processes (see Proposition 5), we have
!
k
X
√
(ρδk )1−α/2
1−α/2
1∨ √
εi f (Xi ) ≤ C k(ρδk )
E sup
k(ρδk )2
P f 2 ≤ρ2 δk2 i=1
(6.14)
α
≤ Ck 2+α ρ1−α/2 (1 ∨ ρ−(1+α/2) )
α
≤ C(ρ1−α/2 ∨ ρ−α ) · k 2+α .
MULTIPLIER INEQUALITIES
27
Applying Theorem 2 we see that
E
sup
n
X
2
P f 2 ≤ρ2 δn
i=1
α
ξi f (Xi ) ≤ C(ρ1−α/2 ∨ ρ−α ) · n 2+α 1 ∨ kξ1 k1+2/α,1 .
1
1
If p < 1 + 2/α, then δn = n− 2 + 2p . In this case,
(6.15)
!
k
1−α/2
X
√
(ρδ
)
k
E sup
εi f (Xi ) ≤ C k(ρδk )1−α/2 1 ∨ √
2
2
2
2
k(ρδ
P f ≤ρ δk i=1
k)
1 1
1 α p−1
1
+α
· p−1
1−α/2
−(1+α/2) 2 − p + 2 · p
2 p
2
p
≤ Cρ
·k
1+ρ
k
1
≤ C(ρ1−α/2 ∨ ρ−α )k p
where the last inequality follows from p1 > α2 · p−1
p by the assumed relationship
between p and α. Now apply Theorem 2 we have
sup
E
n
X
2
P f 2 ≤ρ2 δn
i=1
as desired.
1
ξi f (Xi ) ≤ C(ρ1−α/2 ∨ ρ−α ) · n p (kξ1 kp,1 ∨ 1)
Proof of Theorem 4. The claim follows immediately from Lemma 9 by noting that the rate δn chosen therein corresponds to the condition (3.3) in
Proposition 2, along with Proposition 5 handling (3.4).
6.5. Proof of Theorem 5.
6.5.1. The strategy. The proof of Theorem 5 is technically rather involved;
here we give a brief outline. There are mainly two steps:
(1) We first show that (see Lemma 10): the risk of some LSE corresponds to the extreme value δn∗ of the map
(6.16) δ 7→ Fn (δ) ≡
sup
f ∈F −f0
:P f 2 ≤δ2
(Pn − P )(2ξf − f 2 ) − δ2 ≡ En (δ) − δ2 .
This step is similar in spirit to [16, 66] (e.g. Theorem 1.1 of [16];
Lemma 3.1 of [66]); we will deal with the fact that the LSE may not
be unique, and the map does not enjoy good geometric properties
such as convexity in [16], or uniqueness of the extreme value of the
map (6.16) as in [66].
(2) Step 1 reduces the problem of finding δn∗ to that of finding δ1 < δ2
such that En (δ1 ) < Fn (δ2 ) [This means, Fn (δ) < Fn (δ2 ) for δ ≤ δ1 ,
implying that the extreme value δn∗ ≥ δ1 ]. Hence our task will be
to find δ1 , δ2 with matching order such that En (δ1 ) is smaller than
Fn (δ2 ) up to a constant order under a specific function class. The
construction of such an underlying regression function is inspired
by the one used in Theorem 3. The main technical job involves (i)
developing a problem-specific approach to derive an upper bound
28
Q. HAN AND J. A. WELLNER
for En (ρδ1 ) for small ρ > 0 (corresponding to the Poisson (smallsample) domain of the empirical process where general tools fail),
(ii) using a Paley-Zygmund moment argument to produce a sharp
lower bound for Fn (δ2 ) and (iii) handling the delicate fact that the
Lp,1 norm is slightly stronger than the Lp norm.
6.5.2. The reduction scheme.
Lemma 10. Fix ε > 0. Let δn∗ ≡ inf{δ∗ ≥ 0 : Fn (δ∗ ) ≥ supδ∈[0,∞) Fn (δ)−ε}.
Then there exists a 2ε-approximate LSE fn∗ such that
kfn∗ − f0 k2L2 (P ) ≥ (δn∗ )2 − ε.
Proof. Without loss of generality we assume f0 = 0. Let fn∗ be such that
δ02 ≡ P (fn∗ )2 ≤ (δn∗ )2 and En (δn∗ ) ≤ (Pn − P )(2ξfn∗ − (fn∗ )2 ) + ε. Note that
for any f ∈ F,
(Pn − P )(2ξf − f 2 ) − P f 2 ≤ Fn (kf kL2 (P ) ) ≤ Fn (δn∗ ) + ε
= En (δn∗ ) − (δn∗ )2 + ε
≤ (Pn − P )(2ξfn∗ − (fn∗ )2 ) − P (fn∗ )2 + 2ε,
where in the last inequality we used the definition of fn∗ and the fact that
(δn∗ )2 ≥ P (fn∗ )2 . This implies that for any f ∈ F,
kY − f k2n = Pn ξ 2 − Pn (2ξf − f 2 )
≥ Pn ξ 2 − Pn (2ξfn∗ − (fn∗ )2 ) − 2ε = kY − fn∗ k2n − 2ε.
Hence fn∗ is a 2ε-approximate LSE. The claim follows if we can show δ02 ≥
(δn∗ )2 − ε. This is valid: if (δn∗ )2 > δ02 + ε, then
Fn (δ0 ) ≥ (Pn − P )(2ξfn∗ − (fn∗ )2 ) − δ02
≥ En (δn∗ ) − ε − δ02 = Fn (δn∗ ) − ε + ((δn∗ )2 − δ02 ) > Fn (δn∗ ),
a contradiction to the definition of δn∗ by noting that δ0 ≤ δn∗ .
Since ε > 0 in Lemma 10 can be arbitrarily small, in the following analysis
we will assume without loss of generality that ε = 0.
The following simple observation summarizes the strategy for finding a
lower bound on the rate of some least squares estimator.
Proposition 6. Suppose that 0 < δ1 < δ2 are such that En (δ1 ) < Fn (δ2 ).
Then there exists a LSE fn∗ such that kfn∗ − f0 kL2 (P ) ≥ δ1 .
Proof. The condition implies that Fn (δ) = En (δ) − δ2 ≤ En (δ) ≤ En (δ1 ) <
Fn (δ2 ) for any δ ≤ δ1 and hence a maximizer δn∗ of the map δ 7→ Fn (δ)
cannot lie in [0, δ1 ], i.e. δn∗ ≥ δ1 . The claim now follows by Lemma 10.
In the next few subsections, we first prove claim (1) of Theorem 5. The
proof of claim (2) follows the similar proof strategy as that of claim (1); the
details will be delayed until the last subsection.
MULTIPLIER INEQUALITIES
29
6.5.3. Upper bound. The regression function class F we consider will be the
Hölder class constructed in Lemma 6 with X = [0, 1]d , and f0 ≡ 0. We
first handle the upper bound part of the problem. Lemma 9 is awkward in
this regard because general tools (Proposition 5) cannot handle the Poisson
(small-sample) domain of the empirical process, and hence the resulting
bound is insensitive with respect to small ρ > 0. The following lemma
remedies this for our special function class F.
Lemma 11. Suppose that X = [0, 1]d and P is a probability distribution on
X with Lebesgue density bounded away from 0 and ∞. Suppose ξi ’s are i.i.d.
mean-zero and kξ1 kp,1 < ∞. Suppose p ≥ 1 + 2/α. Then for any ρ ∈ (0, 1),
if
n ≥ min{n ≥ 3 : ρ2 ≥ log n(n−α/(2+α) )},
(6.17)
1
then with δn ≡ n− 2+α we have,
E
sup
n
X
2
f ∈F :P f 2 ≤ρ2 δn
i=1
2ξi f (Xi ) − f 2 (Xi ) + P f 2
α
≤ K̄P,α ρ1−α/2 n 2+α (1 ∨ kξ1 kp,1 ).
Note that in the above lemma we may choose ρ small as long as the sample
size condition (6.17) is satisfied. The key idea of the proof is to compare
supf ∈F :P f 2 <σ2 Pn f 2 with σ 2 directly for (nearly) the whole range of σ 2
including the Poisson (small-sample) domain by exploiting the geometry of
F. To this end we need the following.
Lemma 12. Let X1 , . . . , Xn be i.i.d. P on [0, 1] where P has Lebesgue
density bounded away from 0 and ∞. Set γn = κP log n/n where κP ≥ 1
is a constant large enough depending only on P . Let Ij ≡ [(j − 1)γn , jγn )
for j = 1, . . . , n/(κP log n) ≡ N . Then for some cP > 0, nP large enough
depending on P , if n ≥ nP , with probability at least 1 − 2n−2 , all intervals
{Ij } contain at least one and at most cP log n samples.
Proof. Without loss of generality we assume that P is uniform on [0, 1].
The general case where P has Lebesgue density bounded away from 0 and
∞ follows from minor modification. Let E1 (E2 ) be the event that all intervals
{Ij } contain at least one sample(at most c log n samples). Then for κP = 6,
P (E1c ) = P (∪1≤j≤N {Ij contains no samples})
κP log n n
≤ N e−κP log n ≤ n−5 .
≤N · 1−
n
On the other hand,
!
n
X
c
P(E2 ) = P max
1Ij (Xi ) > c log n
1≤j≤N
≤
N
X
j=1
P
i=1
n
X
i=1
1Ij (Xi ) − γn
!
> (c − 6) log n .
30
Q. HAN AND J. A. WELLNER
Now P
we use Bernstein inequality
following form (cf. (2.10) of [9]): for
P in the
2 where |Z | ≤ b for all 1 ≤ i ≤ n, we
S = ni=1 (Zi − EZi ), v = ni=1 EZ
i
i
2
t
. We apply this with Zi ≡ 1Ij (Xi ) and
have P(S > t) ≤ exp − 2(v+bt/3)
Pn
hence γn = EZi and v = i=1 γn = 6 log n, b = 1, to see that right side of
the above display can be further bounded by
N
X
(c − 6)2 log2 n
exp −
≤ N e−3 log n ≤ n−2
2(6 log n + (c − 6) log n/3)
j=1
by choosing c = 13. Combining two cases we have completed the proof.
We also need Dudley’s entropy integral bound, recorded below for the
convenience of the reader.
Lemma 13 ([21]). Let (Xt )t∈T be a centered separable Gaussian process,
and kt − sk2 := E(Xt − Xs )2 . Then
Z diam(T ) q
log N (ε, T, k·k) dε.
E sup Xt ≤ C
t∈T
0
Here C is a universal constant.
Proof of Lemma 11. By the contraction principle, we only need to handle
E
sup
n
X
2
P f 2 ≤ρ2 δn
i=1
ξi f (Xi ) ,
E
sup
n
X
2
P f 2 ≤ρ2 δn
i=1
εi f (Xi ) .
Let F be the Hölder class constructed in Lemma 6. We first claim that on
an event En with probability at least 1 − 2n−2 , for any f ∈ F,
log n
2
2
.
(6.18)
Pn f ≤ cP P f +
n
To see this, we only need to consider the first dimension by definition of F.
By Lemma 12, we see that on an event En with probability at least 1 − 2n−2 ,
n
N
N
1X 2
1X X 2
1X
f (Xi ) =
f (Xi ) ≤
cP log n · max f 2 (Xi ).
Xi ∈Ij
n
n
n
i=1
j=1 Xi ∈Ij
j=1
Here N = n/(κP log n) is the number of intervals {Ij }. The trick now is
to observe that since f is at least 1/2-Hölder, we have maxXi ∈Ij f (Xi ) ≤
√
minx∈Ij f (x) + γn , where γn = κP log n/n is the length for each interval
Ij . Hence on the same event as above, the right side of the above display
can be further bounded by
N
N
2cP log n
2cP X
2cP log n X
2
γn min f 2 (x) +
min f (x) + γn =
x∈Ij
x∈Ij
n
κP
n
j=1
j=1
Z
2cP log n
2cP 1 2
f (x) dx +
,
≤
κP 0
n
MULTIPLIER INEQUALITIES
31
where the inequality follows from the definition of Riemann integral. The
claim (6.18) is thus proven by noting that the intergal in the above display
is equivalent to P f 2 up to a constant depending on P only. Now using
Dudley’s entropy integral (see Lemma 13) and (6.18), we have for the choice
p
1
σn = ρ(n− 2+α ) ≥ log n/n [the inequality holds when n ≥ min{n ≥ 3 :
ρ2 ≥ log n(n−α/(2+α) )}],
n
1 X
√
εi f (Xi )
n
sup
E
2
f ∈F :P f 2 ≤σn
i=1
Z 2√supf ∈F Pn f 2 p
≤ CE
0
Z √
2
2 +log n/n)
cP (σn
.
0
Since
√
ρ2
2
n 2+α
ε−α/2 dε + J(∞, F, L2 )P(Enc ) .P,α σn1−α/2 + n−2 .
α
1 1 2−α
√
= ρ1−α/2 n 2 − 2 2+α = ρ1−α/2 n 2+α and n · n−2 ≤ n−1 ≤
α
≤ ρ2 ≤ ρ1−α/2 n 2+α , in this case Theorem 2 along with the
1−α/2
nσn
(log n)2
log N (ε, F, L2 (Pn )) dε
assumption p ≥ 1 + 2/α yields that
sup
E
n
X
2
f ∈F :P f 2 ≤σn
i=1
α
α
ξi f (Xi ) .P,α ρ1−α/2 n 2+α kξ1 k1+2/α,1 ≤ ρ1−α/2 n 2+α kξ1 kp,1 .
The proof is complete.
6.5.4. Lower bound. Next we turn to the lower bound part of the problem.
We will first consider a lower bound in expectation, then a Paley-Zygmund
type argument translates the claim from in expectation to in probability.
Lemma 14. Let X = [0, 1]d , and P be a probability measure on X with
Lebesgue density bounded away from 0 and ∞. Let ξ1 , . . . , ξn be i.i.d. meanzero random variables such that kξ1 k1 > 0, and F be the Hölder class con1
structed in Lemma 6. Then with δn ≡ n− 2+α , if kξ1 k1 > Kα,P and ϑ ≥ 1
for some Kα,P > 0 depending only on α, P ,
E
sup
n
X
2
P f 2 ≤ϑ2 δn
i=1
2ξi f (Xi ) − f 2 (Xi ) + P f 2
α
≥ K P,α kξ1 k1 ϑ1−α/2 · n 2+α .
Proof. By Lemmas 5 and 7, and the α-fullness of F̃ , we have
E
sup
n
X
2
P f 2 ≤ϑ2 δn
i=1
ξi f (Xi ) ≥
n
X
α
1
εi f (Xi ) ≥ C1 kξ1 k1 ϑ1−α/2 n 2+α .
kξ1 k1 E sup
2
2
P f 2 ≤ϑ2 δn
i=1
32
Q. HAN AND J. A. WELLNER
On the other hand, during the proof of Lemma 9 (see (6.14)) we see that
α
Pk
2
1−α/2 ∨1)·n 2+α . By de-symmetrization,
E supP f 2 ≤ϑ2 δ2
i=1 εi f (Xi ) ≤ C2 (ϑ
k
sup
E
k
X
P f 2 ≤ϑ2 δk2 i=1
α
(f 2 (Xi ) − P f 2 ) ≤ 2C2 (ϑ1−α/2 ∨ 1) · n 2+α .
Here C1 , C2 are constants depending on α, P only. Now for kξ1 k1 ≥ 2C2 /C1 ,
since ϑ ≥ 1, by the triangle inequality we see that
E
sup
n
X
2
P f 2 ≤ϑ2 δn
i=1
as desired.
2ξi f (Xi ) − f 2 (Xi ) + P f 2
α
≥ C1 kξ1 k1 ϑ1−α/2 n 2+α ,
The following Paley-Zygmund lower bound is standard.
Lemma 15 (Paley-Zygmund). Let Z be any non-negative random variable.
q′
(1−ε)EZ
, where q, q ′ ∈ (1, ∞) are
Then for any ε > 0, P(Z > εEZ) ≥ (EZ
q )1/q
conjugate indices: 1/q + 1/q ′ = 1.
Now we turn the lower bound in expectation in Lemma 14 to a probability
bound by a Paley-Zygmund argument.
Lemma 16. Consider
as in Lemma 14 with p ≥ 2, and let
Pn the same setup
2 (X ) + P f 2 . Suppose p ≥ 1 + 2/α. If
Z = supP f 2 ≤ϑ2 δn2
2ξ
f
(X
)
−
f
i
i
i
i=1
kξ1 kp < ∞, kξ1 k1 > Kα , ϑ ≥ 1 and 1 < q ≤ p. Then
α
1
−1/(q−1)
1−α/2
2+α
·n
≥ 2−q/(q−1) L̄α,ξ,ϑ,q,P > 0.
P Z ≥ K P,α kξ1 k1 ϑ
2
The constant in the probability estimate is defined in (6.19) below.
α
Proof. Lemma 14 entails that EZ ≥ K P,α kξ1 k1 ϑ1−α/2 · n 2+α . By the moment inequality Lemma 1, if ξi ’s have finite p-th moments, and q ≤ p,
(6.19)
√
EZ q
1 ∨ E max1≤i≤n |ξi |q
( n(kξ1 k2 ∨ 1)ϑδn )q
≤ Cq 1 +
+
(EZ)q
(EZ)q
(EZ)q
−q
≤ Cq 1 + 2 K P,α kξ1 k1 ∧ 1
q
− qα
− qα
× ϑαq/2 (kξ1 k2 ∨ 1)q n 2(2+α) ∨ ϑq(α/2−1) kξ1 kqp n p 2+α
i
h
−q αq/2
ϑ
(kξ1 kp ∨ 1)q ≡ L̄α,ξ,ϑ,q,P .
≤ Cq 1 + 2 K P,α kξ1 k1 ∧ 1
In the second inequality we used kmaxi |ξi |kq ≤ kmaxi |ξi |kp ≤ n1/p kξ1 kp ,
and the third inequality follows by noting ϑ ≥ 1 and the assumption p ≥
1 + 2/α. The proof is complete.
MULTIPLIER INEQUALITIES
33
6.5.5. Putting the pieces together.
Proposition 7. Suppose kξ1 kp,1 < ∞ for p ≥ max{2, 1 + 2/α}. If kξ1 k1 >
Kα,P for some (large) constant Kα,P > 0, then for n large enough, there
exist constants ρξ,α,P < ϑξ,α,P such that on an event with positive probability
p1 = p1 (α, ξ, P ) > 0 independent of n,
1
1
Fn ϑξ,α,P · n− 2+α > En ρξ,α,P · n− 2+α .
Proof. Lemma 11 and Markov’s inequality entail that for any ρ > 0, if
n ≥ min{n ≥ 3 : ρ2 ≥ log n(n−α/(2+α) )}, then on an event with probability
at least 1 − 1/M ,
(6.20)
α
nEn (ρδn ) ≤ M C̄ξ,α,P · ρ1−α/2 n 2+α .
We will choose ρ, M later on. On the other hand, apply Lemma 16 with
2
q = 2 (since p ≥ 2) and ϑ = (K P,α kξ1 k1 /4) 2+α (we may increase kξ1 k1 to
ensure ϑ ≥ 1 if necessary) we see that on an event with probability at least
2p1 ≡ 2p1 (α, ξ, P ), we have
α
α
1
nFn (ϑδn ) ≥ K P,α kξ1 k1 ϑ1−α/2 · n 2+α − ϑ2 n 2+α
2
(6.21)
α
α
1
≥ K P,α kξ1 k1 ϑ1−α/2 · n 2+α ≡ C ξ,α,P · n 2+α .
4
First we choose M = 1/p1 so that with probability at least p1 , (6.20) and
2
(6.21) hold simultaneously. Then we choose ρ = min{(p1 C ξ,α,P /2C̄ξ,α,P ) 2−α , ϑ/2}
to conclude Fn (ϑδn ) > En (ρδn ) with probability at least p1 .
Now we have completed the program outlined in Proposition 6.
Proof of Theorem 5: claim (1). Recall that the regression function class is
taken from Lemma 6 with f0 ≡ 0. Combining the proof outline Proposition
6, with Proposition 7, we see that there exists an event with probability at
least p1 = p1 (α, ξ, P ) > 0, on which at least one least squares estimator fn∗
1
over F satisfying kfn∗ − f0 kL2 (P ) ≥ ρ · n− 2+α , where ρ > 0 is a (small) constant independent of n. The claim now follows by bounding the expectation
from below on this event.
6.5.6. Proof of the remaining claim. In this subsection we prove the second
claim of Theorem 5. Without loss of generality, we only consider d = 1, and
the probability measure P is assumed to be uniform for simplicity. To this
end, let F̃ ≡ F̃n ≡ {1[a,b] : a, b ∈ [0, 1] ∩ Q, b − a ≥ δn2 } ∪ {0} and f0 = 0,
′
where δn = ρn−1/2+1/2p with 1p − p1′ = ε for some numeric constants ε, ρ > 0
to be specified later.
In the current case En (δ) = 0 for δ < δn and hence our goal is to give a
lower bound for Fn (δn ). We mimic the proof strategy of the first claim of
Theorem 5 by (i) giving a lower bound for the multiplier empirical process
in expectation, and then (ii) using the Paley-Zygmund moment argument
34
Q. HAN AND J. A. WELLNER
to translate the lower bound in probability. The arguments are somewhat
delicate due to the fact that the Lp,1 norm is stronger than the Lp norm.
P
For notational simplicity, let Z ≡ supf ∈F̃n :P f 2 ≤δ2 | ni=1 ξi f (Xi )|, and Z̃ ≡
n
Pn
2
2
supf ∈F̃n :P f 2 ≤δ2
i=1 2ξi f (Xi ) − f (Xi ) + P f .
n
Lemma 17. Suppose ρ ≤ (64/e3 )1/6 and ξ1 , . . . , ξn are i.i.d. mean-zero
random variables. Then for n ≥ 2, we have EZ ≥ 21 E max1≤j≤4−1 n1−1/p′ |ξj | .
Proof. Let Ij ≡ [(j − 1)δn2 , jδn2 ] ⊂ [0, 1] for j = 1, . . . , N where N = δn−2 ≤
ρ−2 n. Note that for any c ∈ (0, 1),
P X1 , . . . , Xn lie in at most cN intervals among {Ij }N
j=1
= P ∪|I|=cN {X1 , . . . , Xn ∈ ∪i∈I Ii }
−2
N n
≤
c ≤ ecN log(e/c)−n log(1/c) ≤ e(cρ log(e/c)−log(1/c))n .
cN
By choosing c = ρ2 /4, the exponent in the above display can be further
bounded by 41 log(e/c) − log(1/c) = 41 log(c3 e) = 14 log(eρ6 /64) ≤ − 12 where
the last inequality follows by the assumption that ρ ≤ (64/e3 )1/6 . Hence we
conclude that on an event E with probability at least 1 − 0.61n , the samples
X1 , . . . , Xn must occupy at least ρ2 N/4 many intervals among {Ij }N
j=1 . This
implies that
EZ = E
sup
n
X
2
f ∈F̃n :P f 2 ≤δn
i=1
ξi f (Xi ) ≥ E max
1≤j≤N
X
i∈Ij
ξi 1Ij (Xi ) 1E ≥
1
E max |ξj |
2 1≤j≤ρ2 N/4
where we used the same arguments as in the proof of Theorem 3. The claim
′
now follows by noting ρ2 N/4 = ρ2 δn−2 /4 = n1−1/p /4.
Lemma 18. Suppose that ξ1 , . . . , ξn are i.i.d. symmetric random variables
′
with P(|ξ1 | > t) = 1/(1 + |t|p ). Further suppose that ρ ≤ (64/e3 )1/6 ,
−1/2 ∨ 3 ≤ p ≤ log n/ log log n and n ≥ min{n ≥ 2 : δ = ρn−1/2+1/2p′ ≥
εp
n
log n/n}. Then there exists some absolute constant C1 > 0, and for any
C2 > 0, there exists some constant C3 = C3 (C2 ) > 0 such that
p
2
1 1/p′ −1/(p′ )2
1/2p′
P Z̃ ≥ n
− C3 n
log n ≥ C1 (εp′ ) ∧ 1 n−4ε − e−C2 log n .
16
We need some auxiliary results before the proof of Lemma 18.
Lemma 19. Suppose that ξ1 , . . . , ξn are i.i.d. symmetric random variables
with p(t) ≡ P(|ξ1 | > t) = 1/(1 + |t|p ). Then for any 1 ≤ q < p, and n ≥ 2,
1/q
p + q 1/q 1/p
1 1/p
q
n .
n ≤ E max |ξi | ≤ E max |ξi |
≤
1≤i≤n
1≤i≤n
4
p−q
We need the following exact characterization concerning the size of maxima of a sequence of independent random variables due to [25], see also
Corollary 1.4.2 of [19].
MULTIPLIER INEQUALITIES
35
Lemma 20. Let ξ1 , . . . , ξn be a sequence of independent non-negative random variables suchP
that kξi kr < ∞ for all 1 ≤ i ≤ n. For λ > 0, set
δ0 (λ) ≡ inf {t > 0 : ni=1 P(ξi > t) ≤ λ}. Then
n
n
i=1
i=1
1 X r
1 X r
Eξi 1ξi >δ0 ≤ E max ξir ≤
Eξi 1ξi >δ0 .
1≤i≤n
1+λ
1∧λ
Proof of Lemma 19. For λ ≡ 1 in Lemma 20, δ0 = inf{t > 0 : np(t) ≤ 1} =
(n − 1)1/p . Lemma 20 now yields that for q < p,
E max |ξi |q ≤ nE |ξ1 |q 1|ξ1 |>δ0
1≤i≤n
Z
= n P(|ξ1 | > δ0 )
nδ0q
+ qn
1 + δ0p
0
∞
δ0
q−1
qu
du +
Z
∞
δ0
q−1
qu
P(|ξ1 | > u) du
1
du
p−q+1
u
δ0
q
n
p + q q/p
= (n − 1)q/p +
(n − 1)q/p ≤
n
p−qn−1
p−q
≤
Z
since n ≥ 2. For a lower bound for E max1≤i≤n |ξi |, we proceed similarly as
above by using 1 + up ≤ 2up on [δ0 , ∞) for n ≥ 2:
Z ∞
1
δ0
(n − 1)1/p
1
n
+
du
=
≥ n1/p .
E max |ξi | ≥
p
p
1≤i≤n
2 1 + δ0
2
4
δ0 2u
This completes the proof.
We also need Talagrand’s concentration inequality [58] for the empirical
process in the form given by [42], recorded as follows.
Lemma 21. Let F be a countable class of real-valued measurable functions
such that supf ∈F kf k∞ ≤ b. Then
!
√
x
≤ e−x ,
P sup |Gn f | ≥ 2E sup |Gn f | + 8σ 2 x + 34.5b √
n
f ∈F
f ∈F
√
where σ 2 ≡ supf ∈F VarP f , and Gn ≡ n(Pn − P ).
Proof of Lemma 18. The proof is divided into two steps. In the first step,
we handle Z, i.e. the multiplier empirical process part. In the second step
we handle the residual term, i.e. the purely empirical process part.
(Step 1) We first claim that there exists some absolute constant C1 > 0
such that
2
1 1/p′ −1/(p′ )2
n
≥ C1 (εp′ ) ∧ 1 n−4ε .
(6.22)
P Z≥
32
Let G be the class of indicators functions 1[a,b] with 0 ≤ a ≤ b ≤ 1. Then
p
since δn ≥ log n/n, by local maximal inequalities for the empirical process
36
Q. HAN AND J. A. WELLNER
(Proposition 5),
(6.23)
E
sup
n
X
2
g∈G:P g 2 ≤δn
i=1
εi g(Xi ) .
p
√
′p
nδn log(1/δn ) . n1/2p log n
where in the last inequality we used ρ .√1. Applying Theorem
1 and noting
√
′
1/p , we see
that p ≤ log n/ log log n implying n1/2p log n ≤ n1/2p log n ≤
Pn
n
that for some absolute constant C > 0, EZ ≤ E supg∈G:P g2 ≤δn2 | i=1 ξi g(Xi )| ≤
Cn1/p kξ1 kp,1 . By the moment inequality Lemma 1, we have for any q ≥ 1,
√
q
q
q
q
EZ ≤ Cq (EZ) + ( n kξ1 k2 δn ) + E max |ξi |
1≤i≤n
′
≤ Cq′ nq/p kξ1 kqp,1 + nq/2p kξ1 kq2 + E max |ξi |q
1≤i≤n
q
′
q/p
q
≤ 2Cq n (kξ1 kp,1 ∨ kξ1 k2 ) + E max |ξi | .
1≤i≤n
Now using the Paley-Zygmund inequality (Lemma 15) and Lemma 17, we
see that
q ′
EZ
1
′
P Z > EZ ≥ 2−q
2
(EZ q )1/q
!q ′
E
max
−1 n1−1/p′ |ξj |
1≤j≤4
≥ Cq′′
.
n1/p (kξ1 kp,1 ∨ kξ1 k2 ) + (E max1≤i≤n |ξi |q )1/q
By Lemma 19, the above display can be further estimated from below by
!q ′
′
′ 2
n1/p −1/(p )
′′′
Cq
n1/p (kξ1 kp,1 ∨ kξ1 k2 ) + [(p′ + q)/(p′ − q)]1/q n1/p′
!2
1/p′ −1/(p′ )2
n
(choose q = q ′ = 2)
≥ C′
′
n1/p (kξ1 kp,1 ∨ kξ1 k2 ) + n1/p
2 − 2 + 2 − 2
2
≥ C ′′ (εp′ ) ∧ 1 n p p′ (p′ )2 ≥ C ′′ (εp′ ) ∧ 1 n−4ε
where in the last line we use the following facts: (i) p1 − p1′ = ε, (ii) p′ ≥ p ≥
ε−1/2 and (iii)
Z ∞
Z ∞
1
1/p
dt
kξ1 kp,1 =
P(|ξ1 | > t)
dt =
(1 + tp′ )1/p
0
0
Z ∞
dt
1
1
≍1+
= 1 + ′,
′ /p = 1 +
′
p
p /p − 1
εp
t
1
1
≍ 1.
kξ1 k2 ≤ kξ1 k2,1 ≍ 1 + ′
(p /2) − 1
MULTIPLIER INEQUALITIES
37
The proof of (6.22) is complete by noting that
1 −1/p′ 1/p′ −1/(p′ )2
1
−1 1/p′ −1/(p′ )2
.
≥
16
n
|ξ
|
≥
n
max
4
EZ ≥ E
j
2 1≤j≤4−1 n1−1/p′
8
(Step 2) We next claim that for any C2 > 0, there exists some absolute
constant C3 > 0 such that
(6.24)
!
n
X
p
1/2p′
2
2
(f (Xi ) − P f ) ≤ C3 n
P
sup
log n ≥ 1 − e−C2 log n .
2
f ∈F̃n :P f 2 ≤δn
i=1
Note that by contraction principle and (6.23),
E
sup
n
X
(f 2 (Xi ) − P f 2 ) . E
2
f ∈F̃n :P f 2 ≤δn
i=1
n
X
sup
2
g∈G:P g 2 ≤δn
i=1
εi g(Xi ) . n1/2p
′
p
log n.
The claim (6.24) now follows from the above display combined with Talagrand’s concentration inequality (Lemma 21) applied with x = C2 log n.
Now the claimed inequality in the lemma follows by considering the event
that is the intersection of the events indicated in (6.22) and (6.24).
Now we are in a good position to prove the second claim of Theorem 5.
′ 2
n
Proof of Theorem 5: claim (2). Suppose that n1/(p ) ≤ 2, p′ < 2 log(64Clog
3 )+log log n
and ρ = 1/8. Then
′p
1 1/p′ −1/(p′ )2
n
− C3 n1/2p log n − nδn2
16
p
p
1
1 1/p′
1/2p′
2
1/2p′
1/p′
=
log
n
≥
log n > 0.
−
C
n
−
ρ
n
−
C
n
n
3
3
′
2
64
16n1/(p )
o n
o
n
′√
1 1/p′ −1/(p′ )2
− C3 n1/2p log n ⊂ Z̃ − nδn2 > 0 ⊂ {Fn (δn ) > En (δn /2)},
n
Since Z̃ ≥ 16
it follows from Lemma 18 that
2
P(En ) ≡ P (Fn (δn ) > En (δn /2)) ≥ C1 (εp′ ) ∧ 1 n−4ε − e−C2 log n ,
′
provided further ε ≥ 1/p2 , p ≥ 3, n ≥ 2 and n1/p ≥ 64 log n. Equivalently,
p
log n
ε ≥ 1/p2 , p ≥ 3, n ≥ 2,
.
log2 n ≤ p′ ≤
log(64 log n) + 2 log(64C3 )
Furthermore, since p′ = p/(1 − pε) ≤ 2p if ε ≤ 1/2p, it suffices to require
p
log n
2
64C32
1−δ
1/p ≤ ε ≤ 1/2p, n ≥ nδ ∨ e
,
,
log2 n ≤ p ≤ (log n)
≤
2 log(64 log n)
where nδ ≡ min{n ≥ 2 : (log n)δ ≥ 2 log(64 log n)}. Hence for n in the
indicated range (i.e. large enough depending on δ, C3 ), we have
P(En ) ≥ C1 (log n)−2 n−4ε − e−C2 log n ≥ C(log n)−2 n−4ε ,
38
Q. HAN AND J. A. WELLNER
where the first inequality follows from εp′ ≥ εp ≥ 1/p ≥ 1/ log n, and
the second inequality follows for n large enough by choosing C2 = 3 since
n−4ε ≥ n−2 . By Proposition 6,
E kfn∗ − f0 kL2 (P ) ≥ E kfn∗ − f0 kL2 (P ) 1En
δn
· C(log n)−2 n−4ε ≥ C ′ n−1/2+1/2p−4.5ε (log n)−2 .
2
For any 1/(1 − δ) < a < 2 so that p = (log n)1/a , we may choose ε = p−a .
The claim then follows by noting n−ε = n−1/ log n = e−1 , and kξ1 kp,1 ≍
1 + (εp)−1 = 1 + (log n)1−1/a . log n.
≥
7. Remaining proofs
7.1. Proof of Lemma 3. We need some auxiliary lemmas before the proof
of Lemma 3.
Pd
Lemma 22. Let Fβ : Rd → R be defined by Fβ (x) = β −1 log
exp(βxi ) .
i=1
P
Then supx∈Rd dj,k,l,m=1 |∂jklm Fβ (x)| ≤ 25β 3 .
P
Proof. Let πj (x) = ∂j Fβ (x) = exp(βxj )/ di=1 exp(βxi ) and δij = 1i=j .
Then it is easy to verify that (see Lemma 4.3 of [18])
∂jk Fβ = ∂j (πk ) = β(δjk πj − πj πk ),
∂jkl Fβ = β 2 δjk δjl πl − δjk πj πl + 2πj πk πl − (δjl + δkl )πk πl .
Further taking derivative with respect to xm , we have
∂jklmFβ = β 2 δjk δjl ∂m (πl ) − δjk ∂m (πj πl ) + 2∂m (πj πk πl ) − (δjl + δkl )∂m (πk πl )
= β 2 (I) − (II) + 2(III) − (IV ) ,
where (I) = βδjk δjl δlm πl − πl πm , (II) = βδjk (δjm + δlm )πj πl − 2πj πl πm ,
(III)
= β (δjm + δkm + δlm )πj πk πl − 3πj πk πl πm , and (IV ) = β(δjl +
δkl ) (δ
Pjm + δkm )πj πk − 2πj πk πm . Now summing over j, k, l, m by noting
that j πj = 1 yields the claim.
Proof of Lemma 3. In the proof we write k ≡ n to line up with the notation
used in [17]. Slightly abusing the notation, we use simply Xi ’s to denote
εi Xi ’s. Let Y1 , . . . , Yn be centered i.i.d. Gaussian random vectors in Rd such
that EY1 Y1⊤ = EX1 X1⊤ . We first claim that it suffices to prove that,
4
4
3 −1
(7.1) |EFβ (X) − Fβ (Y )| . β n E max |X1j | ∨ |Y1j |
≡ β 3 n−1 M̄4 ,
1≤j≤d
Pn
d
d
Y =
i=1 Yi ∈ R , and Fβ : R → R is
Pd
defined by Fβ (x)
i=1 exp(βxi ) . Once (7.1) is proved, we use
the inequality 0 ≤ Fβ (x) − max1≤j≤d xj ≤ β −1 log d to obtain that
where X =
√1
n
Pn
Rd ,
i=1 Xi ∈
= β −1 log
n
√1
n
n
1 X
1 X
E max √
Xij − E max √
Yij
1≤j≤d
1≤j≤d
n
n
i=1
i=1
. β 3 n−1 M̄4 + β −1 log d.
MULTIPLIER INEQUALITIES
39
1/4
and
The conclusion of Lemma 3 follows by taking β = n log d/M̄4
controlling the size of the Gaussian maxima.
The proof of (7.1) proceeds in similar lines as in Lemma I.1 of [17] by
a fourth moment argument instead of a√third moment
one P
used therein.
√
n
We provide details below. Let Z(t) = tX + 1 − tY =
i=1 Z
√i (t) be
the Slepian’s interpolation between X and Y , where Zi (t) = √1n ( tXi +
√
1 − tYi ). Let Z (i) (t) = Z(t) − Zi (t). Then,
(7.2)
EFβ (X) − EFβ (Y ) = EFβ (Z(1)) − EFβ (Z(0))
Z 1X
Z 1
d
n X
h
i
d
E ∂j Fβ (Z(t))Żij (t) dt
EFβ (Z(t)) dt =
=
0 i=1 j=1
0 dt
1
where Żij (t) = 2√1 n √1t Xij − √1−t
Yij . Now using Taylor expansion for
∂j Fβ (·) at Z (i) (t), we have
(7.3)
X
∂j Fβ (Z(t)) = ∂j Fβ (Z (i) (t)) +
∂jk Fβ (Z (i) (t))Zik (t)
k
+
X
(i)
∂jkl Fβ (Z (t))Zik (t)Zil (t)
k,l
+
XZ
1
∂jklm Fβ (Z (i) (t) + τ Zi (t))Zik (t)Zil (t)Zim (t) dτ.
k,l,m 0
Hence (7.2) can be split into four terms according to (7.3). Now the key
observation here is that Z (i) (t) is independent of Zi· , Żi· . Since EŻij (t) = 0,
the contribution of the first order term in (7.3) vanishes. Similar observation holds for the second and third order terms. For the second order
term, we only need to verify EŻij (t)Zik (t) = 0; this follows from the construction
of Y that matches
second moments of X: EŻij (t)Zik (t) =
√ the √
1
1
1
1
√ Xij − √
(EXij Xik − EYij Yik ) =
tXik + 1 − tYik = 2n
Y
2n E
1−t ij
t
0. For the third order term,
EŻij (t)Zik (t)Zil (t)
√
√
√
√
1
1
1
= 3/2 E √ Xij − √
tXik + 1 − tYik
tXil + 1 − tYil
Yij
2n
1−t
t
√
√
3/2 −1
= (2n )
tEXij Xik Xil − 1 − tEYij Yik Yil .
Cross terms in the calculation of the last line in the above display all vanish
by the independence and centeredness of X and Y . The first term of the
above display is 0 since (recall Xi stands for εi Xi throughout the proof)
Eε3i Xij Xik Xil = Eε3i · EXij Xik Xil = 0 by the independence between the
Rademacher εi and Xi . The second term is also zero by a similar argument:
since Yi =d εi Yi for a Rademacher random variable εi independent of Yi ,
40
Q. HAN AND J. A. WELLNER
EYij Yik Yil = Eε3i · EYij Yik Yil = 0. Hence the only non-trivial contribution of
(7.3) in (7.2) is the fourth order term:
(7.4)
|EFβ (X) − Fβ (Y )|
Z 1Z 1
d
n
X
X
E ∂jklmFβ (Z (i) (t) + τ Zi (t))Żij (t)Zik (t)Zil (t)Zim (t) dτ dt
≤
≤
i=1 j,k,l,m=1 0
0
n Z
X
d
X
i=1
0
1Z 1
E
0
∂jklmFβ (Z (i) (t) + τ Zi (t))
j,k,l,m=1
×
≤ 25β
3
n Z
X
i=1
max
1≤k,l,m≤d
Żij (t)Zik (t)Zil (t)Zim (t)
1
E
0
max
1≤j,k,l,m≤d
dτ dt
Żij (t)Zik (t)Zil (t)Zim (t) dt
where the last inequality follows from the dimension free property of the
third derivatives of soft max function Fβ (Lemma 22). Now the claim (7.1)
follows by noting that
3/4
4 1/4
4
E max
E max |Zij |
Żij (t)Zik (t)Zil (t)Zim (t) ≤ E max Żij
1≤j≤d
1≤j,k,l,m≤d
1≤j≤d
1
1
4
4
−2
√
√
E max |X1j | ∨ |Y1j |
∨
.n
1≤j≤d
1−t
t
R1
1
dt < ∞ converges.
and the fact that the integral 0 √1t ∨ √1−t
7.2. Proof of Proposition 3.
Proof of Proposition 3. We will construct a sequence of {fn } ⊂ F such that
for any κ > 0,
(7.5)
P(|fn | ≥ κ kfn kL2 (P ) ) → 0,
which implies that the small-ball condition fails. √
For (1), let fn ≡ 1[0,1/n] . Then kfn kL2 (P ) = 1/ n, and hence for any
κ > 0,
√
P(|fn | ≥ κ kfn kL2 (P ) ) = P(|fn (X)| ≥ κ/ n) = P(X ∈ [0, 1/n]) = 1/n → 0,
thereby verifying (7.5).
Next consider
√ (2). We let fn (x) ≡ (1/n − x)+ . Then it is easy to see that
kfn kL2 (P ) = 1/ 3n3 . Hence for any κ > 0,
P(|fn | ≥ κ kfn kL2 (P ) ) ≤ P(|fn (X)| ≥ 0) = 1/n → 0,
thereby verifying (7.5). (3) follows from the same construction as in (2)
by replacing fn therein with smooth versions. This can be done, e.g., by
2
choosing fn (x) ≡ φ(x − 1/n) where φ(x) = e−1/x 1x≤0 [since the derivatives
of φ is uniformly bounded on [−1, 0]].
MULTIPLIER INEQUALITIES
41
7.3. Proof of Proposition 4.
Proof of Proposition 4. Consider F = {1[a,b] : 0 ≤ a ≤ b ≤ 1, b − a ≤
δ2 , a, b ∈ Q}. It is easy to see that supf ∈F (δ) |G1 f | ≥ 1 − δ2 . On the other
hand, by the local maximal inequality for the empirical process (see Proposiclass (seeSection 2.6 of [68]), we
tion 5) and the fact that F is a VC-subgraph
√
p
log(1/δ)
have E supf ∈F (δ) |Gn f | . δ log(1/δ) 1 + √nδ
. Talagrand’s concen-
tration inequality (Lemma 21) then implies that with probability exceeding
1 − exp(−u2 2s0 ),
!
p
p
√
log(1/δ)
34.5u2 2s0
√
√
sup |Gn f | ≤ Cδ log(1/δ) 1 +
+ 8u2s0 /2 δ +
.
nδ
n
f ∈F (δ)
p
Note that for δ ≥ log n/n and n ≥ 3, the right side of the above display
can be further bounded from above up to an absolute constant by
p
p
p
δ log(1/δ) + u2s0 /2 δ + u2 2s0 δ log(1/δ) . u2 2s0 δ log(1/δ).
Nowpwe choose n0 ≡ n0 (ε, u, s0 ) so large such that for any n ≥ n0 , if
δ ≡ log n/n, then on an event with probability exceeding 1− exp(−u2 2s0 ),
p
sup |Gn f | ≤ C ′ u2 2s0 δ log(1/δ) < ε(1 − δ2 ) ≤ ε sup |G1 f | .
f ∈F (δ)
f ∈F (δ)
The claim now follows from (5.3) applied for n = 1.
7.4. Proof of Lemma 4.
Proof of Lemma 4. Without loss of generality we assume that an ≤ 1 for
all n = 0, 1, . . .. For any ε ∈ (0, 1), since an vanishes asymptotically, there
exists some Nε for which an ≤ ε as long as n ≥ Nε . Consider
(
ϕ(t),
t ≤ Nε ;
ψε (t) ≡
(1 − ε)ϕ(Nε ) + εϕ(t), t > Nε .
Then it is easy to verify that ψε is a concave function and majorizes n 7→
an ϕ(n) [since ψε (n) ≥ εϕ(n) ≥ an ϕ(n) for n ≥ Nε and ψε (n) = ϕ(n) ≥
an ϕ(n) for n < Nε by the assumption that an ≤ 1]. Hence by definition of
ψ, it follows that lim supt→∞ ψ(t)/ϕ(t) ≤ limt→∞ ψε (t)/ϕ(t) = ε. The claim
now follows by taking ε → 0.
Acknowledgements
The authors would like to thank Vladimir Koltchinskii and Richard Samworth for helpful comments and references on an earlier version of the paper.
42
Q. HAN AND J. A. WELLNER
References
[1] K. S. Alexander. The non-existence of a universal multiplier moment for the central
limit theorem. In Probability in Banach Spaces V, pages 15–16. Springer, 1985.
[2] N. T. Andersen, E. Giné, and J. Zinn. The central limit theorem for empirical processes under local conditions: the case of Radon infinitely divisible limits without
Gaussian component. Trans. Amer. Math. Soc., 308(2):603–635, 1988.
[3] J.-Y. Audibert and O. Catoni. Robust linear least squares regression. Ann. Statist.,
39(5):2766–2794, 2011.
[4] Z. D. Bai, J. W. Silverstein, and Y. Q. Yin. A note on the largest eigenvalue of
a large-dimensional sample covariance matrix. J. Multivariate Anal., 26(2):166–168,
1988.
[5] Z. D. Bai and Y. Q. Yin. Limit of the smallest eigenvalue of a large-dimensional
sample covariance matrix. Ann. Probab., 21(3):1275–1294, 1993.
[6] P. L. Bartlett, O. Bousquet, and S. Mendelson. Local Rademacher complexities. Ann.
Statist., 33(4):1497–1537, 2005.
[7] P. L. Bartlett and S. Mendelson. Empirical minimization. Probab. Theory Related
Fields, 135(3):311–334, 2006.
[8] L. Birgé and P. Massart. Rates of convergence for minimum contrast estimators.
Probab. Theory Related Fields, 97(1-2):113–150, 1993.
[9] S. Boucheron, G. Lugosi, and P. Massart. Concentration inequalities. Oxford University Press, Oxford, 2013. A nonasymptotic theory of independence, With a foreword
by Michel Ledoux.
[10] C. Brownlees, E. Joly, and G. Lugosi. Empirical risk minimization for heavy-tailed
losses. Ann. Statist., 43(6):2507–2536, 2015.
[11] S. Bubeck, N. Cesa-Bianchi, and G. Lugosi. Bandits with heavy tail. IEEE Trans.
Inform. Theory, 59(11):7711–7717, 2013.
[12] P. Bühlmann and S. van de Geer. Statistics for high-dimensional data. Springer Series
in Statistics. Springer, Heidelberg, 2011. Methods, theory and applications.
[13] B. Carl. Entropy numbers of embedding maps between Besov spaces with an application to eigenvalue problems. Proc. Roy. Soc. Edinburgh Sect. A, 90(1-2):63–70,
1981.
[14] O. Catoni. Challenging the empirical mean and empirical variance: a deviation study.
Ann. Inst. Henri Poincaré Probab. Stat., 48(4):1148–1185, 2012.
[15] O. Catoni. Pac-bayesian bounds for the gram matrix and least squares regression
with a random design. arXiv preprint arXiv:1603.05229, 2016.
[16] S. Chatterjee. A new perspective on least squares under convex constraint. Ann.
Statist., 42(6):2340–2381, 2014.
[17] V. Chernozhukov, D. Chetverikov, and K. Kato. Gaussian approximations and multiplier bootstrap for maxima of sums of high-dimensional random vectors. Ann. Statist.,
41(6):2786–2819, 2013.
[18] V. Chernozhukov, D. Chetverikov, and K. Kato. Gaussian approximation of suprema
of empirical processes. Ann. Statist., 42(4):1564–1597, 2014.
[19] V. H. de la Peña and E. Giné. Decoupling. Probability and its Applications (New
York). Springer-Verlag, New York, 1999. From dependence to independence, Randomly stopped processes. U -statistics and processes. Martingales and beyond.
[20] L. Devroye, M. Lerasle, G. Lugosi, and R. I. Oliveira. Sub-Gaussian mean estimators.
Ann. Statist., 44(6):2695–2725, 2016.
[21] R. M. Dudley. The sizes of compact subsets of Hilbert space and continuity of Gaussian processes. J. Functional Analysis, 1:290–330, 1967.
[22] E. Giné and V. Koltchinskii. Concentration inequalities and asymptotic results for
ratio type empirical processes. Ann. Probab., 34(3):1143–1216, 2006.
MULTIPLIER INEQUALITIES
43
[23] E. Giné, R. Latala, and J. Zinn. Exponential and moment inequalities for U -statistics.
In High dimensional probability, II (Seattle, WA, 1999), volume 47 of Progr. Probab.,
pages 13–38. Birkhäuser Boston, Boston, MA, 2000.
[24] E. Giné and R. Nickl. Mathematical foundations of infinite-dimensional statistical
models, volume 40. Cambridge University Press, 2015.
[25] E. Giné and J. Zinn. Central limit theorems and weak laws of large numbers in certain
Banach spaces. Z. Wahrsch. Verw. Gebiete, 62(3):323–354, 1983.
[26] E. Giné and J. Zinn. Some limit theorems for empirical processes. Ann. Probab.,
12(4):929–998, 1984. With discussion.
[27] E. Giné and J. Zinn. Lectures on the central limit theorem for empirical processes.
In Probability and Banach spaces (Zaragoza, 1985), volume 1221 of Lecture Notes in
Math., pages 50–113. Springer, Berlin, 1986.
[28] D. Hsu and S. Sabato. Loss minimization and parameter estimation with heavy tails.
J. Mach. Learn. Res., 17:Paper No. 18, 40, 2016.
[29] E. Joly, G. Lugosi, and R. I. Oliveira. On the estimation of the mean of a random
vector. Electron. J. Stat., 11(1):440–451, 2017.
[30] V. Koltchinskii. Local Rademacher complexities and oracle inequalities in risk minimization. Ann. Statist., 34(6):2593–2656, 2006.
[31] V. Koltchinskii and D. Panchenko. Rademacher processes and bounding the risk of
function learning. In High dimensional probability, II (Seattle, WA, 1999), volume 47
of Progr. Probab., pages 443–457. Birkhäuser Boston, Boston, MA, 2000.
[32] G. Lecué and S. Mendelson. Learning subgaussian classes: Upper and minimax
bounds. arXiv preprint arXiv:1305.4825, 2013.
[33] G. Lecué and S. Mendelson. Sparse recovery under weak moment assumptions. arXiv
preprint arXiv:1401.2188, 2014.
[34] G. Lecué and S. Mendelson. Regularization and the small-ball method i: sparse recovery. arXiv preprint arXiv:1601.05584, 2016.
[35] M. Ledoux and M. Talagrand. Conditions d’intégrabilité pour les multiplicateurs dans
le TLC banachique. Ann. Probab., 14(3):916–921, 1986.
[36] M. Ledoux and M. Talagrand. Un critère sur les petites boules dans le théorème limite
central. Probab. Theory Related Fields, 77(1):29–47, 1988.
[37] M. Ledoux and M. Talagrand. Probability in Banach Spaces. Classics in Mathematics.
Springer-Verlag, Berlin, 2011. Isoperimetry and processes, Reprint of the 1991 edition.
[38] G. Lugosi and S. Mendelson. Risk minimization by median-of-means tournaments.
arXiv preprint arXiv:1608.00757, 2016.
[39] G. Lugosi and S. Mendelson. Regularization, sparse recovery, and median-of-means
tournaments. arXiv preprint arXiv:1701.04112, 2017.
[40] G. Lugosi and S. Mendelson. Sub-gaussian estimators of the mean of a random vector.
arXiv preprint arXiv:1702.00482, 2017.
[41] D. M. Mason. The asymptotic distribution of weighted empirical distribution functions. Stochastic Process. Appl., 15(1):99–109, 1983.
[42] P. Massart. About the constants in Talagrand’s concentration inequalities for empirical processes. Ann. Probab., 28(2):863–884, 2000.
[43] P. Massart and E. Nédélec. Risk bounds for statistical learning. Ann. Statist.,
34(5):2326–2366, 2006.
[44] P. Massart and E. Rio. A uniform Marcinkiewicz-Zygmund strong law of large numbers for empirical processes. In Asymptotic methods in probability and statistics (Ottawa, ON, 1997), pages 199–211. North-Holland, Amsterdam, 1998.
[45] S. Mendelson. Learning without concentration. J. ACM, 62(3):Art. 21, 25, 2015.
[46] S. Mendelson. ‘local’ vs. ‘global’ parameters–breaking the gaussian complexity barrier.
arXiv preprint arXiv:1504.02191, 2015.
[47] S. Mendelson. On aggregation for heavy-tailed classes. arXiv preprint
arXiv:1502.07097, 2015.
44
Q. HAN AND J. A. WELLNER
[48] S. Mendelson. On multiplier processes under weak moment assumptions. arXiv
preprint arXiv:1601.06523, 2016.
[49] S. Mendelson. Upper bounds on product and multiplier empirical processes. Stochastic
Process. Appl., 126(12):3652–3680, 2016.
[50] S. Minsker. Geometric median and robust estimation in Banach spaces. Bernoulli,
21(4):2308–2335, 2015.
[51] R. Nickl and S. van de Geer. Confidence sets in sparse regression. Ann. Statist.,
41(6):2852–2876, 2013.
[52] G. Raskutti, M. J. Wainwright, and B. Yu. Restricted eigenvalue properties for correlated Gaussian designs. J. Mach. Learn. Res., 11:2241–2259, 2010.
[53] T. Robertson, F. T. Wright, and R. L. Dykstra. Order restricted statistical inference.
Wiley Series in Probability and Mathematical Statistics: Probability and Mathematical Statistics. John Wiley & Sons, Ltd., Chichester, 1988.
[54] M. Rudelson and S. Zhou. Reconstruction from anisotropic random measurements.
IEEE Trans. Inform. Theory, 59(6):3434–3447, 2013.
[55] G. R. Shorack and J. A. Wellner. Empirical processes with applications to statistics,
volume 59 of Classics in Applied Mathematics. Society for Industrial and Applied
Mathematics (SIAM), Philadelphia, PA, 2009.
[56] V. Sivakumar, A. Banerjee, and P. K. Ravikumar. Beyond sub-gaussian measurements: High-dimensional structured estimation with sub-exponential designs. In Advances in neural information processing systems, pages 2206–2214, 2015.
[57] V. Strassen and R. M. Dudley. The central limit theorem and ε-entropy. In Probability
and Information Theory (Proc. Internat. Sympos., McMaster Univ., Hamilton, Ont.,
1968), pages 224–231. Springer, Berlin, 1969.
[58] M. Talagrand. New concentration inequalities in product spaces. Invent. Math.,
126(3):505–563, 1996.
[59] M. Talagrand. Upper and lower bounds for stochastic processes, volume 60 of Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys
in Mathematics. Springer, Heidelberg, 2014. Modern methods and classical problems.
[60] R. Tibshirani. Regression shrinkage and selection via the lasso. J. Roy. Statist. Soc.
Ser. B, 58(1):267–288, 1996.
[61] H. Tribel′. Interpolation properties of ε-entropy and of diameters. Geometric characteristics of the imbedding of function spaces of Sobolev-Besov type. Mat. Sb. (N.S.),
98(140)(1 (9)):27–41, 157, 1975.
[62] A. B. Tsybakov. Introduction to nonparametric estimation. Springer Series in Statistics. Springer, New York, 2009. Revised and extended from the 2004 French original,
Translated by Vladimir Zaiats.
[63] S. van de Geer. Estimating a regression function. Ann. Statist., 18(2):907–924, 1990.
[64] S. van de Geer, P. Bühlmann, Y. Ritov, and R. Dezeure. On asymptotically optimal
confidence regions and tests for high-dimensional models. Ann. Statist., 42(3):1166–
1202, 2014.
[65] S. A. van de Geer. Applications of Empirical Process Theory, volume 6 of Cambridge
Series in Statistical and Probabilistic Mathematics. Cambridge University Press, Cambridge, 2000.
[66] S. A. van de Geer and M. Wainwright. On concentration for (regularized) empirical
risk minimization. arXiv preprint arXiv:1512.00677, 2015.
[67] A. van der Vaart and J. A. Wellner. A local maximal inequality under uniform entropy.
Electron. J. Stat., 5:192–203, 2011.
[68] A. W. van der Vaart and J. A. Wellner. Weak Convergence and Empirical Processes.
Springer Series in Statistics. Springer-Verlag, New York, 1996.
[69] Y. Yang. Nonparametric regression with dependent errors. Bernoulli, 7(4):633–655,
2001.
MULTIPLIER INEQUALITIES
45
[70] Y. Yang and A. Barron. Information-theoretic determination of minimax rates of
convergence. Ann. Statist., 27(5):1564–1599, 1999.
(Q. Han) Department of Statistics, Box 354322, University of Washington,
Seattle, WA 98195-4322, USA.
E-mail address: royhan@uw.edu
(J. A. Wellner) Department of Statistics, Box 354322, University of Washington, Seattle, WA 98195-4322, USA.
E-mail address: jaw@stat.washington.edu
| 10math.ST |
IEEE SIGNAL PROCESSING LETTERS, ACCEPTED NOVEMBER 2016
1
Deep Convolutional Neural Networks and Data
Augmentation for Environmental Sound
Classification
arXiv:1608.04363v2 [cs.SD] 28 Nov 2016
Justin Salamon and Juan Pablo Bello
Abstract—The ability of deep convolutional neural networks
(CNN) to learn discriminative spectro-temporal patterns makes
them well suited to environmental sound classification. However,
the relative scarcity of labeled data has impeded the exploitation
of this family of high-capacity models. This study has two
primary contributions: first, we propose a deep convolutional
neural network architecture for environmental sound classification. Second, we propose the use of audio data augmentation for
overcoming the problem of data scarcity and explore the influence
of different augmentations on the performance of the proposed
CNN architecture. Combined with data augmentation, the proposed model produces state-of-the-art results for environmental
sound classification. We show that the improved performance
stems from the combination of a deep, high-capacity model and
an augmented training set: this combination outperforms both
the proposed CNN without augmentation and a “shallow” dictionary learning model with augmentation. Finally, we examine
the influence of each augmentation on the model’s classification
accuracy for each class, and observe that the accuracy for each
class is influenced differently by each augmentation, suggesting
that the performance of the model could be improved further by
applying class-conditional data augmentation.
Index Terms—Environmental sound classification, deep convolutional neural networks, deep learning, urban sound dataset.
I. I NTRODUCTION
T
HE problem of automatic environmental sound classification has received increasing attention from the research
community in recent years. Its applications range from context
aware computing [1] and surveillance [2] to noise mitigation
enabled by smart acoustic sensor networks [3].
To date, a variety of signal processing and machine learning
techniques have been applied to the problem, including matrix
factorization [4]–[6], dictionary learning [7], [8], wavelet filterbanks [8], [9] and most recently deep neural networks [10],
[11]. See [12]–[14] for further reviews of existing approaches.
In particular, deep convolutional neural networks (CNN) [15]
are, in principle, very well suited to the problem of environmental sound classification: first, they are capable of capturing
energy modulation patterns across time and frequency when
applied to spectrogram-like inputs, which has been shown
to be an important trait for distinguishing between different,
often noise-like, sounds such as engines and jackhammers [8].
Second, by using convolutional kernels (filters) with a small
J. Salamon (justin.salamon@nyu.edu) is with the Music and Audio Research Laboratory (MARL) and the Center for Urban Science and Progress
(CUSP) at New York University, USA. J. P. Bello (jpbello@nyu.edu) is with
the Music and Audio Research Laboratory at New York University, USA.
receptive field, the network should, in principle, be able to
successfully learn and later identify spectro-temporal patterns
that are representative of different sound classes even if part
of the sound is masked (in time/frequency) by other sources
(noise), which is where traditional audio features such as MelFrequency Cepstral Coefficients (MFCC) fail [16]. Yet the
application of CNNs to environmental sound classification has
been limited to date. For instance, the CNN proposed in [11]
obtained comparable results to those yielded by a dictionary
learning approach [7] (which can be considered an instance of
“shallow” feature learning), but did not improve upon it.
Deep neural networks, which have a high model capacity,
are particularly dependent on the availability of large quantities of training data in order to learn a non-linear function
from input to output that generalizes well and yields high
classification accuracy on unseen data. A possible explanation
for the limited exploration of CNNs and the difficulty to
improve on simpler models is the relative scarcity of labeled
data for environmental sound classification. While several new
datasets have been released in recent years (e.g., [17]–[19]),
they are still considerably smaller than the datasets available
for research on, for example, image classification [20].
An elegant solution to this problem is data augmentation,
that is, the application of one or more deformations to a
collection of annotated training samples which result in new,
additional training data [20]–[22]. A key concept of data augmentation is that the deformations applied to the labeled data
do not change the semantic meaning of the labels. Taking an
example from computer vision, a rotated, translated, mirrored
or scaled image of a car would still be a coherent image of a
car, and thus it is possible to apply these deformations to produce additional training data while maintaining the semantic
validity of the label. By training the network on the additional
deformed data, the hope is that the network becomes invariant
to these deformations and generalizes better to unseen data.
Semantics-preserving deformations have also been proposed
for the audio domain, and have been shown to increase model
accuracy for music classification tasks [22]. However, in the
case of environmental sound classification the application of
data augmentation has been relatively limited (e.g., [11], [23]),
with the author of [11] (which used random combinations of
time shifting, pitch shifting and time stretching for data
augmentation) reporting that “simple augmentation techniques
proved to be unsatisfactory for the UrbanSound8K dataset
given the considerable increase in training time they generated
and negligible impact on model accuracy”.
IEEE SIGNAL PROCESSING LETTERS, ACCEPTED NOVEMBER 2016
2
In this paper we present a deep convolutional neural network
architecture with localized (small) kernels for environmental
sound classification. Furthermore, we propose the use of data
augmentation to overcome the problem of data scarcity and explore different types of audio deformations and their influence
on the model’s performance. We show that the proposed CNN
architecture, in combination with audio data augmentation,
yields state-of-the-art performance for environmental sound
classification.
II. M ETHOD
A. Deep Convolutional Neural Network
The deep convolutional neural network (CNN) architecture
proposed in this study is comprised of 3 convolutional layers
interleaved with 2 pooling operations, followed by 2 fully connected (dense) layers. Similar to previously proposed feature
learning approaches applied to environmental sound classification (e.g., [7]), the input to the network consists of timefrequency patches (TF-patches) taken from the log-scaled melspectrogram representation of the audio signal. Specifically,
we use Essentia [24] to extract log-scaled mel-spectrograms
with 128 components (bands) covering the audible frequency
range (0-22050 Hz), using a window size of 23 ms (1024
samples at 44.1 kHz) and a hop size of the same duration.
Since the excerpts in our evaluation dataset (described below)
are of varying duration (up to 4 s), we fix the size of the input
TF-patch X to 3 seconds (128 frames), i.e. X ∈ R128×128 . TFpatches are extracted randomly (in time) from the full log-melspectrogram of each audio excerpt during training as described
further down.
Given our input X, the network is trained to learn the
parameters Θ of a composite nonlinear function F(·|Θ) which
maps X to the output (prediction) Z:
Z = F(X|Θ) = fL (· · · f2 (f1 (X|θ1 )|θ2 )|θL ),
(1)
where each operation f` (·|θ` ) is referred to as a layer of the
network, with L = 5 layers in our proposed architecture. The
first three layers, ` ∈ {1, 2, 3}, are convolutional, expressed
as:
Z` = f` (X` |θ` ) = h(W ∗ X` + b),
θl = [W, b]
(2)
where X` is a 3-dimensional input tensor consisting of N
feature maps, W is a collection of M 3-dimensional kernels
(also referred to as filters), ∗ represents a valid convolution,
b is a vector bias term, and h(·) is a point-wise activation
function. Thus, the shapes of X` , W , and Z` are (N, d0 , d1 ),
(M, N, m0 , m1 ) and (M, d0 −m0 +1, d1 −m1 +1) respectively.
Note that for the first layer of our network d0 = d1 = 128,
i.e., the dimensions of the input TF-patch. We apply strided
max-pooling after the first two convolutional layers ` ∈ {1, 2}
using a stride size equal to the pooling dimensions (provided
below), which reduces the dimensions of the output feature
maps and consequently speeds up training and builds some
scale invariance into the network. The final two layers, ` ∈
{4, 5}, are fully-connected (dense) and consist of a matrix
product rather than a convolution:
Z` = f` (X` |θ` ) = h(W X` + b),
θ` = [W, b]
(3)
where X` is flattened to a column vector of length N , W
has shape (M, N ), b is a vector of length M and h(·) is a
point-wise activation function.
The proposed CNN architecture is parameterized as follows:
• `1 : 24 filters with a receptive field of (5,5), i.e., W has the
shape (24,1,5,5). This is followed by (4,2) strided maxpooling over the last two dimensions (time and frequency
respectively) and a rectified linear unit (ReLU) activation
function h(x) = max(x, 0).
• `2 : 48 filters with a receptive field of (5,5), i.e., W has
the shape (48, 24, 5, 5). Like `1 , this is followed by (4,2)
strided max-pooling and a ReLU activation function.
• `3 : 48 filters with a receptive field of (5,5), i.e., W has
the shape (48, 48, 5, 5). This is followed by a ReLU
activation function (no pooling).
• `4 : 64 hidden units, i.e., W has the shape (2400, 64),
followed by a ReLU activation function.
• `5 : 10 output units, i.e., W has the shape (64,10),
followed by a softmax activation function.
Note that our use of a small receptive field (5, 5) in `1
compared to the input dimensions (128, 128) is designed to
allow the network to learn small, localized patterns that can
be fused at subsequent layers to gather evidence in support
of larger “time-frequency signatures” that are indicative of the
presence/absence of different sound classes, even when there
is spectro-temporal masking by interfering sources.
For training, the model optimizes cross-entropy loss via
mini-batch stochastic gradient descent [25]. Each batch consists of 100 TF-patches randomly selected from the training
data (without repetition). Each 3 s TF-patch is taken from a
random position in time from the full log-mel-spectrogram
representation of each training sample. We use a constant
learning rate of 0.01. Dropout [26] is applied to the input
of the last two layers, ` ∈ {4, 5}, with probability 0.5. L2regularization is applied to the weights of the last two layers
with a penalty factor of 0.001. The model is trained for 50
epochs and is checkpointed after each epoch, during which
it is trained on random minibatches until 1/8 of all training
data is exhausted (where by training data we mean all the
TF-patches extracted from every training sample starting at all
possible frame indices). A validation set is used to identify the
parameter setting (epoch) achieving the highest classification
accuracy, where prediction is performed by slicing the test
sample into overlapping TF-patches (1-frame hop), making a
prediction for each TF-patch and finally choosing the samplelevel prediction as the class with the highest mean ouptut
activation over all frames. The CNN is implemented in Python
with Lasagne [27], and we used Pescador [28] to manage and
multiplex data streams during training.
B. Data Augmentation
We experiment with 4 different audio data augmentations
(deformations), resulting in 5 augmentation sets, as detailed
below. Each deformation is applied directly to the audio signal
prior to converting it into the input representation used to
train the network (log-mel-spectrogram). Note that for each
augmentation it is important that we choose the deformation
IEEE SIGNAL PROCESSING LETTERS, ACCEPTED NOVEMBER 2016
C. Evaluation
To evaluate the proposed CNN architecture and the influence of the different augmentation sets we use the UrbanSound8K dataset [17]. The dataset is comprised of 8732 sound
clips of up to 4 s in duration taken from field recordings. The
clips span 10 environmental sound classes: air conditioner,
car horn, children playing, dog bark, drilling, engine idling,
gun shot, jackhammer, siren and street music. By using this
dataset we can compare the results of this study to previously
published approaches that were evaluated on the same data,
including the dictionary learning approach proposed in [7]
(spherical k-means, henceforth SKM) and the CNN proposed
in [11] (PiczakCNN) which has a different architecture to ours
and did not employ augmentation during training. PiczakCNN
1 We
ensured these scenes did not contain any of the target sound classes.
2 https://github.com/justinsalamon/UrbanSound8K-JAMS
Classification Accuracy
0.90
0.85
0.80
0.75
0.70
0.65
0.60
SKM
PiczakCNN
SB-CNN
SKM (aug) SB-CNN (aug)
Fig. 1. Left of the dashed line: classification accuracy without augmentation
– dictionary learning (SKM [7]), Piczak’s CNN (PiczakCNN [11]) and the
proposed model (SB-CNN). Right of the dashed line: classification accuracy
for SKM and SB-CNN with augmentation.
(a)
AI 489 10 35 52 85 262 2
1
4
AI
CA
4 379 2
CA
-1 34 -8 -6 -4
0
3
7
1 -26
17 1 830 36 18 24 6
4 13 51
CH
5
6
5
0
2 -8
DO
12 3 34 904 5
3 15 15
DO
-8 -6 -13 37 -5
2 -2
3
1 -9
DR
38 0 16 39 802 8
DR
15 -3 -11 -8 40 -19 -2 15 -26 -1
EN
62 7
5
0
1 19 1
JA
26 1
0
1
5
3 16 1 13
(b)
78 -1 -17 -40 -79 141 1 -4 -25 -54
CH
GU
6
25 15 25
4
2 69 3 23
8 37 798 5 60 5 13
0 352 0
0
1
0 175 48 1 673 61 14
1 38 28 26 15 0
True label
True label
parameters such that the semantic validity of the label is
maintained. The deformations and resulting augmentation sets
are described below:
• Time Stretching (TS): slow down or speed up the audio
sample (while keeping the pitch unchanged). Each sample
was time stretched by 4 factors: {0.81, 0.93, 1.07, 1.23}.
• Pitch Shifting (PS1): raise or lower the pitch of the
audio sample (while keeping the duration unchanged).
Each sample was pitch shifted by 4 values (in semitones):
{−2, −1, 1, 2}.
• Pitch Shifting (PS2): since our initial experiments indicated that pitch shifting was a particularly beneficial
augmentation, we decided to create a second augmentation set. This time each sample was pitch shifted by 4
larger values (in semitones): {−3.5, −2.5, 2.5, 3.5}.
• Dynamic Range Compression (DRC): compress the
dynamic range of the sample using 4 parameterizations, 3
taken from the Dolby E standard [29] and 1 (radio) from
the icecast online radio streaming server [30]: {music
standard, film standard, speech, radio}.
• Background Noise (BG): mix the sample with another
recording containing background sounds from different
types of acoustic scenes. Each sample was mixed with
4 acoustic scenes: {street-workers, street-traffic, streetpeople, park}1 . Each mix z was generated using z =
(1−w)·x+w·y where x is the audio signal of the original
sample, y is the signal of the background scene, and w is
a weighting parameter that was chosen randomly for each
mix from a uniform distribution in the range [0.1, 0.5].
The augmentations were applied using the MUDA library
[22], to which the reader is referred for further details about
the implementation of each deformation. MUDA takes an
audio file and corresponding annotation file in JAMS format
[31], [32], and outputs the deformed audio together with an
enhanced JAMS file containing all the parameters used for the
deformation. We have ported the original annotations provided
with the dataset used for evaluation in this study (see below)
into JAMS files and made them available online along with
the post-deformation JAMS files.2
3
1 23 -31 -3
EN
3
0 -34 -29 15 132 2 -34 -45 -10
GU
-1
0
JA -37
1 -9 -3 -1 14 -1
0
0
1 -17 -2 -23 -48 0 141 0 -15
SI
6
1 797 17
SI
0
1 -1 -25 14 6
0
0 16 -11
ST
24 5 54 21 4 21 0 13 14 844
ST
7
0 -39 -3 -5 13 0
6 -6 27
AI CA CH DO DR EN GU JA
Predicted label
SI
ST
AI CA CH DO DR EN GU JA
SI
ST
Predicted label
Fig. 2. (a) Confusion matrix for the proposed SB-CNN model with augmentation. (b) Difference between the confusion matrices yielded by SB-CNN with
and without augmentation: negative values (red) off the diagonal mean the
confusion is reduced with augmentation, positive values (blue) off the diagonal
mean the confusion is increased with augmentation. The positive values (blue)
along the diagonal indicate that overall the classification accuracy is improved
for all classes with augmentation.
has 2 convolutional layers followed by 3 dense layers, the
filters of the first layer are “tall” and span almost the entire
frequency dimension of the input, and the network operates
on 2 input channels: log mel-spectra and their deltas.
The proposed approach and those used for comparison in
this study are evaluated in terms of classification accuracy. The
dataset comes sorted into 10 stratified folds, and all models
were evaluated using 10-fold cross validation, where we report
the results as a box plot generated from the accuracy scores of
the 10 folds. For training the proposed CNN architecture we
use 1 of the 9 training folds in each split as a validation set
for identifying the training epoch that yields the best model
parameters when training with the remaining 8 folds.
III. R ESULTS
The classification accuracy of the proposed CNN model
(SB-CNN) is presented in Figure 1. To the left of the dashed
line we present the performance of the proposed model on
the original datast without augmentation. For comparison, we
also provide the accuracy obtained on the same dataset by the
dictionary learning approach proposed in [7] (SKM, using the
best parameterization identified by the authors in that study)
and the CNN proposed by Piczak [11] (PiczakCNN, using the
best performing model variant (LP) proposed by the author).
To the right of the dashed line we provide the performance
IEEE SIGNAL PROCESSING LETTERS, ACCEPTED NOVEMBER 2016
4
TS
PS1
PS2
DRC
BG
All
air_conditioner
car_horn
children_playing
dog_bark
Class
drilling
engine_idling
gun_shot
jackhammer
siren
street_music
All classes
0.10
0.05
0.00
0.05
0.10
0.15
0.20
∆Classification Accuracy
Fig. 3. Difference in classification accuracy for each class as a function of the
augmentation applied: Time Shift (TS), Pitch Shift (PS1 and PS2), Dynamic
Range Compression (DRC), Background Noise (BG) and all combined (All).
of the SKM model and the proposed SB-CNN once again,
this time when using the augmented dataset (all augmentations
described in Section II-B combined) for training.
We see that the proposed SB-CNN performs comparably to
SKM and PiczakCNN when training on the original dataset
without augmentation (mean accuracy of 0.74, 0.73 and 0.73
for SKM, PiczakCNN and SB-CNN, respectively). The original dataset is not large/varied enough for the convolutional
model to outperform the “shallow” SKM approach. However,
once we increase the size/variance in the dataset by means of
the proposed augmentations, the performance of the proposed
model increases significantly, yielding a mean accuracy of
0.79. The corresponding per-class accuracies (with respect to
the list of classes provided in Section II-C) are 0.49, 0.90,
0.83, 0.90, 0.80, 0.80, 0.94, 0.68, 0.85, 0.84. Importantly, we
note that while the proposed approach performs comparably
to the “shallow” SKM learning approach on the original
dataset, it significantly outperforms it (p = 0.0003 according
to a paired two-sided t-test) using the augmented training set.
Furthermore, increasing the capacity of the SKM model (by
increasing the dictionary size from k = 2000 to k = 4000) did
not yield any further improvement in classification accuracy.
This indicates that the superior performance of the proposed
SB-CNN is not only due to the augmented training set, but
rather thanks to the combination of an augmented training set
with the increased capacity and representational power of the
deep learning model.
In Figure 2(a) we provide the confusion matrix yielded by
the proposed SB-CNN model using the augmented training
set, and in Figure 2(b) we provide the difference between
the confusion matrices yielded by the proposed model with
and without augmentation. From the latter we see that overall
the classification accuracy is improved for all classes with
augmentation. However, we observe that augmentation can
also have a detrimental effect on the confusion between
specific pairs of classes. For instance, we note that while
the confusion between the air conditioner and drilling classes
is reduced with augmentation, the confusion between the air
conditioner and the engine idling classes is increased.
To gain further insight into the influence of each augmentation set on the performance of the proposed model for each
sound class, in Figure 3 we present the difference in classification accuracy (the delta) when adding each augmentation
set compared to using only the original training set, broken
down by sound class. At the bottom of the plot we provide the
delta scores for all classes combined. We see that most classes
are affected positively by most augmentation types, but there
are some clear exceptions. In particular, the air conditioner
class is negatively affected by the DRC and BG augmentations.
Given that this sound class is characterized by a continuous
“hum” sound, often in the background, it makes sense that
the addition of background noise that can mask the presence
of this class will deteriorate the performance of the model.
In general, the pitch augmentations have the greatest positive
impact on performance, and are the only augmentation sets
that do not have a negative impact on any of the classes. Only
half of the classes benefit from applying all augmentations
combined more than they would from the application of a
subset of augmentations. This suggests that the performance
of the model could be improved further by the application
of class-conditional augmentation during training – one could
use the validation set to identify which augmentations improve
the model’s classification accuracy for each class, and then
selectively augment the training data accordingly. We intend
to explore this idea further in future work.
IV. C ONCLUSION
In this article we proposed a deep convolutional neural
network architecture which, in combination with a set of
audio data augmentations, produces state-of-the-art results
for environmental sound classification. We showed that the
improved performance stems from the combination of a deep,
high-capacity model and an augmented training set: this
combination outperformed both the proposed CNN without
augmentation and a “shallow” dictionary learning model with
augmentation. Finally, we examined the influence of each augmentation on the model’s classification accuracy. We observed
that the performance of the model for each sound class is
influenced differently by each augmentation set, suggesting
that the performance of the model could be improved further
by applying class-conditional data augmentation.
ACKNOWLEDGMENT
The authors would like to thank Brian McFee and Eric
Humphrey for their valuable feedback, and Karol Piczak for
providing details on the results reported in [11]. This work
was partially supported by NSF award 1544753.
IEEE SIGNAL PROCESSING LETTERS, ACCEPTED NOVEMBER 2016
R EFERENCES
[1] S. Chu, S. Narayanan, and C.-C. Kuo, “Environmental sound recognition
with time-frequency audio features,” IEEE Trans. on Audio, Speech, and
Language Processing, vol. 17, no. 6, pp. 1142–1158, Aug. 2009.
[2] R. Radhakrishnan, A. Divakaran, and P. Smaragdis, “Audio analysis
for surveillance applications,” in IEEE Worksh. on Apps. of Signal
Processing to Audio and Acoustics (WASPAA’05), New Paltz, NY, USA,
Oct. 2005, pp. 158–161.
[3] C. Mydlarz, J. Salamon, and J. P. Bello, “The implementation of lowcost urban acoustic monitoring devices,” Applied Acoustics, vol. In Press,
2016.
[4] A. Mesaros, T. Heittola, O. Dikmen, and T. Virtanen, “Sound event
detection in real life recordings using coupled matrix factorization of
spectral representations and class activity annotations,” in International
Conference on Acoustics, Speech and Signal Processing (ICASSP),
Brisbane, Australia, Apr. 2015, pp. 151–155.
[5] E. Benetos, G. Lafay, M. Lagrange, and M. D. Plumbley, “Detection of
overlapping acoustic events using a temporally-constrained probabilistic
model,” in IEEE International Conference on Acoustics, Speech and
Signal Processing (ICASSP), Shanghai, China, Mar. 2016, pp. 6450–
6454.
[6] V. Bisot, R. Serizel, S. Essid, and G. Richard, “Acoustic scene classification with matrix factorization for unsupervised feature learning,” in IEEE
International Conference on Acoustics, Speech and Signal Processing
(ICASSP), Shanghai, China, Mar. 2016, pp. 6445–6449.
[7] J. Salamon and J. P. Bello, “Unsupervised feature learning for urban
sound classification,” in IEEE Int. Conf. on Acoustics, Speech and Signal
Processing (ICASSP), Brisbane, Australia, Apr. 2015, pp. 171–175.
[8] ——, “Feature learning with deep scattering for urban sound analysis,”
in 2015 European Signal Processing Conference, Nice, France, Aug.
2015.
[9] J. T. Geiger and K. Helwani, “Improving event detection for audio
surveillance using gabor filterbank features,” in 23rd European Signal
Processing Conference (EUSIPCO), Nice, France, Aug. 2015, pp. 714–
718.
[10] E. Cakir, T. Heittola, H. Huttunen, and T. Virtanen, “Polyphonic sound
event detection using multi label deep neural networks,” in 2015
International Joint Conference on Neural Networks (IJCNN), July 2015,
pp. 1–7.
[11] K. J. Piczak, “Environmental sound classification with convolutional
neural networks,” in 25th International Workshop on Machine Learning
for Signal Processing (MLSP), Boston, MA, USA, Sep. 2015, pp. 1–6.
[12] D. Giannoulis, E. Benetos, D. Stowell, M. Rossignol, M. Lagrange, and
M. D. Plumbley, “Detection and classification of acoustic scenes and
events: An IEEE AASP challenge,” in IEEE Workshop on Applications
of Signal Processing to Audio and Acoustics (WASPAA), New Paltz, NY,
USA, Oct. 2013, pp. 1–4.
[13] D. Stowell, D. Giannoulis, E. Benetos, M. Lagrange, and M. D.
Plumbley, “Detection and classification of acoustic scenes and events,”
IEEE Transactions on Multimedia, vol. 17, no. 10, pp. 1733–1746, Oct.
2015.
[14] S. Sigtia, A. Stark, S. Krstulovic, and M. Plumbley, “Automatic environmental sound recognition: Performance versus computational cost,”
IEEE/ACM Transactions on Audio, Speech, and Language Processing,
vol. PP, no. 99, pp. 1–1, 2016.
[15] Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based learning
applied to document recognition,” Proceedings of the IEEE, vol. 86,
no. 11, pp. 2278–2324, Nov. 1998.
[16] C. V. Cotton and D. P. W. Ellis, “Spectral vs. spectro-temporal features
for acoustic event detection,” in IEEE Worksh. on Apps. of Signal
Processing to Audio and Acoustics (WASPAA), New Paltz, NY, USA,
Oct. 2011, pp. 69–72.
[17] J. Salamon, C. Jacoby, and J. P. Bello, “A dataset and taxonomy for urban
sound research,” in 22nd ACM International Conference on Multimedia
(ACM-MM’14), Orlando, FL, USA, Nov. 2014, pp. 1041–1044.
[18] K. J. Piczak, “ESC: Dataset for environmental sound classification,” in
23rd ACM International Conference on Multimedia, Brisbane, Australia,
Oct. 2015, pp. 1015–1018.
[19] A. Mesaros, E. Fagerlund, A. Hiltunen, T. Heittola, and T. Virtanen,
“TUT sound events 2016, development dataset,” Available online:
http://dx.doi.org/10.5281/zenodo.45759 (accessed 10 August 2016),
2016. [Online]. Available: http://dx.doi.org/10.5281/zenodo.45759
[20] A. Krizhevsky, I. Sutskever, and G. Hinton, “ImageNet classification
with deep convolutional neural networks,” in Advances in neural information processing systems (NIPS), 2012, pp. 1097–1105.
5
[21] P. Y. Simard, D. Steinkraus, and J. C. Platt, “Best practices for
convolutional neural networks applied to visual document analysis.” in
International Conference on Document Analysis and Recognition, vol. 3,
Edinburgh, Scottland, UK, Aug. 2003, pp. 958–962.
[22] B. McFee, E. Humphrey, and J. Bello, “A software framework for
musical data augmentation,” in 16th Int. Soc. for Music Info. Retrieval
Conf., Malaga, Spain, Oct. 2015, pp. 248–254.
[23] G. Parascandolo, H. Huttunen, and T. Virtanen, “Recurrent neural
networks for polyphonic sound event detection in real life recordings,”
in International Conference on Acoustics, Speech and Signal Processing
(ICASSP), Shanghai, China, Mar. 2016, pp. 6440–6444.
[24] D. Bogdanov, N. Wack, E. Gómez, S. Gulati, P. Herrera, O. Mayor,
G. Roma, J. Salamon, J. Zapata, and X. Serra, “ESSENTIA: an audio
analysis library for music information retrieval,” in 14th Int. Soc. for
Music Info. Retrieval Conf., Curitiba, Brazil, Nov. 2013, pp. 493–498.
[25] L. Bottou, “Large-scale machine learning with stochastic gradient
descent,” in 19th International Conference on Computational Statistics
(COMPSTAT), Paris, France, Aug. 2010, pp. 177–186. [Online].
Available: http://dx.doi.org/10.1007/978-3-7908-2604-3 16
[26] N. Srivastava, G. E. Hinton, A. Krizhevsky, I. Sutskever, and
R. Salakhutdinov, “Dropout: a simple way to prevent neural networks
from overfitting,” Journal of Machine Learning Research, vol. 15, no. 1,
pp. 1929–1958, 2014.
[27] S. Dieleman, J. Schlüter, C. Raffel, E. Olson, S. Sønderby, D. Nouri,
D. Maturana, M. Thoma, E. Battenberg, and J. Kelly, “Lasagne: First
release,” https://github.com/Lasagne/Lasagne, 2015. [Online]. Available:
http://dx.doi.org/10.5281/zenodo.27878
[28] B. McFee and E. J. Humphrey, “pescador: 0.1.0,” https://github.com/
bmcfee/pescador, 2015. [Online]. Available: http://dx.doi.org/10.5281/
zenodo.32468
[29] Dolby Labortories, Inc., “Standards and practices for authoring Dolby
Digital and Dolby E bitstreams,” 2002.
[30] “Icecast streaming media server forum,” (accessed 12 August 2016).
[Online]. Available: http://icecast.imux.net/viewtopic.php?t=3462
[31] E. J. Humphrey, J. Salamon, O. Nieto, J. Forsyth, R. Bittner, and J. P.
Bello, “JAMS: A JSON annotated music specification for reproducible
MIR research,” in 15th Int. Soc. for Music Info. Retrieval Conf., Taipei,
Taiwan, Oct. 2014, pp. 591–596.
[32] B. McFee, E. J. Humphrey, O. Nieto, J. Salamon, R. Bittner, J. Forsyth,
and J. P. Bello, “Pump up the JAMS: V0.2 and beyond,” Music and
Audio Research Laboratory, New York University, Tech. Rep., Oct.
2015.
| 9cs.NE |
Parallelizing Stochastic Approximation Through
Mini-Batching and Tail-Averaging
Prateek Jain∗, Sham M. Kakade†, Rahul Kidambi‡, Praneeth Netrapalli§, Aaron Sidford¶
arXiv:1610.03774v2 [stat.ML] 27 Oct 2016
Abstract
This work characterizes the benefits of averaging techniques widely used in conjunction with stochastic
gradient descent (SGD). In particular, this work sharply analyzes: (1) mini-batching, a method of averaging
many samples of the gradient to both reduce the variance of a stochastic gradient estimate and for parallelizing
SGD and (2) tail-averaging, a method involving averaging the final few iterates of SGD in order to decrease
the variance in SGD’s final iterate. This work presents the first tight non-asymptotic generalization error
bounds for these schemes for the stochastic approximation problem of least squares regression.
Furthermore, this work establishes a precise problem-dependent extent to which mini-batching can be
used to yield provable near-linear parallelization speedups over SGD with batch size one. These results are
utilized in providing a highly parallelizable SGD algorithm that obtains the optimal statistical error rate
with nearly the same number of serial updates as batch gradient descent, which improves significantly over
existing SGD-style methods.
Finally, this work sheds light on some fundamental differences in SGD’s behavior when dealing with
agnostic noise in the (non-realizable) least squares regression problem. In particular, the work shows that
the stepsizes that ensure optimal statistical error rates for the agnostic case must be a function of the noise
properties.
The central analysis tools used by this paper are obtained through generalizing the operator view of
averaged SGD, introduced by Défossez and Bach [1] followed by developing a novel analysis in bounding
these operators to characterize the generalization error. These techniques may be of broader interest in
analyzing various computational aspects of stochastic approximation.
1
Introduction and problem setup
With the ever increasing size of modern day datasets, practical algorithms for machine learning are increasingly
constrained to spend less time and use less memory. This makes it particularly desirable for simple streaming
algorithms that generalize well in just a few passes over the dataset. Stochastic gradient descent (SGD) is
perhaps the simplest and most well studied algorithm that meets these constraints. The algorithm simply
repeatedly samples an instance from the stream of data and updates the current parameter estimate using
the gradient of the sampled instance. Nevertheless, SGD has been impressively successful and is the de-facto
method of choice for large scale learning problems.
While a powerful machine learning tool, unfortunately SGD in its simplest forms is inherently serial.
Over the past years as dataset sizes have grown there have been remarkable developments in processing
capabilities, with multi-core/distributed/GPU based computing infrastructure available in relative abundance.
The presence of this computing power has triggered the development of parallel/distributed machine learning
algorithms [2, 3, 4, 5, 6, 7] that possess the capability to utilize multiple cores/machines. However, despite this
exciting work, it is yet unclear how to best parallelize SGD and fully utilize these computing infrastructures.
This paper takes a step towards answering this question, by characterizing the behavior of constant
stepsize SGD for the problem of strongly convex stochastic least square regression (LSR) under two averaging
techniques widely believed to improve the performance of SGD. In particular, this work considers the natural
∗ Microsoft
Research, India, e-mail: prajain@microsoft.com
of Washington, Seattle, e-mail: sham@cs.washington.edu
‡ University of Washington, Seattle, e-mail: rkidambi@uw.edu
§ Microsoft Research, India, e-mail: praneeth@microsoft.com
¶ Stanford University, Palo Alto CA, e-mail: sidford@stanford.edu
† University
1
parallelization technique of mini-batching, where multiple points are processed simultaneously and the current
iterate is updated by the average gradient over these points, and combine it with variance reducing technique
of tail-averaging, where the average of many of the final iterates are returned as SGD’s estimate of the solution.
In this work, the arguments about parallelization are structured through the lens of a work-depth tradeoff:
work refers to the total amount of computation required to reach a certain generalization error, and depth
refers to the amount of serial computation. Depth, defined in this manner, is equal to the runtime of the
algorithm on a large multi-core architecture with shared memory, where there is no communication overhead,
and has strong implications for parallelizability on other architectures as well.
The main contributions of this paper are as follows:
• This work shows that mini-batching yields near-linear parallelization speedups over the standard serial
SGD (i.e. with batch size 1), as long as the mini-batch size is smaller than a problem dependent quantity
(denoted by bthresh ). When batch-sizes increase beyond bthresh , mini-batching is inefficient (owing to the
lack of serial updates), thus obtaining only sub-linear speedups over and above mini-batching with a
batch size bthresh . A by-product of this analysis sheds light on how the step sizes naturally interpolate
from ones used by standard serial SGD (with batch size 1) to ones used by batch gradient descent.
• While the final iterate of SGD decays initial error at a geometric rate but does not obtain optimal
statistical error rates, the average iterate [8, 1] decays the initial error at a sublinear rate while
preserving statistical optimality. This work rigorously shows that tail-averaging obtains the best of both
worlds – decaying initial error at a geometric rate and obtaining near-optimal statistical error rates.
• Combining the above results, this paper provides a mini-batching and tail-averaging version of SGD that
is highly parallelizable – the number of serial steps (which is a proxy for the un-parallelizable time)
of this algorithm nearly matches that of offline gradient descent and is lower than the serial time of all
existing streaming LSR algorithms. See Table 1 for comparison. We note that these results are obtained
by providing a tight finite-sample analysis of the effects of mini-batching and tail-averaging with large
constant learning rate schemes.
• All the results in this paper are established for the general agnostic noise case of the streaming LSR
problem. This reveals a fundamental difference in the behavior of SGD when dealing with the agnostic
case as opposed to the realizable case. In particular, this analysis reveals a surprising insight that the
maximal stepsizes that are sufficient to obtain statistically optimal rates are a function of the properties
of the agnostic noise. The main result of this analysis is that the maximal step sizes for the agnostic case
could be much lower than ones that are used in the realizable case: indeed, an instance that yields such
a separation between the maximal learning rates for the realizable and agnostic case is provided.
The tool employed in obtaining these results is Algorithm 1 and its generalization guarantees as in
Theorem 1, which builds upon and generalizes the analysis of [1] and provides the first tight bound on the
finite-sample generalization error for mini-batching, tail-averaged constant step-size SGD. We note that the
work of [1] does not establish statistically optimal rates while working with large constant step size schemes;
this shortcoming is remedied through coming up with a novel sharp analysis that is used to rigorously establish
statistically optimal rates while working with large constant step sizes (this can be viewed as a special case of
Theorem 2 by setting batch size b = 1 and through averaging iterates from the start by setting s = 0).
Applying Theorem 1 to the general agnostic case of LSR, we obtain Theorems 2 and 3, which shows
an interesting trade-off between the convergence rate or generalization error (work) of SGD, the amount
of parallelization (depth of computation) achieved by increasing the mini-batch size, and the amount of
tail-averaging performed. As with previous SGD-style algorithms, [8, 9, 10, 11, 1, 12], this trade-off stems
from the two terms of the generalization error of SGD: (i) the bias term which represents how quickly the
initial conditions are forgotten, and (ii) the variance term which represents the dependence of the excess
risk on the level of noise (due to streaming data/random sampling/model mismatch) that is present in the
problem. We note that existing results [9, 1] dealing with averaged SGD with batch size 1 applied to the
realizable case of LSR generalize to this paper’s results which establishes (a) the finite-sample effects of
mini-batching, tail averaging in the context of stochastic approximation, and (b) the impact of agnostic noise
on the learning rates employed by SGD to guarantee statistically optimal error rates.
The results of this paper are obtained through building on and generalizing the operator view of averaged
SGD with batch size one, as introduced by [1] to the mini-batched tail-averaged case and through subsequently
2
Algorithm
Batch
gradient
descent
Stochastic
Dual
Co-ordinate
Ascent [13]
Averaged
SGD [1, 9]
Streaming
SVRG (with
initial error
oracle) [12]
Final error
2
O σnd
O
O
R4
λ2min n2
σ2 d
n
Runtime/Work
κnd log
R2 t
kHk2 n
κ log
n·init
σ2 d
2
(n+ λRmin d)d·log
2
· init + σnd
t
κ log(dκ)
σ2 d
n
σ2 d
n
Algorithm 2
O
Algorithm 2
with initial
error oracle
min (H)
· init +
O exp − Rnλ
2 ·log(dκ)
· init +
n·init
σ2 d
2
(n+ λRmin d)·log
nd
(H)
· init +
O exp − nλmin
R2
n·init
σ2 d
Depth
nd
n
2
( λminR(H) ) · log
σ2 d
n
nd
nd
n·init
σ2 d
t log
n·init
σ2 d
nkHk2
tR2
κ log(dκ) log
n·init
σ2 d
Streaming
Agnostic
×
X
×
X
X
×
X
X
X
X
X
X
Table 1: Comparison of Algorithm 2 with existing algorithms given n samples for LSR, with init = R(w0 ) − R(w∗ ). Note that
the bounds have been provided assuming realizable noise; for the algorithms with support for agnostic noise, these can be
appropriately modified. Refer to Section 3 for the definitions of all quantities. We do not consider accelerated variants in this
table. Note that the accelerated variants serve to improve running times of the offline algorithms. Furthermore, offline algorithms
possess the advantage that they do not require the knowledge of R2 . The quantity t in the results for Algorithm 2 can be chosen
arbitrarily and trades off the rate of decay in initial error with the depth of the algorithm. Finally, we note that streaming
SVRG [12] does not conform to the standard first order oracle model [14].
providing a new sharp analysis in order to bound the resulting operators. In particular, straightforward
operator norm bounds of these matrix operators suffices to show convergence of SGD updates, but turn out to
be pretty loose bounds. In order to obtain stronger bounds, this paper employs a much more fine grained
analysis that directly bounds the trace of these operators applied to the relevant matrices, which results in
obtaining much tighter bounds on the performance of mini-batched tail-averaged SGD while simultaneously
shedding light on the role of agnostic noise on the behavior of SGD.
While this paper’s results focus on the case of strongly convex streaming least square regression, we believe
that our techniques extend much more broadly. We hope that this paper will serve as the basis for future work
on analyzing SGD and improving the parallelization of large scale algorithms for machine learning.
Paper organization: A brief overview of related works is presented in Section 2. Then, the problem setup
and notations are introduced in Section 3. The main results of this paper as well as the mini-batched tail
averaging SGD algorithm are presented in Section 5. In Section 6, simulation results are presented using
a synthetic dataset, and these results corroborate well with the main results on the mini-batch thresholds
established by the paper as well as the advantages of mini-batch tail averaged SGD algorithm (i.e. algorithm 1).
The proofs of all the claims and theorems are provided in the appendix.
2
Related work
We split the work related to this paper’s contributions into several algorithm classes:
Stochastic Approximation: Seminal works of Ruppert [15], and Polyak and Juditsky [8] showed that
averaging iterates of SGD applied to stochastic optimization problems like stochastic least squares regression
achieves asymptotically optimal statistical rates. A more recent line of work [16, 9, 10, 11, 17, 18, 1] has
provided rigorous non-asymptotic analysis (with explicit constants) of SGD for stochastic approximation
problems including linear and logistic regression. Equipped with averaging, these algorithms achieve (up to
constants) minimax optimal statistical rates. Note that all these results assume that the noise is realizable when
stating the generalization error bounds. We are unaware of any results that provide sharp non-asymptotic
analysis of SGD and the related step size issues in the general agnostic noise setting. Several specialized
3
stochastic algorithms such as SVRG [19], SDCA [13] and SAGA [20] have been proposed for the agnostic
setting, among which, only the streaming version of SVRG [12] is known to provide minimax optimal
statistical rates and is highly parallelizable. Note that offline SVRG [19] is parallelizable as well, but is not
streaming and incurs additional log factors without carefully employing doubling tricks. Note that streaming
SVRG [12] does not take advantage of minibatching and its depth depends on a stronger fourth moment
condition on the input whereas our mini-batching bounds depend only on the condition number of the
covariance matrix. Moreover, our constant-step size SGD algorithm is arguably simpler than streaming SVRG
and can be run in a more general setting requiring less information from a single sample. In particular, SGD
based algorithms conform to the standard first order oracle model [14], whereas the streaming SVRG [12]
does not, since it requires gradient information at two different points from a single sample.
Another key aspect of this paper’s main result is a precise characterization of the effects of mini-batching,
which is also studied extensively in this literature [21, 22, 23, 6, 24]. Previous results on mini-batching within
the context of stochastic approximation yielded results that suggest increasing minibatch sizes tend to hurt
the convergence rate [22, 21]. [6, 24, 21, 23] study mini-batching for various algorithms and show faster
decay in the bias term but do not show optimal statistical error rates. Moreover, all of these results claim
speed-up from mini-batching by comparing the upper bounds on the training/generalization error. In contrast,
this paper’s result for mini-batching compares the exact generalization error for different batch sizes and
demonstrates an improvement with larger sizes.
The other scheme for parallelizing stochastic approximation is through model averaging [2, 3, 25, 26].
These schemes are communication efficient since they average the final output of independent runs of the
algorithm, while being minimax optimal. However, as noted in [26], these schemes do not yield improvements
to the bias and it is unclear how these results relate to this work’s depth bounds. Other efforts [25, 2, 3]
provide guarantees on upperbounds of the error, with only [25] guaranteeing minimax optimality.
Non-Stochastic Approximation: In order to utilize the recent developments in computing frameworks,
there has been a number of efforts in developing distributed/parallel machine learning algorithms that utilize
multi-core [4, 5, 21, 22, 27, 24] and multi-machine [2, 3, 6, 23, 7] computing frameworks. As in the stochastic
case, parallelization speedups are argued only in terms of upper bounds of training/generalization error. On
the other hand, a distinctive feature of this paper’s mini-batching result is that the prescribed parallelization
speedups are on the exact generalization error on a per problem basis, and do not rely on providing linear
speedups on weaker upper bounds. Moreover, none of the existing parallelization results characterize regimes
where there are unconditional linear speedups on the overall generalization error while maintaining minimax
optimal statistical error rates. Indeed, model averaging results are known to be minimax optimal [25, 26],
however, the bias term is unimprovable using these schemes, as noted by [26].
Asynchronous Stochastic Approximation: There exists a recent line of work [5, 28, 29] that provides
linear speedups of SGD in the context of asynchronous stochastic optimization. However, in contrast to
our work, these efforts deal with the much harder case of asynchronous optimization, and moreover, these
results provide speedups on upper bounds on the training/generalization error while relying on specific input
structures such as hard sparsity. Moreover, while [29] proves statistical optimality in an asymptotic sense,
none of the other efforts claim statistical optimality.
3
Preliminaries
Problem formulation: Let D be a distribution on tuples (x, y) ∈ Rd × R. Let R : Rd → R be defined as the
expected square loss over the pairs (x, y):
R(w) =
1
2
E(x,y)∼D (y − hw, xi) ∀ w ∈ Rd .
2
(1)
Our goal is to compute the predictor w∗ = arg minw R(w). In this paper, we provide the maiden sharpest
analysis of SGD with batch sizes ≥ 1, applied to the general agnostic case of LSR (also referred to as the
def
unrealizable case) where y = hw∗ , xi + where ∈ R is the noise.
Notation: We use boldface small letters (x, w etc.) for vectors and boldface capital letters (A, H etc.)
for matrices. We use ⊗ to denote the outer product of two vectors or matrices. S(d) denotes the space of
real symmetric d × d matrices and M(d) denotes the space of linear transformations from Rd×d to Rd×d .
Every element in M(d) is a d × d × d × d tensor, which is represented using normal font scripted letters
4
Algorithm 1 Minibatch-TailAveraging-SGD
Input: Initial point w0 , stepsize γ, minibatch size b, initial iterations s, total number of samples m.
1: for t = 1, 2, .., s do
2:
Sample “b” tuples {(xti , yti )}bi=1 ∼ Db
Pb
3:
wt ← wt−1 − γb i=1 (hwt−1 , xti i − yti )xti
4: w̄ = 0
5: for t = s + 1, s + 2, .., m
b do
6:
Sample “b” tuples {(xti , yti )}bi=1 ∼ Db
Pb
7:
wt ← wt−1 − γb i=1 (hwt−1 , xti i − yti )xti
1
8:
w̄ = t−s
wt + t−s−1
t−s w̄
Output: w̄
(M, T etc.).PA tensor M applied to a matrix A gives a matrix MA such that its ij th element is given by
(MA)ij = k,`∈[d] Mik`j Ak` . Any d × d matrix A can be lifted to AL , AR ∈ M(d), which represents the
left/right multiplication operator of the matrix A respectively. More concretely, AL , AR ∈ M(d) are defined
so that AL W = AW and AR W = WA ∀ W ∈ Rd×d . We let I ∈ M(d) represent the identity map.
We let k·k or k·k2 denote `2 norm for vectors and spectral or operator norm for matrices and tensors. More
concretely, we define
def
kAk = max
x∈Rd
kT AkF
kAxk
def
∀ T ∈ M(d),
∀ A ∈ Rd×d , and kT k = max
kxk
A∈S(d) kAkF
where k·kF denotes the Frobenius norm of a matrix. When we refer to the largest/smallest eigenvalues
λmax (S) or λmin (S) of a tensor S, this is obtained by viewing the operator S as a member of M(d).
Relevant quantities: We denote by H and M the population second moment matrix and fourth moment
tensor of the covariates x, i.e.:
def
def
H = E(x,y)∼D [x ⊗ x], M = E(x,y)∼D [x ⊗ x ⊗ x ⊗ x]
We assume
that both H and M are finite and are related by MI R2 H. This in particular implies that
2
E kxk ≤ R2 . We further assume that H is non-singular, with λmin (H) > 0 denoting the smallest eigenvalue
of H. This makes R(w) strongly convex and implies the existence of a unique minimizer w∗ of the objective
R(w). The condition number κ of the population covariance matrix H is defined to be the ratio of the largest
def λmax (H)
λmin (H) .
to smallest eigenvalues of H, i.e., κ =
Next, we define the noise covariance matrix Σ as:
def
Σ = E(x,y)∼D (y − hw∗ , xi)2 · x ⊗ x
In the realizable case, the noise is independent of the covariate x, thus implying Σ = σ 2 H.
Note that M
can also be thought of as an operator in M(d) which acts on matrix A to produce MA = E (x> Ax)x ⊗ x .
For a given γ > 0 and b ∈ Z+ , a key quantity in our results turns out to be the operator Tbγ defined as
def
Tbγ = HL + HR − γb M −
Tbγ .
γ(b−1)
HL HR .
b
In most cases we drop γ for notational convenience and use Tb for
def
We let T = HL + HR − γHL HR denote the limit of Tb as b → ∞. For a given γ > 0, we also use G to
def
denote the operator G = HL −1 + HR −1 − γI.
Assumptions: For the realizable case, the generalization error estimate in Theorem 1 is exact on a
per-problem basis if E [yi |xi ] = hw∗ , xi i (which is true if we have a constant component
inx, as is typically
the case in practice). For the agnostic case, we note that this result is exact if E x(i) x(j) x(k) = 0 ∀ (i, j, k) ∈
⊗3 {1, 2, .., d} with x(i) referring to the ith dimension of x. If this is not the case, the generalization error in
Theorem 1 are off the exact value by a factor of at most 2 (as noted by [9, 1], owing to the application of
Minkowski’s inequality).
4
Viewing Stochastic Gradient Descent as a Stochastic Process
In this section, we consider a stochastic process view of Stochastic Gradient Descent updates and shed some
light on why this view allows us to provide a sharp characterization of SGD as a tool to solve the general LSR
5
Algorithm 2 MinibatchDoublingPartialAveragingSGD
Input: Initial point w0 , stepsize γ, initial minibatch size b, number of iterations in each epoch s, number of
samples n.
1: /*Run logarithmic number of epochs where each epoch runs t iterations of minibatch SGD (with out
averaging). Double minibatch size after each epoch.*/
n
2: for ` = 1, 2, · · · , log bt
− 1 do
`−1
3:
b` ← 2 b
4:
w` ← Minibatch-TailAveraging-SGD(w`−1 , γ, b` , t − 1, t · b` )
5: /*For the last epoch, run tail averaged minibatch SGD with initial point wt , stepsize γ, minibatch size
n
2log bt −1 · b = n/2t, number of initial iterations t/2 and number of samples n/2.*/
6: w ← Minibatch-TailAveraging-SGD(ws , γ, n/2t, t/2, n/2)
Output: w
problem. In particular, let us consider the mini-batch SGD update that moves iterate wt−1 to wt :
wt = wt−1 −
b
γX
(hwt−1 , xti i − yti )xti
b i=1
def
By defining ηt = wt − w∗ , we write the recursion relating ηt to ηt−1 as:
γX
γX
xti ⊗ xti ηt−1 +
ti xti
ηt = I −
b i
b i
Now, since we deal with the square loss case, the generalization error is written as:
E [R(wt )] − R(w∗ ) =
1
Tr H E [ηt ⊗ ηt ]
2
We thus consider E [ηt ⊗ ηt ], which can be written in terms of E [ηt−1 ⊗ ηt−1 ] as:
E [ηt ⊗ ηt ] = (I − γTb ) E [ηt−1 ⊗ ηt−1 ] + γ 2 Σ
Divergent Learning Rates: This operator view of mini-batch SGD immediately sheds light on the maximal
div
stepsizes γb,max
that prevents divergence of the resulting mini-batch SGD updates: in particular, a necessary
condition that the set of non-divergent learning rates γ ≥ 0 should satisfy is Tbγ 0. This condition yields the
div
following generalized eigenvector problem for computing γb,max
:
2
div
γb,max
=
hW, MWi + (b − 1) · Tr WHWH
b · Tr WHW
W∈S(d)
sup
This characterization generalizes the divergent stepsize characterization of [1] for batch sizes > 1. We note
that such a characterization automatically sheds light on how the maximal divergent learning rates interpolate
batch sizes of 1 (which is ≤ Tr2H [1]) to the batch gradient descent learning rate (setting b to ∞), which
div
turns out to be λmax2(H) . A property of γb,max
that is worth noting is that it does not depend on properties of
the noise covariance Σ, and depends solely on the statistics of the covariate x.
div
We note that in this paper, our interest does not lie in the non-divergent stepsizes 0 ≤ γ ≤ γb,max
, but in
div
the set of (maximal) step sizes 0 ≤ γ ≤ γb,max (< γb,max ) that are sufficient to guarantee statistically optimal
error rates of O( n1 Tr H−1 Σ). We show that our characterization of γb,max is indeed tight in the sense that
div
there are problem instances, where γb,max and γb,max
are off by a factor of d and γb,max is still the largest step
size which provides statistically optimal error rates. A critical property of γb,max is that it is dependent on the
properties of the noise covariance Σ. We return to the characterization of γb,max in section 5.4.
5
Main Results
This section presents the main Algorithms analyzed by this paper and their associated guarantees. In
particular, the standard mini-batch tail-averaged SGD Algorithm for the streaming LSR problem is presented
6
as Algorithm 1. Theorem 1 presents an exact characterization of the error achieved by Algorithm 1. Theorem 2
establishes a scalar estimate of the generalization error while dealing with the general agnostic case of LSR,
in order to better illustrate the performance of Algorithm 1. Theorem 2 shows that increasing minibatch sizes
(up to some threshold) and tail averaging achieves the same work (number of first order oracle calls to get
gradient at a point) and statistical error performance as the standard averaged SGD but decreases the depth
significantly, thereby increasing its parallelizability.
Theorem 3 presents the guarantees of Algorithm 2, which is a doubling based Algorithm that decays
the initial error at a superpolynomial rate (and a geometric rate in the presence of an initial error oracle)
while performing the same work and achieving same statistical accuracy (up to constant factors) compared to
standard sequential SGD (with batch size 1). Algorithm 2’s depth nearly matches the depth of standard offline
gradient descent, while providing the same generalization error (upto constants) as offline gradient descent
on every problem instance, while being a single pass streaming Algorithm, thus performing substantially
lower work compared to the standard offline gradient descent. Refer to table 1 for the associated bounds.
The paper’s final result is lemma 4, which presents an instance that shows learning rates necessary to retain
statistical optimality in the agnostic case could be much lower than ones required by realizable case.
5.1
Characterization of SGD with minibatching and tail averaging
This section presents Theorem 1, which provides an exact characterization of the error incurred by minibatched tail-averaged SGD, as described in Algorithm 1, working with a mini-batch size b and step size
γ. Algorithm 1 consists of two phases. The first phase (lines 1-3) is the burn-in phase that consists of “s”
iterations where mini-batch SGD updates are performed without averaging the iterates. The second phase
(lines 5-8) is the tail-averaging phase, where, iterates of mini-batch SGD are averaged. Intuitively, the burn-in
phase is used to decrease the bias which is dependent on initial error kw0 − w∗ k and the averaging phase is
to decrease the variance. Note that the bias exhibits a geometric decay during the unaveraged phase.
−1
def
Σ
2b
R)
, 2b Tr (HL +H
Theorem 1. For all b ≥ 1, let γb,max be defined as γb,max = min{ (b−1)kHk
−1 Σk R2 d }. Suppose we
2 k(HL +HR )
2
run Algorithm 1 with initial point w0 , stepsize 0 < γ < γb,max , minibatch size b, initial iterations s, total number
of samples n, and ∃ universal constants C1 , C2 such that the final error w − w∗ satisfies:
i
h
>
E (w − w∗ )(w − w∗ ) =
n/b
ρb
1
kE0 kF · M1
GTb −1 (I − γTb )s E0 + C1 ·
− s)2 γ 2
γλmin (H) nb − s
{z
} |
{z
}
( nb
|
"
+
1
I−
n − sb
γ
|
T1
T2
#
1
(HL −1 + HR −1 ) GTb −1 Σ
−s
{z
}
n
b
T3
n
2
+
γb
|
def
n
b
−1
2 HL HR
−s
{z
−1
Tb
−1
−s+1
ρbb
kΣkF · M2 ,
Σ + C2 ·
(n − bs)λmin (H)λmin (Tb )
{z
}
} |
(2)
T5
T4
>
def
where E0 = (w0 − w∗ )(w0 − w∗ ) and ρb = max(kI − γHk , kI − γTb k) < 1 and M1 , M2 are matrices
whose frobenius norm is less than or equal to 1.
Under the fourth moment assumption on the covariates MI R2 H, the stepsize γb,max is the maximal
step size we can use with SGD for the general agnostic LSR problem while being able to guarantee minimax
optimal statistical error rates. We return to this step size characterization after presenting this paper’s main
results. Lemma 7 (in the appendix) provides details regarding why this learning rate is a maximal; briefly put,
this stepsize limit characterizes the regime where the operators that govern the SGD update rules possess
favorable contractive properties while guaranteeing minimax optimal statistical rates.
Though (2) looks quite complicated, each of the terms in the equation has a very intuitive conceptual
meaning. The terms dependent on E0 (i.e., T1 and T2 ) are the bias terms while those depending on Σ (i.e.,
T3 , T4 and T5 ) are the variance terms. We observe the geometric decay of the bias terms T1 , T2 during the “s”
initial iterations when the iterates are not averaged. The term T3 (which is the leading order variance term)
1
contains a coefficient n−sb
, which is the number of samples that have been used in the averaging phase.
7
5.2
Benefits of minibatching and tail-averaging
This section presents a means to combine minibatching and tail averaging in order to obtain better performance
than standard averaged SGD for the general agnostic case of LSR. We note that these results naturally
interpolate to results for the realizable case by setting Σ = σ 2 H. Tail averaging allows a burn-in phase
when iterates are not averaged, thereby decaying the initial error geometrically during this phase, while
minibatching decreases the depth (i.e. enables parallelizing SGD) while performing the same amount of work
(up to constants). By the same amount of work, we mean that after using the same number of samples n, the
generalization error is the same as SGD with a batch-size 1 (up to constant factors): i.e., we obtain near linear
parallelization speedups via mini-batching.
Theorem 2. Consider the general agnostic case of LSR in the setting of Theorem 1. If a minibatch size
2
−1
def R dk(HL +HR ) Σk
γ
1
≤ b,max
of b ≥ bthresh = kHk Tr(HL +HR )−1 Σ2 and stepsize γ = 3kHk
is chosen, we obtain the following
6
2
2
generalization error bound on the performance of Algorithm 1:
h
i
>
R(w) − R(w∗ ) = Tr HE (w − w∗ )(w − w∗ )
√
s
n 3κ2 √d
36 dκ2
∗
{R(w0 ) − R(w )} + exp −
≤
{R(w0 ) − R(w∗ )}
2 exp −
n
3κ
3bκ nb − s
−
s
{z
}
{z
} |
|b
T1
+
T2
√
3 Tr H−1 Σ 9κ Tr H−1 Σ 2 dκ2 Tr H−1 Σ
+
exp −
2 +
b( nb − s)
b( nb − s)
b nb − s
{z
} |
{z
|
{z
} |
T3
( nb
− s + 1)
3κ
T5
T4
.
(3)
}
Tail averaging enables the bias terms T1 and T2 to decay exponentially with number of initial iterations
s, which is burn-in phase when iterates of SGD are not averaged. Compared to standard SGD which reduces
the initial error
R(w0 ) as O( s12 ) [1], Algorithm 1 reduces the dependence on R(w0 ) − R(w∗ ) geometrically
s
as exp − κ . We note that this contraction factor is reminiscent of the contraction factor that is observed
with deterministic batch gradient descent. This paper’s analysis of minibatching thus characterizes how the
behavior of mini-batch SGD smoothly
from vanilla
SGD (with a batch size 1) to batch gradient
interpolates
R2 dk(HL +HR )−1 Σk
2
descent with mini-batch sizes of O kHk Tr((HL +HR )−1 Σ) .
2
Minibatching with a batchsize b = bthresh allows us reduce the depth (serial work) while achieving (up to
−1
constants) the statistically optimal rate of Tr Hn Σ . Note that while standard averaged SGD has a depth of n,
Algorithm 1 has a depth of n/b. We note that this reduction in depth does not incur an extra overhead in terms
of work (the extra work overhead implies that the Algorithm is wasteful of samples, or in other words, we no
longer achieve linear parallelization speedups as a function of increasing the mini-batch size). We refer the
reader to Table 2, which compares the work, depth and accuracy achieved by minibatching and tail-averaging
R2 dk(HL +HR )−1 Σk
as compared to averaged SGD with n samples where minibatch size b = bthresh = kHk Tr(HL +HR )−1 Σ2 and
2
n
number of initial steps s = 2b
.
5.3
(Nearly) Matching the depth of Batch Gradient Descent
The result of section 5.2 establishes a scalar generalization error bound of Algorithm 1 for the general agnostic
case of LSR and showed that the depth (number of sequential updates in our algorithm) is decreased to n/b.
This section builds upon this result to present a simple and intuitive doubling based streaming algorithm that
works in epochs and processes a total of n/2 points. In each epoch, the minibatch size is increased by a factor
of 2 while thus applying Algorithm 1 with twice as many samples as the previous epoch. After running over
n/2 samples using this epoch based Algorithm, we run Algorithm 1 with the remaining n/2 points. Intuitively
each of these epochs decreases the bias linearly and then resets the bias of the next epoch as the total risk of
previous epoch, and then again decays the bias. The final application of the tail-averaging algorithm is to
ensure that the variance is small.
8
Final error
Sequential
averaged SGD
(b = 1)
MinibatchTailAveragingSGD (Algorithm 1)
O
O
√
√
dR4
λ2min n2
dR4
{R(w0 )
λ2min n2
− R(w∗ )} +
σ2 d
n
Runtime
Depth
O (nd)
n
O (nd)
n
bthresh
min
exp − nλ
{R(w0 ) − R(w∗ )} +
R2
σ2 d
n
Table 2: Comparison of minibatching and tail averaging in Algorithm 1 with with averaged SGD, both using
n samples (assuming realizable noise; the agnostic noise bound for Algorithm 1 can be derived similarly).
R2 dk(HL +HR )−1 Σk
1
n
R2
Algorithm 1 uses minibatch size bthresh = kHk Tr(HL +HR )−1 Σ2 = kHk
, γ = 3kHk
and s = 2b
while averaged
2
SGD uses the n samples sequentially. We assume that n >
make any progress [12].
R2
λmin
2
2
which is the minimum number of samples to
The next theorem formalizes
this intuition and shows that Algorithm 2 indeed improves the depth
∗
0 )−R(w )}
) in the presence of an error oracle that
exponentially from n/bthresh to O κ log(dκ) log( n{R(w
Tr H−1 Σ
provides us with the initial excess risk R(w0 ) − R(w∗ ) and the noise level Tr H−1 Σ.
Theorem 3. Consider the general Agnostic case of LSR. Suppose in MinibatchDoublingPartialAveragingSGD (AlR2 dk(HL +HR )−1 Σk
γ
1
gorithm 2), we use initial minibatchsize of b = kHk Tr(HL +HR )−1 Σ2 , stepsize γ = 3kHk
≤ b,max
and number of
6
2
2
iterations in each epoch being t ≥ 16κ log(dκ). We obtain the following bound on the excess risk of w:
E [R(w)] − R(w∗ ) ≤
2bt
n
t
3κ log(dκ)
{R(w0 ) − R(w∗ )} + 40
Tr H−1 Σ
.
n
Remarks: The final error again has two parts – the bias term that depends on the initial error R(w0 ) and
the variance term that depends on the statistical noise Tr H−1 Σ.
Algorithm 2 decays the bias at a superpolynomial rate by choosing t large enough. If Algorithm 2 has
access to an initial error oracle that provides R(w0 ) − R(w∗ ), we can run Algorithm 2 with a batch size bthresh
until the excess risk drops to the noise level Tr H−1 Σ and subsequently begin doubling the batch size. Such
an algorithm indeed gives geometric convergence with a generalization error bound as:
Tr(HL + HR )−1 Σ
Tr H−1 Σ
nλmin
E [R(w)] − R(w∗ ) ≤ exp −( 2
)·
{R(w0 ) − R(w∗ )} + 40
,
−1
R · log(dκ) d k(HL + HR ) Σk2
n
∗
0 )−R(w )}
with a depth of O κ log(dκ) log n{R(w
. The proof of this claim follows relatively straightforwardly
Tr H−1 Σ
from the proof of Theorem 3. We note that this depth nearly matches (up to log factors), the depth of standard
offline gradient descent. Our final observation is that the statistical error decays at a rate of O Tr H−1 Σ/n
which is statistically optimal up to constant factors.
5.4
Tradeoffs between mini-batch sizes, learning rates and the role of agnostic noise
While existing sharp non-asymptotic analyses of SGD with large learning rates [9, 10, 17, 1] focus on the
realizable case (i.e. the case with independent noise, so that Σ = σ 2 H), this paper provides results in the
agnostic noise case of LSR that point towards fundamental differences in the behavior of SGD between the
realizable and agnostic noise setting. Moreover, this paper’s results in general agnostic noise case naturally
specialize to existing results in the realizable case with batch size 1. Before discussing these differences, we
first understand mini-batching and related learning rate issues better. We start by writing out the maximal
learning rate that permits statistical optimality:
γb,max = min
n
2b
2b Tr (HL + HR )−1 Σ o
,
(b − 1) kHk2 k(HL + HR )−1 Σk2 R2 d
9
We note that the maximal learning rates γb,max tends to increase linearly as a function of the batch size b, until
R2 dk(HL +HR )−1 Σk
the point that the batch sizes hit a problem dependent limit bthresh = kHk Tr(HL +HR )−1 Σ2 ; this is precisely the
2
range of mini-batch sizes that yield a provably near-linear parallelization speedup over standard SGD with
batch size 1, while retaining statistical optimality. Moreover, when the batch sizes hit bthresh , the first term in
the definition of γb,max becomes active, and we begin using learning rates that mimic batch gradient descent,
i.e., with γb,max = O( λmax1(H) ). Beyond this mini-batching threshold bthresh , mini-batching is inefficient in that
while the variance term still improves as the batch size increases, the rate of decay of the bias does not show
a linear improvement as a function of the batch size (owing to the lack of serial updates), thus leading to
work inefficiency.
2b
Next, we note that in the realizable case, γb,max = min{ (b−1)kHk
, R2b2 }. This fact, coupled with the case of
2
SGD with batch size 1 (as considered in [9, 10, 17]) provides γ1,max = R22 , which matches the best learning
rates known to provide statistical optimality. The corresponding mini-batching threshold for the realizable
R2
case is bthresh = kHk
.
2
This paper’s final result deals with shedding light on some fundamental differences in the behavior of
SGD when dealing with the realizable case as opposed to the agnostic noise case. In particular, the following
lemma presents an instance working with agnostic noise case, wherein, SGD cannot employ the large learning
rates that typically are used in the realizable case while retaining statistical optimality; the proof of this lemma
can be found in the appendix. Note that this behavior is in stark contrast to algorithms such as streaming
SVRG [12], which work with the same large learning rates in the agnostic case as in the realizable case, while
guaranteeing statistically optimal rates.
Lemma 4. Consider a Streaming LSR example with Gaussian covariates with a diagonal second moment matrix
defined by:
(
1
if i = 1
Hii =
1/d
if i > 1
Further, let the noise covariance matrix Σ be diagonal as well, with the following entries:
(
1
if i = 1
Σii =
1/(d − 1)d
if i > 1
For this problem instance, γ1,max ≤
4
1
(d+2)(1+ d
)
is necessary for retaining statistical optimality, while the realizable
version of this problem permits a maximal learning rate ≤
learning rates between the realizable and agnostic case.
d
1 ,
(d+2)(1+ d
)
thus implying an O(d) separation in
The proof of lemma 4 can be found in the appendix in section A.4. As a final remark, note that in the
realizable case, we are unaware of whether there is a separation between O( R12 ) learning rate and the
divergent step size schema as introduced by [1]. Furthermore, it is unclear whether it is possible if we can
prove statistical optimality when working with step sizes that are larger than O( R12 ). However, we believe this
question is relatively of less consequence given the separation that exists between the realizable and agnostic
case, where a much smaller learning rate is necessary for obtaining statistically optimal rates.
6
Experimental simulations
We conduct experiments using a synthetic example to illustrate the implications of our theoretical results. The
data is sampled from a 50− dimensional Gaussian with eigenvalues decaying as { k1 }50
k=1 (condition number
κ = 50), and the variance σ 2 of the noise is 0.01. In this case, our estimated batch size according to Theorem 2
is bthresh = 11. In all our plots, the x-axis is the log of number of iterations (at each iteration, we use “b”
samples in the SGD update rule), and y-axis is the log of the generalization error. We assume that each
iteration takes constant time for all batch sizes; this is just in order to present evidence regarding the tightness
of our mini-batching characterization of the regime of mini-batch sizes where we obtain problem dependent
linear speedups while achieving the same generalization error as SGD with mini-batch size of 1.
We first consider the effect of mini-batching with batch sizes of 1, 4, bthresh = 11, 2 ∗ bthresh = 22 and d = 50.
In figure 1, we observe the generalization error curves obtained by averaging 100 runs of the Algorithm, each
10
(a) Bias term
(b) Variance term
(c) Total Risk
Figure 1: Effect of increasing batch sizes on the Algorithm’s generalization error. As expected, variance
decreases monotonically with increasing batch size. But the bias term shows that the rate of decay increases
till the optimal bthresh . Total risk shows that with optimal batch size, our method indeed obtains the optimal
statistical rate but using smaller number of iterations compared to sub-optimal batch sizes.
run utilizing the same set of samples for all mini-batch sizes. Averaging begins after observing a fixed number
of samples (set as 5κ). We see that the rate of bias decay (figure 1a) increases until reaching a mini-batch size
of bthresh , saturating thereafter; this implies we are wasteful in terms of sample size. As expected, the rate of
decay of variance (figure 1b) is monotonic as a function of mini-batch size. Finally, the overall generalization
error (figure 1c) shows the tightness of our mini-batching characterization. This is because, with a batch size
of bthresh , we obtain a generalization error that is the same as using batch size of 1 with the number of (serial)
iterations that is an order of magnitude smaller. Subsequently, we note that larger mini-batch sizes tend to
worsen generalization error thus depicting the tightness of our characterization of bthresh .
In the next experiment, we fix the batch size (as bthresh ) as well as the number of iterations that we
run mini-batch SGD, and thus hold the total number of samples as 10κ. We then consider the effect of
when averaging begins; in particular, we consider averaging iterates right from the start, as prescribed by
Dèfossez and Bach [1]. We also consider averaging iterates after a quarter/half/three-fourth of total number
of iterations, and consider unaveraged SGD as well. In figure 2, we see that the bias (figure 2a) exhibits
a geometric decay in the unaveraged phase while switching to an inferior O( t12 ) rate with averaging. The
σ2
variance (figure 2b) tends to increase and stabilize at O( bthresh
) in the absence of averaging, while switching to
1
a O( N ) decay rate when averaging begins. The overall generalization error (figure 2c) shows the superiority
of the scheme where averaging after a burn-in period allows us decay the bias towards the noise level at a
geometric rate, following which tail-averaging allows us to decay the variance term, thus providing credence
to our theoretical results that averaging after a burn-in time allows us to obtain better generalization error as
a function of sample size.
Finally, we consider the effect of step size choice; in particular, we fix the number of samples to be 10κ
and begin averaging after observing 2κ samples. We employ a mini-batch size of bthresh = 11 and compute
the corresponding γb,max . We then run mini-batch SGD with a step size γ = c · γbthresh ,max with c varied
as 0.25/0.5/0.75/1/1.1. The results, as observed in figure 3 presents the effect of increasing stepsizes. In
particular, for the bias term (figure 3a), the bias is decays progressively faster during the unaveraged phase as
the step size is increased, until we hit γb,max , where the behavior tends to be composed of a combination of
divergent and non-divergent steps. The trend is very similar with regard to the behavior of the variance term
(figure 3b) and the total risk (figure 3c).
7
Concluding remarks
The framework established as a part of Theorem 1 can be extended in a very straightforward manner to provide
sharp finite sample generalization error bounds for parameter mixing/model averaging schemes [2, 3, 25, 26].
Note that parameter mixing/model averaging is a parallelization scheme that can be used in conjunction with
mini-batching and tail-averaging. Furthermore, as noted by [26], these schemes do not offer any improvement
on the bias part of the error, while improving the variance term owing to the effect of averaging: this suggests
11
(a) Bias term
(b) Variance term
(c) Total Risk
Figure 2: Effect of tail-averaging with mini-batch size of bthresh = 11.
(a) Bias term
(b) Variance term
(c) Total Risk
Figure 3: Effect of increasing step sizes, i.e. with γ = c · γbthresh ,max with c = 0.25/0.5/0.75/1/1.1.
the effectiveness of these schemes when the bias error is roughly at the noise level.
While this paper provided the first rigorous analysis of the effect of mini-batching with SGD for the general
agnostic case of the streaming LSR problem, we believe the effect of mini-batching holds in much more
generality, with corresponding algorithm/problem dependent thresholds on the mini-batch size. In particular,
possible future directions could include understanding mini-batching thresholds and effects for stochastic
approximation with the Logistic Loss using the framework of [16], more general convex functions [9] and for
PCA using the framework of [30]. Moreover, similar mini-batching thresholds could be derived for algorithms
other than SGD, such as that of SVRG, by building on the framework established by [12].
8
Acknowledgments
Sham Kakade acknowledges funding from the Washington Research Foundation Fund for Innovation in
Data-Intensive Discovery. Rahul Kidambi thanks James Saunderson for useful discussions on matrix operator
theory.
12
References
[1] Alexandre Défossez and Francis R. Bach. Averaged Least-Mean-Squares: Bias-Variance Trade-offs and Optimal
Sampling Distributions. In AISTATS, volume 38, 2015.
[2] Gideon Mann, Ryan T. McDonald, Mehryar Mohri, Nathan Silberman, and Dan Walker. Efficient Large-Scale
Distributed Training of Conditional Maximum Entropy Models. In NIPS 22, 2009.
[3] Martin A. Zinkevich, Alex Smola, Markus Weimer, and Lihong Li. Parallelized stochastic gradient descent. In NIPS
24, 2011.
[4] Joseph K. Bradley, Aapo Kyrola, Danny Bickson, and Carlos Guestrin. Parallel Coordinate Descent for L1-Regularized
Loss Minimization. In ICML, 2011.
[5] Feng Niu, Benjamin Recht, Christopher Re, and Stephen J. Wright. Hogwild: A Lock-Free Approach to Parallelizing
Stochastic Gradient Descent. In NIPS 24, 2011.
[6] Mu Li, Tong Zhang, Yuqiang Chen, and Alexander J. Smola. Efficient mini-batch training for stochastic optimization.
In KDD, 2014.
[7] Yuchen Zhang and Lin Xiao. DiSCO: Distributed Optimization for Self-Concordant Empirical Loss. In ICML, 2015.
[8] B. T. Polyak and A. B. Juditsky. Acceleration of Stochastic Approximation by Averaging. SIAM J Control Optim,
volume 30, 1992.
[9] Francis Bach and Eric Moulines. Non-Asymptotic Analysis of Stochastic Approximation Algorithms for Machine
Learning. In NIPS 24, 2011.
[10] Francis R. Bach and Eric Moulines. Non-strongly-convex smooth stochastic approximation with convergence rate
O(1/n). In NIPS 26, 2013.
[11] Francis R. Bach. Adaptivity of averaged stochastic gradient descent to local strong convexity for logistic regression.
JMLR, volume 15, 2014.
[12] Roy Frostig, Rong Ge, Sham M. Kakade, and Aaron Sidford. Competing with the Empirical Risk Minimizer in a
Single Pass. In COLT, 2015.
[13] Shai Shalev-Shwartz and Tong Zhang. Stochastic Dual Coordinate Ascent Methods for Regularized Loss Minimization.
CoRR, abs/1209.1873, 2012.
[14] Alekh Agarwal, Peter L. Bartlett, Pradeep Ravikumar, and Martin J. Wainwright. Information-theoretic lower bounds
on the oracle complexity of stochastic convex optimization. IEEE Transactions on Information Theory, 2012.
[15] D. Ruppert. Efficient Estimations from a Slowly Convergent Robbins-Monro Process. Tech. Report, ORIE, Cornell
University, 1988.
[16] Francis R. Bach. Self-concordant analysis for logistic regression. Elec. J. of Statistics, volume 4, 2010.
[17] Aymeric Dieuleveut and Francis Bach. Non-parametric Stochastic Approximation with Large Step sizes. The Annals
of Statistics, 2015.
[18] Nicolas Flammarion and Francis R. Bach. From Averaging to Acceleration, There is Only a Step-size. In COLT,
volume 40, 2015.
[19] Rie Johnson and Tong Zhang. Accelerating Stochastic Gradient Descent using Predictive Variance Reduction. In
NIPS 26, 2013.
[20] Aaron Defazio, Francis R. Bach, and Simon Lacoste-Julien. SAGA: A Fast Incremental Gradient Method With Support
for Non-Strongly Convex Composite Objectives. In NIPS 27, 2014.
[21] Andrew Cotter, Ohad Shamir, Nati Srebro, and Karthik Sridharan. Better Mini-Batch Algorithms via Accelerated
Gradient Methods. In NIPS 24, 2011.
[22] Ofer Dekel, Ran Gilad-Bachrach, Ohad Shamir, and Lin Xiao. Optimal Distributed Online Prediction Using MiniBatches. JMLR, volume 13, 2012.
[23] Shai Shalev-Shwartz and Tong Zhang. Accelerated Mini-Batch Stochastic Dual Coordinate Ascent. In NIPS 26, 2013.
[24] Martin Takác, Peter Richtárik, and Nati Srebro. Distributed Mini-Batch SDCA. CoRR, abs/1507.08322, 2015.
[25] Yuchen Zhang, John C. Duchi, and Martin Wainwright. Divide and Conquer Ridge Regression: A Distributed
Algorithm with Minimax Optimal Rates. JMLR, volume 16, 2015.
[26] Jonathan Rosenblatt and Boaz Nadler. On the Optimality of Averaging in Distributed Statistical Learning. CoRR,
abs/1407.2724, 2014.
[27] Martin Takác, Avleen Singh Bijral, Peter Richtárik, and Nati Srebro. Mini-Batch Primal and Dual Methods for SVMs.
In ICML, volume 28, 2013.
[28] Horia Mania, Xinghao Pan, Dimitris Papailiopoulos, Benjamin Recht, Kannan Ramchandran, and Michael I. Jordan.
Perturbed Iterate Analysis for Asynchronous Stochastic Optimization. CoRR, abs/1507.06970, 2015.
[29] John C. Duchi, Sorathan Chaturapruek, and Christopher Ré. Asynchronous stochastic convex optimization. CoRR,
abs/1508.00882, 2015.
[30] Prateek Jain, Chi Jin, Sham M. Kakade, Praneeth Netrapalli, and Aaron Sidford. Streaming PCA: Matching Matrix
Bernstein and Near-Optimal Finite Sample Guarantees for Oja’s Algorithm. In COLT, 2016.
13
A
Appendix
A brief note on the organization of the appendix:
Theorem 1 establishes the generalization error of Algorithm 1. The proof of Theorem 1 can be found
in A.1 and consists of three parts, namely:
• Lemma 7, which rigorously establishes convergence (in section A.1.6) and provides an estimate of the
(scalar) contraction factor (in section A.1.7).
• Theorem 5 provides the asymptotic expansion of the finite-sample generalization error of Algorithm 1
on the noiseless problem (also referred to as the bias term), in section A.1.4.
• Theorem 6 provides the asymptotic expansion of the finite-sample generalization error of running
Algorithm 1 by initializing it at the solution w∗ (also referred to as the variance term), in section A.1.5.
The proofs of Theorem 2 and Theorem 3 are established in section A.3.
• Theorem 2 decomposes the tensor bound of Theorem 1 and provides a scalar estimate of the generalization performance of the mini-batch tail averaged SGD based Algorithm (Algorithm 1). The proof of
Theorem 2 can be found in section A.3.1.
• Theorem 3 provides the generalization error guarantee of the doubling based Algorithm (Algorithm 2).
The proof of Theorem 3 can be found in section A.3.2.
The proofs of both Theorem 2 and Theorem 3 utilize the claims established by lemmas in section A.2
along with the tensor bound established in Theorem 1.
Finally, the proof of lemma 4, which describes an instance that yields separation between learning rates
required for the realizable and agnostic case is described in section A.4.
Definition A.1. We introduce some notations relating the current iterate wt and the averaged iterate
Pt+N −1
def
def
wt,N to the optimum w∗ , namely: ηt = wt − w∗ , and η̄t,N = wt,N − w∗ = N1 s=t
ηs . Let Lt =
E [(wt − w∗ ) ⊗ (wt − w∗ )], Lt,N = E [(wt,N − w∗ ) ⊗ (wt,N − w∗ )] denote the covariance of the respective
def
def
parameter errors. Furthermore, we define the following contraction factors, ρT b = kI − γTb (γ)k2 , ρb =
max{ρT b , kI − γHk}.
Theorem 5. (Asymptotic Expansion of the error covariance for the bias term) Let E0 = η0 ⊗ η0 be the covariance
of the initial parameter error. If 0 < γ < γb,max and if t,i = 0 ∀ i = {1, ..., b}, t > 0 (i.e. we deal with the
noiseless problem), then, by averaging the iterates of “b” sample mini-batch SGD for N iterations starting from
the sth iteration, the covariance term of the error has the following asymptotic expansion:
Lbias
s,N = E[η̄s,N ⊗ η̄s,N ] =
−1
ρs+N
1 −1
−1
s
b
H
+
H
kE0 kF · M1 (4)
−
γI
T
(I
−
γT
)
E
+
C
·
b
b
0
1
L
R
N 2γ2
γλmin (H)N
Where C1 is a universal constant and M1 is a matrix with Frobenius norm ≤ 1.
Theorem 6. (Asymptotic Expansion of the error covariance for the variance term) Let Σ = E[2 x ⊗ x] and let
η0 = 0 (i.e. we start at the solution). If 0 < γ < γb,max , then, by averaging the iterates of “b” sample mini-batch
SGD for N iterations starting from the sth iteration, the covariance term of the error has the following asymptotic
expansion:
1
1
−1
−1
variance
Lt,N
= E[η̄s,N ⊗ η̄s,N ] =
I−
(HL + HR ) (HL )−1 + (HR )−1 − γI Tb −1 Σ
bN
γN
+
+1
ρN
2
b
HL −1 HR −1 Tb −1 Σ + C2 ·
kΣkF · M2
2
γbN
bN λmin (H)λmin (Tb )
Where C2 is a universal constant and M2 is a matrix with Frobenius norm ≤ 1.
14
(5)
Lemma 7. If 0 < γ < γb,max , then, Tb 0 and ρb < 1. Furthermore, for dimension d ≥ 2, the contraction factor
ρb can be upper bounded as:
ρb ≤ max(1 − 2γ(1 −
γ
γ
γ 2
)λmin (H), −1 + 2γ(1 −
)λmax (H) + 2(
) , 1 − γλmin (H))
γb,max
γb,max
γb,max
while in dimension d = 1, a very similar statement holds, i.e.:
ρb ≤ max(1 − 2γ(1 −
γ
γb,max
)λmin (H), −1 + 2γ(1 −
γ
γb,max
)λmax (H) + 2(
γ
γb,max
)2 , |1 − γλmin (H)|)
Remark 1. We note that γ ≤ γb,max is sufficient to ensure Algorithm 1 with constant step sizes achieves
statistically optimal rates. Note, our characterization of γb,max ensures that Tb 0 (refer to the proof of
lemma 7), thus ensuring contraction of all operators that characterize both the burn-in phase as well as
averaging phases of the Algorithm.
A.1
Proofs of Theorems 5, 6 and Lemma 7
We begin by detailing the notation and setup required for proofs of Theorems 5, 6 and lemma 7.
A.1.1
Deriving and setting up the basic recursion
At each iteration t of Algorithm 1, we are provided with b fresh samples {(xti , yti )}bi=1 drawn i.i.d. from the
distribution D. We start by recounting the mini-batch gradient descent update rule that allows us to move
from iterate wt−1 to wt :
b
γX
(hwt−1 , xti i − yti )xti
wt = wt−1 −
b i=1
where, 0 < γ < γb,max is the constant step size that is set to a value less than the maximum allowed learning
rate γb,max . We also recount the definition of wt,N which is the iterate obtained by averaging for N iterations
starting from the tth iteration, i.e.,
t+N −1
1 X
wt,N =
ws
N s=t
Let us first denote the residual error term by i = yi − hw∗ , xi i. By the first order optimality conditions of
w∗ , we observe that and x are uncorrelated, i.e, E(x,y)∼D [x] = 0. However, independence of and x is
guaranteed only iff E(x,y)∼D [y|x] = hw∗ , xi, i.e., in other words, if the noise is well specified. Note that the
def
risk of the iterate wt,N is written as R̄t,N = R(w̄t,N ). We can now write out the the generalization error as:
E[R̄t,N ] − R∗ =
1
1
Tr H E η̄t,N ⊗ η̄t,N = Ehη̄t,N , H η̄t,N i
2
2
We now write out the main recursion governing the mini-batch SGD updates in terms of η. :
b
b
γX
γX
ηt = I −
xti ⊗ xti ηt−1 +
ti xti
b i=1
b i=1
b
b
γX
γX
= I−
xti ⊗ xti ηt−1 +
ξti
b i=1
b i=1
def
Where, Ptb =
I−
γ
b
= Ptb ηn−1 + γζtb
Pb
Pb
def
x
⊗
x
and ζtb = 1b i=1 ξti =
ti
ti
i=1
(6)
1
b
Pb
i=1 ti xti .
Equation 6 automatically brings
out the “operator” view of analyzing the (expected) covariance of the centered estimate E [ηt ⊗ ηt ]. We now
note the following about the covariance of ζtb :
E ζtb ⊗ ζt0 b =
1 X
E ξti ⊗ ξt0 j
b2 i,j
15
=
b
h1 X
i
1
E
ξ
⊗
ξ
1[t = t0 ] = Σ
ti
ti
2
b i=1
b
1[t = t0 ]
(7)
Where, 1[.] is the indicator function, and equals 1 if the argument inside [.] is true and 0 otherwise. We
note that the expectation of the cross terms in equation 7 is zero owing to independence of the samples
{xti , yti }bi=1 as well as between {xti , yti }bi=1 , {xt0 i , yt0 i }bi=1 ∀ t 6= t0 and owing to the first order optimality
conditions. Owing to the invariance of ζtb on the iteration t, context permitting, we sometimes drop the
iteration index t from ζtb and simply refer to it as ζb . Q
t
Next we expand out the recurrence (6). Let Qj,t = ( k=j Pkb )T with the convention that Qt0 ,t = I ∀ t0 > t.
With this notation we have:
ηt = Ptb ηn−1 + γζtb
= Ptb Pt−1,b ...P1,b η0 + γ
t−1
X
{Ptb ....Pt−j+1,b }ζt−j,b
j=0
= Q1,t η0 + γ
t−1
X
Qt−j+1,t ζt−j,b
j=0
= Q1,t η0 + γ
t
X
Qj+1,t ζj,b .
(8)
j=1
A.1.2
The terms in the averaged iterate
Let us consider beginning averaging after a certain number of iterations “s”, i.e for t > s. In particular,
after averaging the updates for the next “N ” steps, we examine the quantity η̄s+1,N = ws+1,N − w∗ , where
ws+1,N is the iterate we obtain by averaging for “N ” iterations starting from iteration s + 1. We write out the
expression for η̄s+1,N starting out from equation 8:
η̄s+1,N =
A.1.3
s+N
1 X
ηt
N t=s+1
=
s+N
s+N
t
o
γ X X
1n X
Q1,t η0 +
Qj+1,t ζj,b
N t=s+1
N t=s+1 j=1
=
s+N
s s+N
o
s+N
o
X s+N
X
1n X
γ n X X
Q1,t η0 +
Qj+1,t ζj,b +
Qj+1,t ζj,b
N t=s+1
N
j=1 t=s+1
j=s+1 t=j
=
s+N
s
s+N
s+N
s+N
o
o
o
1n X
γ Xn X
γ X nX
Q1,t η0 +
Qj+1,t ζj,b +
Qj+1,t ζj,b
N t=s+1
N j=1 t=s+1
N j=s+1 t=j
(9)
Computing the outerproduct
Firstly, the outer product of η̄s+1,N with itself contains three terms, each involving the outerproduct of each
term in equation 9 with itself. The 6 other terms are the cross terms all of whose expectation are zero. We
note that E[ξt,i ] = E[t,i xt,i ] = 0 ∀ i, t, and using this, it is clear to see the outer product of the cross term
between terms 2 and 3 is zero, since they deal with disjoint samples j = 1, ..., s and j = s + 1, ..., s + N and
the q.,. doesn’t deal with the same sample index j. The outer product of the 1st term with the 2nd and 3rd
term, E[x(i) x(j) x(k) ] = 0 (as noted by [9, 1], and if this is not satisfied we lose at most a constant factor two
in our generalization error estimate), and as noted in section 4. This leaves us to deal with the 3 of the terms,
involving outer product of the each of the terms with itself. Finally, we note that the bound in the theorem 5
refers to the outerproduct computation of the first term, and that of theorem 6 refers to the outerproduct
computation of the second and third term.
16
A.1.4
Proof of Theorem 5
We begin by noting that this proof generalizes the proof of Défossez and Bach [1] to account for mini-batching
and tail-averaged SGD. In this proof, we assume that the problem is noiseless, i.e., we assume j,k = 0 a.s.
∀ j, k.
" s+N
#
n X
o
n s+N
oT
X
1
E[η̄s+1,N ⊗ η̄s+1,N ] = 2 E
Q1,t η0 ⊗ η0
Q1,t0
N
t=s+1
t0 =s+1
s+N
X
1
Q1,t η0 ⊗ η0 QT1,t0
= 2E
N
t,t0 =s+1
"
#
s+N
s+N
o
X n
1 X
T
T
T
T
E Q1,t η0 ⊗ η0 Q1,t +
= 2
Q1,t η0 ⊗ η0 Q1,t Qt+1,t0 + Qt+1,t0 Q1,t η0 ⊗ η0 Q1,t
N t=s+1
t0 =t+1
"
#
s+N
s+N
X
1 X
(t0 −t)
(t0 −t)
I+
(I − γH)R
+ (I − γH)L
E[Q1,t η0 ⊗ η0 QT1,t ] (∗)
= 2
N t=s+1
0
t =t+1
"
#
s+N
s+N
X
X
1
(t0 −t)
(t0 −t)
I+
= 2
(I − γH)R
+ (I − γH)L
(I − γTb )t E0 (∗∗)
N t=s+1
0
t =t+1
=
1
N2
s+N
X
h
I + (γHR )−1 {(I − γH) − (I − γH)s+N −t+1 }R
t=s+1
i
+ (γHL )−1 {(I − γH) − (I − γH)s+N −t+1 }L (I − γTb )t E0
(∗ ∗ ∗)
0
where, (∗) follows since E [Qt+1,t0 ] = (I − γH)t −t , and (∗ ∗ ∗) follows by summing up the geometric series
1
(which converges when γ ≤ kHk
, which is the case as discussed in section A.1.6), and finally, (∗∗) follows
2
through the following argument:
E[Q1,t η0 ⊗ η0 QT1,t ] = E[Ptb Pt−1,b ...P1b η0 ⊗ η0 P1b ...Pt−1,b Ptb ]
= E[E[Ptb (Pt−1,b ...P1b η0 ⊗ η0 P1b ...Pt−1,b )Ptb |Ft−1 ]]
where, Ft−1 refers to the sigma-field defined by samples {{(xt0 ,i , yt0 ,i )}bi=1 }t−1
t0 =1 . We note that (Pt−1,b ...P1b η0 ⊗
η0 P1b ...Pt−1,b ) is a constant w.r.t. the inner expectation because of the conditioning Ft−1 . Now, the observation can be seen clearly using the argument below (with C being a matrix that is independent w.r.t. the
expectation):
b
b
h
i
γX
γX
E[Ptb CPttb ] = E (I −
xti ⊗ xti )C(I −
xti ⊗ xti )
b i=1
b i=1
γ2
γ 2 (b − 1)
E[xT Cx x ⊗ x] +
E[xT1 Cx2 x1 ⊗ x2 ]
b
b
γ(b − 1)
γ
HL HR ))C
= (I − γ(HL + HR − M −
b
b
= (I − γTb )C
= C − γCH − γHC +
(10)
where, we apply the fact that the samples {(xti , yti )}bi=1 are independent and identically distributed draws
from the distribution D. The final expression is obtained by applying the argument recursively for all
t0 = t, t − 1, .., 1.
Next we bound the exponential term, in particular:
AN =
s+N
i
−1 X h
−t+1
s+N −t+1
(γHR )−1 (I − γH)s+N
+ (γHL )−1 (I − γH)L
(I − γTb )t E0
R
2
N t=s+1
=⇒ kAN kF ≤
2
ρs+N +1 kE0 kF
γλmin (H)N b
17
The remaining terms are bounded as the following:
E[η̄s+1,N
s+N
1 X
⊗ η̄s+1,N ] =
[HL −1 + HR −1 − γI](I − γTb )t E0 + AN
γN 2 t=s+1
=
1
[HL −1 + HR −1 − γI]Tb−1 [(I − γTb )s+1 − (I − γTb )s+N +1 ]E0 + AN
γ2N 2
Considering the exponential terms separately, we write out the covariance operator as:
E[η̄s+1,N ⊗ η̄s+1,N ] =
1
[HL −1 + HR −1 − γI]Tb−1 (I − γTb )s+1 E0 + AN + BN
γ2N 2
with BN defined as:
−1
[HL −1 + HR −1 − γI]Tb−1 (I − γTb )s+N +1 E0
γ2N 2
2
1
+1
=⇒ kBN kF ≤ 2 2
− γ ρs+N
kE0 kF
Tb
γ N λmin (Tb ) λmin (H)
BN =
The last equation above assumes that γ < λmin2(H) . Finally, using triangle inequality that kAN + BN kF ≤
kAN kF + kBN kF and hence, when we begin averaging starting at iteration s + 1 and average for N iterations
subsequently, we get a geometric decay of the bias term. The final covariance operator of the bias term thus
has the following asymptotic expansion:
E[η̄s+1,N ⊗ η̄s+1,N ] =
ρbs+N +1
1
−1
−1
−1
s+1
[H
+
H
−
γI]T
(I
−
γT
)
E
+
O(
kE0 kF )
L
R
b
0
b
γ2N 2
γλmin (H)N
The above bound requires the following conditions to hold: (i) Tb (γ) 0, (ii) γ < λmax1(H) . This wraps up
the proof of Theorem 5 and indicates the geometric decay of the bias term during the unaveraged phase and
the O( N12 ) decay during the averaging phase.
A.1.5
Proof of Theorem 6
We begin by noting that this proof generalizes the proof of Défossez and Bach [1] to account for mini-batching
and tail-averaged SGD. For bounding the variance part of the generalization error, we go over a sharp analysis
of the stochastic process defined by the mini-batch tail averaged SGD initialized at the solution w∗ , i.e.
w0 = w∗ , and thus the process is driven solely by the noise. We note that we require bounding the behavior
of two terms, the first of which is defined by the noise introduced during the unaveraged phase, and the
second being one that deals with the noise introduced during the tail averaging phase.
s
s+N
s+N
s+N
2
X
X
X
X
γ
E[η̄s+1,N ⊗ η̄s+1,N ] = 2
E[Qj+1,t ζjb ⊗ ζjb QTj+1,t0 ] +
E[Qj+1,t ζjb ⊗ ζjb QTj+1,t0 ]
N
j=1 0
j=s+1 0
t,t =j
t,t =s+1
2
=
γ
(K1 + K2 )
N2
(11)
We consider the first term, and specifically the inner two terms of the summation:
K1j =
s+N
X
s+N
X
E[Qj+1,t ζjb ⊗ ζjb QTj+1,t0 ]
t=s+1 t0 =s+1
=
s+N
X
t=s+1
s+N
t−1
h
i
X
X
E Qj+1,t ζjb ⊗ ζjb QTj+1,t +
Qj+1,t ζjb ⊗ ζjb QTj+1,t QTt+1,t0 +
Qt0 +1,t Qj+1,t0 ζjb ⊗ ζjb QTj+1,t0
t0 =t+1
t0 =s+1
18
=
s+N
X
s+N
t−1
i
h
X
X
0
0
E Qj+1,t ζjb ⊗ ζjb QTj+1,t +
Qj+1,t ζjb ⊗ ζjb QTj+1,t (I − γH)t −t +
(I − γH)t−t Qj+1,t0 ζjb ⊗ ζjb QTj+1,t0
t0 =t+1
t=s+1
=
s+N
X
t0 =s+1
h
i
I + (γHR )−1 {(I − γH) − (I − γH)s+N −t+1 }R + (γHL )−1 {(I − γH) − (I − γH)s+N −t+1 }L (I − γTb )t−j E[ζjb ⊗ ζjb ]
t=s+1
=
s+N
i
1 X h
I + (γHR )−1 {(I − γH) − (I − γH)s+N −t+1 }R + (γHL )−1 {(I − γH) − (I − γH)s+N −t+1 }L (I − γTb )t−j Σ
b t=s+1
Note, the second to last line above is obtained by seeing that we can swap the summations for the third term
and then replace the roles of t and t0 and applying equation 10, following which we note that E[ζjb ⊗ζjb ] = 1b Σ
from equation 7 for the equation in the last line above. And hence, the first term K1 is:
K1 =
s
X
K1j
j=1
=
s+N
s
i
1X X h
I + (γHR )−1 {(I − γH) − (I − γH)s+N −t+1 }R + (γHL )−1 {(I − γH) − (I − γH)s+N −t+1 }L (I − γTb )t−j Σ
b j=1 t=s+1
(12)
In a very similar manner, it is easy to observe the second term is:
K2 =
s+N s+N
i
1 X Xh
I + (γHR )−1 {(I − γH) − (I − γH)s+N −t+1 }R + (γHL )−1 {(I − γH) − (I − γH)s+N −t+1 }L (I − γTb )t−j Σ
b j=s+1 t=j
(13)
2
γ
If K1 and K2 can be bounded, then it is clear that the variance decays as N
2 (K1 + K2 ) as is seen from
equation 11. Let us begin with K1 (from equation 12). We first exchange the two summations and sum up
the resulting series w.r.t. j.
s+N
1 X h
I + (γHR )−1 {(I − γH) − (I − γH)s+N −t+1 }R + (γHL )−1 {(I − γH)−
K1 =
b t=s+1
i
h
i
(I − γH)s+N −t+1 }L (γTb )−1 (I − γTb )t−s − (I − γTb )t Σ
N
1 Xh
I + (γHR )−1 [(I − γH) − (I − γH)N −t+1 ]R + (γHL )−1 [(I − γH)−
b t=1
i
h
i
(I − γH)N −t+1 ]L (γTb )−1 (I − γTb )t − (I − γTb )t+s Σ
=
Next we bound the exponential terms, i.e.
N
i
h
i
1 Xh
(γHL )−1 + (γHR )−1 − I (γTb )−1 (I − γTb )t − (I − γTb )t+s Σ + A0N
b t=1
i
h
ih
i
1h
= (γHL )−1 + (γHR )−1 − I (γTb )−2 (I − γTb ) − (I − γTb )s+1 I − (I − γTb )N Σ + A0N
b
K1 =
where A0N is:
A0N = −
N
i
h
i
1 Xh
−t+1
−t+1
(γHR )−1 (I − γH)N
+ (γHL )−1 (I − γH)N
(γTb )−1 (I − γTb )t − (I − γTb )t+s Σ
R
L
b t=1
whose norm is bounded as:
kA0N kF ≤
2N
γ 2 bλmin (H)λmin (Tb )
19
+1
+s+1
(ρN
+ ρN
)kΣkF
b
b
0
Again, we accumulate the exponential terms existing in K1 in BN
, leading to:
i
h
i
1h
(γHL )−1 + (γHR )−1 − I (γTb )−2 (I − γTb ) − (I − γTb )s+1 Σ + B0N + A0N
b
h
i
i
1h
= (γHL )−1 + (γHR )−1 − I (γTb )−2 (I − γTb ) Σ + C0N + B0N + A0N
b
K1 =
(14)
where, B0N is:
B0N = −
i
h
i
1h
(γHL )−1 + (γHR )−1 − I (γTb )−2 (I − γTb ) − (I − γTb )s+1 (I − γTb )N Σ
b
whose norm is bounded using triangle inequality as:
kB0N kF ≤
1
bγ 3 (λmin (Tb ))2
(
2
N
− γ)(ρT b + ρs+1
T b )ρT b kΣkF
λmin (H)
finally, C0N is:
C0N = −
i
1h
(γHL )−1 + (γHR )−1 − I (γTb )−2 (I − γTb )s+1 Σ
b
whose norm is bounded as:
kC0N k ≤
2
1
(
− γ)ρs+1
T b kΣkF
bγ 3 (λmin (Tb ))2 λmin (H)
which rounds up bounding all terms of equation 14, except for the leading order term. This term ends up
canceling with a term involved in bounding K2 in equation 17.
We now begin bounding the terms in K2 (from equation 13). Towards this effect, we begin by interchanging
the summations in K2 :
t
s+N
1 X X h
I + (γHR )−1 [(I − γH) − (I − γH)s+N −t+1 ]R +
b t=s+1 j=s+1
i
(γHL )−1 [(I − γH) − (I − γH)s+N −t+1 ]L (I − γTb )t−j Σ
K2 =
s+N
1 X h
I + (γHR )−1 [(I − γH) − (I − γH)s+N −t+1 ]R +
b t=s+1
i
(γHL )−1 [(I − γH) − (I − γH)s+N −t+1 ]L (γTb )−1 [I − (I − γTb )t−s ]Σ
=
N
1 Xh
I + (γHR )−1 [(I − γH) − (I − γH)N −t+1 ]R +
b t=1
i
h
i
(γHL )−1 [(I − γH) − (I − γH)N −t+1 ]L (γTb )−1 I − (I − γTb )t Σ
=
We begin considering the exponential terms, i.e.,:
N
K2 =
i
h
i
1 Xh
(γHR )−1 + (γHL )−1 − I (γTb )−1 I − (I − γTb )t Σ + A00N
b t=1
where A00N is:
N
A00N = −
i
h
i
1 Xh
N −t+1
−t+1
−1
t
(γHR )−1 (I − γH)R
+ (γHL )−1 (I − γH)N
(γT
)
I
−
(I
−
γT
)
Σ
b
b
L
b t=1
= B00N + C00N
20
where,
N
i
1 Xh
N −t+1
−t+1
(γHR )−1 (I − γH)R
+ (γHL )−1 (I − γH)N
(γTb )−1 Σ
L
b t=1
h
i
h
ii
1h
+1
+1
+ (γHL )−2 (I − γH)L − (I − γH)N
(γTb )−1 Σ
= − (γHR )−2 (I − γH)R − (I − γH)N
R
L
b
= F00N + G00N
B00N = −
Where F00N is:
1
(γHR )−2 (I − γH)R + (γHL )−2 (I − γH)L (γTb )−1 Σ
b
1
= − 3 (HR −2 − γHR −1 ) + (HL −2 − γHL −1 ) Tb −1 Σ
bγ
1
= − 3 (HL −1 + HR −1 − γI)(HL −1 + HR −1 ) − HL −1 HR −1 − HR −1 HL −1 Tb −1 Σ
bγ
1
= − 3 (HL −1 + HR −1 − γI)(HL −1 + HR −1 ) − 2HL −1 HR −1 Tb −1 Σ
bγ
F00N = −
(15)
The norm of F00N is bounded along with the leading order variance term. G00N is then written out as:
G00N =
1
+1
+1
(γHR )−2 (I − γH)N
+ (γHL )−2 (I − γH)N
(γTb )−1 Σ
R
L
b
whose norm is bounded as:
kG00N kF ≤
+1
ρN
Hb
kΣkF
bγ 3 (λmin (H))2 λmin (Tb )
N
C00N =
i
1 Xh
−t+1
N −t+1
−1
(γHR )−1 (I − γH)N
+
(γH
)
(I
−
γH)
(γTb )−1 (I − γTb )t Σ
L
R
L
b t=1
whose norm is bounded as:
kC00N kF ≤
2
+1
(N ρN
)kΣkF
b
γ 2 bλmin (H)λmin (Tb )
Now, K2 is written as the following:
i
Nh
K2 =
(γHR )−1 + (γHL )−1 − I (γTb )−1 Σ + D00N + A00N
b
(16)
where the N upfront is the time for which the iterates have been averaged, and this is the term that leads to
γ2
1
00
N rate of decay for the variance term (which has the factor of N 2 K2 ), and DN is obtained through summing
the resulting geometric series:
i
h
i
1h
D00N = − (γHR )−1 + (γHL )−1 − I (γTb )−2 (I − γTb ) − (I − γTb )N +1 Σ
b
i
1h
= − (γHR )−1 + (γHL )−1 − I (γTb )−2 (I − γTb )Σ + E00N
(17)
b
where, E00N is:
E00N =
i
1h
(γHR )−1 + (γHL )−1 − I (γTb )−2 (I − γTb )N +1 Σ
b
whose norm is bounded as:
kE00N kF ≤
1
γ 3 b(λ
min (Tb
))2
(
2
+1
− γ)ρN
T b kΣkF
λmin (H)
21
We note that the leading term in D00N in equation 17 is exactly the negative of the leading order term in
equation 14, thus canceling out without requiring us to bound their behavior. This allows us to provide the
γ2
final theorem statement, combining the leading order terms from equations 15 and 16 along with the N
2
outside the summation,
1
(γHR )−1 + (γHL )−1 − I (γTb )−1 Σ
bN
1
−
(HL −1 + HR −1 − γI)(HL −1 + HR −1 ) − 2HL −1 HR −1 Tb −1 Σ
bγN 2
E[η̄s,N ⊗ η̄s,N ] =
+1
ρN
b
kΣkF )
bN λmin (H)λmin (Tb )
1
1
=
(HL −1 + HR −1 ) (HR )−1 + (HL )−1 − γI Tb −1 Σ
I−
bN
γN
+ O(
+
+1
ρN
2
−1
−1
−1
b
H
H
T
Σ
+
O(
kΣkF )
L
R
b
γbN 2
bN λmin (H)λmin (Tb )
with which the proof of Theorem 6 concludes.
A.1.6
Some intermediate Lemmas required to prove Lemma 7
We establish a sequence of lemmas in this subsection in order to prove that the operators I − γH and I − γTb
contract with the choice of our stepsizes as described in lemma 7. We then provide a scalar bound for the
contraction factor, and this is done by generalizing the lemma appearing in Défossez and Bach [1] to the
mini-batching case.
Lemma 8. Without loss of generality, let us assume that the eigenvalues {λi }di=1 > 0 of H be arranged in
decreasing order of their values. If γ > 0 is chosen such that Tb (γ) 0, i.e.:
hA, Tb Ai = 2 Tr AT HA −
γ
b−1
Tr AT MA − γ
Tr AT HL HR A > 0 ∀ A ∈ S(d)
b
b
then the following upper bound on the stepsize γ holds: γ <
λi +λj
b
b+1 ( λi λj )
holds ∀ {i, j} ∈ {1, 2, ..., d}×{1, 2, ..., d}
Proof. This proof of this lemma generalizes the proof appearing in Défossez and Bach [1] to establish bounds
on the stepsize for the mini-batching case.
Expanding out the inner product, we have:
γ
b−1
M − γ(
)HL HR )A)
b
b
b−1
γ
= 2 Tr(AT HA) − Tr(AT MA) − γ(
) Tr(AT HL HR A)
b
b
γ
b−1
= 2 Tr(AT HA) − E(xT Ax)2 − γ(
)E(xT1 Ax2 )2
b
b
hA, Tb Ai = Tr(AT (HL + HR −
Where, (x1 , y1 ), (x2 , y2 ) and (x, y) are independent draws from the distribution D. Next, we consider the
following:
2 Tr (AT HA) −
γ
b−1
γ
b−1
[Tr (AH)]2 − γ(
) Tr(AT HL HR A) = 2 Tr (AT HA) − [Tr AE [x ⊗ x]]2 − γ(
) Tr(AHAH)
b
b
b
b
γ
b−1
≥ 2 Tr (AT HA) − E(xT Ax)2 − γ(
) Tr(AT HAH)
b
b
= hA, Tb Ai > 0
Where, we note that the second line follows through the application of Jensen’s inequality. Further, Let
H = UDUt to be the eigen decomposition of H so that D contains the eigenvalues {λi }di=1 in its diagonal. For
any (i, j) ∈ {1, ..., d} × {1, ..., d}, we set the matrix A = ci ui ⊗ ui + cj uj ⊗ uj so as to observe the following:
2 Tr (AT HA) −
By setting ci =
γ
b−1
γ
b−1 2 2
[Tr AH]2 − γ(
) Tr(AT HL HR A) = 2(c2i λi + c2j λj ) − (ci λi + cj λj )2 − γ(
)(ci λi + c2j λ2j )
b
b
b
b
>0
q
λj
λi
and cj =
q
λi
λj ,
the result follows.
22
b
Corollary 8.1. By setting i = j = 1 in lemma 8, we observe that the condition Tb 0 allows γ < ( b+1
)·
2
λmax (H) .
Remark 2. We note that an immediate consequence of corollary 8.1 is that by setting γ such that Tb 0,
−I ≺ I − γH ≺ I.
Lemma 9. With Tb = HL + HR − γb M − γ( b−1
b )HL HR defined as before, if γ > 0 is chosen such that
hZ, Tb Zi > 0 ∀ Z ∈ S(d), then, ∀ d ≥ 2, we have:
hZ, (I − γTb ) Zi ≥ −hZ, Zi ∀ Z ∈ S(d).
while in dimension d = 1 we have the following condition:
I − γTb ≥ 0
Before we prove the above lemma, we prove a simpler version of this lemma where a stronger conclusion
holds, but only ∀ Z ∈ S + (Rd ) (as opposed to the statement of lemma 9 where Z ∈ S(d)).
Lemma 10. If γ <
1
kHk2 ,
the following statement holds:
hZ, (I − γ(HL + HR − γHL HR )) Zi ≥ 0 ∀ Z ∈ S(d).
Proof. Assuming the eigenvalues of H are {λi }di=1 arranged in decreasing order, the proof requires the
observation that the eigenvalues of (HL + HR − γbHL HR ) are respectively {λi + λj − γλi λj }di,j=1 . This
is because if {ui }di=1 are the eigenvectors of H, then, {ui ⊗ uj }di,j=1 are the eigenvectors for HL , HR and
HL HR with corresponding eigenvalues being λi , λj and λi λj respectively. This is because:
HL ui ⊗ uj = Hui ⊗ uj = λi ui ⊗ uj
HR ui ⊗ uj = ui ⊗ uj H = λj ui ⊗ uj
HL HR ui ⊗ uj = Hui ⊗ uj H = λi λj ui ⊗ uj
If we prove that the eigenvalues of (HL + HR − γHL HR ) are all upper bounded by γ1 , the lemma is proven
since this implies HL + HR − γHL HR ≺ γ1 I.
We consider two cases, (i) when i = j, in which case we deal with eigenvalues of the form 2λi − γλ2i . If we
require this to be less than γ1 , then, (γλi − 1)2 > 0 which holds automatically ∀ γ. (ii) When i 6= j, we require
λi + λj − γλi λj < γ1 , or, γ 2 λi λj − γ(λi + λj ) + 1 > 0. If we assume without loss of generality, λi > λj , then,
the solution of interest is γ < λ1i . Taking the intersection of all the solutions obtained from solving equations
∀ i, j, we obtain, γ < λmax1(H) . We consider lemma 8 with i = 1 and j 6= 1. In this case, we know that
b(λ
(H)+λ )
b(λ
(H)+λ )
λ +λ
(H)
λ
max
j
max
j
j
max
γ < (b+1)λ
. The proof completes iff λmax1(H) < (b+1)λ
or b+1
implying b > λmax j(H) ,
b <
λj
max (H)λj
max (H)λj
which is true ∀ b ≥ 1. We also note that the result holds even in the case that there exists eigenvalues with
multiplicity ≥ 1 since that is covered through case(i).
Lemma 11. For any PSD matrix A ∈ S + (Rd ), the following inequality holds:
hA, HL HR Ai ≤ hA, MAi
Proof. The claim is proven through the following argument:
hA, HL HR Ai = hA, HAHi
= Ex1 ,x2 (x1 > Ax2 )2
2
≤ Ex x> Ax
≤ Ex (x> Ax)2
= hA, MAi.
23
Lemma 12. Under the hypothesis of Lemma 9, we have:
hZ, (I − γTb ) Zi ≥ 0 ∀ Z ∈ S + (Rd ).
Proof of Lemma 12. The first observation is that for any PSD matrix Z ∈ S + (Rd ), using lemma 11, hZ, HL HR Zi ≤
hZ, MZi.
This implies that hZ, Tb Zi = hZ, (HL + HR − γb M − γ( b−1
b )HL HR )Zi ≤ hZ, (HL + HR − γHL HR )Zi.
The proof rounds up by by employing Lemma 10.
Proof of Lemma 9. The proof goes via a contradiction. Let us assume that there exists Z ∈ S(d) such that the
conclusion of Lemma 9 is false. So we have hZ, (I − γTb ) Zi < −hZ, Zi, which implies that
2
hZ, Tb Zi > 2 kZkF /γ
(18)
Using the spectral decomposition theorem, let Z = Z1 −Z2 where Z1 , Z2 ∈ S + (Rd ) and Z1 Z2 = Z2 Z1 = 0. Let
2
2
2
2
b def
b Tb Zi
b > 0,
Z
= Z1 + Z2 . This means that kZk = Zb
= kZ1 k + kZ2 k . Since by hypothesis we have hZ,
F
adding this to (18), we obtain
F
F
F
b Tb Zi
b > 2 kZk2 /γ
hZ, Tb Zi + hZ,
F
2
2
⇒ 2hZ1 , Tb Z1 i + 2hZ2 , Tb Z2 i > 2(kZ1 kF + kZ2 kF )/γ
2
⇒ either
hZ1 , Tb Z1 i > kZ1 kF /γ
⇒ either
hZ1 , (I − γTb )Z1 i < 0
or
or
2
hZ2 , Tb Z2 i > kZ2 kF /γ
hZ2 , (I − γTb )Z2 i < 0.
This contradicts with the conclusion of Lemma (12) which shows that for every Z ∈ S + (Rd ), we have
hZ, (I − γTb )Zi ≥ 0. This proves that there does not exist Z ∈ S(d) for which (18) holds, thus concluding the
proof of Lemma 9.
Note that in dimension d = 1, we have:
γ
b − 1 2 2
I − γTb = 1 − γ(2E x2 − E x4 − γ(
)E x )
b
b
γ2
b − 1 2 2
= 1 − 2γE x2 + E x4 + γ 2 (
)E x
b
b
2
≥ 1 − 2γE x2 + γ 2 E x2
= (1 − γE x2 )2 > 0
2
with the inequality relating the second to the fourth moment namely, E x2 ≤ E x4 being through the
application of Jensen’s inequality.
T
Lemma 13. With γ > 0, Tb = HL + HR − γb M − γ( b−1
b )HL HR , we have ∀A ∈ S(d), 2 Tr(A HA) −
γ
b−1
T
2
T
b E(x Ax) − γ( b ) Tr A HL HR A > 0, then,
• kI − γTb k < 1
• kI − γHk < 1
Proof. This lemma generalizes the lemma governing contraction of all associated operators appearing in DéfosT
sez and Bach [1]. We note that the condition ∀ A ∈ S(d), 2 Tr(AT HA)− γb E(xT Ax)2 −γ( b−1
b ) Tr A HL HR A >
b
2
2
0 implies Tb 0, with the step size set as 0 < γ < γb,max < b+1 ( λmax (H) ) < λmax (H) .
Next, to prove kI − γTb k < 1, we require −I ≺ I − γTb ≺ I, the right hand side of which follows from
the fact that Tb 0. The left hand side follows using lemma 9.
We know that H 0. Next, to ensure kI − γHk < 1, we require −I ≺ I − γH ≺ I. The right side of the
inequality holds owing to the fact that H 0. The left side is equivalent to having H ≺ γ2 I, or equivalently,
γ < λmax2(H) which follows from corollary 8.1 derived from lemma 8.
24
A.1.7
Proof of Lemma 7
We first note that with our choice of γb,max , we have Tb ≥ 0. This is because
γ(b − 1)
γ
E hx, Axi2 −
hH, AHAi
b
b
h
i
γ
γ(b − 1)
2
2
≥ 2 Tr (AHA) − E kxk kAxk −
kHk Tr (AHA)
b
b
h
i
γ
γ(b − 1)
2
≥ 2 Tr (AHA) − R2 E kAxk −
kHk Tr (AHA)
b
b
γ
R2 + (b − 1) kHk Tr (AHA) ,
≥ 2−
b
hA, Tb Ai = 2 Tr (AHA) −
and using the definition of γb,max finishes the claim.
The rest of the proof is a generalization of the bound on the contraction factor provided by Défossez and
Bach [1]. As we operate with a step size 0 < γ < γb,max , all the operators contract, i.e., ρb < 1 and Tbγ 0 as
per lemma 13. Furthermore, given the largest and smallest eigenvalues of H as λmax (H) and λmin (H), then,
we denote the largest and smallest eigenvalues of HL + HR as 2λmax (H) and 2λmin (H).
We first provide bounds on the operator norm of Tbγ . Let λmax (Tbγ ) and λmin (Tbγ ) represent the largest and
smallest eigenvalues of the operator Tbγ . Then,
ρTbγ = max{1 − γλmin (Tbγ )), γλmax (Tbγ ) − 1}
For a particular step size γ, we define α =
Tbγ = HL + HR −
γ
γb,max
implying the following:
b−1
γ
M − γ(
)HL HR
b
b
= (1 − α)(HL + HR ) + α(HL + HR −
γb,max
b−1
M − γb,max (
)HL HR )
b
b
γb,max
= (1 − α)(HL + HR ) + αTb
γ
Since Tb b,max 0, Tbγ 2(1 − α)λmin (H)I, implying λmin (Tbγ ) ≥ 2(1 − α)λmin (H).
γ
γ
2
I. This implies that,
Next, −I ≺ I − γb,max Tb b,max implying Tb b,max ≺ γb,max
γ
Tbγ = (1 − α)(HL + HR ) + αTb b,max
2α
≺ 2(1 − α)λmax (H)I +
I
γb,max
implying λmax (Tbγ ) ≤ 2(1 − α)λmax (H) +
2α
γb,max .
This when plugged into the operator norm of Tbγ yields,
α
)}
γb,max
γ
γ
γ
= max{1 − 2γ(1 −
)λmin (H), −1 + 2γ(1 −
)λmax (H) + 2(
)2 }
γb,max
γb,max
γb,max
ρTbγ = max{1 − 2α(1 − α)γb,max λmin (H), −1 + 2αγb,max ((1 − α)λmax (H) +
Similarly, the operator norm of I − γH is bounded as: ρHb = 1 − γλmin (H), since in dimension d ≥ 2,
1 − γλmin (H) ≥ γλmax (H) − 1. In d=1, ρHb = |1 − γλmin (H)|. This rounds up the claim in lemma 7.
A.1.8
Proof of Theorem 1
The proof of Theorem 1 builds on the results of lemmas 7, 13 and Theorems 5, 6.
By instantiating Equation 4 of Theorem 5 and Equation 5 of Theorem 6 with N =
the output of the Algorithm 1 is w, the result follows, i.e.:
h
∗
∗ >
E (w − w )(w − w )
i
1
= n
GTb −1 (I − γTb )s E0 + O
( b − s)2 γ 2
25
n
b
− s, and noting that
!
n/b
ρb
γλmin (H)
n
b
kE0 kF
−s
"
1
+
I−
n − sb
γ
#
1
−1
−1
(HL + HR ) GTb −1 Σ
n
−
s
b
n
2
+
γb
A.2
n
b
2 HL
−s
−1
HR
−1
Tb
−1
Σ+O
−s+1
ρbb
kΣkF
(n − bs)λmin (H)λmin (Tb )
!
Proofs of some useful Lemmas
Lemma 14.
i
h
2
E kxk xx>
2
· H.
γ1,max
Proof. The proof of the claim above is straightforward given that Ekxk2 x ⊗ x R2 H and owing to the
definition of γ1,max = R22 .
Lemma 15. Denoting the assumption (A) γ ≤ γb,max /2,
1. With (A) in place, Tb −1 W 0 for every W ∈ S(d), W 0
2
2. With (A) in place, Tr Tb −1 W ≤ λmin
· Tr (W) ∀ W ∈ S(d), W 0
3
· kWk∗ for every W ∈ Rd×d
3. With (A) in place, HL −1 + HR −1 − γI W ∗ ≤ λmin
4. With (A) in place, Tr HL −1 + HR −1 − γI Tb −1 W ≤ λ26 · Tr (W) ∀ W ∈ S(d), W 0
min
5. (I − γTb ) W 0 ∀ W ∈ S(d), W 0
6. With (A) in place, Tr ((I − γTb ) W) ≤ (1 − γλmin /2) Tr (W) ∀ W ∈ S(d), W 0
−1
7. Tr (HR + HL ) A = 21 Tr H−1 A ∀ A ∈ S + (Rd )
8. With (A) in place,
3
Tr Tb −1 Σ ≤ Tr H−1 Σ
2
Proof. Proof of claim 1 in Lemma 15: We require to prove Tb −1 operating on a PSD matrix produces a PSD
matrix, or in other words, Tb −1 is a PSD map.
Tb −1 = [HL + HR −
γ
(M + (b − 1)HL HR )]−1
b
1
1
= (HL + HR )− 2 (HL + HR ) 2 [HL + HR −
1
= (HL + HR )− 2 [I −
1
1
γ
(M + (b − 1)HL HR )]−1 (HL + HR ) 2 (HL + HR )− 2
b
1
1
1
γ
(HL + HR )− 2 (M + (b − 1)HL HR )(HL + HR )− 2 ]−1 (HL + HR )− 2
b
1
1
(19)
Now, we prove that k γb (HL + HR )− 2 (M + (b − 1)HL HR )(HL + HR )− 2 k < 1. Given γ < γb,max /2, we employ
lemma 8 to note that Tb 0.
Tb 0
⇒ HL + HR −
γ
(M + (b − 1)HL HR ) 0
b
γ
(M + (b − 1)HL HR ) ≺ HL + HR
b
1
1
γ
⇒ (HL + HR )− 2 (M + (b − 1)HL HR )(HL + HR )− 2 ≺ I
b
1
1
γ
⇒ k (HL + HR )− 2 (M + (b − 1)HL HR )(HL + HR )− 2 k < 1
b
⇒
26
With this fact in place, we employ Taylor series to expand Tb −1 in equation 19, i.e.:
1
Tb −1 = (HL + HR )− 2
=
∞
X
1
1
1
γ
( (HL + HR )− 2 (M + (b − 1)HL HR )(HL + HR )− 2 )i (HL + HR )− 2
b
i=0
∞
X
γ
( (HL + HR )−1 (M + (b − 1)HL HR ))i (HL + HR )−1
b
i=0
The proof completes by employing the following facts: Using Lyapunov’s theorem, we know (HL + HR )−1
is a PSD map, i.e. if (HL + HR )−1 (A) = B, then, if A is PSD =⇒ B is PSD. Furthermore, M is also a PSD
map, i.e. if A1 is PSD, M(A1 ) = E[(xT A1 x)x ⊗ x] is PSD as well. Finally, HL HR is also a PSD map, since, if
A2 is PSD, then, HL HR (A2 ) = HA2 H which is PSD as well. With all these facts in place, we note that each
term in the Taylor’s expansion above is a PSD map implying the overall map is PSD as well, thus rounding up
the proof to claim 1 in Lemma 15.
Proof of claim 2 in Lemma 15: Since we know that Tb −1 is a PSD map, it suffices to prove the following
stronger statement where we replace the condition W 0 with Tb −1 W 0.
Tr Tb −1 W ≤
1
λmin
· Tr (W) ∀ W ∈ S(d), Tb −1 W 0.
This is equivalent to showing that
Tr (U) ≤
Since Tb U = HU + UH − γb MU −
1
· Tr (Tb U) ∀ U ∈ S(d), U 0.
λmin
γ(b−1)
HUH,
b
(20)
we bound the following quantities:
h
i
h
i
2
2
Tr (MU) = E kxk x> Ux = hE kxk xx> , Ui ≤
2
γ1,max
hH, Ui,
where the last step follows from Lemma 14 and the fact that U 0. We also have
Tr (HUH) = hH, H1/2 UH1/2 i ≤ kHk2 Tr H1/2 UH1/2 = kHk2 Tr (HU) .
Using (21) and (22), we have
γ(b − 1)
γ
Tr (MU) −
Tr (HUH)
b
b
γ kHk2 (b − 1)
γ
Tr (HU) −
Tr (HU)
≥ 2 Tr (HU) −
bγ1,max
b
γ kHk2 (b − 1)
γ
= 2−
−
Tr (HU) .
bγ1,max
b
Tr (Tb U) = 2 Tr (HU) −
Since
γ
bγ1,max
<
γb,max
2bγ1,max
≤ 12 , and γ kHk2 ≤ γb,max · kHk2 /2 ≤
Tr (Tb U) ≥
2
kHk2
· kHk2 /2 = 1, we have:
1
λmin
Tr (HU) ≥
· Tr (U) .
2
2
This proves the claim.
Proof of claim 3 in Lemma 15: Again using Lemma 16, we know that
HL −1 + HR −1 − γI W ∗ = H−1 W + WH−1 − γW ∗
3
2
−1
−1
≤ H W ∗ + WH
+ γ kWk∗ ≤
kWk∗ ,
+ γ kWk∗ ≤
∗
λmin
λmin
1
where we used the fact that γ < γb,max /2 ≤ kHk
.
2
Proof of claim 4 in Lemma 15: This follows easily from combining Claims 1, 2 and 3.
Tr HL −1 + HR −1 − γI Tb −1 W ≤ HL −1 + HR −1 − γI Tb −1 W ∗ ( from Lemma 16)
27
(21)
(22)
3
· Tb −1 W ∗ (from Claim 3)
λmin
3
=
· Tr Tb −1 W (from Claim 1)
λmin
6
≤ 2 · Tr (W) (from Claim 2).
λmin
≤
Proof of claim 5 in Lemma 15: Since (I − γTb ) W = (I − γH) W (I − γH) + γ 2 /b (MW − HWH), it suffices
to show that (MW − HWH) 0. So we need to show
(MW − HWH) 0
⇔MW HWH
⇔u> (MW)u ≥ u> HWHu ∀ u ∈ Rd
⇔Ex (u> x)2 (x> Wx) ≥ Ex1 ,x2 (x1 > Wx2 )(u> x1 )(u> x2 ) ∀ u ∈ Rd .
This last statement is proved using Cauchy-Schwarz inequality as follows:
E (x1 > Wx2 )(u> x1 )(u> x2 )
h
i
≤ E W1/2 x1 W1/2 x2 u> x1 u> x2
1/2
≤ E (x1 > Wx1 )(x2 > Wx2 )(u> x1 )2 (u> x2 )2
= E (u> x)2 (x> Wx) .
This finishes the proof.
Proof of claim 6 in Lemma 15: Since in the proof of Claim 2 we proved (20) which says that
Tr (Tb U) ≥ λmin · Tr (U) ∀ U ∈ S(d), U 0,
Claim 6 follows.
Proof of claim 7 in Lemma 15:
−1
We know that the operator (HR + HL ) is a PSD map, i.e, it maps PSD matrices to PSD matrices. Since
−1
A 0, we replace this condition with U = (HR + HL ) A 0 implying, we need to show the following:
Tr (U) =
1
Tr H−1 U ∀ U 0
2
Examining the right hand side, we see the following:
1
1
Tr H−1 U = Tr H−1 (HL + HR )U
2
2
1
= Tr H−1 HU + H−1 UH
2
= Tr (U)
thus wrapping up the proof of claim 8.
Proof of claim 8 in Lemma 15:
We first denote U = HL + HR − γ b−1
b HL HR . Then, we note that the taylor’s expansion holds since
γ < γb,max and hence Tb 0.
γ −1
Σ
Tb −1 Σ = U − M
b
γ
= U−1/2 (I − U−1/2 MU−1/2 )−1 U−1/2 Σ
b
∞
X
γ
= U−1 Σ +
( )i (U−1 M)i (U−1 Σ)
(23)
b
i=1
Before we proceed further, we provide some upper bounds on the spectral properties of the map U−1 as
applied to any psd matrix B 0:
−1
b−1
−1
U B = HL + HR − γ
HL HR
B
b
28
−1
b−1
−1/2
−1/2
(HL + HR )
(HL + HR )−1/2 B
= (HL + HR )
I −γ
HL HR (HL + HR )
b
i
∞
X
b−1
=
(HL + HR )−1 HL HR B̂
γ
b
i=0
−1/2
where, B̂ = (HL + HR )−1 B. Next,
(HL + HR )−1 HL HR B̂ = (HL + HR )−1 HB̂H
1
= B̂H
2
kHk2
B̂
2
implying,
i
∞
X
γ b−1
( )(
) kHk2 B̂
2
b
i=0
!
2b
1
B̂ , γ <
kHk2
=
b−1
b−1
1 − γ 2b kHk2
U−1 B
Next, we establish some lower bounds on the spectral properties of the map U−1 when applied to any psd
matrix B 0. In particular, our main claim is:
(HL + HR )−1 B U−1 B
In order to prove this claim, let us consider C1 = (HL + HR )−1 B, and C2 = U−1 B. Given that B 0, we
know that C1 0 (by Lyapunov’s Theorem), and C2 0 (using an argument similar to claim 1, with the step
2b
size γ < (b−1)kHk
chosen in order to have U 0 for the arguments similar to claim 1 to hold). Then, we
2
establish the claim using the following series of equations:
b−1
HL HR )C2
b
b−1
HC2 H
HC1 + C1 H = HC2 + C2 H − γ
b
b−1
γ
HC2 H = H(C2 − C1 ) + (C2 − C1 )H
b
b−1
γ
HC2 H = (HL + HR )(C2 − C1 )
b
B = (HL + HR )C1 = (HL + HR − γ
γ
b−1
(HL + HR )−1 HC2 H = C2 − C1
b
We know that the left hand side is psd (using Lyapunov’s theorem), and hence, C2 C1 which establishes
the claim. Now, we combine the resulting lower bounds with the upper bounds to obtain:
Tr(HL + HR )−1 B ≤ Tr U−1 B ≤ {
(HL + HR )−1 B
with the condition that γ <
2
≤ U−1 B
2b
(b−1)kHk2 .
2
1
1−
γ(b−1)kHk2
2b
1
≤{
1−
γ(b−1)kHk2
2b
} Tr(HL + HR )−1 B
} (HL + HR )−1 B
2
(24)
(25)
Next, denoting A = U−1 Σ, we have by equation 23:
Tb −1 Σ = A +
∞
X
(U−1 M)i A
i=1
29
(26)
With this in place, we resort to bounding the behavior of equation 26 in a psd sense, with the following
arguments:
U−1 MA = U−1 E(xt Ax)x ⊗ x
kAk2 R2 U−1 H
= kAk2 R2 A0
where, A0 = U−1 H. We note that U−1 is a psd map (following arguments similar to claim 1 in lemma 15
2b
with the step size γ < (b−1)kHk
), which is necessary for the above sequence of inequalities to hold. This
2
implies,
(U−1 M)2 A = U−1 M(U−1 MA)
kAk2 R2 U−1 MA0
kAk2 R2 (kA0 k2 R2 )A0
implying, the ith term in equation 26 is bounded as:
(U −1 M)i A kAk2 R2 (kA0 k2 R2 )i−1 A0
Using these bounds in equation 26 and summing up the resulting geometric series and applying trace on both
sides, we get the following bound:
Tr Tb −1 Σ ≤ Tr U−1 Σ +
γ
b
U−1 Σ
1−
γ
b
R2 Tr U−1 H
kU−1 Hk2 R2
2
,γ<
b
kU−1 Hk2 R2
(27)
To bound the behavior of the first term in equation 27, we employ the bound in equation 24 (which ensures
2b
statistically optimal rates), and is satisfied if γ < (b−1)kHk
. Next, the denominator of the second term
2
in equation 27 holds iff the geometric series converges, which is if γ <
equation 25 is upperbounded as γ <
b
R2 kU−1 Hk2 ,
which by employing
= R2b2 . Finally, for the numerator of the second term
b Tr (HL +HR )−1 Σ
γ < kU
−1 Σk R2 Tr U−1 H , which by employing the bounds
2
2b Tr(HL +HR )−1 Σ
k(HL +HR )−1 Σk2 R2 d . Combining these three bounds, the
b
R2 k(HL +HR )−1 Hk2
in equation 27 to be statistically optimal, we require
in equations 25 and 24, we obtain a bound as γ <
step size that suffices for the statistically optimal rates are:
2b 2b Tr (HL + HR )−1 Σ
2b
,
,
γ < min
(b − 1) kHk2 R2 k(HL + HR )−1 Σk2 R2 d
We also note that the third term is always ≤ the second term, implying,
2b
2b Tr (HL + HR )−1 Σ
γ < min
,
(b − 1) kHk2 k(HL + HR )−1 Σk2 R2 d
In particular, with the choice of step size that is half of this maximum, by employing claim 7, we have
Tr Tb −1 Σ ≤
3
Tr H−1 Σ
2
Lemma 16. Recall that the nuclear norm kUk∗ of a matrix U is defined to be the sum of its singular values. The
following properties are satisfied by the nuclear norm:
• kUAk∗ ≤ kAk2 kUk∗ for all U, A ∈ Rd×d .
• kUk∗ = Tr (U) for all PSD matrices U.
• kUk∗ = max A∈Rd×d hU, Ai for all U ∈ Rd×d . Moreover, the optimizer A∗ satisfies A∗ U = UA∗ 0.
kAk2 ≤1
30
A.3
Proofs of Theorems 2 and 3
A.3.1
Proof of Theorem 2
Proof. Since γ ≤
γb,max
3 ,
we first invoke lemma 7 to get the contraction factor ρb :
ρb ≤ max(1 − 2γ(1 −
γ
γb,max
)λmin (H), −1 + 2γ(1 −
γ
γb,max
)λmax (H) + 2(
γ
γb,max
)2 , 1 − γλmin (H))
We now wish to get a finer bound on the contraction factor, for which, we first define α =
equation 28 we subtract the second term from the first, to yield:
γ
γb,max
(28)
in the
1 − 2γ(1 − α)λmin (H) − (−1 + 2γ(1 − α)λmax (H) + 2α2 )
= 2 − 2αγb,max (1 − α)(λmin (H) + λmax (H)) − 2α2
≥ 2 − 4αγb,max (1 − α)λmax (H) − 2α2
= 2(1 − 2α(1 − α)γb,max λmax (H) − α2 )
≥ 2(1 − 4α(1 − α) − α2 )
= 2(1 − 4α + 3α2 )
= 2(1 − α)(1 − 3α)
≥0
implies that α ≤ 1 and α ≤ 13 , which is indeed true from the statement of the theorem. Once this is indeed
true, we bound the contraction factor from equation 28 as:
ρb ≤ max(1 − 2γ(1 − α)λmin (H), 1 − γλmin (H))
By setting α = 61 , we get,
5
1
1
,1 −
)=1−
9κ
3κ
3κ
With a bound on the contraction factor established explicitly, we now turn towards providing a scalar bound
by taking the innerproduct of each term in equation 2 with H, which helps in establishing the statement of
theorem 2.
ρb ≤ max(1 −
hH, T1 i =
=
≤
=
≤
≤
≤
≤
≤
1
hH, GTb −1 (I − γTb )s E0 i
− s)2 γ 2
1
hTb −1 (I − γTb )s E0 , (HL −1 + HR −1 − γI)Hi
n
( b − s)2 γ 2
2
hTb −1 (I − γTb )s E0 , Ii
( nb − s)2 γ 2
2
Tr Tb −1 (I − γTb )s E0
( nb − s)2 γ 2
√
2 d
Tb −1 (I − γTb )s E0 F
( nb − s)2 γ 2
√
2 d
s
Tb −1 kI − γTb k kE0 kF
( nb − s)2 γ 2
√
s
2 d
e(− 3κ ) kTb −1 kkE0 kF
n
2
2
( b − s) γ
√
s
2 d
e(− 3κ ) kTb −1 k Tr (E0 )
n
2
2
( b − s) γ
√
s
2 d
1
(− 3κ
) kT −1 k
hH,
E
i
e
b
0
γ 2 ( nb − s)2
λmin (H)
( nb
31
√
s
1
4 d
(− 3κ
) kH−1 k
e
hH,
E
i
0
γ 2 ( nb − s)2
λmin (H)
√
s
4 d
= 2 n
e(− 3κ ) hH, E0 i
2
2
γ ( b − s) (λmin (H))
≤
which wraps up bounding hH, T1 i by plugging in the value of γ. We switch our attention to the second term.
hH, T2 i = kHk Tr (T2 )
√
≤ d kHk kT2 kF
√
n/b
ρb
kE0 kF
γλmin (H) nb − s
√
3 exp(−n/3bκ)κ2 d
≤
hH, E0 i,
( nb − s)
≤
d kHk ·
which bounds the second term. Next, we proceed to bounding the third term:
"
#
1
1
−1
−1
(HL + HR ) GTb −1 Σi
hH, I −
hH, T3 i =
n − sb
γ nb − s
"
#
1
1
−1
−1
−1
(HL + HR ) Hi
=
hΣ, Tb G I −
n − sb
γ nb − s
1
hΣ, Tb −1 (HL −1 + HR −1 − γI)Hi
n − sb
2
hΣ, Tb −1 Ii
≤
n − sb
2
hI, Tb −1 Σi
=
n − sb
3 Tr H−1 Σ
≤
n − sb
≤
Next, we proceed to the fourth term:
2
−1
−1 −1
2 hH, HL HR Tb Σi
γb
−s
2
−1
−1
−1
=
2 hΣ, Tb HR HL Hi
γb nb − s
2
−1 −1
=
2 hΣ, Tb H i
n
γb b − s
2
−1
≤
2 hI, Tb Σi
γbλmin (H) nb − s
hH, T4 i =
≤
n
b
3 Tr H−1 Σ
n
b
γbλmin (H)
2
−s
which wraps up bounding hH, T4 i. We consider the final term:
hH, T5 i = kHk Tr (T5 )
√
≤ d kHk kT5 kF
≤
≤
n
−s+1
√
ρbb
d kHk ·
kΣkF
(n − bs)λmin (H)λmin (Tb )
√
2ρbb
d kHk ·
kΣkF
(n − bs)λmin (H)2
n
32
−s+1
(29)
n
−s+1
√
ρb
≤ 2 dκ2 · b
Tr H−1 Σ
(n − bs)
which wraps up bounding all terms to establish the proof of Theorem 2.
A.3.2
Proof of Theorem 3
n
− 1.
Proof. We first prove a recursive bound on the error R(we ) in epoch e using Theorem 2 for e ≤ log bt
2
−1
R
d
(H
+H
)
Σ
k
k
L
R
Since in epoch e, we use a minibatch size of be = 2e−1 b = 2e−1 kHk Tr(HL +HR )−1 Σ2 , number of initial iterations
2
def
s = t − 1 and total samples ne = be t, we obtain from Theorem 2 the following bound on the error of the
iterate we :
√
2√
s
ne
36 dκ2
3κ d
∗
∗
{R(we−1 ) − R(w )} + exp −
R(we ) − R(w ) ≤
{R(we−1 ) − R(w∗ )}
2 exp −
ne
3κ
3b
κ
−s
ne
e
b
e
be − s
!
( nbee − s + 1)
2κ2 Tr H−1 Σ
9κ Tr H−1 Σ
3 Tr H−1 Σ
+
exp −
+
2 +
ne − be s
(ne − be s)
3κ
be nbee − s
√
√
3 dκ2 Tr H−1 Σ
t
≤ 54 dκ2 exp −
{R(we−1 ) − R(w∗ )} +
using be = 2e−1 b, ne = be t, s = t − 1
e−1
3κ
2 b
√ 2
t
3 dκ Tr H−1 Σ
≤ exp −
( since t ≥ 16κ log (dκ)) .
{R(we−1 ) − R(w∗ )} +
3κ log(dκ)
2e−1 b
By expanding out the recursion, we obtain:
√
e
X
te
1
3 dκ2 Tr H−1 Σ
∗
R(we ) − R(w ) ≤ exp −
· e−`
{R(w0 ) − R(w )} +
`−1
3κ log(dκ)
2 b
16
`=1
√
e
te
3σ 2 dκ2 Tr H−1 Σ X 1
= exp −
{R(w0 ) − R(w∗ )} +
3κ log(dκ)
2e−1 b
8e−`
`=1
√
te
3 dκ2 Tr H−1 Σ 8
∗
·
≤ exp −
{R(w0 ) − R(w )} +
3κ log(dκ)
2e−1 b
7
√
2
−1
te
16 dκ Tr H Σ
≤ exp −
{R(w0 ) − R(w∗ )} +
.
3κ log(dκ)
2e+1 b
∗
(30)
n
In the final step of Algorithm 2, we have a minibatch size of b = 2t
, number of samples n = n/2 and
t
number of initial iterations s = 2 . Plugging this into Theorem 2 again, we obtain the following bound on the
final error.
√
√
t
144κ2 d
t
6κ2 d
∗
∗
−
R(w) − R(w ) ≤
exp
−
{R(w
)
−
R(w
)}
+
exp
{R(we ) − R(w∗ )}
e
t2
6κ
t
3κ
!
√
( nb − s + 1)
3 Tr H−1 Σ 9κ Tr H−1 Σ 2κ2 d Tr H−1 Σ
+
+
exp −
2 +
3κ
n − bs
(n − bs)
b nb − s
√
√
t
4 Tr H−1 Σ
18κ
t
≤ 3 dκ2 exp −
{R(we ) − R(w∗ )} +
3+
+ 2κ2 d exp −
3κ
n
t
6κ
n
using b = , n = n/2, s = t/2
2t
√ 2
t
20 Tr H−1 Σ
≤ 3 dκ exp −
{R(we ) − R(w∗ )} +
( since t ≥ 16κ log (dκ)) .
(31)
3κ
n
33
n
Since in the final iteration, we have e = log bt
− 1, using (30), we have
√
te
16 dκ2 Tr H−1 Σ
∗
∗
R(we ) − R(w ) ≤ exp −
{R(w0 ) − R(w )} +
3κ log(dκ)
2e+1 b
√ 2
t
3κ log(dκ)
2bt
16t dκ Tr H−1 Σ
{R(w0 ) − R(w∗ )} +
≤
.
n
n
Plugging the above inequality into (31), we obtain
√
t
20 Tr H−1 Σ
R(w) − R(w∗ ) ≤ 3 dκ2 exp −
{R(we ) − R(w∗ )} +
3κ
n
!
√
t
3κ log(dκ)
√ 2
t
2bt
16t dκ2 Tr H−1 Σ
20 Tr H−1 Σ
∗
{R(w0 ) − R(w )} +
≤ 3 dκ exp −
+
3κ
n
n
n
t
3κ log(dκ)
2bt
40 Tr H−1 Σ
≤
{R(w0 ) − R(w∗ )} +
( since t > 16κ log(dκ)) .
n
n
This proves the theorem.
A.4
Proof of Lemma 4
For this problem instance, we begin by noting that (HL + HR )−1 Σ is diagonal as well, with entries:
(
1/2
if i = 1
1
−1
−1
{(HL + HR ) Σ}ii = {H Σ}ii =
2
1/2(d − 1)
if i > 1
Let us consider the case with batch size b = 1. With the appropriate choice of step size γ that ensure
contracting operators, we require considering Tr[Tb −1 Σ] as in equation 29, which corresponds to bounding
the leading order term in the variance. We employ the taylor’s expansion (just as in claim 1 of lemma 15) to
expand the term of interest Tb −1 Σ:
Tb −1 Σ =
∞
X
i
γ(HL + HR )−1 M (HL + HR )−1 Σ
i=0
= (HL + HR )−1 Σ +
∞
X
i
γ(HL + HR )−1 M (HL + HR )−1 Σ
i=1
∞
X
⇒ Tr Tb −1 Σ = Tr(HL + HR )−1 Σ +
Tr
h
i
i
γ(HL + HR )−1 M (HL + HR )−1 Σ
i=1
∞
h
i
X
i
1
Tr γ(HL + HR )−1 M (HL + HR )−1 Σ
Tr Tb −1 Σ = Tr H−1 Σ +
2
i=1
We observe that the term corresponding to i = 0 works out regardless of the choice of stepsize γ; we then
switch our attention to the second term, i.e., the term corresponding to i = 1:
d+2
Tr γ(HL + HR )−1 M (HL + HR )−1 Σ =
· Tr Σ
4
We require that this term should be ≤ Tr(HL + HR )−1 Σ, implying,
γ<
4 Tr(HL + HR )−1 Σ
(d + 2) Tr Σ
For this example, we observe that this yields γ <
4
1 ,
(d+2)(1+ d
)
d
1 ,
(d+2)(1+ d
)
which clearly is off by a factor d compared to
the realizable case which requires γ <
thus indicating a clear separation between the step sizes
required by SGD for the realizable and agnostic noise cases.
34
| 8cs.DS |
Uncountable locally free groups
and their group rings
arXiv:1601.00295v1 [math.GR] 3 Jan 2016
TSUNEKAZU NISHINAKA
University of Hyogo
8-2-1 Gakuen Nishimachi Nishiku
Kobe-City 651-2197 Japan
Email: nishinaka@econ.u-hyogo.ac.jp
Abstract. In this note, we show that an uncountable locally free group, and
therefore every locally free group, has a free subgroup whose cardinality is the
same as that of G. This result directly improve the main result in [4] and establish
the primitivity of group rings of locally free groups.
1. INTRODUCTION
A group G is called locally free if all of its finitely generated subgroups are free. As a consequence of Nielsen-Schreier theorem, a free
group is always locally free. If the cardinality |G| of G is countable,
then G is locally free if and only if G is an ascending union of free
groups. In particular, G is a locally free group which is not free provided that it is a properly ascending union of non-abelian free groups
of bounded finite rank. In fact, in this case, G is infinitely generated
and Hopfian and so it is not free (also see [5] and [1]). If |G| is uncountable, that is |G| > ℵ0 , then it was studied in the context of almost free
groups, and it is also known that there exists an uncountable locally
free group which is not free ([2]).
Now, clearly, if G is a locally free group with |G| = ℵ0 , then G has
a free subgroup whose cardinality is the same as that of G. In the
present note, we shall show that it is true for locally free groups of any
cardinality. In fact, we shall prove the following theorem:
Theorem 1.1. If G is a locally free group with |G| > ℵ0 , then for each
finitely generated subgroup A of G, there exists a subgroup H of G with
|H| = |G| such that AH ≃ A ∗ H, the free product of A and H.
In particular, G has a free subgroup of the same cardinality as that
of G.
2000 Mathematics Subject Classification: 20E05, 20E25, 20C07
Key words and phrases: uncountable locally free group, primitive group ring.
This research was partially supported by Grants-in-Aid for Scientific Research
(KAKEN) under grant no. 26400055
1
2
Theorem 1.1 means that there is no need to assume the existence of
free subgroups in [4, Theorem 1]. That is, we can improve the theorem
and establish the primitivity of group rings of locally free groups, where
a ring R is (right) primitive provided it has a faithful irreducible (right)
R-module.
2. Proof of the theorem
In order to prove Theorem 1.1, we prepare necessary notations and
some lemmas which include a result due to Mal’cev [3]. Some of them
might be trivial for experts but we include their proofs for completeness.
For a finitely generated subgroup H of a locally free group G, µG (H)
is defined to be the least positive integer m such that H ⊆ Fm for some
free subgroup Fm of rank m > 0 in G. The rank r(G) of G is defined
to be the maximum element in
{µG (H) | H is a finitely generated subgroup in G}
or r(G) = ∞. We should note that r(G) is finite if and only if for each
finitely generated subgroup H of G, there exists a free subgroup N of
rank r(G) such that H ⊆ N. We should also note that for subgroups
H, M and N of G with H ⊆ M ⊆ N, µM (H) ≥ µN (H) holds.
Lemma 2.1. Let G be a locally free group, D a finitely generated subgroup of G with µG (D) = m and M a subgroup of G such that D ⊆ M
and r(M) = m. For g ∈ G \ M, set Mg = Mhgi; the subgroup of G
generated by g and the elements in M, and let r(Mg ) = n.
Then, if n > m, then n = m + 1 and there exists a free subgroup Fm
of rank m in M such that D ⊆ Fm and Fm hgi is isomorphic to the free
product Fm ∗ hgi.
Proof. Since r(Mg ) = n, there exists a finitely generated subgroup C of
Mg such that µMg (C) = n. Since C is finitely generated, it can be easily
seen that there exists finite number of elements a1 , . . . , al in M such
that C ⊆ ha1 , . . . , al , gi. We have then that there exists a free subgroup
Fm of rank m in M such that ha1 , . . . , al iD ⊆ Fm because of r(M) = m.
Since C ⊆ Fm hgi and µMg (C) = n, we see that r(Fm hgi) ≥ n. On the
other hand, r(Fm hgi) ≤ m + 1 because of g 6∈ Fm . Combining these
with the assumption n > m, we get that n = m + 1 = r(Fm hgi), which
implies Fm hgi ≃ Fm ∗ hgi.
For a locally free group of finite rank, the following result due to
Mal’cev is well-known.
Lemma 2.2. (See [3]) If G is a locally free group of finite rank, then
the cardinality of G is countable; namely |G| = ℵ0 .
3
On the other hand, if G is not of finite rank, then we have the
following property:
Lemma 2.3. If G is a locally free group whose rank is not finite, then
for each finitely generated subgroup A of G, there exists an element
x ∈ G with x 6∈ A, such that Ahxi ≃ A ∗ hxi, the free product of A and
hxi.
Proof. Let A be a finitely generated subgroup of G. We have then that
A is a free group of finite rank, because G is locally free. Since the rank
of G is not finite, there exists a free subgroup F of G such that A ( F
and r(F ) > r(A). Let A = hy1, . . . , yl i and F = hx1 , . . . , xm i, where
l = r(A) and m = r(F ). If for A1 = hy1, . . . , yl , x1 i, r(A1 ) = l + 1,
then A1 ≃ A ∗ hx1 i. If r(A1 ) ≤ l then there exists i ∈ {2, . . . , m} such
that r(Ai ) = r(Ai−1 ) + 1, where Ai = hy1, . . . , yl , x1 , . . . , xi i. We have
then that Ai ≃ Ai−1 ∗ hxi i. Since A ⊆ Ai−1 , we have thus seen that
Ahxi ≃ A ∗ hxi for some x ∈ G \ A.
Let G = A ∗ B be the free product of A 6= 1 and B 6= 1. Clearly,
if |G| = ℵ0 , then G has a free subgroup whose cardinality is the same
as that of G. If |G| > ℵ0 , then either |A| = |G| or |B| = |G|, say
|A| = |G|. Let I be a set with |I| = |A|, and for each i ∈ I, let ai be
in A such that ai 6= aj for i 6= j. We have then that for 1 6= b ∈ B,
the elements (ai b)2 over i ∈ I freely generate the subgroup of G whose
cardinality is the same as that of G. Hence we have
Lemma 2.4. If G = A ∗ B is the free product of A and B, then G has
a free subgroup whose cardinality is the same as that of G.
We are now read to prove Theorem 1.1.
Proof of Theorem 1.1. Let A be a finitely generated subgroup of G.
We set
B = {B | B is a non-trivial subgroup of G such that AB ≃ A ∗ B}.
Since G is locally free, A is a free group of finite rank. By assumption,
|G| > ℵ0 , and so the rank of G is not finite by Lemma 2.2. Hence, by
Lemma 2.3, there exists an element g ∈ G \ A such that Ahgi ≃ A ∗ hgi,
whence hgi ∈ B; thus B =
6 ∅.S Let B1 ⊆ B2 ⊆ · · · ⊆ Bi ⊆ · · · be a chain
∗
of Bi ’s in B, and let B ∗ = ∞
i=1 Bi . We can see that B belongs to B.
∗
∗
In fact, if not so, then AB 6≃ A ∗ B , and so there exists a finitely
generated subgroup C of B ∗ such that AC 6≃ A ∗ C. However, because
C is finitely generated in B ∗ , we have C ⊆ Bi for some i, which implies
ABi 6≃ A ∗ Bi , a contradiction. We have thus shown that (B, ⊆) is
an inductively ordered set. By Zorn’s lemma, there exists a maximal
4
element H in (B, ⊆). We shall show |H| = |G|, which completes the
proof of the theorem. In fact, AH ≃ A ∗ H and it has also a free
subgroup whose cardinality is the same as that of G by Lemma 2.4.
Suppose, to the contrary, that |H| < |G|. Set N = AH(≃ A∗H), and
for a finitely generated subgroup D of N with µG (D) = m, let M(D)
be the set of subgroups M of G such that D ⊆ M and r(M) = m.
We can see that (M(D), ⊆) is an inductively ordered set as follows:
Since m = µG (D), there exists a free subgroup Fm of rank m in G such
that D ⊆ Fm . Hence Fm ∈ M(D), whence M(D) 6= ∅. Let
S∞M1 ⊆
∗
M2 ⊆ · · · ⊆ Mi ⊆ · · · be a chain of Mi ’s in M, and let M = i=1 Mi .
Clearly, D ⊆ M ∗ . By the definition of the rank, r(M ∗ ) ≥ µM ∗ (D).
Since M ∗ ⊆ G, we have that µM ∗ (D) ≥ µG (D) = m; thus r(M ∗ ) ≥ m.
On the other hand, by the definition of r(M ∗ ), there exists a finitely
generated subgroup C of M ∗ such that µM ∗ (C) = r(M ∗ ). Since C
is finitely generated, there exists i > 0 such that C ⊆ Mi , and then
µMi (C) ≤ r(Mi ), which implies that r(M ∗ ) = µM ∗ (C) ≤ µMi (C) ≤
r(Mi ) = m; thus r(M ∗ ) ≤ m. Hence we have r(M ∗ ) = m. We have
thus proved that M ∗ ∈ M(D) and that (M(D), ⊆) is an inductively
ordered set.
Again by Zorn’s lemma,
S there exists a maximal element M(D) in
(M(D), ⊆). Let L = D∈D M(D), where D is the set consisting of
all finitely generated subgroups of N. Since r(M(D)) is finite for each
D ∈ D, it follows from Lemma 2.2 that |M(D)| = ℵ0 for each D ∈ D.
Hence we have |L| < |G| because |D| = |N| < |G|. In particular, there
exists g ∈ G such that g 6∈ L. Note that g 6∈ N because of N ⊂ L, and
so g 6∈ H. We shall show that Nhgi ≃ N ∗ hgi. In order to do this, it
suffices to show that for each D ∈ D, Dhgi ≃ D ∗ hgi holds.
Let r(M(D)) = m for D ∈ D and let Mg = M(D)hgi. Since D ⊆ Mg
and µMg (D) ≥ µG (D) = m, it follows that r(Mg ) ≥ m. Moreover,
since g 6∈ M(D), we have M(D) ( Mg . Hence the maximality of
M(D) implies r(Mg ) > m. It follows from Lemma 2.1 that there
exists a free subgroup Fm of rank m in M(D) such that D ⊆ Fm and
Fm hgi ≃ Fm ∗ hgi. Hence we have Dhgi ≃ D ∗ hgi.
We have thus shown that Nhgi ≃ N ∗ hgi, which contradicts the
maximality of H because N ∗ hgi = A ∗ (H ∗ hgi). This completes the
proof of the theorem.
Theorem 1.1 shows that the assumption on existence of free subgroups in [4, Theorem 1] can be dropped. That is, we have the following
theorem:
5
Theorem 2.5. Let G be a non-abelian locally free group. If R is a
domain with |R| ≤ |G|, then RG is primitive.
In particular, KG is primitive for any field K.
References
[1] G. Higman, A finitely related group with an isomorphic factor group, J. London
Math.Soc., 26(1951), 59-61
[2] G. Higman, Almost free groups, Proc. London Math.Soc., 1(3)(1951), 284-290
[3] A. I. Mal’cev, On groups of finite rank, Mat. Sbornik N.S., 22(64)(1948), 351352.
[4] T. Nishinaka, Group rings of countable non-abelian locally free groups are primitive, Int. J. algebra and computation, 21(3) (2011), 409-431
[5] M. Takahashi, Note on locally free groups, Osaka, City, Univ., 1(1950), 65-70.
| 4math.GR |
A Unified Framework for Pricing Credit and
Equity Derivatives
arXiv:0712.3617v2 [cs.CE] 20 Sep 2008
Erhan Bayraktar
‡
∗†
Bo Yang
Abstract
We propose a model which can be jointly calibrated to the corporate bond term
structure and equity option volatility surface of the same company. Our purpose is to
obtain explicit bond and equity option pricing formulas that can be calibrated to find
a risk neutral model that matches a set of observed market prices. This risk neutral
model can then be used to price more exotic, illiquid or over-the-counter derivatives. We
observe that the model implied credit default swap (CDS) spread matches the market
CDS spread and that our model produces a very desirable CDS spread term structure.
This is observation is worth noticing since without calibrating any parameter to the CDS
spread data, it is matched by the CDS spread that our model generates using the available
information from the equity options and corporate bond markets. We also observe that
our model matches the equity option implied volatility surface well since we properly
account for the default risk premium in the implied volatility surface. We demonstrate
the importance of accounting for the default risk and stochastic interest rate in equity
option pricing by comparing our results to Fouque et al. (2003), which only accounts for
stochastic volatility.
Keywords: Credit Default Swap, Defaultable Bond, Defaultable Stock, Equity Options,
Stochastic Interest Rate, Implied Volatility, Multiscale Perturbation Method.
Contents
1 Introduction
2
∗
Department of Mathematics, University of Michigan, Ann Arbor, MI 48109, USA, email:{erhan,boy}@umich.edu.
†
We would like to thank the two anonymous referees and the anonymous AE for their constructive comments, which helped us improve our paper.
‡
This work is supported in part by the National Science Foundation, under grant DMS-060449.
1
2 A Framework for Pricing Equity and Credit Derivatives
2.1
The model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5
2.2
Equity and credit derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . .
7
3 Explicit Pricing Formulas for Credit and Equity Derivatives
9
3.1
Pricing equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9
3.2
Asymptotic expansion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10
3.3
Explicit pricing formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
13
4 Calibration of the Model
1
5
18
4.1
Data description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
19
4.2
The parameter estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
20
4.3
Model implied CDS spread matches the observed CDS spread . . . . . . . . .
22
4.4
Fitting Ford’s implied volatility . . . . . . . . . . . . . . . . . . . . . . . . . .
22
4.5
Fitting the implied volatility of the index options . . . . . . . . . . . . . . . .
23
Introduction
Our purpose is to build an intensity-based modeling framework that can be used in trading
and calibrating across the credit and equity markets. The same company has stocks, stock
options, bonds, credit default swaps on these bonds, and several other derivatives. When this
company defaults, the payoffs of all of these instruments are affected; therefore, their prices
all contain information about the default risk of the company.
We build a model that can be jointly calibrated to corporate bond prices and stock
options, and can be used to price more exotic derivatives. In our framework we use the
Vasicek model for the interest rate, and use doubly stochastic Poisson process to model the
default of a given company. We assume that the bonds have recovery of market value and
that stocks become valueless at the time of default. Using the multi-scale modeling approach
of Fouque et al. (2003) we obtain explicit bond pricing equation with three free parameters
which we calibrate to the corporate bond term structure. On the other hand, stock option
pricing formula contain seven parameters, three of which are common with the bond option
pricing formula. (The common parameters are multiplied with the loss rate in the bond
2
pricing formula.) We calibrate the remaining set of parameters to the stock option prices.
This hybrid model, therefore, is able to account for the default risk premium in the implied
volatility surface.
The calibration results reveal that our model is able to produce implied volatility surfaces
that match the data closely. We compare the implied volatility surfaces that our model
produces to those of Fouque et al. (2003). We see that even for longer maturities our model
has a prominent skew: compare Figures 5 and 6. Even when we ignore the stochastic volatility
effects, our model fits the implied volatility of the Ford Motor Company well and performs
better than the model of Fouque et al. (2003); see Figure 4. This points to the importance
of accounting for the default risk for companies with low ratings.
Once the model parameters are calibrated, the model can be used to compute the prices
of more exotic options. To test whether our model produces correct prices we use the CDS
spread data and show that the model implied CDS spread matches the “out of sample” CDS
data. To compute the CDS spread, under our assumption on the recovery, one needs to
reconstruct the term structure of the treasury and the corporate bonds. Moreover, one needs
to separate the loss rate from the other parameters in the bond pricing formula (see (2.14) or
(4.7) for the CDS spread formula). This separation is possible since we calibrate our model
to corporate bond data and stock option data jointly as described above. The model-implied
CDS spread time series matches the observed CDS spread time series of Ford Motor Company
for over a long period of time; see Figures 1 and 2. This is an interesting observation since
we did not make use of the CDS spread data in our calibration. This observation also shows
that one can use our model to trade across different markets that contain information about
the default risk of a given firm.
Our model has three building blocks: (1) We model the default event using the multi-scale
stochastic intensity model of Papageorgiou and Sircar (2008). We also model the interest rate
using an Ornstein-Uhlenbeck process (Vasicek model). As it was demonstrated in Papageorgiou and Sircar (2008), these modeling assumptions are effective in capturing the corporate
yield curve; (2) We take the stock price process to follow a stochastic volatility model which
jumps to zero when the company defaults. This stock price model was considered in Bayraktar (2008). Our model specification for the stock price differs from the jump to default models
for the stock price considered by Carr and Linetsky (2006) and Linetsky (2006), which take
the volatility and the default intensity to be functions of the stock price; (3) We also account
for the stochastic volatility in the modeling of the stocks since even the index options (when
there is no risk of default) possess implied volatility skew. We model the volatility using
the fast scale stochastic volatility model of Fouque et al. (2000). We demonstrate on index
options (when there is no risk of default) that (see Section 4.5), we match the performance
3
of the two time scale volatility model Fouque et al. (2003). The latter model extends Fouque
et al. (2000) by including a slow factor in the volatility to get a better fit to longer maturity
option. We see from Section 4.5 that when one assumes the interest rate to be stochastic, the
calibration performance of the stochastic volatility model with only the fast factor is as good
as the two scale stochastic volatility model, which is why we choose the volatility to be driven
by only the fast factor. Even though the interest rate is stochastic in our model, we are able
to obtain explicit asymptotic pricing formulas for stock options. Thanks to these explicit
pricing formulas the inverse problem that we face in calibrating to the corporate bond and
stock data can be solved with considerable ease. Our modeling framework can be thought of
as a hybrid of the models of Fouque et al. (2000), which only considers pricing options in a
stochastic volatility model with constant interest rate, and Papageorgiou and Sircar (2008),
which only considers a framework for pricing derivatives on bonds. Neither of these models
has the means to transfer information from the equity markets to bond market or vice versa,
which we are set to do in this paper. We should also note that our model also takes input
from the treasury yield curve, historical stock prices, and historical spot rate data to estimate
some of its parameters (see Section 4).
Our model extends Bayraktar (2008) by taking the interest rate process to be stochastic,
which leads to a richer theory and more calibration parameters, and therefore, better fit
to data: (i) When the interest rate is deterministic the corporate bond pricing formula
turns out to be very crude and does not fit the bond term structure well (compare (2.57)
in Bayraktar (2008) and (4.1)); (ii) With deterministic interest rates the bond pricing and
the stock option pricing formulas share only one common term, “the average intensity of
default” (this parameter is multiplied by the loss rate in the bond pricing equation, under
our loss assumptions). Therefore, the default premium in the implied volatility surface is not
accounted for as much as it should be. And our calibration analysis demonstrates that this
has a significant impact. When the volatility is taken to be constant, both our new model
and the model in Bayraktar (2008) have three free parameters. The model in Bayraktar
(2008) produces a below par fit to the implied volatility surface (see e.g. Figure 5 in that
paper), whereas our model produces an excellent fit (see Section 4.4 and Figure 4); (iii) To
calculate the CDS spread, in the constant interest rate model, one needs to separate the loss
rate and the average intensity of default. This is again established calibrating the model to
the bond term structure data and the stock option implied volatility surface. The estimates
for the average intensity and the loss rate are not as accurate in Bayraktar (2008) as it is in
our model because of (i) and (ii). This crude estimation leads to a poor out of sample match
to the CDS spread time series.
The other defaultable stock models are those of Carr and Linetsky (2006), Linetsky (2006)
and Carr and Wu (2006), which assume that the interest rate is deterministic. Carr and
4
Linetsky (2006), Linetsky (2006) take the volatility and the intensity to be functions of the
stock price and obtain a one-dimensional diffusion for the pre-default stock price evolution.
Using the fact that the resolvents of particular Markov processes can be computed explicitly,
they obtain pricing formulas for stock option prices. On the other hand Carr and Wu (2006)
uses a CIR stochastic volatility model and also models the intensity to be a function of
the volatility and another endogenous CIR factor. The option prices in this framework are
computed numerically using inverse the Fourier transform. We, on the other hand, use
asymptotic expansions to provide explicit pricing formulas for stock options in a framework
that combines a) the Vasicek interest rate model, b) fast-mean reverting stochastic volatility
model, c) defaultable stock price model, d) multi-scale stochastic intensity model.
Our calibration exercise differs from that of Carr and Wu (2006) since they perform a
time series analysis to obtain the parameters of the underlying factors (from the the stock
option prices and credit default swap spread time series), whereas we calibrate our pricing
parameters to the daily implied volatility surface and bond term structure data. Our purpose
is to find a risk neutral model that matches a set of observed market prices. This risk neutral
model can then be used to price more exotic, illiquid or over-the-counter derivatives. For
further discussion of this calibration methodology we refer to Cont and Tankov (2004) (see
Chapter 13), Fouque et al. (2000), Fouque et al. (2003) and Papageorgiou and Sircar (2008).
We also provide daily prediction of the CDS spread only using the data from the bond term
structure and implied volatility surface of the options.
The rest of the paper is organized as follows: In Section 2, we introduce our modeling
framework and describe the credit and equity derivatives we will consider and obtain an
expression for the CDS spread under the assumption that the recovery rate of a bond that
defaults is a constant fraction of its predefault value. In Section 3, we introduce the asymptotic expansion method. We obtain explicit (asymptotic) prices for bonds and equity options
in Section 3.3. In Section 4, we describe the calibration of our parameters and discuss our
empirical results. Figures, which show our calibration results, are located after the references.
2
2.1
A Framework for Pricing Equity and Credit Derivatives
The model
Let (Ω, H, P) be a complete probability space supporting (i) correlated standard Brownian
~ t = (W 0 , W 1 , W 2 , W 3 , W 4 ), t ≥ 0, with
motions W
t
t
t
t
t
E[Wt0 , Wti ] = ρi t,
E[Wti , Wtj ] = ρij t,
5
i, j ∈ {1, 2, 3, 4}, t ≥ 0,
(2.1)
~ . Let us
for some constants ρi , ρij ∈ (−1, 1), and (ii) a Poisson process N independent of W
Rt
introduce the Cox process (time-changed Poisson process) Ñt , N ( 0 λs ds), t ≥ 0, where
λt = f (Yt , Zt ),
√
1
ν 2
dYt = (m − Yt )dt + √ dWt2 , Y0 = y,
√
dZt = δc(Zt )dt + δg(Zt )dWt3 , Z0 = z,
(2.2)
in which , δ are (small) positive constants and f is a strictly positive, bounded, smooth
function. We also assume that the functions c and g satisfy Lipschitz continuity and growth
conditions so that the diffusion process for Zt has a unique strong solution. We model the
time of default as
τ = inf{t ≥ 0 : Ñt = 1}.
(2.3)
We also take interest rate to be stochastic and model it as an Ornstein-Uhlenbeck process
drt = (α − βrt )dt + ηdWt1 ,
r0 = r,
(2.4)
for positive constants α, β, and η.
We model the stock price as the solution of the stochastic differential equation
Z
dX̄t = X̄t rt dt + σt dWt0 − d Ñt −
t∧τ
λu du
,
X̄0 = x,
(2.5)
0
where the volatility is stochastic and is defined through
σt = σ(Ỹt );
dỸt =
!
√
√
1
ν̃ 2
ν̃ 2
(m̃ − Ỹt ) − √ Λ(Ỹt ) dt + √ dWt4 ,
Ỹ0 = ỹ.
(2.6)
Here, Λ is a smooth, bounded function of one variable which represents the market price of
volatility risk. The function σ is also a bounded, smooth function. Note that the discounted
stock price is a martingale under the measure P, and at the time of default, the stock price
jumps down to zero. The pre-banktruptcy stock price coincides with the solution of
dXt = (rt + λt )Xt dt + σt Xt dWt0 ,
X0 = x.
(2.7)
It will be useful to keep track of different flows of information. Let F = {Ft , t ≥ 0} be the
~ . Denote the default indicator process by It = 1{τ ≤t} , t ≥ 0, and let
natural filtration of W
I = {It , t ≥ 0} be the filtration generated by I. Finally, let G = {Gt , t ≥ 0} be an enlargement
of F such that Gt = Ft ∨ It , t ≥ 0.
6
Since we will take and δ to be small positive constants, the processes Y and Ỹ are
fast mean reverting, and Z evolves on a slower time scale. See Fouque et al. (2003) for
an exposition and motivation of multi-scale modeling in the context of stochastic volatility
models.
We note that our specification of the intensity of default coincides with that of Papageorgiou and Sircar (2008), who considered only a framework for pricing credit derivatives.
Our stock price specification is similar to that of Linetsky (2006) and Carr and Linetsky
(2006) who considered a framework for only pricing equity options on defaultable stocks.
Our volatility specification, on the other hand, is in the spirit of Fouque et al. (2000).
Bayraktar (2008) considered a similar modeling framework to the one considered here,
but the interest rate was taken to be deterministic. In this paper, by extending this modeling
framework to incorporate stochastic interest rates, we are able to consistently price credit
and equity derivatives and produce more realistic yield curve and implied volatility surfaces.
We are also able to take the equity option surface and the yield curve data as given and
predict the credit default swap spread on a given day. Testing our model prediction against
real data demonstrates the power of our pricing framework.
2.2
Equity and credit derivatives
In our framework, we will price European options, bonds, and credit default swaps of the
same company in a consistent way.
1. The price of a European call option with maturity T and strike price K is given by
Z T
+
C(t; T, K) = E exp −
rs ds (X̄T − K) 1{τ >T } Gt
t
Z T
+
= 1{τ >t} E exp −
(rs + λs )ds (XT − K) Ft ,
(2.8)
t
in which the equality follows from Lemma 5.1.2 of Bielecki and Rutkowski (2002). (This
lemma, which lets us write a conditional expectation with respect to Gt in terms of conditional
expectations with respect to Ft , will be used in developing several identities below). Also,
see Linetsky (2006) and Carr and Linetsky (2006) for a similar computation.
7
On the other hand, the price of a put option with the same maturity and strike price is
"
!
T
Z
#
"
Z
+
#
rs ds K1{τ ≤T } Gt
rs ds (K − XT ) 1{τ >T } Gt + E exp −
Put(t; T ) = E exp −
!
T
t
t
"
!
T
Z
#
+
(rs + λs )ds (K − XT ) Ft
= 1{τ >t} E exp −
t
"
Z
+ KE exp −
T
!
rs ds
#
"
!
T
Z
Ft − KE exp −
(rs + λs )ds
t
#!
Ft
.
t
(2.9)
2. Consider a defaultable bond with maturity T and par value of 1 dollar. We assume the
recovery of the market value, introduced by Duffie and Singleton (1999). In this model, if
the issuer company defaults prior to maturity, the holder of the bond recovers a constant
fraction 1 − l of the pre-default value, with l ∈ [0, 1]. The price of such a bond is
"
Z
c
B (t; T ) = E exp −
T
Z
rs ds 1{τ >T } + exp −
t
Z
= E exp −
τ
#
rs ds 1{τ ≤T } (1 − l)B c (τ −; T ) Gt
t
T
(rs + l λs )ds Ft ,
t
(2.10)
on {τ > t}, see Duffie and Singleton (1999) and Schönbucher (1998).
3. Consider a credit default swap (CDS) written on B c , which is a insurance against
losses incurred upon default from holding a corporate bond. The protection buyer pays a
fixed premium, the so-called CDS spread, to the protection seller. The premium is paid on
fixed dates T = (T1 , · · · , TM ), with TM being the maturity of the CDS contract. We denote
the CDS spread at time t by cds (t; T ). Our purpose is to determine a fair value for the CDS
spread so that what the protection buyer buyer is expected to pay, the value of the premium
leg of the contract, is equal to what the protection seller is expected to pay, the value of
the protection leg of the contract. For a more detailed description of the CDS contract, see
Bielecki and Rutkowski (2002) or Schönbucher (2003).
The present value of the premium leg of the contract is
"
Premium(t; T ) = cds (t; T ) E
ds
M
X
Z
exp −
m=1
M
X
= 1{τ >t} c (t; T )
t
Tm
#
rs ds 1{τ >Tm } Gt
(2.11)
Z
E exp −
t
m=1
8
Tm
(rs + λs )ds
Ft ,
in which we assumed that t < T1 . The present value of the protection leg of the contract
under our assumption of recovery of market value is
τ
Z
Protection(t; T ) = 1{τ >t} E exp −
t
c
rs ds 1{τ ≤TM } l B (τ −; TM ) Gt
(2.12)
Adding (2.10) and (2.12), we obtain
"
Z
c
Protection(t; T ) + B (t; TM ) = E exp −
T
Z
rs ds 1{τ >T } + exp −
τ
#
rs ds 1{τ ≤T } B c (τ −; T ) Gt
t
t
TM
Z
= 1{τ >t} E exp −
rs ds Ft ,
t
(2.13)
where the last equality is obtained by setting l = 0 in (2.10).
Now, the CDS spread can be determined, by setting Protection(t; T ) = Premium(t; T )
and using equations (2.11) and (2.13), as
cds (t; T ) = 1{τ >t}
R
R
TM
TM
E exp − t rs ds Ft − E exp − t (rs + l λs )ds Ft
M
X
Z
E exp −
(rs + λs )ds
(2.14)
Ft
t
m=1
3
.
Tm
Explicit Pricing Formulas for Credit and Equity Derivatives
3.1
Pricing equation
Let P ,δ denote
P
,δ
Z
(t, Xt , rt , Yt , Ỹt , Zt ) = E exp −
T
(rs + lλs )ds h(XT ) Ft .
(3.1)
t
When l = 1 and h(XT ) = (XT − K)+ , P ,δ is the price of a call option (on a defaultable
stock). On the other hand, when h(XT ) = 1, P ,δ becomes the price of a defaultable bond.
Using the Feynman-Kac formula, we can characterize P ,δ as the solution of
L,δ P ,δ (t, x, r, y, ỹ, z) = 0,
P ,δ (T, x, r, y, ỹ, z) = h(x),
9
(3.2)
where the partial differential operator L,δ is defined as
,δ
L
√
1
1
, L0 + √ L1 + L2 + δM1 + δM2 +
r
δ
M3 ,
(3.3)
in which
2
∂
∂
∂2
∂2
2 ∂
+
(m
−
y)
+
(
m̃
−
ỹ)
+
ν̃
+
2ρ
vṽ
,
24
∂y 2
∂y
∂ ỹ 2
∂ ỹ
∂y∂ ỹ
√
√ ∂2
√
√ ∂2
√ ∂
∂2
∂2
L1 , ρ2 σ(ỹ)ν 2x
+ ρ12 ην 2
+ ρ4 σ(ỹ)ν̃ 2x
+ ρ14 ην̃ 2
− Λ(ỹ)ν̃ 2 ,
∂x∂y
∂r∂y
∂x∂ ỹ
∂r∂ ỹ
∂ ỹ
2
2
2
∂
1
∂
∂
∂
∂
1 ∂
L2 ,
+ σ 2 (ỹ)x2 2 + (r + f (y, z))x
+ (α − βr)
+ σ(ỹ)ηρ1 x
+ η 2 2 − (r + l f (y, z))·,
∂t 2
∂x
∂x
∂r
∂x∂r 2 ∂r
L0 , ν 2
∂2
∂
1
∂2
∂2
+ ηρ13 g(z)
, M2 , c(z)
+ g 2 (z) 2 ,
∂x∂z
∂r∂z
∂z 2
∂z
2
2
√
√
∂
∂
M3 , ρ23 ν 2g(z)
+ ρ34 ν̃ 2g(z)
.
∂y∂z
∂ ỹ∂z
M1 , σ(ỹ)ρ3 g(z)x
3.2
Asymptotic expansion
We construct an asymptotic expansion for P ,δ as , δ → 0. First, we consider an expansion
√
of P ,δ in powers of δ
√
(3.4)
P ,δ = P0 + δP1 + δP2 + · · ·
By inserting (3.4) into (3.2) and comparing the δ 0 and δ terms, we obtain that P0 satisfies
1
1
L0 + √ L1 + L2 P0 = 0,
(3.5)
P0 (T, x, r, y, ỹ, z) = h(x),
and that P1 satisfies
1
1
1
√
√
L0 +
L1 + L2 P1 = − M1 +
M3 P0 ,
(3.6)
P1 (T, x, y, ỹ, z, r) = 0.
Next, we expand the solutions of (3.5) and (3.6) in powers of
P0 = P0 +
√
√
P1,0 + P2,0 + 3/2 P3,0 + · · ·
√
P1 = P0,1 + P1,1 + P2,1 + 3/2 P3,1 + · · ·
10
(3.7)
(3.8)
Inserting the expansion for P0 into (3.5) and matching the 1/ terms gives L0 P0 = 0. We
choose P0 not to depend on y and ỹ because the other solutions have exponential growth at
√
infinity (see e.g. Fouque et al. (2003)). Similarly, by matching the 1/ terms in (3.5) we
obtain that L0 P1,0 + L1 P0 = 0. Since L1 takes derivatives only with respect to y and ỹ, we
observe that L0 P1,0 = 0. We choose P1,0 not to depend on y and ỹ.
Now equating the order-one terms in the expansion of (3.5) and using the fact that
L1 P1,0 = 0, we get that
L0 P2,0 + L2 P0 = 0,
(3.9)
which is a Poisson equation for P2,0 (see e.g. Fouque et al. (2000)). The solvability condition
for this equation requires that
hL2 iP0 = 0,
(3.10)
where h·i denotes the averaging with respect to the invariant distribution of (Yt , Ỹt ), whose
density is given by
(
"
#)
1
y−m 2
ỹ − m̃ 2
(y − m)(ỹ − m̃)
1
exp −
+
− 2ρ24
.
Ψ(y, ỹ) =
2πν ν̃
ν
ν̃
ν ν̃
2(1 − ρ224 )
(3.11)
Let us denote
σ̄1 , hσ(ỹ)i, σ̄22 , hσ 2 (ỹ)i, λ̄(z) = hf (y, z)i.
(3.12)
To demonstrate the effect of averaging on L2 , let us write
∂
∂
∂2
1 2 ∂2
∂ 1 2 2 ∂2
+ σ̄2 x
+
(r
+
λ̄(z))x
+
(α
−
βr)
+
σ̄
ηρ
x
+
η
− (r + l λ̄(z))·
1
1
∂t 2
∂x2
∂x
∂r
∂x∂r 2 ∂r2
(3.13)
Together with the terminal condition
hL2 i :=
P0 (T, x, r, z) = h(x),
(3.14)
equation (3.10) defines the leading order term P0 . On the other hand from (3.9), we can also
deduce that
P2,0 = −L−1
(3.15)
0 (L2 − hL2 i)P0 .
√
Matching the order terms in the expansion of (3.5) yields
L0 P3,0 + L1 P2,0 + L2 P1,0 = 0,
(3.16)
which is a Poisson equation for P3,0 . The solvability condition for this equation requires that
hL2 P1,0 i = −hL1 P2,0 i = hL1 L−1
0 (L2 − hL2 i)iP0 ,
11
(3.17)
which along with the terminal condition
P1,0 (T, x, r, z) = 0,
(3.18)
completely identifies the function P1,0 . To obtain the second equality in (3.17) we used (3.15).
Next, we will express the right-hand side of (3.17) more explicitly. To this end, let ψ, κ,
and φ be the solutions of the Poisson equations
L0 ψ(ỹ) = σ(ỹ) − σ̄1
L0 κ(ỹ) = σ 2 (ỹ) − σ̄22 ,
and L0 φ(y, z) = (f (y, z) − λ̄(z)),
(3.19)
respectively. First observe that
2
1 2
∂ 2 P0
∂P0
2 2 ∂ P0
(L2 −hL2 i)P0 = (σ (ỹ)−σ̄2 )x
+(σ(ỹ)−σ̄1 )ηρ1 x
+l (f (y, z)−λ̄(z)) x
− P0 .
2
∂x2
∂x∂r
∂x
(3.20)
Now, along with (3.19), we can write
L−1
0 (L2
2
1
∂ 2 P0
∂P0
2 ∂ P0
− hL2 i)P0 = κ(ỹ)x
+ ψ(ỹ)ηρ1 x
+ l φ(y, z) x
− P0 .
2
∂x2
∂x∂r
∂x
(3.21)
Applying the differential operator L1 to the last expression yields
hL1 L−1
0 (L2
√
√
∂P0
∂
− hL2 i)iP0 = l ρ2 ν
+ l ρ12 ην 2hφy i
x
− P0
∂r
∂x
2
2
√
1
∂
∂
∂ P0
2 ∂ P0
+ ρ4 ν̃ 2
hσκỹ ix
x
+ hσψỹ iηρ1 x
x
2
2
∂x
∂x
∂x
∂x∂r
2
2
√
∂
∂ P0
∂
∂ P0
1
hκỹ i
x2
+ hψỹ iηρ1
x
+ ρ14 ην̃ 2
2
∂r
∂x2
∂r
∂x∂r
√
1
∂ 2 P0
2 ∂P0
− ν̃ 2
hΛκỹ ix
+ hΛψỹ iηρ1 x
.
2
∂x2
∂x∂r
(3.22)
∂P0
2hσφy ix2 2
∂x
Finally, we insert the expression for P1 in (3.8) into (3.6) and collect the terms with the
same powers of . Arguing as before, we obtain that P0,1 is independent of y and ỹ and
satisfies:
hL2 iP0,1 = −hM1 iP0 ,
12
P0,1 (T, x) = 0.
(3.23)
3.3
Explicit pricing formula
We approximate P ,δ defined in (3.1) by
Pe,δ = P0 +
√
√
P1,0 +
δP0,1 .
(3.24)
Since the Vasicek interest rate process is unbounded, which implies that the potential term
in L2 or the discounting term in (3.1) is unbounded, the arguments of Fouque et al. (2003)
can not be directly used. However as in Cotton et al. (2004) and Papageorgiou and Sircar
(2008), one can write
P
,δ
T
(t, Xt , rt , Yt , Ỹt , Zt ) = B(t, T )E
Z
exp −
T
lλs ds h(XT ) Ft := B(t, T )F ,δ (t, Xt , rt , Yt , Ỹt , Zt ),
t
(3.25)
in which
dPT
dP
=
R
T
exp − 0 rs ds
B(0, T )
and
Z
B(t, T ) = E exp −
T
,
rs ds
(3.26)
Ft .
(3.27)
t
Now, the analysis of Fouque et al. (2003) can be used to approximate F ,δ (t, x, r, y, ỹ, z). As a
result of this analysis for each (t, x, r, y, ỹ, z), there exists a constant C such that |P ,δ − Pe,δ | ≤
√
C · ( + δ) when h is smooth, and |P ,δ − Pe,δ | ≤ C · ( log() + δ + δ) when h is a put or a
call pay-off. In what follows, we will obtain P0 , P1,0 and P0,1 explicitly.
Our first objective is to develop a closed-form expression for P0 , the solution of (3.10)
and (3.14).
Proposition 3.1. The leading order term P0 in (3.24) is given by:
P0 (t, x, z, r) = B0c (t, r; z, T, l)
∞
1
(u − m(t, T ))2
h(exp(u)) p
exp(−
)du,
2v(t, T )
2πv(t, T )
−∞
Z
(3.28)
where
B0c (t, r; z, T, l) , exp − lλ̄(z)(T − t) + a(T − t) − b(T − t)r),
(3.29)
in which the functions a(s) and b(s) are defined as:
a(s) =
α
η2
−
2
2β
β
s+
η2
α
− 2
3
β
β
(exp(−βs) − 1) −
13
η2
(exp(−2βs) − 1)
4β 3
(3.30)
and b(s) = (1 − exp(−βs))/β. On the other hand,
2ηρ1 σ̄1 2η 2
=
(T − t) +
+ 3
β2
β
2
2
η
2ηρ1 σ̄1
3η
− 3 exp(−2β(T − t)) −
+ 3 ,
2
2β
β
2β
vt,T
and
σ̄22
2ηρ1 σ̄1
η2
+
+ 2
β
β
exp(−β(T − t))
(3.31)
1
mt,T = log(x) + λ̄ · (T − t) − a(T − t) + b(T − t)r − v(t, T ).
2
(3.32)
Proof. By applying the Feynman-Kac theorem to (3.10) and (3.14) we have that
T
Z
P0 (t, x, z, r) = E exp −
(rs + lλ̄(z))ds h(ST ) St = x, rt = r ,
(3.33)
t
where the dynamics of S is given by
ft0 ,
dSt = (rt + λ̄(z))St dt + σ̄2 St dW
f 0 is a Wiener process whose correlation with W 1 is ρ̄1 =
in which W
(3.34)
σ̄1
σ̄2 ρ1 .
Let us define
Z
P̃0 (t, x, r) = E exp −
T
rs ds h(SeT ) Set = x, rt = r ,
(3.35)
t
in which
ft0 .
dSet = rt Set dt + σ̄2 Set dW
(3.36)
Then
P0 (t, x, z, r) = e−lλ̄(z) (T −t) P̃0 (t, x exp(λ̄(z)(T − t)), r).
(3.37)
Now, by following Geman et al. (1995), we change the probability measure P to the
forward measure PT through the Radon-Nikodym derivative (3.26)
We can obtain the following representation of P˜0 using the T -forward measure
h
i
P̃0 (t, Set , rt ) = B(t, T )ET h(SeT )|Ft = B(t, T )ET [h(FT )|Ft ] ,
(3.38)
in which
Ft ,
Set
,
B(t, T )
(3.39)
which is a PT martingale. Note that an explicit expression for B(t, T ) is available since rt is
14
a Vasicek model, and it is given in terms of the functions a and b:
B(t, T ) = exp(a(T − t) − b(T − t)rt ).
(3.40)
By applying Itô’s formula to (3.39), we observe that the dynamics of F are
f 0 + b(T − t)ηdW
f 1 ),
dFt = Ft (σ̄1 dW
t
t
(3.41)
f 1 is a PT Brownian motion whose correlation with the W
f 0 (which is still a
in which W
Brownian motion under PT ) is ρ̄1 . Given Xt and B(t, T ), the random variable log FT is
normally distributed with variance
Z T
Z T
b2 (T − s)ds + 2η ρ̄1 σ̄2
b(T − s)ds
vt,T = σ̄22 (T − t) + η 2
t
t
2η ρ̄1 σ̄2
η2
2η ρ̄1 σ̄2 2η 2
2
= σ̄2 +
+ 2 (T − t) +
+ 3 exp(−β(T − t))
β
β
β2
β
2
2
η
2η ρ̄1 σ̄2
3η
− 3 exp(−2β(T − t)) −
+ 3 ,
2
2β
β
2β
(3.42)
and mean
1
m(t, T ) = log Ft −
2
Z
T
(σ̄22 +b2 (T −s)η 2 + ρ̄1 σ̄2 b(T −s)η)ds = log
t
!
Set
1
− vt,T . (3.43)
B(t, T )
2
Now the result immediately follows.
An immediate corollary of the last proposition is the following:
Corollary 3.1. i) When l = 1, h(x) = (x − K)+ , then (3.28) becomes
C0 (t, x, z, r) = xN (d1 ) − KB0c (t, r; z, T, 1)N (d2 ),
(3.44)
in which N is the standard normal cumulative distribution function and
d1,2 =
x
± 21 vt,T
log KB c (t,r;z,T,1)
0
.
√
vt,T
(3.45)
ii) When l = 1, and h(x) = (K − x)+ , then (3.28) becomes
Put0 (t, x, z, r) = −x + xN (d1 ) − KB0c (t, r; z, T, 1)N (d2 ) + KB0c (t, r; z, T, 0).
(3.46)
iii) When h(x) = 1, then (3.28) coincides with (3.30) in Papageorgiou and Sircar (2008).
15
Proposition 3.2. The correction term
√
√
P1,0 is given by
2
2
2 ∂ P0
∂
2 ∂ P0
P1,0 = −(T − t) V1 x
+ V2 x
x
∂x2
∂x
∂x2
∂ 3 P0
∂P0
∂ 2 P0
∂ 2 P0
∂ 2 P0
+ V4 x2 2
−
+ V5 x
+ V6 x
,
+ l V3 −x
∂x∂α
∂α
∂x ∂α
∂η∂x
∂x∂α
(3.47)
in which
√
√ 1
√
1√
l ρ2 ν 2hσφy i − ν̃ 2 hΛκỹ i , V2 =
ρ4 ν̃ 2hσκỹ i,
2
2
√
√
√
√
√
√ 1
2
V3 = (ρ12 ην 2hφy i), V4 = −
ρ14 ην̃ 2hκỹ i − ρ4 ν̃ 2hσψỹ iηρ1 + ρ14 ην̃ 2hψỹ iσ̄1 ρ1 ,
2
√
√
√
√
√
√
V5 = − (ρ14 ην̃ 2hψỹ iρ1 ), V6 = (−ρ4 ν̃ 2hσψỹ iηρ1 + ρ14 ην̃ 2hψỹ iσ̄1 ρ21 − ν̃ 2hΛψỹ iηρ1 ).
V1 =
√
(3.48)
Proof. Recall that P1,0 is the solution of (3.17) and (3.18) and that the right-hand-side of
(3.17) is given by (3.22). The result is a simple algebraic exercise given the following four
observations:
n
∂
1) xn ∂x
n commutes with hL2 i.
n
∂
2) −(T − t)(xn ∂x
n )P0 solves:
n
n ∂
P0 ,
hL2 iu = x
∂xn
u(T, x, r; z) = 0.
(3.49)
0
3) By differentiating (3.14) with respect to α, we see that − ∂P
∂α also solves
hL2 iu =
∂P0
,
∂r
u(T, x, r; z) = 0.
(3.50)
4) Using 1) and 2) above and the equation we obtain by differentiating (3.10) with respect
∂ 2 P0
0
to η, we can show that 1/η · (σ̄1 ρ1 x ∂x∂α
− ∂P
∂η ) solves
hL2 iu =
∂ 2 P0
,
∂r2
u(T, x, r; z) = 0.
(3.51)
Remark 3.1. By differentiating (3.10) with respect to r, we obtain
hL2 i
∂
∂P0
∂P0
= −x P0 + β
+ P0 .
∂r
∂x
∂r
16
(3.52)
∂P0
0
Using observation 2 in the proof of Proposition 3.2, we see that β1 −(T − t)(x ∂P
−
P
)
+
0
∂x
∂r
solves
∂P0
hL2 iu =
, u(T, x, r; z) = 0.
(3.53)
∂r
Now, it follows from observation 3 in the proof of Proposition 3.2 that
∂P0
1
−
=
∂α
β
∂P0
−(T − t) x
− P0
∂x
∂P0
+
∂r
.
(3.54)
Using this identity, we can express (3.47) only in terms of the “Greeks”.
Next, we obtain an explicit expression for P0,1 , the solution of (3.23). We need some
0
preparation first. By differentiating (3.10) with respect to z, we see that ∂P
∂z solves
hL2 iu = −λ̄0 (z)x
∂P0
+ l λ̄0 (z)P0 ,
∂x
u(T, x, r; z) = 0.
(3.55)
As a result (see Observation 2 in the proof of Propostion 3.2)
∂P0
∂P0
0
= (T − t)λ̄ (z) x
− l P0 ,
∂z
∂x
(3.56)
from which it follows that −hM1 iP0 can be represented as
2
2
∂P0
∂ P0
∂P0
2 ∂ P0
−hM1 iP0 = −(T −t)λ̄ (z) σ̄1 ρ3 g(z) x
+ (1 − l)x
+ ηρ13 g(z) x
−l
.
∂x2
∂x
∂x∂r
∂r
(3.57)
√
Proposition 3.3. The correction term δP0,1 is given by
0
√
2
2
− t)2
∂P0
∂ P0
∂P0
2 ∂ P0
δ1
δP0,1 =
x
+ (1 − l)x
+ V2
x
−l
2
2
∂x
∂x
β ∂α∂x
∂α
2
2
2
(T − t)
∂P0
∂ P0
∂P0
2 ∂ P0
+
x
−lx
+ l P0 − (T − t) x
−l
,
2
∂x2
∂x
∂r∂x
∂r
(T
V1δ
in which
V1δ =
√
δ λ̄0 (z)σ̄1 ρ3 g(z),
V2δ =
√
δ λ̄0 (z)ηρ13 g(z).
(3.58)
(3.59)
Proof. We construct the solution from the following observations and superposition since
hL2 i is linear:
1) We first observe that
(T −t)2
∂n
(xn ∂x
n )P0
2
solves
n
n ∂
hL2 iu = −(T − t) x
P0 ,
∂xn
17
u(T, x, r; z) = 0.
(3.60)
0
2) Next, we apply hL2 i on (T − t) ∂P
∂r and obtain
∂P0
∂P0
∂P0
∂P0
=−
+ (T − t) −x
+β
+ P0 ,
hL2 i (T − t)
∂r
∂r
∂x
∂r
as a result of which we see that
1
∂P0 (T − t)2
∂P0
∂P0
−
x
−
− P0 + (T − t)
β
∂α
2
∂x
∂r
solves
hL2 iu = (T − t)
4
∂P0
,
∂r
u(T, x, r; z) = 0.
(3.61)
(3.62)
(3.63)
Calibration of the Model
In this section, we will calibrate the loss rate l and the parameters
{λ̄, V1 , V2 , V3 , V4 , V5 , V6 , V1δ , V2δ },
which appear in the expressions (3.28), (3.47), and (3.58) on a daily basis (see, e.g., Fouque
et al. (2003) and Papageorgiou and Sircar (2008) for similar calibration exercises carried out
only for the option data or only for the bond data). We demonstrate this calibration on Ford
Motor Company. Note that there are some common parameters between equity options and
corporate bonds. Therefore, our model will be calibrated simultaneously to both of these
data sets. We will also calibrate the parameters of the interest rate and stock models to the
yield curve data, historical spot rate data and historical stock price data. Next, we test our
model by using the estimated parameters to construct an out-of-sample CDS spread time
series (3 year and 5 year), which matches real quoted CDS spread data over the time period
(1/6/2006 − 6/8/2007) quite well.
We also look at how our model-implied volatility matches the real option implied volatility.
We compare our results against those of Fouque et al. (2003). We see that even when we
make the unrealistic assumption of constant volatility, our model is able to produce a very
good fit.
Finally, in the context of index options (when λ = 0), using SPX 500 index options data,
we show the importance of accounting for stochastic interest rates by comparing our model
to that of Fouque et al. (2000, 2003).
18
4.1
Data description
• The daily closing stock price data is obtained from finance.yahoo.com.
• The stock option data is from OptionMetrics under WRDS database, which is the same
database used in Carr and Wu (2006).
– For index options, SPX 500 in our case, we use the data from their Volatility
Surface file. The file contains information on standardized options, both calls and
puts, with expirations of 30, 60, 91, 122, 152, 182, 273, 365, 547, and 730 calender
days. Implied volatilities there are interpolated data using a methodology based
on kernel smoothing algorithm. The interpolated implied volatilities are very close
to real data because there are a great number of options each day for SPX 500
with different maturities and strikes. The calibration results for index options are
presented in Figure 7 and only the data set on the June 8, 2007 is used.
– On September 15, 2006 (Friday) Ford announced that it would not be paying dividends (see e.g., http://money.cnn.com/2006/09/15/news/companies/ford/index.htm).
Therefore, call options on Ford after that date do not have early exercise premium
starting from Sep 18, 2006. We use Ford’s implied volatility surface data from
9/18/2006 to 6/8/2007. when we consider Ford Motor Company’s options. We
excluded the observations with zero trading volume or with maturity less than 9
days. The calibration results are used to construct Figures 1-6. In particular, the
implied volatility surface data from 9/18/2006 to 6/8/2007 is used to construct 2.
As opposed to the options on the index there are not as many individual company
options; and we find that the results given by using interpolated implied volatilities
in the Volatility Surface File and data implied volatilities differ. This may be due
to the fact that there are a limited number of option prices available for individual
companies; i.e., there may not be enough data points for the implied volatilities to
be accurately interpolated. Therefore, we use the Option Price file, which contains
the historical option price information, of the OptionMetrics database
• For each day we U.S government Treasury yield data with maturities: 1 month, 3 monts,
6 months, 1 year, 2 years, 3 years, 5 years, 7 years, 10 years, 20 years. This data set
is available at: www.treasury.gov/offices/domestic-finance/debt-management/interestrate/yield.shtml.
• Corporate bond and CDS data is obtained from Bloomberg. Number of available bond
quotes and bond maturities vary. Typically there are around 15 data points, for example, on June 8th, we have the following maturities: 0.60278,1.0222,1.1861,1.3139,1.4083,
1.5944, 2.3889, 2.6028, 3.0194, 3.2694, 3.3972, 3.6472, 4.1722, 4.3806, 6.3139, 9.5194.
19
4.2
The parameter estimation
The following parameters can be directly estimated from the spot-rate and stock price historical data:
1. The parameters of the interest rate model {α, β, η} are obtained by a least-square fitting
to the Treasury yield curve as in Papageorgiou and Sircar (2008).
2. ρ̄1 = σ̄σ̄12 ρ1 , the “effective” correlation between risk-free spot rate r (we use the onemonth treasury bonds as a proxy for r) and stock price in (3.34) is estimated from
historical risk-free spot rate and stock price data.
3. σ̄2 , the “effective” stock price volatility in (3.34) is estimated from the historical stock
price data.
Now, we detail the calibration method for l, λ̄(z) and {V1 , V2 , V3 , V4 , V5 , V6 , V1δ , V2δ }. We
will minimize the in-sample quadratic pricing error using non-linear least squares to calibrate
these parameters on a daily basis. This way we find a risk neutral model that matches a set
of observed market prices. This risk neutral model can then be used to price more exotic,
illiquid or over-the-counter derivatives. This practice is commonly employed; and for further
discussion of this calibration methodology we refer to Cont and Tankov (2004) (see Chapter
13 and the references therein).
Our calibration is carried out in two steps in tandem:
Step 1. Estimation of lλ̄ and {lV3 , lV2δ } from the corporate bond price data.
The approximate price formula in (3.24) for a defaultable bond is
ec = Bc +
B
0
√
c
B1,0
+
√
c
δB0,1
,
(4.1)
in which B0c is given by (3.29) and
√
√
∂B0c
,
∂α
∂B0c (T − t)2 c
∂B0c
δ1
= lV2
−
+
B0 + (T − t)
.
β
∂α
2
∂r
c
B1,0
= lV3
c
δB0,1
(4.2)
We obtain {lλ̄(z), lV3 , lV2δ } from least-squares fitting, i.e. by minimizing
n
X
c
c
(Bobs
(t, Si ) − Bmodel
(t, Si ; lλ̄, lV3 , lV2δ ))2 ,
i=1
20
(4.3)
c (t, S ) is the observed market price of a bond that matures at time S and
where Bobs
i
i
c
δ
Bmodel (t, Si ; lλ̄, lV3 , lV2 ) is the corresponding model price obtained from (4.1). Here, n is
the number of bonds that are traded at time t. For a fixed value of lλ̄(z) it follows from (4.1)
that {lV3 , lV2δ } can be determined as the least squares solution of
∂B0c
(t, S1 ),
∂α
1
β
∂B0c
∂α (t, Sn ),
1
β
h
−
∂B0c
∂α
+
(S1 −t)2 c
B0
2
..
.
+ (S1 − t)
∂B0c
∂r
+ (Sn − t)
∂r
..
.
h
−
∂B0c
∂α
+
(Sn −t)2 c
B0
2
i
lV
3
δ
i
lV
2
∂B c
0
!
c (t, S ) − B c (t, S ; lλ̄)
Bobs
1
1
0
..
.
=
.
c
c
Bobs (t, Sn ) − B0 (t, Sn ; lλ̄)
Now, we vary lλ̄(z) ∈ [0, M1 ] and choose the point {lλ̄, lV3 , lV2δ } that minimizes (4.3). Here,
we take M1 = 1 guided by the results of Papageorgiou and Sircar (2008).
Step 2. Estimation of {l, V1 , V2 , V4 , V5 , V6 , V1δ } from the equity option data: These
parameters are calibrated from the stock options data by a least-squares fit to the observed
implied volatility. We choose the parameters to minimize
n
X
(Iobs (t, Ti , Ki ) − Imodel (t, Ti , Ki ; model parameters))2
i=1
n
X
(Pobs (t, Ti , Ki ) − Pmodel (t, Ti , Ki ; model parameters))2
≈
vega2 (Ti , Ki )
(4.4)
i=1
in which Iobs (t, Ti , Ki ) and Imodel (t, Ti , Ki ; model parameters) are observed Black-Scholes implied volatility and model Black-Scholes implied volatility, respectively. The right hand side
of (4.4) is from Cont and Tankov (2004), page 439. Here, Pobs (t, Ti , Ki ) is the market price
of a European option (a put or a call) that matures at time Ti and with strike price Ki and
Pmodel (t, Ti , Ki ; model parameters) is the corresponding model price which is obtained from
(3.24). As in Cont and Tankov (2004), vega(Ti , Ki ) is the market implied Black-Scholes vega.
Let P0 (t, Ti , Ki ; λ̄(z)) be either of (3.44) and (3.46) with K = Ki and T = Ti . Let us
introduce the Greeks,
g1 = −(T − t)x
∂ 3 P0
,
∂x2 ∂α
2
1
∂ P0
g8 =
x
β
∂α∂x
g4 = x2
2P
0
,
2
∂x
2
∂
∂
∂P0
2 ∂ P0
g2 = −(T − t)x
x
, g3 =
x
− P0 ,
∂x
∂x2
∂α
∂x
∂ 2 P0
∂ 2 P0
(T − t)2 2 ∂ 2 P0
g5 = x
, g6 = x
, g7 =
x
,
∂η∂x
∂α∂x
2
∂x2
2
∂P0 (T − t)2
∂P0
∂ 2 P0
∂P0
2 ∂ P0
−
+
x
+ P0 − (T − t) x(
∂x) −
,
−x
∂α
2
∂x2
∂x
∂r
∂r
(4.5)
2∂
in which each term can be explicitly evaluated (see Appendix).
21
Now from (3.24) and the results of Section 3.3 (with l = 1), we can write
Pmodel (t, Ti , Ki ; model parameters) = P0 (t, Ti , Ki ; λ̄) + V1 g1 (Ti , Ki ; λ̄) + V2 g2 (Ti , Ki ; λ̄)
+ V3 g3 (Ti , Ki ; λ̄) + V4 g4 (Ti , Ki ; λ̄) + V5 g5 (Ti , Ki ; λ̄)
+ V6 g6 (Ti , Ki ; λ̄) + V1δ g7 (Ti , Ki ; λ̄) + V2δ g8 (Ti , Ki ; λ̄).
(4.6)
First, let us fix the value of l. Then, from Step 1, we can infer the values of {λ̄, V3 , V2δ }.
Now the fitting problem in (4.4) is a linear least squares problem for {V1 , V2 , V4 , V5 , V6 , V1δ }.
Next, we vary l ∈ [0, 1] and choose {l, V1 , V2 , V4 , V5 , V6 , V1δ } so that (4.4) is minimized.
4.3
Model implied CDS spread matches the observed CDS spread
e c (t, T ; l) denote the approximation for the price at time t of a defaultable bond that
Let B
matures at time T , and has loss rate l (see (4.1)). Let B(t, T ) be the price of a risk-free bond.
Then, the model implied CDS spead with maturity TM is
cds
model (t, TM ) =
e c (t, TM ; l)
B(t, TM ) − B
.
M
X
c
e (t, Tm ; 1)
B
(4.7)
m=1
Recall that we have already estimated all of the model parameters in Section 4.2 using both
corporate term structure data and the stock option implied volatility surface. Therefore,
using (4.7) we can plot the model implied CDS spread over time and compare it with the
CDS spread data available in the market. This is precisely what we do in Figures 1 and 2. We
ds
look at the time series cds
model (t, 3) and cmodel (t, 5) and compare them to the CDS spread time
series of the Ford Motor Company. The match seems to be extremely good, which attests to
the power of our modeling framework.
By varying TM in (4.7) we can obtain the model implied term structure of the CDS
spread. Figure 3 shows the range of shapes we can produce.
4.4
Fitting Ford’s implied volatility
We will compare how well our model fits the implied volatility against the model of Fouque
et al. (2003), which does not account for the default risk and for the randomness of the
interest rates. Although, we only calibrate seven parameters (hence we refer to our model
as the 7-parameter model) to the option prices (see the second step of the estimation in
22
Section 4.2), we have many more parameters than the model of Fouque et al. (2003), which
only has four parameters (we refer to this model as the 4-parameter model). Therefore, for
a fair comparison, we also consider a model in which the volatility is a constant. In this
case, as we shall see below, there are only three parameters to calibrate to the option prices,
therefore we call it the 3-parameter model.
Constant Volatility Model In this case, we take σ̄1 = σ̄2 = σ in the expression for P0
√
√
in Corollary 3.1 . The expression for δP0,1 remains the same as before. However, P1,0
simplifies to
√
P1,0 = −(T −
2P
0
∂x2
∂
t)V1 x2
+
V3
∂ 2 P0
∂P0
.
−x
+
∂α∂x
∂α
(4.8)
This model has only three parameters, l, V1 , V1δ that need to be calibrated to the options
prices, as opposed to the 4-parameter model of Fouque et al. (2003).
As it can be seen from Figure 4 as expected our 7-parameter model outperforms the
4-parameter model of Fouque et al. (2003) as expected and fits the implied volatility data
well. But, what is surprising is that the 3-parameter model, which does not account for the
volatility but accounts for the default risk and stochastic interest rate, has almost the same
performance as the 7-parameter model.
The 7-parameter model has a very rich implied volatility surface structure, the surface has
more curvature than that of the 4-parameter model of Fouque et al. (2003), whose volatility
surface is more flat; see Figures 5 and 6. (The parameters to draw these figures are obtained
by calibrating the models to the data implied volatility surface on June 8 2007.) The 7parameter model has a recognizable skew even for longer maturities and has a much sharper
skew for shorter maturities.
4.5
Fitting the implied volatility of the index options
The purpose of this section is to show the importance of accounting for stochastic interest
rates in fitting the implied volatility surface. Interest rate changes should, indeed, be accounted for in pricing long maturity options. When we price index options, we set λ̄ = 0 and
our approximation in (3.24) simplifies to
P ,δ ≈ P0 +
23
√
P1,0 ,
(4.9)
in which P0 is given by Corollary 3.1 after settiing λ̄(z) = 0, and
2
2
3
2
2
∂
2 ∂ P0
2 ∂ P0
∂ P0
∂ P0
2 ∂ P0
P1,0 = −(T − t) V1 x
+
V
x
x
+
V
x
+
V
x
+
V
x
2
4
5
6
∂x2
∂x
∂x2
∂x2 ∂α
∂η∂x
∂α∂x
(4.10)
Note that the difference of (4.9) with the model of Fouque et al. (2003) is that the latter
allows for a slow evolving volatility factor to better match the implied volatility at the longer
maturities. This was an improvement on the model of Fouque et al. (2000), which only has
a fast scale component in the volatility model. We, on the other hand, by accounting for
stochastic interest rates, capture the same performance by using only a fast scale volatility
model.
√
From Figure 7, we see that both (4.9) and Fouque et al. (2003) outperform the model of
Fouque et al. (2000), especially at the longer maturities (T = 9 months, 1 year, 1.5 years and
2 years), and their performance is very similar. This observation emphasizes the importance
of accounting for stochastic interest rates for long maturity contracts.
Appendix: Explicit formulae for the Greeks in (4.5)
When h(x) = (x − K)+ , we can explicitly express the Greeks in (4.5) in terms of f (x) =
√1 exp(−x2 /2) as
2π
2
xf (d1 )
∂
xf (d1 )
d1
2 ∂ C0
= √
, x
x
= √
1− √
,
x
vt,T
∂x
∂x2
vt,T
vt,T
∂
∂C0
T − t exp(−β(T − t)) − 1
f (d2 )
c
x
− C0 = −K B̄ (t, T )
+
N (d2 ) − √
,
∂α
∂x
β
β2
vt,T
2
∂
−xf (d1 )d1 T − t exp(−β(T − t)) − 1
2 ∂ C0
x
=
+
,
∂α
∂x2
vt,T
β
β2
∂ ∂C0
xf (d1 ) T − t exp(−β(T − t)) − 1
x
= √
+
,
∂x ∂α
vt,T
β
β2
∂
∂C0
1 − exp(−β(T − t))
f (d2 )
x
− C0 = −K B̄ c (t, T )
N (d2 ) − √
,
∂r
∂x
β
vt,T
∂ ∂C0
1
η
2η
η
x
= xf (d1 ) − √
(T
−
t)
+
(exp(−β(T
−
t))
−
1
− 3 (exp(−2β(T − t)) − 1))
∂x ∂η
vt,T β 2
2β 3
2β
!
x
1
1
−3/2
+ − log
vt,T + √
×
c
2
4 vt,T
K B̄t,T
2ρ̄1 σ̄2 2η
2ρ̄1 σ̄2 4η
2ρ̄1 σ̄2 3η
η
+ 2 (T − t) +
+
exp(−β(T
−
t))
−
exp(−2β(T
−
t))
−
+
.
β
β
β2
β3
β3
β2
β3
2∂
2C
0
∂x2
24
References
Bayraktar, E. (2008).
(3): 277–304.
Pricing options on defaultable stocks, Applied Mathematical Finance 15
Bielecki, T. R. and Rutkowski, M. (2002). Credit Risk: Modeling, Valuation and Hedging, Springer,
New York.
Carr, P. and Linetsky, V. (2006). A jump to default extended CEV model: An application of Bessel
processes, Finance and Stochastics 10: 303–330.
Carr, P. and Wu, L. (2006). Stock options and credit default swaps: A joint framework for valuation
and estimation, Technical report, NYU. Available at http://faculty.baruch.cunny.edu/lwu.
Cont, R. and Tankov, P. (2004). Financial Modeling with Jump Processes, Chapman & Hall, Boca
Raton, FL.
Cotton, P., Fouque, J.-P., Papanicolaou, G. and Sircar, R. (2004). Stochastic volatility corrections for
interest rate derivatives, Math. Finance 14(2): 173–200.
Duffie, D. and Singleton, K. (1999). Modeling term structure of defaultable bonds, Review of Financial
Studies 12 (4): 687–720.
Fouque, J.-P., Papanicolaou, G. and Sircar, K. R. (2000). Derivatives in Financial Markets with
Stochastic Volatility, Cambridge University Press, New York.
Fouque, J. P., Papanicolaou, G., Sircar, R. and Solna, K. (2003). Multiscale stochastic volatility
asymptotics, SIAM J. Multiscale Modeling and Simulation 2 (1): 22–42.
Geman, H., Karoui, N. E. and Rochet, J. C. (1995). Changes of numéraire, changes of probability
measures and option pricing, Journal of Applied Probability 32: 443–458.
Linetsky, V. (2006). Pricing equity derivatives subject to bankruptcy, Mathematical Finance 16
(2): 255–282.
Papageorgiou, E. and Sircar, R. (2008). Multiscale intensity based models for single name credit
derivatives, Applied Mathematical Finance 15 (1): 73–105.
Schönbucher, P. J. (1998). Term structure of defaultable bond prices, Review of Derivatives Research
2 (2/3): 161–192.
Schönbucher, P. J. (2003). Credit Derivatives Pricing Models: Model, Pricing and Implementation,
Wiley, New York.
25
400
350
Ford 3yr CDS spread(bps)
300
250
200
150
100
50
0
9/18/06
11/7/06
12/27/06
2/15/07
4/6/07
Date
Figure 1: Ford 3 year CDS annual spread time series from 9/18/2006-3/13/2007.
Spread implied by model is pink solid line, real quoted spread is blue broken line. Ford’s 3
year CDS spread time series is not available from 3/13/2007 until 8/31/07 in our data source.
500
450
400
Ford 5yr CDS spread(bps)
350
300
250
200
150
100
50
0
9/18/06
11/7/06
12/27/06
2/15/07
4/6/07
Date
Figure 2: Ford 5 year CDS annual spread time series from 9/18/2006-6/8/2007.
Spread implied by model is pink solid line, real quoted spread is blue broken line.
26
5/26/07
450
400
350
CDS spread (bps)
300
250
200
150
100
50
0
1
2
3
4
5
6
7
8
9
10
Time to maturity (years)
Figure 3: CDS Term Structures (4.7) can produce:
Legend
-x-, blue (The parameters are obtained from calibration to 11/13/2006): α=0.0037, β=0.0872
η = 0.0001, r = 0.0516, l(loss rate) = 0.283, λ̄(z) = 0.0459, [V3 , V2δ ] = [0.0425, 0.0036].
-squares-, black (The parameters correspond to 6/18/2006): α = 0.0045, β = 0.0983,
η = 0.0002, r = 0.0516, l = 1, λ̄ = 0.012, [V3 , V2δ ] = [0.0185, 0.0025],
-diamonds-, red (The parameters correspond to 9/22/2006): α = 0.0039, β = 0.0817,
η = 0.0012, r = 0.0496, l=1, λ̄(z) = 0.017, [V3 , V2δ ] = [0.0067, 0.0005]
27
maturity = 45 days
maturity = 17 days
0.55
0.9
0.8
implied volatility
implied volatility
0.5
0.45
0.4
0.35
0.7
0.6
0.5
0.4
0.3
7
7.2
7.4
7.6
strikes
7.8
0.2
8
5
0.44
0.42
0.42
0.4
0.38
0.36
0.34
0.32
0.3
7
strikes
8
9
maturity = 168 days
0.44
implied volatility
implied volatility
maturity = 72 days
6
0.4
0.38
0.36
0.34
0.32
7
7.5
8
strikes
8.5
0.3
9
7
8
9
10
strikes
maturity = 285 days
maturity = 643 days
0.6
0.6
implied volatility
implied volatility
0.55
0.5
0.4
0.3
0.5
0.45
0.4
0.35
0.2
4
6
8
10
strikes
12
14
5
10
strikes
15
Figure 4:
Implied volatility fit to the Ford call option data with maturities of
[17,45,72,168,285,643] calender days on April 4, 2007.
Model is calibrated aross all maturities but we plotted the implied volatilities for each maturity, separately. Here, stock price (x) = 8.04, historical volatility (σ̄2 ) = 0.3827, one month
treasury rate (r) = 0.0516, estimated correlation between risk-free spot rate(one month treasury) and stock price (ρ̄1 ) = −0.0327. Also α = 0.0037, β = 0.0872, η = 0.0001 which are
obtained with a least-square fitting to the Treasury yield curve on the 4th of April.
Legend:
’o’, empty circles = observed data;
’x’, green = stochastic vol+stochastic hazard rate+stochastic interest rate = the 7-parameter
model;
small full circle, blue = constant vol+stochastic hazard rate+ stochastic interest rate = the
3-parameter model
’*’, red = The model of Fouque et al. (2003) which has constant interest rate+stochastic vol
(slow and fast scales) = the 4 parameter model.
28
0.7
implied volatility
0.65
0.6
0.55
0.5
0.45
0.4
0.35
1
0.8
0.4
0.6
0.6
0.8
0.4
1
1.2
1.4
0.2
time to maturity
moneyness
Figure 5: Implied volatility surface corresponding to (4.6), the 7-parameter model.
Here,
α
=
0.0063,
β
=
0.1034,
η
=
0.012,
r
=
0.0476
σ̄2
=
0.2576,
λ̄(z)
=
0.027,
(V1 , V2 , V3 , V4 , V5 , V6 , V1δ , V2δ )
=
(0.9960, −0.0014, 0.0009, 0.0104, −0.6514, 0.3340, −0.1837, −0.0001).
0.65
0.6
implied volatility
0.55
0.5
0.45
0.4
0.35
0.3
0.25
0.4
1
0.6
0.8
0.8
0.6
1
1.2
0.4
1.4
0.2
time to maturity
moneyness
Figure 6: Implied Volatility Surface corresponding to the 4-parameter model of Fouque et al.
(2003).
Here, r = 0.046, average volatility=0.2546, and the parameters in (4.3) of Fouque et al.
(2003) are choosen to be (V2 , V3 , V0δ , V1δ ) = (−0.0164, −0.1718, 0.0006, 0.0630). Note that
the parameters here and Figure 5 are both obtained by calibrating the models to the data
implied volatility surface of Ford Motor Company on June 8, 2007.
29
maturity = 60 days
implied volatility
implied volatility
maturity = 30 days
0.25
0.2
0.15
0.1
1450
1500
1550
1600
0.2
0.15
0.1
1400
1450
strikes
0.15
0.1
1400
1450
1500
strikes
1550
1600
0.15
0.1
1400
1450
implied volatility
implied volatility
1500
strikes
1600
1700
implied volatility
implied volatility
1500
strikes
1600
1700
0.15
0.1
1300
1400
implied volatility
implied volatility
1600
strikes
1800
1600
1700
0.15
0.1
1300
1400
1500
1600
strikes
1700
1800
maturity = 730 days
0.15
1400
1500
strikes
0.2
maturity = 547 days
0.2
0.1
1200
1650
maturity = 365 days
0.15
1400
1600
0.2
maturity = 273 days
0.2
0.1
1300
1500
1550
strikes
maturity = 182 days
0.15
1400
1600
0.2
maturity = 152 days
0.2
0.1
1300
1550
maturity = 122 days
0.2
implied volatility
implied volatility
maturity = 91 days
1500
strikes
2000
0.25
0.2
0.15
0.1
1200
1400
1600
strikes
1800
2000
Figure 7: The fit to the Implied Volatility Surface of SPX on June 8, 2007 with maturities
[30,60,91,122,152,182,273,365,547,730] calender days. Recall from Section 4.1 that we use
standardized options from the OptionMetrics.
Models are calibrated aross all maturities, but we plot the implied volatility fits separately.
The parameters are: stock price (x) = 1507.67, dividend rate = 0.0190422, historical volatility
(σ̄2 ) = 0.1124, one month treasury rate (r) = 0.0476, estimated correlation between risk-free
spot rate(one month treasury) and stock price (ρ̄1 ) = 0.020454. Also, α = 0.0078, β = 0.1173,
η = 0.0241, which are obtained from a least-square fitting to the Treasury yield curve of the
same day.
Legend
’o’, empty cirles = observed data,
’x”, green = Implied volatility of (4.9),
’*’, red = Implied volatility of Fouque et al. (2003),
small full circle, blue = Implied volatility of Fouque et al. (2000).
30
| 5cs.CE |
Time complexity of concurrent programs⋆
– a technique based on behavioural types –
arXiv:1511.05104v1 [cs.PL] 16 Nov 2015
Elena Giachino1 , Einar Broch Johnsen2, Cosimo Laneve1, and Ka I Pun2
1
Dept. of Computer Science and Engineering, University of Bologna – INRIA FOCUS
2
Dept. of Informatics, University of Oslo
Abstract. We study the problem of automatically computing the time complexity of concurrent object-oriented programs. To determine this complexity we
use intermediate abstract descriptions that record relevant information for the
time analysis (cost of statements, creations of objects, and concurrent operations), called behavioural types. Then, we define a translation function that takes
behavioural types and makes the parallelism explicit into so-called cost equations, which are fed to an automatic off-the-shelf solver for obtaining the time
complexity.
1 Introduction
Computing the cost of a sequential algorithm has always been a primary question for
every programmer, who learns the basic techniques in the first years of their computer
science or engineering curriculum. This cost is defined in terms of the input values to
the algorithm and over-approximates the number of the executed instructions. In turn,
given an appropriate abstraction of the CPU speed of a runtime system, one can obtain
the expected computation time of the algorithm.
The computational cost of algorithms is particularly relevant in mainstream architectures, such as the cloud. In that context, a service is a concurrent program that must
comply with a so-called service-level agreement (SLA) regulating the cost in time and
assigning penalties for its infringement [3]. The service provider needs to make sure
that the service is able to meet the SLA, for example in terms of the end-user response
time, by deciding on a resource management policy and determining the appropriate
number of virtual machine instances (or containers) and their parameter settings (e.g.,
their CPU speeds). To help service providers make correct decisions about the resource
management before actually deploying the service, we need static analysis methods
for resource-aware services [6]. In previous work by the authors, cloud deployments
expressed in the formal modeling language ABS [8] have used a combination of cost
analysis and simulations to analyse resource management [1], and a Hoare-style proof
system to reason about end-user deadlines has been developed for sequential executions [7]. In contrast, we are here interested in statically estimating the computation
time of concurrent services deployed on the cloud with a given dynamic resource management policy.
⋆
Supported by the EU projects FP7-610582 Envisage: Engineering Virtualized Services
(http://www.envisage-project.eu).
Technically, this paper proposes a behavioural type system expressing the resource
costs associated with computations and study how these types can be used to soundly
calculate the time complexity of parallel programs deployed on the cloud. To succinctly
formulate this problem, our work is developed for tml, a small formally defined concurrent object-oriented language which uses asynchronous communications to trigger
parallel activities. The language defines virtual machine instances in terms of dynamically created concurrent object groups with bounds on the number of cycles they can
perform per time interval. As we are interested in the concurrent aspects of these computations, we abstract from sequential analysis in terms of a statement job(e), which
defines the number of processing cycles required by the instruction – this is similar to
the sleep(n) operation in Java.
The analysis of behavioural types is defined by translating them in a code that is
adequate for an off-the-shelf solver – the CoFloCo solver [4]. As a consequence, we are
able to determine the computational cost of algorithms in a parametric way with respect
to their inputs.
Paper overview. The language is defined in Section 2 and we discuss restrictions that
ease the development of our technique in Section 3. Section 4 presents the behavioural
type system and Section 5 explains the analysis of computation time based on these
behavioural types. In Section 6 we outline our correctness proof of the type system
with respect to the cost equations. In Section 7 we discuss the relevant related work and
in Section 8 we deliver concluding remarks.
2 The language tml
The syntax and the semantics of tml are defined in the following two subsections; the
third subsection discusses a few examples.
Syntax. A tml program is a sequence of method definitions T m(T x){ F y ; s }, ranged
over by M, plus a main body { F z ; s′ } with k. In tml we distinguish between simple types T which are either integers Int or classes Class (there is just one class in
tml), and types F, which also include future types Fut<T >. These future types let asynchronous method invocations be typed (see below). The notation T x denotes any finite
sequence of variable declarations T x. The elements of the sequence are separated by
commas. When we write T x ; we mean a sequence T 1 x1 ; · · · ; T n xn ; when the
sequence is not empty; we mean the possibly empty sequence otherwise.
The syntax of statements s, expressions with side-effects z and expressions e of tml
is defined by the following grammar:
s ::= x = z | if e { s } else { s } | job(e) | return e | s ; s
z ::= e | e!m(e) | e.m(x) | e.get | new Class with e | new local Class
e ::= this | se | nse
A statement s may be either one of the standard operations of an imperative language
or the job statement job(e) that delays the continuation by e cycles of the machine
executing it.
An expression z may change the state of the system. In particular, it may be an asynchronous method invocation of the form e!m(e), which does not suspend the caller’s
execution. When the value computed by the invocation is needed, the caller performs a
non-blocking get operation: if the value needed by a process is not available, then an
awaiting process is scheduled and executed, i.e., await-get. Expressions z also include
standard synchronous invocations e.m(e) and new local Class, which creates a new
object. The intended meaning is to create the object in the same machine – called cog
or concurrent object group – of the caller, thus sharing the processor of the caller: operations in the same virtual machine interleave their evaluation (even if in the following
operational semantics the parallelism is not explicit). Alternatively, one can create an
object on a different cog with new Class with e thus letting methods execute in parallel. In this case, e represents the capacity of the new cog, that is, the number of cycles
the cog can perform per time interval. We assume the presence of a special identifier
this.capacity that returns the capacity of the corresponding cog.
A pure expression e can be the reserved identifier this or an integer expression.
Since the analysis in Section 5 cannot deal with generic integer expressions, we parse
expressions in a careful way. In particular we split them into size expressions se, which
are expressions in Presburger arithmetics (this is a decidable fragment of Peano arithmetics that only contains addition), and non-size expressions nse, which are the other
type of expressions. The syntax of size and non-size expressions is the following:
nse ::=
|
se ::=
ve ::=
k ::=
k | x | nse ≤ nse | nse and nse | nse or nse
nse + nse | nse − nse | nse × nse | nse/nse
ve | ve ≤ ve | se and se | se or se
k | x | ve + ve | k × ve
rational constants
In the paper, we assume that sequences of declarations T x and method declarations
M do not contain duplicate names. We also assume that return statements have no
continuation.
Semantics. The semantics of tml is defined by a transition system whose states are
configurations cn that are defined by the following syntax.
cn ::=
|
p ::=
q ::=
ε | fut( f, val) | ob(o, c, p, q) | invoc(o, f, m, v)
cog(c, act, k) | cn cn
{ l | s } | idle
∅ | {l | s} | q q
act
val
l
v
::=
::=
::=
::=
o|ε
v|⊥
[· · · , x 7→ v, · · · ]
o| f |k
A configuration cn is a set of concurrent object groups (cogs), objects, invocation
messages and futures, and the empty configuration is written as ε. The associative and
commutative union operator on configurations is denoted by whitespace. A cog is given
as a term cog(c, act, k) where c and k are respectively the identifier and the capacity of
the cog, and act specifies the currently active object in the cog. An object is written
as ob(o, c, p, q), where o is the identifier of the object, c the identifier of the cog the
object belongs to, p an active process, and q a pool of suspended processes. A process
is written as { l | s }, where l denotes local variable bindings and s a list of statements.
An invocation message is a term invoc(o, f, m, v) consisting of the callee o, the future
f to which the result of the call is returned, the method name m, and the set of actual
parameter values for the call. A future fut( f, val) contains an identifier f and a reply
value val, where ⊥ indicates the reply value of the future has not been received.
(Assign-Local)
x ∈ dom(l) v = [[e]]l
ob(o, c, { l | x = e ; s }, q)
→ ob(o, c, { l [x 7→ v] | s }, q)
(Cond-True)
true = [[e]]l
ob(o, c, { l | if e { s1 } else { s2 } ; s }, q)
→ ob(o, c, { l | s1 ; s }, q)
(New)
c′ = fresh( ) o′ = fresh( ) k = [[e]]l
ob(o, c, { l | x = new Class with e ; s }, q)
→ ob(o, c, { l | x = o′ ; s }, q) ob(o′ , c′ , idle, ∅) cog(c′ , o′ , k)
(Get-True)
f = [[e]]l v , ⊥
ob(o, c, { l | x = e.get ; s }, q) fut( f, v)
→ ob(o, c, { l | x = v ; s }, q) fut( f, v)
(New-Local)
o′ = fresh( )
ob(o, c, { l | x = new local Class ; s }, q)
→ ob(o, c, { l | x = o′ ; s }, q) ob(o′ , c, idle, ∅)
(Get-False)
f = [[e]]l
ob(o, c, { l | x = e.get ; s }, q) fut( f, ⊥)
→ ob(o, c, idle, q ∪ { l | x = e.get ; s }) fut( f, ⊥)
(Self-Sync-Call)
o = [[e]]l v = [[e]]l f ′ = l(destiny)
f = fresh( ) { l′ | s′ } = bind(o, f, m, v)
ob(o, c, { l | x = e.m(e) ; s }, q)
→ ob(o, c, { l′ | s′ ; cont( f ′ ) }, q ∪ { l | x = f.get ; s }) fut( f, ⊥)
(Cog-Sync-Call)
o′ = [[e]]l v = [[e]]l f ′ = l(destiny)
f = fresh( ) { l′ | s′ } = bind(o′ , f, m, v)
ob(o, c, { l | x = e.m(e) ; s }, q)
ob(o′ , c, idle, q′ ) cog(c, o, k)
→ ob(o, c, idle, q ∪ { l | x = f.get ; s }) fut( f, ⊥)
ob(o′ , c, { l′ | s′ ; cont( f ′ ) }, q′ ) cog(c, o′ , k)
(Async-Call)
o′ = [[e]]l v = [[e]]l f = fresh( )
ob(o, c, { l | x = e!m(e) ; s }, q)
→ ob(o, c, { l | x = f ; s }, q) invoc(o′ , f, m, v) fut( f, ⊥)
(Release-Cog)
ob(o, c, idle, q) cog(c, o, k)
→ ob(o, c, idle, q) cog(c, ε, k)
(Return)
v = [[e]]l f = l(destiny)
ob(o, c, { l | return e }, q) fut( f, ⊥)
→ ob(o, c, idle, q) fut( f, v)
(Cond-False)
false = [[e]]l
ob(o, c, { l | if e { s1 } else { s2 } ; s }, q)
→ ob(o, c, { l | s2 ; s }, q)
(Self-Sync-Return-Sched)
f = l′ (destiny)
ob(o, c, { l | cont( f ) }, q ∪ { l′ | s })
→ ob(o, c, { l′ | s }, q)
(Cog-Sync-Return-Sched)
f = l′ (destiny)
ob(o, c, { l | cont( f ) }, q) cog(c, o, k)
ob(o′ , c, idle, q′ ∪ { l′ | s′ })
→ ob(o, c, idle, q) cog(c, o′ , k)
ob(o′ , c, { l′ | s′ }, q′ )
(Bind-Mtd)
{ l | s } = bind(o, f, m, v)
ob(o, c, p, q) invoc(o, f, m, v)
→ ob(o, c, p, q ∪ { l | s })
(Activate)
ob(o, c, idle, q ∪ { l | s }) cog(c, ε, k)
→ ob(o, c, { l | s }, q) cog(c, o, k)
(Job-0)
[[e]]l = 0
ob(o, c, { l | job(e) ; s }, q)
→ ob(o, c, { l | s }, q)
(Context)
cn → cn′
cn cn′′ → cn′ cn′′
Fig. 1. The transition relation of tml – part 1.
The following auxiliary function is used in the semantic rules for invocations. Let
T ′ m(T x){ F x′ ; s } be a method declaration. Then
bind(o, f, m, v) = { [destiny 7→ f, x 7→ v, x′ 7→ ⊥] | s{o/this} }
The transition rules of tml are given in Figures 1 and 2. We discuss the most relevant ones: object creation, method invocation, and the job(e) operator. The creation
of objects is handled by rules New and New-Local: the former creates a new object
inside a new cog with a given capacity e, the latter creates an object in the local cog.
Method invocations can be either synchronous or asynchronous. Rules Self-Sync-Call
and Cog-Sync-Call specify synchronous invocations on objects belonging to the same
cog of the caller. Asynchronous invocations can be performed on every object.
(Tick)
strongstable t (cn)
cn → Φ(cn, t)
where
ob(o, c, {l′ | job(k′ ) ; s}, q) Φ(cn′ , t)
Φ(cn, t) =
ob(o, c, idle, q) Φ(cn′ , t)
cn
if cn = ob(o, c, {l | job(e) ; s}, q) cn′
and cog(c, o, k) ∈ cn′
and k′ = [[e]]l − k ∗ t
if cn = ob(o, c, idle, q) cn′
otherwise.
Fig. 2. The transition relation of tml – part 2: the strongly stable case
In our model, the unique operation that consumes time is job(e). We notice that the
reduction rules of Figure 1 are not defined for the job(e) statement, except the trivial
case when the value of e is 0. This means that time does not advance while non-job
statements are evaluated. When the configuration cn reaches a stable state, i.e., no other
transition is possible apart from those evaluating the job(e) statements, then the time
is advanced by the minimum value that is necessary to let at least one process start.
In order to formalize this semantics, we define the notion of stability and the update
operation of a configuration cn (with respect to a time value t). Let [[e]]l return the value
of e when variables are bound to values stored in l.
Definition 1. Let t > 0. A configuration cn is t-stable, written stable t (cn), if any object
in cn is in one of the following forms:
1. ob(o, c, { l | job(e); s }, q) with cog(c, o, k) ∈ cn and [[e]]l /k ≥ t,
2. ob(o, c, idle, q) and
i. either q = ∅,
ii. or, for every p ∈ q, p = { l | x = e.get; s } with [[e]]l = f and fut( f, ⊥),
iii. or, cog(c, o′ , k) ∈ cn where o , o′ , and o′ satisfies Definition 1.1.
A configuration cn is strongly t-stable, written strongstable t (cn), if it is t-stable and
there is an object ob(o, c, { l | job(e); s }, q) with cog(c, o, k) ∈ cn and [[e]]l /k = t.
Notice that t-stable (and, consequently, strongly t-stable) configurations cannot progress
anymore because every object is stuck either on a job or on unresolved get statements.
The update of cn with respect to a time value t, noted Φ(cn, t) is defined in Figure 2.
Given these two notions, rule Tick defines the time progress.
The initial configuration of a program with main method { F x; s } with k is
ob(start, start, { [destiny 7→ fstart , x 7→ ⊥] | s }, ∅)
cog(start, start, k)
where start and start are special cog and object names, respectively, and fstart is a fresh
future name. As usual, →∗ is the reflexive and transitive closure of →.
Examples. To begin with, we discuss the Fibonacci method. It is well known that the
computational cost of its sequential recursive implementation is exponential. However,
this is not the case for the parallel implementation. Consider
Int fib(Int n) {
if (n<=1) { return 1; }
else { Fut<Int> f; Class z; Int m1; Int m2;
job(1);
z = new Class with this.capacity ;
f = this!fib(n-1); g = z!fib(n-2);
m1 = f.get; m2 = g.get;
return m1 + m2;
}
}
Here, the recursive invocation fib(n-1) is performed on the this object while the
invocation fib(n-2) is performed on a new cog with the same capacity (i.e., the object
referenced by z is created in a new cog set up with this.capacity), which means
that it can be performed in parallel with the former one. It turns out that the cost of the
following invocation is n.
Class z; Int m; Int x;
x = 1;
z = new Class with x;
m = z.fib(n);
Observe that, by changing the line x = 1; into x = 2; we obtain a cost of n/2.
Our semantics does not exclude paradoxical behaviours of programs that perform
infinite actions without consuming time (preventing rule Tick to apply), such as this
one
Int foo() { Int m; m = this.foo(); return m; }
This kind of behaviours are well-known in the literature, (cf. Zeno behaviours) and
they may be easily excluded from our analysis by constraining recursive invocations
to be prefixed by a job(e)-statement, with a positive e. It is worth to observe that this
condition is not sufficient to eliminate paradoxical behaviours. For instance the method
below does not terminate and, when invoked with this.fake(2), where this is in a
cog of capacity 2, has cost 1.
Int fake(Int n) {
Int m; Class x;
x = new Class with 2*n; job(1); m = x.fake(2*n); return m;
}
Imagine a parallel invocation of the following method
Int one() { job(1); }
on an object residing in a cog of capacity 1. At each stability point the job(1) of the
latter method will compete with the job(1) of the former one, which will win every
time, since having a greater (and growing) capacity it will require always less time. So at
the first stability point we get job(1−1/2) (for the method one), then job(1−1/2−1/4)
and so on, thus this sum will never reach 0.
In the examples above, the statement job(e) is a cost annotation that specifies how
many processing cycles are needed by the subsequent statement in the code. We notice
that this operation can also be used to program a timer which suspends the current
execution for e units of time. For instance, let
Int wait(Int n) { job(n); return 0; }
Then, invoking wait on an object with capacity 1
Class timer; Fut<Class> f; Class x;
timer = new Class with 1;
f = timer!wait(5); x = f.get;
one gets the suspension of the current thread for 5 units of time.
3 Issues in computing the cost of tml programs
The computation time analysis of tml programs is demanding. To highlight the difficulties, we discuss a number of methods.
Int wrapper(Class x) {
Fut<Int> f; Int z;
job(1) ; f = x!server(); z = f.get;
return z;
}
Method wrapper performs an invocation on its argument x. In order to determine the
cost of wrapper, we notice that, if x is in the same cog of the carrier, then its cost
is (assume that the capacity of the carrier is 1): 1 + cost(server) because the two
invocations are sequentialized. However, if the cogs of x and of the carrier are different,
then we are not able to compute the cost because we have no clue about the state of the
cog of x. Next consider the following definition of wrapper
Int wrapper_with_log(Class x) {
Fut<Int> f; Fut<Int> g; Int z;
job(1) ; f = x!server(); g = x!print_log(); z = f.get;
return z;
}
In this case the wrapper also asks the server to print its log and this invocation
is not synchronized. We notice that the cost of wrapper_with_log is not anymore
1+cost(server) (assuming that x is in the same cog of the carrier) because print_log
might be executed before server. Therefore the cost of wrapper_with_log is 1 +
cost(server) + cost(print log).
Finally, consider the following wrapper that also logs the information received from
the server on a new cog without synchronising with it:
Int wrapper_with_external_log(Class x) {
Fut<Int> f; Fut<Int> g; Int z; Class y;
job(1) ; f = x!server(); g = x!print_log(); z = f.get;
y = new Class with 1;
f = y!external_log(z) ;
return z;
}
What is the cost of wrapper_with_external_log? Well, the answer here is debatable: one might discard the cost of y!external_log(z) because it is useless for the
value returned by wrapper_with_external_log, or one might count it because one
wants to count every computation that has been triggered by a method in its cost. In
this paper we adhere to the second alternative; however, we think that a better solution
should be to return different cost for a method: a strict cost, which spots the cost that
is necessary for computing the returned value, and an overall cost, which is the one
computed in this paper.
Anyway, by the foregoing discussion, as an initial step towards the time analysis of
tml programs, we simplify our analysis by imposing the following constraint:
– it is possible to invoke methods on objects either in the same cog of the caller or on
newly created cogs.
The above constraint means that, if the callee of an invocation is one of the arguments of
a method then it must be in the same cog of the caller. It also means that, if an invocation
is performed on a returned object then this object must be in the same cog of the carrier.
We will enforce these constraints in the typing system of the following section – see
rule T-Invoke.
4 A behavioural type system for tml
In order to analyse the computation time of tml programs we use abstract descriptions,
called behavioural types, which are intermediate codes highlighting the features of tml
programs that are relevant for the analysis in Section 5. These abstract descriptions
support compositional reasoning and are associated to programs by means of a type
system. The syntax of behavioural types is defined as follows:
t ::= -- | se | c[se]
x ::= f | t
a ::= e | νc[se] | m(t) → t | ν f : m(t) → t
b ::= a ⊲ Γ | a # b | (se){ b } | b + b
| fX
basic value
extended value
atom
behavioural type
where c, c′ , · · · range over cog names and f , f ′ , · · · range over future names. Basic
values are either generic (non-size) expressions -- or size expressions se or the type
c[se] of an object of cog c with capacity se. The extended values add future names to
basic values.
Atoms define creation of cogs (νc[se]), synchronous and asynchronous method
invocations (m( ) → and ν f : m( ) → , respectively), and synchronizations on asynchronous invocations ( f X ). We observe that cog creations always carry a capacity,
which has to be a size expression because our analysis in the next section cannot deal
t
a
t
t
t
t
b
b b
a b
with generic expressions. Behavioural types are sequences of atoms # ′ or conditionals, typically (se){ } + (¬se){ ′ } or + ′ , according to whether the boolean
guard is a size expression that depends on the arguments of a method or not. In order to
type sequential composition in a precise way (see rule T-Seq), the leaves of behavioural
types are labelled with environments, ranged over by Γ, Γ ′ , · · · . Environments are maps
from method names m to terms ( ) → , from variables to extended values , and from
future names to values that are either or X .
The abstract behaviour of methods is defined by method behavioural types of the
form: m( t , ){ } : r , where t is the type value of the receiver of the method, are the
type value of the arguments, is the abstract behaviour of the body, and r is the type
value of the returned object. The subterm t , of the method contract is called header;
r is called returned type value. We assume that names in the header occur linearly.
Names in the header bind the names in and in r . The header and the returned type
value, written ( t , ) → r , are called behavioural type signature. Names occurring in
or r may be not bound by header. These free names correspond to new cog creations
and will be replaced by fresh cog names during the analysis. We use to range over
method behavioural types.
The type system uses judgments of the following form:
b
b
t
t t b t
t
b
t
t t
t
t
t
t t
x
t
t t
b
t
t
b
C
x
t t
x a
t
t
– Γ ⊢ e : for pure expressions e, Γ ⊢ f : or Γ ⊢ f : X for future names f , and
Γ ⊢ m( ) : for methods.
– Γ ⊢ z : , [ ⊲ Γ ′ ] for expressions with side effects z, where is the value, ⊲ Γ ′
is the corresponding behavioural type, where Γ ′ is the environment Γ with possible
updates of variables and future names.
– Γ ⊢ s : , in this case the updated environments Γ ′ are inside the behavioural type,
in correspondence of every branch of its.
x
a
b
Since Γ is a function, we use the standard predicates x ∈ dom(Γ) or x < dom(Γ).
Moreover, we define
(
def
if y = x
Γ[x 7→ ](y) =
Γ(y) otherwise
x
x
The multi-hole contexts C[ ] are defined by the following syntax:
a # C[ ] | C[ ] + C[ ] | (se){ C[ ] }
and, whenever b = C[a1 ⊲ Γ1 ] · · · [an ⊲ Γn ], then b[x 7→ x] is defined as C[a1 ⊲ Γ1 [x 7→
x]] · · · [an ⊲ Γn[x 7→ x]].
C[ ] ::=
[]
|
The typing rules for expressions are defined in Figure 3. These rules are not standard
because (size) expressions containing method’s arguments are typed with the expressions themselves. This is crucial to the cost analysis in Section 5. In particular, cog creation is typed by rule T-New, with value c[se], where c is the fresh name associated with
the new cog and se is the value associated with the declared capacity. The behavioural
type for the cog creation is νc[se] ⊲ Γ[c 7→ se], where the newly created cog is added
to Γ. In this way, it is possible to verify whether the receiver of a method invocation is
within a locally created cog or not by testing whether the receiver belongs to dom(Γ) or
not, respectively (cf. rule T-Invoke). Object creation (cf. rule T-New-Local) is typed as
the cog creation, with the exception that the cog name and the capacity value are taken
(T-Method)
(T-Var)
x ∈ dom(Γ)
Γ ⊢ x : Γ(x)
(T-Se)
(T-Nse)
Γ ⊢ se : se
Γ ⊢ nse : --
t
t
t
t
Γ(m) = ( ) → ′
fv( ′ ) \ fv( ) , ∅ implies σ( ′ ) fresh
t
t
t
Γ ⊢ m(σ( )) : σ( ′ )
(T-New)
(T-New-Local)
Γ ⊢ e : se
Γ ⊢ this : c[se]
c fresh
Γ ⊢ new Class with e : c[se],
[νc[se] ⊲ Γ[c 7→ se]]
(T-Invoke-Sync)
Γ ⊢ new local Class : c[se],
[0 ⊲ Γ ]
(T-Invoke)
Γ ⊢ e : c[se] Γ(this) = c[se]
Γ⊢e:
Γ ⊢ m(c[se], ) : ′
Γ ⊢ e : c[se] (c ∈ dom(Γ) or Γ(this) = c[se])
Γ⊢e:
Γ ⊢ m(c[se], ) : ′ f fresh
t
t t
Γ ⊢ e.m(e) : t , [m(c[se], t) → t ⊲ Γ ]
′
t
t t
Γ ⊢ e!m(e) : f, [ν f : m(c[se], t) → t
′
(T-Get)
Γ⊢e: f
t
Γ ⊢ e.get : ,
Γ( f ) =
[ f X ⊲ Γ[ f
′
t
7→
(T-Get-Top)
t
Γ⊢e: f
X
]]
Γ( f ) =
t
Γ ⊢ e.get : ,
⊲ Γ[ f 7→
t ]]
′
t
X
[0 ⊲ Γ ]
Fig. 3. Typing rules for expressions
from the local cog and the behavioural type is empty. Rule T-Invoke types method invocations e!m(e) by using a fresh future name f that is associated to the method name, the
cog name of the callee and the arguments. In the updated environment, f is associated
with the returned value. Next we discuss the constraints in the premise of the rule. As
we discussed in Section 2, asynchronous invocations are allowed on callees located in
the current cog, Γ(this) = c[se], or on a newly created object which resides in a fresh
cog, c ∈ dom(Γ). Rule T-Get defines the synchronization with a method invocation that
corresponds to a future f . The expression is typed with the value of f in the environment and behavioural type f X . Γ is then updated for recording that the synchronization
has been already performed, thus any subsequent synchronization on the same value
would not imply any waiting time (see that in rule T-Get-Top the behavioural type is
0). The synchronous method invocation in rule T-Invoke-Sync is directly typed with the
return value ′ of the method and with the corresponding behavioural type. The rule
enforces that the cog of the callee coincides with the local one.
The typing rules for statements are presented in Figure 4. The behavioural type in
rule T-Job expresses the time consumption for an object with capacity se′ to perform se
processing cycles: this time is given by se/se′ , which we observe is in general a rational
number. We will return to this point in Section 5.
The typing rules for method and class declarations are shown in Figure 5.
t
t
Examples The behavioural type of the fib method discussed in Section 2 is
fib (c[x],n) {
(n ≤ 1){ 0 ⊲ ∅ }
+
(T-Assign)
x [a ⊲ Γ ]
Γ ⊢ x = rhs : a ⊲ Γ [x 7→ x]
(T-Job)
Γ ⊢ e : se Γ ⊢ this : c[se′ ]
Γ ⊢ job(e) : se/se′ ⊲ Γ
′
Γ ⊢ rhs : ,
′
(T-Seq)
t
(T-Return)
Γ⊢e:
Γ ⊢ destiny :
Γ ⊢ return e : 0 ⊲ Γ
b
(T-If-Nse)
t
Γ ⊢ s ; s′
b
b b
Γ ⊢ s′ :
Γ ⊢ e : -- Γ ⊢ s :
′
Γ ⊢ if e { s } else { s } : +
a
a ⊲Γ ]
b
: C[a # b ] · · · [a # b ]
Γ ⊢ s : C[ 1 ⊲ Γ1 ] · · · [
Γi ⊢ s′ : ′i
n
n
′
1
1
n
b
(T-If-Se)
′
n
b
Γ ⊢ e : se Γ ⊢ s :
Γ ⊢ s′ : ′
′
Γ ⊢ if e { s } else { s } : (se){ } + (¬se){ ′ }
′
′
b
b
Fig. 4. Typing rules for statements
(T-Method)
Γ(m) = (t , t) → t
t ][destiny 7→ t ][x 7→ t] ⊢ s : C[a ⊲ Γ ] · · · [a
Γ ⊢ T m (T x) { s } : m(t , t){ C[a ⊲ ∅] · · · [a ⊲ ∅] } : t
t
Γ[this 7→
t
r
r
1
t
(T-Class)
Γ⊢M:
C
1
n
t
Γ[this 7→ start[k]][x 7→ ] ⊢ s : C[
Γ ⊢ M { T x ; s } with k :
1
C, C[a
1
a
1
n
⊲ Γn ]
r
⊲ Γ1 ] · · · [
⊲ ∅] · · · [
a
n
a
n
⊲ Γn ]
⊲ ∅]
Fig. 5. Typing rules for declarations
(n ≥ 2){
1/x # d[x] # ν f : fib (c[x],n -1)→ -- # νg: fib (d[x],n -2)→ -- #
f X # gX #0 ⊲ ∅
}
} : --
5 The time analysis
The behavioural types returned by the system defined in Section 4 are used to compute
upper bounds of time complexity of a tml program. This computation is performed by
an off-the-shelf solver – the CoFloCo solver [4] – and, in this section, we discuss the
translation of a behavioural type program into a set of cost equations that are fed to the
solver. These cost equations are terms
m(x) = exp
[se]
where m is a (cost) function symbol, exp is an expression that may contain (cost) function symbol applications (we do not define the syntax of exp, which may be derived
by the following equations; the reader may refer to [4]), and se is a size expression
whose variables are contained in x. Basically, our translation maps method types into
cost equations, where (i) method invocations are translated into function applications,
and (ii) cost expressions se occurring in the types are left unmodified. The difficulties
of the translation is that the cost equations must account for the parallelism of processes
in different cogs and for sequentiality of processes in the same cog. For example, in the
following code:
x = new Class with c; y = new Class with d;
f = x!m(); g = y!n(); u = g. get ; u = f. get ;
the invocations of m and n will run in parallel, therefore their cost will be max(t, t′ ),
where t is the time of executing m on x and t′ is the time executing n on y. On the
contrary, in the code
x = new local Class ; y = new local Class;
f = x!m(); g = y!n(); u = g. get ; u = f. get ;
the two invocations are queued for being executed on the same cog. Therefore the time
needed for executing them will be t +t′ , where t is time needed for executing m on x, and
t′ is the time needed for executing n on y. To abstract away the execution order of the
invocations, the execution time of all unsynchronized methods from the same cog are
taken into account when one of these methods is synchronized with a get-statement.
To avoid calculating the execution time of the rest of the unsynchronized methods in
the same cog more than necessary, their estimated cost are ignored when they are later
synchronized.
In this example, when the method invocation y!n() is synchronized with g.get,
the estimated time taken is t + t′ , which is the sum of the execution time of the two
unsynchronized invocations, including the time taken for executing m on x because both
x and y are residing in the same cog. Later when synchronizing the method invocation
x!m(), the cost is considered to be zero because this invocation has been taken into
account earlier.
The translate function. The translation of behavioural types into cost equations is carried out by the function translate, defined below. This function parses atoms, behavioural types or declarations of methods and classes. We will use the following auxiliary function that removes cog names from (tuples of) terms:
t
⌊⌋=
⌊e⌋ = e
⌊c[e]⌋ = e
t
⌊ 1, . . . ,
tn⌋ = ⌊t1⌋, . . . , ⌊tn⌋
We will also use translation environments, ranged over by Ψ , Ψ ′ , · · · , which map future
names to pairs (e, m( )) that records the (over-approximation of the) time when the
method has been invoked and the invocation.
In the case of atoms, translate takes four inputs: a translation environment Ψ ,
the cog name of the carrier, an over-approximated cost e of an execution branch, and
the atom . In this case, translate returns an updated translation environment and the
t
a
cost. It is defined as follows.
a
translate(Ψ, c, e, ) =
(Ψ, e + e′ )
(Ψ, e)
(Ψ, e + m(⌊ ⌋))
(Ψ [ f 7→ (e, m( ))], e)
(Ψ
\ F, e + e1 )))
(Ψ \ F, max(e, e1 + e2 ))
(Ψ, e)
t
t
a=e
when a = νc[e ]
when a = m(t) → t
when a = (ν f : m(t) → t )
and Ψ ( f ) = (e , m (c[e ], t ))
when a = f
let F = { g | Ψ (g) = (e , m (c[e ], t )) } then
P
and e = { m (⌊t ⌋) | (e , m (t )) ∈ Ψ (F) }
when a = f and Ψ ( f ) = (e , m (c [e ], t )) and c , c
let F = { g | Ψ (g) = (e , m (c [e ], t )) } then
P
e = { m (⌊t ⌋) | (e , m (t )) ∈ Ψ (F) }
and e = max{ e | (e , m (t )) ∈ Ψ (F) }
when a = f and f < dom(Ψ )
′
when
′
′
′
X
f
g
1
g
′
g
g
′
g
g
′
g
g
2
g
g
g
g
g
g
f
g
g
′
g
f
′
′
X
1
′
f
f
′
′
′
f
g
′
g
′
g
X
The interesting case of translate is when the atom is f X . There are three cases:
1. The synchronization is with a method whose callee is an object of the same cog.
In this case its cost must be added. However, it is not possible to know when the
method will be actually scheduled. Therefore, we sum the costs of all the methods
running on the same cog (worst case) – the set F in the formula – and we remove
them from the translation environment.
2. The synchronization is with a method whose callee is an object on a different cog c′ .
In this case we use the cost that we stored in Ψ ( f ). Let Ψ ( f ) = (e f , m f (c′ [e′ ], f )),
then e f represents the time of the invocation. The cost of the invocation is therefore
e f + m f (e′ , ⌊ f ⌋). Since the invocation is in parallel with the thread of the cog c, the
overall cost will be max(e, e f + m f (e′ , ⌊ f ⌋)). As in case 1, we consider the worst
scheduler choice on c′ . Therefore, instead of taking e f + m f (e′ , ⌊ f ⌋), we compute
the cost of all the methods running on c′ – the set F in the formula – and we remove
them from the translation environment.
3. The future does not belong to Ψ . That is the cost of the invocation which has been
already computed. In this case, the value e does not change.
t
t
t
t
In the case of behavioural types, translate takes as input a translation environment, the cog name of the carrier, an over-approximated cost of the current execution
branch (e1 )e2 , where e1 indicates the conditions corresponding to the branch, and the
behavioural type .
a
translate(Ψ, c, (e1 )e2 ,
{ (Ψ ′ , (e1 )e′2 ) }
C
C ∪ C′
C
b)
=
when
b = a⊲Γ
when b = a # b
a
and translate(Ψ, c, e , a) = (Ψ , e )
and translate(Ψ , c, (e )e , b ) = C
when b = b + b and translate(Ψ, c, (e )e , b ) = C
and translate(Ψ, c, (e )e , b ) = C
when b = (e){ b } and translate(Ψ, c, (e ∧ e)e , b ) = C
and
translate(Ψ, c, e2 , ) = (Ψ ′ , e′2 )
′
1
′
′
2
1
2
1
1
′
′
2
′
2
2
2
′
2
1
′
1
2
′
The translation of the behavioural types of a method is given below. Let dom(Ψ ) =
def
{ f1 , · · · , fn }. Then we define Ψ X = f1 X # · · · # fn X .
t b } : t)
translate(m(c[e], ){
=
b
m(e, e) = e′1 + e′′1
..
.
m(e, e) = e′n + e′′n
[e1 ]
[en ]
t
where translate(∅, c, 0, ) = { Ψi , (ei )e′i | 1 ≤ i ≤ n }, and e = ⌊ ⌋,
and e′′i = translate(Ψi , c, 0, Ψi X ⊲ ∅) .
In addition, [ei ] are the conditions for branching the possible execution paths of method
m(e, e), and e′i + e′′i is the over-approximation of the cost for each path. In particular,
e′i corresponds to the cost of the synchronized operations in each path (e.g., jobs and
gets), while e′′i corresponds to the cost of the asynchronous method invocations triggered by the method, but not synchronized within the method body.
Examples We show the translation of the behavioural type of fibonacci presented in
= (se){0 ⊲ ∅} + (¬se){ ′ }, where se = (n ≤ 1) and ′ = 1/e #
Section 4. Let
ν f : fib(c[e], n − 1) → -- # νg: fib(c′ [e], n − 2) → -- # f X # gX # 0 ⊲ ∅}. Let also
Ψ = Ψ1 ∪Ψ2 , where Ψ1 = [ f 7→ (1/e, fib(e, n−1))] and Ψ2 = [g 7→ (1/e, fib(e, n−2))].
The following equations summarize the translation of the behavioural type of the
fibonacci method.
b
b
b
b
translate(∅, c, 0, )
= translate(∅, c, 0, (se) { 0 ⊲ ∅ }) ∪ translate(∅, c, 0, (¬se) { ′ })
= translate(∅, c, (se)0, { 0 ⊲ ∅ }) ∪ translate(∅, c, (¬se)0, { 1/e # . . . })
= { (se)0 } ∪ translate(∅, c, (¬se)(1/e), { ν f : fib(c[e], n − 1) → -- # . . . })
= { (se)0 } ∪ translate(Ψ1 , c, (¬se)(1/e), { νg: fib(c′ [e], n − 2) → -- # . . . })
= { (se)0 } ∪ translate(Ψ, c, (¬se)(1/e), { f X # gX # . . . })
= { (se)0 } ∪ translate(Ψ2 , c, (¬se)(1/e + fib(e, n − 1)), { gX # . . . })
= { (se)0 } ∪ translate(∅, c, (¬se)(1/e + max(fib(e, n − 1), fib(e, n − 2))), { 0 ⊲ ∅ })
= { (se)0 } ∪ { (¬se)(1/e + max(fib(e, n − 1), fib(e, n − 2))) }
b
translate(∅, c, 0, 0) = (∅, 0)
translate(∅, c, 0, 1/e) = (∅, 1/e)
translate(∅, c, 1/e, ν f : fib(c[e], n − 1) → -- ) = (Ψ1 , 1/e)
translate(Ψ1 , c, 1/e, νg: fib(c′ [e], n − 2) → -- ) = (Ψ, 1/e)
translate(Ψ, c, 1/e, f X ) = (Ψ2 , 1/e + fib(e, n − 1))
translate(Ψ2 , c, 1/e + fib(e, n − 1), gX ) = (∅, 1/e + max(fib(e, n − 1), fib(e, n − 2)))
b
translate(fib
(c[e], n){ } : --) =
fib(e, n) = 0
fib(e, n) = 1/e + max(fib(e, n − 1), fib(e, n − 2))
[n ≤ 1]
[n ≥ 2]
Remark 1. Rational numbers are produced by the rule T-Job of our type system. In
particular behavioural types may manifest terms se/se′ where se gives the processing
cycles defined by the job operation and se′ specifies the number of processing cycles
per unit of time the corresponding cog is able to handle. Unfortunately, our backend
solver – CoFloCo – cannot handle rationals se/se′ where se′ is a variable. This is the
case, for instance, of our fibonacci example, where the cost of each iteration is 1/x,
where x is a parameter. In order to analyse this example, we need to determine a priori
the capacity to be a constant – say 2 –, obtaining the following input for the solver:
eq(f(E,N),0,[],[-N>=1,2*E=1]).
eq(f(E,N),nat(E),[f(E,N-1)],[N>=2,2*E=1]).
eq(f(E,N),nat(E),[f(E,N-2)],[N>=2,2*E=1]).
Then the solver gives the following upper bound:
nat(N-1)* (1/2).
It is worth to notice that fixing the fibonacci method is easy because the capacity
does not change during the evaluation of the method. This is not always the case, as in
the following alternative definition of fibonacci:
Int fib_alt ( Int n) {
if (n <=1) { return 1; }
else { Fut <Int > f; Class z; Int m1; Int m2;
job (1);
z = new Class with ( this . capacity *2) ;
f = this ! fib_alt (n -1); g = z! fib_alt (n -2);
m1 = f. get ; m2 = g. get ;
return m1+m2; }
}
In this case, the recursive invocation z!fib alt(n-2) is performed on a cog with
twice the capacity of the current one and CoFloCo is not able to handle it. It is worth to
observe that this is a problem of the solver, which is otherwise very powerful for most
of the examples. Our behavioural types carry enough information for dealing with more
complex examples, so we will consider alternative solvers or combination of them for
dealing with examples like fib alt.
6 Properties
In order to prove the correctness of our system, we need to show that (i) the behavioural
type system is correct, and (ii) the computation time returned by the solver is an upper
bound of the actual cost of the computation.
The correctness of the type system in Section 4 is demonstrated by means of a
subject reduction theorem expressing that if a runtime configuration cn is well typed and
cn → cn′ then cn′ is well-typed as well, and the computation time of cn is larger or equal
to that of cn′ . In order to formalize this theorem we extend the typing to configurations
and we also use extended behavioural types with the following syntax
k ::= b
| [
b
]cf
|
kkk
k
runtime behavioural type
b
k
The type [ ]cf expresses the behaviour of an asynchronous method bound to the future f
and running in the cog c; the type k ′ expresses the parallel execution of methods
in and in ′ .
We then define a relation Dt between runtime behavioural types that relates types.
The definition is algebraic, and Dt ′ is intended to mean that the computational time
of is at least that of ′ + t (or conversely the computational time of ′ is at most that
of − t). This is actually the purpose of our theorems.
k
k k
k
k
k
k k
k
k
k′
k′
Theorem 1 (Subject Reduction). Let cn be a configuration of a tml program and let
be its behavioural type. If cn is not strongly t-stable and cn → cn′ then there exists
typing cn′ such that D0 ′ . If cn is strongly t-stable and cn → cn′ then there exists
typing cn′ such that Dt ′ .
k k
k k
The proof of is a standard case analysis on the last reduction rule applied.
The second part of the proof requires an extension of the translate function to
runtime behavioural types. We therefore define a cost of the equations Ek returned by
translate( ) – noted cost(Ek ) – by unfolding the equational definitions.
k
Theorem 2 (Correctness). If
k Dt k′, then cost(Ek)
≥ cost(Ek′ ) + t.
As a byproduct of Theorems 1 and 2, we obtain the correctness of our technique, modulo the correctness of the solver.
7 Related work
In contrast to the static time analysis for sequential executions proposed in [7], the paper
proposes an approach to analyse time complexity for concurrent programs. Instead of
using a Hoare-style proof system to reason about end-user deadlines, we estimate the
execution time of a concurrent program by deriving the time-consuming behaviour with
a type-and-effect system.
Static time analysis approaches for concurrent programs can be divided into two
main categories: those based on type-and-effect systems and those based on abstract
interpretation – see references in [9]. Type-and-effect systems (i) collect constraints on
type and resource variables and (ii) solve these constraints. The difference with respect
to our approach is that we do not perform the analysis during the type inference. We
use the type system for deriving behavioural types of methods and, in a second phase,
we use them to run a (non compositional) analysis that returns cost upper bounds. This
dichotomy allows us to be more precise, avoiding unification of variables that are performed during the type derivation. In addition, we notice that the techniques in the literature are devised for programs where parallel modules of sequential code are running.
The concurrency is not part of the language, but used for parallelising the execution.
Abstract interpretation techniques have been proposed addressing domains carrying
quantitative information, such as resource consumption. One of the main advantages of
abstract interpretation is the fact that many practically useful optimization techniques
have been developed for it. Consequently, several well-developed automatic solvers for
cost analysis already exist. These techniques either use finite domains or use expedients
(widening or narrowing functions) to guarantee the termination of the fix-point generation. For this reason, solvers often return inaccurate answers when fed with systems
that are finite but not statically bounded. For instance, an abstract interpretation technique that is very close to our contribution is [2]. The analysis of this paper targets
a language with the same concurrency model as ours, and the backend solver for our
analysis, CoFloCo, is a slightly modified version of the solver used by [2]. However the
two techniques differ profoundly in the resulting cost equations and in the way they are
produced. Our technique computes the cost by means of a type system, therefore every
method has an associated type, which is parametric with respect to the arguments. Then
these types are translated into a bunch of cost equations that may be composed with
those of other methods. So our approach supports a technique similar to separate compilation, and is able to deal with systems that create statically an unbounded but finite
number of nodes. On the contrary, the technique in [2] is not compositional because it
takes the whole program and computes the parts that may run in parallel. Then the cost
equations are generated accordingly. This has the advantage that their technique does
not have any restriction on invocations on arguments of methods that are (currently)
present in our one.
We finally observe that our behavioural types may play a relevant role in a cloud
computing setting because they may be considered as abstract descriptions of a method
suited for SLA compliance.
8 Conclusions
This article presents a technique for computing the time of concurrent object-oriented
programs by using behavioural types. The programming language we have studied features an explicit cost annotation operation that define the number of machine cycles
required before executing the continuation. The actual computation activities of the program are abstracted by job-statements, which are the unique operations that consume
time. The computational cost is then measured by introducing the notion of (strong)
t-stability (cf. Definition 1), which represents the ticking of time and expresses that up
to t time steps no control activities are possible. A Subject Reduction theorem (Theorem 1), then, relates this stability property to the derived types by stating that the
consumption of t time steps by job statements is properly reflected in the type system.
Finally, Theorem 2 states that the solution of the cost equations obtained by translation
of the types provides an upper bound of the execution times provided by the type system
and thus, by Theorem 1, of the actual computational cost.
Our behavioural types are translated into so-called cost equations that are fed to a
solver that is already available in the literature – the CoFloCo solver [4]. As discussed
in Remark 1, CoFloCo cannot handle rational numbers with variables at the denominator. In our system, this happens very often. In fact, the number pc of processing
cycles needed for the computation of a job(pc) is divided by the speed s of the machine
running it. This gives the cost in terms of time of the job(pc) statement. When the capacity is not a constant, but depends on the value of some parameter and changes over
time, then we get the untreatable rational expression. It is worth to observe that this
is a problem of the solver (otherwise very powerful for most of the examples), while
our behavioural types carry enough information for computing the cost also in these
cases. We plan to consider alternative solvers or a combination of them for dealing with
complex examples.
Our current technique does not address the full language. In particular we are still
not able to compute costs of methods that contain invocations to arguments which do not
live in the same machine (which is formalized by the notion of cog in our language). In
fact, in this case it is not possible to estimate the cost without any indication of the state
of the remote machine. A possible solution to this issue is to deliver costs of methods
that are parametric with respect to the state of remote machines passed as argument. We
will investigate this solution in future work.
In this paper, the cost of a method also includes the cost of the asynchronous invocations in its body that have not been synchronized. A more refined analysis, combined
with the resource analysis of [5], might consider the cost of each machine, instead of
the overall cost. That is, one should count the cost of a method per machine rather than
in a cumulative way. While these values are identical when the invocations are always
synchronized, this is not the case for unsynchronized invocation and a disaggregated
analysis might return better estimations of virtual machine usage.
References
1. E. Albert, F. S. de Boer, R. Hähnle, E. B. Johnsen, R. Schlatte, S. L. Tapia Tarifa, and P. Y. H.
Wong. Formal modeling of resource management for cloud architectures: An industrial
case study using Real-Time ABS. Journal of Service-Oriented Computing and Applications,
8(4):323–339, 2014.
2. E. Albert, J. C. Fernández, E. B. Johnsen, and G. Román-Dı́ez. Parallel cost analysis of
distributed systems. In Proceedings of SAS 2015, volume 9291 of Lecture Notes in Computer
Science. Springer, 2015. To appear.
3. R. Buyya, C. S. Yeo, S. Venugopal, J. Broberg, and I. Brandic. Cloud computing and emerging
IT platforms: Vision, hype, and reality for delivering computing as the 5th utility. Future
Generation Comp. Sys., 25(6):599–616, 2009.
4. A. Flores Montoya and R. Hähnle. Resource analysis of complex programs with cost equations. In Proceedings of 12th Asian Symposium on Programming Languages and Systems,
volume 8858 of Lecture Notes in Computer Science, pages 275–295. Springer, 2014.
5. A. Garcia, C. Laneve, and M. Lienhardt. Static analysis of cloud elasticity. In Proceedings of
PPDP 2015, 2015.
6. R. Hähnle and E. B. Johnsen. Resource-aware applications for the cloud. IEEE Computer,
48(6):72–75, 2015.
7. E. B. Johnsen, K. I. Pun, M. Steffen, S. L. Tapia Tarifa, and I. C. Yu. Meeting deadlines,
elastically. In From Action Systems to Distributed Systems: the Refinement Approach. CRC
Press, 2015. To Appear.
8. E. B. Johnsen, R. Schlatte, and S. L. Tapia Tarifa. Integrating deployment architectures and
resource consumption in timed object-oriented models. Journal of Logical and Algebraic
Methods in Programming, 84(1):67–91, 2015.
9. P. W. Trinder, M. I. Cole, K. Hammond, H. Loidl, and G. Michaelson. Resource analyses for
parallel and distributed coordination. Concurrency and Computation: Practice and Experience, 25(3):309–348, 2013.
| 6cs.PL |
Coverage and Rate of Downlink Sequence Transmissions with
Reliability Guarantees
arXiv:1704.05296v2 [cs.IT] 24 Jun 2017
Jihong Park and Petar Popovski
Abstract—Real-time distributed control is a promising application of 5G in which communication links should satisfy
certain reliability guarantees. In this letter, we derive closedform maximum average rate when a device (e.g. industrial
machine) downloads a sequence of n operational commands
through cellular connection, while guaranteeing a certain signalto-interference ratio (SIR) coverage for all n messages. The result
is based on novel closed-form n-successive SIR coverage bounds.
The proposed bounds provide simple approximations that are
increasingly accurate in the high reliability region.
Index Terms—Sequence transmissions, ultra-reliable communication, coverage, rate, stochastic geometry.
I. M OTIVATION AND C ONTRIBUTION
previous
commands
current
commands
interference
static machine
mobile machine
Fig. 1. A sequence of 3 command receptions for mobile and static UEs.
UE locations, the n-successive SIR coverage has been known
as (Theorem 2, [4]):
One of the focus in the emerging 5G wireless systems
will be communication with reliability guarantees. The most
challenging variant is ultra-reliable low-latency communication (URLLC), with extreme reliability guarantees. A representative use case for URLLC is industrial Internet of Things
(IoT) [1], in which a certain user equipment (UE), such as an
actuator, is controlled in real time by downloading operational
commands for high-precision cloud manufacturing [2]. A
loss of a command sequence may cause critical malfunction
or shutdown of the process. Therefore, the network should
guarantee wireless coverage with very high reliability, e.g.
≥ 99% in order to decode the n successive commands.
In this letter we rely on stochastic geometry to derive nsuccessive SIR (Signal-to-Interference Ratio) coverage pn (t)
for a target SIR threshold t and n ≥ 1, defined as pn (t) :=
Pr(SIR1 ≥ t, SIR2 ≥ t, · · · , SIRn ≥ t). We thereby derive
the closed-form maximum average rate Rn (η) that ensures η
reliability for n-successive message receptions, i.e. pn (t) ≥ η.
Regardless of whether the reliability is extreme as in URLLC
or more relaxed, the introduction of a reliability guarantee
poses a challenging problem. This precludes us from directly
exploiting previously known closed-form SIR coverage and
rate expressions [3]–[5]. The reason comes from the following
two issues: spatial correlation during n-successive receptions
and the trade-off between coverage and rate.
Regarding spatial correlation, if UE locations are independent and identically distributed (i.i.d.) during n message
receptions, then pn (t) = p1 (t)n straightforwardly.1 However,
a URLLC UE should receive n messages within a short
time span, during which all UEs and the interferers can be
considered static, see Fig. 1. For a cellular downlink with fixed
Although (2) is a simple convex optimizaition with an inequality condition, it cannot be analytically solved due to the
intractable hypergeometric function of pn (t) in (1).
A suitable way to avoid the difficulties brought by the hypergeometric function is to utilize more tractable SIR coverage
J. Park and P. Popovski are with Department of Electronic Systems, Aalborg
University, Denmark (email: {jihong, petarp}@es.aau.dk).
This work has been in part supported by the European Research Council
(ERC Consolidator Grant Nr. 648382 WILLOW) within the Horizon 2020
Program, and partly by Innovation Fund Denmark via the Virtuoso project.
1 For a typical UE’s SIR, i.i.d. mobility is almost surely identical to
any Markovian mobility such as a random walk and Lévy flight thanks to
displacement theorem [6] as utilized in [7] (see [8] for general
mobility).
2
2
2 A special case is p (t) = 1/ F
1 R
2 1 1, − α ; 1 − α ; −t in [4], which is
identical to 1/(1 + t2/α u>t−2/α du/[1 + uα/2 ]) presented in [3].
pn (t) = 1/2 F1 (n, −2/α; 1 − 2/α; −t)
(1)
2
at a randomly picked (i.e. a typical) UE. The denominator
2 F1 (a, b; c; z) is a Gauss hypergeometric function that can
only be numerically computed, defined as 2 F1 (a, b; c; z) :=
P∞ Γ(a+n)Γ(b+n)Γ(c) zn
n=0 Γ(a)Γ(b)Γ(c+n) n! where Γ(x) denotes the gamma
function. It is thus difficult to establish a direct relation
between n and pn (t) due to the analytical intractability of
the hypergeometric function.
Next, the average rate expression E[log(1+SIR)] in previous
works [3]–[5] bore no relationship to the reliability constraint
pn (t) ≥ η. The reason is because in those works it was
implicitly assumed that each BS knows the full channel state
information (CSI) of all the desired and interfering links. This
enables the BS to adaptively adjust its modulation and coding
scheme (MCS) such that t = SIR resulting in pn (SIR) = 1.
The full CSI unfortunately cannot be exchanged within a low
target latency, and is therefore not affordable for URLLC.
Without CSI, the SIR threshold is kept constant during n
transmissions, while satisfying pn (t) ≥ η. This induces the
trade-off between a typical UE’s coverage pn (t) and its quasiconcave average rate pn (t) log(1 + t) with respect to t. We
focus on the maximum average rate Rn (η) given as
Rn (η) := arg max pn (t) log(1 + t) s.t. pn (t) ≥ η.
(2)
t≥0
n=3, a=4
n=3, a=4
11
1
0.7
0.6
+
0.5
LBa , UBa Lemma 1
LBb , UBb Lemma 2
LBc , UBc Theorem 1
LB⇥
Corollary 1
LB+
0.99
Exact
Bound in [5]
0.4
0.3
0.97
0.6
0.96
0.93
0.91
0.9
0
0.5
0.5 0.8
SIR Coverage, pn (t)
0
1
1.51
1.5 2.5
3 2
3.5
2.5 4
4.5 3
0
0.9
00
5
LB
0.5
0.005
0.01
1
0.015
1.5 0.02
LBc
LB
+
b
2
0.025
LB
2.5
0.03
⇥
0.035 3
0.04
SIRSIR
Threshold,
t t
Threshold,
SIR
Threshold,
t t
SIR
Threshold,
0.7
0.6
2
Exact
0.4
0.94
1
0.1
UBc
0.95
0.2
0.92
0.2
UBa
0.8
0.98
Bound in [5]
SIR
pnn(t)
(t)
SIR Coverage,
Coverage, pp
(t)
n
0.8
+
(t)
SIR Coverage, ppn(t)
n
0.9
(a) for the range within [0, 1].
(b) for the range within [0.9, 1].
0.5
Fig. 2. 3-successive
SIR coverage bounds (Theorem 1, α = 4), derived via the bounds after Pfaff transformation (Lemma 1) and the asymptotic bounds
0.4
(Lemma 2). They are compared to the existing bounds 1/(1 + cn t2/α ) in [5] and cn −1 t−2/α (= UBa ) in [9], [10]. The exact curve follows from (1).
0.3
+
0.2
0.1
bounds, as done
in our recent work [11]. Nonetheless, the
0
0
1
1.5
3
4
bounds in [11]
are0.5limited
only
to2 the 2.5
case n
= 3.51. Other
SIR Threshold,
existing SIR coverage bounds [5], [9],
[10] aret too loose for
ultra-reliable scenarios, within the range [0.9, 1] for low SIR
threshold t, as Fig. 2 shows.
In this letter we propose novel closed-form bounds of nsuccessive SIR coverage pn (t) (see LBc and UBc in Theorem 1), and thereby derive the closed-form bounds of maximum rate Rn (η) guaranteeing pn (t) ≥ η (Proposition 1).
For high reliability η ≥ 0.9, the closed-form results show that
the impact of spatial UE correlation during n receptions is
negligibly small. In this case the proposed lower bound under
i.i.d. UE locations tightly approximates the exact rate (LB× in
Corollaries 1 and 2). For moderate reliability 0.3 < η < 0.9,
a slightly modified lower bound becomes accurate as an
approximation (LB+ in Corollaries 1 and 2).
II. S YSTEM M ODEL
Consider a downlink cellular network in a two-dimensional
plane. The network comprises uniformly located: BSs with
density λb and UEs with density λu . This leads to two
independent homogeneous Poisson point processes Φ(λb ) and
Φ(λu ), representing BS and UE locations. Each UE associates
with its nearest BS, and requests n messages. All messages
are of unit size. UE locations are fixed during the n-successive
message receptions, and randomly and independently changed
for the next n-successive message receptions.
Each BS serves a single UE with unit transmission power
and frequency bandwidth per time slot. We assume that
λu λb such that any BS has at least a single associated
UE to be served. Transmitted signals experience path loss
attenuation with the exponent α > 2 and Rayleigh fading. A
typical UE is located at the origin, capturing any locations
without SIR changes according to Slyvnyak’s theorem [6].
The i-th message, 1 P
≤ i ≤ n, at the typical UE is received
with SIRi := g i r−α / rj >r gji rj −α , where rj := |xj | denotes
the j-th nearest interfering BS distance for its location xj ∈
Φ(λb ). Fading power gji at the i-th reception is exponentially
distributed with unit mean. We drop the subscript j for the
desired signal link. When i changes from 1 to n, the fading
5
powers
g i and gji vary with each i, while the desired BS and
interferer distances r and rj are fixed. In this case, pn (t) has
been known as (1) according to Theorem 2 in [4].
4.5
III. n-S UCCESSIVE SIR C OVERAGE B OUNDS
In this section we propose closed-form bounds of nsuccessive SIR coverage pn (t). The bounds are not only much
tighter for URLLC but also as simple as the existing bounds.
The desired bounds are derived by utilizing a transformed
function of pn (t) and pn (t)’s asypmtotic curves as follows.
We define the lower and the upped bounds of pn (t) through
UBi
i
pLB
n (t) ≤ pn (t) ≤ pn (t) for i ∈ {a, b, c, ×, +}. We then
apply a hypergeometric function transform [12] to (1), yielding
the following lemma.
Lemma 1. (Pfaff Transformed Bounds) At a static typical UE,
pn (t) has the following lower and upper bounds.
2
−1
a
pLB
(1 + t)− α
n (t) = cn
a
pUB
n (t)
(3)
2
−α
= (1 + t)
(4)
where cn := Γ 1 − α Γ n + α2 /Γ(n)
Proof: By using Pfaff’s hypergeometric
function transform
2
[12], 2 F1 n, − α2 ; 1− α2 , −t = (1 + t) α G(α, n,
t) where
2
2
2
t
G(α, n, t) := 2 F1 1 − α + n , − α ; 1 − α ; 1+t that has
the range [1, cn ]. Applying this to (1) leads to the result.
2
The term cn ≥ 1 is a monotone increasing/decreasing
function of n and
α respectively. Its minimum value is thus
2π α→∞
csc
= 1.
c1 = 2π
α
α
It is remarkable that both bounds identically share the
2
simple term (1 + t)− α only with the constant multiplication
difference. Nevertheless, they have the following limitations.
−1
a
First, pLB
] of which the
n (t) has the limited range [0, cn
maximum is much lower than our target minimum reliability
a
0.9 under 2 < α ≤ 6 for all n ≥ 1. Second, pUB
n (t) is
independent of n, and thus becomes less accurate for larger n.
For these reasons, we aim at further improving the bounds
in Lemma 1 such that the bounds behave similar to the original
pn (t) with respect to n and t under the entire range [0, 1] while
converging to the asymptotic curve of pn (t) for t → 0. To
this end, we first derive the asymptotic bounds as follows.
Lemma 2. (Asymptotic Bounds) At a static typical UE, nsuccessive SIR coverage pn (t) follows the asymptotic curves:
2n
t
α−2
2
−1 − α
b
pUB
t .
n (t) = cn
b
pLB
n (t) = 1 −
(5)
(6)
Proof: Applying Taylor expansion to (1) with the hypergeometric function definition, its first derivative satisfies:
limt→0 pn (t)0 = −(2nt)/(α − 2) and limt→∞ pn (t)0 =
2
t− α /cn . By the hypergeometric function definition, pn (0) = 1
and limt→∞ pn (t) = 0. The asymptotic curves thus become
UBb
b
pLB
respectively as t → 0 and ∞.
n (t) and pn
To prove the asymptotic curves become the bounds of
pn (t), we further identify the behavior of pn (t) by directly considering its first derivative given as: pn (t)0 =
−2pn (t) [1 − (1 + t)−n pn (t)] /(αt). This shows pn (t)0 < 0
for all t < ∞. Since pn (t) is continuous by the hypergeometric function definition, its functional and derivative
values at the extreme t’s indicate pn (t) is a function that is
(strictly) decresing as well as (strictly) convex for all t (strict
b
conditions hold for t < ∞). The asymptotic curve pLB
n (t)
LBb
shares pn (0) = pn (0) = 1 while having limited domain,
b
thus becoming the lower bound. The curve pUB
n (t) satisfies
UBb
limt→∞ pn (t) = limt→∞ pn (t) = 0 while having infinite
maximum range as t → 0, thus becoming the upper bound.
b
Note that pUB
n (t) in Lemma 2 generalizes its special case
that has been derived in [9], [10]. It is also noted
a
that the previous Lemma 1’s pLB
n (t) converges to the derived
UBb
asymptotic curve pn (t) as t → ∞.
b
Next, we combine Lemma 1 and pLB
n (t) in Lemma 2,
resulting in the following Theorem.
b
pUB
1 (t)
Theorem 1. (Coverage Bounds) At a static typical UE, the
n-successive SIR coverage pn (t) is bounded by:
−
2
c
α−2
pLB
n (t) = (1 + nt)
− α2
nα
UBc
pn (t) = 1 +
t
.
α−2
(7)
(8)
2
Proof: We choose the common term (1 + t)− α in Lemma 1
as the baseline, which is also a strictly decreasing and convex
function for t < ∞. We modify this, and aim at its convergence to the asymptotic curve 1 − 2nt/(α − 2) in Lemma 2 as
t → 0. For this purpose, we consider its generalized function
2B
(1 + At)− α for constants A, B > 0, and consider its Taylor
expanded form 1 − 2ABt/α for an infinitesimal t, which
should satisfy AB = nα/ (α − 2) for the desired convergence.
We select {A, B} = {n, α/ (α − 2)} and {nα/ (α − 2) , 1}
c
c
respectively yielding pLB
and pUB
n
n .
c
To verify the lower bound, showing pn (t) ≥ pLB
n (t) only
at a single point t suffices thanks to their strictly decreasing
and convex-shaped behaviors (see the proof of Lemma 2). In
this respect, consider a sufficiently large constant M > 0 such
b
that pUB
pn (M ). In this case,
n (M ) in Lemma 2 approximates
α
2
c
α ≥ c 2 −1 /n, which can
pn (M ) ≥ pLB
(M
)
becomes
M
n
n
always be satisfied by choosing a sufficiently large M .
Similarly, we prove the upper bound by showing
UBb
c
pUB
n (M ) ≥ pn (M ) ≈ p2n (M ). This is identical to proving
α
≥ 1, which is a stricly increasD(n, α) := cn α−2
nα
ing function of n ≥ 1 and a quasi-concave function of
α > 2 by the definition of cn . Therefore, it suffices to
2
verify D(1, α) = 1 − α2 α c1 ≥ 1 as α goes to either ∞
α→∞
or 2 as follows. Firstly, recall c1 = 1 discussed after
Lemma 1, thereby yielding limα→∞ D(1, α) = 1. Secondly,
(b)
(a)
limα→2 D(1, α) = limα→2 1 − α2 Γ 1 − α2 Γ 1 + α2 =
limα→2 Γ 2 − α2 Γ 1 + α2 = 1, where (a) and (b) follow
2
from limα→2 1 − α2 α / 1 − α2 = 1 and Γ(x+1) = xΓ(x).
The results indicate that the minimum of D(n, α) is unity, and
consequently D(n, α) ≥ 1 for all α > 2 and n ≥ 1.
UBc
c
have the full range [0, 1],
As Fig. 2 illustrates, pLB
n and pn
LBb
and are even tighter than pn within the range [0.9, 1].
Along the same lines, we additionally consider different
choices of {A, B} leading to the following lower bounds.
Corollary 1. At a static typical UE, pn (t)’s lower bounds are:
2n
− α−2
×
pLB
n (t) = (1 + t)
+
pLB
n (t)
= 1 + cn
α
2
(9)
− 2
t α.
(10)
α
Proof: Applying {A, B} = {1, nα/(α − 2)} and {cn 2 , 1} to
LB
LB
the proof of Theorem 1 results in pn × (t) and pn + (t).
LB
The first bound pn × (t) is tight under high reliabilty, i.e. low
LB
LB
t as Fig. 2 shows. In addition, it satisfies pn × (t) = p1 × (t)n .
This equality indicates that static (LHS) and i.i.d. (RHS) UE
locations during n receptions identically affect SIR almost
LB
surely. Knowing that pn × (t) ≤ pn (t), we conclude spatial
UE correlation during n receptions increases successive SIR
coverage, compared to the coverage under i.i.d. mobility. Such
a spatial correlation gain diminishes under high reliability.
LB
The second bound pn + (t) accuracy increases with t as
Fig. 2 shows, which may complement Theorem 1 for less
reliable applications. This is achieved by utilizing the opposite
b
asymptotic curve pUB
n (t) in Lemma 2 for high t accuracy,
b
instead of pLB
(t)
exploited
in Theorem 1.
n
IV. M AXIMUM AVERAGE R ATE B OUNDS UNDER
n-S UCCESSIVE SIR C OVERAGE C ONSTRAINTS
In this section we derive the closed-form Rn (η), a typical
UE’s maximum average rate ensuring pn (t) ≥ η. For brevity,
rate is measured in nats/sec (1 nat = 1/ log 2 bits).
Recall the optimization (2). We define t̂ as the t value when
the reliability constraint’s equality holds, i.e. pn (t̂) = η. We
consider t̂ is the optimal SIR
threshold maximizing average
rate, i.e. Rn (η) = pn t̂ log(1 + t̂). To show this, we
additionally define t̃ as the t maximizing Rn (η) without
the reliable constraint, i.e. t̃ := arg maxt≥0 pn (t) log(1 + t).
0.6
0.4
0.1
0.7
0.09
0.6
0.08
n=1
+
+
0.5
LBc , UBc Proposition 1
LB⇥
Corollary 2
LB+
LBb
Exact
Maximum Average Rate,R
Rnn(⌘)
( ) (nats/sec)
n=1
+
Maximum Average Rate,R
Rnn(⌘)
( ) (nats/sec)
0.7
+
0.5
0.07
0.06
0.4
LBc , UBc Proposition 1
LB⇥
Corollary 2
LB+
LBb
Exact
0.05
0.3
Baseline: n = 3, ↵ = 4
n = 3, ↵ = 4
0.3
0.04
0.2
0.1
0
0.3
↵ = 2.5
0.4
0.5
0.6
0.7
0.8
0.9
1
Reliability,
(a) for the reliability η within [0.3, 1].
0.03
0.2
Baseline: n = 3, ↵ = 4
n = 3, ↵ = 4
0.02
0.1
0.01
↵ = 2.5
00
0.3
0.9
0.91 0.4
0.92
0.5
0.93
0.6
0.94
0.95
0.70.96
0.8
0.97
0.98 0.9 0.99
11
Reliability,
(b) for the reliability η within [0.9, 1].
Fig. 3. The bounds of maximum average rate Rn (η) ensuring n-successive coveraege with probability η, under different n and the path loss exponent α > 2.
If t̂ ≤ t̃, then the quasi-concave objective function strictly
increases, having the maximum at t = t̂. This is true for
sufficiently high reliability, e.g. when η ≥ 0.29 for all n ≥ 1
and 2 < α ≤ 4, which is much less than our target η ≥ 0.9.
For this reason, Rn (η) derivation is identical to deriving t̂.
For URLLC, it can be analytically derived via Theorem 1. The
original Rn (η) is then bounded by their corresponding results
UBi
i
RLB
n (η) ≤ Rn (η) ≤ Rn (η) given as follows.
Proposition 1. (Rate Bounds) At a static typical UE, maximum
average rate Rn (η) satisfying pn (t) ≥ η is bounded by:
!
α
η −( 2 −1) − 1
LBc
Rn (η) = η log 1 +
(11)
n
α
2 η− 2 − 1
UBc
Rn (η) = η log 1 + 1 −
.
(12)
α
n
Proof: The optimal SIR threshold t̂ is derived by taking the
LBc
c
and pUB
inverse pn (t) into η, i.e. t̂ = p−1
n
n (η). Applying pn
in Theorem 1 completes the proof.
Exploiting Corollary 1, we also derive the following bounds.
Corollary 2. At a static typical UE, Rn (η)’s lower bounds
are:
1 α
− n ( 2 −1)
×
(η)
=
η
log
η
RLB
(13)
n
α
2
−α
+
RLB
η− 2 − 1 .
(14)
n (η) = η log 1 + cn
As reliability increases, Fig. 3 shows that the exact Rn (η)
curve from (1) asymptotically converges to RnLBc (η), RnUBc (η),
UB
and Rn × (η), as well as RnLBb (η) = α−2
2n (1 − η) similarly
b
derived by applying pLB
(t)
in
Lemma
2.
Especially, the conn
UB×
vergence to Rn (η) implies the spatial UE correlation impact
becomes negligible in URLLC as stated after Corollary 1.
LB
For moderate reliability 0.3 < η < 0.9, Rn + (η) becomes
b
accurate. Note that RLB
n (η) can also be an approximation if
n is large; otherwise, a huge gap exists as shown by the curve
in Fig. 3 when n = 1 (yellow circle).
Additionally, it is noted that the proposed closed-form
bounds are also applicable for conventional average rate
calculation under the full CSI assumption as in [3]–[5]. For
when n = 1, average rate
+ SIR)] =
Rexample,
R ∞ is E[log(1
∞
x
Pr
(log(1
+
SIR)
>
x)
dx
=
p
(e
−
1)
dx.
Apply1
0
0
UBa
c
ing the simplest bounds pLB
and
p
to
this
yields
the
1
1
average rate bounds (α − 2)/2 and α/2 respectively. Their
linear combination with the equal weight factor leads to the
rate approximation (α − 1)/2. For α = 4, this corresponds
to 1.5 nats/sec. The result is fit well with the exact calcu simplifying the√known expression
Rlation 1.49 nats/sec
√ x while
dx/
1
+
e
−
1
π/2 − arctan(1/ ex − 1) in [3].
x>0
V. C ONCLUSION
We proposed closed-form n-successive SIR coverage
bounds, and thereby derived the closed-form maximum average rate guaranteeing such a successive coverage requirement.
R EFERENCES
[1] Ericsson Business Review, Manufacturing Reengineered: Robots, 5G
and the Industrial IoT, no. 4, 2015.
[2] 5G-PPP, 5G and the Factories of the Future, White Paper, Oct. 2015.
[3] J. G. Andrews, F. Baccelli, and R. K. Ganti, “A Tractable Approach
to Coverage and Rate in Cellular Networks,” IEEE Trans. Commun.,
vol. 59, pp. 3122–3134, Nov. 2011.
[4] X. Zhang and M. Haenggi, “A Stochastic Geometry Analysis of InterCell Interference Coordination and Intra-Cell Diversity,” IEEE Trans.
Wireless Commun., vol. 13, pp. 6655–6669, Dec. 2014.
[5] J. Park, S.-L. Kim, and J. Zander, “Tractable Resource Management with
Uplink Decoupled Millimeter-Wave Overlay in Ultra-Dense Cellular
Networks,” IEEE Trans. Wireless Commun., vol. 15, pp. 4362–4379,
Jun. 2016.
[6] M. Haenggi, Stochastic Geometry for Wireless Networks. Cambridge
Univ. Press, 2013.
[7] J. Park, S. Jung, S.-L. Kim, M. Bennis, and M. Debbah, “User-Centric
Mobility Management in Ultra-Dense Cellular Networks under SpatioTemporal Dynamics,” Proc. IEEE GLOBECOM, Washington, DC, USA,
Dec. 2016.
[8] S. Krishnan and H. S. Dhillon, “Spatio-Temporal Interference Correlation and Joint Coverage in Cellular Networks,” [Online]. Available:
https://arxiv.org/abs/1606.05332.
[9] B. Blaszczyszyn, M. K. Karray, and H. P. Keeler, “Using Poisson
Processes to Model Lattice Cellular Networks,” Proc. IEEE INFOCOM,
Turin, Italy, Apr. 2013.
[10] H. S. Dhillon, R. K. Ganti, F. Baccelli, and J. G. Andrews, “Modeling
and Analysis of K-Tier Downlink Heterogeneous Cellular Networks,”
IEEE J. Sel. Areas Commun., vol. 30, pp. 550–560, Apr. 2012.
[11] J. Park, D. Kim, P. Popovski, and S.-L. Kim, “Revisiting Frequency
Reuse towards Supporting Ultra-Reliable Ubiquitous-Rate Communication,” Proc. IEEE WiOpt Wksp. SpaSWiN, Paris, France, May 2017.
[12] G. E. Andrews, R. Askey, and R. Roy, Special Functions. Cambridge
Univ. Press, 1999.
| 7cs.IT |
arXiv:1706.01554v2 [cs.CV] 27 Oct 2017
Best of Both Worlds: Transferring Knowledge from
Discriminative Learning to a Generative Visual
Dialog Model
Jiasen Lu1∗, Anitha Kannan2∗, Jianwei Yang1 , Devi Parikh3,1 , Dhruv Batra3,1
1
Georgia Institute of Technology, 2 Curai, 3 Facebook AI Research
{jiasenlu, jw2yang, parikh, dbatra}@gatech.edu
Abstract
We present a novel training framework for neural sequence models, particularly
for grounded dialog generation. The standard training paradigm for these models
is maximum likelihood estimation (MLE), or minimizing the cross-entropy of the
human responses. Across a variety of domains, a recurring problem with MLE
trained generative neural dialog models (G) is that they tend to produce ‘safe’
and generic responses (‘I don’t know’, ‘I can’t tell’). In contrast, discriminative
dialog models (D) that are trained to rank a list of candidate human responses
outperform their generative counterparts; in terms of automatic metrics, diversity,
and informativeness of the responses. However, D is not useful in practice since it
can not be deployed to have real conversations with users.
Our work aims to achieve the best of both worlds – the practical usefulness of
G and the strong performance of D – via knowledge transfer from D to G. Our
primary contribution is an end-to-end trainable generative visual dialog model,
where G receives gradients from D as a perceptual (not adversarial) loss of the
sequence sampled from G. We leverage the recently proposed Gumbel-Softmax
(GS) approximation to the discrete distribution – specifically, a RNN augmented
with a sequence of GS samplers, coupled with the straight-through gradient estimator to enable end-to-end differentiability. We also introduce a stronger encoder for
visual dialog, and employ a self-attention mechanism for answer encoding along
with a metric learning loss to aid D in better capturing semantic similarities in
answer responses. Overall, our proposed model outperforms state-of-the-art on the
VisDial dataset by a significant margin (2.67% on recall@10). The source code
can be downloaded from https://github.com/jiasenlu/visDial.pytorch
1
Introduction
One fundamental goal of artificial intelligence (AI) is the development of perceptually-grounded
dialog agents – specifically, agents that can perceive or understand their environment (through vision,
audio, or other sensors), and communicate their understanding with humans or other agents in natural
language. Over the last few years, neural sequence models (e.g. [47, 44, 46]) have emerged as the
dominant paradigm across a variety of setting and datasets – from text-only dialog [44, 40, 23, 3] to
more recently, visual dialog [7, 9, 8, 33, 45], where an agent must answer a sequence of questions
grounded in an image, requiring it to reason about both visual content and the dialog history.
The standard training paradigm for neural dialog models is maximum likelihood estimation (MLE)
or equivalently, minimizing the cross-entropy (under the model) of a ‘ground-truth’ human response.
Across a variety of domains, a recurring problem with MLE trained neural dialog models is that they
tend to produce ‘safe’ generic responses, such as ‘Not sure’ or ‘I don’t know’ in text-only dialog [23],
and ‘I can’t see’ or ‘I can’t tell’ in visual dialog [7, 8]. One reason for this emergent behavior is that
∗
Work was done while at Facebook AI Research.
the space of possible next utterances in a dialog is highly multi-modal (there are many possible paths
a dialog may take in the future). In the face of such highly multi-modal output distributions, models
‘game’ MLE by latching on to the head of the distribution or the frequent responses, which by nature
tend to be generic and widely applicable. Such safe generic responses break the flow of a dialog and
tend to disengage the human conversing with the agent, ultimately rendering the agent useless. It is
clear that novel training paradigms are needed; that is the focus of this paper.
One promising alternative to MLE training proposed by recent work [36, 27] is sequence-level
training of neural sequence models, specifically, using reinforcement learning to optimize taskspecific sequence metrics such as BLEU [34], ROUGE [24], CIDEr [48]. Unfortunately, in the case
of dialog, all existing automatic metrics correlate poorly with human judgment [26], which renders
this alternative infeasible for dialog models.
In this paper, inspired by the success of adversarial training [16], we propose to train a generative
visual dialog model (G) to produce sequences that score highly under a discriminative visual dialog
model (D). A discriminative dialog model receives as input a candidate list of possible responses and
learns to sort this list from the training dataset. The generative dialog model (G) aims to produce a
sequence that D will rank the highest in the list, as shown in Fig. 1.
Note that while our proposed approach is inspired by adversarial training, there are a number of
subtle but crucial differences over generative adversarial networks (GANs). Unlike traditional GANs,
one novelty in our setup is that our discriminator receives a list of candidate responses and explicitly
learns to reason about similarities and differences across candidates. In this process, D learns a
task-dependent perceptual similarity [12, 19, 15] and learns to recognize multiple correct responses
in the feature space. For example, as shown in Fig. 1 right, given the image, dialog history, and
question ‘Do you see any bird?’, besides the ground-truth answer ‘No, I do not’, D can also assign
high scores to other options that are valid responses to the question, including the one generated by
G: ‘Not that I can see’. The interaction between responses is captured via the similarity between the
learned embeddings. This similarity gives an additional signal that G can leverage in addition to the
MLE loss. In that sense, our proposed approach may be viewed as an instance of ‘knowledge transfer’
[17, 5] from D to G. We employ a metric-learning loss function and a self-attention answer encoding
mechanism for D that makes it particularly conducive to this knowledge transfer by encouraging
perceptually meaningful similarities to emerge. This is especially fruitful since prior work has
demonstrated that discriminative dialog models significantly outperform their generative counterparts,
but are not as useful since they necessarily need a list of candidate responses to rank, which is only
available in a dialog dataset, not in real conversations with a user. In that context, our work aims to
achieve the best of both worlds – the practical usefulness of G and the strong performance of D – via
this knowledge transfer.
Our primary technical contribution is an end-to-end trainable generative visual dialog model, where
the generator receives gradients from the discriminator loss of the sequence sampled from G. Note
that this is challenging because the output of G is a sequence of discrete symbols, which naïvely is not
amenable to gradient-based training. We propose to leverage the recently proposed Gumbel-Softmax
(GS) approximation to the discrete distribution [18, 30] – specifically, a Recurrent Neural Network
(RNN) augmented with a sequence of GS samplers, which when coupled with the straight-through
gradient estimator [2, 18] enables end-to-end differentiability.
Our results show that our ‘knowledge transfer’ approach is indeed successful. Specifically, our
discriminator-trained G outperforms the MLE-trained G by 1.7% on recall@5 on the VisDial dataset,
essentially improving over state-of-the-art [7] by 2.43% recall@5 and 2.67% recall@10. Moreover,
our generative model produces more diverse and informative responses (see Table 3).
As a side contribution specific to this application, we introduce a novel encoder for neural visual
dialog models, which maintains two separate memory banks – one for visual memory (where do
we look in the image?) and another for textual memory (what facts do we know from the dialog
history?), and outperforms the encoders used in prior work.
2
Related Work
GANs for sequence generation. Generative Adversarial Networks (GANs) [16] have shown to be
effective models for a wide range of applications involving continuous variables (e.g. images) c.f
[10, 35, 22, 55]. More recently, they have also been used for discrete output spaces such as language
generation – e.g. image captioning [6, 41], dialog generation [23], or text generation [53] – by either
viewing the generative model as a stochastic parametrized policy that is updated using REINFORCE
2
Generator
Do you see any birds?
Question Qt
A gray tiger cat sitting underneath
a metal bench.
Answer
Decoder
𝑒"
Is it in color? Yes it is.
Is it day time? Yes.
Is the tiger big? No, it’s a regular
cat.
HCIAE
Encoder
No bird
I do not see any birds
No
No , I do not
Answer
Encoder
𝑓(𝑎)
Nope
…
yes
Mangoes
White
I see small shops
t rounds of history
HQI
No bird
I do not see any
birds
Image I
Option answers (D)
Not that I can see
HCIAE
Encoder
Deep Metric Loss
Gumbel
Sampler
Discriminator
𝑎∗
Score
No
No , I do not
No ,
Nope
Not that I can see
yes
…
…
Mangoes
𝑒"
White
I see small shops
HQI
HQI
Figure 1: Model architecture of the proposed model. Given the image, history, and question, the discriminator
receives as additional input a candidate list of possible responses and learns to sort this list. The generator aims
to produce a sequence that discriminator will rank the highest in the list. The right most block is D’s score for
different candidate answers. Note that the multiple plausible responses all score high. Image from the COCO
dataset [25].
with the discriminator providing the reward [53, 6, 41, 23], or (closer to our approach) through
continuous relaxation of discrete variables through Gumbel-Softmax to enable backpropagating the
response from the discriminator [21, 41].
There are a few subtle but significant differences w.r.t. to our application, motivation, and approach.
In these prior works, both the discriminator and the generator are trained in tandem, and from scratch.
The goal of the discriminator in those settings has primarily been to discriminate ‘fake’ samples
(i.e. generator’s outputs) from ‘real’ samples (i.e. from training data). In contrast, we would like
to transfer knowledge from the discriminator to the generator. We start with pre-trained D and G
models suited for the task, and then transfer knowledge from D to G to further improve G, while
keeping D fixed. As we show in our experiments, this procedure results in G producing diverse
samples that are close in the embedding space to the ground truth, due to perceptual similarity learned
in D. One can also draw connections between our work and Energy Based GAN (EBGAN) [54] –
without the adversarial training aspect. The “energy” in our case is a deep metric-learning based
scoring mechanism, instantiated in the visual dialog application.
Modeling image and text attention. Models for tasks at the intersection of vision and language
– e.g., image captioning [11, 13, 20, 49], visual question answering [1, 14, 31, 37], visual dialog
[7, 9, 8, 45, 33] – typically involve attention mechanisms. For image captioning, this may be attending
to relevant regions in the image [49, 51, 28]. For VQA, this may be attending to relevant image
regions alone [4, 50, 52] or co-attending to image regions and question words/phrases [29].
In the context of visual dialog, [7] uses attention to identify utterances in the dialog history that may
be useful for answering the current question. However, when modeling the image, the entire image
embedding is used to obtain the answer. In contrast, our proposed encoder HCIAE (Section 4.1)
localizes the region in the image that can help reliably answer the question. In particular, in addition
to the history and the question guiding the image attention, our visual dialog encoder also reasons
about the history when identifying relevant regions of the image. This allows the model to implicitly
resolve co-references in the text and ground them back in the image.
3
Preliminaries: Visual Dialog
We begin by formally describing the visual dialog task setup as introduced by Das et al. [7]. The
machine learning task is as follows. A visual dialog model is given as input an image I, caption c
describing the image, a dialog history till round t − 1, H = ( c , (q1 , a1 ), . . . , (qt−1 , at−1 )), and
|{z} | {z }
|
{z
}
H0
H1
Ht−1
the followup question qt at round t. The visual dialog agent needs to return a valid response to the
question.
Given the problem setup, there are two broad classes of methods – generative and discriminative
models. Generative models for visual dialog are trained by maximizing the log-likelihood of the
ground truth answer sequence agt
t ∈ At given the encoded representation of the input (I, H, qt ).
3
On the other hand, discriminative models receive both an encoding of the input (I, H, qt ) and as
(1)
(100)
additional input a list of 100 candidate answers At = {at , . . . , at
}. These models effectively
learn to sort the list. Thus, by design, they cannot be used at test time without a list of candidates
available.
4
Approach: Backprop Through Discriminative Losses for Generative
Training
In this section, we describe our approach to transfer knowledge from a discriminative visual dialog
model (D) to generative visual dialog model (G). Fig. 1 (a) shows the overview of our approach.
Given the input image I, dialog history H, and question qt , the encoder converts the inputs into a
joint representation et . The generator G takes et as input, and produces a distribution over answer
sequences via a recurrent neural network (specifically an LSTM). At each word in the answer
sequence, we use a Gumbel-Softmax sampler S to sample the answer token from that distribution.
The discriminator D in it’s standard form takes et , ground-truth answer agt
t and N − 1 “negative”
gt
N −1
answers {a−
}
as
input,
and
learns
an
embedding
space
such
that
similarity(e
t , f (at )) >
t,i i=1
−
similarity(et , f (at,· )), where f (·) is the embedding function. When we enable the communication
between D and G, we feed the sampled answer ât into discriminator, and optimize the generator G
to produce samples that get higher scores in D’s metric space.
We now describe each component of our approach in detail.
4.1
History-Conditioned Image Attentive Encoder (HCIAE)
…
An important characteristic in dialogs is the use of co-reference to avoid repeating entities that can
be contextually resolved. In fact, in the VisDial dataset [7] nearly all (98%) dialogs involve at least
one pronoun. This means that for a model to correctly answer a question, it would require a reliable
mechanism for co-reference resolution.
A common approach is to use an encoder architecture with an attention mechanism that implicitly
performs co-reference resolution by identifying the portion of the dialog history that can help in
answering the current question [7, 38, 39, 32]. while using a holistic representation for the image.
Intuitively, one would also expect that the answer is also localized to regions in the image, and be
consistent with the attended history.
With this motivation, we propose a novel encoder architecture (called HCIAE) shown in Fig. 2. Our
encoder first uses the current question to attend to the exchanges in the history, and then use the
question and attended history to attend to the image, so as to obtain the final encoding.
Specifically, we use the spatial image features V ∈
Rd×k from a convolution layer of a CNN. qt is
encoded with an LSTM to get a vector mqt ∈
CNN
𝐼
Rd . Simultaneously, each previous round of history
(H0 , . . . , Ht−1 ) is encoded separately with another
LSTM
𝑄"
LSTM as Mth ∈ Rd×t . Conditioned on the question
embedding, the model attends to the history. The at𝑒"
tended representation of the history and the question
LSTM
𝐻$
embedding are concatenated, and used as input to
attend to the image:
LSTM
𝐻"%&
αht
= softmax(zth )
(2)
…
= waT tanh(Wh Mth + (Wq mqt )1T ) (1)
…
…
zth
Encoder
Figure 2: Structure of the proposed encoder.
where 1 ∈ R is a vector with all elements set to 1.
Wh , Wq ∈ Rt×d and wa ∈ Rk are parameters to be learned. α ∈ Rk is the attention weight over
history. The attended history feature m̂ht is a convex combination of columns of Mt , weighted
appropriately by the elements of αth . We further concatenate mqt and m̂ht as the query vector and get
the attended image feature v̂t in the similar manner. Subsequently, all three components are used to
obtain the final embedding et :
t
et = tanh(We [mqt , m̂ht , vˆt ])
where We ∈ Rd×3d is weight parameters and [·] is the concatenation operation.
4
(3)
4.2
Discriminator Loss
Discriminative visual dialog models produce a distribution over the candidate answer list At and
maximize the log-likelihood of the correct option agt
t . The loss function for D needs to be conducive
for knowledge transfer. In particular, it needs to encourage perceptually meaningful similarities.
Therefore, we use a metric-learning multi-class N-pair loss [43] defined as:
logistic loss
z
− N −1
LD = Ln−pair {et , agt
t , {at,i }i=1 }, f = log 1 +
N
X
}|
gt
−
>
exp e>
t f (at,i ) − et f (at )
|
{z
}
i=1
!{
(4)
score margin
where f is an attention based LSTM encoder for the answer. This attention can help the discriminator
better deal with paraphrases across answers. The attention weight is learnt through a 1-layer MLP
over LSTM output at each time step. The N-pair loss objective encourages learning a space in which
the ground truth answer is scored higher than other options, and at the same time, encourages options
similar to ground truth answers to score better than dissimilar ones. This means that, unlike the
multiclass logistic loss, the options that are correct but different from the correct option may not be
overly penalized, and thus can be useful in providing a reliable signal to the generator. See Fig. 1 for
an example. Follwing [43], we regularize the L2 norm of the embedding vectors to be small.
4.3
Discriminant Perceptual Loss and Knowledge Transfer from D to G
At a high-level, our approach for transferring knowledge from D to G is as follows: G repeatedly
queries D with answers ât that it generates for an input embedding et to get feedback and update
itself. In each such update, G’s goal is to update its parameters to try and have ât score higher than
the correct answer, agt
t , under D’s learned embedding and scoring function. Formally, the perceptual
loss that G aims to optimize is given by:
gt
>
>
LG = L1−pair {et , ât , agt
},
f
=
log
1
+
exp
e
f
(a
)
−
e
f
(â
)
(5)
t
t
t
t
t
where f is the embedding function learned by the discriminator as in (4). Intuitively, updating
generator parameters to minimize LG can be interpreted as learning to produce an answer sequence
ât that ‘fools’ the discriminator into believing that this answer should score higher than the human
response agt
t under the discriminator’s learned embedding f (·) and scoring function.
While it is straightforward to sample an answer ât from the generator and perform a forward pass
through the discriminator, naïvely, it is not possible to backpropagate the gradients to the generator
parameters since sampling discrete symbols results in zero gradients w.r.t. the generator parameters.
To overcome this, we leverage the recently introduced continuous relaxation of the categorical
distribution – the Gumbel-softmax distribution or the Concrete distribution [18, 30].
At an intuitive level, the Gumbel-Softmax (GS) approximation uses the so called ‘Gumbel-Max trick’
to reparametrize sampling from a categorical distribution and replaces argmax with softmax to obtain
a continuous relaxation of the discrete random variable. Formally, let x denote a K-ary categorical
K
random variable with parameters denoted by (p1 , . . . pK ), or x ∼ Cat(p). Let gi 1 denote K
−g
IID samples from the standard Gumbel distribution, gi ∼ F (g) = e−e . Now, a sample from the
Concrete distribution can be produced via the following transformation:
e(log pi +gi )/τ
yi = PK (log p +g )
j
j /τ
j=1 e
∀i ∈ {1, . . . , K}
(6)
where τ is a temperature parameter that control how close samples y from this Concrete distribution
approximate the one-hot encoding of the categorical variable x.
As illustrated in Fig. 1, we augment the LSTM in G with a sequence of GS samplers. Specifically,
at each position in the answer sequence, we use a GS sampler to sample an answer token from
that conditional distribution. When coupled with the straight-through gradient estimator [2, 18]
this enables end-to-end differentiability. Specifically, during the forward pass we discretize the GS
samples into discrete samples, and in the backward pass use the continuous relaxation to compute
gradients. In our experiments, we held the temperature parameter fixed at 0.5.
5
5
Experiments
Dataset and Setup. We evaluate our proposed approach on the VisDial dataset [7], which was
collected by Das et al. by pairing two subjects on Amazon Mechanical Turk to chat about an image.
One person was assigned the role of a ‘questioner’ and the other of ‘answerer’. One worker (the
questioner) sees only a single line of text describing an image (caption from COCO [25]); the image
remains hidden to the questioner. Their task is to ask questions about this hidden image to “imagine
the scene better”. The second worker (the answerer) sees the image and caption and answers the
questions. The two workers take turns asking and answering questions for 10 rounds. We perform
experiments on VisDial v0.9 (the latest available release) containing 83k dialogs on COCO-train and
40k on COCO-val images, for a total of 1.2M dialog question-answer pairs. We split the 83k into 82k
for train, 1k for val, and use the 40k as test, in a manner consistent with [7]. The caption is
considered to be the first round in the dialog history.
Evaluation Protocol. Following the evaluation protocol established in [7], we use a retrieval setting
to evaluate the responses at each round in the dialog. Specifically, every question in VisDial is
coupled with a list of 100 candidate answer options, which the models are asked to sort for evaluation
purposes. D uses its score to rank these answer options, and G uses the log-likelihood of these
options for ranking. Models are evaluated on standard retrieval metrics – (1) mean rank, (2) recall
@k, and (3) mean reciprocal rank (MRR) – of the human response in the returned sorted list.
Pre-processing. We truncate captions/questions/answers longer than 24/16/8 words respectively. We
then build a vocabulary of words that occur at least 5 times in train, resulting in 8964 words.
Training Details In our experiments, all 3 LSTMs are single layer with 512d hidden state. We use
VGG-19 [42] to get the representation of image. We first rescale the images to be 224 × 224 pixels,
and take the output of last pooling layer (512 × 7 × 7) as image feature. We use the Adam optimizer
with a base learning rate of 4e-4. We pre-train G using standard MLE for 20 epochs, and D with
supervised training based on Eq (4) for 30 epochs. Following [43], we regularize the L2 norm of the
embedding vectors to be small. Subsequently, we train G with LG + αLM LE , which is a combination
of discriminative perceptual loss and MLE loss. We set α to be 0.5. We found that including LM LE
(with teacher-forcing) is important for encouraging G to generate grammatically correct responses.
5.1
Results and Analysis
Baselines. We compare our proposed techniques to the current state-of-art generative and discriminative models developed in [7]. Specifically, [7] introduced 3 encoding architectures – Late Fusion (LF),
Hierarchical Recurrent Encoder (HRE), Memory Network (MN) – each trained with a generative
(-G) and discriminative (-D) decoder. We compare to all 6 models.
Our approaches. We present a few variants of our approach to systematically study the individual
contributions of our training procedure, novel encoder (HCIAE), self-attentive answer encoding
(ATT), and metric-loss (NP).
• HCIAE-G-MLE is a generative model with our proposed encoder trained under the MLE objective. Comparing this variant to the generative baselines from [7] establishes the improvement
due to our encoder (HCIAE).
• HCIAE-G-DIS is a generative model with our proposed encoder trained under the mixed MLE
and discriminator loss (knowledge transfer). This forms our best generative model. Comparing
this model to HCIAE-G-MLE establishes the improvement due to our discriminative training.
• HCIAE-D-MLE is a discriminative model with our proposed encoder, trained under the standard discriminative cross-entropy loss. The answer candidates are encoded using an LSTM
(no attention). Comparing this variant to the discriminative baselines from [7] establishes the
improvement due to our encoder (HCIAE) in the discriminative setting.
• HCIAE-D-NP is a discriminative model with our proposed encoder, trained under the n-pair
discriminative loss (as described in Section 4.2). The answer candidates are encoded using an
LSTM (no attention). Comparing this variant to HCIAE-D-MLE establishes the improvement
due to the n-pair loss.
• HCIAE-D-NP-ATT is a discriminative model with our proposed encoder, trained under the
n-pair discriminative loss (as described in Section 4.2), and using the self-attentive answer
encoding. Comparing this variant to HCIAE-D-NP establishes the improvement due to the
self-attention mechanism while encoding the answers.
6
Table 1: Results (generative) on VisDial dataset. “MRR” Table 2: Results (discriminative) on VisDial dataset.
is mean reciprocal rank and “Mean” is mean rank.
Model
MRR R@1 R@5 R@10 Mean
Model
MRR R@1 R@5 R@10 Mean
LF-D [7]
0.5807 43.82 74.68 84.07 5.78
0.5868 44.82 74.81 84.36 5.66
LF-G [7]
0.5199 41.83 61.78 67.59 17.07 HREA-D [7]
0.5965 45.55 76.22 85.37 5.46
HREA-G [7]
0.5242 42.28 62.33 68.17 16.79 MN-D [7]
MN-G [7]
0.5259 42.29 62.85 68.88 17.06
HCIAE-D-MLE
0.6140 47.73 77.50 86.35 5.15
0.6182 47.98 78.35 87.16 4.92
HCIAE-G-MLE 0.5386 44.06 63.55 69.24 16.01 HCIAE-D-NP
HCIAE-G-DIS 0.5467 44.35 65.28 71.55 14.23 HCIAE-D-NP-ATT 0.6222 48.48 78.75 87.59 4.81
Results. Tables 1, 2 present results for all our models and baselines in generative and discriminative
settings. The key observations are:
1. Main Results for HCIAE-G-DIS: Our final generative model with all ‘bells and whistles’,
HCIAE-G-DIS, uniformly performs the best under all the metrics, outperforming the
previous state-of-art model MN-G by 2.43% on R@5. This shows the importance of the
knowledge transfer from the discriminator and the benefit from our encoder architecture.
2. Knowledge transfer vs. encoder for G: To understand the relative importance of the proposed history conditioned image attentive encoder (HCIAE) and the knowledge transfer, we
compared the performance of HCIAE-G-DIS with HCIAE-G-MLE, which uses our proposed encoder but without any feedback from the discriminator. This comparison highlights
two points: first, HCIAE-G-MLE improves R@5 by 0.7% over the current state-of-art
method (MN-D) confirming the benefits of our encoder. Secondly, and importantly, its
performance is lower than HCIAE-G-DIS by 1.7% on R@5, confirming that the modifications to encoder alone will not be sufficient to gain improvements in answer generation;
knowledge transfer from D greatly improves G.
3. Metric loss vs. self-attentive answer encoding: In the purely discriminative setting, our
final discriminative model (HCIAE-D-NP-ATT) also beats the performance of the corresponding state-of-art models [7] by 2.53% on R@5. The n-pair loss used in the discriminator
is not only helpful for knowledge transfer but it also improves the performance of the
discriminator by 0.85% on R@5 (compare HCIAE-D-NP to HCIAE-D-MLE). The improvements obtained by using the answer attention mechanism leads to an additional, albeit
small, gains of 0.4% on R@5 to the discriminator performance (compare HCIAE-D-NP to
HCIAE-D-NP-ATT).
5.2
Does updating discriminator help?
Recall that our model training happens as follows: we independently train the generative model
HCIAE-G-MLE and the discriminative model HCIAE-D-NP-ATT. With HCIAE-G-MLE as the
initialization, the generative model is updated based on the feedback from HCIAE-D-NP-ATT and
this results in our final HCIAE-G-DIS.
We performed two further experiments to answer the following questions:
• What happens if we continue training HCIAE-D-NP-ATT in an adversarial setting? In particular, we continue training by maximizing the score of the ground truth answer agt
t and
minimizing the score of the generated answer ât , effectively setting up an adversarial training regime LD = −LG . The resulting discriminator HCIAE-GAN1 has significant drop in
performance, as can be seen in Table. 4 (32.97% R@5). This is perhaps expected because
HCIAE-GAN1 updates its parameters based on only two answers, the ground truth and the
generated sample (which is likely to be similar to ground truth). This wrecks the structure that
HCIAE-D-NP-ATT had previously learned by leveraging additional incorrect options.
• What happens if we continue structure-preserving training of HCIAE-D-NP-ATT? In addition
to providing HCIAE-D-NP-ATT samples from G as fake answers, we also include incorrect
options as negative answers so that the structure learned by the discriminator is preserved.
HCIAE-D-NP-ATT continues to train under loss LD . In this case (HCIAE-GAN2 in Table. 4),
we find that there is a small improvement in the performance of G. The additional computational
overhead to training the discriminator supersedes the performance improvement. Also note that
HCIAE-D-NP-ATT itself gets worse at the dialog task.
7
Table 3: Qualitative comparison. “Ours” are samples from G-DIS model with different gumbel noise z. Images
from the COCO dataset [25].
Q: Is it a home or restaurant?
A: I think restaurant.
G-MLE: I can’t tell.
Ours (z1): Hard to say.
Ours (z2): It looks like a restaurant.
Ours (z3): I can’t tell because it
is too close.
Q: Can you see his face?
A: I am not sure.
G-MLE: Yes.
Ours (z1): I can only see the
back of his body.
Ours (z2): No.
Ours (z3): No , he’s too far
away.
Q: How old does the man seem to
be?
A: 20’s.
G-MLE: Late teens.
Ours (z1): On his 20s.
Ours (z2): In his twenties.
Ours (z3): Mid 20’s.
Q: Can you see broccoli?
A: Yes, 3 larger pieces and some
small.
G-MLE: No, just the broccoli.
Ours (z1): I can see broccoli,
slightly butter.
Ours (z2): Yes, there is broccoli.
Ours (z3): Yes, broccoli is
green.
Table 4: Adversarial training results on VisDial dataset.
Discriminative
Generative
Model
MRR
R@1
R@5
R@10
Mean
MRR
R@1
R@5
R@10
Mean
HCIAE-D-NP-ATT
HCIAE-G-DIS
0.6222
-
48.48
-
78.75
-
87.59
-
4.81
-
0.5467
44.35
65.28
71.55
14.23
HCIAE-GAN1
HCIAE-GAN2
0.2177
0.6050
8.82
46.20
32.97
77.92
52.14
87.20
18.53
4.97
0.5298
0.5459
43.12
44.33
62.74
65.05
68.58
71.40
16.25
14.34
One might wonder, why not train a GAN for visual dialog? Formulating the task in a GAN setting
would involve G and D training in tandem with D providing feedback as to whether a response that
G generates is real or fake. We found this to be a particularly unstable setting, for two main reasons:
First, consider the case when the ground truth answer and the generated answers are the same. This
happens for answers that are typically short or ‘cryptic’ (e.g. ‘yes’). In this case, D can not train itself
or provide feedback, as the answer is labeled both positive and negative. Second, in cases where the
ground truth answer is descriptive but the generator provides a short answer, D can quickly become
powerful enough to discard generated samples as fake. In this case, D is not able to provide any
information to G to get better at the task. Our experience suggests that the discriminator, if one were
to consider a ‘GANs for visual dialog’ setting, can not merely be focused on differentiating fake
from real. It needs to be able to score similarity between the ground truth and other answers. Such a
scoring mechanism provides a more reliable feedback to G. In fact, as we show in the previous two
results, a pre-trained D that captures this structure is the key ingredient in sharing knowledge with G.
The adversarial training of D is not central.
5.3
Qualitative Comparison
In Table 3 we present a couple of qualitative examples that compares the responses generated by
G-MLE and G-DIS. G-MLE predominantly produces ‘safe’ and less informative answers, such as
‘Yes’ and or ‘I can’t tell’. In contrast, our proposed model G-DIS does so less frequently, and often
generates more diverse yet informative responses.
6
Conclusion
Generative models for (visual) dialog are typically trained with an MLE objective. As a result, they
tend to latch on to safe and generic responses. Discriminative (or retrieval) models on the other hand
have been shown to significantly outperform their generative counterparts. However, discriminative
models can not be deployed as dialog agents with a real user where canned candidate responses
are not available. In this work, we propose transferring knowledge from a powerful discriminative
visual dialog model to a generative model. We leverage the Gumbel-Softmax (GS) approximation to
the discrete distribution –specifically, a RNN augmented with a sequence of GS samplers, coupled
with a ST gradient estimator for end-to-end differentiability. We also propose a novel visual dialog
encoder that reasons about image-attention informed by the history of the dialog; and employ a
metric learning loss along with a self-attentive answer encoding to enable the discriminator to
learn meaningful structure in dialog responses. The result is a generative visual dialog model that
significantly outperforms state-of-the-art.
8
References
[1] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick,
and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE International Conference
on Computer Vision, pages 2425–2433, 2015.
[2] Yoshua Bengio, Nicholas Léonard, and Aaron C. Courville. Estimating or propagating gradients through
stochastic neurons for conditional computation. CoRR, abs/1308.3432, 2013.
[3] Antoine Bordes and Jason Weston.
arXiv:1605.07683, 2016.
Learning end-to-end goal-oriented dialog.
arXiv preprint
[4] Kan Chen, Jiang Wang, Liang-Chieh Chen, Haoyuan Gao, Wei Xu, and Ram Nevatia. Abc-cnn: An attention based convolutional neural network for visual question answering. arXiv preprint arXiv:1511.05960,
2015.
[5] Tianqi Chen, Ian Goodfellow, and Jonathon Shlens. Net2net: Accelerating learning via knowledge transfer.
arXiv preprint arXiv:1511.05641, 2015.
[6] Bo Dai, Dahua Lin, Raquel Urtasun, and Sanja Fidler. Towards diverse and natural image descriptions via
a conditional gan. arXiv preprint arXiv:1703.06029, 2017.
[7] Abhishek Das, Satwik Kottur, Khushi Gupta, Avi Singh, Deshraj Yadav, José M.F. Moura, Devi Parikh,
and Dhruv Batra. Visual Dialog. In Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition, 2017.
[8] Abhishek Das, Satwik Kottur, José MF Moura, Stefan Lee, and Dhruv Batra. Learning cooperative visual
dialog agents with deep reinforcement learning. arXiv preprint arXiv:1703.06585, 2017.
[9] Harm de Vries, Florian Strub, Sarath Chandar, Olivier Pietquin, Hugo Larochelle, and Aaron Courville.
Guesswhat?! visual object discovery through multi-modal dialogue. arXiv preprint arXiv:1611.08481,
2016.
[10] Emily L. Denton, Soumith Chintala, Arthur Szlam, and Robert Fergus. Deep generative image models
using a laplacian pyramid of adversarial networks. Neural Information Processing Systems, 2015.
[11] Jeff Donahue, Lisa Anne Hendricks, Sergio Guadarrama, Marcus Rohrbach, Subhashini Venugopalan,
Kate Saenko, and Trevor Darrell. Long-term Recurrent Convolutional Networks for Visual Recognition
and Description. In CVPR, 2015.
[12] Alexey Dosovitskiy and Thomas Brox. Generating images with perceptual similarity metrics based on
deep networks. In Advances in Neural Information Processing Systems, pages 658–666, 2016.
[13] Hao Fang, Saurabh Gupta, Forrest N. Iandola, Rupesh Kumar Srivastava, Li Deng, Piotr Dollár, Jianfeng
Gao, Xiaodong He, Margaret Mitchell, John C. Platt, C. Lawrence Zitnick, and Geoffrey Zweig. From
Captions to Visual Concepts and Back. In CVPR, 2015.
[14] Haoyuan Gao, Junhua Mao, Jie Zhou, Zhiheng Huang, Lei Wang, and Wei Xu. Are you talking to a
machine? dataset and methods for multilingual image question answering. In NIPS, 2015.
[15] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. A neural algorithm of artistic style. arXiv preprint
arXiv:1508.06576, 2015.
[16] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron
Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in neural information processing
systems, pages 2672–2680, 2014.
[17] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv
preprint arXiv:1503.02531, 2015.
[18] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv
preprint arXiv:1611.01144, 2016.
[19] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and
super-resolution. In European Conference on Computer Vision, pages 694–711. Springer, 2016.
[20] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In
CVPR, 2015.
[21] Matt J. Kusner and José Miguel Hernández-Lobato. Gans for sequences of discrete elements with the
gumbel-softmax distribution. CoRR, abs/1611.04051, 2016.
[22] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew P. Aitken, Alykhan Tejani, Johannes
Totz, Zehan Wang, and Wenzhe Shi. Photo-realistic single image super-resolution using a generative
adversarial network. CoRR, abs/1609.04802, 2016.
[23] Jiwei Li, Will Monroe, Tianlin Shi, Alan Ritter, and Dan Jurafsky. Adversarial learning for neural dialogue
generation. arXiv preprint arXiv:1701.06547, 2017.
[24] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In ACL 2004 Workshop, 2004.
9
[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár,
and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on
Computer Vision, pages 740–755. Springer, 2014.
[26] Chia-Wei Liu, Ryan Lowe, Iulian V Serban, Michael Noseworthy, Laurent Charlin, and Joelle Pineau.
How not to evaluate your dialogue system: An empirical study of unsupervised evaluation metrics for
dialogue response generation. arXiv preprint arXiv:1603.08023, 2016.
[27] Siqi Liu, Zhenhai Zhu, Ning Ye, Sergio Guadarrama, and Kevin Murphy. Optimization of image description
metrics using policy gradient methods. arXiv preprint arXiv:1612.00370, 2016.
[28] Jiasen Lu, Caiming Xiong, Devi Parikh, and Richard Socher. Knowing when to look: Adaptive attention
via A visual sentinel for image captioning. CoRR, abs/1612.01887, 2016.
[29] Jiasen Lu, Jianwei Yang, Dhruv Batra, and Devi Parikh. Hierarchical question-image co-attention for
visual question answering. In Advances In Neural Information Processing Systems, pages 289–297, 2016.
[30] Chris J Maddison, Andriy Mnih, and Yee Whye Teh. The concrete distribution: A continuous relaxation of
discrete random variables. arXiv preprint arXiv:1611.00712, 2016.
[31] Mateusz Malinowski, Marcus Rohrbach, and Mario Fritz. Ask your neurons: A neural-based approach to
answering questions about images. In ICCV, 2015.
[32] Hongyuan Mei, Mohit Bansal, and Matthew R Walter. Coherent dialogue with attention-based language
models. arXiv preprint arXiv:1611.06997, 2016.
[33] Nasrin Mostafazadeh, Chris Brockett, Bill Dolan, Michel Galley, Jianfeng Gao, Georgios P Spithourakis,
and Lucy Vanderwende. Image-grounded conversations: Multimodal context for natural question and
response generation. arXiv preprint arXiv:1701.08251, 2017.
[34] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation
of machine translation. In ACL, 2002.
[35] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep
convolutional generative adversarial networks. CoRR, abs/1511.06434, 2015.
[36] Marc’Aurelio Ranzato, Sumit Chopra, Michael Auli, and Wojciech Zaremba. Sequence level training with
recurrent neural networks. arXiv preprint arXiv:1511.06732, 2015.
[37] Mengye Ren, Ryan Kiros, and Richard Zemel. Exploring models and data for image question answering.
In NIPS, 2015.
[38] Iulian V Serban, Alessandro Sordoni, Yoshua Bengio, Aaron Courville, and Joelle Pineau. Building end-to-end dialogue systems using generative hierarchical neural network models. arXiv preprint
arXiv:1507.04808, 2015.
[39] Iulian Vlad Serban, Alessandro Sordoni, Ryan Lowe, Laurent Charlin, Joelle Pineau, Aaron Courville,
and Yoshua Bengio. A hierarchical latent variable encoder-decoder model for generating dialogues. arXiv
preprint arXiv:1605.06069, 2016.
[40] Iulian Vlad Serban, Alessandro Sordoni, Ryan Lowe, Laurent Charlin, Joelle Pineau, Aaron Courville,
and Yoshua Bengio. A hierarchical latent variable encoder-decoder model for generating dialogues. In
Thirty-First AAAI Conference on Artificial Intelligence, 2017.
[41] Rakshith Shetty, Marcus Rohrbach, Lisa Anne Hendricks, Mario Fritz, and Bernt Schiele. Speaking the
same language: Matching machine to human captions by adversarial training. CoRR, abs/1703.10476,
2017.
[42] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014.
[43] Kihyuk Sohn. Improved deep metric learning with multi-class n-pair loss objective. In Advances in Neural
Information Processing Systems, pages 1849–1857, 2016.
[44] Alessandro Sordoni, Michel Galley, Michael Auli, Chris Brockett, Yangfeng Ji, Margaret Mitchell, JianYun Nie, Jianfeng Gao, and Bill Dolan. A neural network approach to context-sensitive generation of
conversational responses. arXiv preprint arXiv:1506.06714, 2015.
[45] Florian Strub, Harm de Vries, Jeremie Mary, Bilal Piot, Aaron Courville, and Olivier Pietquin. End-to-end
optimization of goal-driven and visually grounded dialogue systems. arXiv preprint arXiv:1703.05423,
2017.
[46] Sainbayar Sukhbaatar, Jason Weston, Rob Fergus, et al. End-to-end memory networks. In Advances in
neural information processing systems, pages 2440–2448, 2015.
[47] Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural networks. In
Advances in neural information processing systems, pages 3104–3112, 2014.
[48] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description
evaluation. In CVPR, 2015.
10
[49] Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. Show and tell: A neural image
caption generator. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,
pages 3156–3164, 2015.
[50] Huijuan Xu and Kate Saenko. Ask, attend and answer: Exploring question-guided spatial attention for
visual question answering. In European Conference on Computer Vision, pages 451–466. Springer, 2016.
[51] Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron C. Courville, Ruslan Salakhutdinov, Richard S.
Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention.
CoRR, abs/1502.03044, 2015.
[52] Zichao Yang, Xiaodong He, Jianfeng Gao, Li Deng, and Alex Smola. Stacked attention networks for image
question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,
pages 21–29, 2016.
[53] Lantao Yu, Weinan Zhang, Jun Wang, and Yong Yu. Seqgan: Sequence generative adversarial nets with
policy gradient. AAAI Conference on Artificial Intelligence, 2017.
[54] Junbo Jake Zhao, Michaël Mathieu, and Yann LeCun. Energy-based generative adversarial network. CoRR,
abs/1609.03126, 2016.
[55] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using
cycle-consistent adversarial networks. arXiv preprint arXiv:1703.10593, 2017.
11
| 2cs.AI |
CITATION: Kruse, J., Sosa, R. & Connor, A.M. (2016) Procedural urban environments for FPS games. Paper presented at Interactive
Entertainment 2016, 2-5 February, Canberra, Australia.
Procedural Urban Environments for FPS Games
Jan Kruse
Ricardo Sosa
Andy M. Connor
Auckland University of Technology
Private Bag 92006, Wellesley Street
Auckland 1142, New Zealand
+64 9 921 9999
Auckland University of Technology
Private Bag 92006, Wellesley Street
Auckland 1142, New Zealand
+64 9 921 9999
Auckland University of Technology
Private Bag 92006, Wellesley Street
Auckland 1142, New Zealand
+64 9 921 9999
jan.kruse@aut.ac.nz
ricardo.sosa@aut.ac.nz
andrew.connor@aut.ac.nz
This paper presents a novel approach to procedural generation of
urban maps for First Person Shooter (FPS) games. A multi-agent
evolutionary system is employed to place streets, buildings and
other items inside the Unity3D game engine, resulting in playable
video game levels. A computational agent is trained using machine
learning techniques to capture the intent of the game designer as
part of the multi-agent system, and to enable a semi-automated
aesthetic selection for the underlying genetic algorithm.
generation are designed to replace the human game designer. Such
approaches offer both benefits and disadvantages typical of
applying Artificial Intelligence (AI) approaches to cognitive task.
AI approaches can work much faster than a human level designer,
and are in some cases able to explore the design space automatically
to find levels with desirable qualities. But they aren't able to capture
the human creativity that produces the most interesting level
designs, and they are usually very specific to their particular
domain [6].
CCS Concepts
2.1 Evolving FPS Maps
ABSTRACT
• Software and its engineering➝Software organization and
properties➝ Contextual software domains➝Virtual worlds
software➝Interactive games
• Applied computing➝Computers in other domains➝ Personal
computers and PC applications➝ Computer games
Keywords
Procedural Environment; Computer Games; First Person Shooter,
Urban Environment; Genetic Algorithm; Agents; Artificial
Intelligence; Unity.
1. INTRODUCTION
First Person Shooter (FPS) games have been one of the most
successful and fastest growing genres in the computer games
industry over the past decade [1]. Titles such as Battlefield,
Insurgency, Halo, Call of Duty and many others sell millions of
copies every year. Creating engaging playable maps for FPS games
is a challenging and laborious task [2] which is usually achieved
manually which ensures that the level is both playable and matches
the intent of the designer. Urban environments are particularly
challenging as they are characterized by high complexity. The large
number of individual items such as houses, streets, cars or similar
obstacles and smaller props pose huge challenges for game level
designers. This paper outlines a novel semi-automated approach to
urban map design, which draws on Interactive Evolutionary
Computation and Autonomous Agents to produce fully playable
FPS game levels. The approach has the potential to significantly
reduce the degree of manual labour, making the design process
more efficient.
2. BACKGROUND AND RELATED WORK
Procedural level creation is a popular application for creative
computational systems [3] and has been used since the 1980s to
produce generative content for video games [4]. It gained new
momentum in recent years due to the availability of computational
resources in form of faster CPUs and general-purpose computing
on GPUs [5]. However, most approaches to procedural content
In this paper procedural or generative design denotes the automatic
or semi-automatic generation of game levels, including the terrain,
position of buildings, streets and other items. Instead of manually
placing game assets into the map, significant parts of this task are
automated. While there are some similarities to work conducted for
instance by Cardamone et al. [7] in context of FPS content
generation, and also by Cook and Colton [8] for general procedural
3D game design, this approach differs in that it heavily considers
the (human) game designers intent and feeds this deliberate choice
back into the automated selection process of the evolutionary
algorithm. Our approach is not about getting computer to produce
content or whole game procedurally, but more an attempt to support
the game designer in an otherwise very costly and time consuming
process. Currently, as a result of high design and development
costs, FPS games offer only a few hours of single-player gameplay
and oftentimes just a small number of multiplayer maps [7]. This
paper offers a semi-automated approach that should allow level
designers to produce a higher number of maps at significantly
reduced costs which has the potential to boost the popularity of a
new title as a means to commercial success.
Another important benefit of the proposed approach is the
reduction of game testing costs by use of automated content
evaluators in addition to pure content generators. The algorithm
allows simple integration of additional agents, which provides
flexible approaches to design and evaluation. Testing levels to fix
issues around playability, technical flaws, enjoyableness and fair
balance in case of multiplayer team-based gameplay is currently
another significant issue for commercial game development. Some
companies tend to release maps as open betas to get the gamer
community to evaluate the maps and identify technical and ludic
issues before the official release of the game or game extension.
Other companies rely on “free to play” models to test the base
content of the game before selling additional items commercially.
Examples include Battlefield Community Testing Environment
(free beta play), Team Fortress 2 (free to play) and the recently
announced Rainbow Six Siege (‘closed’ beta with access for anyone
with an NVidia graphics card). Eventually, this work will also lead
to reduced cost of map testing by use of additional evaluation
agents, resulting in cost effective generation and evaluation of
game content.
Evolving FPS game levels, also often called FPS maps, is not an
entirely novel idea. Previous work includes Cardamone et al. [7]
who also based the map generation on evolutionary algorithms,
with the main difference that their work used a reasonably simple
mathematical fitness function based on the average fighting time of
the player plus the free space on the map. The fitness function is
not only purely theory-driven rather than modelled on data-driven
player capabilities and preferences, but it also reflects a purely
player-oriented approach, which excludes the game designer’s
intent. Here the computational agent is the single designing entity.
Our approach differs from this in that it considers the map designer
and their preference. We use a design-driven instead of user-driven
approach to map generation. Potential expansions of this idea is
outlined below under Human-based Genetic Algorithms.
Another recent example for procedurally generated FPS maps is
work conducted by Lanzi et al. [9] which in turn utilizes Searchbased Procedural Content Generation [10]. Their work aims to
evolve game levels that provide basic match balancing properties
based on player skill and strategies. The content generator uses
genetic algorithms and is not dissimilar to the level generation
outlined above [7]. The novelty of our approach and main
difference to their work lies in the addition of a multi-agent system
and the ability to capture a human game designer’s preferences.
2.2 Human-Based Genetic Algorithms
Human-based Genetic Algorithms (HBGA) are a variation of
Interactive Genetic Algorithms [11]. The human user is replaced by
a multi-agent system consisting of human and computational agents
[12]. A generic example of a hybrid human and computational
HBGA is shown in Error! Reference source not found., where
the human works in combination with an agent to perform selection
from the population. An additional agent undertakes the
recombination processes.
intent. An agent based on this classifier provides additional
(computational) selections, which reduces the selection runs for the
human in the HBGA. This approach has shown modest, but
promising results in previous applications [13].
While it could be argued that calling an interactive genetic
algorithm with additional agents a Human-based Genetic
Algorithm is merely an academic exercise, the idea of a multi-agent
system that feeds into a genetic algorithm goes beyond that. It poses
a paradigm shift, where selection and recombination are separated
and performed by different entities. This allows subsequently to
add a variety of agents to the algorithm as we will show in the
following sections of this paper. It is envisaged that such agents
could focus on different aspects of the design, so potentially could
include a diversity agent that drives the design towards novel
layouts. This inclusion of multiple agents transforms canonical
Interactive Evolutionary Computation as introduced by Takagi and
Iba [14] from a purely generative algorithm into a flexible system
for generation and evaluation. Therefore, using the concept of
Human-based Genetic Algorithm proves to be a successful
framework for procedural content for video games.
But the HBGA opens up additional possibilities, for instance
running design intent and user preference concurrently as the
driving forces behind map generation. The HBGA allows both
design- and user-driven paradigms to be implemented by choosing
selection agents accordingly. The multi-agent system could be
made of both (human) designer agent for selection as well as
(human) player agent for selection, both in additional to
computational agents. Some games such as Battlefield or
Insurgency allow users at the start of the round to select the map
that is going to be played next. With HBGA map generation in
mind, this pre-game ‘vote’ could be used to feed back into the
HBGA to capture the player’s intent and therefore increase the
chance of more enjoyable gameplay for the target group.
3. URBAN FPS GENERATOR
We present an Urban FPS Generator, which is a multi-agent
evolutionary system that integrates human and computational
agents into the selection process of a genetic algorithm. It is capable
of addressing different aspects of Computational Creativity, and in
its current version targets placement of game assets to form
playable FPS maps. Figure 2 shows a typical game level in Unity3D
developed with the approach.
Figure 1. Human-Based Genetic Algorithm.
HBGA have been used to successfully counteract user fatigue, one
of the implicit problems of Interactive Evolutionary Systems when
applied to creative design tasks [13]. A machine learning classifier
(J48) is trained using the designer’s input over a number of GA
generations to gauge the aesthetic preference and capture the design
Figure 2. Game Level in Unity3D.
Urban FPS Generator extends the generic HBGA framework
shown in Figure 1. The specific agents involved are shown in
Figure 3.
Initial
Random
Population
Design Intent
Agent
Additional
Agents
Play Testers
Interactive
GA
Playable
Game Level
Initial
Level(s)
Design
Player Model
Agent
Game Designer
Figure 3. Urban FPS Generator System Overview.
The usage of the system would typically start with the manual
design of a number of different game levels that are used in the
development of the player model. These levels are played
extensively by a group of suitable play testers and this gameplay is
used to generate the player model. This development of the player
model is performed “offline” from the game design process. When
the game designer wishes to start the design of a new level, the
interactive Genetic Algorithm generates a new initial random
population. The creation of the next generation is enacted through
a combination of the designer’s choices and the input of the player
model agent which evaluates the playability of each level. As the
designer evolves the new level the design intent agent observes the
designer’s choices and learns about the goal of the designer,
whether such a goal is even implicit. The design intent agent can
accelerate the evolution of new levels by acting on the behalf of the
designer and therefore reducing the fatigue associated with typical
interactive evolutionary computation approaches. Whilst not yet
developed, the agent architecture allows for additional agents to
also influence the evolution of the new level. For example, a
novelty agent may inject candidate solutions that are distinct from
those generated by the designer and/or the design intent agent as a
means to prevent premature convergence. The Human-Based
Genetic Algorithm is therefore the combination of these agents
with the interactive Genetic Algorithm and the designer.
Previous work on procedural generation of cityscapes [13] was
implemented in Java, assisted by the core libraries of the
Processing programming language as imported jar files to visualize
the underlying map layouts. In contrast, Urban FPS Generator is
written as an extension to Unity3D, a commercially available game
development environment using C# (and optionally JavaScript).
Unity3D offers cross-platform development for browser games,
native PC platforms, game consoles and handheld devices in a
highly versatile and extensible way. It is also a very efficient video
game engine with support for the latest graphics options (DirectX
11 and OpenGL 4), which allows games to run at a playable
framerate with minimal lag while the graphics are attractive and
aesthetically pleasing. These attributes make it a good choice for
FPS game research, and given that one of the aims of this project is
a fully functional, distributable and simple to install application.
This application is intended to serve as a platform for future
research into player behaviour and agent systems, therefore
Unity3D seems to be a better choice, superior to fully custom
developments in Java.
The systems makes use of prefabs, which are prebuilt game assets
inside the Unity3D project. Prefabs can be used as instances inside
a game level (or scene). The current version of Urban FPS
Generator uses 12 building prefabs, a range of street sections, a
number of simple props such as chairs, tables, leaves, containers
and plants, as well as 5 weapons from the UFPS package from the
Unity3D asset store. This has allows the system to be suitably
robust and have the ability to produce fully playable levels. One
such level is shown in Figure 4.
Figure 4. Fully Playable FPS Map.
We are currently working on extending this basic library of prefabs
by adding additional items, but also even smaller building blocks.
Our main intent is the generation of maps, but it is also possible to
evolve the building blocks for these maps from even smaller units
such as parts of buildings (windows, framing, doors, and wall
pieces). Currently, Urban FPS Generator is not capable of creating
interiors of multi-level buildings or complex maze-like structures
such as interior office spaces or cluttered warehouses. Using partial
prefabs to evolve buildings and other structures allows us to address
these issues.
Further, we are building terrain support into the evolutionary
system, which enables rugged environments such as mountain
ranges and elevated city layouts to be part of the map generation.
While this means an increase in parameters for the HBGA, the user
is factually only interacting with the same interface and requires no
additional skills or knowledge. At the same time, with an increase
of the number of parameters, the need for additional iterations in
the HBGA will rise and therefore compensation for user fatigue
through learning computational agents is of even higher
significance for the overall system performance.
3.1 User Interaction
The system presents a selection of 9 maps to the designer using the
inbuilt UI system in Unity3D. The maps are rendered top-down
views of three-dimensional game levels. The user selects the two
most preferred maps, which are fed back into the evolutionary
breeding pool. In parallel, the agent receives all 9 maps classified
by the user selection as training data. The process is repeated for a
maximum number of times or when terminated by the designer. The
final map is used as the active game level for the FPS round.
3.2 Map Encoding
Unity3D supports 3D game levels of nearly any complexity, only
limited by computer performance and creation time. We take a
much simpler approach and limit the maps to currently only a single
level. Urban FPS Generator is based on a two-dimensional matrix
(512x512 units) which roughly follows Unity3D’s terrain system
with a few exceptions. We use zero elevation for the urban
environments. Our street system also uses large sections of 25 units,
which limits the number of streets and intersections in each
dimension. This can easily be extended by increasing the base
terrain size, but in order to keep computation times for the map
generation to a minimum, we have opted for these rather small
game levels. From experience with existing maps in AAA titles,
this is not a disadvantage. Battlefield has an additional content
package that explicitly focusses on close quarter, infantry only
warfare. Given that Urban FPS Generator has no inbuilt
functionality for vehicles, we believe this choice is sufficient and
has no impact on playability or enjoyableness. This will be verified
in future work.
The chromosomes of the maps are represented as a strongly typed
list where all parameters such as building height, type of content
(building, street, free space) and position of props (leaves,
containers, obstacles, fences, plants) are held as numerical values.
The system is able to produce environments where props are well
placed in context, as shown in Figure 5.
Figure 5. Placement of Props.
We use real coded values to counteract some of the known issues
when dealing with variables in continuous domains [15].
Accordingly, crossover and mutation are performed on real values
instead of binary strings.
3.3 Agents
The computational agents in our HBGA are currently based on J48
(open source C4.5 implementation) classifiers that feed into simple
decision trees. Previous work [13] successfully utilised J48 as a
classifier of designer intent, though we acknowledge that this might
not be the ideal classifier for this type of machine learning
application due to the inherent system noise based on the low
number of training samples (user inputs) in relation to the number
of attributes being classified. J48 has been acknowledged as
problematic in this regard in other applications [16]. As further
attributes are introduced to the classifier as Urban FPS Generator
is extended is expected that J48 will reach a threshold level of
performance. We are looking at alternatives in a different context
and will replace this algorithm within Urban FPS Generator at
some point in time.
The agent captures user input through the selection process as
indicated above. A set of abstract features such as occupied/free
space, density of street system and number of (obstacle) props used
are used as values for the training set. The agent uses all 9 maps of
each iteration, with the preferred selection classified as positive,
and the remaining 7 maps as rejected training samples. Over the
course of a defined number of iterations (we use either 10 or 20 for
testing purposes), the agent receives a small training set of selected
and rejected candidates. We are aware of the limitations due to the
relatively small number of samples in the training set, given that
J48 classifiers are often trained using hundreds or thousands of
samples. But even under these circumstances the contribution of
the agent has shown improvements of the overall system
performance by reducing the required number of iterations by the
designer [13].
Additional agents can be added into the multi-agents system to
support the game design process even further. For example
evaluation agents are commonly modelled on behaviours similar to
non-player characters (NPCs), which make use of simple artificial
intelligence (AI) techniques such as decision trees and pathfinding.
Keeping the AI simple allows for real time performance, an
important property of NPCs. The capability of such evaluation
agents is in turn relatively simple as well. We are currently looking
at using ray casting to create advanced evaluation agents. Ray
casting essentially allows for basic ‘vision’ so that agents can detect
potential areas that serve as cover for players by identifying
obstacles (obstructions to the cast rays). This process also enables
agents to find areas that are lacking cover and could potentially be
exploited by players and lead to so called choke points during
multi-player matches. Further, ray casting works in three
dimensional space, whereas path finding is limited to simple two
dimensional maps. Ray casting is conveniently part of the Unity3D
physics engine. It is therefore efficient and simple to use. Overall
ray casting is a mechanism that provides significant improvements
over simple path finding and advances common AI techniques used
for NPCs. Figure 6 demonstrates the use of these techniques.
[3]
[4]
[5]
[6]
[7]
Figure 6. NPC Using Advanced AI.
4. FUTURE WORK
The work shown here serves as the foundation for future research
into player behavior. We intent to use Cognitive Modelling [17]
based on data gathered through in-game metric as well as player
profile, demographic and other static factors. For the former we will
utilize Urban FPS Generator as a testbed to extract data in a
controlled way. Many commercial games, while providing the
technical facilities to do so, allow no in-game measurements
legally. Urban FPS Generator avoids these implications. The
HBGA offers sufficient flexibility to customize the type and
number of agents in a very economical way. It therefore provides a
viable platform for future research into different generation and
evaluation agents. That includes alternative machine learning
algorithms to improve response to human fatigue during interactive
iterations. But it also allows for more advanced evaluation agents
testing game levels before they reach the breeding pool of the
genetic algorithm. Finally, there is the possibility to model adaptive
in-game AI agents that are used as NPCs to continuously adjust to
human players, for instance as training bots or competitive partners.
The main benefit of our work is the flexibility of the HBGA that
allows fast development of different computational agents as part
of the overall multi-agent system.
5. CONCLUSIONS
This paper introduces Urban FPS Generator a design tool for FPS
map creation within the Unity3D game development environment
using HBGA. We demonstrate its potential to create fully playable
game levels with the option to run multi-player matches without
any additional involvement of game designers or programmers.
The system uses computational agents to compensate for
limitations normally found in interactive evolutionary systems.
Finally, Urban FPS Generator provides a platform for future
research into player behaviour and team play including the option
to derive in-game metrics during multi-player matches.
6. REFERENCES
[1]
[2]
F. Cifaldi, “Analysts: FPS ‘Most Attractive’ Genre for
Publishers,” Gamasutra, 21-Feb-2006. [Online]. Available:
http://www.gamasutra.com/phpbin/news_index.php?story=8241. [Accessed: 08-Sep-2015].
K. Hullett and J. Whitehead, “Design patterns in FPS
levels,” in proceedings of the Fifth International Conference
on the Foundations of Digital Games, 2010, pp. 78–85.
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
A. Liapis, G. N. Yannakakis, and J. Togelius,
“Computational game creativity,” in Proceedings of the
Fifth International Conference on Computational
Creativity, 2014, pp. 285–292.
M. Toy, G. Wichman, K. Arnold, and J. Lane, “Rogue,”
Comput. Sci. Res. Group UC Berkeley, 1980.
B. Mark, T. Berechet, T. Mahlmann, and J. Togelius,
“Procedural Generation of 3D Caves for Games on the
GPU,” in Proceedings of the 10th International Conference
on the Foundations of Digital Games, Pacific Grove, CA,
2015.
P. Mawhorter and M. Mateas, “Procedural level generation
using occupancy-regulated extension,” in 2010 IEEE
Symposium on Computational Intelligence and Games
(CIG), 2010, pp. 351–358.
L. Cardamone, G. N. Yannakakis, J. Togelius, and P. L.
Lanzi, “Evolving Interesting Maps for a First Person
Shooter,” in Applications of Evolutionary Computation, C.
D. Chio, S. Cagnoni, C. Cotta, M. Ebner, A. Ekárt, A. I.
Esparcia-Alcázar, J. J. Merelo, F. Neri, M. Preuss, H.
Richter, J. Togelius, and G. N. Yannakakis, Eds. Springer
Berlin Heidelberg, 2011, pp. 63–72.
M. Cook and S. Colton, “Ludus ex machina: Building a 3D
game designer that competes alongside humans,” in
Proceedings of the 5th International Conference on
Computational Creativity, 2014.
P. L. Lanzi, D. Loiacono, and R. Stucchi, “Evolving maps
for match balancing in first person shooters,” in 2014 IEEE
Conference on Computational Intelligence and Games
(CIG), 2014, pp. 1–8.
J. Togelius, G. N. Yannakakis, K. O. Stanley, and C.
Browne, “Search-Based Procedural Content Generation: A
Taxonomy and Survey,” IEEE Trans. Comput. Intell. AI
Games, vol. 3, no. 3, pp. 172–186, Sep. 2011.
H. Takagi, “Interactive evolutionary computation: fusion of
the capabilities of EC optimization and human evaluation,”
Proc. IEEE, vol. 89, no. 9, pp. 1275 –1296, Sep. 2001.
A. Kosorukoff, “Human based genetic algorithm,” in 2001
IEEE International Conference on Systems, Man, and
Cybernetics, 2001, vol. 5, pp. 3464 –3469 vol.5.
J. Kruse and A. M. Connor, “Multi-agent evolutionary
systems for the generation of complex virtual worlds,” EAI
Endorsed Trans. Creat. Technol., vol 2, no. 5, pp. e5, Oct.
2015.
H. Takagi and H. Iba, “Interactive Evolutionary
Computation,” New Gener. Comput., vol. 23, no. 2, pp. 113–
114, Apr. 2005.
F. Herrera, M. Lozano, and J. L. Verdegay, “Tackling RealCoded Genetic Algorithms: Operators and Tools for
Behavioural Analysis,” Artif. Intell. Rev., vol. 12, no. 4, pp.
265–319, Aug. 1998.
J. Finlay, A. M. Connor, and R. Pears, “Mining Software
Metrics from Jazz,” in 2011 9th International Conference
on Software Engineering Research, Management and
Applications (SERA), 2011, pp. 39–45.
G. N. Yannakakis, P. Spronck, D. Loiacono, and E. André,
“Player Modeling.,” Artif. Comput. Intell. Games, vol. 6, pp.
45–59, 201.
| 9cs.NE |
Finding Root Causes of Floating Point Error
Alex Sanchez-Stern
Pavel Panchekha
University of California San Diego
alexss@eng.ucsd.edu
University of Washington
pavpan@cs.washington.edu
Sorin Lerner
Zachary Tatlock
University of California San Diego
lerner@cs.ucsd.edu
University of Washington
ztatlock@cs.washington.edu
arXiv:1705.10416v3 [cs.PL] 8 Mar 2018
Abstract
However, after floating-point error has been detected,
there are no tools to help a developer diagnose and debug its
root cause. The root cause is the part of a computation whose
improvement would reduce error in the program’s outputs.
Previous work in the area [4] has called out root cause analysis as a significant open problem that needs to be addressed.
This paper addresses this problem with a tool that works on
large, real, expert-crafted floating-point applications written
in C/C++, and Fortran.
In practical settings, root causes are difficult to identify
precisely. Root causes often involves computations that cross
function boundaries, use the heap, or depend on particular
inputs. As such, even though root causes are part of the computation, they rarely appear as delineable syntactic entities
in the original program text. The key challenge, then, is identifying root causes and suitably abstracting them to a form
that enables numerical analysis and facilitates improving the
accuracy of the program.
From a developer’s perspective, identifying and debugging
numerical issues is difficult for several reasons. First, floatingpoint error is non-compositional: large intermediate errors
may not impact program outputs and small intermediate
errors may blow up in a single operation (due to, e.g., cancellation or overflow). Second, floating-point errors are often
non-local: the source of an error can be far from where it is
observed and may involve values that cross function boundaries and flow through heap-allocated data structures. Third,
floating-point errors are non-uniform: a program may have
high error for certain inputs despite low or non-existent
error on other inputs.
This paper presents Herbgrind, a dynamic binary analysis
that identifies candidate root causes for numerical error in
large floating-point applications. First, to address the noncompositionality of error, Herbgrind records operations with
intermediate error and tracks the influence of these operations on program outputs and control flow. Second, to address the non-locality of error, Herbgrind provides symbolic
expressions to describe erroneous computation, abstracting
the sequence of operations that cause error to program fragments which facilitate numerical analysis. Finally, to address
the non-uniformity of error, Herbgrind characterizes both
the set of all inputs to a computation and the subset that
cause significant error.
Floating-point arithmetic plays a central role in science, engineering, and finance by enabling developers to approximate real arithmetic. To address numerical issues in large
floating-point applications, developers must identify root
causes, which is difficult because floating-point errors are
generally non-local, non-compositional, and non-uniform.
This paper presents Herbgrind, a tool to help developers identify and address root causes in numerical code written in
low-level C/C++ and Fortran. Herbgrind dynamically tracks
dependencies between operations and program outputs to
avoid false positives and abstracts erroneous computations
to simplified program fragments whose improvement can
reduce output error. We perform several case studies applying Herbgrind to large, expert-crafted numerical programs
and show that it scales to applications spanning hundreds of
thousands of lines, correctly handling the low-level details of
modern floating point hardware and mathematical libraries
and tracking error across function boundaries and through
the heap.
Keywords floating point, debugging, dynamic analysis
1
Introduction
Large floating-point applications play a central role in science, engineering, and finance by enabling engineers to approximate real number computations. Ensuring that these
applications provide accurate results (close to the ideal real
number answer) has been a challenge for decades [17–21].
Inaccuracy due to rounding errors has led to market distortions [25, 30], retracted scientific articles [1, 2], and incorrect
election results [36].
Floating-point errors are typically silent: even when a
grievous error has invalidated a computation, it will still
produce a result without any indication things have gone
awry. Recent work [3, 4] has developed dynamic analyses
that detect assembly-level operations with large intermediate
rounding errors, and recent static error analysis tools have
been developed to verify the accuracy of small numerical
kernels [11, 33].
Conference’17, July 2017, Washington, DC, USA
2018.
1
Conference’17, July 2017, Washington, DC, USA Alex Sanchez-Stern, Pavel Panchekha, Sorin Lerner, and Zachary Tatlock
We demonstrate Herbgrind’s effectiveness by identifying the root causes of error in three expert-written numerical applications and benchmarks. We find that Herbgrind
handles the tricky floating-point manipulations found in
expert-written numerical code, and can identify real sources
of floating-point error missed by experts when writing important software. To further characterize the impact of Herbgrind’s key components, we carried out a set of smaller
experiments with the FPBench floating-point benchmark
suite [10], and find that each of Herbgrind’s components is
crucial to its accuracy and performance.
To the best of our knowledge, Herbgrind provides the
first approach to identifying and summarizing root causes
of error in large numerical programs. Herbgrind is implemented in the Valgrind binary instrumentation framework,
and achieves acceptable performance using several key optimizations.1 Building Herbgrind required developing the
following contributions:
• An analysis that identifies candidate root causes by
tracking dependencies between sources of error and
program outputs, abstracting the responsible computations to an improvable program fragment, and characterizing the inputs to these computations (Section 4).
• An implementation of this analysis that supports numerical code written in low-level languages like C/C++
and Fortran and handles the complexities of modern
floating point hardware and libraries (Section 5).
• Key design decisions and optimizations required for
this implementation to achieve acceptable performance
when scaling up to applications spanning hundreds of
thousands of lines (Section 6).
• An evaluation of Herbgrind including bugs found “in
the wild” and measurements of the impact of its various subsystems (Section 7 and Section 8).
2
some error is necessarily produced. Thus, the floating-point
sum of x and y corresponds to the real number x +y +(x +y)ϵ,
where ϵ is some small value induced by rounding error.2
Floating-point operations implemented in libraries also tend
to bound error to a few ulps for each operation, but are generally not guaranteed to be the closest result. However error
grows when multiple operations are performed. For example,
consider the expression (x + 1) − x = 1. The addition introduces error ϵ1 and produces x + 1 + (x + 1)ϵ1 . The subtraction
then introduces ϵ2 and produces
1 + (x + 1)ϵ1 + ϵ2 + (x + 1)ϵ1ϵ2 .
Since x can be arbitrarily large, the (x +1)ϵ1 term can be large;
in fact, for values of x on the order of 1016 , the expression
(x +1)−x evaluates to 0, not 1. The influence of intermediate
errors on program output can be subtle; not only can accurate
intermediate operations compute an inaccurate result, but
intermediate error can cancel to produce an accurate result.
Experts often orchestrate such cancellation, which poses a
challenge to dynamic analysis tools trying to minimize false
positives.
Non-local error Floating-point error can also be non-local;
the cause of a floating point error can span functions and
thread through data structures. Consider the snippet:
double foo(struct Point a, struct Point b) {
return ((a.x + a.y) - (b.x + b.y)) * a.x;
}
double bar(double x, double y, double z) {
return foo(mkPoint(x, y), mkPoint(x, z));
}
The foo and bar functions individually appear accurate. However, bar’s use of foo causes inaccuracy. For example for inputs x=1e16, y=1, z=0, the correct output of bar is 1e16, yet
bar instead computes 0. However, this combination of foo
and bar can be computed more accurately, with the expression (y − z) · x. Note that in this example, understanding the
error requires both reasoning across function boundaries
and through data structures.
Background
A floating-point number represents a real number of the
form ±(1 + m)2e , where m is a fixed-point value between
0 and 1 and e is a signed integer; several other values, including two zeros, two infinities, not-a-number error values,
and subnormal values, can also be represented. In doubleprecision floating point, m is a 52-bit value, and e is an 11-bit
value, which together with a sign bit makes 64 bits. Simple
operations on floating-point numbers, such as addition and
multiplication, are supported in hardware on most computers.
2.1
Non-uniform error For a given computation, different inputs can cause vastly different amounts of floating-point
error. Effectively debugging a numerical issue requires characterizing the inputs which lead to the root cause impacting
output accuracy. For example, consider the snippet:
double baz(double x){
double z = 1 / (x - 113);
return (z + M_PI) - z;
}
When debugging baz, it’s important to know what inputs
baz is called on in the context of the larger program. For most
inputs, baz is accurate; if baz is only called on inputs far from
Floating-Point Challenges
Non-compositional error Individual floating-point instructions are always evaluated as accurately as possible, but since
not all real numbers are represented by a floating-point value,
1 The
2 For
implementation of Herbgrind is publicly available at [removed for
anonymized submission]
cases involving subnormal numbers, the actual error formula is more
complex; these details are elided here.
2
Finding Root Causes of Floating Point Error
Conference’17, July 2017, Washington, DC, USA
113, then a programmer need not consider it problematic.
However, for values of x near 113, baz suffers significant
rounding error, because z becomes very large, and then most
of the bits of π are lost to catastrophic cancellation. To diagnose the root cause of error in a program containing baz,
programmers need to know whether baz is called on inputs
near 113; if not, they may waste time investigating baz’s
behavior on inputs near 113 when those inputs are never
seen in practice. In this example, understanding the error
requires reasoning about the inputs on which the fragment
of code will be executed.
2.2
Figure 1. Complex plotter output before (left) and after
(right) diagnosing and fixing a floating-point error.
Existing Debugging Tools
Table 1 summarizes features of the most closely related tools
to Herbgrind.
3
Overview
We illustrate Herbgrind by detailing its use on a simple complex function plotter. Given function f : C → C, region
R = [x 0 , x 1 ] × [y0 , y1 ] in the complex plane, and a resolution,
the plotter tiles R with a grid of pixels, evaluates z = f (x +iy)
at the center (x, y) of each pixel, and colors that pixel based
on arg(z) (also known as atan2).3 Since the output of the
program is an image, minor errors in the evaluation of f
can usually be ignored. However, floating-point error can
compound in unexpected ways. Consider the function4
p
p
f (z) = 1/ ℜ(z) − ℜ(z) + i exp(−20z) .
Error detection All of the tools compared are dynamic
analyses which attempt to detect error. Like Herbgrind,
FpDebug[4] uses high-precision shadow values to track the
error of floating-point computations; Verrou[12] and BZ[3]
use heuristic methods to detect possible instances of error.
While individual hardware floating-point operations are
accurate, more complex operations, like trigonometric functions, are generally implemented in low-level math libraries
which make use of hundreds of floating-point instructions
and bit manipulations for each operation, and are painstakingly crafted by experts [16, 27]. Even higher level operations, like those on matrices and vectors, are implemented
in thousands of lower-level operations, often building on
both hardware floating-point instructions and lower-level
math libraries [22]. Previous tools report error locations as
individual opcode addresses, which may be deep within the
internals of the implementation of a sophisticated operation.
This is unfortunate, as most users are unwilling and/or unable to modify a libm or BLAS implementation; instead such
operations should be treated as atomic black boxes so that
users can focus on ensuring such operations are accurately
used at a higher level of abstraction. In contrast to previous tools, Herbgrind supports abstracting over such library
calls which enables more useful error location reporting and
permits meaningful shadowing in high-precision.
To evaluate this function, the plotter must provide codes
for evaluating the square root of a complex number. The
standard formula is
q
qp
p
√
p
2
2
2
2
x + iy =
x +y +x +i
x + y − x / 2,
where the square roots in the definitions are all square roots
of real numbers (typically provided by standard math libraries).
Implementing f using this formula and plotting the region
R = [0, 41 ] × [−3, 3] results in the left image of Figure 1. The
speckling is not an accurate representation of f ; in fact, f is
continuous in both x and y throughout R.
Herbgrind uses three key components (detailed in Section 4) to identify the root cause of error in the plotting
program: (1) a shadow taint analysis, which tracks the ways
that erroneous operations influence important program locations called spots; (2) a shadow symbolic execution, which
builds expressions representing the computations that produced each value; and (3) an input characterization system,
which maintains information about the set of inputs to each
computation.
Herbgrind detects that the plotter computes wrong pixel
values due to significant error from a subtraction with high
local error:
Root cause analysis There are two ways in which floatingpoint error can affect observable program behavior: either
by flowing directly to an output, or changing control flow.
BZ [3] can detect changes in control flow due to error; but
cannot reason about how error affects program outputs. Verrou and FpDebug have no support for detecting observable
program behavior affected by error. When an error is detected, FpDebug reports the binary address where the error
was found, while Verrou and BZ only report that something
has gone wrong, not where it has gone wrong. In contrast to
previous work, Herbgrind characterizes the inputs to problematic computations.
3 The
plotter implementation is included in supplementary material.
indicates the real (non-imaginary) part of z
4 ℜ(z)
3
Conference’17, July 2017, Washington, DC, USA Alex Sanchez-Stern, Pavel Panchekha, Sorin Lerner, and Zachary Tatlock
Feature
FpDebug
BZ Verrou
Herbgrind
Error Detection
Dynamic
✓
✓
✓
✓
Detects Error
✓
✓
✓
✓
Shadow Reals
✓
✗
✗
✓
Local Error
✗
✗
✗
✓
Library Abstraction
✗
✗
✗
✓
Root Cause Analysis
Output-Sensitive Error Report
✗
✗
✗
✓
Detect Control Divergence
✗
✓
✗
✓
Localization
Opcode Address None None Abstracted Code Fragment
Characterize Inputs
✗
✗
✗
✓
Other
Automatically Re-run in High Precision
✗
✓
✗
✗
Overhead*
395x
7.91x
7x
574x
Table 1. Comparison of floating-point error detection tools. Note that all tools are run on distinct benchmark suites.
Compare @ main.cpp:24 in run(int, int)
231878 incorrect values of 477000
Influenced by erroneous expressions:
(FPCore (x y)
:pre (and (<= -2.061152e-9 x 2.497500e-1)
(<= -2.619433e-9 y 2.645912e-9))
(- (sqrt (+ (* x x) (* y y))) x))
Example problematic input: (2.061152e-9, -2.480955e-12)
4
The implementation of Herbgrind’s analysis requires handling complex machine details including different value precisions, multiple distinct storage types, and bit-level operations. To first clarify the core concepts, this section describes
Herbgrind’s analysis in terms of an abstract float machine.
Herbgrind’s analysis consists of three components: a spotsand-influences system to determine which operations influence which program outputs (Section 4.2), a symbolic
expression system to track computations across function
and heap data structure boundaries (Section 4.3), and an input characteristics system to determine on which inputs the
computation is erroneous or accurate (Section 4.4).
This report shows that at line 24 of the main.cpp source file,
inaccuracy is caused by the expression:
p
x 2 + y2 − x .
Passing the above expression into Herbie produces the
more accurate computation
p
2
2
if x ≤ 0
p
x +y −x
2
2
x +y −x ⇝ 2 p
.
2
2
if x > 0
y / x + y + x
Substituting this expression back into the original complex
square root definition (and simplifying) yields
qp
qp
2 + y2 − x + i
|y|/
x
x 2 + y2 − x
p
1
q
q
x + iy = √
p
p
2
x 2 + y 2 + x + i |y|/
x 2 + y2 + x
Analysis
4.1
Abstract machine semantics
The abstract machine has floating-point values and operations as well as memory and control flow (Figure 2). A
machine contains mutable memory M : Z → (F | Z)
which stores floating-point values or integers and a program counter pc : Z that indexes into the list of program
statements. Statements include: computations, control operations, and outputs. A program is run by initializing the
memory and program counter, and then running statements
until the program counter becomes negative.
Herbgrind’s analysis describes candidate root causes for
programs on this abstract machine (see Figure 3 and Figure 4)
by updating analysis information for each instruction in
the program. Each computation instruction with floatingpoint output is associated with an operation entry, which
describes the computation that led up to that operation, and
a summary of the values that computation takes. All other
instructions have a spot entry, which lists the error at that
location and the erroneous computations that influence the
location (Section 4.2).
if x ≤ 0
if x > 0
for the complex square root. Replacing the complex square
root computation in the plotter with this alternative implementation fixes the inaccurate computation, as confirmed
by running Herbgrind on the repaired program. The fixed
code produces the right graph in Figure 1.
While error in the complex plotter is primarily in a single
compilation unit, real world numerical software often has
many numerical components which interact, and root causes
often cannot be isolated to one component. We evaluate
Herbgrind on several large numerical systems which exhibit
these issues in Section 7.
4
Finding Root Causes of Floating Point Error
Conference’17, July 2017, Washington, DC, USA
−x )K(M, pc) =
Jy ← f (→
−−−−→
r = Jf K(M[x])
M[y 7→ r ], pc + 1
Addr = PC = Z
M[a : Addr] : F | Z
pc : PC
−−−→
prog[n : PC] : Addr ← f (Addr)
−−−→
| if P(Addr) goto PC
| out Addr
−x ) goto nK(M, pc) =
Jif P(→
−−−−→
pc ′ = if JPK(M[x]) then n else pc + 1
M, pc ′
run(prog) =
M[−] = 0, pc = 0
while pc ≥ 0
M, pc = Jprog[pc]K(M, pc)
Jout xK(M, pc) =
print M[x]
M, pc + 1
Figure 2. The abstract machine semantics for low-level floating-point computation. A machine contains memory and a
program counter which indexes into a list of statements. Statements either compute values and store the result in memory,
perform a conditional jump, or output a value.
−−−→
Expr = R | f (Expr)
MR [a : Addr] : R
M I [a : Addr] : Set PC
M E [a : Addr] : Expr
ops[n : PC] :
Set (Expr × List (Set F) × (Position → Set F))
spots[n : PC] : (Set R) × (Set PC)
analyze(prog) =
M[−] = MR [−] = M E [−] = 0, M I [−] = ∅
pc = 0, spots[−] = (∅, ∅), ops[−] = ∅
while pc ≥ 0
M ′, pc ′ = Jprog[pc]K(M, pc)
MR′ = Jprog[pc]KR (M, MR )
M I′ = Jprog[pc]KI (MR , M I , pc)
M E′ = Jprog[pc]KE (MR′ , M E )
record(prog, pc, ops, spots, M ′, MR , M I′ , M E′ )
pc = pc ′, M = M ′, MR = MR′ , M I = M I′ , M E = M E′
return spots, ops
Figure 3. The Herbgrind analysis for finding root causes for floating-point error. Herbgrind maintains shadow memories for
real values (MR ), influences (M I ), and concrete expressions (M E ). Additionally, Herbgrind tracks concrete expressions and
input sets (both total and problematic) for operations in ops and error and influences for spots in spots.
Shadow Reals Floating-point errors are generally silent:
even when error invalidates a computation, that computation
still produces a floating-point value without any indication
that the value is erroneous. Herbgrind follows prior work in
detecting the floating-point errors in a program by computing a real number shadow for every floating-point value in
ordinary program memory.5 For each such statement, Herbgrind executes the statement’s operation in the reals on
real-number shadow inputs, and stores the resulting real
number to the real-number shadow memory. Herbgrind’s
handling of mathematical libraries and similar details is discussed in Section 5.
4.2
at a spot can be far from the spot itself. To overcome this,
Herbgrind identifies candidate root causes and tracks their
influence on spots using a taint analysis: every floating-point
value has a “taint” set of influencing instructions which is
propagated by computation.
Herbgrind uses local error to determine which floatingpoint operations cause error (Figure 4). Local error [28] measures the error an operation’s output would have even if
its inputs were accurately computed, and then rounded to
native floats. Using local error to assess operations avoids
blaming innocent operations for erroneous operands. Any
operation whose local error passes a threshold T is treated
as a candidate root cause. In the case of the complex plotter
from Section 3, the subtraction at the top of the reported
expression was determined to have high local error, resulting
in it being tracked.
Spots and Influence Shadows
Herbgrind uses the real-number execution to measure the
floating-point error at program outputs, conditional branches,
and conversions from floating-point values to integers; these
three types of program locations are collectively called spots.
Since error is non-compositional, the root cause of error
4.3 Symbolic Expressions
Identifying the operation that introduced error, as the influences system does, is not sufficient to understand the error,
because floating-point error is non-local: to understand why
an operation is erroneous requires understanding how its
5 While
the abstract analysis is defined in terms of computing over reals,
the implementation must settle merely for high precision (e.g., 1000-bit
mantissa) approximations (Section 5.1).
5
Conference’17, July 2017, Washington, DC, USA Alex Sanchez-Stern, Pavel Panchekha, Sorin Lerner, and Zachary Tatlock
record(prog, pc, ops, spots, M, MR , M I , M E ) =
−c , ĉ = ops[pc]
e,→
ε, i = spots[pc]
match prog[pc] with
−x )) when M[y] ∈ F =⇒
| (y ← f (→
′
e = M E [y] : : e
→
−c ′ = update-total-inputs(→
−x ,→
−c )
−−−−−→
if local-error(f , MR [x]) > Tℓ then
ĉ ′ = update-problematic-inputs(M E [y], ĉ)
−c ′, ĉ ′ )
ops[pc] = (e ′,→
→
−
| (y ← f ( x )) when M[y] ∈ Z =⇒
−−−−−→
if Jf KR (MR [x]) = M[y] then
Ð −−−−−→
spots[pc] = (1 : : ε, i M I [x])
else
spots[pc] = (0 : : ε, i)
−x ) goto y) =⇒
| (if P(→
−−−−−→
−−−−→
if JPKR (MR [x]) = JPK(M[x]) then
Ð −−−−−→
spots[pc] = (1 : : ε, i M I [x])
else
spots[pc] = (0 : : ε, i)
| (out x) =⇒
r = E(M[x], MR [x])
if r > Tm then
spots[pc] = (r : : ε, i ∪ M I [x])
else
spots[pc] = (r : : ε, i)
// Reals
−x )K (M, M ) when M [y] ∈ R =
Jy ← f (→
R
R
R
−−−−−→
−−−−−→
−−−−→
→
−
v = if MR [x] ∈ R then MR [x] else M[x]
−
M [y 7→ Jf K (→
v )]
R
R
// Influences
−x )K (M , M , pc) when M [y] ∈ R =
Jy ← f (→
I
I
R
R
Ð −−−−−→
s = M I [x]
−−−−−→
if local-error(f , MR [x]) > Tℓ then
s = pc : : s
M I [y 7→ s]
−
local-error(f ,→
v)=
−
r R = F(Jf KR (→
v ))
−−−→
r F = Jf KF (F(v))
E(r R , r F )
// Expressions
−x )K (M , M ) when M [y] ∈ R =
Jy ← f (→
E
E
R
R
−−−−−→
−−−−−→
e = if MR [x] ∈ R then f (M E [x]) else MR [y]
M E [y 7→ e]
update-problematic-inputs(e, ĉ)
nodes, positions =
get-all-descendant-nodes(e)
ĉ ′ = make-table()
for node, position in zip(nodes, positions) :
ĉ ′ [position] = ĉ[position] + {node.value}
−
−c ) =
update-total-inputs(→
v ,→
→
−c ′ = []
−
−c ) :
for v, c in zip(→
v ,→
→
−c ′ = (c + {v}) :: →
−c ′
→
−c ′ = reverse(→
−c ′ )
Figure 4. On the left, the real-number execution, influence propagation, and concrete expressions building in Herbgrind;
shadow executions not shown are no-ops. On the right, how Herbgrind updates the operation and spot information on every
statement. Below are helper functions.
inputs were computed. In most cases those inputs are separated from the erroneous operation by function boundaries
and heap data structure manipulations. Herbgrind analyzes
through those boundaries by providing symbolic expressions
for the erroneous operation and its inputs.
Symbolic expressions represent an abstract computation
that contains a candidate root cause. Each symbolic expression contains only floating-point operations: it therefore abstracts away function boundaries and heap data structures. A
symbolic expression must be general enough to encompass
any encountered instance of a computation, while still being
specific enough to be as helpful as possible to the user.
Herbgrind constructs symbolic expressions by first recording a concrete expression for every floating-point value and
then using anti-unification to combine these concrete expressions into symbolic expressions. Each concrete expression
tracks the floating-point operations that were used to build
a particular value. Concrete expression are copied when a
value is passed to a function, entered in a heap data structure,
or stored and retrieved later, but these operations themselves
are not recorded. A single concrete expression (and the resulting symbolic expression), might encompass a subtraction
from one function, a series of multiplications from another,
and an exponentiation which occurs after the value is stored
in a hash table and later retrieved.
From the concrete expression for every value, Herbgrind
computes symbolic expressions using a variant of the classic anti-unification algorithm [29] for computing the mostspecific generalization of two trees. Symbolic expressions
are much like concrete expressions, but include variables
which can stand in for any subtree; variables which stand in
for equivalent subtrees are the same. To produce expressions
which are more useful for program improvement, Herbgrind
6
Finding Root Causes of Floating Point Error
Conference’17, July 2017, Washington, DC, USA
5.1
uses a modified version of anti-unification. These modifications are described in Appendix A. Reporting detailed
symbolic expressions is essential for diagnosing the root
causes of error; in the case of the plotter, the full extracted
expression was essential for producing an improved complex
square root definition.
4.4
Unlike the abstract machine, VEX has values of different
sizes and semantics. Floating-point values come in different
precisions, so the Herbgrind implementation makes a distinction between single- and double-precision floating-point
values. Both types of values are shadowed with the same
shadow state, but their behaviors in the client program are
different, and they have different sizes in memory. Client
programs also have file descriptors, pointers, and integers
of various sizes, but this does not affect Herbgrind, since it
does not analyze non-floating-point computations.
The algorithm in Section 4 tracks the exact value of computations by shadowing floating-point values with real numbers. Herbgrind approximates the real numbers using the
MPFR library [13] to shadow floating-point values with
arbitrary-precision floating-point.7 As an alternative, we
could use an efficient library for the computable reals [5, 23,
26].8
Input Characteristics
Because floating-point error is non-uniform, the error of
a computation is highly dependent on its inputs. In many
cases, a developer must know on which inputs a computation
is run on in order to improve its error behavior. Herbgrind
satisfies this need by computing input characteristics for each
symbolic expression it produces. These input characteristics
summarize both the inputs that the expression is called on,
and the ones it has high error on.6
Herbgrind’s input characteristics system is modular, and
Herbgrind comes with three implementations. In the first
kind of input characteristic, a representative input is selected
from the input. Since input characteristics are done once
for all inputs and again for only inputs that lead to high
error, this means this system provides two inputs: one with
high error and one with low error. In the second kind of
input characteristic, ranges are tracked for each variable in a
symbolic expression. In the third kind of input characteristic,
ranges are tracked separately for positive and negative values
of each variable.
5
Values
5.2
Storage
The abstract machine model of Section 4 represents storage
as a single map from locations to values. However, VEX
has three different types of storage— temporaries, thread
state, and memory— and the latter two store unstructured
bytes, not values directly. Herbgrind uses slightly different
approaches for each.
To support SIMD instructions in the SSE instruction set,
temporaries can contain multiple floating-point values, unlike the memory locations in the abstract machine. Herbgrind attaches a shadow temporary to each temporary: a
type-tagged unit which can store multiple shadow values.
The shadow values stored in a shadow temporary correspond
to the individual floating-point values inside a SIMD vector. Temporaries that only store a single value have trivial
shadow temporaries.
Thread state in VEX, which represents machine registers,
is an unstructured array of bytes, so it does not use shadow
temporaries. Each floating-point value consumes multiple
bytes, and floating-point values of different sizes take up different numbers of bytes. This means that for reads and writes
to memory, Herbgrind must be careful to check whether they
overwrite nearby memory locations with shadow values.
Herbgrind also supports writing SIMD results to memory
and reading them back at an offset, as long as the boundaries
of individual values are respected. In the rare cases where
client programs make misaligned reads of floating-point values, Herbgrind conservatively acts as if the read computes a
floating-point value from non-floating-point inputs.
Like thread state, memory is an unstructured array of
bytes, with the complication that it is too large to shadow
Implementation
The previous section described Herbgrind’s analysis in terms
of an abstract machine; however, important numerical software is actually written in low level languages like C, C++,
and Fortran—sometimes a polyglot of all three.
To support all these use cases, we refine the algorithm
presented in Section 4 to operate on compiled binaries instead of abstract machine programs. Herbgrind does this by
building upon Valgrind, a framework for dynamic analysis
through binary instrumentation. Building upon Valgrind requires mapping the abstract machine described in Section 4
to the VEX machine internal to Valgrind.
Analyses built upon Valgrind receive the instructions of
the client program translated to VEX, and can add instrumentation to them freely before they are compiled back to
native machine code and executed.
Implementing the algorithm from Section 4 with Valgrind
requires adapting the abstract machine’s notions of values,
storage, and operations to those provided by Valgrind.
6
To run Herbgrind, a developer usually runs their code on a set of representative inputs. Herbgrind’s input characteristics apply not to these inputs,
but to the inputs of symbolic expressions, usually intermediate values in
the program.
7 The
precision used is configurable, set to 1000 by default.
treats real computation as an abstract data type and alternate
strategies could easily be substituted in.
8 Herbgrind
7
Conference’17, July 2017, Washington, DC, USA Alex Sanchez-Stern, Pavel Panchekha, Sorin Lerner, and Zachary Tatlock
Laziness Since program memory is untyped, it is initially
impossible to tell which bytes in the program correspond
to floating-point values. Herbgrind therefore tracks floatingpoint values in memory lazily: as soon as the client program
executes a floating-point operation on bytes loaded from a
memory location, that location is treated as a floating-point
location and shadowed by a new shadow value.
Besides lazily shadowing values in the client program, Herbgrind also minimizes instrumentation. Some thread state
locations can always be ignored, such as CPU flag registers.
VEX also adds a preamble to each basic block representing
the control flow effects of the architecture, which Herbgrind
also ignores.
For more fine-grained instrumentation minimization, Herbgrind makes use of static superblock type analysis. Values
known to be integers do not have to be instrumented, and
values known to be floating point can have type checking
elided. This combination of static type analysis and dynamic
error analysis is crucial for reducing Herbgrind’s overhead.
Unfortunately, Herbgrind must still instrument many simple
memory operations, since values that are between storage
locations but not operated on could have shadow values.
The static type analysis is also used to reduce reduce calls
from the instrumentation into Herbgrind C functions. Valgrind allows the instrumentation to call into C functions
provided by Herbgrind, which then compute shadow values,
build concrete expressions, and track influences. However,
calls from client program to host functions are slow. The
static type analysis allows inlining these computations directly into VEX, avoiding a client-host context switch, because the type system tracks the size of values in thread
state. Knowing the size means no type or size tests need to
be done, so instrumentation can be inlined without requiring
branches and thus crossing superblock boundaries. Inlining
is also used for copies between temporaries, and for some
memory accesses, where the inlined code must also update
reference counts.
completely. Herbgrind shadows only memory that holds
floating-point values; memory is shadowed by a hash table
from memory addresses to shadow values.
5.3
Operations
In the abstract machine model, all floating-point operations
are handled by specialized instructions. However, few machines support complex operations such as logarithms or
tangents in hardware. Instead, client programs evaluate these
functions by calling libraries, such as the standard libm.
Shadowing these internal calculations directly would miscompute the exact value of library calls; Herbgrind therefore
intercepts calls to common library functions before building
concrete expressions and measuring error. For example, if
a client program calls the tan function, Herbgrind will intercept this call and add tan to the program trace, not the
actual instructions executed by calling tan.
Expert-written numerical code often uses “compensating”
terms to capture the error of a long chain of operations, and
subtract that error from the final result. In the real numbers,
this error term would always equal zero, since the reals don’t
have any error with respect to themselves. Yet in floating
point, these “compensating” terms are non-zero and computations that produce them therefore have high local error. A
naive implementation of Herbgrind would therefore report
spots influenced by every compensated operation used to
compute it, even though the compensating terms increase
the accuracy of the program.
Instead, Herbgrind attempts to detect compensating operations, and not propagate influence from the compensating
term to the output of the compensated operation. Herbgrind
identifies compensating operations by looking for additions
and subtractions which meet two criteria: they return one of
their arguments when computed in the reals; and the output
has less error than the argument which is passed through.
The influences for the other, compensating, term, are not
propagated.
While most floating-point operations in real programs are
specialized floating-point instructions or library calls, some
programs use bitwise operations to implement a floatingpoint operation. Programs produced by gcc negate floatingpoint values by XORing the value with a bitmap that flips the
sign bit, and a similar trick can be used for absolute values.
Herbgrind detects and instruments these bitwise operations,
treating them as the operations they implement (including
in concrete expressions).
6
Sharing Many floating-point values are copies of each
other, scattered in temporaries, thread state, and memory.
Though copying floating-point values is cheap on most architectures, copying shadow values requires copying MPFR
values, concrete expressions, and influence sets. To save
time and memory, shadow values are shared between copies.
Shadow values are reference counted to ensure that they can
be discarded once they no longer shadow any floating-point
values. The trace nodes stored in a shadow value are not
freed along with the shadow value, since traces also share
structure. Traces are therefore reference counted as well,
with each shadow value holding a reference to its trace node,
and each trace node holding references to its children.
Optimization
A direct implementation of the algorithm in Section 5 is prohibitively expensive. Herbgrind improves on it by using the
classic techniques of laziness, sharing, incrementalization,
and approximation.
8
Finding Root Causes of Floating Point Error
Conference’17, July 2017, Washington, DC, USA
Many shadow values are freed shortly after they are created. Related data structures, like trace nodes, are also allocated and freed rapidly, so memory allocation quickly becomes a bottleneck. Herbgrind uses custom stack-backed
pool allocators to quickly allocate and free many objects of
the same size.
In the Gram-Schmidt orthonormalization kernel, Herbgrind detected a floating-point problem which it reported to
have 64 bits of error, surprising for a vetted numerical benchmark. Upon investigating, we found that Gram-Schmidt decomposition is not well defined on the given inputs, resulting
in a division by zero; Herbgrind reports the resulting NaN
value as having maximal error. The fundamental problem is
not in the Gram-Schmidt procedure itself but in its invocation on an invalid intermediate value. Luckily, Herbgrind provides, as its example problematic input to the computation in
question, a zero vector, an invalid input to Gram-Schmidt orthonormalization. Herbgrind’s input characteristics system
was able to link the error in the program output to the root
cause of an input violating the precondition of the orthnormalization procedure. Note that there was nothing wrong
with the procedure itself, but rather its interaction with the
program around it. Upon understanding the bug (and fixing
it ourselves), we tested version 4.2.0 of Polybench and confirmed that this more recent version fixed the problem by
changing the procedure that generated the vectors to ensure
that valid inputs to Gram-Schmidt orthonormalization are
produced.
Incrementalization The algorithm in Section 4 accumulates errors, concrete expressions, and operation inputs perinstruction, and summarizes all the results after the program
finishes running. For long-running programs, this approach
requires storing large numbers of ever-growing concrete
expressions. The implementation of Herbgrind avoids this
problem by aggregating errors (into average- and maximumtotal and local errors) concrete expressions (into symbolic
expressions) and inputs (into input characteristics) incrementally, as the analysis runs. This leads to both large memory
savings and significant speed-ups.
This incrementalization does not change the analysis results since our implementation of anti-unification, used to
aggregate concrete expressions, and summation, used to aggregate error, are associative.
6.1
Approximation
PID Controller A proportional-integral-derivative controller
is a control mechanism widely used in industrial control systems. The controller attempts to keep some measure at a fixed
value. It runs in a loop, receiving the current value of the
measure as input and outputting the rate at which to increase
or decrease the measure. We investigated an adaptation of
a simple PID controller which runs for a fixed number of
iterations and with a fixed rate of change to the measure [9].
We initially ran Herbgrind on the PID controller expecting
to find, perhaps, some floating-point error in the controller
code itself. Instead, we found that Herbgrind was detecting a
problem in the loop condition. To run the PID controller for
a limited number of seconds, the program tests the condition
(t < N), where N is the number of seconds. The variable t
is stored as a double-precision floating-point number, and
is incremented by 0.2 on every iteration through the loop.
As we experimented with different loop bounds, Herbgrind
noticed that the condition, for some loop bounds, iterates
once too many times. For example, if the loop bound is set
to 10.0, the loop executes 51 times, not 50 times, because
adding 0.2 to itself 50 times produces a value 3.5 · 10−15 less
than 10. This bug is closely related to one that occurred in
the Patriot missile defense system in 1992, resulting in the
death of 28 people [34]. Herbgrind’s automatic marking of
all control flow operations as spots was necessary to detect
the bug and link the inaccurate increment to its affect on
control flow. Herbgrind was successfully able to trace back
from error detected in the output of the program to the
root cause of the error, the inaccurate increment; the output
contained the source location of the erroneous compare and
reported that it was influenced by the inaccurate increment.
Herbgrind makes a sound, but potentially incomplete, approximation to the standard anti-unification algorithm to
speed it up. Anti-unification requires knowing which pairs
of nodes are equivalent, so that those nodes could be generalized to the same variable. Symbolic expressions can be
trees hundreds of nodes deep, and this equivalence information must be recomputed at every node, so computing
these equivalence classes for large trees is a significant portion of Herbgrind’s runtime. To limit the cost, Herbgrind exactly computes equivalence information to only a bounded
depth for each node, 5 by default. In practice, this depth
suffices to produce high-quality symbolic expressions. This
depth bound also allows freeing more concrete program trace
nodes, further reducing memory usage.
7
Case Studies
This section describes three examples where Herbgrind was
used to identify the root causes of floating-point errors in
numerical programs: in all three cases, the bugs were fixed
in later versions of the software. Herbgrind’s three major
subsystems were crucial in detecting and understanding
these bugs.
Gram-Schmidt Orthonormalization Gram-Schmidt orthonormalization transforms a collection of vectors into a orthogonal basis of unit vectors for their span. We investigated
an implementation of Gram-Schmidt orthonormalization
provided by the Polybench benchmark suite for numerical
kernels. Polybench is provided in several languages; we used
the C version of Polybench 3.2.1.
9
Conference’17, July 2017, Washington, DC, USA Alex Sanchez-Stern, Pavel Panchekha, Sorin Lerner, and Zachary Tatlock
We notified the authors of the adapted PID controller and
they confirmed the bug and identified a fix: incrementing
the t variable by 1 instead of 0.2, and changing the test to
(t * 0.2 < N).
Running on Triangle’s example inputs, we found that Herbgrind’s compensation detection correctly handles all but
14 of 225 compensating terms with local error and does not
present these false positives to the user.
The 14 remaining compensated operations are not detected, because the compensating term affects control flow:
Triangle checks whether the compensating term is too large,
and if so runs the same computation in a different way. Herbgrind’s real-number execution computes the accurate value
of a compensating term to be 0, so these branches often
go the “wrong way”. Fortunately, given Herbgrind’s candidate root cause, this behavior is always easy to check in the
Triangle source.
Gromacs Gromacs is a molecular dynamics package used
for simulating proteins, lipids, and nucleic acids in drug
discovery, biochemistry, and molecular biology. We investigated the version of Gromacs that ships with the SPEC
CPU 2006 benchmark suite. Gromacs is a large program—
42 762 lines of C, with the inner loops, which consume approximately 95% of the runtime, written in 21 824 lines of
Fortran. We tested Gromacs on the test workload provided
by CPU 2006, which simulates the protein Lysozyme in a
water-ion solution.
During this run, Herbgrind reported an error in the routine
that computes dihedral angles (the angle between two planes,
measured in a third, mutual orthogonal plane). For inputs
where the dihedral angle is close to flat, corresponding to
four colinear molecules, the dihedral angle computation was
returning values with significant error due to cancellation
in the computation of a determinant. These cases, though
a small subset of all possible angles, were important. First,
collections of four colinear molecules are common, for example in triple-bonded organic compounds such as alkynes.
Second, molecular dynamics is chaotic, so even small errors
can quickly cause dramatically different behavior.
Herbgrind’s symbolic expression system was crucial in
understanding the root cause of this bug. The dihedral angle
procedure invokes code from multiple source files, across
both C and Fortran, moving data into and out of vector data
structures. The symbolic expression gathered together the
slivers of computation that contributed to the high rounding
error. From the expression reported by Herbgrind the potential for cancellation was clear and the input characteristics
provided by Herbgrind allowed us to narrow our investigations to flat angles. We identified the problem and developed
a fix based on the numerical analysis literature [32]. After
we contacted the developers, they confirmed the bug and
explained that they had deployed a similar fix in recent Gromacs versions.
8
8.1
8.2
Improvability
Methodology We define the root cause of floating point
error as the part of a program whose improvement would
reduce error in the outputs. To evaluate Herbgrind’s effectiveness, we measure the “improvability” of the root causes
that it identifies in various programs.
Improvability is hard to measure directly, since there are
many approaches to improving floating point code. As a
proxy for the improvability of program fragments, we picked
one tool, Herbie [28], and ran it on the program fragments
produced by Herbgrind.9 We evaluated Herbgrind on benchmarks from the FPBench general-purpose floating-point
benchmark suite [10].
Since Herbgrind is a tool for diagnosing floating-point
error in binary programs, the first step to using FPBench
benchmarks is to compile them to native code. We do so
by using the publicly available FPCore-to-C compiler provided with FPBench, and then compiling this C code, along
with some driver code which exercises the benchmarks on
many inputs, using the GNU C Compiler. We then run it
under Herbgrind, and pass the resulting FPCore output to
Herbie. For some performance numbers, we also timed the
benchmarks.10
All experiments were run on an Intel Core i7-4790K processor at 4GHz with 8 cores, running Debian 9 with 32 Gigabytes
of memory. Herbgrind introduces a 574x overhead on the
FPBench suite.
Benchmarks in the FPBench benchmark suite are, for the
most part, minimized test cases which exhibit floating point
error. This means that in most cases, the perfect expression for Herbgrind to extract is the one present in the original benchmark. We compare Herbgrind against an “oracle”
which directly extracts the relevant symbolic expression
from the source benchmark.
Evaluation
Handling expert tricks
Detecting compensating terms (see Section 5.3) in client code
is important for reducing false positives in Herbgrind’s output. To test the compensation detection system, we applied
Herbgrind to analyze Triangle, an expert-written numerical
program.
Triangle [31], written in C by Jonathan Shewchuk, is a
mesh generation tool, which computes the Delaunay triangulation of a set of input points, and can add additional points
so that the triangulation produced satisfies various stability properties, such as avoiding particularly sharp angles.
9 We
include in our artifact the tool, and scripts to connect it to the benchmark suite, so that these results can be easily reproduced.
10 Our original timing code actually had a floating point bug, which we
discovered when Herbgrind included it in its output.
10
Finding Root Causes of Floating Point Error
Conference’17, July 2017, Washington, DC, USA
Results We found that 34 of our 57 benchmarks were determined by Herbgrind to have significant error (> 5 bits). To
test Herbgrind’s assessment of the error of these problematic
computations, we take the symbolic expression it produced,
and ask Herbie to estimate its average error through sampling.
In 21 of the benchmarks with computations reported by
Herbgrind to have error, Herbie could independently detect
error in the expressions produced by Herbgrind. Using the
FPCore oracle produces computations that Herbie can independently determine to be erroneous for 31 benchmarks.
In some cases, the error reported by Herbgrind and the
error estimated by Herbie are similar. Where they are not, it is
often because the pre-conditions produced by Herbgrind are
too coarse; the current implementation of Herbgrind’s input
characterization can miss more fine-grained aspects of their
distribution. As discussed in Section 4.4, the type of input
characterization is orthogonal to Herbgrind’s approach, and
can easily be specialized as needed.
Finally, for 18 benchmarks Herbgrind extracted expressions which Herbie could automatically significantly improve. Using the FPCore oracle produces Herbie-improvable
expressions for 31 benchmarks. This shows that Herbgrind
is useful for diagnosing the root causes of floating point error.
That is, using Herbgrind yields code fragments with significant error that can be improved in isolation to improve the
end-to-end accuracy of the program, even with no human
interaction.
A maximum expression depth of 1 node deep effectively
disables symbolic expression tracking, and only reports the
operation where error is detected, much like FpDebug and
similar floating-point debugging tools. However, unlike those
tools, it still tracks the influence of error and the range of
inputs to the operation. As you can see from the figure, not
tracking operations before the one that produced error results in a speedup over the normal configuration, but at a
high cost: none of the expressions produced are significantly
improvable.
Finally, to test the effectiveness of input characterization,
we measured the improvability of our benchmarks in three
configurations: with ranges turned off , with a single range
for all inputs, and with separate ranges for positive and negative inputs (see Figure 5b). Adding input ranges significantly
increases the improvability of Herbgrind’s output, but using
more detailed negative and positive ranges doesn’t help (and
tends to confuse the improver).
Library Wrapping Herbgrind instruments calls to mathematical library functions such as sqrt and tan to correctly
evaluate the exact result of a computation and provide simpler symbolic expressions. With this wrapping behavior
turned off, Herbgrind finds significantly more complex expressions, representing the internals of library functions: the
largest expressions are not 9 but 31 operations in size, and
133 expressions11 have more than 9 operations. For example,
instead of e x − 1, Herbgrind finds 17 expressions such as
(x−0.6931472(y−6.755399e15)+2.576980e10)−2.576980e10.
8.3
Subsystems
Furthermore, as discussed in Section 5, without wrapping
calls to mathematical libraries, Herbgrind measures output
accuracy incorrectly, though on the FPBench examples the
inaccuracy is slight.
Herbgrind’s analysis consists of three main subsystems, influence tracking, symbolic expressions, and input characteristics. In this section we will demonstrate the effect of these
subsystems.
Figure 5a shows the result of running Herbgrind with
various error thresholds for the influences system. The error threshold selected determines how much local error an
expression has to have before it is marked as “significantly
erroneous”, and tracked as a source of error.
A higher threshold means that fewer computations are
reported as being problematic. Users might want to use a
higher error threshold on certain applications when there
are too many expressions that are somewhat erroneous to
address them all. In highly critical applications, where it
is important to address even a small source of error, users
might choose to lower the error threshold to catch even more
errors.
Because of the non-local nature of floating-point error, the
root cause of error is often spread across many floating-point
operations. To measure how far root causes are spread, we
varied the maximum expression depth Herbgrind tracked.
In Figure 5c and Figure 5d, we measured the runtime and effectiveness of Herbgrind using various maximum expression
depths.
9
Related Work
There is a rich literature on analyzing and mitigating floatingpoint error. Below we discuss the most closely related work.
Recently, work on statically analyzing error for floating
point programs has made tremendous progress [6, 8, 11, 14,
15, 24, 33]. Broadly, this work focuses on providing sound,
though conservative, error bounds for small numerical kernels. This work is useful for reasoning about the expressions
returned by Herbgrind, but does not on its own scale to large
numerical software.
Several papers in recent years have used dynamic analyses
to analyze floating-point error.
FpDebug [4] uses Valgrind to build a dynamic analysis of
floating point error. Like Herbgrind, it uses MPFR shadow
values to measure the error of individual computations. Unlike FpDebug however, Herbgrind’s shadow real execution
11
With library wrapping disabled, Herbgrind identifies 848 problematic
expressions, mostly corresponding to false positives in the internals of the
math library.
11
Conference’17, July 2017, Washington, DC, USA Alex Sanchez-Stern, Pavel Panchekha, Sorin Lerner, and Zachary Tatlock
Figure 5. In (a) we compare the number of computations flagged with various error thresholds. In (b) we show how many
benchmarks can be improved with various types of ranges. (c) and (d) explore the effect of different maximum expression
depths on runtime and number of benchmarks improved.
is based on a model of full programs, including control flow,
conversions, and I/O, as opposed to FpDebug’s model of
VEX blocks. This enables a rigorous treatment of branches
as spots, and leads to extensions such as wrapping library
functions, sharing shadow values, SIMD operations, and
bit-level transformations. All these features required significant design and engineering to scale to 300 KLOC numerical
benchmarks from actual scientific computing applications.
In addition to an improved real execution, Herbgrind departs from FpDebug with its spots and influences system,
symbolic expressions, and input ranges, which allow it to
connect inaccurate floating-point expressions to inaccurate
outputs produced by the program. Herbgrind’s use of local
error, symbolic expressions, and input ranges, help the user
diagnose the parts of the program that contributed to the
detected error.
Similarly to FpDebug and Herbgrind, Verrou [12] is a dynamic floating point analysis built on Valgrind. Verrou’s aim
is also to detect floating point error in numerical software,
but attempts to do so at much lower overhead. The resulting approach uses very little instrumentation to perturb the
rounding of a floating point program, thus producing a much
more conservative report of possible rounding errors.
Recent work by [3] also attempts to detect floating-point
error with low overhead, with the goal of determining when
floating-point error flows into what they call “discrete factors”. The tool is designed to detect the possible presence
of inaccuracy with very low runtime overhead to enable
re-running in higher precision. In this context, a very high
false positive rate (> 80-90% in their paper) is acceptable,
but it is not generally acceptable as a debugging technique.
Bao and Zhang’s discrete factors address only floating-point
errors that cause changes in integer or boolean values (hence
“discrete”). Unlike [3], Herbgrind tracks all factors (not just
discrete ones), including changes in floating-point values
that lead to changes in floating-point outputs. Re-running
programs in higher precision is untenable in many contexts,
but may work for some.
Herbie [28] is a tool for the automatically improving the
accuracy of small floating point expressions (≈ 10 LOC). Herbie uses randomly sampled input points and an MPFR-based
ground truth to evaluate expression error. This statistical, dynamic approach to error cannot give sound guarantees, but
is useful for guiding a search process. Herbie’s main focus
is on suggesting more-accurate floating-point expressions
to program developers. Herbie can be combined with Herbgrind to improve problematic floating point code in large
numerical programs, by feeding the expressions produced
by Herbgrind directly into Herbie to improve them.
Wang, Zou, He, Xiong, Zhang, and Huang [35] develop
a heuristic to determine which instructions in core mathematical libraries have an implicit dependence on the precision of the floating-point numbers. A ground truth for such
precision-specific operations cannot be found by evaluating
the operations at higher precision. These results justify Herbgrind detecting and abstracting calls to the floating-point
math library.
CGRS [7] uses evolutionary search to find inputs that
cause high floating-point error; these inputs can be used for
debugging or verification. Unlike Herbgrind, these inputs
can be unrealistic for the program domain, and CGRS does
not help the developer determine which program expressions created the high error. However, users who want to
analyze the behavior of their programs on such inputs can
use Herbgrind to do so.
10
Conclusion
Floating point plays a critical role in applications supporting
science, engineering, medicine, and finance. This paper presented Herbgrind, the first approach to identifying candidate
root causes of floating point errors in such software. Herbgrind does this with three major subsystems: a shadow taint
analysis which tracks the influence of error on important program locations, a shadow symbolic execution which records
the computations that produced each value, and an input
characterization system which reports the inputs to problematic computations. Herbgrind’s analysis is implemented on
top of the Valgrind framework, and finds bugs in standard
numerical benchmarks and large numerical software written
by experts.
12
Finding Root Causes of Floating Point Error
Conference’17, July 2017, Washington, DC, USA
References
[1] Micah Altman, Jeff Gill, and Michael P. McDonald. 2003. Numerical
Issues in Statistical Computing for the Social Scientist. Springer-Verlag.
1–11 pages.
[2] Micah Altman and Michael P. McDonald. 2003. Replication with
attention to numerical accuracy. Political Analysis 11, 3 (2003), 302–
307. http://pan.oxfordjournals.org/content/11/3/302.abstract
[3] Tao Bao and Xiangyu Zhang. 2013. On-the-fly Detection of Instability
Problems in Floating-point Program Execution. SIGPLAN Not. 48, 10
(Oct. 2013), 817–832. DOI:http://dx.doi.org/10.1145/2544173.2509526
[4] Florian Benz, Andreas Hildebrandt, and Sebastian Hack. 2012. A
Dynamic Program Analysis to Find Floating-point Accuracy Problems
(PLDI ’12). ACM, New York, NY, USA, 453–462. http://doi.acm.org/10.
1145/2254064.2254118
[5] Hans-J. Boehm. 2004. The constructive reals as a Java Library. J. Log.
Algebr. Program 64 (2004), 3–11.
[6] Wei-Fan Chiang, Mark Baranowski, Ian Briggs, Alexey Solovyev,
Ganesh Gopalakrishnan, and Zvonimir Rakamarić. 2017. Rigorous Floating-point Mixed-precision Tuning. In Proceedings of the
44th ACM SIGPLAN Symposium on Principles of Programming Languages (POPL 2017). ACM, New York, NY, USA, 300–315. DOI:http:
//dx.doi.org/10.1145/3009837.3009846
[7] Wei-Fan Chiang, Ganesh Gopalakrishnan, Zvonimir Rakamarić, and
Alexey Solovyev. 2014. Efficient Search for Inputs Causing High
Floating-point Errors. ACM, 43–52.
[8] Nasrine Damouche, Matthieu Martel, and Alexandre Chapoutot. 2015.
Formal Methods for Industrial Critical Systems: 20th International
Workshop, FMICS 2015 Oslo, Norway, June 22-23, 2015 Proceedings.
(2015), 31–46.
[9] N. Damouche, M. Martel, and A. Chapoutot. 2015. Transformation of a
{PID} Controller for Numerical Accuracy. Electronic Notes in Theoretical
Computer Science 317 (2015), 47 – 54. DOI:http://dx.doi.org/10.1016/j.
entcs.2015.10.006 The Seventh and Eighth International Workshops
on Numerical Software Verification (NSV).
[10] Nasrine Damouche, Matthieu Martel, Pavel Panchekha, Jason Qiu,
Alex Sanchez-Stern, and Zachary Tatlock. 2016. Toward a Standard
Benchmark Format and Suite for Floating-Point Analysis. (July 2016).
[11] Eva Darulova and Viktor Kuncak. 2014. Sound Compilation of Reals
(POPL ’14). ACM, New York, NY, USA, 235–248. http://doi.acm.org/10.
1145/2535838.2535874
[12] François Févotte and Bruno Lathuilière. 2016. VERROU: Assessing
Floating-Point Accuracy Without Recompiling. (Oct. 2016). https:
//hal.archives-ouvertes.fr/hal-01383417 working paper or preprint.
[13] Laurent Fousse, Guillaume Hanrot, Vincent Lefèvre, Patrick Pélissier,
and Paul Zimmermann. 2007. MPFR: A Multiple-Precision Binary
Floating-Point Library with Correct Rounding. ACM Trans. Math.
Software 33, 2 (June 2007), 13:1–13:15. http://doi.acm.org/10.1145/
1236463.1236468
[14] Eric Goubault and Sylvie Putot. 2011. Static Analysis of Finite Precision
Computations (VMCAI’11). Springer-Verlag, Berlin, Heidelberg, 232–
247. http://dl.acm.org/citation.cfm?id=1946284.1946301
[15] Nicholas J. Higham. 2002. Accuracy and Stability of Numerical Algorithms (2nd ed.). Society for Industrial and Applied Mathematics,
Philadelphia, PA, USA.
[16] Andreas Jaeger. 2016. OpenLibm. (2016). http://openlibm.org/
[17] W. Kahan. 1965. Pracniques: Further Remarks on Reducing Truncation
Errors. Commun. ACM 8, 1 (Jan. 1965), 40–. DOI:http://dx.doi.org/10.
1145/363707.363723
[18] William Kahan. 1971. A Survey of Error Analysis.. In IFIP Congress
(2). 1214–1239. http://dblp.uni-trier.de/db/conf/ifip/ifip71-2.html#
Kahan71
[19] W. Kahan. 1987. Branch Cuts for Complex Elementary Functions
or Much Ado About Nothing’s Sign Bit. In The State of the Art in
Numerical Analysis (Birmingham, 1986), A. Iserles and M. J. D. Powell
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
13
(Eds.). Inst. Math. Appl. Conf. Ser. New Ser., Vol. 9. Oxford Univ. Press,
New York, 165âĂŞ211.
W. Kahan. 1998. The Improbability of Probabilistic Error Analyses for
Numerical Computations. Technical Report. 34 pages. http://www.cs.
berkeley.edu/~wkahan/improber.pdf
William Kahan. 2005. Floating-Point Arithmetic Besieged by “Business
Decisions”. World-Wide Web lecture notes.. (5 July 2005). http://www.
cs.berkeley.edu/~wkahan/ARITH_17.pdf
C. L. Lawson, R. J. Hanson, D. R. Kincaid, and F. T. Krogh. 1979. Basic
Linear Algebra Subprograms for Fortran Usage. ACM Trans. Math.
Softw. 5, 3 (Sept. 1979), 308–323. DOI:http://dx.doi.org/10.1145/355841.
355847
Vernon A. Lee, Jr. and Hans-J. Boehm. 1990. Optimizing Programs
over the Constructive Reals. In Proceedings of the ACM SIGPLAN
1990 Conference on Programming Language Design and Implementation (PLDI ’90). ACM, New York, NY, USA, 102–111. DOI:http:
//dx.doi.org/10.1145/93542.93558
Matthieu Martel. 2009. Program Transformation for Numerical Precision (PEPM ’09). ACM, New York, NY, USA, 101–110. http://doi.acm.
org/10.1145/1480945.1480960
B. D. McCullough and H. D. Vinod. 1999. The Numerical Reliability
of Econometric Software. Journal of Economic Literature 37, 2 (1999),
633–665.
Marvin L. Minsky. 1967. Computation: Finite and Infinite Machines.
Prentice-Hall, Inc., Upper Saddle River, NJ, USA.
Dr. K-C Ng. 1993. FDLIBM. (1993). http://www.netlib.org/fdlibm/
readme
Pavel Panchekha, Alex Sanchez-Stern, James R. Wilcox, and Zachary
Tatlock. 2015. Automatically Improving Accuracy for Floating Point
Expressions. In Proceedings of the 36th ACM SIGPLAN Conference on
Programming Language Design and Implementation (PLDI ’15). ACM.
Gordon D. Plotkin. 1970. A note on inductive generalization. Machine
Intelligence 5 (1970), 153–163.
Kevin Quinn. 1983. Ever Had Problems Rounding Off Figures? This
Stock Exchange Has. The Wall Street Journal (November 8, 1983), 37.
Jonathan Richard Shewchuk. 1996. Triangle: Engineering a 2D Quality
Mesh Generator and Delaunay Triangulator. In Applied Computational Geometry: Towards Geometric Engineering, Ming C. Lin and
Dinesh Manocha (Eds.). Lecture Notes in Computer Science, Vol. 1148.
Springer-Verlag, 203–222. From the First ACM Workshop on Applied
Computational Geometry.
Hang Si. 2015. TetGen, a Delaunay-Based Quality Tetrahedral Mesh
Generator. ACM Trans. Math. Softw. 41, 2, Article 11 (Feb. 2015),
36 pages. DOI:http://dx.doi.org/10.1145/2629697
Alexey Solovyev, Charlie Jacobsen, Zvonimir Rakamaric, and Ganesh
Gopalakrishnan. 2015. Rigorous Estimation of Floating-Point Roundoff Errors with Symbolic Taylor Expansions (FM’15). Springer.
U.S. General Accounting Office. 1992. Patriot Missile Defense: Software
Problem Led to System Failure at Dhahran, Saudi Arabia. (1992).
http://www.gao.gov/products/IMTEC-92-26
Ran Wang, Daming Zou, Xinrui He, Yingfei Xiong, Lu Zhang, and
Gang Huang. 2015. Detecting and Fixing Precision-Specific Operations
for Measuring Floating-Point Errors (FSE’15).
Debora Weber-Wulff. 1992. Rounding error changes Parliament
makeup. (1992). http://catless.ncl.ac.uk/Risks/13.37.html#subj4
Conference’17, July 2017, Washington, DC, USA Alex Sanchez-Stern, Pavel Panchekha, Sorin Lerner, and Zachary Tatlock
A
A.1
Appendix
trees that would not otherwise be apparent. For example,
consider the computation:
Changes to Anti-unification
To produce expressions which are more useful for program
improvement, Herbgrind uses a modified version of antiunification.
p
(1)
This expression is inaccurate for values of x on the order
of 1014 . However, the particular constant used in the multiplication is not central to this inaccuracy, it merely changes
the range of x for which the computation is inaccurate. The
same expression, with y = x · 12345.67
p
√
y+1− y
(2)
Generalizing multiple paths Herbgrind treats concrete
expressions that compute identical values in different ways
as equivalent for variable assignment, and when they are
generalized together they produce a constant node. For example, when generalizing the expressions:
5/(3 + 2)
p
(x · 12345.67) + 1 − (x · 12345.67)
6/(5 · 1)
exhibits the same erroneous behavior.
In this case, the expression is small, and so the simplified
version is only marginally more useful. However, in large
numerical computations expressions can become massive
and obscure the relationship between different subexpressions. If x · 12345.67 were instead a much larger expression,
it would be tedious and error prone to identify by hand that
both occurrences were syntactically identical.
Here, all equivalent subexpressions were syntactically
identical, so making the appropriate substitution requires
only a syntactic check. However, in other cases there may
be expressions which are non-trivially equivalent, either
because of identities (commutativity, etc.), or because of limitations on the expression inference (Section 6.1).
Because Herbgrind is a dynamic analysis, it can infer the
equivalence of internal nodes of the expression tree using
their runtime values, without relying on syntactic rules or
the equivalence of subexpressions. Herbgrind uses this information to prune equivalence trees to reveal extra equivalence
information. In the above example, Herbgrind will produce
√
√
the second form, y + 1 − y, since it detects that both occurrences of (x · 12345.67) are equivalent for the dynamic
traces observed.
Pruning expressions to an equivalence class can also obscure other equivalences which were present in the original
expression. For instance, in the computation
p
√
y + 1 − y · (y + 1) ,
(3)
classical anti-unification would produce
x/y
, because the · and + operators are distinct, so the expressions are generalized to a variable. However, in Herbgrind’s
modified version of anti-unification, the expressions are generalized to:
x/5
because both expressions on the right hand side of the division produce 5. This is a natural extension of the fact that
Herbgrind does not track integer computations and thus
must treat different integer computations with identical results as equivalent. This tweak to anti-unification often produces a symbolic expression that shows more dependencies
between variables, and exposes more constant nodes.
Replacing independent computations Herbgrind replaces
repeated subexpressions with simple variables when they do
not share variables with the rest of the symbolic expression.
This accounts for the intuition that it is the reuse of floatingpoint values (and the rounding errors inherent in them) that
causes floating-point error; without reuse, error can slowly
build but, but cannot grow rapidly and non-compositionally.
This tweak to anti-unification often produces a symbolic
expression that is shorter and simpler.
By building on anti-unification, Herbgrind’s symbolic expression satisfy a useful soundness property: the symbolic
expressions Herbgrind generates can, by substitution, produce all concrete expressions seen during program execution.
Proof of this property follow directly from [29]. Classical
anti-unification includes a completeness property, that the
symbolic expressions generated are the most specific generalization (i.e., preserve as much common structure as possible).
However, in Herbgrind we chose to eschew this property in
favor of a more subtle one: where the symbolic expressions
generated are not the most specific generalization, the extra
generality does not obscure dependencies between different
parts of the expression 12 . This extra freedom allows us to expose dependencies between internal nodes of the expression
both occurrences of (y + 1) are equivalent, and can be substituted with z = y + 1 to produce
√
√
z − y · z.
(4)
However, this second expression obscures the relationship
between the two sides of the subtraction, and prevents the
user from improving the accuracy of the expression.
To prevent such undesirable simplifications, Herbgrind
checks the following criteria of a group of equivalent variables E before using it to prune the expression:
1. The group has more than one member
2. For any other group of equivalent variables E ′ in the
expression, either all members of E ′ are within the
12 Herbgrind
also supports anti-unification with the original completeness
property, using a flag.
14
Finding Root Causes of Floating Point Error
Conference’17, July 2017, Washington, DC, USA
members of E, or they are all outside the members of
E.
√
√
In the y + 1 − y · (y + 1) example, the second condition is violated for E = {(y +1), (y +1)}, because the members
of the equivalence class E ′ = {y, y, y} appear both outside
of members of E (on the right hand of the subtraction), and
inside members of E (in the left side of both additions). Therefore, Herbgrind will not make this substitution, and will leave
the relationship between the sides of the subtraction intact.
Evaluating changes to anti-unification Herbgrind makes
two significant changes to the classical anti-unification algorithm: it generalizes multiple paths that compute the same
value to a constant value, and it replaces independent computations with variables to expose more equivalencies. In this
15
experiment, we measure how much each of these changes
affected the improvability of symbolic expressions produced
by Herbgrind. When we remove the ability to replace independent computations to expose equivalencies, the number
of benchmarks we could significantly improve drops from
18 to 16. When we remove the ability to generalize multiple
paths to a constant, and instead generalize them to a variable
as in [29], the number of benchmarks that can be significantly improved stays at 18. We believe that there are some
cases where this feature helps, but this is not reflected in our
benchmark suite, so they are likely rare.
| 6cs.PL |
arXiv:1801.08719v1 [math.RT] 26 Jan 2018
ON UNIFORM ADMISSIBILITY OF UNITARY AND SMOOTH
REPRESENTATIONS
URIYA A. FIRST AND THOMAS RÜD
Let G be a reductive algebraic group over a non-Archimedean local field F , or more precisely,
its group of F -points. The study of complex representations of G often divides into investigating
two classes of representations: unitary representations, i.e. representations on Hilbert spaces
via unitary operators, and smooth representations, i.e. representation on an abstract C-vector
space such that the stabilizers of all vectors are open. Historically, the unitary representations
served as the original object of interest, and the smooth representations were introduced in as
a tool to study them. Interplay between the two types of representations is abundant and is
now part of the standard theory.
Building on the work of Harish-Chandra [7], Bernstein [2] proved that the irreducible unitary
representations of G are uniformly admissible. That is to say, for any compact open subgroup
K ≤ G, there exists an integer n = n(K, G) such that the dimension of the K-fixed vectors
in any irreducible unitary representation of G is at most n. Here, “irreducible” should be
interpreted in the category of unitary representations, namely, the irreducible unitary representations are those representations admitting no proper closed subrepresentations other than 0.
Bernstein later showed that the irreducible smooth representations of G are uniformaly admissible as well; here, “irreducible” means irreducible as a complex G-module. A standard
source for the latter result is Section II.2.2 in Bernstein’s lecture notes [4]. Both results are
fundamental in the theory of reductive p-adic Lie groups.
It was noted by Bernstein and Zelevinsky in [3] that by combining [3, §2.5, Prp. 2.12] with a
modification of results by Godement [6, Lem. 4], one can formally derive uniform admissiblity of
the irreducible unitary representations of G from uniform admissiblity of its irreducible smooth
representations. The purpose of this note is to extend this to arbitrary ℓ-groups, provide a
converse, and give self-contained proofs in the process. We show:
Theorem A. Let G be a totally disconnected locally compact group such that |G/K| < 2ℵ0 for
some compact open subgroup K ≤ G. Then the following conditions are equivalent:
(a) The irreducible unitary representations of G are uniformly admissible.
(b) The irreducible smooth representations of G are uniformly admissible.
In fact, (b) implies (a) without the assumption |G/K| < 2ℵ0 .
The assumption |G/K| < 2ℵ0 is necessary in general; see Remark 2.5. The proof makes use
of polynomial identities, a technique which goes back at least as far as Kaplansky [9].
Theorem A was applied in [12, §4.2] to show that irreducible unitary representations of
quasi-reductive groups over a non-Archimedean local field are uniformly admissible.
In addition to Theorem A, we show, separately, that uniform admissibility of the irreducible
smooth representations is invariant under passing to a subgroup or an overgroup of finite index.
Theorem B. Let G be a totally disconnected locally compact group and let H ≤ G be an open
subgroup of finite index. Then the following conditions are equivalent:
(a) The irreducible smooth representations of G are uniformly admissible.
(b) The irreducible smooth representations of H are uniformly admissible.
1
2
URIYA A. FIRST AND THOMAS RÜD
This statement also holds upon replacing “uniformly admissible” with “admissible”.
We do not know whether Theorem B holds under the milder assumption that H is a group
such that there exists a continuous group homomorphism H ! G with finite kernel and cokernel.
The paper is organized in three sections, the first of which is preliminary, whereas the second
and the third are concerned with proving Theorems A and B, respectively.
We thank Maarten Solleveld for encouraging us to publish this note.
1. Preliminaries
We begin with recalling some necessary facts and setting notation for the sequel. Throughout,
all vector spaces are over C, and an algebra means a unital associative C-algebra.
1.1. ℓ-Groups. Throughout this paper, G denotes a locally compact totally disconnected
group, or an ℓ-group for short, and µ is a fixed left-invariant measure on G. The modular
character of G is denoted δ : G ! R>0 ; it is determined by the identity µ(Sg) = µ(S)δ(g),
holding for any compact S ⊆ G and g ∈ G.
We write K ≤c.o. G to denote that K is a compact open subgroup of G. It is well known
that such subgroups form a basis of neighborhoods to the identity element 1G ; see, for instance,
[10, §2.3].
As usual, a unitary representation of G consists of a complex Hilbert space V endowed with
a continuous G-action G × V ! V such that hgu, gvi = hu, vi for all u, v ∈ V , g ∈ G. We
say that V is topologically irreducible, or just irreducible, if V has no proper nonzero closed
G-submodules.
A smooth representation of G is an abstract complex vector space V together with
S a G-action
G×V ! V such that the G-stabilizer of every v ∈ V is open, or equivalently, V = K≤c.o. G V K ,
where V K := {v ∈ V : gv = v for all g ∈ K}. We say that V is algebraically irreducible, or
just irreducible, if V is contains no proper nonzero G-modules.
1.2. Unital ∗-Algebras. By a (unital) ∗-algebra, we mean a unital associative C-algebra A
together with an involution ∗ : A ! A, i.e., an additive map satisfying a∗∗ = a, (ab)∗ = b∗ a∗
and (αa)∗ = αa∗ for all a ∈ A, α ∈ C.
A unitary representation of a ∗-algebra A is a Hilbert space V endowed with an A-module
structure such that hau, vi = hu, a∗ vi for all a ∈ A, u, v ∈ V . A unitary representation is
topologically irreducible, or just irreducible, if it has no proper nonzero closed A-submodules.
We shall need the following versions of Schur’s lemma.
Proposition 1.1. Let A be a ∗-algebra with dimC A < 2ℵ0 and let M be a simple left A-module.
Then EndA (M ) = C.
Proof. Fix some 0 6= m ∈ M . Then Am = M and hence any element f ∈ EndA (M ) is uniquely
determined by f (m). As a result, dimC EndA (M ) ≤ dimC M ≤ dimC A < 2ℵ0 . Thus, EndA (M )
is a division algebra over C of C-dimension < |C|. It is well known that such algebras are
algebraic over C, and hence coincide with C. We conclude that EndA (M ) = C.
Proposition 1.2. Let A be a ∗-algebra and let V be an irreducible unitary representation.
Then all continuous A-endomorphisms of V are scalar.
Proof. See, for instance, [11, Thm. 9.6.1]
ON UNIFORM ADMISSIBILITY
3
1.3. The Relative Hecke Algebra. The most important example of a ∗-algebra that we
shall consider is the relative Hecke algebra H(G, K) associated to a compact open subgroup
K ≤c.o. G. Recall that the underlying vector space of H(G, K) consists of the bi-K-invariant
compactly supported functions f : G ! C and that its multiplication is given by convolution
Z
f (x)g(x−1 y) dµ .
(f ⋆ g)(y) :=
x∈G
The unity of H(G, K) is eK := µ(K)−1 1K , where 1K : G ! C denotes the characteristic
function of K. We make H(G, K) into a ∗-algebra by setting
f ∗ (g) = δ(g)f (g −1 )
for all f ∈ H(G, K) and g ∈ G (recall that δ is the modular character of G).
Given a unitary or smooth representation V of G, the space V K carries a left H(G, K)module structure given by
Z
f (g) · gv dµ
∀ f ∈ H(G, K), v ∈ V K
f · v :=
g∈G
P
−1
Note that in
P the smooth case, if we write f = i αi µ(K) 1gi K , then the integral is just the
finite sum i αi gi v.
When V is a unitary representation, we further have hf u, vi = hu, f ∗ vi for all u, v ∈
V and f ∈ H(G, K), and so V K becomes a unitary representation of H(G, K). We remark
that the equalityR hf u, vi = hu, f ∗ vi follows easily from the definitions and the identity
R
f
(x−1 )δ(x)dµ = x∈G f (x)dµ, which holds whenever the integrals exist.
x∈G
We shall need the following two easy propositions.
Proposition 1.3. Let V be an irreducible smooth representation of G and let K ≤c.o. G. Then
either V K = 0, or V K is a simple H(G, K)-module.
K
Proof. We need to show that H(G, K) · v = V K for all 0 6= v ∈ V K . Let
P u ∈ V . Since
V is irreducible,
there are g1 , . . . , gn ∈ G and α1 , . . . , αn ∈ C such that i αi gi v = u. Let
P
f = eK ⋆ ( i αi µ(K)−1 1gi K ) ∈ H(G, K). Then
X
X
αi gi v = eK u = u
αi µ(K)−1 1gi K ) · v = eK ·
f · v = eK · (
i
i
and we conclude that u ∈ H(G, K) · v.
Proposition 1.4. Let V be an irreducible unitary representation of G and let K ≤c.o. G. Then
either V K = 0, or V K is an irreducible unitary representation of H(G, K).
Proof. Similarly to the proof of Proposition 1.3, we need to show that V K is the closure of
H(G, K) · v for all 0 6= v ∈ V K . Since V is topologically irreducible, Span {gv | g ∈ G} is
dense in V . For all g ∈ G, we have µ−1 (K)1gK · v = gv and thus, Span {1gK · v | g ∈ G} is
dense in V . This in turn implies that Span {eK · 1gK · v | g ∈ G} is dense in eK V = V K . Since
eK ⋆ 1gK ∈ H(G, K), it follows that H(G, K) · v is dense in V .
1.4. Miscellaneous Results. We finish this section with recalling the Amistur–Levitzky theorem and two density theorems due to Jacobson and von Neumann. These will be needed in
proving Theorem A.
To state the Amitsur–Leviski theorem, let Zhx1 , . . . , xn i denote the ring of polynomials in n
non-commuting variables over Z. The n-th standard polynomial is define by
X
Sn (x1 , . . . , xn ) =
sgn(σ)xσ(1) . . . xσ(n)
σ
4
URIYA A. FIRST AND THOMAS RÜD
where σ runs over all permutations on {1, . . . , n}. Recall that a polynomial f ∈ Zhx1 , . . . , xn i
is called an identity of a ring R if f vanishes on all inputs from R.
Theorem 1.5 (Amitsur–Levitzki [1]). Let V be a C-vector space and let n ∈ N. The polynomial
Sn (x1 , . . . , xn ) is a polynomial identity of EndC (V ) if and only if n ≥ 2 dim V .
Theorem 1.6 (Jacobson density theorem). Let R be a ring, let M be a simple left R-module,
and let D be the R-endomorphism ring of M . Give M the discrete topology and EndD (M ) the
topology induced by the product topology on M M (the point-wise convergence topology). Then
the image of R in EndD (M ) under the map sending r to (m 7! rm) is dense in EndD (M ).
Proof. See, for instance, [8, Theorem 13.14].
Given a Hilbert space V , let B(V ) denote the algebra of bounded linear automorphisms of
V . For a subset S ⊆ B(V ), we let S ′ denote the commutant of S, namely, the set of elements
of B(V ) that commute with all elements in S.
Theorem 1.7 (Von Neumann density theorem). Let V be a Hilbert space and let A ⊂ B(V ) be
a unital subalgebra closed under taking adjoints. Then A is dense in A′′ in the strong operator
topology (and hence also in the weak operator topology).
Proof. See, for instance, [11, Theorem 9.3.3].
2. Proof of Theorem A
We begin with several results about C-algebras and ∗-algebras. The Jacobson radical of a
ring R is denoted Jac(R).
Proposition 2.1. Let A be a unital C-algebra. Consider the conditions:
(a) Every simple A-module has C-dimension at most n.
(b) A/ Jac(A) has S2n as a polynomial identity.
Then (a) implies (b), and the converse holds provided dimC A < 2ℵ0 .
Proof. Suppose (a) holds. Let S be a set of representative for the isomorphism classes of
simple left A-modules. By the definition of Jac(A), the map
Y
A/ Jac(A) !
EndC (M )
M∈S
sending a + Jac(A) ∈ A/ Jac(A) to (x 7! ax) ∈ EndC (M ) in the M -component, is injective.
Condition (a) and Theorem 1.5 imply that S2n is a polynomial identity of EndC (M ) for all
S ∈ S , and hence also of A/ Jac(A).
Suppose now that (b) holds and dimC A < 2ℵ0 . Let M be a simple A-module. Consider
M as a discrete topological space and make EndC (M ) into a topological ring by giving it the
topology induced from the product topology on M M . By Proposition 1.1, EndA (M ) = C, and
hence, by Theorem 1.6, the image of A in EndC (M ) is dense. Since the map A ! EndC (M )
factors through A/ Jac(A), condition (b) implies that S2n vanishes on all inputs from dense
subset of EndC (M ). Since the topology on EndC (M ) is Hausdorff, it follows that S2n is a
polynomial identity of EndC (M ). Thus, dim M ≤ n by Theorem 1.5.
Lemma 2.2. Let A be a unitary ∗-algebra and let V be a unitary representation of A. Then
Jac(A) · V = 0.
Proof. Let a ∈ Jac(A). Suppose first that a∗ = a. Then the operator Ta : V ! V given
by Ta (v) = av is self-adjoint. Since a ∈ Jac(A), the element λ − a is invertible in A for all
ON UNIFORM ADMISSIBILITY
5
0 6= λ ∈ C. As a result, Spec(Ta ) = {0}, and since T is self-adjoint, this forces T = 0 and
a · V = 0.
∗
∗
The general case follows by noting that any a ∈ Jac(A) satisfies a = a+a
+ i( a−a
) with
2
2i
a−a∗
a+a∗
and 2i being elements of Jac(A) fixed by ∗.
2
Proposition 2.3. Let A be a unital ∗-algebra. Consider the conditions:
(a′ ) Every irreducible unitary representation of A has C-dimension at most n.
(b) A/ Jac(A) has S2n as a polynomial identity.
Then (b) implies (a′ ), and the converse holds provided A/ Jac(A) admits a faithful unitary
representation.
Proof. Suppose (b) holds and let V be an irreducible unitary representation of A. We may
replace A with its image in B(V ) and hence assume that Jac(A) = 0, by Lemma 2.2. Now, by
Proposition 1.2, the commutant A′ is C, and hence A′′ = B(V ). By Theorem 1.7, A is dense in
B(V ) in the strong operator topology. Condition (b) therefore implies that S2n is a polynomial
identity of B(V ). It is easy to check that this is impossible if V is infinite-dimensional, whereas
in the finite-dimensional case, Theorem 1.5 implies that dim V ≤ n, as required.
Suppose now that (a′ ) holds and A/ Jac(A) admits a faithful unitary representation V . By
Lemma 2.2, we may replace A with its image in B(V ) and assume that Jac(A) = 0 hereafter.
Let 0 6= a ∈ A. The the operator v 7! av : V ! V is nonzero, and hence, by [5, Corollary 2.30],
there exists an irreducible unitary representation Va of A such that the image of a in B(Va ) is
nonzero. Consequently, the ring homomorphism
Y
A ֒−!
B(Va )
a6=A
is injective. By condition (a′ ) and Theorem 1.5, S2n is a polynomial identity of each of the
factors B(Va ), and a fortiori of A as well.
We finally prove Theorem A by establishing the following slightly stronger version. Notice
that if |G/K| < 2ℵ0 holds for some K ≤c.o. G, then it holds for all K ≤c.o. G.
Theorem 2.4. Let G be an ℓ-group, let K be a compact open subgroup with |G/K| < 2ℵ0 , and
let n ∈ N. Then the following conditions are equivalent:
(a) All irreducible unitary representations V of G satisfy dim V K ≤ n.
(b) All irreducible smooth representations V of G satisfy dim V K ≤ n.
In fact, the implication (b)=⇒(a) holds without assuming |G/K| < 2ℵ0 .
Proof. Let A = H(G, K). By Propositions 1.3 and 1.4, it is enough to show that every simple
A-module is of dimension ≤ n if and only if every irreducible unitary representation of A is of
dimension ≤ n. This would follow from Propositions 2.1 and 2.3 if we check that dimC A < 2ℵ0
and A admits a faithful unitary representation. The former condition holds since |G/K| < 2ℵ0 ,
whereas for the second condition one can consider the left action of A = H(G, K) on L2 (K\G)
by convolution. Indeed, if a ⋆ f = 0 for all f ∈ L2 (K\G), then a = a ⋆ eK = 0.
Remark 2.5. (i) The assumption that |G/K| < 2ℵ0 in Theorems A and 2.4 is necessary in
general. For instance, consider the function field C(t) and let G be the multiplicative group
C(t)× with the discrete topology. Since G is commutative and locally compact, any unitary
representation of G is 1-dimensional. However, C(t) with the G-action induced by the product
in C(t) is an infinite-dimensional irreducible smooth representation of G.
(ii) The assumption that dimC A < 2ℵ0 in Proposition 2.1 cannot be dropped in general. For
example, take A = C(t) and note that S2 (x1 , x2 ) = x1 x2 − x2 x1 is a polynomial identity of A,
while A itself is a simple A-module of infinite C-dimension.
6
URIYA A. FIRST AND THOMAS RÜD
(iii) The assumption that A/ Jac(A) admits a faithful unitary representation in Proposition 2.3 is necessary in general. A counterexample can be constructed as follows: Let ∗ denote
the unique involution of C(t) extending the complex conjugation and fixing t. Let A = M2 (C(t))
∗
and let ∗ : A ! A be the involution given by (fij ) 7! (fji
). Then S2 (x1 , x2 ) = x1 x2 − x2 x1
is not a polynomial identity of A, but all irreducible unitary representations of A have dimension ≤ 1. In fact, the latter holds vacuously, since A has no nonzero unitary representations.
To see this, note that if V is a Hilbert space and A ! B(V ) is a ∗-homomorphism, then the
image of [ 0t 0t ] in B(V ) is a self-adjoint operator with an empty spectrum, forcing V = 0.
3. Proof of Theorem B
Theorem B follows readily from the following theorem.
Theorem 3.1. Let H be an open subgroup of G of finite index and let K be a compact open
subgroup of H. Then there is n ∈ N and a compact open subgroup L ≤ K such that:
(i) For any irreducible smooth representation V of H, there are
P irreducible smooth representations U1 , . . . , Um (m ≤ n) of G such that dim V K ≤ i dim UiL .
(ii) For any irreducible smooth representation U of G, there are
P irreducible smooth representations V1 , . . . , Vm (m ≤ n) of H such that dim U K ≤ i dim ViL .
Proof. Since [G : H] is finite, H has only finitely many conjugates. Their intersection, N , is
an open normal subgroup of finite index. We may replace K with K ∩ N . The theorem would
now follow if we establish it for the pairs (G, N ) and (H, N ). It is therefore enough to prove
the theorem when H is normal in G. Under this extra assumption, we shall see that L = K
and n := [G : H] will suffice.
(i) Let g1 , . . . , gn be representatives for the H-cosets in G. Let Vi be the irreducible smooth
representation of H obtained by giving V the H-action (h, v) 7! (gi hgi−1 )v (note that gi hgi−1 ∈
H since H ⊳ G).
Consider the (non-normalized) induction
IndG
H (V ) := {f : G ! V | f (hg) = h · f (g) for all h ∈ H, g ∈ G} .
′
As usual, the group G acts on IndG
H (V ) by translations on the right. Let Vi denote the functions
G
′
in IndH (V ) supported on gi H = Hgi . It is easy to see that Vi is an H-subrepresention of
Ln
G
′
′
IndG
H (V ), that IndH (V ) =
i=1 Vi and that f 7! f (gi ) : Vi ! Vi defines an isomorphism of
H-modules. Thus, the length of IndG
H (V ) as a complex H-module is n, and so the length of
IndG
H (V ) as a complex G-module is at most n.
Let U1 , . . . , Um denote the composition factors of IndG
H (V ), regarded as a G-module. Then
P
K
K
(V
)
≥
dim
V
,
so
we have proved (i).
m ≤ n, and i dim UiK = dim IndG
H
(ii) Let U be an irreducible smooth representation of G, and let g1 , . . . , gn be as in (i).
Since U is a simple G-module, it is finitely generated. As [G : H] < ∞, it follows that U
is also finitely generated as a complex H-module. Thus, there exists a maximal complex Hsubmodule M0 ≤ U . Since H is normal in G, the space gM0 is a maximal H-submodule for
all g ∈ G. Let M = {gM0 | g ∈ G}. Then M consists
[G : H] = n elements, call
Lm of at mostL
them M1 , . . . , Mm . Write Vi = U/Mi and set L
V = i=1 U/Mi = LM∈M U/M . Then action
of H on V extends
L to a G-action by setting g · M∈M (vM + M ) = M∈M (gvg−1 M + M ) and
the map u 7! M∈M (u + M ) : U ! V defines a nontrivial G-homomorphism, which must be
P
K
injective since V is irreducible. It follows that dim U K ≤ dim V K = m
i=1 dim Vi , and the
proof is complete.
Remark 3.2. The proof of part (ii) also shows that any irreducible complex G-module has
finite length as a complex H-module.
ON UNIFORM ADMISSIBILITY
7
References
[1] J. Levitzki A. S. Amitsur. Minimal identities for algebras. Proceedings of the American Mathematical
Society, 1(4):449–463, 1950.
[2] I. N. Bernšteı̆n. All reductive p-adic groups are of type I. Funkcional. Anal. i Priložen., 8(2):3–6, 1974.
[3] I. N. Bernstein and A. V. Zelevinskii. Representations of the group GL(n, F ), where F is a local nonArchimedean field. Uspehi Mat. Nauk, 31(3(189)):5–70, 1976.
[4] J. Bernstein. Draft of: Reperesentations of p-adic groups. 1992. Notes taken by E. Rumelhart. The notes
can be downloaded from the author’s webpage.
[5] Uriya A. First. The ramanujan property for simplicial complexes. 2016. Preprint. Available at
https://arxiv.org/abs/1605.02664.
[6] Roger Godement. A theory of spherical functions. I. Trans. Amer. Math. Soc., 73:496–556, 1952.
[7] Harish-Chandra. Harmonic analysis on reductive p-adic groups. Lecture Notes in Mathematics, Vol. 162.
Springer-Verlag, Berlin-New York, 1970. Notes by G. van Dijk.
[8] I. Martin Isaacs. Algebra: a graduate course, volume 100 of Graduate Studies in Mathematics. American
Mathematical Society, Providence, RI, 2009. Reprint of the 1994 original.
[9] Irving Kaplansky. Groups with representations of bounded degree. Canadian J. Math., 1:105–112, 1949.
[10] Deane Montgomery and Leo Zippin. Topological transformation groups. Interscience Publishers, New YorkLondon, 1955.
[11] Theodore W. Palmer. Banach algebras and the general theory of ∗-algebras. Vol. 2, volume 79 of Encyclopedia of Mathematics and its Applications. Cambridge University Press, Cambridge, 2001. ∗-algebras.
[12] Maarten Solleveld. Pseudo-reductive and quasi-reductive groups over non-archimedean local fields. 2017.
Preprint. Available at https://arxiv.org/abs/1706.07644.
E-mail address: uriya.first@gmail.com
Department of Mathematics, University of Haifa, Mount Carmel, Haifa, Israel 31905
E-mail address: thomas@math.ubc.ca
Department of Mathematics, University of British Columbia, BC, Canada V6T 1Z2
| 4math.GR |
The ’Core’ of Symmetric Homogeneous Polynomial
Inequalities of Degree Four of Three Real Variables
Mariyan Milev 1 , Nedelcho Milev
2
arXiv:1604.00993v1 [math.CA] 26 Mar 2016
1
Department of Mathematics and Physics, UFT-Plovdiv, bul. Maritza 26, 4000 Plovdiv,
Bulgaria, tel. +359 32 603744 email: marianmilev2002@gmail.com
2
Department of Mathematical Analysis, Plovdiv University ’Paisii Hilendarski’, Tzar
Asen 24, 4000 Plovdiv, Bulgaria, phone: +359-32 519542, email: milev0@abv.bg
Abstract. In this paper we explore inequalities between symmetric homogeneous polynomials of degree four of three real variables and three nonnegative real variables. The main
theorems describe the cases in which the smallest possible coefficient is not expressed by
the other coefficients. The problem is resolved by introducing a parametric representation.
Key words: symmetric, homogeneous, polynomial, inequalities, degree three and four.
Mathematics Subject Classification: 26D05.
Introduction
In recent years, inequalities between symmetric homogeneous polynomials are studied
in several articles [1]-[10]. We will explore the inequality
x4 + y 4 + z 4 + a(xy 3 + yz 3 + zx3 + x3 y + y 3z + z 3 x) + b(x2 y 2 + y 2 z 2 + z 2 x2 )+
+ c xyz(x + y + z) ≥ 0,
where a, b, c are real constants, and x, y, z are real variables or nonnegative real variables.
The proof of the main theorem 3 in this paper is structured by setting of auxiliary functions
(2), lemma 4 and the representation (7). For brevity we will denote the main symmetric
homogeneous polynomials of degree four
w4 = x4 + y 4 + z 4 ,
w3 = x3 y + y 3z + z 3 x + xy 3 + yz 3 + zx3 , w2 = x2 y 2 + y 2z 2 + z 2 x2
w1 = xyz(x + y + z) and the inequality is written in the form
f (x, y, z) = w4 + aw3 + bw2 + cw1 ≥ 0.
In Section I we find b min such that the inequality f (x, y, z) = w4 + aw3 + bw2 + cw1 ≥ 0
holds for arbitrary x, y, z ∈ R if and only if when b ≥ b min .
1
In the particular case a = − the upper inequality holds if and only if when
2
c 9
(Theorem 2).
b min = max − c ; −
2 8
When c ≤ −a2 − 2a (Theorem 1) b min = −2a − c − 1. This statement follows also
from [4], Theorem 2.1.
2
When c > −a2 − 2a the situation radically changes. In some particular cases b min
2
could beexpressed
in radicals, for example when a = 2 and c = 4 > −a − 2a we find
√
√
5 5−7
1+ 5
. When b = b min the equality holds when x = y = 1 and z = −
b min =
2
2
and of course
for any
permutation. When a = −1 and c = 2 > −a2 − 2a we find
√
√
2 + 5 3 100 − 10 3 10
b =
= 0.305299773.... The equality holds when x = y = 1 and
12
√
√
3
3
100 − 2 10 − 2
= −0.27788....
z=
6
We overcome this problem by introducing the rational parameterization (2) (see Theorem 3).
We note that this parametrization is in line with [5], Theorem 2.2: ’The inequality
f (x, y, z) = w4 + aw3 + bw2 + cw1 ≥ 0 holds if and only if when f (x, 1, 1) ≥ 0 for every
real x.’
In Section II we will explore the inequality
f (x, y, z) = w4 + aw3 + bw2 + cw1 ≥ 0 ,
where x, y, z are nonnegative real variables.
First we find the necessarily conditions for the coefficient b (Lemma 5) and then we
search cmin such that the previous inequality holds if and only if when c ≥ cmin . We note
that w1 is the smallest symmetric polynomial 2w4 ≥ w3 ≥ 2w2 ≥ 2w1 ≥ 0.
When a ≥ −1 ⇒ b ≥ −2(a + 1) and cmin = −2a − b − 1. This statement follows
also from [4], Theorem 2.6. According to this theorem the inequality f (x, y, z) ≥ 0
holds also when a < −1, but when c > −2a − b − 1 the condition b ≥ a2 − 1 is not
necessary. When a < −1 and b < a2 − 1, only in some particular cases cmin could be
expressedh in radicals. For example,
when a i= −4 and b = 14 (−2(a + 1) < b < a2 − 1)
√
√
√ p
cmin = 4 − 9 + 8 2 − 2 3 − 2
3 − 2 = −6.72076... When c = cmin the equality
p
√
√
holds when x = y = 1 and z = 2 + 6 − 2 2 = 3.1951... and of course for any
permutation. Another example, when a = −6, b = 31
a2
√ p
√
√
2
40 − 2 15 = −18.131094...
+2 < b < a −1 .
cmin = −58 + 20 15 + 5 − 3 15
p
√
√ 4
1 + 15 + 40 − 2 15
The equality holds when x = y = 1 and z =
= 5.276...
2
We again solve the problem by using the rational parametrization (2) (see Theorem
7).
We note that this do not contradict to [2] and [6], Theorem 2.1: ’The inequality
f (x, y, z) = w4 + aw3 + bw2 + cw1 ≥ 0 holds if and only if when f (x, 1, 0) ≥ 0 and
f (x, 1, 1) ≥ 0 for every x ≥ 0.’
Remark. All identities are verified via the Maplesoft platform.
Main Results
Section I. We will explore the inequality
f (x, y, z) = w4 + aw3 + bw2 + cw1 ≥ 0 ,
where a, b, c are real constants and x, y, z are real variables.
(1)
3
Theorem 1. One necessary condition to be true the inequality (1) is b ≥ −2a − c − 1.
When c ≤ −a2 − 2a this condition is also a sufficient condition (This statement follows
also from [4], Theorem 2.1).
1
Theorem 2. When a = − the inequality (1) holds if and only if when
2
c 9
b ≥ max − c, −
.
2 8
1
When a 6= − and c ≥ −a2 − 2a we introduce the rational functions:
2
c(t) =
2t4 + 2(a + 1)t3 + (2a + 4)t2 + (3a + 4)t − a
,
2t + 1
b(t) =
t4 + (a + 4)t3 + (5a + 4)t2 + 4(a + 1)t + 2a + 2
−(2t + 1)
(2)
and we will explore the different cases about the parameter a when the rational function
c(t) always takes values in the interval [−a2 − 2a ; +∞).
1
1
< a ≤ 4 and t ∈ [−a − 1; − ) the continuous function c(t) is monotone in2
2
1
2
creasing and −a − 2a = c(−a − 1) ≤ c(t) < c(− − 0) = +∞.
2
1
1
When −2 ≤ a < − and t ∈ (− ; −a − 1] the continuous function c(t) is monotone
2
2
1
decreasing and +∞ = c(− + 0) > c(t) ≥ c(−a − 1) = −a2 − 2a.
2
When −
Let us set
p
(a + 2)(a − 4)
and t2 =
.
t1 =
2
1
When a > 4 we obtain two intervals t ∈ [−a − 1 ; t1 ] ∪ [ t2 ; − ,
2
1
−a − 1 < t1 < −2 < t2 < − . The function c(t) is monotone increasing and
2
−a −
p
(a + 2)(a − 4)
2
− a2 − 2a = c(−a − 1) < c(t1 ) = c(t2 ) =
−a +
(3)
a2
1
+ 2 < c(− − 0) = +∞ .
2
2
1
When a < −2 we obtain again two intervals: t ∈ (− ; t1 ] ∪ [−a − 1 ; t2 ]. The function
2
1
c(t) is monotone decreasing in (− ; t1 ] and monotone increasing in [−a − 1 ; t2 ]:
2
a2
1
+ 2 > c(−a − 1) = −a2 − 2a .
+ ∞ = c(− + 0) > c(t1 ) = c(t2 ) =
2
2
1
Theorem 3. When a 6= − and c ≥ −a2 − 2a in the following four cases:
2
1
1
(3.1.) − < a ≤ 4 and t ∈ [−a − 1 ; − );
2
2
1
1
(3.2.) −2 ≤ a < − and t ∈ (− ; −a − 1];
2
2
4
1
a > 4 and t ∈ [−a − 1 ; t1 ] ∪ [t2 ; − ;
2
1
(3.4.)
a < −2 and t ∈ − ; t1 ] ∪ [−a − 1 ; t2 ]
2
the inequality f (x, y, z) = w4 + aw3 + bw2 + cw1 ≥ 0 holds if and only if when b ≥ b(t).
When b = b(t) we have f (t, 1, 1) = f (1, t, 1) = f (1, 1, t) = 0.
(3.3.)
For the proof of this theorem 3 and theorem 7 below we will use the following known
statement.
Lemma 4. For arbitrary real numbers X, Y and Z the inequality
g(X, Y, Z) = X 2 + Y 2 + Z 2 + k(XY + Y Z + ZX) ≥ 0
(4)
holds if and only if when −1 ≤ k ≤ 2.
Section II. We will explore the inequality
f (x, y, z) = w4 + aw3 + bw2 + cw1 ≥ 0 ,
(5)
where a, b, c are real constants and x, y, z are nonnegative real variables.
Lemma 5. To be true the inequality (5) it is necessary:
(
1 2
a +2
when a ≤ −4;
b≥
and c ≥ −2a − b − 1.
4
−2(a + 1) when a ≥ −4
Theorem 6. ([4], Theorem 2.6). When a ≥ −1 the inequality (5) holds if and only if
when b ≥ −2(a + 1) and c ≥ −2a − b − 1.
Remark. According to ([4], Theorem 2.6) the inequality (5) is true also when a < −1,
b ≥ a2 − 1 and c ≥ −2a − b − 1, but when c > −2a − b − 1 the condition b ≥ a2 − 1 is not
necessary.
When a < −1 and b ≤ a2 − 1 we will describe the different cases about the parameter
a again with the rational function b(t) and c(t), defined by (2).
When −2 ≤ a < −1 and t ∈ [0; −a − 1] the rational function b(t) is monotone increasing and takes values in the intervala [−2(a + 1) ; a2 − 1], i.e.
−2(a + 1) = b(0) ≤ b(t) ≤ b(−a − 1) = a2 − 1.
When −4 < a < −2 there are two intervals for t.
When t ∈ [0 ; t1 ] the function b(t) is monotone increasing and takes values in the
a2
a2
+ 2], i.e. −2(a + 1) = b(0) ≤ b(t) ≤ b(t1 ) =
+ 2, and when
interval [−2(a + 1) ;
4
4
t ∈ [−a − 1 ; t2 ] the function b(t) is monotone decreasing and takes values in the interval
a2
a2
[ + 2 ; a2 − 1], i.e.
+ 2 = b(t2 ) ≤ b(t) ≤ b(−a − 1) = a2 − 1, where t1 and t2 are
4
4
defined in formulas (3).
When a ≤ −4 and t ∈ [−a − 1; t2 ] the function b(t) is monotone decreasing and takes
a2
a2
+ 2 = b(t2 ) ≤ b(t) ≤ b(−a − 1) = a2 − 1.
values in the interval [ + 2 ; a2 − 1], i.e.
4
4
5
Theorem 7. In the three cases:
(7.1.) −2 ≤ a < −1 and t ∈ [0 ; −a − 1] ;
(7.2.) −4 < a < −2 and t ∈ [0 ; t1 ] ∪ [−a − 1 ; t2 ];
(7.3.)
a ≤ −4 and t ∈ [−a − 1 ; t2 ]
the inequality f (x, y, z) = w4 + aw3 + b(t)w2 + cw1 ≥ 0 holds if and only if when c ≥ c(t).
When c = c(t) we have f (t, 1, 1) = f (1, t, 1) = f (1, 1, t) = 0.
Proves of Theorems
Section I (real variables)
Remark. In the proves we have in mind that the following basic symmetric homogeneous polinomial inequalites of degree four of three real variables are true:
2w4 ≥ w3 , w4 ≥ w2 ≥ 0 , w2 ≥ w1 and w2 + 2w1 ≥ 0.
Proof of Theorem 1.
The necessity follows from 0 ≤ f (1, 1, 1) = 3 + 6a + 3b + 3c, i.e. b ≥ −2a − c − 1.
The sufficiency follows from the relation
w4 + aw3 + (a2 − 1)w2 − (a2 + 2a)w1 = u2 + v 2 − uv ≥ 0,
(6)
where u = (x − y)(x + y + az) and v = (x − z)(x + z + ay).
When c ≤ −a2 − 2a ⇒ b ≥ −2a − c − 1 ≥ a2 − 1 and w4 + aw3 + bw2 + cw1 =
= w4 +aw3 +(a2 −1)w2 −(a2 +2a)w1 +(2a+b+c+1)w2−(c+a2 +2a)(w2 −w1 ) ≥ 0.
Remark. When a = −1, c = −a2 − 2a = 1 and b = −2a − c − 1 = 0 we obtain the
famous inequality of I. Schur [11], w4 + w1 ≥ w3 .
Proof Theorem 2.
1
3
3
and c ≤ −a2 − 2a =
⇒ b ≥ −2a − c − 1 = −c ≥ − the basic
2
4
4
3
3
1
inequality (6) is w4 − w3 − w2 + w1 ≥ 0 and the statement follows from Theorem 1.
2
4
4
c 9
From 0 ≤ f (2, 2, −1) = 24b − 12c + 27 ⇒ b ≥ − .
2 8
c 9
3
and b ≥ − we have
When c ≥
4
2 8
c 9
1
3
3 1 3
1
w4 − w3 +bw2 +cw1 = w4 − w3 − w2 + w1 + c− (w2 +2w1 )+ b− + w2 ≥ 0.
2
2
4
4
2
4
2 8
When a = −
Before the proof of Theorem 3 we will prove Lemma 4.
Proof of Lemma 4.
From g(1, 1, 1) = 3 + 3k ≥ 0 and from g(1, −1, 0) = 2 − k ≥ 0 it follows that −1 ≤ k ≤ 2.
2 2 − k
k+1
And g(X, Y, Z) =
X +Y +Z +
X 2 + Y 2 + Z 2 − XY − Y Z − ZX ≥ 0.
3
3
6
Proof of Theorem 3. Necessity.
f (x, y, z) = w4 + aw3 + bw2 + c(t)w1 = w4 + aw3 + b(t)w2 + c(t)w1 + b − b(t) w2 .
From 0 ≤ f (t, 1, 1) = 0 + b − b(t) (2t2 + 1) follows that b ≥ b(t).
Sufficiency follows from the identity:
w4 + aw3 + b(t)w2 + c(t)w1 = X 2 + Y 2 + Z 2 + k XY + Y Z + ZX , where
X = x2 + pxy + pxz + qyz, Y = y 2 + pyz + pyx + qzx, Z = z 2 + pzx + pzy + qxy, (7)
t2 + t + 1
t2 + 2t
2t2 + 2t(a + 1) + a + 2
, q=
and k =
2t + 1
2t + 1
t−1
When t = 1 we obtain that p = −1, q = 1 and the indentity do not depend on k because
XY + Y Z + ZX = 0.
Having in mind Lemma 4 it remains to check that −1 ≤ k ≤ 2 for the different cases.
1
1
(3.1) When − < a ≤ 4 and t ∈ [−a − 1 ; − ) ⇒ 2t + 1 < 0, t − 1 < 0, t + a + 1 ≥ 0 ⇒
2
2
p=−
k+1=
(2t + 1)(t + a + 1)
(2t + a)2 + (4 − a)(a + 2)
≥ 0 and k − 2 =
< 0.
t−1
2(t − 1)
1
(3.2) When −2 < a < −1 and t ∈ (− ; −a − 1] ⇒ 2t + 1 > 0, t − 1 ≤ −a − 2 < 0,
2
t+a+1<0 ⇒
k+1=
(2t + a)2 + (4 − a)(a + 2)
(2t + 1)(t + a + 1)
≥ 0 and k − 2 =
< 0.
t−1
2(t − 1)
1
When a = −2 and t ∈ (− ; −a − 1) ⇒ k = 2t ∈ (−1; 2).
2
Remark. The case t = 1 is possible only if a = −2 ⇒ b(1) = 3, c(1) = 0 and
2(w4 − 2w3 + 3w2 ) = (y − z)4 + (x − z)4 + (x − y)4 = 2(x2 + y 2 + z 2 − xy − yz − xz)2 .
1
(3.3) When a > 4 and t ∈ [−a − 1; t1 ] ∪ [t2 ; −
⇒ 2t + 1 < 0, t − 1 < 0, t + a + 1 ≥ 0
2
(2t + 1)(t + a + 1)
2(t − t1 )(t − t2 )
≥ 0 and k − 2 =
≤ 0.
t−1
t−1
1
(3.4) Case: a < −2 and t ∈ − ; t1 ] ∪ [−a − 1 ; t2 ].
2
1
When t ∈ − ; t1 ] ⇒ t ≤ t1 < 1, 2t + 1 > 0, t + a + 1 < 0, i.e.
2
⇒ k+1=
k+1=
(2t + 1)(t + a + 1)
2(t − t1 )(t − t2 )
≥ 0 and k − 2 =
≤ 0.
t−1
t−1
When t ∈ [−a − 1 ; t2 ] ⇒ t > 1, t + a + 1 ≥ 0 and again
k+1=
2(t − t1 )(t − t2 )
(2t + 1)(t + a + 1)
≥ 0 and k − 2 =
≤ 0.
t−1
t−1
7
Remark. The function b(t) is not monotone, e.g. when
a = 2, t = −1 ⇒ c(−1) = 8, b(−1) = 3, k = 0, p = q = 1 and
w4 +2w3 +3w2 +8w1 = (x2 +xy +xz +yz)2 +(y 2 +xy +xz +yz)2 +(z 2 +xy +xz +yz)2 ≥ 0
while when t = −3 ⇒ c(−3) = −8, b(−3) = 3 = b(−1), p =
k = −1 from (7)
w4 + 2w3 + 3w2 − 8w1 =
3
7
, q=−
5
5
and as
i
1h
(X − Y )2 + (Y − Z)2 + (Z − X)2 =
2
i
1h
(x−y)2 (x+y+2z)2 +(y−z)2 (y+z+2x)2 +(z−x)2 (z+x+2y)2 ≥ 0.
2
19
Analogously when a = , t = −1 and t = −4 we obtain the inequalities:
4
19
17
19
17
49
w4 +
w3 +
w2 + 19w1 ≥ 0 and w4 +
w3 +
w2 +
w1 ≥ 0.
4
2
4
2
4
=
Remark. The parameter t generates different optimal inequalities.
1. When t = −a − 1 we have
w4 + aw3 + (a2 − 1)w2 − (a2 + 2a)w1 =
1
2
2
2
2
2
2
(x − y) (x + y + az) + (y − z) (y + z + ax) + (z − x) (z + x + ay) .
=
2
a2
a2
2. When t = t1 or t = t2 and a ∈ (−∞ ; −2] ∪ [4 ; +∞) ⇒ b =
+ 2, c =
+ a,
4
2
a
and according to (7) we have
k = 2, 2p + q =
2
a2
a2
w4 + aw3 +
+ 2 w2 −
+ a w1 =
4
2
2
2
a
2
2
2
= X + Y + Z = x + y + z + (xy + yz + zx) .
2
3. When t = −2 and a ∈ [1 ; 4] we have
w4 + aw3 + 2(a − 1)w2 + (5a − 8)w1 =
2
= x+y+z
x2 + y 2 + z 2 + (a − 2) (xy + yz + zx) .
4. When t = −1 and a ∈ [0 ; 6] ⇒ p = q = 1, k = −1 +
a
2
and we have
w4 + aw3 + (2a − 1)w2 + 4aw1 ≥ 0.
5. When t = 0 and a ∈ [−4 ; −1] ⇒ p = −1, q = 0, k = −a − 2 and we have
w4 + aw3 − 2(a + 1)w2 − aw1 ≥ 0.
8
Section II. (nonnegative real variables)
Remark. In the proof we have in mind that the following fundamental symmetric
homogeneous polinomial inequalities of degree four for three nonnegative real variables
hold: 2w4 ≥ w3 ≥ 2w2 ≥ w1 ≥ 0, and the inequality I. Schur [11] w4 + w1 ≥ w3 .
Proof of Lemma 5. From f (1,
1, 1) ≥
0 ⇒ c ≥ −2a − b − 1.
4
3
From 0 ≤ f (x, 1, 0) = x + 1 + a x + x + bx2 ⇒ x2 + x−2 + a x + x−1 + b ≥ 0.
We set u = x + x−1 ≥ 2. When a ≥ −4 and u = 2 ⇒ b ≥ −2(a + 1).
a2
a
+ 2.
When a ≤ −4 and u = − ≥ 2 ⇒ b ≥
2
4
Proof of Theorem 6. The necessity follows from Lemma 5.
Using the inequalities: w4 + w1 ≥ w3 , w3 ≥ 2w2 and w2 ≥ w1 ≥ 0 we see that the
conditions 2a + 2 + b ≥ 0 and 1 + 2a + b + c ≥ 0 are also sufficient:
f = w4 + aw3 + bw2 + cw1 = (w4 + w1 − w3 )+
+ (a + 1)(w3 − 2w2 ) + (2a + 2 + b)(w2 − w1 ) + (2a + b + c + 1)w1 ≥ 0.
Proof of Theorem 7. Necessity.
f (x, y, z) = w4 + aw3 + b(t)w2 + cw1 = w4 + aw3 + b(t)w2 + c(t)w1 + c − c(t) w1 .
From 0 ≤ f (t, 1, 1) = 0 + c − c(t) t(t + 2) when t > 0 follows that c ≥ c(t).
When t = 0 then c(0) = −a, b(0) = −2(a + 1), a ∈ (−4 ; −1).
f (x, 1, 1)
From 0 ≤ lim
= 2(c + a) it follows that c ≥ −a = c(0).
x→+0
x
Sufficiency follows analogously as in Theorem 3 from the indentity:
w4 + aw3 + b(t)w2 + c(t)w1 = X 2 + Y 2 + Z 2 + k(XY + Y Z + ZX),
defined with formulas (7), where −1 ≤ k ≤ 2, because each of the three cases are
subregions of these of Theorem 3:
(7.1) −2 ≤ a < −1 and t ∈ [ 0 ; −a − 1] contains itself in
1
1
(3.2) that is −2 ≤ a < − and t ∈ − ; −a − 1];
2
2
(7.2) −4 < a < −2 and t ∈ [0; t1 ] ∪ [−a − 1; t2 ] and (7.3) a ≤ −4 and t ∈ [−a − 1; t2 ]
1
contain itself in (3.4) a < −2 and t ∈ (− ; t1 ] ∪ [−a − 1 ; t2 ].
2
References
[1]
T. Ando, Some Homogeneous Cyclic Inequalities of Three Variable of Degree Three
and Four, The Australian Journal of Math. Anal. and Appl. 7, (2), art. 12, (2011).
[2]
T. Ando, Cubic and quartic cyclic homogeneous inequalities of Three Variables,
Math. Inequal. Appl. 16, 127-142 (2013).
[3]
T. Ando, Discriminants of Cyclic Homogeneous Inequalities of Three Variables,
Preprint, (29 Sep. 2015) http://www.math.s.chiba-u.ac.jp/ando/ineq17.pdf
[4]
V. Cirtoaje, On the Cyclic Homogeneous Polynomial Inequalities of Degree Four,
Journal of Inequalities in Pure and Applied Mathematics 10 (3), art. 67, (2009).
9
[5]
V. Cirtoaje, Necessary and sufficient conditions for symmetric homogeneous polynomial inequalities of degree four and six in real variables, J. Nonlinear Sci. Appl.,
5, 307-320 (2012).
[6]
V. Cirtoaje, Necessary and Sufficient Conditions for Symmetric Homogeneous Polynomial Inequalities in Nonnegative Real Variables, Math. Inequal. Appl., 16, 413-438
(2013).
[7]
V. Cirtoaje, A Strong Method for Symmetric Homogeneous Polynomial Inequalities
of Degree Six in Nonnegative Real Variables, Brit. J. Math. & Comp. Sci., 4, 685703 (2014).
[8]
V. Cirtoaje, Y. Zhou, Necessary and sufficient conditions for cyclic homogeneous
polynomial inequalities of degree four in real variables, The Australian Journal of
Mathematical Analysis and Applications, 9 (1) Article 15, pp. 1-17, (2012).
[9]
M. Milev, N. Milev, On the Cyclic Homogeneous Polynomial Inequalities of Degree
Four of Three Nonnegative Real Variables, 2015, accepted in Journal of Mathematical Inequalities, http : //jmi.ele − math.com/.
[10] Y. Zhou, V. Cirtoaje, Some strong sufficient conditions for cyclic homogeneous
polynomial inequalities of degree four in nonnegative variables, J. Nonlinear Sci.
Appl., 6 (2013).
[11] I. Schur: http : //en.wikipedia.org/wiki/Schur%27s inequality.
| 0math.AC |
STRONG TOPOLOGIES FOR SPACES OF SMOOTH MAPS WITH
INFINITE-DIMENSIONAL TARGET
arXiv:1603.09127v3 [math.GN] 7 Apr 2016
EIVIND OTTO HJELLE AND ALEXANDER SCHMEDING
A BSTRACT. In this article we study two “strong” topologies for spaces of smooth functions from a finite-dimensional manifold to a (possibly infinite-dimensional) manifold
modelled on a locally convex space. Namely, we construct Whitney type topologies for
these spaces and a certain refinement corresponding to Michor’s F D-topology. Then
we establish the continuity of certain mappings between spaces of smooth mappings,
e.g. the continuity of the joint composition map. As a first application we prove that the
bisection group of an arbitrary Lie groupoid (with finite-dimensional base) is a topological group (with respect to these topologies).
For the reader’s convenience the article includes also a proof of the folklore fact that
the Whitney topologies defined via jet bundles coincide with the ones defined via local
charts.
Keywords: topologies on spaces of smooth functions, Whitney topologies, continuity
of composition, bisections of a Lie groupoid, manifolds of smooth maps
MSC2010: 54C35 (primary); 26E20, 58D15, 58H05 (secondary)
Acknowledgment: This research was part of StudForsk at NTNU, Trondheim. We
acknowledge financial support by Olav Thon Stiftelsen.
C ONTENTS
Introduction and statement of results
2
1. The very strong topology
6
2. Composition of maps in the very strong topology
13
3. The fine very strong topology
21
4. The manifold structure on smooth vector valued functions
24
5. Application to bisection groups
28
E-mail addresses: eivindhj@stud.ntnu.no and alexander.schmeding@math.ntnu.no.
1
STRONG TOPOLOGIES FOR SMOOTH MAPS
2
Appendix A. Calculus in the locally convex setting
31
Appendix B.
Auxiliary results for Sections 1 and 2
34
Appendix C.
Alternative description of the topology via jet bundles
38
References
45
I NTRODUCTION
AND STATEMENT OF RESULTS
This paper gives a systematic treatment of two topologies for spaces of smooth functions from a finite-dimensional manifold to a (possibly infinite-dimensional) manifold
modeled on a locally convex space.
In particular, we establish the continuity of certain mappings between spaces of smooth
mappings, e.g. the continuity of the joint composition map. As a first application we
prove that the bisection group of an arbitrary Lie groupoid (with finite-dimensional
base) is a topological group. For the most part, these results are generalizations of well
known constructions to spaces of smooth functions with infinite-dimensional range. We
refer to [Ill03, Mic80, Hir94] for topologies on spaces of smooth functions between
finite-dimensional manifolds.
To understand these results of the present article, recall first the situation for spaces of
smooth functions between finite-dimensional manifolds. For 0 ≤ r ≤ ∞, let Cr (M, N)
denote the set of r-times continuously differentiable functions between manifolds M and
N In the case where r is finite, the standard choice for a topology on Cr (M, N) is the well
known Whitney Cr -topology (cf. [Ill03, Hir94]). For r = ∞ and M non-compact there
are several choices for a suitable topology. One can for example choose the topology
generated by the union of all Whitney Cr -topologies. We call this topology the strong
C∞ -topology and write CS∞ (M, N) for the smooth functions with this topology.1 Note
that each basic neighborhood of the strong C∞ -topology allows one to control derivatives of functions only up to a fixed upper bound. However, in applications one wants to
control the derivatives of up to arbitrary high order (this is made precise in Section 1).
To achieve this one has to refine the strong topology, obtaining the very strong topol∞ (M, N) the smooth
ogy2 in the process (cf. [Ill03] for an exposition). We denote by CvS
functions with the very strong topology and note that this topology is fine enough for
many questions arising from differential topology.
1The strong topology is in the literature often also called the “Whitney C∞ -topology”. Following
Illman in [Ill03], we will not use this naming convention as it can be argued that the strong C∞ -topology
is not a genuine C∞ -topology. See ibid. for more information.
2In [Mic80] this topology is called the D-topology.
STRONG TOPOLOGIES FOR SMOOTH MAPS
3
Unfortunately, as is argued in [Mic80] this topology is still not fine enough, if one
wants to obtain manifold structures on C∞ (M, N) (and subsequently on the group of
diffeomorphisms Diff(M)). Hence Michor constructed a further refinement of the very
strong topology, called the F D-topology. In the present paper, we call this topology the
fine very strong topology and denote the space of smooth functions with this topology
∞ (M, N).
by CfS
Note that the topologies discussed so far coincide if the source manifold is compact. In
fact in this case, all of these topologies coincide with the compact open C∞ -topology (see
e.g. [Nee06, Definition 5.1]). The compact open C∞ -topology for infinite-dimensional
target manifolds is already well understood and has been used in many investigations,
for example in infinite-dimensional Lie theory, e.g. [Glö02]. Hence our investigation
will only turn up new results for non-compact source manifolds and infinite-dimensional
target manifolds.
We will now go into some more detail and explain the main results of the present paper.
Our aim is now to generalize the construction of the very strong and fine very strong
topology to the set of smooth functions C∞ (M, X ), where X is a locally convex manifold.
Here smooth maps are understood in the sense of Bastiani’s calculus [Bas64] (often also
called Keller’s Cr -theory [Kel74]). We refer to [Mil83, Glö02, Nee06] for streamlined
expositions, but have included a brief recollection in Appendix A.
Working in this framework we construct the very strong and the fine very strong topology for C∞ (M, X ), where M is finite-dimensional and X is a locally convex manifold.
Our exposition mostly follows Illman’s article [Ill03] and we adapt his arguments to
our setting. In particular, we describe the topology in terms of local charts as in [Ill03]
(cf. also [Hir94]). For finite-dimensional manifolds one can alternatively introduce the
topology using jet bundles and it is well known that both approaches yield the same
topology. This fact seems to be a folklore theorem, but we were not able to locate a
proof in the literature. As this fact is needed later on, a proof is given in Appendix C.
The advantage of the approach using local charts can be summarized as follows: Arguments and proofs often split into two distinct steps. First one establishes a property
of the function space topology only for the (easier) special case of vector space valued
smooth mappings. Then a localization argument involving manifold charts allows one
to establish the result for smooth maps between manifolds.
To our knowledge the topologies discussed in the present paper have so far only been
studied for finite-dimensional manifolds. A topology somewhat similar to the very
strong topology but for infinite-dimensional manifolds can be found in [KM97, Section 41]. Albeit the similar look, be aware that the jet bundles used in the construction
are only manifolds in the inequivalent convenient setting of calculus. In particular, the
topology in loc.cit. does not coincide with the one constructed here if M is non-compact
STRONG TOPOLOGIES FOR SMOOTH MAPS
4
(cf. [KM97, 42.2 Remarks]). We refer to Remark 1.12 for related topologies on function spaces between Banach manifolds. For finite-dimensional manifolds, our construction recovers exactly the ones in the literature. We exploit this and recall that the set
Prop(N, N) ⊆ C∞ (M, N) of all proper maps is open in the very strong and the fine very
strong topology. Then one can establish continuity of certain composition mappings, in
particular our results subsume the following theorem.
Theorem A Let M, N be finite-dimensional manifolds, X and Y be (possibly infinitedimensional) manifolds. In the following, endow all function spaces either with the very
strong or the fine very strong topology. Then the joint composition
Γ : Prop(M, N) ×C∞ (N, X ) → C∞ (M, X ),
( f , g) 7→ g ◦ f
is continuous.
Further, for any smooth map h : X → Y , the pushforward
h∗ : C∞ (M, X ) → C∞ (M,Y ),
f 7→ h ◦ f
is continuous.
Having this theorem at our disposal, we construct an interesting class of topological
groups: Suppose G = (G ⇒ M) is a Lie groupoid. This means that G, M are smooth
manifolds, equipped with submersions α , β : G → M and an associative and smooth
multiplication G×α ,β G → G that admits a smooth identity map 1 : M → G and a smooth
inversion ι : G → G. Then the bisections Bis(G ) of G are the sections σ : M → G of α
such that β ◦ σ is a diffeomorphism of M. This becomes a group with respect to
(σ ⋆ τ )(x) := σ ((β ◦ τ )(x))τ (x) for x ∈ M.
Many interesting groups from differential geometry such as diffeomorphism groups,
automorphism groups and gauge transformations of principle bundles can be realised
as bisection groups of suitable Lie groupoids. By construction Bis(G ) ⊆ C∞ (M, G) and
with respect to the topologies on the space of smooth functions we obtain the following.
Theorem B Let G = (G ⇒ M) be a Lie groupoid with finite-dimensional base M. Then
(Bis(G ), ⋆) is a topological group with respect to the subspace topology induced by
either the very strong or the fine very strong topology on C∞ (M, G).
This result is a first step needed to turn the bisection group into an infinite-dimensional
Lie group. In fact, it turns out that one can establish this result quite easily (see below)
once Theorem B is available. The key step to establish the applications mentioned
below, is to work out the continuity of certain composition mappings (which has been
done in Theorem A). Then Proposition C and Theorem D below can be established using
standard techniques from the literature. In the present paper we will be only concerned
with properties of the topology on function spaces. Hence the next results are stated
STRONG TOPOLOGIES FOR SMOOTH MAPS
5
without a proof. We provide only some references to the literature and hope to provide
details in future work.
Proposition C Let M be a finite-dimensional manifold and X be a possibly infinite∞ (M, X ) can be turned
dimensional manifold which admits a local addition.3 Then CfS
into a manifold modeled on spaces of compactly supported sections of certain bundles.
It turns out that once the space of smooth functions is endowed with the correct topology
it is not hard to prove Proposition C. More details and references to literature containing the necessary auxiliary facts can be found at the end of Section 1. Proposition C
generalizes [Mic80, Theorem 10.4] in so far as it admits arbitrary infinite-dimensional
manifolds as target manifolds (whereas loc.cit. was confined to finite-dimensional targets). We remark that in [KM97, 42.4 Theorem] the smooth functions C∞ (M, X ) for
M and X as in Proposition C have been endowed with a manifold structure in the inequivalent convenient setting of calculus. However, following [KM97, 42.2 Remarks]
the topology on C∞ (M, X ) used in the construction does not coincide with the fine very
strong topology if M is non-compact. Hence both constructions are inequivalent even if
both M and X are finite-dimensional (and M is non-compact).
The manifold structure provided by Proposition C allows one to establish the Lie group
structure for a general class of bisection groups. Adapting arguments from [Mic80] and
[SW15a] one can prove that
Theorem D The group of bisections of a Lie groupoid G = (G ⇒ M) with M finitedimensional and G a Banach manifold4 is an infinite-dimensional Lie group.
This generalizes the construction from [SW15a], where the group of bisections of a Lie
groupoid with compact base was turned into an infinite-dimensional Lie group. Thus
one obtains a conceptual approach to the Lie group structures of many groups which are
of interest in differential geometry (e.g. automorphism groups and gauge transformation
groups of principle bundles over a non compact base). Moreover, Theorem D is a crucial
ingredient if one wants to extend the strong connection between Lie groupoids and
infinite-dimensional Lie groups which was developed in [SW15b].
3This is for example satisfied if X is a Lie group, see also [KM97, Section 42.4] for a definition of
local additions and more examples.
4Assuming certain mild conditions on G (i.e. an adapted local addition, cf. [SW15a]), it is not necessary to assume that G is a Banach manifold.
STRONG TOPOLOGIES FOR SMOOTH MAPS
1. T HE
6
VERY STRONG TOPOLOGY
In this section, we introduce the very strong topology on the space C∞ (M, X ) of smooth
maps from a finite-dimensional smooth manifold M to a possibly infinite-dimensional
smooth manifold X . The very strong topology allows us to control derivatives of smooth
maps up to arbitrarily high order on certain families of compact sets. This is a straightforward generalization of the very strong topology on the space of smooth maps between
finite-dimensional manifolds as described in [Ill03].
Notation and conventions. We write N := {1, 2, . . .} and N0 := {0, 1, . . .}, and will
only work with vector spaces over the field of real numbers R. Finite-dimensional manifolds are always assumed to be σ -compact, i.e. a countable union of compact subspaces
(which for finite-dimensional manifolds is equivalent to being second countable). We
always endow Rn with the supremum norm k · k∞ unless otherwise stated. We define
Bnε (x) := {y ∈ Rn : ky − xk∞ < ε }. Notation and conventions regarding locally convex
vector spaces, smooth maps, and infinite-dimensional manifolds is covered in Appendix
A. Typically, M and N will be finite-dimensional smooth manifolds, X a smooth manifold modeled on a locally convex vector space, and E a locally convex vector space.
Definition 1.1. Let E be a locally convex vector space, p a continuous seminorm on E,
f : Rm → E smooth, A ⊆ Rm compact, r ∈ N0 , and e1 , . . . , em the standard basis vectors
in Rm . Then define
k f k(r, A, p) = sup{p(d(k) f (a; α )) : a ∈ A, α ∈ {e1 , . . ., em }k , 0 ≤ k ≤ r}.
Remark 1.2. The symbol d(k) f is defined in Definition A.4. Elsewhere in the literature,
d(k) f (x; y) = d(k) f (x; y1 , . . ., yk ) is often denoted
∂k
f (x)
∂ yk · · · ∂ y1
where y = (y1 , . . ., yk ).
or
∂
f (x),
∂y
In the definition above we require α ∈ {e1 , . . . , em }k . But for any α ∈ Bn1 (0) and a ∈ A
and k ≤ r we have p(d(k) f (a; α )) ≤ Kk f k(r, A, p) for some constant K depending only
on r and m, by (3) in Proposition A.7.
If E = Rn , any norm generates the topology on E and norms are in particular seminorms.
By Proposition 1.13, the very strong topology is not affected if we always assume that
the seminorm p on Rn is the supremum norm k · k∞ . In this case we simply write
k f k(r, A) for k f k(r, A, k · k∞).
Lemma 1.3 (Triangle inequality). Let E, p, A, r be as in Definition 1.1. Then the map
k · k(r, A, p) : C∞ (Rm , E) → R
satisfies the triangle inequality. In fact it is a seminorm on C∞ (Rm , E).
STRONG TOPOLOGIES FOR SMOOTH MAPS
7
Proof. Use linearity of d(−)(a, α ) for fixed (a, α ), and the fact that p satisfies the triangle inequality.
Definition 1.4 (Elementary neighborhood). Let E, p, and r be as in Definition 1.1, M
an m-dimensional smooth manifold, X a smooth manifold modeled on E. Consider
f : M → X smooth, (U, φ ) a chart on M, (V, ψ ) a chart on X , A ⊆ U compact such that
f (A) ⊆ V , and ε > 0. Define
N r ( f ; A, (U, φ ), (V, ψ ), p, ε ) = {h ∈ C∞ (M, E) : h(A) ⊆ V and
kψ ◦ h ◦ φ −1 − ψ ◦ f ◦ φ −1 k(r, φ (A), p) < ε }.
We call this set an elementary Cr -neighborhood of f in C∞ (M, E).
Conventions for elementary neighborhoods If X = Rn , we will assume that p is the
supremum norm and omit the p when writing down the elementary neighborhoods.
When there is a canonical choice of charts for our manifolds, e.g. if X = E is a locally convex vector space, we omit the obvious charts when writing down elementary Cr -neighborhoods. Thus for f : M → E we write e.g. N r ( f ; A, (U, φ ), p, ε ) :=
N r ( f ; A, (U, φ ), (X , id), p, ε ).
Remark 1.5.
(1) The conditions f (A) ⊆ V and h(A) ⊆ V ensure that the map ψ ◦
h ◦ φ −1 − ψ ◦ f ◦ φ −1 makes sense. Further, the conditions enable us to control the open sets into which a (given) compact set is mapped, i.e. the kind of
control provided by the well known compact open topology (cf. [Nee06, Definition I.5.1]). Indeed, by restricting to elementary C0 -neighborhoods, one would
recover a subbase of the compact open topology on C∞ (M, X ).
(2) We define elementary neighborhoods only for finite-dimensional source manifolds as the seminorms in Definition 1.1 make only sense for these manifolds.
Compare Remark 1.12 for more information on alternative approaches to the
topology which avoid this problem.
We now define what will become the basis sets in the very strong topology on C∞ (M, X ).
Definition 1.6 (Basic neighborhood). Let f : M → X be a smooth map from a finitedimensional smooth manifold M to a smooth manifold X modeled on a locally convex
vector space E. A basic neighborhood of f in C∞ (M, X ) is a set of the form
\
N ri ( f ; Ai , (Ui, φi ), (Vi , ψi ), pi , εi ),
i∈Λ
where Λ is a possibly infinite indexing set, for all i the other parameters are as in Definition 1.4, and {Ai }i∈Λ is locally finite. We call {Ai }i∈Λ the underlying compact family
of the neighborhood.
STRONG TOPOLOGIES FOR SMOOTH MAPS
8
Without loss of generalization, Λ = N, since every locally finite family over a σ -compact
space is countable.
As Proposition 1.8 show, the basic neighborhoods in C∞ (M, X ) form a basis for a topology on C∞ (M, X ). In order to prove the proposition we need the following lemma.
Lemma 1.7. Let f : M → X be smooth, and g ∈ N := N r ( f ; A, (U, φ ), (V, ψ ), p, ε ).
Then there exists ε ′ > 0 such that N ′ := N r (g; A, (U, φ ), (V, ψ ), p, ε ′) ⊆ N .
Proof. For h, h̃ ∈ C∞ (M, X ) with h(A), h̃(A) ⊆ V , let
d(h, h̃) = kψ ◦ h̃ ◦ φ −1 − ψ ◦ h ◦ φ −1 k(r, φ (A), p).
Note that d satisfies the triangle inequality by Lemma 1.3, and that h ∈ N is equivalent
to d( f , h) < ε .
Set ε ′ = ε − d( f , g), and let N ′ be as in the statement of the lemma. If h ∈ N ′ , then
d( f , h) ≤ d( f , g) + d(g, h) < d( f , g) + (ε − d( f , g)) = ε .
Hence h ∈ N , and N ′ ⊆ N .
Proposition 1.8. Let U and U ′ be basic neighborhoods of f and f ′ in C∞ (M, X ),
respectively. If g ∈ U ∩ U ′ , then there exists a basic neighborhood V of g such that
V ⊆ U ∩ U ′.
Hence the basic neighborhoods form a basis for a topology on C∞ (M, E), called the
very strong topology on C∞ (M, E).
Proof. We may write
U =
\
Ni
and
U′=
\
N j′
j∈Λ′
i∈Λ
for some sets Λ and Λ′ , where Ni and Ni ′ are elementary neighborhoods of f and f ′ ,
respectively. For all i ∈ Λ and j ∈ Λ choose as in Lemma 1.7 elementary neighborhoods
Mi and M j′ of g such that Mi ⊂ Ni and M j′ ⊂ N j ′ . Then
!
V :=
\
i∈Λ
Mi ∩
\
j∈Λ′
Mi′ ⊆ U ∩ U ′ .
It remains to check that V is in fact a basic neighborhood of g. The set V is a basic
neighborhood of g provided that the underlying compact family of V is locally finite.
This is indeed the case since the underlying compact families of U and U ′ are locally
finite and finite unions of locally finite families are locally finite.
The preceding proposition justifies the following definition.
STRONG TOPOLOGIES FOR SMOOTH MAPS
9
Definition 1.9 (Very strong topology). The very strong topology on C∞ (M, X ) is the
topology on C∞ (M, X ) with basis the basic neighborhoods in C∞ (M, X ).
∞ (M, X ).
The set C∞ (M, X ) equipped with the very strong topology will be denoted by CvS
∞ (M, E), where E is a locally convex space.
Remark 1.10. We will work later on with CvS
To this end, we considered E as a manifold with the canonical atlas given by the identity.
This may seem artificial at first glance as one in principle needs to take all “manifold
charts” of E into account. Note however that by Lemma B.2 the very strong topology
on C∞ (M, E) is generated by all basic neighborhoods of the form
(1)
\
N ri ( f ; Ai , (Ui , φi ), (E, idE ), pi , εi ),
i∈Λ
i.e. it suffices to consider elementary neighborhoods with respect to the identity chart.
Hence the topology on C∞ (M, E) is quite natural.
Similarly for C∞ (Rn , E) the charts (Ui , φi ) in (1) can be replaced by (Rn , idRn ) by
Lemma B.3. In the following, we will always assume that our elementary and basic
neighborhoods are constructed with respect to the identity if one (or both) of the manifolds are a locally convex space.
Remark 1.11. There are other well-known topologies on C∞ (M, X ). The strong topology (or Whitney C∞ -topology) and the compact open C∞ -topology (or weak topology)
have as bases neighborhoods of the form described in Definition 1.6, with some additional restrictions. For the strong topology the collection {ri }i∈Λ of indices giving
differentiation order is bounded, and for the compact open C∞ -topology we require that
the indexing set Λ is finite.
The very strong topology is finer than the strong topology which is finer than the compact open C∞ -topology, and in the case that M is compact all of these topologies coincide
(since every locally finite family meets a compact set only finitely many times). We refer the reader to section 2.1 in [Hir94] for information about the strong and compact
open C∞ topologies in the case that X is finite-dimensional. A comparison of the strong
topology and the very strong topology can be found in the introduction of [Ill03].
Since the very strong topology is finer than the strong topology, subsets of C∞ (M, X )
that are open in the strong topology are also open in the very strong topology. [Hir94,
Section 2.1] has several results stating that certain subsets of C∞ (M, N) are open in the
strong topology, consequently also in the very strong topology. In particular, the set
∞ (M, N). We write Prop (M, N) for
Prop(M, N) of proper smooth maps is open in CvS
vS
∞
the subspace Prop(M, N) of CvS (M, N) equipped with the subspace topology.
Remark 1.12. One can also define the very strong topology on the space C∞ (X ,Y )
where X and Y are Banach manifolds (i.e. modeled on Banach spaces). To this end one
needs to redefine the seminorms generating the topology, which in the vector space case
will take the following form:
STRONG TOPOLOGIES FOR SMOOTH MAPS
10
If X ,Y are Banach spaces, f : X → Y smooth, r ∈ N0 , and A ⊆ X compact define
n
o
(2)
k f k(r, A) = sup kDk f (x)kY : 0 ≤ k ≤ r and x ∈ A ,
where Dk f denotes the k-th Fréchet derivative of f .
Here we use that every smooth Bastiani map is also smooth in the sense of Fréchet
differentiability by [Mil83, Lemma 2.10]. It is easy to see that all statements made on
elementary neighborhoods in the present section remain valid. Hence we obtain a very
strong topology on smooth functions between Banach manifolds.
Note that one can prove as in Appendix C that the “very strong topology” constructed
with respect to the seminorms (2) induces again the (original) very strong topology on
C∞ (X ,Y ) if X is finite-dimensional. Unfortunately, for an infinite-dimensional Banach
manifold X this topology does not allow us to control the behavior of functions “at
infinity” (or anywhere for that matter since compact subsets of infinite-dimensional Banach spaces have empty interior). To see this recall that manifolds modeled on infinitedimensional Banach space don’t have a locally finite compact exhaustion by the Baire
category theorem.
Recall however, that one can define a Whitney C∞ -topology for X ,Y Banach manifolds
via jet bundles (see e.g. [Mic80, KM97] or Appendix C for a short exposition). As
shown in [MROD92, Chapter 9], this topology then allows one to control the behavior
of a function on all of X . The key difference is that the Whitney topology defined in this
way controls the behavior of jets on locally finite families of closed sets. Obviously, one
can not hope to describe it via the seminorms as the existence of the suprema in the seminorms is tied to the compactness of the sets. Even worse, for an infinite-dimensional
manifold X and Y = E a locally convex space, the largest topological vector space contained in C∞ (X , E) with respect to this topology is trivial (cf. [KM97, 437]). For these
reasons we work exclusively with the very strong topology for finite-dimensional source
manifolds.
Additional facts about the very strong topology. Sometimes it is convenient to assume that the continuous seminorms p used in constructing very strong neighborhoods
are of a certain form, as we have already remarked. There is no loss of generality in
making such assumptions if the family of seminorms that we restrict to is “big enough”.
Proposition 1.13. Let M be a finite-dimensional smooth manifold and X a smooth manifold modeled on a locally convex vector space E. Suppose P is a generating family of
seminorms for E (see Definition A.2).
If we replace every instance of “p is a continuous seminorm on E” in the definitions
and results earlier in this section with “p ∈ P”, then the resulting very strong topology
on C∞ (M, X ) is unaffected.
STRONG TOPOLOGIES FOR SMOOTH MAPS
11
Proof. Let T be the very strong topology on C∞ (M, X ) constructed with respect to
all continuous seminorms on E, and let T ′ be the very strong topology on C∞ (M, X )
obtained by restricting to seminorms in P. Then T ′ is obviously coarser than T since
every p ∈ P is continuous, so it suffices to show that T is coarser
than T ′ . This will be
T
the case if for every basic T -very strong neighborhood U = i∈Λ Ni of f ∈ C∞ (M, X ),
where each Ni is an elementary T -very strong neighborhood
Ni = N ri ( f ; Ai , (Ui , φi ), (Vi, ψi ), pi , εi ),
there exists a basic T ′ -very strong neighborhood U ′ of f such that U ′ ⊆ U .
Fix i ∈ Λ. By (2) in Proposition A.3 there exist ni ∈ N and pi,1 , . . ., pi,ni ∈ P and ci > 0
such that pi ≤ ci sup1≤ j≤ni pi, j . And then
ni
\
εi
ri
N
f ; Ai , (Ui , φi ), (Vi, ψi ), pi, j ,
Vi :=
⊆ Ni .
2ci
j=1
Indeed, if g ∈ Vi , then for a ∈ Ai , α ∈ {e1 , . . . , em }k , 0 ≤ k ≤ ri , and 0 ≤ j ≤ ni , we have
εi
ci pi, j (d(ψi ◦ g ◦ φi−1 − ψi ◦ f ◦ φi−1 )(k) (a, α )) < ,
2
which together with pi ≤ ci sup pi, j clearly implies that g ∈ Ni .
Now set U ′ :=
U′⊆U.
T
i∈Λ Vi .
This is a basic T ′ -very strong neighborhood of f such that
The following lemma is useful when constructing certain basic neighborhoods. The
proof given here is fairly detailed, but throughout the remainder of this text the details
of similar arguments will be omitted.
Lemma 1.14. Let M be a finite-dimensional smooth manifold, X a locally convex manifold, and f : M → X a smooth map. Suppose {Kn }n∈N is a locally finite family of
compact subsets of M. Then there exist families of charts {(Vi , ψi )}i∈N for X and
{(Ui , φi )}i∈N for M, and a locally finite family {Ai }i∈N of compact subsets of M such
that
(1) i∈N Ai = n∈N Kn ,
(2) Ai ⊆ Ui for all i ∈ N,
(3) f (Ui ) ⊆ Vi for all i ∈ N.
S
S
Proof. Fix n ∈ N. For every x ∈ Kn choose a chart (Vn,x , ψn,x ) around f (x) and a chart
(Un,x , φn,x ) around x. By shrinking Un,x we may assume that f (Un,x ) ⊆ Vn,x . Since M is
locally compact there exists a compact neighborhood A′n,x around x such that A′n,x ⊆ Un,x .
Now set An,x = Kn ∩A′n,x . By compactness of Kn there exist finitely many xn,1 , . . . , xn,kn ∈
n
Kn such that {An,xn, j }ki=1
covers Kn .
STRONG TOPOLOGIES FOR SMOOTH MAPS
12
The families {(Vn,xn,i , ψn,xn,i )}n,i , {(Un,xn,i , φn,xn,i )}n,i and {An,xn,i }n,i have the desired
properties. By relabeling the indices we can take the indexing set to be N.
Lemma 1.15. Let M be a finite-dimensional smooth manifold, X a smooth manifold
modeled on a locally convex vector space E, and let U ⊆ M and V ⊆ X be open subsets.
∞
∞ (M, X ).
Consider the subspace CvS,sub
(U,V ) := { f ∈ C∞ (M, X ) : f (U ) ⊆ V } ⊆ CvS
∞
∞ (M, X ).
(1) CvS,sub
(U,V ) is an open subset of CvS
∞
∞ (U,V ) is continuous.
(2) The restriction resvS : CvS,sub
(U,V ) → CvS
∞
r
(3) If f ∈ C (U,V ) and N ( f ; A, (Uφ , φ ), (Vψ , ψ ), p, ε ) is an elementary neighborhood of f such that ψ (Vψ ) is a convex set, then there exists g ∈ C∞ (M, X ) with
r
r
res−1
vS (N ( f ; A, (Uφ , φ ), (Vψ , ψ ), p, ε )) = N (g; A, (Uφ , φ ), (Vψ , ψ ), p, ε )
(3)
Proof.
(1) Suppose f ∈ C∞ (U,V ). Since U is an open subset of M, the subspace U is
metrizable and locally compact, hence σ -compact. Combining Lemma B.4 and
Lemma 1.14, we find a locally finite exhaustion {An }n∈N of U by compact sets,
charts {(Un, φn )}n∈N for M, and charts {(Vn , ψn )}n∈N for X such that An ⊆ Un
and f (An ) ⊆ Vn for all n ∈ N. Since f (An ) ⊆ V , shrink the Vn if necessary such
that Vn ⊆ V for all n ∈ N (while still f (An ) ⊆ Vn ). Take any continuous seminorm
p on E and define
\
U =
N 0 ( f ; An , (Un, φn ), (Vn , ψn ), p, 1) .
n∈N
If gS∈ U , then g(An ) ⊆ Vn ⊆ V for all n ∈ N, from which it follows that g(U ) =
∞ (M, X ) such that U ⊆
g ( An ) ⊆ V . So U is a neighborhood of f in CvS
C∞ (U,V ).
(2) Take an arbitrary basic neighborhood
U =
\
N ri ( f ; Ai , (Ui , φi ), (Vi, ψi ), pi , εi )
i∈Λ
∞ (U,V ). We will show that given g ∈ res−1 (U ), there exists a basic neighin CvS
vS
∞ (M, X ) such that V ⊆ res−1 (U ), i.e. res−1 (U ) is open.
borhood V of g in CvS
vS
vS
By Lemma 1.8 there are elementary neighborhoods of resvS (g) such that
V =
\
N ri (resvS (g); Ai , (Ui , φi ), (Vi , ψi ), pi , δi ) .
i∈Λ
is contained in U . Now clearly i∈Λ N ri (g; Ai , (Ui , φi ), (Vi , ψi ), pi , δi ) is con−1
tained in res−1
vS (V ) ⊆ resvS (U ).
(3) Since M is finite-dimensional, whence paracompact, we can choose a neighborhood W of A and a smooth cutoff function ρ : M → R with ρ |W ≡ 1 and
ρM\ f −1 (Vψ )∩Uφ ≡ 0. Composing with a suitable translation, we may assume without loss of generality that ψ (Vψ ) is a convex 0-neighborhood. Suppressing the
T
STRONG TOPOLOGIES FOR SMOOTH MAPS
13
translation we can thus define
(
ψ −1 (ρ (x) · ψ ◦ f (x)) if x ∈ Uφ ,
g : M → X , x 7→
ψ −1 (0)
else.
Now as g|W = f |W the identity (3) is satisfied.
2. C OMPOSITION
OF MAPS IN THE VERY STRONG TOPOLOGY
Throughout this section, M and N are finite-dimensional smooth manifolds and X denotes a smooth manifold modeled on a locally convex vector space E.
It is a desirable property of the very strong topology on C∞ (M, X ) that composition
∞
∞
∞
Γ : CvS
(M, N) ×CvS
(N, X ) → CvS
(M, X )
( f , h) 7→ h ◦ f
is continuous. But this is not the case in general, a counterexample can be found in
Example 2.1. However, the restriction of the composition map
∞
∞
Γ : PropvS (M, N) ×CvS
(N, X ) → CvS
(M, X )
∞ (M, N) consisting of all
is continuous, where PropvS (M, N) denotes the subspace of CvS
the proper maps. This is precisely what Theorem 2.5 says, and proving it is the main
goal of this section.
As we will see, the crucial property of proper maps needed is the fact that if {Ai } is a
locally finite family of subsets of M and f : M → N is proper, then { f (Ai )} is locally
finite. This will enable us to choose for a basic neighborhood V of some composition
∞ (M, X ) basic neighborhoods U and U ′ of f and h, respectively, such that
h ◦ f in CvS
Γ(U × U ′ ) ⊆ V . The challenge is to choose U ′ in such a way that the underlying
compact family of the neighborhood is locally finite.
We now give the promised counterexample to the statement that composition of maps in
the very strong topology is continuous in general. This example is inspired by the proof
of [Glö06, Proposition 2.2(b)].
Example 2.1. The composition map
∞
∞
∞
Γ : CvS
(R, R) ×CvS
(R, R) → CvS
(R, R),
( f , h) 7→ h ◦ f
is not continuous.
∞ (R, R) there exists a basic
Proof. Note that for every basic neighborhood U of f ∈ CvS
′
neighborhood U of f with underlying compact family {[2n − 1, 2n + 1]}n∈Z such that
f ∈ U ′ ⊆ U , since each compact interval [2n − 1, 2n + 1] intersects only finitely many
sets belonging to the locally finite underlying compact family of U .
STRONG TOPOLOGIES FOR SMOOTH MAPS
14
To show discontinuity of Γ it suffices to show discontinuity at (0, 0). Let V be the basic
neighborhood of 0 given by
\
V :=
N n (0; [2n − 1, 2n + 1], 1).
n∈N
We will show that for any pair of basic neighborhoods
U =
\
N
rn
(0; [2n − 1, 2n + 1], εn),
N
rn′
(0; [2n − 1, 2n + 1], εn′ ),
n∈Z
U′=
\
n∈Z
there exists a pair of functions ( f , h) ∈ U ′ × U such that h ◦ f ∈
/V.
Construct h ∈ C∞ (R, R) such that in a neighborhood of 0, h is given by the equation
h(x) = xr0 +1 , and such that supp h ⊆] − 1, 1[. For some sufficiently small k > 0 we will
have kh ∈ U . For every m ∈ N define
hm (x) :=
k
h(mx)
mr0
and note that hm ∈ U , since
( j)
|hm (x)| =
km j ( j)
|h (mx)| ≤ k|h( j) (mx)| < ε0
mr0
for j ≤ r0 , where we use the notation g( j) (y) = d( j) g(y; 1, . . ., 1) for smooth maps
g : R → R.
Let 2n ≥ r0 + 1 and construct f˜ ∈ C∞ (R, R) such that f˜(x) = x − 2n in a neighborhood
of 2n and supp f˜ ⊆]2n − 1, 2n + 1[. Then for some sufficiently small s > 0 we have
f := s f˜ ∈ U ′ .
So far we have a sequence {hm }m∈N ⊂ U and f ∈ U ′ . By construction, hm ◦ f (x) =
kmsr0 +1 (x − 2n)r0+1 in a neighborhood of 2n. Hence
|(hm ◦ f )(r0 +1) (2n)| = kmsr0 +1 (r0 + 1)! ≥ 1
for large enough m, in which case hm ◦ f ∈
/V.
Having given the example above, we return our focus to the main task of the section,
which is proving Theorem 2.5. Leading up to the theorem is a sequence of lemmata.
Although we are actually interested in mapping spaces between manifolds, we first give
a lemma that only applies to vector spaces. In a sense, this lemma resolves the main
difficulty, and generalizing to manifolds is only a matter of dealing with charts.
STRONG TOPOLOGIES FOR SMOOTH MAPS
15
Lemma 2.2. Consider the composition map
∞
∞
∞
Γ : CvS
(Rm , Rn ) ×CvS
(Rn , E) → CvS
(Rm , E),
( f , h) 7→ h ◦ f .
∞ (Rm , Rn ) ×C ∞ (Rn , E), and consider an arbitrary elementary neighborLet ( f , h) ∈ CvS
vS
∞ (Rm , E) of h ◦ f . For all compact neighborhoods A′
hood N = N r (h ◦ f ; A, p, ε ) ⊆ CvS
′
of f (A) there exist δ , δ > 0 such that the elementary neighborhoods
M = N r ( f ; A, δ )
and
M ′ = N r (h; A′ , p, δ ′ )
satisfy Γ(M × M ′ ) ⊆ N .
Proof. Let A′ be any compact neighborhood of f (A). We proceed in several steps.
Step 1. Our first goal is to define M . We want a δ > 0 such that for all fˆ ∈ C∞ (Rm , Rn ),
the inequality k fˆ − f k(r, A) < δ implies
ε
and
fˆ(A) ⊆ A′ .
kh ◦ ( fˆ − f )k(r, A, p) <
2
By Lemma B.1 it is possible to choose δ such that the first property holds. We may
choose δ such that the second property also holds, because fˆ(A) = ( fˆ − f )(A) + f (A) ⊆
Bδn (0) + f (A). Pick such a δ and define
M := N r ( f ; A, δ ).
Observe that by the triangle inequality (Lemma 1.3) there exists an R > 0 such that
every fˆ ∈ M satisfies k fˆk(r, A) ≤ R.
Step 2. Our second goal is to define M ′ . We want a δ ′ > 0 such that for all ĥ ∈
C∞ (Rn , E) and all fˆ ∈ M ,
ε
kĥ − hk(r, A′, p) < δ ′
=⇒
k(ĥ − h) ◦ fˆk(r, A, p) < .
2
′
A δ having this property exists by Lemma B.1 and the observation at the end of step 1.
Now define
M ′ := N r (h; A′ , p, δ ′ ).
Step 3. Now we must show that M and M ′ have the desired property. Let fˆ ∈ M and
ĥ ∈ M ′ . By the triangle inequality,
kĥ ◦ fˆ − h ◦ f k(r, A, p) ≤ k(ĥ − h) ◦ fˆk(r, A, p) + kh ◦ ( fˆ − f )k(r, A, p) < ε .
So ĥ ◦ fˆ ∈ N . Thus Γ(M × M ′ ) ⊆ N .
A version of the preceding lemma still holds if we replace Rm and Rn with finitedimensional smooth manifolds and E with an infinite-dimensional smooth manifold.
This is our next result.
STRONG TOPOLOGIES FOR SMOOTH MAPS
16
Lemma 2.3. Given f ∈ C∞ (M, N) and h ∈ C∞ (N, X ) and an arbitrary elementary neighborhood
∞
(M, X )
N := N r (h ◦ f ; A, (U, φ ), (W, η ), p, ε ) ⊆ CvS
of h ◦ f there exist finitely many
M j := N r ( f ; A j , (U, φ ), (V j , ψ j ), δ j ) and M j′ := N r (h; A′j , (V j , ψ ), (W, η ), p, δ j′ )
such that
(1) Γ
\
Mj ×Mj
j
(2)
[
′
A j = A,
!
⊆N ,
j
(3) f (A j ) ⊆ int A′j ⊆ A′j ⊆ V j for all j.
Moreover, given any neighborhood Q of f (A) we may choose the V j such that all V j ⊆ Q.
Proof. Since f (A) is compact we may choose finitely many sets
D j ⊆ int A′j ⊆ A′j ⊆ V j ⊆ N
such that
f (A) ⊆
[
D j,
j
where D j and A′j are compact, and V j is a chart domain for a chart (V j , ψ j ) on N.
ing the V j we may assume that every V j ⊆ Q. Set
Shrink-
A j := A ∩ f −1 (D j ),
to obtain compact sets that satisfy
[
Aj = A
and
f (A j ) ⊆ D j , for all j.
j
Let N j := j N r (h ◦ f ; A j , (U, φ ), (W, η ), p, ε ), and note that N = j N j . For each j
apply Lemma 2.2 to the maps ψ j ◦ f ◦ φ −1 and η ◦ h ◦ ψ −1
j and the elementary neighborhood
N˜j := N r (η ◦ h ◦ f ◦ φ −1 ; φ (A j ), p, ε )
to obtain elementary neighborhoods
M˜j = N r (ψ j ◦ f ◦ φ −1 ; φ (A j ), δ j ) and M˜j′ = N r (η ◦ h ◦ ψ −1 ; ψ j (A′j ), p, δ j′ )
T
T
j
such that
Γ(M˜j × M˜j′ ) ⊆
N˜j .
The elementary neighborhoods M˜j and M˜j′ of ψ j ◦ f ◦ φ −1 and η ◦h◦ ψ −1
j , respectively,
induce elementary neighborhoods
M j = N r ( f ; A j , (U, φ ), (V j , ψ j ), δ j ) and M j′ = N r (h; A′j , (V j , ψ j ), (W, η ), p, δ j′ )
STRONG TOPOLOGIES FOR SMOOTH MAPS
17
of f and h, respectively. These neighborhoods correspond to each other in the sense that
˜′
fˆ ∈ M j if and only if ψ j ◦ fˆ ◦ φ −1 ∈ M˜j , and ĥ ∈ M j′ if and only if η ◦ ĥ ◦ ψ −1
j ∈ Mj.
Similarly for N˜j and N j . Since Γ(M˜j × M˜j′ ) ⊆ N˜j , one has by the correspondence
described here that Γ(M j × M j′ ) ⊆ N j .
Now just observe that
Γ
\
!
′
Mj ×Mj
j
⊆
\
j
\
Γ M j × M j′ ⊆ N j = N .
j
∞ (M, N) and h ∈ C ∞ (N, X ) and a basic
Lemma 2.4. Consider
smooth maps f ∈ CvS
vS
T
neighborhood U = i∈Λ Ni , where each
Ni = N ri (h ◦ f ; Ai , (Ui , φi ), (Wi , ηi ), pi , εi )
is an elementary neighborhood of h ◦ f .
If { f (Ai )}i∈Λ is locally finite, then there exist basic neighborhoods V and V ′ of f and
h, respectively, such that Γ(V × V ′ ) ⊆ U .
Proof. Since { f (Ai )}i∈Λ is locally finite, there exist compact neighborhoods Qi of f (Ai )
such that {Qi }i∈Λ is locally finite, by [Čec66, 30.C.10]. Here we use our assumption
that finite-dimensional manifolds are σ -compact.
For each i ∈ Λ, Lemma 2.3 implies that there exist
Wi :=
Wi ′ :=
ni
\
j=1
ni
\
N ri ( f ; Ai, j , (Ui , φi ), (Vi, j , ψi, j ), δi, j ),
N ri (h; A′i, j , (Vi, j , ψi, j ), (Wi , ηi ), pi , δi,′ j )
j=1
such that
(1) Γ(Wi × Wi ′ ) ⊆ Ni ,
(2)
(3)
ni
[
j=1
A′i, j
Ai, j = Ai ,
⊆ Vi, j ⊆ Qi for all j.
Then {Ai, j }i, j is locally finite by (2) and since {Ai }i is locally finite, and {A′i, j }i, j is
locally finite by (3) and since {Qi }i is locally finite. Hence
V :=
\
i∈Λ
Wi
and
V ′ :=
\
i∈Λ
Wi ′
STRONG TOPOLOGIES FOR SMOOTH MAPS
18
are basic neighborhoods of f and h, respectively, such that
!
\
\
\
′
′
⊆
Wi × Wi
Γ(Wi × Wi ′ ) ⊆
Ni = U .
Γ V ×V = Γ
i∈Λ
i∈Λ
i∈Λ
Theorem 2.5. Let M and N be finite-dimensional smooth manifolds and let X be a
smooth manifold modeled on a locally convex vector space E. Then the composition
map
∞
∞
Γ : PropvS (M, N) ×CvS
(N, X ) → CvS
(M, X )
sending ( f , h) to h ◦ f is continuous.
∞ (N, E) and a
Proof. It suffices to show that given maps f ∈ PropvS (M, N) and h ∈ CvS
basic neighborhood
U =
\
N ri (h ◦ f ; Ai , (Ui, φi ), (Vi , ψi ), pi , εi )
i∈Λ
∞ (M, X ), there exist basic neighborhoods V and V ′ of f and h, respectively,
of h ◦ f in CvS
such that Γ(V × V ′ ) ⊆ U .
So suppose that we are given f , h and U as above. Then { f (Ai )}i∈Λ is locally finite
since f is proper, by [Eng89, Lemma 3.10.11]. Thus we may apply Lemma 2.4 to
obtain the desired neighborhoods V and V ′ .
Unfortunately, precomposition is not continuous in general as an examination of Example 2.1 reveals. However, precomposition by a proper map is continuous.
Proposition 2.6. Let f ∈ PropvS (M, N). Then the following map is continuous
∞
∞
f ∗ : CvS
(N, X ) → CvS
(M, X ),
h 7→ h ◦ f
∞ (N, X ) → Prop(M, N) × C ∞ (N, X ) given by ι (h) = ( f , h) is
Proof. The map ι f : CvS
f
vS
∗
continuous. Hence f , which is the composition
ιf
Γ
∞
∞
∞
→ PropvS (M, N) ×CvS
(N, X ) −
→ CvS
(M, X ),
CvS
(N, X ) −
is also continuous.
We will now prove that postcomposition is always continuous. This result is needed
even for postcomposition by a map f : X → Y between infinite-dimensional manifolds.
Thus the next proposition can not readily be deduced from Theorem 2.5 (or the other
results in this section)5. Instead we have to take a detour using results on the (coarser)
compact open C∞ -topology (cf. to the proof of Lemma B.2).
5For the finite-dimensional case, a proof along these lines can be found in [Ill03].
STRONG TOPOLOGIES FOR SMOOTH MAPS
19
Proposition 2.7. Let f : X → Y be smooth, where Y is a (possibly infinite-dimensional)
manifold. Then for any finite-dimensional manifold M the following map is continuous
∞
∞
f∗ : CvS
(M, X ) → CvS
(M,Y ),
h 7→ f ◦ h
Proof. To see that f∗ is continuous, we proceed in several steps.
Step 1 Special elementary neighborhoods. Consider first an arbitrary elementary neigh∞ (M,Y ). Since h(A) is compact,
borhood N = N r ( f ◦ h; A, (U, φ ), (V, ψ ), q, ε ) in CvS
S
there are finitely many manifold charts (Wi , κi ) of X with h(A) ⊆ i Wi . Now the
openSsets h−1 (Wi ) cover A and thus there are finitely many compact sets K j such that
A = j K j and h(K j ) ⊆ Wi j . Thus we replace A by the finitely many compact sets. Note
that this will ensure that the families of compact sets considered later remain locally
finite. To shorten the notation, assume without loss of generality that there is a manifold
chart (W, κ ) of X such that h(A) ⊆ W and f (W ) ⊆ V . In particular, we can thus consider
ψ
the mapping fκ := ψ ◦ f ◦ κ −1 : κ (W ) → ψ (V )
Step 2 The preimage of a special elementary neighborhood of f ◦ h is a neighborhood
of h. We work locally in charts. Let Y be modeled on the locally convex space F and X
be modeled on the locally convex space E. Recall that the compact open C∞ -topology
(see [Nee06, Definition I.5.1]) controls the derivatives of functions on compact sets.
Moreover, the elementary neighborhoods of the very strong topology form a subbase of
the compact open C∞ -topology. We denote by C∞ (M, F)co the vector space of smooth
functions with the compact open C∞ -topology.
Choose a compact neighborhood C ⊆ U of A such that h(C) ⊆ W (this entails f ◦ h(C) ⊆
V ). We endow the subset ⌊C, κ (W )⌋ := {g ∈ C∞ (M, E) | g(C) ⊆ κ (W )} with the subspace topology induced by the compact open C∞ -topology. As κ (W ) ⊆ E is open, we
note that ⌊C, κ (W )⌋ is open in C∞ (M, E)co . Now [Glö04, Proposition 4.23 (a)] shows
that
ψ
( fκ )∗ : ⌊C, κ (W )⌋ → C∞ (intC, F)co , h 7→ (ψ ◦ f ◦ κ −1 ) ◦ h
is continuous. Moreover, Nloc := N r (ψ ◦ f ◦ h|intC ; A, (U ∩ intC, φ ), (F, idF ), q, ε ) is
ψ
open in C∞ (intC, F)co . Further, fκ ◦ κ ◦ h|intC = ψ ◦ f ◦ h|intC . Observe that thus κ ◦ h ∈
ψ
(( fκ )∗ )−1 (Nloc). As the elementary neighborhoods form a subbase of the compact
ψ
open C∞ -topology, Lemma B.2 together with continuity of ( fκ )∗ yields
(4)
⌊C, κ (W )⌋ ∩
N
\
ψ
N r (κ ◦ h; Ak , (Uk , φk ), (E, idE ), pk , εk ) ⊆ (( fκ )∗ )−1 (Nloc ).
k=1
Recall from the proof of [Glö04, Proposition 4.23 (a)] that the compact sets Ak are
contained by construction in intC. Thus one easily deduces from (4) that
N 0 (h;C, (U, φ ), (W, κ ), p1, 1) ∩
N
\
k=1
N r (h; Ak , (Uk , φk ), (W, κ ), pk , εk ) ⊆ ( f∗ )−1 (N )
STRONG TOPOLOGIES FOR SMOOTH MAPS
20
Summing up, we see that ( f∗ )−1 (N ) is a neighborhood of h. Further, this finite family
of neighborhoods controls the behavior of mappings only on a pre chosen compact set
C (which depends of course on h).
Step 3 Preimages of basic neighborhoods are open Let M = i∈N Ni be a basic neighborhood of f ◦ h ∈ C∞ (M,Y ) with {Ak }k∈N its the underlying compact family. We will
prove that for arbitrary g ∈ ( f∗ )−1 (M ) the preimage is a neighborhood of g. Choose
with Proposition 1.8 a basic neighborhood of f ◦ g which is contained in M . Replacing
M with this basic neighborhood, it suffices thus to consider the case g = h. Splitting
each Ak as in Step 1 we may assume without loss of generality that each Ni is of the
form considered in Step 2. Use [Čec66, 30.C.10] to construct for every Ak a compact
neighborhood Ck such that {Ck }k∈N is locally finite. Now we proceed for every elementary neighborhood Nk as above (replace C in Step 2 by Ck and shrink Ck if necessary!).
Since the family {Ck }k∈N is locally finite, we thus end up with a basic neighborhood Mh
around h which mapped by f∗ to M . We conclude that ( f∗ )−1 (M) is a neighborhood of
h, whence of every of its elements. Hence preimages of basic neighborhoods under f∗
∞ (M, X ), whence f is continuous.
are open in CvS
∗
T
As an application, we can now identify (as topological spaces) spaces of maps into a
product with products of spaces of mappings to the factors.
Theorem 2.8. Let M be a finite-dimensional manifold, and let X1 and X2 be smooth
manifolds modeled on locally convex vector spaces E1 and E2 , respectively. Then
∞
∞
∞
ι : CvS
(M, X1 × X2 ) → CvS
(M, X1) ×CvS
(M, X2),
f 7→ (pr1 ◦ f , pr2 ◦ f )
is a homeomorphism, where for i ∈ {1, 2} pri : X1 × X2 → Xi is the canonical projection.
Proof. Clearly ι is a bijection, and it is continuous by Proposition 2.7. We will prove
that ι −1 is continuous, i.e. that ι is open. By (3) in Proposition A.3, the set
P := {p ◦ pri : p is a continuous seminorm on Ei }
is a generating family of seminorms on E1 × E2 . Consider a basic neighborhood U =
∞
ri
i∈Λ Ni of f ∈ CvS (M, X1 × X2 ), where each Ni = N ( f ; Ai , (Ui, φi ), (Vi, ψi ), pi , εi ).
By Proposition 1.13 we may assume that each pi ∈ P. Take an arbitrary i ∈ Λ. If
pi = p ◦ pr1 for some continuous seminorm p on E1 , let
T
Mi = N ri (pr1 ◦ f ; Ai , (Ui, φi ), (Vi , ψi ), p, εi )
and
∞
Mi′ = CvS
(M, X2).
If pi = q ◦ pr2 for a continuous seminorm q on E2 , reverse the roles of Mi and Mi′ .
Now suppose without loss of generalization that pi = p ◦ pr1 for some continuous seminorm p on E1 . For g : Rm ⊇ φi (Ai ) → ψi (Vi ) ⊆ E1 × E2 , one has
d(k) g = d(k) (pr1 ◦g, pr2 ◦g) = (d(k) pr1 ◦g, d(k) pr2 ◦g),
STRONG TOPOLOGIES FOR SMOOTH MAPS
21
so the condition pi d(k) g(a; α ) < εi is equivalent to p d(k) (pr1 ◦g)(a; α ) < εi . Hence
Mi × Mi′ = ι (Ni ). Since ι is bijective one has
ι (U ) =
\
ι (Ni ) =
i∈Λ
\
Mi ×
i∈Λ
So ι is open, and a homeomorphism.
\
Mi′ .
i∈Λ
Corollary 2.9. If Q is a compact smooth manifold, then following map is continuous
∞
∞
χ : CvS
(M, X ) → CvS
(Q × M, Q × X ),
f 7→ id × f .
Proof. By Theorem 2.8 it suffices to show that the maps
∞
∞
(M, X ) → CvS
(Q × M, Q)
χ1 : CvS
f 7→ pr1 ◦(id × f )
and
∞
∞
(M, X ) → CvS
(Q × M, X )
χ2 : CvS
f 7→ pr2 ◦(id × f )
are continuous. For (q, m) ∈ Q × M, one has
χ1 ( f )(q, m) = pr1 ◦(id × f )(q, m) = pr1 (q, f (m)) = q = pr1 (q, m),
χ2 ( f )(q, m) = pr2 ◦(id × f )(q, m) = pr2 (q, f (m)) = f (m)
= f ◦ pr2 (q, m) = pr∗2 ( f )(q, m).
The map χ1 is constant in f , hence continuous, and the map χ2 = pr∗2 . Since Q is
compact, pr2 is proper, so Proposition 2.6 implies that χ2 = pr∗2 is also continuous.
3. T HE
FINE VERY STRONG TOPOLOGY
In the end, we would like a structure on C∞ (M, X ) as a locally convex manifold, where
M is a finite-dimensional smooth manifolds and X is a manifold modeled on a locally
convex vector space E, but for this purpose the very strong topology is not fine enough.
A first step in the direction of making C∞ (M, X ) into a locally convex manifold would be
∞ (M, E) itself with the
having a similar structure on C∞ (M, E). One might hope that CvS
vector space structure induced by pointwise operations would be a locally convex vector space. But as Corollary 3.3 points out, this is not the case when E is a (non-trivial)
locally convex vector space and M is a non-compact manifold. However, we will see
∞ (M, E) consisting of maps with compact
in the next section that the subspace of CvS
∞
support, denoted CvS,c (M, E), is a locally convex vector space. Following [Mic80], we
∞ (M, E) to obtain a structure on C ∞ (M, E) as a smooth manirefine the topology on CvS
∞
fold modeled on CvS,c (M, E). The resulting topology on C∞ (M, E), or more generally
C∞ (M, X ), is called the fine very strong topology on C∞ (M, X ). The space C∞ (M, X )
∞ (M, X ).
equipped with the fine very strong topology is denoted CfS
Fortunately, the results of the previous sections are easily extended to hold in the fine
very strong topology. This is done in Proposition 3.9.
STRONG TOPOLOGIES FOR SMOOTH MAPS
22
It is a folklore fact (Proposition C.8) that in the finite-dimensional case, the very strong
topology is equivalent to the D-topology as described in [Mic80, 36].6 Consequently,
the fine very strong topology is equivalent to the F D-topology defined in [Mic80, 40].
Proposition 3.1. Let M be a finite-dimensional smooth manifold and E be a locally
∞ (M, E) which converges in the
convex vector space. Consider a sequence { fn }n∈N ⊆ CvS
very strong topology towards f ∈ C∞ (M, E). Then there exist a compact K ⊆ M and an
N ∈ N such that for all n ≥ N we have
dif( f , fn ) := {y ∈ M : fn (y) 6= f (y)} ⊆ K.
∞ (M, X ), we will show that f cannot be a limit of { f } if for all
Proof. For f ∈ CvS
n
compact K ⊆ M and all N ∈ N there exists n ≥ N such that dif( f , fn ) * K.
Let {An }n∈N be a locally finite exhaustion of MSby compact sets (exists by Lemma B.4
since M is σ -compact), and for n ∈ N set Kn = ni=1 Ai .
Construct a basic neighborhood of f recursively, using the following procedure. Let
n0 = 1, m0 = 1. For i ∈ N, choose ni > ni−1 such that dif( f , fni ) * Kmi−1 . By construction
/ Take any x in this
there exists mi > mi−1 such that dif( f , fni ) ∩ M \ Kmi−1 ∩ Ami 6= 0.
nonempty set. Since f (x) 6= fni (x), there exists a continuous seminorm pi on E such
that 2εi := pi ( fni (x) − f (x)) > 0, and then
fni ∈
/ Ni := N 0 ( f ; Ami , pi , εi ) .
Now U := i∈N Ni is a basic neighborhood of f such that for all N ∈ N there exists
n ≥ N such that fn ∈
/ U . So the sequence { fn }n∈N does not converge to f .
T
Remark 3.2. One can easily prove the proposition above for E a locally convex manifold rather than a locally convex vector space, by “hacking” the compact sets Ai in the
proof into smaller compact sets that are contained in charts.
Corollary 3.3. Let M be a finite-dimensional non-compact manifold and E 6= {0} a
∞ (M, E) with the vector space structure induced by
locally convex vector space. Then CvS
pointwise operations is not a topological vector space.
∞ (M, E) be a non-zero constant map. Then Proposition 3.1 shows that
Proof. Let f ∈ CvS
limλ →0 (λ f ) 6= 0 = (limλ →0 λ ) f , hence scalar multiplication is not continuous.
∞ (M, E) is not a topological vector space, it is a topological
Remark 3.4. Although CvS
∞ (M, R) is a topological ring
group under pointwise addition by Lemma 4.2. And CvS
under the pointwise operations induced by addition and multiplication in R.
6The D-topology was defined using jet bundles (also reviewed in Appendix C). Our treatment of
the topology has the advantage that only elementary arguments are needed. Further, only our approach
generalizes to arbitrary locally convex target manifolds.
STRONG TOPOLOGIES FOR SMOOTH MAPS
23
Definition 3.5 (The fine very strong topology). Define an equivalence relation ∼ on
C∞ (M, X ) by declaring that f ∼ g whenever
supp( f , g) := {y ∈ M : f (y) 6= g(y)}
is compact. Now refine the very strong topology on C∞ (M, X ) by demanding that the
equivalence classes are open in C∞ (M, X ). In other words, equip C∞ (M, X ) with the
topology generated by the very strong topology and the equivalence classes. This is
∞ (M, X ) for C ∞ (M, X ) equipped
the fine very strong topology on C∞ (M, X ). We write CfS
with the fine very strong topology.
Remark 3.6. Here is another way to look at the fine very strong topology. Start with
∞ (M, X ) and equip the equivalence classes [ f ] with the subspace topology. Then
CvS
∞
CfS
(M, X ) =
G
[f]
[ f ]∈C∞ (M,X)/∼
as topological spaces. Taking the family of all sets of the form U ∩ [ f ], where U
runs through the basic neighborhoods in C∞ (M, X ) and [ f ] runs through the equivalence
classes, yields a basis for the fine very strong topology on C∞ (M, X ).
Remark 3.7. If f ∈ C∞ (M, X ) is a proper map and f ∼ fˆ, then fˆ is also proper. Indeed,
if K ⊆ X is compact, then fˆ−1 (K) ⊆ f −1 (K) ∪ supp( f , fˆ). Since closed subspaces of
compact spaces are compact, fˆ−1 (K) is compact.
We would obviously like the results of the previous sections to remain true in the fine
very strong topology. Fortunately, it is easy to extend the results to this case using the
following lemma.
Lemma 3.8. Let T be a topological space, and ζ : T → C∞ (M, X ) a function. If ζ is
∞ (M, X ) and ζ −1 ([ f ]) ⊆ T is open for all equivalence classes
continuous as a map to CvS
∞ (M, X ).
[ f ] ⊆ C∞ (M, X ), then ζ is continuous as a map to CfS
Proof. The map ζ is continuous if preimages of basis elements are open. Basis ele∞ (M, X ) are of the form U ∩ [ f ] for some basic neighborhood U and some
ments for CfS
equivalence class [ f ], and ζ −1 (U ∩ [ f ]) = ζ −1 (U ) ∩ ζ −1 ([ f ]).
Proposition 3.9. Theorem 2.5, Proposition 2.6, Proposition 2.7, Theorem 2.8, and
Corollary 2.9 still hold if we in every case replace the very strong topology with the
fine very strong topology.
In the cases that we consider PropvS (M, N), replace this with PropfS (M, N), by which
∞ (M, N) equipped with the subspace topology.
is meant the subset Prop(M, N) ⊆ CfS
Proof. The proof is case by case. In all cases except for the generalization of Theorem
2.8 and its corollary, it suffices by 3.8 to check that preimages of equivalence classes
STRONG TOPOLOGIES FOR SMOOTH MAPS
24
are open. Unless otherwise stated, letters (such as f or N) are always assumed to have
the same role here as in the statement of the corresponding result.
Theorem 2.5 (the full composition map is continuous). Suppose that f ∼ fˆ and h ∼ ĥ. We
have supp(h ◦ f , ĥ ◦ fˆ) ⊆ supp( f , fˆ) ∪ f −1 supp(h, ĥ) . The right hand side is compact
since f is proper, so supp(h ◦ f , ĥ ◦ fˆ) is a closed subset of a compact space, hence
compact. By definition this means that h ◦ f ∼ ĥ ◦ fˆ.
∞ (M, X ). By what we just observed, if h ◦ f ∼ g
Consider an equivalence class [g] ⊆ CfS
and fˆ ∼ f and ĥ ∼ h, then ĥ ◦ fˆ ∼ g. Hence
Γ−1 ([g]) =
[
[ f ] × [h],
h◦ f ∼g
which is open.
Proposition 2.6 (precomposition is continuous). If h ∼ ĥ, then h ◦ f ∼ ĥ ◦ f by the same
argument as before. So for any equivalence class [g] ⊆ C∞ (M, X ), we have
( f ∗ )−1 ([g]) =
[
[h].
h◦ f ∼g
Proposition 2.7 (postcomposition is continuous). If h, ĥ ∈ C∞ (M, X ), then it is easy to
see that supp( f ◦ h, f ◦ ĥ) ⊆ supp(h, ĥ). So if h ∼ ĥ, then f ◦ h ∼ f ◦ ĥ, since closed
subsets of compact spaces are compact. It follows that for any equivalence class [g] ⊆
S
C∞ (M, X ), we have ( f∗ )−1 ([g]) = f ◦h∼g [h].
Theorem 2.8 (the product theorem). For the same reasons as in the proof of the very
strong version of the theorem, ι is clearly a bijective continuous map. So by Lemma 3.8
it suffices to show that images of equivalence classes are open.
Observe that for f , fˆ ∈ C∞ (M, X1 × X2 ), we have supp( f , fˆ) = supp(pr1 ◦ f , pr1 ◦ fˆ) ∪
supp(pr2 ◦ f , pr2 ◦ fˆ). Hence f ∼ fˆ if and only if pr1 ◦ f ∼ pr1 ◦ fˆ and pr2 ◦ f ∼ pr2 ◦ fˆ.
Another way of stating this fact is ι ([ f ]) = [pr1 ◦ f ] ×[pr2 ◦ f ] for all f ∈ C∞ (M, X1 ×X2 ).
Corollary 2.9. Same proof as in the very strong case.
4. T HE
MANIFOLD STRUCTURE ON SMOOTH VECTOR VALUED FUNCTIONS
Throughout this section, M is a finite-dimensional manifold, E is a locally convex vector
space, and X is a locally convex manifold.
∞ (M, E) with pointwise operations is not a locally conRecall from Corollary 3.3 that CvS
∞ (M, E).
vex vector space, in fact it is not even a topological vector space. Neither is CfS
∞
However, we will in this section make CfS (M, E) into a locally convex manifold. This
STRONG TOPOLOGIES FOR SMOOTH MAPS
25
∞ (M, X ) into a locally convex manifold (but we will not
is a first step towards making CfS
∞ (M, E) as a locally convex manifold is C ∞ (M, E),
do this). The modeling space for CfS
vS,c
defined below.
∞ (M, E) to be the subspace of C ∞ (M, E) consisting of the
Definition 4.1. We define CvS,c
vS
functions with compact support, i.e.
∞
∞
CvS,c
(M, E) = { f ∈ CvS
(M, E) : supp( f , 0) is compact}
∞ (M, E). Note that C ∞ (M, E) = [0] in
equipped with the subspace topology from CvS
vS,c
∞ (M, E).
CfS
∞ (M, E) with pointwise operations is a locally
As a first step towards proving that CvS,c
∞
convex vector space, we show that C (M, E) with pointwise addition is a topological
group in the very strong and fine very strong topologies.
Lemma 4.2. Addition
Σ : C∞ (M, E) ×C∞ (M, E) → C∞ (M, E),
( f , g)
7→ f + h = [m 7→ f (m) + h(m)]
is continuous when C∞ (M, E) is equipped with the very strong topology or fine very
strong topology.
Proof. We prove the assertion only for the very strong topology as the proof carries
over verbatim to the fine very strong topology. By Theorem 2.8 there is a canonical
∞ (M, E) × C ∞ (M, E) ∼ C ∞ (M, E × E). Since addition S : E ×
homeomorphism ι : CvS
= vS
vS
∞ (M, E × E) → C ∞ (M, E) is
E → E in E is smooth, induced postcomposition S∗ : CvS
vS
continuous. Hence Σ = S∗ ◦ ι is continuous.
∞ (M, E)
Once we have established the following proposition, it will be easy to make CfS
∞
into a locally convex manifold modeled on CvS,c (M, E). The hard work lies here.
∞ (M, E) with vector space structure induced
Proposition 4.3. The topological space CvS,c
by pointwise operations in E is a locally convex vector space.
Proof. In Lemma 4.2 we showed that addition is continuous, and the topological space
∞ (M, E) is Hausdorff since the compact open C ∞ -topology on C ∞ (M, E) is Hausdorff
CvS,c
and the very strong topology is finer than the compact open C∞ -topology. It is therefore
only necessary to check that scalar multiplication is continuous in order to conclude that
∞ (M, E) is a topological vector space. Finally, we must verify that this topological
CvS,c
vector space is locally convex.
∞ (M, E) → C ∞ (M, E), (λ , f ) 7→ λ f is continuous.
Scalar multiplication µ : R × CvS,c
vS,c
∞ (M, E), and consider a basic neighborhood V = T
Let (λ , f ) ∈ R ×CvS,c
i∈Λ Ni of λ f ,
r
i
where each Ni = N (λ f ; Ai , (Ui , φi ), pi , εi ) is an elementary neighborhood of λ f . We
∞ (M, E) such that µ (I ×U ) ⊆ V .
will show that there exists open sets I ⊆ R and U ⊆ CvS
STRONG TOPOLOGIES FOR SMOOTH MAPS
26
Since supp( f , 0) is compact, only finitely many Ai intersect supp( f , 0), say only for
i = i1 , . . ., in . Define ε := min(εi1 , . . . , εin ).
o
n
−1
Set m1 := max sup1≤ j≤n k f ◦ φi j k(ri j , φi j (Ai j ), pi j ), 1 and
ε
ε
1
I := B ε (λ ) = λ −
,λ +
.
2m1
2m1
2m1
Define m2 := sup{|t| : t ∈ I}, and set U :=
T
i∈Λ N
ri
εi
f ; Ai , (Ui, φi ), pi , 2m
. Suppose
2
(λ ′ , f ′ ) ∈ I × U . For all i ∈ Λ, x ∈ Ai , 1 ≤ k ≤ ri , and α ∈ {e1 , . . . , edim M }k , we have
−1
−1
(k) ′ ′
pi d (λ f ◦ φi − λ f ◦ φi )(φi (x); α )
≤|λ ′ |pi d(k) ( f ′ ◦ φi−1 − f ◦ φi−1 )(φi (x); α ) + |λ ′ − λ |pi d(k) ( f ◦ φi−1 )(φi (x); α )
εi
ε
(k)
−1
pi d ( f ◦ φi )(φi (x); α ) =: C.
< +
2 2m1
If i ∈
/ {i1 , . . ., in }, then pi d(k) ( f ◦ φi−1 )(φi (x); α ) = 0, in which case C ≤ εi . And if
i ∈ {i1 , . . . , in }, then ε ≤ εi and pi d(k) ( f ◦ φi−1 )(φi (x); α ) ≤ m1 , in which case we still
have C ≤ εi . Hence λ ′ f ′ ∈ V , and µ (I × U ) ⊆ V . Consequently, µ is continuous.
∞
The space is locally convex. We have now established that CvS,c
(M, E) is a topological
vector space. It remains to see that this topological vector space is locally convex. For
r ∈ N0 , (U, φ ) a chart on M, A ⊆ U compact, and p a continuous seminorm on E, define
∞
(M, E) → [0, ∞),
k · k(r, A, (U, φ ), p) : CvS,c
f 7→ k f ◦ φ −1 k(r, φi (Ai ), p)
∞
This is a seminorm on CvS,c
(M, E). Consider a family {k · k(ri , Ai , (Ui , φi ), pi )}i∈Λ
of such seminorms, where {Ai }i∈Λ is locally finite. For some family {εi }i∈Λ define
∞
q : CvS,c
(M, E) → [0, ∞) by
q( f ) = sup εi k f k(ri , Ai , (Ui , φi ), pi ).
i∈Λ
∈ C∞
Every f
vS,c (M, E) has compact support, so supp( f , 0) intersects only finitely many
of the Ai , from which it follows that k f k(ri , Ai , (Ui, φi ), pi ) 6= 0 for only finitely many
i ∈ Λ. Hence q( f ) < ∞, so q is well-defined. Clearly q is a seminorm. Also q is
continuous as for all λ > 0, the preimage q−1 [0, λ ) is a basic neighborhood of 0, e.g.
!
q−1 [0, 1) =
\
∞
(M, E).
N ri (0; Ai , (Ui , φi ), pi , εi ) ∩CvS,c
i∈Λ
So every basic neighborhood of 0 arises as a preimage of a continuous seminorm. Con∞ (M, E) is locally convex (see [Köt69, §18]).
sequently, CvS,c
STRONG TOPOLOGIES FOR SMOOTH MAPS
27
∞ (M, E) as an
We will now provide an alternative description of the topology on CvS,c
inductive limit of certain locally convex spaces. This characterization also implies that
∞ (M, E) is a locally convex space (thus providing an elegant proof of Proposition
CvS,c
4.3). Note however: Though the proof of Proposition 4.3 is a bit cumbersome, it is also
completely elementary and does not use auxiliary results on inductive limits.
Definition 4.4. Let K ⊆ M be a compact subset and E be a locally convex space. Then
we define
CK∞ (M, E) := { f ∈ C∞ (M, E) | supp( f , 0) ⊆ K}
and topologize this space with the compact open C∞ -topology, i.e. the topology generated by the subbase N ∩CK∞ (M, E) where N runs through all elementary neighbor∞ (M, E). Recall from [Glö04, Proposition 4.19] that C ∞ (M, E) is a locally
hoods of CvS
K
convex vector space.
Remark 4.5. Since all functions CK∞ (M, E) have compact support contained in K one
can prove that the compact open C∞ -topology coincides with the subspace topologies
∞ (M, E) and C ∞ (M, E). However, we will not need this.
induced by CvS
fS
∞ (M, E) =
Denote by K (M) the set of compact subsets of M. Observe that as sets CvS,c
S
∞
K∈K (M) CK (M, E). We claim that the topology on the compactly supported functions
is determined by the smaller locally convex spaces: To see this, recall that with respect to
inclusion, K (M) is a directed set. Further, for K, L ∈ K (M) with K ⊆ L the canonical
inclusion ιKL : CK∞ (M, E) → CL∞ (M, E) is continuous linear by definition of the topology.
Hence we can form the locally convex inductive limit lim CK∞ (M, E) (cf. [Köt69, §19 3.])
→
of the family {CK∞ (M, E)}K (M) (with respect to the canonical inclusions).
Lemma 4.6. Let E be a locally convex space, then as locally convex spaces
∞
CvS,c
(M, E) = lim CK∞ (M, E).
→
∞ (M, E) = lim C ∞ (M, E), we only have to prove that the topoloProof. Since as sets CvS,c
→ K
gies coincide. However, since M is σ -compact, [Glö04, Proposition 8.13 (d)] implies
∞ (M, E) = lim C ∞ (M, E) is given by
that a basis for the inductive limit topology on CvS,c
→ K
the basic neighborhoods of the very strong topology.
∞ (M, E) define φ : [ f ] → C ∞ (M, E) by
Proposition 4.7. For each class [ f ] in CfS
[f]
vS,c
∞ (M, E). Hence
φ[ f ] (g) = g − f . Then A = {(φ[ f ] , [ f ])} f ∈C∞(M,E) is a smooth atlas for CfS
∞ (M, E) is a smooth manifold modeled on C ∞ (M, E).
CfS
vS,c
Proof. We will first show that every chart φ[ f ] is a homeomorphism. First of all, note
that φ[ f ] is well-defined since g − f is smooth and compactly supported for g ∈ [ f ]. It is
−1
bijective with inverse φ[−1
f ] (h) = h + f . Both φ[ f ] and φ[ f ] are continuous by Lemma 4.2.
STRONG TOPOLOGIES FOR SMOOTH MAPS
28
∞ , whence we have to check that chart transformations
The chart domains of A cover CfS
/ Then [ f ] = [g] and
are smooth. Let (φ[ f ] , [ f ]) and (φ[g] , [g]) be charts with [ f ] ∩ [g] 6= 0.
−1
∞ (M, E) is so.
φ[g] ◦ φ[ f ] (h) = h + f − g, whence it is smooth in h as addition in CvS,c
∞ (M, E) is just a collection of (affine) copies of C ∞ (M, E).
Structurally, the manifold CfS
vS,c
For this reason, it is also called in [Mic80] a local topological affine space.
∞ (M, X ) for an arbitrary locally convex manifold
To construct a manifold structure on CfS
X one needs a so called local addition on X (cf. [Mic80, KM97]). A local addition
replaces the vector space addition. It allows to “smoothly choose” charts on X (see
[Sta13] for more information). The details are similar to [Mic80, Section 10] but require
certain analytical tools (e.g. a suitable version of the Ω-Lemma, [Glö04, Appendix F])7.
5. A PPLICATION
TO BISECTION GROUPS
In this section we use our results on the very strong and the fine very strong topology to
turn certain groups into topological groups. The groups envisaged here are the bisection
groups associated to certain Lie groupoids. A reference on (finite-dimensional) Lie
groupoids is [Mac05], see [SW15a, SW15b] for infinite-dimensional Lie groupoids.
Definition 5.1 (Lie groupoid). Let M be a finite-dimensional smooth manifold and
G a smooth manifold modeled on a locally convex vector space. Then a groupoid
G = (G ⇒ M) with source projection α : G → M and target projection β : G → M is a
(locally convex) Lie groupoid if α and β are smooth submersions (i.e. locally projections), partial multiplication m : G ×α ,β G → G is smooth, object inclusion 1 : M → G
is smooth, and inversion i : G → G is smooth.
Definition 5.2 (Bisection group). The group of bisections Bis(G ) of a Lie groupoid
G = (G ⇒ M) is the set of sections σ : M → G of α such that β ◦ σ is a diffeomorphism
of M. The group operation ⋆ is given by
(σ ⋆ τ )(x) := σ ((β ◦ τ )(x))τ (x).
With this operation, the object inclusion 1 : M → G becomes the neutral element and
the inverse of a section σ is σ −1 (x) = i(σ ((β ◦ σ )−1 (x))).
Example 5.3.
(1) For a finite-dimensional manifold M, the unit Lie groupoid is the
groupoid (M ⇒ M) with both source and target projection idM . The bisection
group of this groupoid is trivial.
7To apply the Ω-Lemma as stated in [Glö04], one needs a topology on spaces of compactly supported
sections in vector bundles. In ibid. the compact open C∞ -topology is used, however by arguments similar
to Lemma 4.6 one proves that this topology coincides with the very strong topology.
STRONG TOPOLOGIES FOR SMOOTH MAPS
29
(2) Let M be a finite-dimensional smooth manifold. Then P(M) := (M × M ⇒ M)
with source projection α = pr2 and target projection β = pr1 is a Lie groupoid.
Multiplication in the groupoid is given by (x, y)(y, z) = (x, z). Postcomposition
β∗ induces an isomorphism Bis(P(M)) ∼
= Diff(M) of groups, where Diff(M) is
the group of smooth diffeomorphisms of M.
(3) Suppose G is a locally convex Lie group, and ∗ is the one-point space. Then
(G ⇒ ∗) is a Lie groupoid with bisection group G.
To prepare the construction of a topological group structure on bisection groups, recall
the following facts on diffeomorphism groups of finite-dimensional manifolds.
Remark 5.4. Let M be a finite-dimensional manifold and Diff(M) be the group of
smooth diffeomorphisms of M. As Diff(M) ⊆ C∞ (M, M), we can endow Diff(M) either with the subspace topology induced by the very strong topology (write DiffvS (M))
or with respect to the fine very strong topology (we write DifffS (M)). Now as a consequence of [Mic80, Corollary 7.7] and Proposition C.8, both DiffvS (M) and DifffS (M)
are topological groups. Note that DifffS (M) is even a locally convex Lie group by
[Mic80, Theorem 11.11]. In particular, we remark that the (subspace topology induced
by the) fine very strong topology is the Lie group topology of Diff(M).
Proposition 5.5. If Bis(G ) is equipped with the subspace topology with respect to
∞ (M, G) or C ∞ (M, G), then Bis(G ) becomes a topological group.
CvS
fS
Proof. We will prove that Bis(G ) becomes a topological group when equipped with
∞ (M, G). The case where we consider the
the subspace topology with respect to CvS
∞ (M, G) can be proven identically, since we only
subspace topology with respect to CfS
use results that hold in both topologies. Let Ω : Bis(G ) × Bis(G ) → Bis(G ) be the
multiplication map defined by Ω(σ , τ ) = σ ⋆ τ , and let ι be the inclusion Bis(G ) →
∞ (M, G). Observe that we can write
CvS
Ω(σ , τ )(x) = σ ((β ◦ τ )(x))τ (x) = m(Γ(β ◦ τ , σ )(x), τ (x)).
So ι ◦ Ω can be written as a composition of continuous maps; the diagram
Bis(G ) × Bis(G )
ι ◦Ω
(β∗ ◦pr2 ,ι ×ι )
/ C ∞ (M, G)
vS O
∞ (M, G) ×C ∞ (M, G)
PropvS (M, M) ×CvS
vS
m∗
Γ×id
∞ (M, G) ×C ∞ (M, G)
CvS
vS
∼
=
/ C ∞ (M, G × G)
vS
commutes. Here we have used that β∗ (Bis(G )) ⊆ Diff(M) ⊆ Prop(M, M) by definition of bisections. All of the maps represented by normal arrows in the diagram are
continuous by results in the previous sections. Since ι ◦ Ω is continuous, so is Ω.
STRONG TOPOLOGIES FOR SMOOTH MAPS
30
Let Φ : Bis(G ) → Bis(G ) be the inversion map. Inversion Inv : DiffvS (M) → DiffvS (M)
is continuous by [Mic80, Theorem 7.6] and Proposition C.8. The diagram
ι ◦Φ
Bis(G )
/ C ∞ (M, G)
vS O
β∗
DiffvS (M)
i∗
Inv /
DiffvS (M)
σ∗
/ C ∞ (M, G)
vS
commutes as Φ(σ )(x) = i(σ ((β ◦ σ )−1 (x))) = (i∗ (σ∗ (Inv(β∗ (σ ))))(x). All maps represented by normal arrows are continuous. Thus ι ◦ Φ and also Φ are continuous.
As noted in Remark 5.4, Diff(M) is a topological group with respect to the subspace
topologies induced by the (fine) very strong topology on C∞ (M, M). Thus we obtain the
following morphisms of topological groups.
Corollary 5.6. The target projection β : G → M of a locally convex Lie groupoid G =
(G ⇒ M) induces a map β∗ : Bis(G ) → Diff(M) given by postcomposition. This is a
homomorphism of topological groups with respect to the very strong and fine very strong
topologies on both groups.
∞ (M, G) → C ∞ (M, M) is continuous, so is the (co)restriction of β
Proof. Since β∗ : CvS
∗
vS
to Bis(G ) and Diff(M). The same argument holds in the fine very strong topology. The
map β∗ is also a group homomorphism, since
(β∗ (σ ⋆ τ )) (x) = β (σ ((β ◦ τ )(x))τ (x)) = β (σ (β ◦ τ )(x)) = (β∗ (σ ) ◦ β∗(τ )) (x).
The results of this section enable the construction of a Lie group structure on Bis(G ).
It is worth noting that the key step in constructing the Lie group structure is sorting
∞ (M, G)
out the topology of the function spaces. Using the manifold structure on CfS
(see comments in Section 4) one establishes the smoothness of joint composition and
postcomposition with respect to these structures. Since Theorem A and the Ω-Lemma
[Glö04, Appendix F] are at our disposal, one can copy exactly the arguments from the
finite-dimensional case outlined in [Mic80, §10 and §11]. After that one can proceed
as in [SW15a] and establish smoothness of the group operations following the proof of
Proposition 5.5. Again, results of this type are beyond the scope of the present paper.
STRONG TOPOLOGIES FOR SMOOTH MAPS
A PPENDIX A. C ALCULUS
31
IN THE LOCALLY CONVEX SETTING
In this appendix we lay down the definitions, notation and conventions used throughout
this article regarding locally convex vector spaces, infinite-dimensional manifolds, and
smooth maps between such objects.
Locally convex vector spaces.
Definition A.1.
(1) A real vector space E is a topological vector space if E is
equipped with a Hausdorff topology turning both addition and scalar multiplication into continuous maps.
(2) A topological vector space E is called locally convex if every 0-neighborhood
contains a convex 0-neighborhood.
Particularly useful for our purposes (constructing the very strong topology on a space
of smooth functions into a locally convex vector space) is the perspective of a locally
convex vector topology as generated by a suitable family of seminorms on the space.
Definition A.2. Let E be a locally convex vector space.
(1) A seminorm on E is a map p : E → [0, ∞) such that p(x + y) ≤ p(x) + p(y) and
p(λ x) = |λ |p(x) for all x, y ∈ E and all λ ∈ R.
(2) A family P of continuous seminorms on E is called generating if E has the
initial topology with respect to P.
p
(3) For a seminorm p on a vector space E, x ∈ E and ε > 0, we write Bε (x) := {y ∈
p
E : p(y − x) < ε } and Bε (x) := {y ∈ E : p(y − x) ≤ ε }, called open and closed
seminorm balls, respectively.
We list some properties of seminorms and families of such that are particularly useful
for us.
Proposition A.3 (§18 in [Köt69]). Let E be a locally convex vector space.
(1) The family of all continuous seminorms on E is generating.
(2) A family P of continuous seminorms on E is generating if and only if for any
continuous seminorm p on E there exist p1 , . . . , pn ∈ P and a c ∈ R such that
p ≤ c sup pi .
1≤i≤n
(3) Let {Ei }i∈I be a family of locally convex vector spaces. The product ∏i∈I Ei of
vector spaces is a locally convex
vector space when equipped with the product
S
topology. Moreover, the set i∈I {p ◦ pri : p is a continuous seminorm on Ei } is
a generating family of seminorms for ∏i∈I Ei .
STRONG TOPOLOGIES FOR SMOOTH MAPS
32
Smooth maps between locally convex vector spaces. Differential calculus in the
locally convex setting (in our sense) is sometimes known as Bastiani calculus (see
[Bas64]) or Keller Cr -theory (see [Kel74]). The idea is to define the derivative of a
map between locally convex vector spaces using directional derivatives.
Definition A.4. Let E and F be locally convex vector spaces, U ⊆ E open, and f : U →
F a map. If it exists, we define for (x, h) ∈ U × E the directional derivative
d f (x, h) := Dh f (x) := lim
t→0
For r ∈ N ∪ {∞} we say that f
is Cr
f (x + th) − f (x)
.
t
if
d(k) f (x; y1 , . . . , yk ) := Dyk Dyk−1 · · · Dy1 f (x)
exists for all k with 1 ≤ k ≤ r and all yi ∈ E, the d(k) f : U × E k → F define continuous
maps, and f is continuous. We say f is smooth if it is C∞ .
Elsewhere, d(k) f (x; y1 , . . ., yn ) is often denoted
∂k
f (x)
∂ yk · · · ∂ y1
∂
f (x), where y = (y1 , . . ., yk ).
∂y
or
Here is an alternative, but equivalent condition for when a map is Cr for r ≥ 2.
Lemma A.5 (Lemma 1.14 in [Glö02]). Suppose f is C1 in the sense of Definition A.4.
Then f is a Cr -map if and only if d f : U × E → F is Cr−1 . In this case, we define dk f
for k ≤ r recursively by
k −1
dk f := dk−1 (d f ) : U × E 2
→ F.
By convention d0 f = d(0) f = f . Note that whereas d(k) f is obtained by taking iterated
directional derivatives of f with respect to the first variable, dk f is the derivative of
dk−1 f with respect to all variables. The dk f and d(k) f determine each other:
Proposition A.6 (Proof of Lemma 1.14 in [Glö02]). Let k ∈ N. Then
(1) For each j ∈ {1, . . ., k} and each integer j-tuple α with 1 ≤ α1 < · · · α j ≤ 2k − 1
there exists nα ∈ N0 such that for every Ck -map f : E ⊇ U → F, we have
dk f (x; y1 , . . . , y2k −1 ) =
k
∑ ∑
j=1 α1 <···<α j
nα d( j) f (x; yα1 , . . ., yα j )
for all x ∈ U and all y1 , . . . , y2k −1 ∈ E.
(2) There exist numbers α1 < · · · < αk in {1, . . ., 2k − 1} such that for every Ck -map
f : E ⊇ U → F, we have
d(k) f (x; y1 , . . . , yk ) = dk f (x; ια (y1 , . . ., yk ))
STRONG TOPOLOGIES FOR SMOOTH MAPS
33
k
for all x ∈ U and all y1 , . . ., yk ∈ E k . Here, ια : E k → E 2 −1 is the inclusion
sending yi to position αi , i.e. pr j ια (y1 , . . . , yk ) = yi if j = αi for some i and 0
otherwise.
Proposition A.7 (1.12, 1.13, and 1.15 in [Glö02]). Let E, F and G be locally convex
vector spaces, U ⊆ E and V ⊆ F open, f : U → F and g : V → G maps such that
f (U ) ⊆ V , and r ∈ N ∪ {∞}.
(1) If f and g are Cr -maps, then g ◦ f : U → G is Cr .
(2) If f and g are C1 -maps, then d(g ◦ f )(x, y) = d g( f (x), d f (x, y)) for all x ∈ U
and all y ∈ E.
(3) If f is Cr and x ∈ U , then d(k) f (x; ·) : E k → F is a symmetric k-linear map for
every k ≤ r.
The chain rule (2) is difficult to work with directly for higher order derivatives. But
there is a way to circumvent this problem, by defining a map T j f in such a way that we
have the identity d j (g ◦ f ) = d j g ◦ T j f .
Definition A.8. If f : E ⊇ U → F is Ck , we define
T f : U × E → F × F,
(x, y) 7→ ( f (x), d f (x, y))
j
j
Note that T f is Ck−1 . For j ≤ k define T j f := T(T j−1 f ) : U × E 2 −1 → F 2 . We also
j
define T̃ f as the projection of T j f onto the last 2 j − 1 coordinates, so that
j
T j f (x; y1 , . . . , y2 j −1 ) = ( f (x), T̃ f (x; y1 , . . ., y2 j −1 ).
Manifolds modeled on locally convex vector spaces.
Definition A.9. A C0 -manifold modeled on a locally convex vector space E is a Hausdorff topological space X which is locally homeomorphic to E.
By locally homeomorphic to E it is meant that any point x ∈ X has an open neighborhood
that is homeomorphic to an open set in E.
Remark A.10. As opposed to the standard definition of finite-dimensional manifolds,
we don’t require that manifolds are σ -compact. However, we always assume that finitedimensional manifolds are σ -compact. For locally compact and locally metrizable
Hausdorff spaces (which includes finite-dimensional manifolds), being σ -compact is
equivalent to being second countable.
Just like in the finite-dimensional case, a Cr structure on an infinite-dimensional manifold X is given by a choice of a Cr -atlas, or more precisely an equivalence class of
such.
STRONG TOPOLOGIES FOR SMOOTH MAPS
34
Definition A.11. Let X be a C0 -manifolds modeled on E. A Cr -atlas for X is a collection
A = {(φ ,U ) : U ⊆ X open, U ′ ⊆ E open, and φ : U → U ′ a homeomorphism}
/
such that for any two elements (φ ,U ), (ψ ,V ) ∈ A (called charts) for which U ∩V 6= 0,
−1
′
r
the transition map ψ ◦ φ|φ (U∩V ) : φ (U ∩V ) → V is a C -map.
We define an equivalence class on the set of Cr -atlases of X by declaring that A ∼ A ′ if
A ∪ A ′ is a Cr -atlas for X . Using (1) and (2) in Proposition A.7 one may check that this
is indeed an equivalence relation. Given a specified equivalence class [A ] of Cr -atlases
for X , a chart in [A ] (or simply a chart for X ) is a chart in A ′ for some A ′ ∈ [A ].
Definition A.12. A Cr -manifold X modeled on a locally convex vector space E (or just
a locally convex Cr manifold) is a C0 -manifold modeled on E together with a chosen
equivalence class of Cr -atlases for X .
Definition A.13 ( Cr -map). Let f : X → Y be a map. Then f is a Cr -map if for all x ∈ X
and all pairs of charts (U, φ ) and (V, ψ ) on X and Y , respectively, such that x ∈ U and
f (x) ∈ V , the composition ψ ◦ f ◦ φ|−1
: φ ( f −1 (V )) → ψ (V ) is a Cr -map. Denote
φ ( f −1 (V ))
the set of all Cr -maps X → Y by Cr (X ,Y ).
Products of manifolds, tangent spaces etc. can be defined for manifolds modeled on
locally convex spaces as in the finite-dimensional setting. More information on this and
on locally convex Lie groups (see below) can be found in [Nee06].
Definition A.14 (Locally convex Lie group). Let G be a smooth manifold modeled on a
locally convex vector space which is also a group. If these structures are compatible in
the sense that group multiplication G × G → G and inversion G → G are smooth, then
G is a locally convex Lie group.
A PPENDIX B. AUXILIARY
RESULTS FOR
S ECTIONS 1
AND
2
In this appendix we provide details and auxiliary results for the first two sections.
Lemma B.1. Let m, n ∈ N, r ∈ N0 , A ⊆ Rm compact, E a locally convex vector space,
p a continuous seminorm on E, and R > 0.
Then there exists K > 0 such that for all smooth maps f : Rm → Rn and g : Rn → E for
which k f k(r, A) ≤ R, one has kg ◦ f k(r, A, p) ≤ Kkgk(r, f (A), p)k f k(r, A).
Proof. The proof is naturally divided into two parts.
Part 1. Our first goal is to control the size of Tk f (see Definition A.8). Let m, n and r
be given. The precise statement we want to show is that for all k with 0 ≤ k ≤ r and all
STRONG TOPOLOGIES FOR SMOOTH MAPS
35
a ∈ A, and for all z1 , z2 , . . ., z2k −1 ∈ {0, e1 , . . . , em } one has
k Tk f (a; z1 , z2 , . . ., z2k −1 )k∞ ≤ ck f k(r, A) =: C,
where c ≥ 1 is a constant (depending only on r, m and n ) such that
k dk f (a; z1, . . . , z2k −1 )k∞ ≤ ck f k(r, A).
Such a constant c exists by Proposition A.6. We prove the seemingly more general
statement that for all non-negative integers i and j such that 0 ≤ i + j ≤ r one has
k di T j f (a; z1 , . . . , z2i+ j −1 )k∞ ≤ C,
(5)
by induction on i + j. When i + j = 0, both i = 0 and j = 0, so the left hand side in
inequality (5) is k f (a)k∞ ≤ C, by the choice of c.
Now let k be a positive integer with 0 ≤ k ≤ r − 1 and assume that inequality (5) holds
for all non-negative integers i and j with i + j ≤ k. If i + j = k + 1 and 1 ≤ j, then
k di T j f (a; z1 , . . . , z2i+ j −1 )k∞
=k di (T j−1 f (a; z1 , . . ., z2i+ j−1 −1 ), d T j−1 f (a; z1 , . . ., z2i+ j −1 ) k∞
= di T j−1 f (a; z1, . . . , z2i+ j−1 −1 ), di+1 T j−1 f (a; z1 , . . . , z2i+ j −1 )
i
≤ max k d T
j−1
i+1
f (a; z1 , . . ., z2i+ j−1 −1 )k∞ , k d
T
j−1
∞
f (a; z1, . . . , z2i+ j −1 )k∞ ,
by definition of T and the formula d(h1, h2 ) = (d h1 , d h2 ). By the induction hypothesis,
k di T j−1 f k∞ ≤ C, so it remains to check that k di+1 T j−1 f k∞ ≤ C. This last inequality is
just what we started with, but with i increased by one and j decreased by one. Iterating
the argument above, we are left with the case i = k + 1 ≤ r and j = 0. This case holds
by the choice of c.
Part 2. Given 0 ≤ k ≤ r, a ∈ A, and y1 , . . ., yk ∈ {e1 , . . . , em } ⊂ Rm , we would like to
control the size of d(k) (g ◦ f )(a; y1, . . ., yk ). By Proposition A.6 there exist z1 , . . . , z2k −1 ∈
{0, y1 , . . . , yk } such that
d(k) (g ◦ f )(a; y1, . . ., yk ) = dk (g ◦ f )(a; z1, . . ., z2k −1 )
k
= dk g ◦ Tk f (a; z1, . . . , z2k −1 ) = dk g( f (a); T̃ f (a; z1 , . . ., z2k −1 )
k
=∑
∑
j=1 1≤α1 <···<α j ≤2k −1
k
nα d( j) g( f (a); prα (T̃ f (a; z1, . . . , z2k −1 ))),
where the second sum is taken over all increasing j-tuples α = (α1 , . . ., α j ), each of the
nα are fixed numbers as in Proposition A.6, prα (t1 , . . .,t2k −1 ) = (tα1 , . . . ,tα j ), and T̃ is
as in Definition A.8. There exist λ j,i ∈ R such that
!
k
T̃ f (a; z1, . . . , z2k −1 ) =
n
n
i=1
i=1
∑ λ1,iei, . . . , ∑ λ2k −1,iei .
STRONG TOPOLOGIES FOR SMOOTH MAPS
36
Note that every |λi, j | ≤ C by part 1 of the proof and our choice of norm k · k∞ ! We have
k
d( j) g( f (a); prα T̃ f (a; z1 , . . . , z2k −1 )
∑
=
λα1 ,i1 λα2 ,i2 · · · λα j ,i j d( j) g( f (a); ei1 , . . ., ei j )
(i1 ,...,i j )∈[n] j
by (3) in Proposition A.7. Recall that p is a seminorm on E. From the preceding
paragraphs we have
p d(k) (g ◦ f )(a; y1, . . . , yk )
k
=p ∑
∑
j=1 1≤α1 <···<α j ≤2k −1
≤
≤
∑
j,α ,i
nα
∑
(i1 ,...,i j )∈[n] j
λα1 ,i1 λα2 ,i2 · · · λα j ,i j d( j) g( f (a); ei1 , . . . , ei j )
|nα ||λα1,i1 | · · · |λα j ,i j |p d( j) g( f (a); ei1 , . . ., ei j )
∑ |nα |C j
j,α ,i
!
kgk(r, f (A), p) ≤
∑ |nα |c j R j−1
j,α ,i
!
k f k(r, A)kgk(r, f (A), p),
where the last inequality holds since C = ck f k(r, A) ≤ cR (recall that k f k(r, A) ≤ R by
assumption). Setting K = ∑ |nα |c j R j−1 , we obtain
(k)
p d (g ◦ f )(a; y1, . . . , yk ) ≤ Kk f k(r, A)kgk(r, f (A), p).
By definition of kg ◦ f k(r, A, p) the assertion follows.
We now prove that for spaces of smooth functions to locally convex spaces it suffices to
consider basic neighborhoods with respect to the identity charts.
Lemma B.2. Let M be a finite-dimensional manifold and E be a locally convex space.
Consider the topology T on C∞ (M, E) generated by basic neighborhoods which arise
from elementary neighborhoods of the form N r ( f ; A, (U, φ ), (E, idE ), q, ε ). Then T
coincides with the very strong topology on C∞ (M, E).
Proof. Restricting our choice of charts on E to the canonical chart clearly generates a
topology which is coarser than the very strong topology.
To see that it is also finer, consider first an arbitrary elementary neighborhood N =
∞ (M, E). Recall that the compact open C ∞ -topology
N r ( f ; A, (U, φ ), (V, ψ ), q, ε ) in CvS
(see [Nee06, Defintion I.5.1]) allows us to control the derivatives of functions on compact sets. We denote by C∞ (M, E)co the vector space with this topology and note that the
elementary neighborhoods of the very strong topology form a subbase of the compact
open C∞ -topology.
STRONG TOPOLOGIES FOR SMOOTH MAPS
37
Now choose a compact neighborhood C ⊆ U of A such that f (C) ⊆ V . We endow the
subset ⌊C, ψ (V )⌋ := {g ∈ C∞ (M, E) | g(C) ⊆ ψ (V )} with the subspace topology induced
by the compact open C∞ -topology. As ψ (V ) ⊆ E is open, we note that ⌊C, ψ (V )⌋ is open
in C∞ (M, E)co. Now [Glö04, Proposition 4.23 (a)] shows that
(ψ −1 )∗ : ⌊C, ψ (V )⌋ → C∞ (intC, E)co ,
h 7→ ψ −1 ◦ h
is continuous. Moreover, the set NC := N r (ψ ◦ f |intC ; A, (U ∩ intC, φ ), (E, idE ), q, ε ) is
open in C∞ (intC, E)co . Using the description of subbasic neighborhoods of the compact
open C∞ -topology in [Glö02], we obtain elementary neighborhoods
(6)
N
\
N r ( f ; Ak , (Uk , φk ), (E, idE ), qk , ε ) ⊆ ((ψ −1 )∗ )−1 (NC ) = N ∩ ⌊C,V ⌋ ⊆ N .
k=1
Observe that the compact sets Ak are contained by construction in intC.8 Summing up,
(6) together with Lemma 1.7 shows that every elementary neighborhood in C∞ (M, E) is
open in T .
It is now
easy to prove that every basic neighborhood is open in T . To this end let
T
M = k∈N Nk be a basic neighborhood around f ∈ C∞ (M, E) with {Ak }k∈N its the underlying compact family. Use [Čec66, 30.C.10] to construct for every Ak a compact
neighborhood Ck such that {Ck }k∈N is locally finite. Now we proceed for every elementary neighborhood Nk as above (working with Ck !). Since the family {Ck }k∈N is locally
finite, we thus end up with a basic neighborhood around f which is contained in T . As
all basic neighborhoods of the very strong topology are open in T , the topology T is
finer as the very strong topology. This proves the assertion.
Lemma B.3. Let n ∈ N and E be a locally convex space. Consider the topology T on
C∞ (Rn , E) generated by basic neighborhoods which arise from elementary neighborhoods of the form N r ( f ; A, (Rn , idRn ), (E, idE ), q, ε ). Then T coincides with the very
strong topology on C∞ (M, E).
Proof. Clearly restricting our choice of charts to idRn generates a topology coarser than
the very strong topology. To see that T is also finer, consider an elementary neighbor∞ (Rn , E).
hood N = N r ( f ; A, (U, φ ), (E, idE ), q, ε ) in CvS
Now g ∈ N if and only if k f ◦ φ −1 − g ◦ φ −1 k(r, φ (A), q) < ε . Using Lemma B.1, we
obtain a constant K > 0 with
k f ◦ φ −1 − g ◦ φ −1 k(r, φ (A), q) = k( f − g) ◦ φ −1 k(r, φ (A), q)
≤ Kk f − gk(r, A, q)kφ −1k(r, φ (A))
Together with Lemma 1.7 this implies that for every h ∈ N there is δ > 0 with Nh :=
N r (h; A, (Rn, idRn ), (E, idE ), q, δ ) ⊆ N . Hence N is also open in T . As Nh controls
8We invite the reader to verify this in the proof of [Glö04, Proposition 4.23 (a)].
STRONG TOPOLOGIES FOR SMOOTH MAPS
38
functions on the same compact set as N , the argument generalizes to basic neighborhoods. Summing up, basic neighborhoods are open in T which is thus finer as the very
strong topology.
Finally, we provide a standard fact about σ -compact topological spaces.
Lemma B.4. Let M be a σ -compact topological space. Then M admits a locally finite
exhaustion by compactSsets, i.e. there exists a locally finite family {Ai }i∈N of compact
subsets of M such that i∈N Ai = M.
Proof. By [Dug78][Theorem XI.7.2] there exists aSfamily {Ki }i∈N of compact subspaces of M such that Ki ⊆ int Ki+1 for all i ∈ N and Ki = M. Now define A′1 = K1 and
A′i = Ki \ int Ki−1 for i ≥ 2. The resulting family {A′i }i∈N is a locally finite exhaustion of
M by compact sets.
A PPENDIX C. A LTERNATIVE
DESCRIPTION OF THE TOPOLOGY VIA JET BUNDLES
In this section we give a proof of the “folklore” fact that Michor’s D-topology on
C∞ (M, N) (cf. [Mic80, Section 1]) coincides with the very strong topology when the
target N is finite-dimensional. This is claimed in many places in the literature, e.g.
[Hir94] (where the D-topology is called strong Whitney topology), but the authors of
this paper have been unable to locate a proof. The D-topology is constructed using jet
bundles (cf. [Mic80, Section 1] and [KM97, Section 41]). We will briefly recall this
construction here.
Definition C.1. Let U ⊆ Rm and V ⊆ Rn be open. For 0 ≤ r ≤ ∞ we define the space
of r-jets from U to V by
r
J r (U,V ) := U ×V × ∏ Lisym (Rm , Rn ),
i=1
where Lisym (Rm , Rn ) is the space of symmetric linear maps (Rm )i → Rn . We topologize
∏ri=1 Lisym (Rm , Rn ) with respect to the operator norm for multilinear maps.
For a smooth map f : U → V we define
1 (2)
1 (r)
r
(1)
j f (x) := x, f (x), d f (x; ·), d f (x; ·), . . . , d f (x; ·) ,
2!
r!
called the r-jet of f at x.
Definition C.2. Let M and N be finite-dimensional smooth manifolds. We define an rjet from M to N to be an equivalence class of pairs ( f , x), where f : M → N is a smooth
map and x ∈ M. Two pairs ( f , x) and ( f ′ , x′ ) are equivalent if x = x′ and Txr f = Txr f ′ ,
where T r is the r-th tangent mapping. We write jr f (x) for the equivalence class of
( f , x). The set of all r-jets from M to N is denoted J r (M, N).
STRONG TOPOLOGIES FOR SMOOTH MAPS
39
In the case M = Rm and N = Rn , the r-jet of a smooth map f : M → N is represented by
the Taylor polynomial of f . So the different definitions above coincide.
If f : M → M ′ is smooth, then there is an induced map J r (N, f ) : J r (N, M) → J r (N, M ′ )
given by J r (N, f )( jr g(x)) = jr ( f ◦ g)(x). For a diffeomorphism f we obtain in addition
the map J r ( f , N) : J r (M ′ , N) → J r (M, N) given by J r ( f , N)( jr g(x)) = jr (g◦ f )( f −1 (x)).
The set of all r-jets J r (M, N) form a smooth manifold. Suppose (U, φ ) is a chart for M
and (V, ψ ) is a chart for N. The map
J r (φ −1 , ψ ) : J r (U,V ) →J r (φ (U ), ψ (V )) ,
J r (φ −1 , ψ ) := J r (φ −1 , ψ (V )) ◦ J r (U, ψ ) = J r (φ (U ), ψ ) ◦ J r(φ −1 ,V )
is bijective. Note that J r (U,V ) can be identified with a subset of J r (M, N), namely
J r (U,V ) = { f ∈ J r (M, N) : f (U ) ⊆ V }
since ateach point there existsa smooth map with a given Taylor expansion. The collection J r (U,V ), J r (φ −1 , ψ ) , where (U, φ ) and (V, ψ ) runs through the charts for
M and N, respectively, is a smooth atlas for J r (M, N). See [Mic80, 1.8] for details.
Via jet bundles one can define a topology on the space of smooth functions. Using the
embedding j∞ : C∞ (M, N) → C(M, J ∞(M, N)) for M, N finite-dimensional, one pulls
back a certain topology to obtain Michor’s D-topology on C∞ (M, N). We will not describe this construction in detail and refer to [Mic80, 4.7.2] for the following alternative
description of the topology
Definition C.3. Let M and N be finite-dimensional smooth manifolds. We define
Ω(L, U) := { f ∈ C∞ (M, N) | jn f (Ln ) ⊆ Un for all n ∈ N}
where L = {Ln }n∈N is a locally finite family of closed subsets of M, and U = {Un}n∈N
is a family of open subsets Un ⊆ J n (M, N).
Then the family of sets Ω(L, U) where L runs through all locally finite families of
closed sets and U runs through all families of open subsets Un ⊆ J n (M, N) is a basis for
a topology on C∞ (M, N). Following Michor, we call this topology the D-topology and
∞ (M, N) the smooth functions with this topology.
denote by CD
Remark C.4. Let L = {Ln }n∈N be a locally finite family of closed subsets of M, r =
{rn }n∈N be a sequence of natural numbers, U = {Un }n∈N be a family of open subsets
Un ⊆ J rn (M, N), and set
Ω(r, L, U) := { f ∈ C∞ (M, N) : jrn f (Ln ) ⊆ Un for all n ∈ N}.
Taking the family of all sets Ω(r, L, U) as L, r, and U vary also forms a basis for the
D-topology on C∞ (M, N). To see this, note that if Un ⊆ J n (M, N) is open, k ≥ 0, and
π : J n+k (M, N) → J n (M, N) is the “truncation map”, then for all f : M → N smooth and
x ∈ M, we have jn f (x) ∈ Un if and only if jn+k f (x) ∈ π −1Un .
STRONG TOPOLOGIES FOR SMOOTH MAPS
40
∞ (M, N)
Lemma C.5. The family of all sets Ω(r, L, U) is a basis for the D-topology CD
if we require that: r = {rn }n∈N runs through all sequences of natural numbers, L =
{Ln }n∈N runs through all locally finite families of compact subsets of M such that each
Ln is contained in some chart (Vn , ψn ) for M, and U = {Un }n∈N runs through all families
of open subsets Un ⊆ J rn (Vn ,Wn ) where (Wn , φn ) is a chart of N,
Proof. It suffices to show the following: Given a locally finite family L = {Ln }n∈N of
closed subsets of M and a family U = {Un}n∈N of open subsets Un ⊆ J n (M, N), there
exist a sequence r = {rn }n∈N of natural numbers, a locally finite family A = {An }n∈N
of compact subsets of M, a family {(Vn , ψn )}n∈N of charts for M such that An ⊆ Vn for
all n ∈ N, a family {(Wn , φn )}n∈N of charts for N, and a family U′ = {Un′ }n∈N of open
subsets Un′ ⊆ J rn (Vn ,Wn ) such that Ω(r, A, U′) = Ω(L, U).
Since M is σ -compact there exists a compact exhaustion K = {Kn }n∈N of M, and by
dividing each Kn into finitely many compact subsets each contained in some manifold
chart, we may assume that each Kn ⊆ Vn for some chart (Vn , ψn ) for M. We may further
require that there is a chart (Wn , φn ) such that f (Kn ) ⊆ Wn . The family {Ki ∩Lk }(i,k)∈N×N
is a locally finite family of compact sets since the Ki are compact and L is a locally finite
family of closed sets. Note that Ki ∩Lk ⊆ Vi for all (i, k) ∈ N ×N. Moreover, for all k ∈ N
we have jk f (Ki ∩ Lk ) ⊆ Uk ∩ J rk (Vi ,Wi ) for all i ∈ N if and only if jk f (Lk ) ⊆ Uk .
′
Take a bijection b : N × N → N, set Ab(i,k) = Ki ∩ Lk , and rb(i,k) = k, and Ub(i,k)
= Uk ∩
r
k
J (Vi ,Wi ). Then
Ω(r, A, U′)
={ f ∈ C∞ (M, N) : jrn f (An ) ⊆ Un′ for all n ∈ N}
={ f ∈ C∞ (M, N) : jk f (Ki ∩ Lk ) ⊆ Uk ∩ J rk (Vi ,Wi ) for all i ∈ N and all k ∈ N}
={ f ∈ C∞ (M, N) : jk f (Lk ) ⊆ Uk for all k ∈ N}
=Ω(L, U).
Definition C.6. Let M and N be finite-dimensional smooth manifolds. If r ∈ N, A ⊆ M
a compact subset, and O ⊆ J r (M, N) open, then
M r (A, O) := { f ∈ C∞ (M, N) : jr f (A) ⊆ O}
is an elementary D-neighborhood.
These neighborhoods play the same role for the D-topology as the elementary (vS)neighborhoods play in the very strong topology. If r is a sequence of natural numbers,
A is a locally finite family of compact subsets of M, and O is a family of open subsets
On ⊆ J rn (M, N), then
\
Ω(r, A, O) =
M rn (An , On ).
n∈N
STRONG TOPOLOGIES FOR SMOOTH MAPS
41
Lemma C.7. Let M and N be finite-dimensional smooth manifolds. Suppose U ⊆ M
∞
and V ⊆ N are open. Consider the subspace CD,sub
(U,V ) = { f ∈ C∞ (M, N) : f (U ) ⊆
∞ (M, N).
V } ⊆ CD
∞
(1) CD,sub
(U,V ) is an open subset of CD (M, N).
∞
∞ (U,V ) is continuous.
(2) The restriction resD : CD,sub
(U,V ) → CD
(3) For an elementary D-neighborhood M r (A, O) = { f ∈ C∞ (U,V ) : jr f (A) ⊆ O}
∞ (U,V ), we have res−1 (M r (A, O)) = M r (A, O) ⊆ C ∞ (M, N).
in CD
D
D
Proof.
(1) By Lemma B.4 there exists a locally finite exhaustion {An }n∈N of U by
compact sets. Observe that
∞
CD,sub
(U,V ) = f ∈ C∞ (M, N) : j0 f (An ) ⊆ U ×V for all n ∈ N .
(2) The inclusions ιU : U → M and ιV : V → N are continuous embeddings. Con∞
∞ (U, N) : f (U ) ⊆ V }, and the coresider the subspace ĈD,sub
(U,V ) := { f ∈ CD
∞
∞ (U,V ),
striction R : ĈD,sub
(U,V ) → CD
f 7→ f |V .
Then resD = R ◦ (ιU )∗ . The map (ιU )∗ is continuous with respect to the Dtopology by [Mic80, Proposition 7.4]. Recall from [Mic80, Proposition 10.8]
∞ (U,V ) → C ∞ (U, N) is an embedding of topological
that the mapping (ιV )∗ : CD
D
∞
(U,V ) is open by
spaces. Further (ιV )∗ ◦ R = idC∞ (U,N) |Ĉ∞ (U,V ) . Since ĈD,sub
D,sub
(1) and the identity is continuous, we deduce that R is continuous. In conclusion
resD is continuous.
(3) Clear from the definition as we can identify O ⊆ J r (U,V ) via the identification
J r (U,V ) ⊆ J r (M, N) with an open subset in J r (M, N).
Our aim is now to prove the following folklore theorem.
Proposition C.8. If M and N are finite-dimensional smooth manifolds, then the very
strong and the D-topology coincide.
To prove the Proposition we need to prove the following special case first.
∞ (U,V ) = C ∞ (U,V )
Proposition C.9. Let U ⊆ Rm and V ⊆ Rn be open subsets, then CvS
D
Proof. Comparing the bases of the D-topology and vS-topology, it is clearly sufficient to show the following: For each elementary D-neighborhood M = M r (A, O)
and f ∈ M , there exists an ε > 0 such that N r ( f ; A, ε ) ⊆ M , and conversely that
for every elementary vS-neighborhood N = N r ( f ; A, ε ) and g ∈ N there exists a
finite
cover of A by compact sets Kk and open subsets Ok ⊆ J r (U,V ) such that g ∈
T
r
1≤k≤n M (Kk , Ok ) ⊆ N .
STRONG TOPOLOGIES FOR SMOOTH MAPS
42
To this end, we fix r ∈ N, A ⊆ V a compact subset, and O ⊆ J r (U,V ) open. Let us
prove that M r (A, O) is open in the vS-topology. To this end consider f ∈ M r (A, O).
For all x ∈ A we have jr f (x) ∈ O, whence there exist Ox,U ⊆ U open, an open ball
′ ( f (x)) ⊆ V and for every i ∈ {1, . . ., r} open balls
Bx,0 := Bεx,0
)
(
(i)
d
f
(x;
y
,
.
.
.
,
y
)
i
1
′
< εx,i
T (y1 , . . . , yi ) −
Bx,i := T ∈ Lisym (Rm , Rn ) :
sup
i!
k(y1 ,...,yi )k≤1
such that we have
r
j f (x) ∈ Ox,U × ∏ Bx,i ⊆ O ⊆ J r (U,V ).
r
i=0
Recall that a map which is smooth in the sense that all partial derivatives of arbitrarily
high order exist, is also smooth in the sense of Fréchet differentiability (see [Die69,
8.12.8]). Hence jr f : U → J r (U,V ) is continuous and we can thus find for each x a
compact x-neighborhood Kx,U ⊆ Ox,U such that
r
jr f (y) ∈ Ox,U × ∏ Bx,i ⊆ O ⊆ J r (U,V )
for all y ∈ Kx,U .
i=0
Since A is compact there exist x1 , . . . , xn ∈ A with
!
(7)
jr f (A) ⊆
n
[
k=1
r
Kxk ,U × ∏ Bxk ,i
⊆
i=0
n
[
k=1
Set ε ′ = min{εx′ k ,i : 1 ≤ k ≤ n, 0 ≤ i ≤ r}.
r
Oxk ,U × ∏ Bxk ,i
i=0
!
⊆O
Exploiting the triangle inequality and multilinearity of the Fréchet derivative, one constructs constants εk,i > 0 such that: If g : U → V is smooth and k d(i) g(x; α1 , . . . , αi ) −
d(i) f (x; α1 , . . . , αi )k < εk,i for all α1 , . . . , αi ∈ {e1 , . . ., em } and x ∈ Kxk ,U , then
(8)
d(i) g(x; y1 , . . . , yi ) d(i) f (x; y1 , . . . , yi )
< ε′
−
i!
i!
for x ∈ A, y1 , . . ., yi ∈ Rm with k(y1 , . . . , yi )k ≤ 1. Set ε = min{εk,i : 1 ≤ k ≤ n, 0 ≤ i ≤ r}.
If g ∈ N r ( f ; A, ε ), then for 1 ≤ k ≤ n and 0 ≤ i ≤ r, we have for x ∈ A
k d(i) g(x; α1 , . . . , αi ) − d(i) f (x; α1 , . . ., αi )k ≤ kg − f k(r, A) < ε ≤ εk,i ,
which implies (8). By construction of the neighborhoods, we derive from (7) that g ∈
M r (A, O). Hence N r ( f ; A, ε ) ⊆ M r (A, O) and M r (A, O) is open in the vS-Topology.
For the converse, fix g ∈ N r ( f ; A, ε ). Set δ =
x ∈ A the open sets
(
Bi,x :=
T ∈ Lisym (Rm , Rn ) :
sup
k(y1 ,...,yi )k≤1
ε −kg− f k(r,A)
2
and define for 1 ≤ i ≤ r and
d (i) g(x; (y1 , . . . , yi ))
<δ
T (y1 , . . . , yi ) −
i!
)
STRONG TOPOLOGIES FOR SMOOTH MAPS
43
Since d (i) g : U → Lisym (Rm , Rn ) is continuous (cf. [Die69, 8.12.8]), there are finitely
many xk ∈ A, 1 ≤ k ≤ n and compact xk -neighborhoods Kk such that d (i) g(Kk ) ⊆ Bi,xk
for each 1 ≤ i ≤ r. Dividing the compact sets Kk into smaller parts, we may assume that
there are open sets B0,xk such that g(Kk ) ⊆ B0,xk and z ∈ B0,xk implies kz − g(x)k < δ for
all x ∈ Kk . Now define Ok := U × ∏ri=0 Bi,xk and observeTthat this set is open in J r (U,V ).
Further we have jr g(Kk ) ⊆ Ok for each kTand thus g ∈ 1≤k≤n M r (Kk , Ok ). Exploiting
the triangle inequality, we derive for h ∈ 1≤k≤n M r (Kk , Ok ) the estimate
k f − hk(r, A) ≤ k f − gk(r, A) + kg − hk(r, A) ≤ k f − gk(r, A) + sup kg − hk(r, Kk )
1≤k≤n
= k f − gk(r, A) + sup sup sup kd g(x; α ) − d (i) h(x, α )k
(i)
1≤k≤n x∈Kk 0≤|α |≤r
≤ k f − gk(r, A) + sup sup sup kd (i) g(x; ·) − d (i) h(x, ·)kop
1≤k≤n x∈Kk 0≤i≤r
< k f − gk(r, A) + sup δ = ε .
1≤k≤n
The last inequality is derived from the definition of Ok (or Bi,xk , respectively). Summing
up, h ∈ N r ( f ; A, ε ) and we see that N r ( f ; A, ε ) is open in the D-topology.
Proof of Proposition C.8. Observe that by definition of the very strong topology and
the definition of the D-topology it suffices to prove
that every elementary neighborT
r
hood N ( f ; A, (U, φ ), (V, ψ ), p, ε ) is the union of 1≤k≤n M r (Kk , O), where the Kk are
a finite family of compact sets which cover A (cf. proof of Proposition C.9). Then
N r ( f ; A, (U, φ ), (V, ψ ), p, ε ) is open and we can write each basic neighborhood in
∞ (M, N) as a (possibly infinite) union of sets of the form Ω(L, U) (cf. Definition C.3).
CvS
Hence we fix N r ( f ; A, (U, φ ), (V, ψ ), p, ε ). If necessary we shrink U to achieve f (U ) ⊆
V (while still A ⊆ U ). Thus we obtain a corresponding elementary neighborhood
∞
N r ( f |VU ; A, (U, φ ), (V, ψ ), p, ε ) ⊆ CvS
(U,V ).
r
r
V
Clearly res−1
vS (N ( f |U ; A, (U, φ ), (V, ψ ), p, ε )) = N ( f ; A, (U, φ ), (V, ψ ), p, ε ). Now
consider the commutative diagram:
∞
CvS,sub
(U,V )
resvS
/ C ∞ (U,V )
vS
(9)
(φ −1 )∗ ◦ψ∗ ∞
/ C (φ (U ), ψ (V ))
vS
∼
=
∼
=
∞
CD,sub
(U,V )
resD
/ C ∞ (U,V )
D
(φ −1 )∗ ◦ψ∗ ∞
/ C (φ (U ), ψ (V ))
∼
D
=
Since both instances of (φ −1 )∗ ◦ ψ∗ are homeomorphisms, and the rightmost identity map is a homeomorphism by Proposition C.8, the middle identity map is also
a homeomorphism. From the second part of the proof of Proposition C.8 we derive that N r ( f |VU ; A, (U, φ ), (V, ψ ), p, ε ) indeed is a union of open sets of the form
STRONG TOPOLOGIES FOR SMOOTH MAPS
44
∞ (U,V ). Now Lemma C.7 (3) implies together with the comin CD
mutativity
of (9) that also N r ( f ; A, (U, φ ), (V, ψ ), p, ε ) is a union of neighborhoods
T
r
∞
1≤k≤n M (Kk , O) in CD (M, N). We conclude that the very strong topology is coarser
than the D-topology.
T
1≤k≤n M
r (K , O )
k
k
For the converse observe again, that by the definitions of the topologies it suffices to
prove that every elementary D-neighborhood M r (A, O) is the union of elementary
neighborhoods N r ( f ; A, (U, φ ), (V, ψ ), p, ε ). Here we have invoked Lemma C.5 to see
that it is indeed enough to consider the case A ⊆ (U, φ ) and O ⊆ J r (U,V ) for some charts
(U, φ ) and (V, ψ ). Hence, one can argue as in the first case, if one replaces Lemma C.7
(3) with Lemma 1.15 (3).
Summing up, the D-topology is also coarser than the very strong topology, whence they
coincide on C∞ (M, N).
Note that we have not used in any essential way in this Appendix that the target manifold N is a finite-dimensional manifold. Indeed, all arguments readily generalize without changes to the case where N is a Banach manifold (and the source M is finitedimensional). In [Mic80] only finite-dimensional manifolds are considered as target.
As the definition of jet bundles (and their topology) generalizes verbatim to the realm of
Banach manifolds, one could also define a D-topology for the space C∞ (M, N) where
N is a Banach manifold. We conclude that the statement of Proposition C.8 remains
true even if N is a Banach manifold (and M finite-dimensional). Hence we can copy the
following results from [Mic80] by appealing to Proposition C.8:
Corollary C.10. Let M be a finite-dimensional manifold and N be a Banach manifold.
∞ (M, N) is a Baire space.
Then the topological space CvS
Proof. Combine [Mic80, 4.7.6] with Proposition C.8. Note that the proof in [Mic80]
carries over verbatim to the case of N being an infinite-dimensional Banach manifold.
Remark C.11. In [Mic80] a refinement of the D-topology, the so called F D-topology
is constructed. The construction is similar to the construction of the fine very strong
topology, whence Proposition C.8 implies that the fine very strong topology and the
F D-topology on C∞ (M, N) coincide for finite-dimensional manifolds M and N.
∞ (M, N) will in general not be
In particular, [Mic80, Remark 4.11] thus implies that CfS
a Baire space. However by loc.cit., we then derive that for M and N finite-dimensional
∞ (M, N) is paracompact and normal.
the topological space CfS
STRONG TOPOLOGIES FOR SMOOTH MAPS
45
R EFERENCES
[Bas64]
Bastiani, A. Applications différentiables et variétés différentiables de dimension infinie. J.
Analyse Math. 13 (1964):1–114
[Čec66]
Čech, E. Topological spaces. Revised edition (Wiley, London-New York-Sydney, 1966)
[Die69]
Dieudonné, J. Foundations of modern analysis (Academic Press, New York-London, 1969).
Enlarged and corrected printing, Pure and Applied Mathematics, Vol. 10-I
[Dug78]
Dugundji, J. Topology (Allyn and Bacon, 1978). Reprinting of the 1966 original, Allyn and
Bacon Series in Advanced Mathematics
[Eng89]
Engelking, R. General topology, Sigma Series in Pure Mathematics, vol. 6 (Heldermann
Verlag, Berlin, 1989), second edn.
[Glö02]
Glöckner, H. Infinite-dimensional Lie groups without completeness restrictions. In Geometry
and analysis on finite- and infinite-dimensional Lie groups (Bedlewo, 2000), Banach Center
‘
Publ., vol. 55, pp. 43–59 (Polish Acad. Sci., Warsaw, 2002)
[Glö04]
Glöckner,
H. Lie groups over non-discrete topological fields 2004.
arxiv:math/0408008v1
[Glö06]
Glöckner, H. Discontinuous non-linear mappings on locally convex direct limits. Publ. Math.
Debrecen 68 (2006)(1-2):1–13
[Hir94]
Hirsch, M. W. Differential topology, Graduate Texts in Mathematics, vol. 33 (SpringerVerlag, New York, 1994). Corrected reprint of the 1976 original
[Ill03]
Illman, S. The very-strong C∞ topology on C∞ (M, N) and K-equivariant maps. Osaka J.
Math. 40 (2003)(2):409–428
[Kel74]
Keller, H. H. Differential calculus in locally convex spaces. Lecture Notes in Mathematics,
Vol. 417 (Springer-Verlag, Berlin-New York, 1974)
[KM97]
Kriegl, A. and Michor, P. The convenient setting of global analysis, Mathematical Surveys
and Monographs, vol. 53 (American Mathematical Society, Providence, RI, 1997)
[Köt69]
Köthe, G. Topological vector spaces. I (Springer-Verlag New York Inc., New York, 1969)
[Mac05]
Mackenzie, K. C. H. General theory of Lie groupoids and Lie algebroids, London Mathematical Society Lecture Note Series, vol. 213 (Cambridge University Press, Cambridge,
2005)
[Mic80]
Michor, P. W. Manifolds of differentiable mappings, Shiva Mathematics Series, vol. 3 (Shiva
Publishing Ltd., Nantwich, 1980)
[Mil83]
Milnor, J. Remarks on infinite-dimensional Lie groups. In B. DeWitt and R. Stora (Eds.),
Relativity, Groups and Topology II, pp. 1007–1057 (North Holland, New York, 1983)
[MROD92] Margalef Roig, J. and Outerelo Domı́nguez, E. Differential topology, North-Holland Mathematics Studies, vol. 173 (North-Holland, Amsterdam, 1992)
[Nee06]
Neeb, K.-H. Towards a Lie theory of locally convex groups. Jpn. J. Math. 1 (2006)(2):291–
468
[Sta13]
Stacey, A. Yet More Smooth Mapping Spaces and Their Smoothly Local Properties 2013.
arXiv:1301.5493v1
[SW15a]
Schmeding, A. and Wockel, C. The lie group of bisections of a lie groupoid. Annals of Global
Analysis and Geometry 48 (2015)(1):87–123
[SW15b] Schmeding, A. and Wockel, C. (Re)constructing Lie groupoids from their bisections and
applications to prequantisation 2015. arXiv:1506.05415
| 4math.GR |
Logical Methods in Computer Science
Vol. 6 (3:1) 2010, pp. 1–37
www.lmcs-online.org
Submitted
Published
Oct. 12, 2009
Jul. 11, 2010
LAZY EVALUATION AND DELIMITED CONTROL
RONALD GARCIA a , ANDREW LUMSDAINE b , AND AMR SABRY c
a
b,c
Carnegie Mellon University
e-mail address: rxg@cs.cmu.edu
Indiana University
e-mail address: {lums,sabry}@cs.indiana.edu
Abstract. The call-by-need lambda calculus provides an equational framework for reasoning syntactically about lazy evaluation. This paper examines its operational characteristics.
By a series of reasoning steps, we systematically unpack the standard-order reduction relation of the calculus and discover a novel abstract machine definition which, like
the calculus, goes “under lambdas.” We prove that machine evaluation is equivalent to
standard-order evaluation.
Unlike traditional abstract machines, delimited control plays a significant role in the
machine’s behavior. In particular, the machine replaces the manipulation of a heap using
store-based effects with disciplined management of the evaluation stack using control-based
effects. In short, state is replaced with control.
To further articulate this observation, we present a simulation of call-by-need in a callby-value language using delimited control operations.
1. Introduction
From early on, the connections between lazy evaluation (Friedman and Wise, 1976;
Henderson and Morris, 1976) and control operations seemed strong. One of these seminal
papers on lazy evaluation (Henderson and Morris, 1976) advocates laziness for its coroutinelike behavior. Specifically, it motivates lazy evaluation with a solution to the same fringe
problem: how to determine if two trees share the same fringe without first flattening each
tree and then comparing the resulting lists. A successful solution to the problem traverses
just enough of the two trees to tell that they do not match. The same fringe problem
is also addressed in Sussman and Steele’s original exposition of the Scheme programming
1998 ACM Subject Classification: D.3.1.
Key words and phrases: call-by-need, reduction semantics, abstract machines, delimited continuations,
lambda calculus.
a
This work was supported by the National Science Foundation under Grant #0937060 to the Computing
Research Association for the CIFellows Project.
b,c
This work was supported by NSF awards CSR/EHS 0720857 and CCF 0702717.
l
LOGICAL METHODS
IN COMPUTER SCIENCE
c
DOI:10.2168/LMCS-6 (3:1) 2010
CC
R. Garcia, A. Lumsdaine, and A. Sabry
Creative Commons
2
R. GARCIA, A. LUMSDAINE, AND A. SABRY
language (Sussman and Steele Jr., 1975). One of their solutions uses a continuation passingstyle representation of coroutines. More recently, Biernacki et al. (2005) explores a number
of continuation-based solutions to same fringe variants.
Same fringe is not the only programming problem that can be solved using either lazy
evaluation or continuations. For instance, lazy streams and continuations are also used to
implement and reason about backtracking (Wand and Vaillancourt, 2004; Kiselyov et al.,
2005). Strong parallels in the literature have long suggested that lazy evaluation elegantly
embodies a stylized use of coroutines. Indeed, we formalize this connection.
Call-by-need evaluation combines the equational reasoning capabilities of call-by-name
with a more efficient implementation technology that systematically shares the results of
some computations. However, call-by-need’s evaluation strategy makes it difficult to reason
about the operational behavior and space usage of programs. In particular, call-by-need
evaluation obscures the control flow of evaluation. To facilitate reasoning, semantic models (Launchbury, 1993; Sestoft, 1997; Friedman et al., 2007; Nakata and Hasegawa, 2009),
simulations (Okasaki et al., 1994), and tracing tools (Gibbons and Wansbrough, 1996) for
call-by-need evaluation have been developed. Many of these artifacts use an explicit store
or store-based side-effects (Wang, 1990) to represent values that are shared between parts
of a program. Stores, being amorphous structures, make it difficult to establish program
properties or analyze program execution. This representation of program execution loses
information about the control structure of evaluation.
The call-by-need lambda calculus was introduced by Ariola et al. (1995) as an alternative to store-based formalizations of lazy evaluation. It is an equational framework for
reasoning about call-by-need programs and languages. Following Plotkin (1975), these authors present a calculus and prove a standardization theorem that links the calculus to a
complete and deterministic (i.e. standard order) reduction strategy. The calculus can be
used to formally justify transformations, particularly compiler optimizations, because any
terms it proves equal are also contextually equivalent under call-by-need evaluation.
Call-by-need calculi were investigated by two groups (Maraist et al., 1998; Ariola and Felleisen,
1997). The resulting two calculi are quite similar but their subtle differences yield trade-offs
that are discussed in the respective papers. Nonetheless, both papers connect their calculi
to similar standard-order reduction relations.
One notable feature of Ariola and Felleisen’s calculus (and both standard-order reduction relations) is the use of evaluation contexts within the notions of reduction. It has been
observed that evaluation contexts correspond to continuations in some presentations of language semantics (Felleisen and Friedman, 1986; Biernacka and Danvy, 2007). However, in
these systems evaluation contexts are used to model variable references and demand-driven
evaluation, not first-class continuations.
This paper exposes how Ariola et al.’s call-by-need evaluation relates to continuations.
By systematically unpacking the standard-order reduction relation of the calculus, we discover a novel abstract machine that models call-by-need style laziness and sharing without
using a store. Instead, the machine manipulates its evaluation context in a manner that
corresponds to a stylized use of delimited control operations. The machine’s behavior reveals a connection between control operations and laziness that was present but hidden in
the reduction semantics.
To directly interpret this connection in the terminology of delimited control, we construct a simulation of call-by-need terms in the call-by-value language of Dybvig et al.
LAZY EVALUATION AND DELIMITED CONTROL
3
(2007), which provides a general framework for delimited continuations with first-class generative prompts.
Our concrete specifications of the relationship between call-by-need and delimited control firmly establish how lazy evaluation relates to continuations and other control-oriented
language constructs and effects. Implementations of both the machine and the simulation
are available at the following url:
http://osl.iu.edu/~garcia/call-by-need.tgz.
2. The Call-by-need Lambda Calculus
The remainder of this paper examines Ariola and Felleisen’s formalization of call-byneed (Ariola and Felleisen, 1997). The terms of the calculus are standard:
t ::= x | λx.t | t t
The call-by-need calculus, in direct correspondence with the call-by-value and call-by-name
lambda calculi, distinguishes lambda abstractions as values:
v ::= λx.t
Call-by-need is characterized by two fundamental properties: a computation is only performed when its value is needed, and the result of any computation is remembered and
shared so that it only needs to be computed once. This calculus distinguishes two additional subsets of the term language to help represent these properties.
To capture the notion of necessary computations, the calculus distinguishes the set of
lambda abstractions that immediately need the value of their argument. To define this set,
the calculus appeals to a notion of evaluation contexts, a set of terms that each have a
single hole () in them:
E ::= | E t | (λx.E[x]) E | (λx.E) t
Given the evaluation contexts, the set of lambda abstractions in question is defined syntactically as (λx.E[x]), the set of lambda abstractions whose bodies can be decomposed into
an evaluation context and a free instance of the abstracted variable.
The intuition for this definition is as follows. Evaluation contexts are essentially terms
with a single hole in them. When used in the style of Felleisen and Hieb (1992), evaluation
contexts indicate those locations in a program that are subject to evaluation. As such, a
lambda abstractions of the form (λx.E[x]) will immediately refer to the value of its argument
when its body is evaluated. Here, E[x] is not a language construct, but rather metalinguistic
notation for a side condition on the term in the body of the lambda abstraction.
The syntactic structure of call-by-need evaluation contexts give some hint to the nature
of computation captured by this calculus. First, the context production E t indicates that
evaluation can focus on the operator position of an application regardless of the structure
of the operand. This property is also true of call-by-name and call-by-value and reflected
in their respective evaluation contexts. Second, the (λx.E[x]) E production indicates the
operand of an application expression can be evaluated only if the operator is a lambda
abstraction that immediately needs its argument. This restriction does not hold for call-byvalue1, where any lambda abstraction in operator position justifies evaluating the operand.
1assuming left-to-right evaluation of application expressions
4
R. GARCIA, A. LUMSDAINE, AND A. SABRY
This restriction on evaluation corresponds with our intuitive understanding of call-by-need.
Third, the (λx.E) t production indicates that evaluation can proceed under a lambda abstraction when it is in operator position. Though not immediately obvious, this trait is
used by the calculus to capture the process of sharing computations among subterms.
To capture the notion of shared computations, the call-by-need calculus distinguishes
lambda abstractions with explicit bindings for some variables, calling them answers:
a ::= v | (λx.a) t
Answers are a syntactic representation of (partial) closures. An answer takes the form of a
lambda term nested inside some applications. The surrounding applications simulate environment bindings for free variables in the nested lambda term. This representation makes
it possible for the calculus to explicitly account for variable binding and to syntactically
model how call-by-need evaluation shares lazily computed values.
The calculus has three notions of reduction:
(λx.E[x]) v
→need (λx.E[v]) v
(λx.a) t1 t2
→need (λx.a t2 ) t1
(λx1 .E[x1 ]) ((λx2 .a) t1 ) →need (λx2 .(λx1 .E[x1 ]) a) t1
The first reduction rule substitutes a value for a single variable instance in an abstraction.
The rule retains the binding and abstraction so as to share its computation with other
variable references as needed. The second and third reduction rules commute an application
with an answer binding to expose opportunities for reduction without duplicating not-yetneeded computations. These two rules help to ensure that computations will be shared
among references to a common variable.
As popularized by Barendregt (1981), each reduction assumes a hygiene convention.
When combined with the evaluation contexts, the notions of reduction yield a deterministic
standard order reduction relation (7−→sr ) and its reflexive-transitive closure ( 7−→
→sr ).
Definition 2.1. t1 7−→sr t2 if and only if t1 ≡ E[tr ], t2 ≡ E[tc ] and tr →need tc .
Terms of the calculus satisfy unique decomposition, meaning that any program (i.e.
closed term) that is not an answer can be decomposed exactly one way into a context E
and redex tr . This property ensures that 7−→sr is deterministic. Standard order reduction
is an effective specification of call-by-need evaluation: if t is a program (i.e. closed term),
then t call-by-need evaluates to an answer if and only if t 7−→
→sr a for some answer a.
3. From Reduction Semantics to Machine Semantics
Some reduction semantics have been shown to correspond directly to abstract machine
semantics, thereby establishing the equivalence of a reducer and a tail-recursive abstract
machine implementation (Felleisen and Friedman, 1986; Felleisen et al., 2009). In particular, Danvy and Nielsen (2004) introduces a method and outlines criteria for mechanically
transforming reduction semantics into abstract machine semantics. However, proceeding
directly from the reduction semantics for call-by-need to a tail-recursive abstract machine
semantics poses some challenges that do not arise with other reduction semantics like callby-name and call-by-value.
A straightforward call-by-need reducer implementation naı̈vely decomposes a term into
a context and a redex. Any application could be one of three different redexes, each of
LAZY EVALUATION AND DELIMITED CONTROL
5
which is nontrivial to detect, so whenever the decompose function detects an application, it
sequentially applies several recursive predicates to the term in hopes of detecting a redex.
If the term is a redex, it returns; if not, it recursively decomposes the operator position.
Call-by-need redexes require more computational effort to recognize than either callby-name or call-by-value. For instance, given a term t, only a fixed number of terminal
operations are required to detect whether t is a call-by-name redex: one to check if the
term is an application, one to access the operator position, and one to check if the operator
is an abstraction.
Contrast this with the call-by-need redex (λx.E[x])((λy.a)t). Given a call-by-need term
tx , testing whether it matches this redex form requires an unknown number of operations:
check if tx is an application; check if its operator position is a lambda abstraction; check, in
an unknown number of steps, if the operator’s body can be decomposed into E[x], where x
is both free in E and bound by the operator; and finally check, in an unknown number of
steps, if the operand has the inductive structure of an answer.
To make matters worse, some terms can be decomposed into the form E[x] in more
than one way. For instance, consider the term (λx.(λy.y) x). It can be decomposed as both
(λx.E1 [y]) and (λx.E2 [x]) where E1 ≡ (λy.) x and E2 ≡ (λy.y) . As such, a recursive
procedure for decomposing a term cannot stop at the first variable it finds: it must be able
to backtrack in a way that guarantees it will find the right decomposition E[x]—and in turn
the right redex—if there is one.
Recall that one of the evaluation contexts has the form (λx.E)t. This means that redex
evaluation can occur “under binders” (Moggi and Sabry, 2004; Kameyama et al., 2008). All
three call-by-need notions of reduction shuffle lambda abstractions about in unusual ways.
Furthermore, while reducing a recursive routine, a call-by-need evaluator may end up performing reductions under multiple copies of the same lambda abstraction. Call-by-name and
call-by-value evaluators can address hygiene concerns by using environments and closures,
but a call-by-need evaluator must prevent its evaluation context from incorrectly capturing free variable references. Any evaluator that goes under lambdas must pay particular
attention to hygiene (Xi, 1997).
Since this work was originally published, Danvy et al. (2010) have adapted and extended the method of Danvy and Nielsen (2004) to produce a related abstract machine for
call-by-need.
3.1. Towards an Abstract Machine. To find call-by-need redexes tail-recursively, we apply an insight from the CK abstract machine (Felleisen and Friedman, 1986; Felleisen et al.,
2009). The CK machine implements an evaluation strategy for call-by-value based on a reduction semantics using the (inside-out) evaluation contexts , E[ t] and E[(λx.t) ]. To
find a redex, the machine iteratively examines the outermost constructor of a term and uses
the evaluation context to remember what has been discovered. Since call-by-value satisfies
unique decomposition, this approach will find the standard redex if there is one.
To illustrate this in action, we walk through an example. Consider the program
(λx.x) λy.y. Evaluation begins with configuration h[ ], (λx.x) λy.yi. Underlining indicates
subterms that the machine knows nothing about; at the beginning of evaluation, it knows
nothing about the entire term. On the first step of the reduction, the machine detects that
the term is an application (λx.x) λy.y. To examine the term further, the machine must
move its focus to either the operator or operand of this application. Since the machine is tail
recursive, it must also push an evaluation context to store the as-yet uncovered structure
6
R. GARCIA, A. LUMSDAINE, AND A. SABRY
of the term. The only context it can reliably push at this point is [ λy.y]: it cannot push
[(λx.x) ] because it has not yet discovered that the operator position is a value. So the
machine pushes the [ λy.y] context, which serves as a reminder that it is focused on the
operator of an application.
On the second step of the reduction, the machine detects that the operator is an abstraction λx.x, and observes that the innermost context is [ λy.y]. In response, the machine
pops the context, focuses on λy.y, and pushes the context [(λx.x) ], since the operator is
now known to be a value. This context serves as a reminder that evaluation is currently
focused on the operand of an application that can be reduced once that operand becomes
a value.
On the third step, the machine detects the abstraction (λy.y), and remembers that the
innermost context is [(λx.x) ]. At this point, the machine has deduced enough information
to recognize the redex (λx.x) λy.y. This example illustrates how the CK machine uses a
depth-first left-to-right search strategy to detect call-by-value redexes.
Now consider the same term under call-by-need using a similar strategy. As with callby-value, the top-level application can be detected, the operand can be pushed onto the
evaluation context, and the operator can be exposed as the abstraction λx.x. At this
point behavior must diverge from call-by-value because the body of the abstraction is still
unknown and call-by-need does not have [(λx.t) ] contexts for arbitrary t. However, callby-need does have contexts of the form [(λx.) λy.y]. Therefore, it is possible to proceed
under the first lambda abstraction, push the context, and focus on x.
The term is exposed as a variable x, which combines with the context [(λx.) λy.y]
to form the term (λx.E[x]) λy.y (where E ≡ ). At this point, enough information has
been uncovered to push the context [(λx.[x]) ] and focus on λy.y. The abstraction λy.y
is recognized, and with that a call-by-need redex (λx.[x]) λy.y has been found. Success
with this example suggests a promising strategy for implementing call-by-need reduction
tail-recursively.
3.2. An Initial Abstract Machine. In this section, we elaborate the above search strategy into a simple but inefficient tail-recursive abstract machine. We present it without proof
and then by a series of correct transformations we derive an efficient machine that we prove
correct.
This abstract machine uses the same terms, values, and answers as the calculus. However, it introduces two alternate notions. First, the machine uses a more versatile representation of evaluation contexts. As observed in Danvy and Nielsen (2004), evaluation
contexts can be mathematically specified in more than one way. For optimal flexibility, we
define evaluation contexts as lists of frames, where the empty list [ ] and single-frame lists
[f ] are our simple units, and the operator ◦ stands for list concatenation.
f ::= t | (κx.E) | (λx.) t
E ::= [ ] | [f ] ◦ E | E ◦ [f ]
where E ◦ [ ] = [ ] ◦ E = E
and E1 ◦ (E2 ◦ E3 ) = (E1 ◦ E2 ) ◦ E3
When two contexts are composed, the second context is plugged into the hole of the first
context: for example [ t2 ] ◦ [ t1 ] = [( t1 ) t2 ].
LAZY EVALUATION AND DELIMITED CONTROL
7
We call the frame [(λx.) t] a binder frame. It represents a variable binding in the
context. It can be read as [let x = t in ], but we use the former notation to emphasize that
call-by-need evaluation proceeds under lambdas. This observation motivates our analysis
of hygiene in Section 4.5.
We call the frame [(κx.E) ] a cont frame, in reference to continuations. The construction (κx.E) is called a cont and replaces the metalinguistic term notation (λx.E[x]) from
the calculus. We use a different notation for conts than lambda terms to indicate that in
the machine conts are distinct from terms (they are of type Cont rather than type Term
in an implementation). Conts indicate nontrivial structural knowledge that the machine
retains as it searches for a redex. This distinction matters when we establish continuation
semantics for machine states. As we shall see, a cont frame represents a suspended variable
reference.
Finally we call the frame [ t] an operand frame, and it represents a term waiting for
an abstraction.
The abstract machine also introduces a notion of redexes:
r ::= a t | (κx.E) a
Redexes are distinguished from terms in the machine, meaning that in an implementation,
the type Redex is distinct from the type Term. This distinction suggests that conts (κx.E)
are neither terms nor first-class entities in the call-by-need language: they only appear in
evaluation contexts and in redexes. As we discuss below, the machine distinguishes one
more kind of redex than the calculus.
The transition rules for the machine are staged into four distinct groups: refocus,
rebuild, need, and reduce. Each machine configuration can be related to a term in the
language of the calculus. The refocus rules examine the current term and push as many
operand frames [ t] as possible. A refocus configuration hE, tif represents the term E[t].
hE, tif
(Refocus)
hE, xif
hE, λx.tif
hE, t1 t2 if
7→ hE, [ ], xin
7
→
hE, λx.tib
7→ hE ◦ [ t2 ], t1 if
Upon reaching a variable, refocus transitions to the need rules; upon reaching a lambda
abstraction, it transitions to the rebuild rules.
The rebuild rules search up into the context surrounding an answer for the next applicable redex. A rebuild configuration hE, aib represents the term E[a].
hE, aib
(Rebuild)
h[ ], aib
hE ◦ [ t1 ], aib
hE ◦ [(λx.) t1 ], aib
hE1 ◦ [(κx.E2 ) ], aib
7
→
7→
7→
7→
a
hE, a t1 id
hE, (λx.a) t1 ib
hE1 , (κx.E2 ) aid
These rules examine the current context and proceed to build a maximal answer-shaped
term, progressively wrapping binder frames around the current answer. If the entire context
is consumed then evaluation has completed and the entire program is an answer. Upon
8
R. GARCIA, A. LUMSDAINE, AND A. SABRY
reaching an operand or cont frame, a redex has been found, and rebuild transitions to the
reduce rules. These rules resemble the refocusaux rules of Danvy and Nielsen (2004).
The need rules also examine the context, but they search for the binder frame that
corresponds to the variable under focus. A need configuration hE1 , E2 , xin represents the
term E1 [E2 [x]].
hE, E, xin
(Need)
hE1 ◦ [(λx.) t], E2 , xin 7→ hE1 ◦ [(κx.E2 ) ], tif
hE1 ◦ [f ], E2 , xin 7→ hE1 , [f ] ◦ E2 , xin
where, [f ] 6≡ [(λx.) t]
Since input programs are closed, the associated binder must be somewhere in the context. Upon finding the right binder frame, a cont frame [(κx.E) ] is pushed onto the
context and evaluation proceeds to refocus on the operand from the associated binder
frame.
The reduce rules simulate the notions of reduction from the calculus. A reduce configuration hE, rid represents the term E[r] where a cont κx.E represents the term λx.E[x].
hE, rid
(Reduce)
hE1 , (κx.E2 ) vid
hE1 , (κx1 .E2 ) ((λx2 .a) t)id
hE, (λx.a) t1 t2 id
hE, (λx.t1 ) t2 id
7
→
7→
7→
7→
hE1 , (λx.E2 [v]) vif
hE1 , (λx2 .(λx1 .E2 [x1 ]) a) tif
hE, (λx.a t2 ) t1 if
hE ◦ [(λx.) t2 ], t1 if
Each of the first two reduce rules transforms a cont into a lambda abstraction by
plugging its context with a term and abstracting its variable. As such, each reduce rule
transforms a redex into a pure term of the calculus and transitions to a refocus configuration,
which searches for the next redex.
The reduce rules also handle terms of the form (λx.t1 ) t2 , even though such terms are
not call-by-need redexes. Including this rule gives the set of redexes greater uniformity: all
terms of the form a t are redexes, just like the terms of the form (κx.E) a. This symmetry
is not exhibited in the call-by-need calculus. However, Ariola and Felleisen (1997) defines
and uses an auxiliary let calculus that adds the reduction
(λx.t1 ) t2 →need let x = t2 in t1
to the calculus and defines the other reductions in terms of the let expressions. The fourth
reduce rule corresponds to this reduction rule. However, our presentation shows that an
auxiliary let term, though compatible with this model, is not needed to specify call-by-need:
the syntax of pure calculus terms suffices. Furthermore, treating this rule as a reduction
here anticipates a change we make to the machine (Section 4.5) that adds explicit variable
renaming to that rule. Finally, the reduce rules are improved in the next section so that all
reduce rules change their representative terms nontrivially.
Machine evaluation of a program t begins with the refocus configuration h[ ], tif and
terminates if it arrives at an answer a. Its behavior in between can be summarized as follows:
search downwards until a value or variable reference is reached. If a variable reference is
reached, store a cont in the context to remember the variable reference and proceed to
LAZY EVALUATION AND DELIMITED CONTROL
9
evaluate its binding. If an abstraction is reached, accumulate an answer up to the innermost
redex, or the top of the evaluation context if none is found. In short, the machine performs
a depth-first, left-to-right traversal in search of a call-by-need redex. Along the way it
uses the evaluation context to store and retrieve information about program structure,
particularly the location of variable bindings (using binder frames) and variable references
(using cont frames). The refocus, rebuild, and need rules leave the term representation of
their configurations unchanged (e.g. if hE1 , t1 if 7→ hE2 , t2 if then E1 [t1 ] ≡ E2 [t2 ]), and the
reduce rules embody the notions of reduction from the calculus.
Our strategy for producing this machine builds on the strategy of Danvy and Nielsen
(2004), which mechanizes the direct transformation of reduction semantics into abstract machine semantics. That report introduces and verifies a general method for using reduction
semantics that meet certain criteria to construct a function that “refocuses” an arbitrary
term-context pair to a redex-context pair. The resulting function can then be used to induce
an abstract machine semantics. Unfortunately that refocus function construction does not
apply to the call-by-need lambda calculus because the calculus does not meet the required
criteria. In particular, the construction requires that a maximally-decomposed closed term
(i.e. program) will focus on a value or a redex. However, call-by-need evaluation contexts
can decompose down to variable references, which are neither redexes nor values under
call-by-need. There are however other ways to produce an abstract machine from a reduction semantics which may apply to the call-by-need calculus studied here. For instance,
Danvy et al. (2010) devise a variant of the let-based call-by-need reduction semantics, implement it, and use a program-transformation based approach to produce a refocus function
and abstract machine implementation.
The following partial trace demonstrates how the initial abstract machine discovers the
first redex for our running example (λx.x) λy.y:
h[ ], (λx.x) λy.yif
7→
[ λy.y], λx.x
7→
[ ], (λx.x) λy.y
d
7→
[(λx.) λy.y], x
7→
[(κx.[ ]) ], λy.y
7→
[(κx.[ ]) ], λy.y
f
7→
[ λy.y], λx.x
f
7→
[(λx.) λy.y], [ ], x
b
7→
[ ], (κx.[ ]) λy.y
f
b
n
d
4. Refining the Machine
In this section we study the behavior of the abstract machine and make some improvements based on our observations. These changes lead us from the initial machine above to
our final machine specification.
4.1. Grabbing and Pushing Conts. The need rules linearly search the nearest context
for a binder frame that matches the variable under question. This process can be specified
as one step:
hE1 ◦ [(λx.) t] ◦ E2 , xin 7→ hE1 ◦ [(κx.E2 ) ], tif
where [(λx.) t] ∈
/ E2
This evaluation step accumulates a segment of the current evaluation context and stores it.
In general, abstract machines that model control operators represent control capture in a
similar manner. In this particular case, only part of the evaluation context is captured, and
the amount of context captured depends on the dynamic location in the context of a certain
10
R. GARCIA, A. LUMSDAINE, AND A. SABRY
frame. As such, the need rules seem to perform some kind of delimited control capture. This
analogy becomes stronger upon further analysis of the first reduce rule from Section 3.2.
The machine uses its structural knowledge of κx.E to construct the abstraction λx.E[v].
However, the resulting machine configuration no longer retains any of the structure that
had previously been discovered. Recall our example execution trace from Section 3.2. The
machine reduces the redex found at the end of that trace as follows:
[ ], (κx.[ ]) λy.y
d
7→ h[ ], (λx.λy.y) λy.yif
By returning to refocus following the reduction, the machine loses all structural knowledge
of the term. To continue execution, it must examine the structure of the contractum from
scratch. Fortunately, the evaluator can be safely improved so that it retains knowledge of
the contractum’s structure:
Proposition 4.1.
hE1 , (λx.E2 [v]) vif 7−→
→ hE1 ◦ [(λx.) v] ◦ E2 , vib
Proof. Corollary of hE1 , E2 [v]if 7−→
→ hE1 ◦ E2 , vib , which is proven by induction on E2 .
This proposition justifies replacing the first reduce rule with one that pushes the evaluation context embedded in the cont and proceeds to rebuild an answer:
hE1 , (κx.E2 ) vid 7→ hE1 ◦ [(λx.) v] ◦ E2 , vib
This short-circuit rule extends the current evaluation context with a binder frame and the
context E2 that was inside the cont. The rule is suggestive of delimited control because
machine models of control operators generally represent the reinstatement of delimited
continuations by extending the current context with a piece of captured evaluation context.
Of more immediate interest, though, is how reduction of our example now proceeds:
[ ], (κx.[ ]) λy.y
d
7→ [(λx.) (λy.y)], λy.y
b
All knowledge of the contractum’s structure is retained, though much of it is now stored in
the evaluation context.
4.2. Shifting Binder Frames. The second and third reduce rules from Section 3.2 also
discard structural information. Specifically, they both transition to the forgetful refocus
rule. However their information can be preserved.
Proposition 4.2.
hE, (λx.a t2 ) t1 if 7−→
→ hE ◦ [(λx.) t1 ], a t2 id .
Proof. Corollary of hE1 , aif 7−→
→ hE1 , aib , which is proven by induction on a.
Proposition 4.3. If E2 does not capture x1 (Section 4.5), then
hE1 , (λx2 .(λx1 .E2 [x1 ]) a) tif 7−→
→ hE1 ◦ [(λx2 .) t], (κx1 .E2 ) aid .
Proof. Corollary of hE1 , aif 7−→
→ hE1 , aib and
hE1 , (λx1 .E2 [x1 ]) tif 7−→
→ hE1 ◦ [(κx1 .E2 ) ], tif , which is proven by case analysis and induction on E2 .
LAZY EVALUATION AND DELIMITED CONTROL
11
These propositions justify short-circuiting the respective evaluation rules. The new
rules improve the behavior of the abstract machine.
hE1 , (κx1 .E2 ) ((λx2 .a) t)id 7→ hE1 ◦ [(λx2 .) t], (κx1 .E2 ) aid
hE, (λx.a) t1 t2 id 7→ hE ◦ [(λx.) t1 ], a t2 id
By fast-forwarding to reduce, the rules retain the discovered term structure and thereby
avoid retracing the same terms again.
4.3. Answer = Binders × Value. The transition rules repeatedly create binder frames
out of terms and reabsorb those frames into answers. In this section we simplify this
protocol. We distinguish answers from terms by providing them a separate representation:
a ::= TE, vU, where E = [(λxi .) ti ]
An answer is now represented as a tuple containing the nested lambda abstraction and
the sequence of binder frames that are wrapped around it in the old presentation (we use
overlines to indicate sequences). This presentation bears strong similarity to calculi with
explicit substitutions (Abadi et al., 1991) in that each binder frame [(λx.) t] corresponds
to a substitution [t/x]. An answer can be seen as a lambda term nested inside a sequence
of explicit substitutions, v[ti /xi ].
The rebuild rules could be reformulated as a three place configuration, hE, E, vib , but
instead we immediately apply the same improvement that we applied to the need rules in
Section 4.1. For instance, the new transition rule for rebuilding to a cont frame is:
hE1 ◦ [(κx.E2 ) ] ◦ E3 , vib 7→ hE1 , (κx.E2 ) TE3 , vUid
where E3 = [(λxi .) ti ]
Returning to our running example, reduction from its most recent state (at the end of
Section 4.1) transitions to a final answer, signaling the end of execution:
[(λx.) λy.y], λy.y
b
7→ T[(λx.) λy.y], λy.yU
4.4. Aggregate Reduction. Now that answers explicitly carry their binders in aggregate,
the reduce rules can be substantially consolidated. Currently, the second and third reduce
rules iteratively remove the top binder frame from an answer and push it onto the evaluation
context. This process repeats until the answer is just a lambda abstraction. At that point,
the second and third reduce rules defer to the first and fourth reduce rules respectively.
This corresponds exactly with standard-order reduction (cf. Definition 2.1):
Proposition 4.4.
E[((λxn . . . . ((λx1 .((λx0 .v) t0 )) t1 ) . . . ) tn ) t] 7−→
→sr
E[((λxn . . . . ((λx1 .((λx0 .v t) t0 )) t1 ) . . . ) tn )].
E[(λx.E[x]) ((λxn . . . . ((λx1 .((λx0 .v) t0 )) t1 ) . . . ) tn )] 7−→
→sr
E[((λxn . . . . ((λx1 .((λx0 .(λx.E[x]) v) t0 )) t1 ) . . . ) tn )].
Proof. By induction on the structure of the answer term, using the unique decomposition
lemma of Ariola and Felleisen (1997).
12
R. GARCIA, A. LUMSDAINE, AND A. SABRY
Using the new answer representation, each pair of associated reduce rules can be merged
into one omnibus rule that moves all the binder frames at once and simultaneously performs
a reduction using the underlying value.
hE1 , (κx.E2 ) TE3 , vUid 7→ hE1 ◦ E3 ◦ [(λx.) v] ◦ E2 , vib
hE1 , TE2 , (λx.t1 )U t2 id 7→ hE1 ◦ E2 ◦ [(λx.) t2 ], t1 if
As a result of these transformations, both conts and answers contain evaluation contexts. Furthermore, conts and answers are not terms of the calculus, and the machine never
reverts a cont or answer to a term. The rules that create them, rebuild for answers and
need for conts, capture part of the evaluation context, and the rules that consume them,
the reduce rules, reinstate the captured contexts.
4.5. Variable Hygiene. Presentations of calculi often invoke a hygiene convention and
from then on pay little attention to bound or free variables. In this manner, calculi do
not commit to any of the numerous ways that hygiene can be enforced. Many abstract
machines, however, use environments or explicit sources of fresh names to guarantee hygiene
and thereby provide a closer correspondence to concrete implementations. In this section,
we augment the call-by-need machine with simple measures to enforce hygiene.
Our primary hygiene concerns are that evaluation occurs under binders and binders
are shifted about in unusual ways. In order to ensure that binding structure is preserved
throughout evaluation, we need to be able to reason locally, within each machine configuration, about bound variables. To make this possible, we make one last change to the
machine. We add a list of names to each machine configuration.
X ::= xi
Most of the machine rules simply pass the list of names along to the next transition. One
of the reduce rules manipulates the list of names.
hX | E1 , TE2 , λx.t1 U t2 id 7−→nam
X, x′ E1 ◦ E2 ◦ [(λx′ .) t2 ], t1 [x′ /x]
f
x′ ∈
/ X (D.2)
When this rule goes under a lambda, it adds the name of its bound variable to the list X
of variables. The notation X, x expresses adding a new name x to X. If the bound variable
x on the left hand side of the rule is already a member of X, then the variable is renamed
as part of the transition. As such, X can be considered a set.
Now each machine configuration has one of five forms:
hX | E, ?i ::= hX | TE, vUi | hX | E, rid | hX | E, tif
| hX | E, vib | hX | E, xin
We use the notation hX | E, ?i below to uniformly discuss all configuration types, where
X refers to the list of names, E refers to the context, and ? refers to the term or redex.
For a final configuration hX | TE, vUi, ? refers to the answer’s underlying value v, and E
corresponds to the answer’s binder frames E. We use the metavariable C to range over
configurations when the internal structure does not matter.
LAZY EVALUATION AND DELIMITED CONTROL
13
The call-by-need abstract machine uses the set X of names to keep track of active
variables: any variable x whose binding instance has been pushed into a binder frame
[(λx.) t]:
AV ([ ])
= ∅
AV ([(λx.) t] ◦ E)
= { x } ∪ AV (E)
AV ([ t] ◦ E)
= AV (E)
AV ([(κx.E1 ) ] ◦ E) = AV (E1 ) ∪ { x } ∪ AV (E)
Cont-bound variables are counted among the active variables because machine evaluation
must have gone under a binding to construct the cont frame.
The renaming condition on the (D.2) reduce rule ensures that active variables are
mutually distinguishable. This guarantees that the machine’s need rule can never capture
the wrong evaluation context and thus execute the wrong bound expression.
Renaming is not obviously sufficient to ensure bound variable hygiene because of how
the machine manipulates evaluation contexts. For instance, even though the need rule is
guaranteed to only match a variable with the right binder frame, we have no guarantee
that the right binder frame could never be trapped inside a cont frame and hidden from
view while a need transition searches for it. Were this to happen, the machine would get
stuck. Furthermore, the reduction rules flip and shift evaluation contexts that might contain
binder frames. If a binder frame were to end up below another context frame that contains
references to its bound variable, then those references would no longer be bound in the
context; the need rule would exhaust the evaluation context if it attempted to resolve any
of these references.
To verify that machine evaluation is well-formed, we establish well-formedness conditions that suffice to ensure safe evaluation and we show that they hold for evaluation of
all programs. The well-formedness conditions rely on straightforward notions of captured
context variables (CV ) and free context variables (F V ):
CV ([ ])
CV (E ◦ [(λx.) t])
CV (E ◦ [ t])
CV (E ◦ [(κx.E1 ) ])
=
=
=
=
∅
{ x } ∪ CV (E)
CV (E)
CV (E)
F V ([ ])
F V ([(λx.) t] ◦ E)
F V ([ t] ◦ E)
F V ([(κx.E1 ) ] ◦ E)
=
=
=
=
∅
F V (t) ∪ (F V (E) − { x })
F V (E) ∪ F V (t)
F V (E) ∪ (F V (E1 ) − { x })
As expected, these two notions are related.
Lemma 4.5.
(1) F V (E1 ) ⊆ F V (E1 ◦ E2 ).
(2) CV (E1 ◦ E2 ) = CV (E1 ) ∪ CV (E2 ).
(3) F V (E1 ◦ E2 ) = F V (E1 ) ∪ (F V (E2 )\CV (E1 )).
Proof. By induction on the length of E1 , E2 , and E1 respectively.
14
R. GARCIA, A. LUMSDAINE, AND A. SABRY
To establish that binder frames remain properly positioned, we define a notion of wellformed evaluation contexts:
Definition 4.6. A Machine context/name pair is well formed, notation X | E wf, iff:
(1) F V (E) = ∅;
(2) AV (E) = { x | x occurs in X }
(3) Each active variable of E is distinct: if E1 ◦ E2 ∈ E then AV (E1 ) ∩ AV (E2 ) = ∅, and
if [(κx.E1 ) ] ∈ E then x ∈
/ AV (E1 ).
These well-formedness criteria ensure that a context has no unbound variable references,
that active variables cannot interfere with each other, and that X is simply a particular
ordering of the E’s active variables. The captured variables of E are also distinct since
every captured variable is an active variable.
Furthermore, machine configurations also have a notion of well-formedness:
X | E wf F V (?) ⊆ CV (E)
hX | E, ?ic wf
c 6≡ d
X | (E1 ◦ [(κx.E2 ) ] ◦ E3 ) wf F V (v) ⊆ CV (E1 ◦ [(κx.E2 ) ] ◦ E3 )
hX | E1 , (κx.E2 ) TE3 , vUid wf
X | (E1 ◦ [ t] ◦ E2 ) wf F V (v) ⊆ CV (E1 ◦ [ t] ◦ E2 )
hX | E1 , TE2 , vU t2 id wf
These well-formedness conditions ensure that the evaluation contexts E are well-formed,
that the list of names X matches the active variables of E, and that the free variables
of the term under focus are captured by the context. To account for redexes, the wellformedness conditions for each reduce configuration reflect the well-formedness conditions
for the corresponding rebuild configuration. As shown below, well-formed reduce configurations hX | E, rid wf ensure that reduce rules can be safely performed without implicit
renaming.
Well-formedness of the reduce configurations ensures that the reduce transitions require
no implicit bound-variable renaming to preserve hygiene. Well-formedness of the need
configuration guarantees that it cannot be “stuck”: since x ∈ CV (E), a well-formed need
configuration always has a binder frame [(λx.) t] that matches the variable under focus,
so the configuration can transition.
Well-formedness of configurations combined with rule D.2’s name management ensures
that machine evaluation respects variable binding structure.
Theorem 4.7. If t is a closed term of the calculus, then h∅ | [ ], tif wf.
Proof. ∅ | [ ] wf and F V (t) ⊆ CV ([ ]) = ∅.
Theorem 4.8. Let C1 and C2 be configurations. If C1 wf and C1 7−→nam C2 then C2 wf.
Proof. By cases on 7−→nam . The cases are immediate except when C1 7−→nam C2 by a
reduce rule. The proofs for both kinds of reduce configurations are similar, so we present
only one of the cases:
Case (C1 = hX | E1 , (κx.E2 ) TE3 , vUid ).
By definition, C1 7−→nam C2 = hX | E, vib , where E = E1 ◦ E3 ◦ [(λx.) v] ◦ E2 . Since the
LAZY EVALUATION AND DELIMITED CONTROL
15
transition rule introduces no new active variables, X should remain the same. Furthermore,
all the active variables remain distinct, though x is now introduced by the binder frame
[(λx.) v] rather than the cont frame [(κx.E2 ) ]. It remains to show that F V (v) ⊆ CV (E)
and that F V (E) = ∅.
First, since C1 wf, it follows by inversion that X | (E1 ◦ [(κx.E2 ) ] ◦ E3 ) wf and
F V (v) ⊆ CV (E1 ◦ [(κx.E2 ) ] ◦ E3 ) = CV (E1 ◦ E3 ) ⊆ CV (E).
By Lemma 4.5, F V (E1 ) = F V (E1 ◦ [(κx.E2 ) ]) = F V (E1 ◦ [(κx.E2 ) ] ◦ E3 ) = ∅.
Since CV ([(κx.E2 ) ]) = ∅, it follows from Lemma 4.5 that F V (E3 ) ⊆ CV (E1 ) and
from the definition of F V that F V (E2 ) ⊆ CV (E1 ) ∪ { x }. From these it follows that
F V (E) = ∅.
In short, well-formedness of the reduce configurations ensures that the reduce rules
can be safely performed without any implicit renaming. Since the machine preserves wellformedness, this property persists throughout evaluation. The rest of this paper only considers well-formed configurations.
4.6. An Abstract Machine for Call-by-need. Putting together our observations from
the previous section, we now present the specification of the abstract machine. Figure 1
presents its 7−→nam transitions rules. We have derived a heap-less abstract machine for
call-by-need evaluation. It replaces the traditional manipulation of a heap using store-based
effects with disciplined management of the evaluation stack using control-based effects. In
short, state is replaced with control.
Machine evaluation of a program t begins with h∅ | [ ], tif and terminates at hX | TE, vUi.
5. Correctness of the Machine
The previous section proves that the machine manipulates terms in a manner that
preserves variable binding. In this section, we prove that those manipulations correspond
to standard-order call-by-need evaluation.
To proceed, we first establish correspondences between abstract machine configurations and call-by-need terms. As we have alluded to previously, abstract machine contexts
correspond directly to calculus contexts:
C[[[ ]]]
C[[[ t] ◦ E]]
C[[[(κx.E1 ) ] ◦ E2 ]]
C[[[(λx.) t] ◦ E]]
=
=
=
=
C[[E]] t
(λx.C[[E1 ]][x]) C[[E2 ]]
(λx.C[[E]]) t
Redexes also map to call-by-need terms:
C[[TE, vU t]]
= (C[[E]][v]) t
C[[(κx.E1 ) TE2 , vU]] = (λx.C[[E1 ]][x]) (C[[E2 ]][v])
Given that terms map identically to terms, configuration mapping is defined uniformly:
C[[hX | E, ?i]] = C[[E]][ C[[?]] ]
16
R. GARCIA, A. LUMSDAINE, AND A. SABRY
hX | E, rid
(Reduce)
(D.1) hX | E1 , (κx.E2 ) TE3 , vUid
(D.2)
hX | E1 , TE2 , λx.t1 U t2 id
hX | E, tif
(F.1)
(F.2)
(F.3)
hX | E1 ◦ E3 ◦ [(λx.) v] ◦ E2 , vib
hX, x′ | E1 ◦ E2 ◦ [(λx′ .) t2 ], t1 [x′ /x]if
x′ ∈
/X
(Refocus)
hX | E, xif
hX | E, λx.tif
hX | E, t1 t2 if
hX | E, vib
7−→nam
7−→nam
7−→nam
7−→nam
7−→nam
hX | E, xin
hX | E, λx.tib
hX | E ◦ [ t2 ], t1 if
(Rebuild)
(B.1)
hX | Eb , vib
(B.2)
hX | E1 ◦ [ t] ◦ Eb , vib
(B.3) hX | E1 ◦ [(κx.E2 ) ] ◦ Eb , vib
7 →nam
−
7−→nam
7−→nam
hX | TEb , vUi
hX | E1 , TEb , vU tid
hX | E1 , (κx.E2 ) TEb , vUid
where Eb = [(λxi .) ti ]
hX | E, xin
(Need)
(N.1) hX | E1 ◦ [(λx.) t] ◦ E2 , xin
where [(λx.) t] ∈
/ E2
7−→nam
hX | E1 ◦ [(κx.E2 ) ], tif
Figure 1: Call-by-need Machine
Since the calculus is defined over alpha equivalence classes, we reason up to alpha equivalence
when relating terms to machine configurations.
We now state our fundamental correctness theorems. First we guarantee soundness,
the property that every step of the abstract machine respects standard-order reduction.
Theorem 5.1. If t1 = C[[C1 ]] and C1 7−→nam C2 , then t1 7−→
→sr t2 , for some t2 = C[[C2 ]].
Proof. By cases on 7−→nam . Only rules D.1 and D.2 are not immediate. The other rules
preserve equality under C[[C]].
Corollary 1 (Soundness).
If t = C[[C]] and C 7−→
→nam hX | TE, vUi, then t 7−→
→sr a, for some a = C[[hX | TE, vUi]].
Proof. By induction on the length of the 7−→
→nam sequence.
We also prove completeness, namely that abstract machine reduction subsumes standard order reduction.
Theorem 5.2 (Completeness).
If t = C[[C]] and t 7−→
→sr a, then C 7−→
→nam hX | TE, vUi, with a = C[[hX | TE, vUi]].
Proof. This proof proceeds by induction on the length of 7−→
→sr sequences. It utilizes
Proposition 4.4 to accelerate the 7−→sr rules in accordance with 7−→nam . It also relies on a
number of lemmas to establish that 7−→nam will find the unique redex of a term from any
decomposition of a term into a context E and a subterm t.
LAZY EVALUATION AND DELIMITED CONTROL
17
Theorem 5.3 (Correctness). If t = C[[C]], then
t 7−→
→sr a
if and only if
C 7−→
→nam hX | TE, vUi
with a = C[[hX | TE, vUi]].
5.1. Discussion. This abstract machine has nice inductive properties. The refocus rules
always dispatch on the outermost term constructor. The rebuild and need rules dispatch
on a prefix of the context, though each has different criteria for bounding the prefix.
The abstract machine’s evaluation steps should not be seen as merely a desperate search
for a redex. Rather, the machine exposes the fine-grain structure of call-by-need evaluation,
just as the CK machine and the Krivine machine (Krivine, 2007) model evaluation for callby-value and call-by-name respectively. Answers are the partial results of computations, and
the rebuild rules represent the process of reconstructing and returning a result to a reduction
site. Furthermore, the need rules can be viewed as a novel form of variable-lookup combined
with lazy evaluation. The evaluation context captures the rest of computation, but not in
order: variable references cause evaluation to skip around in a manner that is difficult to
predict.
The way that variables behave in these semantics reveals a connection to coroutines.
The reduction rule D.2 binds a variable to a delayed computation; referencing that variable
suspends the current computation and jumps to its associated delayed computation. Upon
completion of that computation, any newly delayed computations (i.e. binder frames) are
added to the evaluation context and the original computation is resumed.
The standard-order reduction relation of the call-by-need lambda calculus defines an
evaluator concisely but abstractly. Surely unique decomposition, standardization, and hygiene ensure the existence of a deterministic evaluator, but these properties do not spell out
the details or implications. Based on a reasoned inspection of standard-order reduction, we
expose its computational behavior and capture it in a novel abstract machine that has no
store. The improvements to the initial machine produce a variant that effectively assimilates computational information, explicitly accounts for variable hygiene and thereby reveals
coarse-grained operational structure implicit in call-by-need standard-order evaluation.
5.2. Extensions. The machine presented above describes evaluation only for the pure
lambda calculus. In this subsection, we introduce some features that are typical of pragmatic programming languages.
5.2.1. Let binding. To help with a proof that the call-by-need calculus can simulate callby-name, Ariola and Felleisen introduce a let-based calculus. The let-calculus adds the
construction let x = t in t to the set of terms and considers a new axiom:
(λx.t1 ) t2 = let x = t2 in t1
The let-calculus is formulated by taking this axiom as a reduction rule running from left to
right, and reformulating the original three axioms of the call-by-need calculus in terms of
let expressions. These new axioms can be justified by proving that they are derivable from
the original three axioms and the new let axiom.
One approach to producing an abstract machine that supports let is to repeat the
entire process described in this section, but focusing on the let-calculus instead of the
18
R. GARCIA, A. LUMSDAINE, AND A. SABRY
lambda calculus. However, let-binding can be retro-fitted to the current lambda calculus
much more simply by reading the new let axiom as a right-to-left reduction rule:
let x = t2 in t1 →need (λx.t1 ) t2
Essentially, this rule indicates that let bindings in the calculus can be understood as equivalent to an immediate application of an abstraction to a term.
The consequences for the naı̈ve machine are the addition of a new kind of redex:
r ::= . . . | let x = t2 in t1
and a new refocus rule:
hE, let x = t2 in t1 if 7→ hE, let x = t2 in t1 id
and a new reduce rule:
hE, let x = t2 in t1 id 7→ hE, (λx.t1 ) t2 if
Then, in the process of improving our machine, the reduce rule can be fast-forwarded to
reduce the introduced abstraction and application immediately:
hX | E, let x = t2 in t1 id 7→ X, x′ E ◦ [(λx′ .) t2 ], t1 [x′ /x]
f
This process confirms that the let expression form can be comfortably treated as a
conservative extension of the call-by-need lambda calculus. Its addition does not force a
radical reconstruction of the abstract machine. As we show in Section 5.4, adding a circular
letrec construct to the language fundamentally alters the system.
5.2.2. Constants. Plotkin (1975) augments the lambda calculus with two sets of constants,
the basic constants b and the function constants f . In that language, the basic constants
are observables and serve as placeholders for real programming language constants like
numbers, strings, and so on. The function constants serve as placeholders for real primitive
functions, generally over basic constants. The function constants are not first-class values
in the language, but instead appear as operators in primitive function expressions.
t ::= . . . | b | f t
v ::= . . . | b
Function expressions always take a single term argument.
The calculus is parametrized on a partial function δ that maps function-value pairs to
values. As is standard, the notions of reduction for the calculus are augmented to handle
function constant applications.
f v1 →need
v2 if δ(f, v1 ) = v2
Since the notions of reduction for function constants require them to be applied to values,
the calculus must account for answers. To handle this, the calculus can be extended with
a rule to commute answer bindings with function expressions, as is done for application
expressions:
f ((λx.a) t) →need (λx.f a) t
If function expressions were only defined for basic constant arguments, then answer bindings could be discarded instead of commuted. However this approach imposes an ad hoc
LAZY EVALUATION AND DELIMITED CONTROL
19
limitation on the semantics. It would be cleaner and more orthogonal to uniformly handle
garbage collection with a specific notion of reduction (Ariola and Felleisen, 1997).
Function constants must be considered in the evaluation contexts. A function constant
must force its argument in order to produce a value, so upon encountering a function
constant application, evaluation must continue in its argument:
E ::= . . . | f E
The other notions of reduction remain exactly the same, but since basic constants are
values, they are subject to rules that manipulate values and answers. For example, if the set
of basic constants includes numbers and the set of function constants includes operations
on numbers, then by the deref rule:
(λx.add1 x) 5 →need (λx.add1 5) 5
The abstract machine requires few changes to accommodate these additions. The set
of redexes is extended to include function expressions:
r ::= . . . | f a
The refocus rules are generalized to create context frames for function constant applications
and to rebuild when any value, lambda abstraction or basic constant, is encountered:
hX | E, f tif
hX | E, vif
7−→nam hX | E ◦ [f ], tif
7−→nam hX | E, vib
Furthermore, rebuild must account for function constant applications:
hX | E ◦ [f ] ◦ Eb , vif 7−→nam hX | E, f TEb , vUid
Finally, the reduce rule for f a, pushes binder frames upwards and appeals to the delta rule
δ(f, v) for its result:
hX | E, f TEb , vUid 7−→nam hX | E ◦ Eb , δ(f, v)ib
Since δ(f, v) yields only values, the reduce configuration for f a can immediately transition
to rebuild from this result.
5.3. Lazy Constructors. As pointed out by Ariola and Felleisen (1997), the call-by-need
lambda calculus can be easily extended with support for lazy constructors. The rules for
constructors and destructors can be inferred from the church encoding for pairs:
cons ≡ λx1 .λx2 .λd.d x1 x2
car ≡ λp.p (λx1 .λx2 .x1 )
cdr ≡ λp.p (λx1 .λx2 .x2 )
To add support for lazy pairs, we first extend the syntax of the language:
t ::= . . . | cons t t | car t | cdr t | hx, xi
v ::= . . . | hx, xi
E ::= . . . | car E | cdr E
The cons t1 t2 expression creates a lazy pair of its two arguments, t1 and t2 . The car t
and cdr t expressions evaluate their respective arguments and extract the first or second
element of the resulting pair. As with function constants in the previous section, the lazy
pair constructor and destructors are second-class. To express first-class constructors, these
20
R. GARCIA, A. LUMSDAINE, AND A. SABRY
may be eta-expanded into, e.g., λx1 .λx2 .cons x1 x2 . The hx1 , x2 i value is a representation
of a lazy pair. It contains variables that refer to shared computations.
According to the evaluation contexts, evaluation may focus on the argument of a car
or cdr expression. However, evaluation never directly operates on the arguments to a
constructor. This is why the evaluation contexts for the language do not examine the
arguments to cons. As a direct consequence, the standard call-by-need reduction rules will
never substitute a value for a variable inside a pair. A variable inside a pair can only be
evaluated after decomposing the pair using car or cdr.
The Church encoding of pairs motivates the following notions of reduction:
cons t1 t2 →need
car hx1 , x2 i →need
cdr hx1 , x2 i →need
(λx1 .(λx2 . hx1 , x2 i) t2 ) t1
x1
x2
These rules model cons as it is applied to two arguments. The car and cdr operations each
expose a previously inaccessible variable reference to evaluation.
To accommodate lazy pairs in the abstract machine, we extend the set of redexes:
r ::= . . . | cons t1 t2 | car a | cdr a
and we introduce several new rules. The first set of rules extends the refocus stage of the
machine to handle our extensions.
hX | E, cons t1 t2 if
hX | E, car tif
hX | E, cdr tif
7 →nam hX | E, cons t1 t2 id
−
7−→nam hX | E ◦ [car ], tif
7−→nam hX | E ◦ [cdr ], tif
A cons expression is immediately ready for reduction. The car and cdr expressions proceed
to evaluate their respective arguments. Another set of rules returns an answer to car or
cdr.
hX | E ◦ [car ] ◦ Eb , vib 7−→nam hX | E, car TEb , vUid
hX | E ◦ [cdr ] ◦ Eb , vib 7−→nam hX | E, cdr TEb , vUid
Finally, a set of rules corresponds to the notions of reduction.
hX | E, cons t1 t2 id −
7 →nam hX, x1 , x2 | E ◦ [(λx1 .) t1 ] ◦ [(λx2 .) t2 ], hx1 , x2 iib
hX | E1 , car TE2 , hx1 , x2 iUid −
7 →nam hX | E1 ◦ E2 , x1 in
hX | E2 , cdr TE2 , hx1 , x2 iUid 7−→nam hX | E1 ◦ E2 , x2 in
Lazy pair construction creates new binder frames for the two terms and produces a pair that
references them. To evaluate a destructor, the binder frames associated with the answer
are pushed upwards and the corresponding variable reference is extracted from the pair and
the value of its associated computation is demanded.
5.4. Circularity. As pointed out by Ariola et al. (1995), the presence of constructors
makes it possible to observe duplicated constructors when the Y combinator is used to
express recursion.
Consider the expression:
Y (λy.cons 1 y) ≡ (λf.(λx.f (x x)) (λx.f (x x))) (λy.cons 1 y)
LAZY EVALUATION AND DELIMITED CONTROL
21
Standard-order reduction in the calculus proceeds as follows, underlining either the
active variable reference or the current redex:
(λf.(λx.f (x x)) (λx.f (x x))) (λy.cons 1 y)
→need (λf.(λx.(λy.cons 1 y) (x x)) (λx.f (x x))) (λy.cons 1 y)
→need (λf.(λx.(λy.(λx1 .(λx2 . hx1 , x2 i) y) 1) (x x)) (λx.f (x x))) (λy.cons 1 y)
Because of the two references to f in Y, the term (λy.cons 1 y) is copied and ultimately
recomputed each time the term is needed. Ideally a recursive hx, xi value could refer to
itself and not recompute its value whenever its cdr is demanded. Duplicated computation
does not correspond to how recursion and lazy constructors interact in typical semantics for
lazy evaluation (Henderson and Morris, 1976). An implementation would create a single
self-referencing cons cell.
Because of the scoping rules for λ, there is no way to explicitly define truly circular
structures in this calculus. To address this, Ariola and Felleisen introduce a letrec-based
calculus, inspired by the circular calculus of Ariola and Klop (1994, 1997). The syntax of
the letrec-calculus follows:
t
D
v
a
E
::=
::=
::=
::=
::=
|
|
D[x, xn ] ::=
x | λx.t | t t | letrec D in t
x1 be t1 , . . . , xn be tn
λx.t
v | letrec D in a
| E t | letrec D in E
letrec D, x be E in E[x]
letrec xn be E, D[x, xn ] in E[x]
x be E[x1 ], x1 be E[x2 ], . . . , xn−1 be E[xn ], D
The letrec-calculus is similar to the let-calculus in how it adds an explicit binding form, but
each letrec expression can contain an unordered set D of mutually recursive bindings x be t
for distinct variables.
The evaluation contexts for the letrec-calculus resemble those for the let-calculus. The
contexts letrec D, x be E in E[x] and letrec xn be E, D[x, xn ] in E[x] represent evaluation taking place in the binding position of a letrec expression. The first form expresses that a
variable has been referenced in the body of the letrec and that variable’s definition is currently under evaluation. The second form expresses that some variables bound in the letrec
form depend on each other, and a variable reference in the body of the letrec has forced
evaluation of this chain of variable references. This chain of dependencies is denoted by the
syntax D[x, xn ]. The definition of the last referenced variable in the chain of dependencies,
xn , is currently being evaluated. For evaluation to proceed, all the variables in a chain of
dependencies must be disjoint. The letrec calculus regards a cyclic dependency chain as a
diverging computation, a stuck expression that is not a valid answer.
22
R. GARCIA, A. LUMSDAINE, AND A. SABRY
The notions of reduction for the letrec calculus follow:
(1)
(2)
(3)
(4)
(5)
(λx.t1 ) t2
letrec D, x be v in E[x]
letrec xn be v, D[x, xn ] in E[x]
(letrec D in a) t
letrec xn be (letrec D in a), D[x, xn ]
in E[x]
(6) letrec D1 , x be (letrec D2 in a) in E[x]
→need
→need
→need
→need
→need
letrec x be t2 in t1
letrec D, x be v, in E[v]
letrec xn be v, D[x, v] in E[x]
letrec D in a t
letrec D, xn be a, D[x, xn ] in E[x]
→need letrec D2 , D1 , x be a in E[x]
The rules operate as follows. Rule (1) is analogous to the equivalent let-calculus rule.
Rule (2) is analogous to the basic dereference rule. Rule (3) resolves the last variable
reference in a chain of dependency. The syntax D[x, v] expresses replacing xn−1 be E[xn ]
with xn−1 be E[v] in the list of bindings. Rules (4) through (6) are associativity rules that
percolate bindings upward to ensure proper sharing. Rules (5) and (6) are important for
recursion. They lift recursive bindings from the definition of a letrec-bound variable x and
incorporate them into the same letrec expression that binds x. In essence they expand the
set of variables bound by the outer letrec expression.
Using the letrec-calculus, the corresponding example using cons is as follows:
letrec y be cons 1 y in y
→need letrec y be (letrec x1 be 1 in (letrec x2 be y in hx1 , x2 i)) in y
→need letrec x1 be 1, y be (letrec x2 be y in hx1 , x2 i) in y
→need letrec x1 be 1, x2 be y, y be hx1 , x2 i in y
→need letrec x1 be 1, x2 be y, y be hx1 , x2 i in hx1 , x2 i
In this reduction, the recursive cons cell has been computed once and for all. No reference
to the original cons lazy constructor remains.
5.4.1. The letrec machine. Now we express the call-by-need letrec-calculus as an abstract
machine. Development of this machine proceeds along the same lines as the non-circular
machine.
As described in the calculus, letrec bindings need not be ordered, but dependency chains
have a natural order imposed by the order of dependencies. Furthermore, cyclic dependency
chains are provable divergences in the letrec-calculus. In the face of such a cycle, a program
is no longer subject to reduction. We can model this by letting the machine get “stuck”.
The basic terms for the machine remain close to those of the calculus:
t ::= x | λx.t | t t | letrec D + in t
D ::= xi be ti
v ::= λx.t
As with the calculus, D refers to sets of recursive bindings. As needed, we distinguish
possibly empty sets of bindings, D ∗ , from nonempty sets D + . This precision is needed to
discuss machine behavior. The values of this machine are the lambda abstractions, but now
answers are defined as values wrapped in zero or more tiers of letrec bindings.
As we did for the prior machine, we introduce some representation changes that help
with presenting the letrec-machine. The evaluation context is once again replaced with a
LAZY EVALUATION AND DELIMITED CONTROL
23
list of context frames.
Pxx0n ::= {hxn , En i :: hxn−1 , En−1 i :: · · · :: hx0 , E0 i}
f
::= t | LR D + in
| LR x be , D ∗ in E
| LR xn+1 be , Pxx0n , D ∗ in E
E
::= [ ] | [f ] ◦ E | E ◦ [f ]
where E ◦ [ ] = [ ] ◦ E = E
and E1 ◦ (E2 ◦ E3 ) = (E1 ◦ E2 ) ◦ E3
The operand frame [ t] and binding frame [LR D + in ] are directly analogous to the corresponding frames in the original abstract machine. The original cont frame, on the other
hand, splits into two variants. The frame [LR xn+1 be , Pxxn0 , D ∗ in E[x0 ]] captures chain of
dependencies that is currently under evaluation. The expression D ∗ stands for the inactive
bindings in the frame, while the expression Pxxn0 indicates a chain of dependencies. While
no particular ordering is imposed on the bindings in D ∗ , the chain of dependencies represented by Pxx0n is ordered. Our machine representation separates dependencies from other
bindings. In the calculus, D[x0 , xn ] also includes any other bindings D. In the machine,
Pxx0n only captures the dependencies, and the other bindings D ∗ are explicitly indicated.
Machine dependencies Pxx0n ≡ {hxn , En i :: hxn−1 , En−1 i :: · · · :: hx0 , E0 i} correspond to calx
culus dependencies D[x0 , xn ] ≡ x be E[x1 ], x1 be E[x2 ], . . . , xn−1 be E[xn ]. We allow Px0−1
to denote an empty chain. Observe that the order of the machine dependencies is reversed.
This ordering expresses that dependencies are resolved in last-in first-out order. When the
value of xn+1 is computed, its value will be used to compute the value of xn and so on
until the value of x0 is computed and its value is returned to the body of the letrec expression. Machine dependencies form a stack of computations that reference one another. This
behavior is clarified in the behavior of the cyclic abstract machine.
The frame [LR x be , D ∗ in E] closely resembles the old cont frame [(κx.E) ], except
that the LR form may bind other variables as well. The absence of Pxx0n in this frame
explicitly indicates that it has no chain of dependencies to be evaluated before substituting
into the body of the LR .
Machine answers are still binding-value pairs, but now each binding is a recursive binding of multiple variables.
a ::= TE, vU where E = [LR Di+ in ]
The set of redexes for the abstract machine follows.
r ::= a t
| LR x be a, D ∗ in E
| LR xn+1 be a, Pxx0n , D ∗ in E
The first redex form is the same as a form from the original machine. The second form is
analogous to the (κx.E) a form from the original machine. The final redex form captures
the case where one link in the chain of dependencies is about to be resolved.
Figure 2 presents the transition rules of the cyclic abstract machine. The machine relies
on an operator F(Eb ), which given a list of binder frames [LR Di+ in ], flattens them into
a single binder frame [LR Di+ in ]. This operation captures in aggregate the treatment of
bindings by rules (5) and (6) of the calculus.
24
R. GARCIA, A. LUMSDAINE, AND A. SABRY
The first three refocus rules are the same as the lambda calculus, while the fourth rule
is analogous to the equivalent rule for the let-calculus. The rebuild rules are also analogous
to the lambda calculus, though now the binder frames have the form [LR Di+ in ].
The letrec-machine has two need rules, (N.1) for a variable reference in the body of
the corresponding letrec , and (N.2) for a variable reference that extends a (possible empty)
chain of dependencies. The need rule does not consider variable bindings that are currently
in a dependency chain, so evaluation will get stuck upon arriving at a cycle.
There are four reduce rules. The (D.1) rule, substitutes a value into the body of a
letrec after its value has been computed. The (D.2) rule resolves the most recent reference
in a chain of dependencies. Having computed the value of xn+1 , it returns to computing the
value of xn , which needed xn+1 ’s value. If the chain of dependencies has only one element
(i.e. Pxx00 ), then the chain is fully resolved. The (D.3) rule handles when an answer is
applied to an expression. It combines calculus rules (4) and (1). Finally, the (D.4) rule
handles letrec occurrences in the source program. In order to address hygiene, this rule
must simultaneously substitute for every bound variable in each binding as it focuses on
the body of the letrec expression.
The following is a machine trace of the cyclic cons example:
7→
7→
7→
7→
7→
7→
7→
7→
7→
h∅ | [ ], letrec y be cons 1 y in yif
h∅ | [ ], letrec y be cons 1 y in yid
h{ y } | [LR y be cons 1 y in ], yif
h{ y } | [LR y be cons 1 y in ], yin
h{ y } | [LR y be in [ ]], cons 1 yif
h{ y } | [LR y be in [ ]], cons 1 yid
h{ y, x1 , x2 } | [LR y be in [ ]] ◦ [LR x1 be 1 in ] ◦ [LR x2 be y in ], hx1 , x2 iib
h{ y, x1 , x2 } | [ ], LR y be T[LR x1 be 1 in ] ◦ [LR x2 be y in ], hx1 , x2 iU in [ ]id
h{ y, x1 , x2 } | [LR y be hx1 , x2 i, x1 be 1, x2 be y in ], hx1 , x2 iib
h{ y, x1 , x2 } | T[LR y be hx1 , x2 i, x1 be 1, x2 be y in ], hx1 , x2 iUi
6. Simulating Call-by-need Using Control
As we allude to above, call-by-need machine evaluation is highly suggestive of delimited
control operations, but the connection is indirect and mixed with the other details of lazy
evaluation. In this section, we directly interpret this connection in the terminology of
delimited control.
Based on the operational behavior of the abstract machine from Figure 1, we derive a
simulation of call-by-need execution under call-by-value augmented with delimited control
operators. In particular, we translate call-by-need terms into the framework of Dybvig et al.
(2007). First we overview the language of delimited control operations. Then we describe
how the abstract machine performs delimited control operations. Next we present the
simulation of call-by-need using delimited control. Finally we show its correctness.
hX | E, rid
(D.1)
(D.2)
(D.3)
(D.4)
hX | E1 , (LR x be TE2 , vU, D∗ in E3 )id
X E1 , (LR xn+1 be TE2 , vU, Pxx0n , D∗ in E3 ) d
D hX | E1 , TE2 , λx.t1 U t2Eid
X E, letrec xi be ti in t
hX | E, tif
7−→nam
7−→nam
7−→nam
7−→nam
(N.1)
f
x′i ∩ X = ∅
hX | E, xin
hX | E, λx.tib
hX | E ◦ [ t2 ], t1 if
hX | E, letrec D+ in tid
(Rebuild)
(B.1)
hX | Eb , vib
(B.2)
hX | E1 ◦ [ t] ◦ Eb , vib
(B.3)
hX | E1 ◦ [(LR x be , D∗ in E2 )] ◦ Eb , vib
(B.4) X E1 ◦ [(LR xn+1 be , Pxx0n , D∗ in E2 ] ◦ Eb , v b
where Eb = [LR Di+ in ]
hX | E, xin
X, x′i E ◦ [LR x′i be ti [x′i /xi ] in ], t
7−→nam
b
(Refocus)
hX | E, xif
hX | E, λx.tif
hX | E, t1 t2 if
hX | E, letrec D+ in tif
hX | E, vib
d
hX | E1 ◦ [LR x be v, F (E2 ), D∗ in ] ◦ E3 , vib
x
X E1 ◦ [LR xn be , Px0n−1 , xn+1 be v, F (E2 ), D∗ in E3 ] ◦ En , v
′
′
′
′
hX,
/X
D x | E1 ◦ E2 ◦ [LR x be t2 in ], t1 [x
E /x]if x ∈
7−→nam
7−→nam
7−→nam
LAZY EVALUATION AND DELIMITED CONTROL
(F.1)
(F.2)
(F.3)
(F.4)
(Reduce)
7−→nam
7−→nam
7−→nam
7−→nam
hX | TEb , vUi
hX | E1 , TEb , vU tid
hX | E1 , (LR x be TEb , vU, D∗ in E2 )id
X E1 , (LR xn+1 be TEb , vU, Pxx0n , D∗ in E2 )
d
(Need)
D
E
X E1 ◦ [LR (x be t, D∗ ) in ] ◦ E2† , x
(N.2) X E1 ◦ [LR xn be
where (x be t) ∈
/ E†
x
, Px0n−1 , (x
be t, D∗ ) in E2 ] ◦ En† , x
n
n
7−→nam
7−→nam
hX | E1 ◦ [LR x be , D∗ in E2 ], tif
X E1 ◦ [LR x be , Pxx0n , D∗ in E2 ], t
f
Figure 2: Letrec Machine
25
26
R. GARCIA, A. LUMSDAINE, AND A. SABRY
6.1. Delimited Control Operators. Dybvig et al. (2007) define a language with delimited control operators. We explain these operators using a simplified variant of the defining
machine semantics.
t
v
E
M
p
::=
|
::=
::=
|
|
::=
∈
x | v | t t | newPrompt | pushPrompt t t
withSubCont t t | pushSubCont t t
λx.t | p | hM i
| E[ t] | E[(λx.t) ] | E[pushPrompt t]
E[withSubCont t] | E[withSubCont p ]
E[pushSubCont t]
[]|E:M |p:M
N
The language extends the call-by-value untyped lambda calculus with the four operators
newPrompt, pushPrompt, withSubCont, and pushSubCont as well as two new values: firstclass prompts p, and first-class delimited continuations hM i. Its control structure is defined
using evaluation contexts E, and metacontexts M , which are lists that interleave prompts
and contexts. Metacontexts use Haskell list notation. Prompts are modeled using natural
numbers.
A program state comprises an expression t, continuation E, metacontinuation M , and
fresh prompt source p. The initial state for a program t is [t], [ ], 0.
E[(λx.t) v], M, p
E[newPrompt], M, p
E[pushPrompt p1 t], M, p2
E[withSubCont p1 λx.t], M1 ++(p1 : M2 ), p2
where p1 ∈
/ M1
E[pushSubCont hM1 i t], M2 , p
[v], E : M, p
[v], p1 : M, p2
7
→
7→
7
→
7
→
E[t[v/x]], M, p
E[p], M, p + 1
[t], p1 : E : M, p2
[t[hE : M1 i /x]], M2 , p2
7
→
[t], M1 ++(E : M2 ), p
7→ E[v], M, p
7
→
[v], M, p2
The four operators manipulate delimited continuations, or subcontinuations, which are
part of an execution context. The withSubCont operator takes a prompt and a function; it
captures the smallest subcontinuation that is delimited by the prompt and passes it to the
function. The non-captured part of the continuation becomes the new continuation. The
prompt instance that delimited the captured subcontinuation is discarded: it appears in
neither the captured subcontinuation nor the current continuation. This operator generalizes F (Felleisen, 1988) and shift (Danvy and Filinski, 1990).
The pushSubCont operator takes a subcontinuation and an expression; it composes the
subcontinuation with the current continuation and proceeds to evaluate its second argument
in the newly extended continuation.
The pushPrompt operator takes a prompt and an expression; it extends the current
continuation with the prompt and evaluates the expression in the newly extended continuation. The newPrompt operator returns a distinguished fresh prompt each time it is
called. These two operators generalize the delimiting operators # (Felleisen, 1988) and
LAZY EVALUATION AND DELIMITED CONTROL
27
reset (Danvy and Filinski, 1990), which extend a continuation with a single common delimiter.
To illustrate these operators in action, we consider a program that uses arithmetic and
conditionals:
let p = newPrompt
in 2 + pushPrompt p
if (withSubCont p
(λk.(pushSubCont k False)+
(pushSubCont k True)))
then 3
else 4
A fresh prompt is bound to p and pushed onto the continuation just prior to evaluation
of the if expression. withSubCont captures the subcontinuation [if then 3 else 4], which
was delimited by p, and binds it to k. The subcontinuation k is pushed twice, given the
value False the first time and True the second. The result of evaluation is the expression
2 + 4 + 3 which yields 9.
6.2. Delimited Control Naı̈vely Simulates the Machine. The call-by-need abstract
machine performs two different kinds of partial control capture. To review, the rebuild and
need rules of the abstract machine both capture some portion of the evaluation context.
In particular, the rebuild rules capture binder frames. If only binder frames remain, then
execution is complete. When either of the other frames is found, then a reduction is performed. On the other hand, the need rule captures the evaluation context up to the binder
that matches the variable whose value is needed.
These actions of the abstract machine can be recast in the language of delimited control
capture. First, the need rule uses the identity of its variable, which must be an active
variable, to delimit the context it captures. The well-formedness conditions from Section 4.5
guarantee that each binder frame binds a unique variable, so each active variable acts as
a unique delimiter. Second, the rebuild rule uses the nearest non-binder frame to delimit
the context it captures. This means that rebuild operates as though the operand frames,
the cont frames, and the top of the evaluation context share a common delimiter. This
guarantees that only binder frames are captured (as is stipulated in the rules).
In short, call-by-need evaluation captures partial evaluation contexts. These partial
evaluation contexts correspond to delimited continuations, and there are two different kinds
of delimitation, redex-based (for rebuild) and binder-based (for need).
It is useful to also consider how the machine manipulates these delimited continuations.
Each reduce rule in Figure 1 immediately pushes the context associated with an answer onto
the current evaluation context. In this manner, binders are consistently moved above the
point of evaluation. The reduce rule then operates on the value part of the answer and the
associated cont (for D.1) or term (for D.2).
Although each reduce rule pushes binders onto the evaluation context, only the D.2 rule
creates new binders. The variable bound by the answer’s underlying lambda abstraction
may already be a member of the set X, in which case it must be alpha-converted to a fresh
name with respect to the set X. Also note that if λx.t is alpha converted to λx′ .t[x′ /x], the
body under call-by-value satisfies the equation t[x′ /x] = (λx.t) x′ . Since we are using the
identifiers x′ as delimiters, and we never turn the binder frame [(λx′ .) t] back into a term,
we can replace fresh variables x′ with fresh prompts (Balat et al., 2004).
28
R. GARCIA, A. LUMSDAINE, AND A. SABRY
Let s be a distinguished identifier:
N P [[t]] = runCC (let s = newPrompt in pushPrompt s N [[t]])
N [[x]] = need x
N [[t1 t2 ]] = do va ⇐ N [[t1 ]]
in let xp = newPrompt
in delay N [[t2 ]] as xp in (va xp )
N [[λx.t]] = return λx.N [[t]]
return va ≡ withSubCont s λka . hka , va i
do x ⇐ t1 in t2 ≡ let hka , xi = pushPrompt s t1
in pushSubCont ka t2
delay t1 as x in t2 ≡ let fk = pushPrompt x t2
in fk λ().t1
force f ≡ f ()
need x ≡ withSubCont x λk.
λfth .do va ⇐ force fth
in delay (return va ) as x
in pushSubCont k (return va )
Figure 3: Translating CBN to CBV+Control
From these observations, we construct the simulation in Figure 3. The simulation can
be understood as a direct encoding of the abstract machine semantics for call-by-need.
To execute a program, N P [[t]], the transformation uses runCC to initiate a control-based
computation, acquires a fresh prompt, and binds it to a distinguished variable s. This
prompt is the redex prompt, which is used to delimit every continuation that denotes a
redex.
To expose the conceptual structure of the simulation, we define five syntactic macros,
do, return, delay, force, and need. We accord no formal properties to them: they merely
simplify the presentation. The return macro captures the nearest subcontinuation that is
delimited by the redex prompt s. Since the s delimiter appears before every reduction,
the captured continuation is guaranteed to contain only code equivalent to binder frames.
The translation returns a tuple containing the subcontinuation and the argument to return,
which must be a value; the tuple represents an answer. So the translation rule for lambda
abstractions, N [[λx.t]], literally simulates the rebuild rules.
The do macro executes a term t1 under the current continuation extended with the
redex prompt. If the term returns an answer hka , xi it immediately pushes the subcontinuation part and continues execution, binding the value part to the variable x. As such, the
translation rule for applications, N [[t1 t2 ]], executes [[t1 ]] and binds the resulting operator to
LAZY EVALUATION AND DELIMITED CONTROL
29
va . The answer binders are pushed by the do macro, which starts the simulation of the D.2
rule.
The remainder of the D.2 rule is subtle. In the abstract machine, binder frame variables
delimit the need rules. Since the delimited continuation framework relies on prompts to
delimit continuations, fresh prompts literally substitute for variables (Kiselyov et al., 2006).
The translation uses newPrompt to acquire a fresh prompt xp and then uses the delay macro
to simulate pushing a binder frame: the context delay t as x in is analogous to the binder
frame [(λx.) t]. The delay macro anticipates that its body returns a function fk that
expects the delayed argument, so it applies fk to a suspension of t. As we see below, the
function fk is a cont (κx.E).
In the context of delay, the simulation executes va xp . Since alpha conversion of λx.t can
be written (λxp .t[xp /x]), the term va xp is analogous to (λx.t) xp = t[xp /x]: it substitutes
a fresh prompt for a fresh variable.
The need macro, which defines the translation rule for variables, N [[x]], captures the
continuation delimited by x (which had better be a prompt!) and returns a function λfth . . . .
that closes over both x and the captured continuation k. This function is the cont κx.E,
with x modeling the bound variable of the same name, and continuation k modeling E.
The function expects the binder frame [(λx.) t], which is now at the top of the current
continuation, to pass it the suspension λ().N [[t]]. The simulation forces the suspension, and
the do macro pushes the resulting answer binders and binds va to the underlying value.
Pushing the answer binders begins the simulation of the D.1 rule.
The simulation of D.1 delays a computation that immediately returns the result va of
evaluating the term t, pushes the continuation k associated with the cont, and returns va
to the extended continuation. Now any subsequent evaluation of x immediately returns the
memoized value va instead of recomputing t. This yields an answer hka , va i where ka is an
empty subcontinuation. The value va is delayed exactly as before and is also returned from
the properly extended continuation. This part of the translation bears close resemblance
to the paradoxical Y combinator (Curry and Feys, 1958), suggesting that the simulation
requires recursive types (Shan, 2007).
7. Correctness of the Simulation
We prove correctness of the simulation relative to the machine semantics. Since we
already proved correctness of the machine semantics relative to standard-order reduction,
the result is a proof that our simulation provides a continuation semantics for call-by-need.
The previous discussion provides an informal justification for the structure of the callby-need simulation. To prove the correctness of the simulation, we appeal to the continuation semantics for delimited control (Dybvig et al., 2007). This semantics is completely
standard for the terms of the lambda calculus. Figure 4 presents the interesting parts of
the semantics. All CPS terms take a standard continuation κ, but the control combinators
also take a metacontinuation γ, which is a list of continuations and prompts, and a global
prompt counter q. The base continuation κ0 delimits each proper continuation and directs
evaluation up the metacontinuation, discarding any intervening prompts. Given a CPS
program t, the expression t κ0 [ ] 0 runs it.
To prove correctness, we compose N [[·]] with the delimited continuation semantics to
produce a translation Λ[[·]] to the λβη calculus augmented with arithmetic, lists, and the
30
R. GARCIA, A. LUMSDAINE, AND A. SABRY
newP romptc = λκ.λγ.λq.κ q γ (q + 1)
withSubContc = λp.λf.λκ.λγ.f (κ : γ↑p ) κ0 γ↓p
pushP romptc = λp.λt.λκ.λγ.t κ0 (p : κ : γ)
pushSubContc = λγ ′ .λt.λκ.λγ.t κ0 (γ ′++(κ : γ))
κ0 = λv.λγλq.K(v, γ, q)
K(v, [ ], q) = v
K(v, p : γ, q) = K(v, γ, q)
K(v, κ : γ, q) = κ v γ q
Figure 4: Delimited Control Combinators
operator K defined in Figure 4. We also give each abstract machine configuration a denotation, defined in terms of name-indexed denotations for its constituents D[[·]]X (see Figures 5
through 8).
D[[t]]X = Λ[[t]][ι(xi , X)/xi ]
ΛP [[t]] = Λ[[t]] κ0 (0 : [ ]) 1
Λ[[x]] =
withSubContc x
λkx .λk1 .k1
λfth .λk2 .
pushPromptc 0 (fth ())
(λ hka , va i.
pushSubContc ka
(λk3 .
pushPromptc x
(pushSubContc kx
(withSubContc 0 λka .λk.k hka , va i))
(λfk .fk (λ().withSubContc 0
λka .λk.k hka , va i)
k3 ))
k2 )
Λ[[t1 t2 ]] = λk1 .pushPromptc 0 Λ[[t1 ]]
(λ hka , va i.
pushSubContc ka
(λk2 .
newPromptc
λxp .pushPromptc xp (va xp )
(λfk .fk (λ().Λ[[t2 ]]) k2 ))
k1 )
Λ[[λx.t]] = withSubContc 0 λka .λk.k hka , λx.Λ[t℄i
Figure 5: Denotations for Terms
LAZY EVALUATION AND DELIMITED CONTROL
31
ι(xi , X) = ι(xi , [x1 , x2 , . . . , xi , . . . , xn ]) = i
|X| = |[x1 , x2 , . . . , xi , . . . , xn ]| = n
Λ[[hX | E, ?ic ]] = D[[?]]X κ0 (D[[E]]X ++(0 : [ ])) (|X| + 1) c ∈ { d, f, b, n }
Λ[[hX | TE, λx.tUi]] = hD[[E]]X , λx.D[[t]]X i
Figure 6: Denotations for Names and Configurations
Denotations of machine configurations are constructed from their components: the
configuration’s focus ?, context E, and list of names X. A machine configuration denotes
the translation of its focus applied to three arguments: the base continuation κ0 as its
starting continuation, the denotation of its context, bounded by the redex delimiter 0,
as the metacontinuation, and the size |X| of X plus 1 as its initial prompt. The redex
delimiter attached to the metacontinuation handles the case when an answer subsumes
the entire context by returning the answer as the result. The denotation of the terminal
machine configuration hX | TE, vUi is treated separately to show how it corresponds directly
to a final answer.
Our semantic translation takes advantage of X being a proper list of unique names. Free
active variables denote prompts in our translation, and since 0 is the redex delimiter, we
assign to each variable its 1-based index in X. We use |X|+1 as the global prompt counter to
ensure that no future prompts conflict with the current active variable denotations, thereby
guaranteeing hygiene (see Section 4.5).
Each evaluation context frame denotes a two-element metacontinuation consisting of a
prompt and a proper continuation. The prompt for a binder frame is the prompt translation
ι(x, X) of the bound variable x. The cont and operand frames have redex prompts 0. These
prompts guarantee that answer building operations will arrive at the innermost redex. Each
continuation function specializes a subexpression of the CPS translation for terms D[[·]]X
with the denotations of the context frame’s parts. Compare, for instance, the denotation
of an application, t1 t2 , to that of an operand frame, [ t2 ]. The application term pushes
the global prompt, and executes t1 in the context of a continuation that receives an answer
hka , va i. The denotation of the operand frame is a metacontinuation containing the same
prompt and continuation.
A redex denotes a CPS’ed term that closes over the denotations of its constituents and
implements the corresponding reduction step.
To facilitate our proof of correctness, we make a slight change to the machine semantics.
In the machine, composing an empty context with the current context is an identity operation. The continuation semantics do not share this property. During execution, an empty
continuation is denoted by the base continuation κ0 . If a continuation is captured or pushed
in the context of an empty continuation, then the empty continuation will be captured as
part of the metacontinuation or pushed onto the current metacontinuation before reinstating the pushed continuation. In short, the call-by-need machine semantics guarantees that
E ◦ [ ] = E, but the continuation semantics do not prove that κ0 : γ = γ. Dybvig et al.
discuss the notion of proper tail recursion for delimited continuations. Their operational
characterization of proper tail recursion corresponds to the latter equation.
32
R. GARCIA, A. LUMSDAINE, AND A. SABRY
D[[E ◦ [f ]]]X = D[[[f ]]]X ++D[[E]]X
D[[[ ]]]X = [ ]
D[[[#]]]X = κ0 : [ ]
D[[[ t2 ]]]X = 0 : k ′ : [ ]
where k ′ = λ hka , va i.
pushSubContc ka
(λk2 .
newPromptc
λxp .pushPromptc xp (va xp )
(λfk .fk (λ().D[[t2 ]]X ) k2 ))
κ0
D[[[(λx.) t2 ]]]X = ι(x, X) : k ′ : [ ]
where k ′ = λfk .fk (λ().D[[t2 ]]X ) κ0
D[[[(κx.E) ]]]X = 0 : k ′ : [ ]
where k ′ =
λ hka , va i.
pushSubContc ka
(λk3 .
pushPromptc ι(x, X)
(pushSubContc D[[E]]X
(withSubContc 0 λka .λk.k hka , va i))
(λfk .fk (λ().withSubContc 0 λka .λk.k hka , va i)
k3 ))
κ0
Figure 7: Denotations for Evaluation Contexts
To remove this mismatch, we add a ghost frame [#] to our definition of evaluation
contexts. The ghost frame denotes the metacontinuation κ0 : [ ]. We also extend the unplug
operation on evaluation contexts such that it discards ghost frames: C[[E ◦ [#]]] = C[[E]].
Finally, we alter the right hand side of transition rules that grab and push continuations
to pair ghost frames with composed evaluation contexts in a manner consistent with the
continuation semantics. For instance, the updated D.1 rule is as follows2:
hX | E1 , (κx.E2 ) TE3 , vUid 7−→nam hX | E1 ◦ [#] ◦ E3 ◦ [(λx.) v] ◦ [#] ◦ E2 , vib
(D.1)
These modifications do not alter the observable behavior of the machine while modeling the
property that pushing empty frames has meaning in the continuation semantics.
Given these denotations, it is straightforward to prove correctness of the simulation
relative to the abstract machine.
Theorem 7.1. If t is a closed term, then ΛP [[t]] = Λ[[h∅ | [ ], tif ]].
Proof. ΛP [[t]] = D[[t]]∅ κ0 (0 : [ ]) 1 = Λ[[h∅ | [ ], tif ]].
2The exact placement of ghost frames falls right out of the correctness proof.
LAZY EVALUATION AND DELIMITED CONTROL
33
D[[(κx1 .E1 ) TE2 , λx2 .tU]]X =
pushSubContc D[[E2 ]]X
(λk3 .
pushPromptc ι(x1 , X)
(pushSubContc D[[E1 ]]X
(withSubContc 0 λka .λk.k hka , λx2 .[t℄X i))
(λfk .fk (λ().withSubContc 0 λka .λk.k hka , λx2 .[t℄X i)
k3 ))
D[[TE, λx.t1 U t2 ]]X =
pushSubContc D[[E]]X
(λk2 .
newPromptc
λxp .pushPromptc xp ((λx.D[[t1 ]]X ) xp )
(λfk .fk (λ().D[[t2 ]]X ) k2 ))
Figure 8: Denotations for Redexes
Theorem 7.2. If C1 7−→nam C2 then Λ[[C1 ]] = Λ[[C2 ]].
Proof. By cases on 7−→nam . The proof utilizes beta, eta and K equivalences to establish
correspondences.
8. Simulating Extensions
Many straightforward language extensions also have straightforward simulations under
the above model.
Simulating let bindings essentially performs the same operations as immediately applying a lambda abstraction.
N [[let x = t1 in t2 ]] = let x = newPrompt
in delay N [[t1 ]] as x
in N [[t2 ]]
In contrast to the application rule, the variable x is directly assigned a fresh prompt, rather
than binding it to an auxiliary variable xp . The body of the let can be interpreted in place
and substitution of the prompt is implicit since x is already free in t2 .
The translation for basic constants is analogous to that for lambda abstractions: the
constant must be returned to the next redex.
N [[c]] = return c
For this translation, we assume that the call-by-value language provides the same constants
as the call-by-need language.
The translation for function constants is as follows:
N [[f t]] = do va ⇐ N [[t]]
in return (f va )
Interpreting a function constant application forces its argument and then acts on the value
that is ultimately produced. Since function expressions yield values, the result is immediately returned.
34
R. GARCIA, A. LUMSDAINE, AND A. SABRY
The translation for cons acquires two fresh prompts, uses them to delay the argument
to cons, and stores them as a pair.
N [[cons t1 t2 ]] = let x1 = newPrompt in
let x2 = newPrompt in
delay N [[t1 ]] as x1 in
delay N [[t2 ]] as x2 in
cons x1 x2
The translations for car and cdr evaluate their argument, retrieve a prompt from the
resulting pair, and demand its value.
N [[car t]] = do vp ⇐ N [[t]]
in need (car vp )
N [[cdr t]] = do vp ⇐ N [[t]]
in need (cdr vp )
9. Conclusions
In this paper, we expose and examine the operational structure of lazy evaluation as
embodied in call-by-need semantics. We present this understanding in two ways: as an
abstract machine whose operational behavior involves control capture, and as a simulation
of call-by-need under call-by-value plus delimited control operations. Delimited control can
be used to simulate a global heap, but our particular simulation uses delimited control
operations to manage laziness locally, just like the calculus reduction rules.
The artifacts of this investigation provide new tools for increasing our understanding
of lazy evaluation and its connections to control. The abstract machine could be used to
establish connections to heap-based implementations of call-by-need, and possibly modern
graph-reduction based formulations (Peyton Jones and Salkild, 1989). In fact it seems that
the calculus and abstract machine may point out new structural and dynamic invariants that
are inherent to call-by-need evaluation but are hidden in the unstructured representations
of heaps.
The abstract machine and simulation might also provide new opportunities for reasoning about the correctness of transformations applied to call-by-need programs. Surely the
calculus provides the same equational reasoning powers as the abstract machine. However
the machine may enable researchers to more easily conceive transformations and justifications that are not as easy to recognize in the reduction semantics. Our simulation might be
connected to that of Okasaki et al. (1994). The simulation might suggest new mechanisms
by which to embed call-by-need evaluation within call-by-value programs.
One significant difference between the two formulations of call-by-need lambda calculi (Maraist et al., 1998; Ariola and Felleisen, 1997) is the status of variables. Maraist
et al. consider variables to be values, whereas Ariola and Felleisen do not. Ultimately,
Maraist et al. (1998) prove standardization against a standard-order relation that does not
consider variables to be values. This paper sheds no light on the inclusion of variables
among the values, however it demonstrates in stark detail the consequences of the latter
design. In the abstract machine, the transition rules for lambda terms, namely the rebuild
rules, differ significantly from the transition rules for variables, the need rules. A similar
LAZY EVALUATION AND DELIMITED CONTROL
35
distinction can be seen simply by observing the complexity of their respective translations.
In short, our semantics interpret variables as memoized computations rather than values.
Variables can be treated as values under deterministic call-by-value and call-by-name reduction; it remains an open question whether the same could be achieved for call-by-need
and if so what its operational implications would be.
Our results reveal that a proliferation of semantic frameworks—reduction semantics,
machine semantics, etc—is a boon and not a crisis. The reduction semantics of call-by-need
elegantly and mysteriously encode a rich semantics whose broad implications can be seen in
equivalent machine semantics and continuation semantics. As such, our work provides new
perspectives from which to reason about call-by-need, delimited control, and their respective
expressive powers.
10. Acknowledgements
We thank Daniel P. Friedman, Roshan James, William Byrd, Michael Adams, and
the rest of the Indiana University Programming Languages Group, as well as Jeremy Siek,
Zena Ariola, Phil Wadler, Olivier Danvy, and anonymous referees for helpful discussions
and feedback on this work.
References
Abadi, M., Cardelli, L., Curien, P.-L., and Lévy, J.-J. 1991. Explicit substitutions.
Journal of Functional Programming 1, 4, 375–416.
Ariola, Z. and Klop, J. W. 1994. Cyclic lambda graph rewriting. In Logic in Computer
Science (LICS ’94). IEEE Computer Society Press, Los Alamitos, Ca., USA, 416–425.
Ariola, Z. M. and Felleisen, M. 1997. The call-by-need lambda calculus. Journal of
Functional Programming 7, 3 (May), 265–301.
Ariola, Z. M. and Klop, J. W. 1997. Lambda calculus with explicit recursion. Information and Computation 139, 2, 154–233.
Ariola, Z. M., Maraist, J., Odersky, M., Felleisen, M., and Wadler, P. 1995. A
call-by-need lambda calculus. In POPL ’95: Proceedings of the 22nd ACM SIGPLANSIGACT Symposium on Principles of Programming Languages. ACM Press, New York,
NY, USA, 233–246.
Balat, V., Di Cosmo, R., and Fiore, M. 2004. Extensional normalisation and typedirected partial evaluation for typed lambda calculus with sums. In POPL ’04: Proceedings of the 31st ACM SIGPLAN-SIGACT Symposium on Principles of Programming
Languages. ACM, New York, NY, USA, 64–76.
Barendregt, H. P. 1981. The Lambda Calculus, its Syntax and Semantics. NorthHolland, Amsterdam, NL. Studies in Logic and the Foundations of Mathematics.
Biernacka, M. and Danvy, O. 2007. A concrete framework for environment machines.
ACM Transactions on Computational Logic 9, 1, 6.
Biernacki, D., Danvy, O., and chieh Shan, C. 2005. On the dynamic extent of
delimited continuations. Information Processing Letters 96, 1, 7 – 17.
Curry, H. B. and Feys, R. 1958. Combinatory Logic, Volume I. Studies in Logic and
the Foundations of Mathematics. North-Holland, Amsterdam. Second printing 1968.
36
R. GARCIA, A. LUMSDAINE, AND A. SABRY
Danvy, O. and Filinski, A. 1990. Abstracting control. In LFP ’90: Proceedings of the
1990 ACM Conference on LISP and Functional Programming. ACM, New York, NY,
USA, 151–160.
Danvy, O., Millikin, K., Munk, J., and Zerny, I. 2010. Defunctionalized interpreters
for call-by-need evaluation. In FLOPS ’10: Proceedings of the Tenth International Symposium on Functional and Logic Programming. Springer-Verlag, London, UK. To appear.
Danvy, O. and Nielsen, L. R. 2004. Refocusing in reduction semantics. Tech. Rep. RS04-26, BRICS, DAIMI, Department of Computer Science, University of Aarhus, Aarhus,
Denmark. November.
Dybvig, R. K., Peyton Jones, S., and Sabry, A. 2007. A monadic framework for
delimited continuations. Journal of Functional Programming 17, 6, 687–730.
Felleisen, M. 1988. The theory and practice of first-class prompts. In POPL ’88: Proceedings of the 15th ACM SIGPLAN-SIGACT Symposium on Principles of Programming
Languages. ACM, New York, NY, USA, 180–190.
Felleisen, M., Findler, R., and Flatt, M. 2009. Semantics Engineering with PLT
Redex. MIT Press, Cambridge, MA.
Felleisen, M. and Friedman, D. P. 1986. Control operators, the SECD-machine, and
the λ-calculus. In Formal Description of Programming Concepts, M. Wirsing, Ed. NorthHolland, Netherlands, 193–217.
Felleisen, M. and Hieb, R. 1992. A revised report on the syntactic theories of sequential
control and state. Theoretical Computer Science 103, 2, 235–271.
Friedman, D. P., Ghuloum, A., Siek, J. G., and Winebarger, O. L. 2007. Improving
the lazy Krivine machine. Higher-Order and Symbolic Computation 20, 3, 271–293.
Friedman, D. P. and Wise, D. S. 1976. CONS should not evaluate its arguments. In
Automata, Languages and Programming, S. Michaelson and R. Milner, Eds. Edinburgh
University Press, Edinburgh, Scotland, 257–284.
Gibbons, J. and Wansbrough, K. 1996. Tracing lazy functional languages. In Proceedings of Conference on Computing: The Australian Theory Symposium, M. E. Houle and
P. Eades, Eds. Australian Computer Science Communications, Townsville, 11–20.
Henderson, P. and Morris, Jr., J. H. 1976. A lazy evaluator. In POPL ’76: Proceedings of the 3rd ACM SIGACT-SIGPLAN Symposium on Principles of Programming
Languages. ACM, New York, NY, USA, 95–103.
Kameyama, Y., Kiselyov, O., and Shan, C. 2008. Closing the stage: From staged code
to typed closures. In PEPM ’08: Proceedings of the 2008 ACM SIGPLAN Symposium on
Partial Evaluation and Semantics-based Program Manipulation. ACM, New York, NY,
USA, 147–157.
Kiselyov, O., Shan, C., Friedman, D. P., and Sabry, A. 2005. Backtracking, interleaving, and terminating monad transformers: (functional pearl). In ICFP ’05: Proceedings of the Tenth ACM SIGPLAN International Conference on Functional Programming.
ACM, New York, NY, USA, 192–203.
Kiselyov, O., Shan, C., and Sabry, A. 2006. Delimited dynamic binding. In ICFP
’06: Proceedings of the eleventh ACM SIGPLAN international conference on Functional
programming. ACM, New York, NY, USA, 26–37.
Krivine, J.-L. 2007. A call-by-name lambda-calculus machine. Higher-Order and Symbolic
Computation 20, 3, 199–207.
Launchbury, J. 1993. A natural semantics for lazy evaluation. In POPL ’93: Proceedings of the 20th ACM SIGPLAN-SIGACT Symposium on Principles of Programming
LAZY EVALUATION AND DELIMITED CONTROL
37
Languages. ACM, New York, NY, USA, 144–154.
Maraist, J., Odersky, M., and Wadler, P. 1998. The call-by-need lambda calculus.
Journal of Functional Programming 8, 3 (May), 275–317.
Moggi, E. and Sabry, A. 2004. An abstract monadic semantics for value recursion.
Theoretical Informatics and Applications 38, 4, 375–400.
Nakata, K. and Hasegawa, M. 2009. Small-step and big-step semantics for call-by-need.
Journal of Functional Programming 19, 06, 699–722.
Okasaki, C., Lee, P., and Tarditi, D. 1994. Call-by-need and continuation-passing
style. Lisp and Symbolic Computation 7, 1 (Jan.), 57–81.
Peyton Jones, S. L. and Salkild, J. 1989. The spineless tagless G-machine. In FPCA
’89: Proceedings of the Fourth International Conference on Functional Programming Languages and Computer Architecture. ACM, New York, NY, USA, 184–201.
Plotkin, G. D. 1975. Call-by-name, call-by-value and the λ-calculus. Theoretical Computer Science 1, 2 (Dec.), 125–159.
Sestoft, P. 1997. Deriving a lazy abstract machine. Journal of Functional Programming 7, 3, 231–264.
Shan, C. 2007. A static simulation of dynamic delimited control. Higher-Order and Symbolic Computation 20, 4, 371–401.
Sussman, G. J. and Steele Jr., G. L. 1975. Scheme: An interpreter for extended lambda
calculus. AI Memo 349, Artificial Intelligence Laboratory, Massachusetts Institute of
Technology, Cambridge, Massachusetts. Dec. Reprinted in Higher-Order and Symbolic
Computation 11(4):405–439, 1998, with a foreword (Sussman and Steele Jr., 1998).
Sussman, G. J. and Steele Jr., G. L. 1998. The first report on Scheme revisited.
Higher-Order and Symbolic Computation 11, 4, 399–404.
Wand, M. and Vaillancourt, D. 2004. Relating models of backtracking. In ICFP
’04: Proceedings of the Ninth ACM SIGPLAN International Conference on Functional
Programming. ACM, New York, NY, USA, 54–65.
Wang, C. 1990. Obtaining lazy evaluation with continuations in Scheme. Information
Processing Letters 35, 2, 93–97.
Xi, H. 1997. Evaluation under lambda abstraction. In PLILP ’97: Proceedings of the
Ninth International Symposium on Programming Languages: Implementations, Logics,
and Programs. Springer-Verlag, London, UK, 259–273.
This work is licensed under the Creative Commons Attribution-NoDerivs License. To view
a copy of this license, visit http://creativecommons.org/licenses/by-nd/2.0/ or send a
letter to Creative Commons, 171 Second St, Suite 300, San Francisco, CA 94105, USA, or
Eisenacher Strasse 2, 10777 Berlin, Germany
| 6cs.PL |
PARALLEL ALGORITHMS AND CONCENTRATION BOUNDS FOR THE
LOVÁSZ LOCAL LEMMA VIA WITNESS DAGS
arXiv:1509.06430v7 [cs.DM] 28 Sep 2017
BERNHARD HAEUPLER1 AND DAVID G. HARRIS2
Abstract. The Lovász Local Lemma (LLL) is a cornerstone principle in the probabilistic method
of combinatorics, and a seminal algorithm of Moser & Tardos (2010) provides an efficient randomized
algorithm to implement it. This can be parallelized to give an algorithm that uses polynomially
many processors and runs in O(log 3 n) time on an EREW PRAM, stemming from O(log n) adaptive
computations of a maximal independent set (MIS). Chung et al. (2014) developed faster local
and parallel algorithms, potentially running in time O(log 2 n), but these algorithms require more
stringent conditions than the LLL.
We give a new parallel algorithm that works under essentially the same conditions as the original
algorithm of Moser & Tardos but uses only a single MIS computation, thus running in O(log 2 n)
time on an EREW PRAM. This can be derandomized to give an NC algorithm running in time
O(log 2 n) as well, speeding up a previous NC LLL algorithm of Chandrasekaran et al. (2013).
We also provide improved and tighter bounds on the run-times of the sequential and parallel
resampling-based algorithms originally developed by Moser & Tardos. These apply to any problem
instance in which the tighter Shearer LLL criterion is satisfied.
1. Introduction
The Lovász Local Lemma (LLL), first introduced in [5], is a cornerstone principle in probability
theory. In its simplest symmetric form, it states that if one has a probability space Ω and a set of
m “bad” events B in that space, and each such event has probability PΩ (B) ≤ p; and each event
depends on at most d events (including itself), then under the criterion
(1)
epd ≤ 1
there is a positive probability that no bad events occurs. If equation (1) holds, we say the symmetric
LLL criterion is satisfied.
Although the LLL applies to general probability spaces, and the notion of dependency for a
general space can be complicated, most applications in combinatorics use a simpler setting in
which the probability space Ω is determined by a series of discrete variables X1 , . . . , Xn , each of
which is drawn independently with PΩ (Xi = j) = pij . Each bad event B ∈ B is a Boolean function
of a subset of variables SB ⊆ [n]. Then events B, B ′ are dependent (denoted B ∼ B ′ ) if they share
a common variable, i.e., SB ∩ SB ′ 6= ∅; note that B ∼ B. We say a set of bad events I ⊆ B is
independent if B 6∼ B ′ for all distinct pairs B, B ′ ∈ I. We say a variable assignment X avoids B if
every B ∈ B is false on X.
There is a more general form of the LLL, known as the asymmetric LLL, which can be stated
as follows. Suppose that there is a weighting function x : B → (0, 1) with the following property:
Y
(2)
∀B ∈ B
PΩ (B) ≤ x(B)
(1 − x(A))
A∼B
A6=B
1
School of Computer Science, Carnegie Mellon University. Research supported in part by NSF Awards CCF1527110 and CCF-1618280. Email: haeupler@cs.cmu.edu.
2
Department of Computer Science, University of Maryland, College Park, MD 20742. Research supported in part
by NSF Awards CNS-1010789 and CCF-1422569. Email: davidgharris29@gmail.com.
1
2
BERNHARD HAEUPLER AND DAVID G. HARRIS
then there is a positive probability of avoiding all bad events. The symmetric LLL is a special case
of this, derived by setting x(B) = ep. Both of these criteria are special cases of a yet more powerful
criterion, known as the Shearer criterion. This criterion requires a number of definitions to state;
we discuss this further in Section 1.3.
The probability of avoiding all bad events, while non-zero, is usually exponentially small; so the
LLL does not directly lead to efficient algorithms. Moser & Tardos [17] introduced a remarkable
randomized procedure, which we refer to as the Resampling Algorithm, which gives polynomial-time
algorithms for nearly all LLL applications:
Algorithm 1 The sequential Resampling Algorithm
1: Draw all variables X ∼ Ω.
2: while some bad events are true do
3:
Choose some true B ∈ B arbitrarily.
4:
Resample the variables in SB , independently from the distribution Ω.
This resampling algorithm terminates with probability one under the same condition as the
probabilistic LLL, viz. satisfying the Shearer criterion. The expected number of resamplings is
typically polynomial in the input parameters.
We note that this procedure can be useful even when the total number of bad events is exponentially large. At any stage of this algorithm, the expected number of bad events which are currently
true (and thus need to be processed), is still polynomial. If we have a subroutine which lists the
currently-true bad events in time poly(n), then the overall run-time of this algorithm can still be
polynomial in n. We refer to such a subroutine as a Bad-Event Checker. These are typically very
problem-specific; see [9] for more details.
1.1. Parallel algorithms for the LLL. Moser & Tardos also gave a simple RNC algorithm for
the LLL, shown below as Algorithm 2. Unlike their sequential algorithm, this requires a small slack
in the LLL criterion. In the symmetric setting, this criterion is
epd(1 + ǫ) ≤ 1
and in the asymmetric setting, it is given by
∀B ∈ B
(1 + ǫ)PΩ (B) ≤ x(B)
Y
A∼B
A6=B
(1 − x(A))
for some parameter ǫ ∈ (0, 1/2). We refer to these stronger criteria as ǫ-slack.
Algorithm 2 The Parallel Resampling Algorithm
1: Draw all variables X ∼ Ω.
2: while some bad events are true do
3:
Choose a maximal independent set I ofSbad events which are currently true.
4:
Resample, in parallel, all the variables B∈I SB from the distribution Ω.
P
x(B)
) rounds
Moser & Tardos showed that this algorithm terminates after O ǫ−1 log(n B∈B 1−x(B)
1
with high probability. In each round, there are two main computational tasks: one must execute
a parallel Bad-Event Checker and one must find a maximal independent set (MIS) among the bad
events which are currently true.
Both of these tasks can be implemented in parallel models of computation. The most natural
complexity parameter in these settings is the number of variables n, since the final output of the
1We say that an event occurs with high probability (abbreviated whp), if it occurs with probability ≥ 1 − n−Ω(1) .
PARALLEL ALGORITHMS AND CONCENTRATION BOUNDS FOR THE LLL VIA WITNESS DAGS
3
algorithm (i.e. a satisfying solution) will require at least n bits. This paper will focus on the
PRAM (Parallel Random Access Machine) model, in which we are allowed poly(n) processors and
polylog(n) time. There are a number of variants of the PRAM model, which differ in (among other
things) the ability and semantics of multiple processors writing simultaneously to the same memory
cell. Two important cases are the CRCW model, in which multiple cells can simultaneously write
(the same value) to a cell, and the EREW model, in which each memory cell can only be used by a
single processor at a time. Nearly all “housekeeping” operations (sorting and searching lists, etc.)
can also be implemented in O(log n) time using standard techniques in either model.
A Bad-Event Checker can typically be implemented in time O(log n). The step of finding an
MIS can potentially become a computational bottleneck. In [15], Luby introduced randomized
algorithms for computing the MIS of a graph G = (V, E) using poly(|V |) processors; in the CRCW
model of computation, this algorithm requires time O(log |V |) while in other models such as EREW
it requires time O(log2 |V |). Luby also discussed a deterministic algorithm using O(log2 |V |) time
(in either model).
Applying Luby’s MIS algorithm to the Resampling Algorithm yields an overall run-time of
P
P
x(B)
x(B)
)) (on EREW) or O(ǫ−1 log2 (n B∈B 1−x(B)
)) (on CRCW) and the overO(ǫ−1 log3 (n B∈B 1−x(B)
P
x(B)
all processor complexity is poly(n, B∈B 1−x(B) )).2
The computation of an MIS is relatively costly. In [4], Chung et al. gave several alternative
algorithms for the symmetric LLL which either avoid this step or reduce its cost. One algorithm,
based on bad events choosing random priorities and resampling a bad event if it has earlier priority
than its neighbors, runs in O(ǫ−1 log m) distributed rounds. This can be converted to a PRAM
algorithm using O(ǫ−1 log2 m) time (in EREW) and O(ǫ−1 log m) time (in CRCW). Unfortunately,
this algorithm requires a stronger criterion than the LLL: namely, in the symmetric setting, it
requires that epd2 ≤ (1 − ǫ). In many applications of the LLL, particularly those based on Chernoff
bounds for the sum of independent random variables, this stricter criterion leads to qualitatively
similar results as the symmetric LLL. In other cases, this loses much critical precision leading to
weaker results. In particular, their bound essentially corresponds to the state of the art [16] before
the break-through results of Moser and Moser & Tardos [16, 17].
Another parallel algorithm of Chung et al. requires only the standard symmetric LLL criterion
and runs in O(ǫ−1 (log2 d)(log m)) rounds, subsequently reduced to O(ǫ−1 (log d)(log m)) rounds by
[8]. When d is polynomial in m, these do not improve on the Moser-Tardos algorithm. More recent
distributed algorithms for the LLL such as [7] do not appear to lead to PRAM algorithms.
In [17], a deterministic parallel (NC) algorithm for the LLL was given, under the assumption
that d = O(1). This was strengthened in [3] to allow arbitrary d under a stronger LLL criterion
epd1+ǫ ≤ 1, with a complexity of O(ǫ−1 log3 (mn)) time and (mn)O(1/ǫ) processors (in either CRCW
or EREW). This can be extended to an asymmetric setting, but there are many more technical
conditions on the precise form of B.
1.2. Overview of our results. In Section 2, we introduce a new theoretical structure to analyze
the behavior of the Resampling Algorithm, which we refer to as the witness DAG. This provides
an explanation or history for some or all of the resamplings that occur. This generalizes the notion
of a witness tree, introduced by Moser & Tardos in [17], which only provides the history of a single
resampling. We use this tool to show stronger bounds on the Parallel Resampling Algorithm given
by Moser & Tardos:
Theorem 1.1. Suppose that the Shearer criterion is satisfied with ǫ-slack. Then whp the Parallel
Resampling Algorithm terminates after O(ǫ−1 log n) rounds.
2The weighting function x(B) plays a somewhat mysterious role in the LLL, and it can be confusing to have it
appear in the complexity bounds for Resampling Algorithm. In most (although not all) applications, the expression
P
x(B)
B∈B 1−x(B) ) can be bounded as poly(n).
4
BERNHARD HAEUPLER AND DAVID G. HARRIS
Suppose furthermore we have a Bad-Event Checker which uses polynomial processors and T
time. Then the total complexity of the Parallel Resampling Algorithm is ǫ−1 nO(1) processors, and
2
O( (log n)(Tǫ+log n) ) time (in EREW model) or O( (log n)(Tǫ +log n) ) time (in CRCW model).
These bounds are independent of the LLL weighting function x(B) and the number of bad events
m. These simplify similar bounds shown in Kolipaka & Szegedy [14], which show that Parallel
Resampling Algorithm terminates, with constant probability, after O(ǫ−1 log(n/ǫ)) rounds.3
In Sections 3 and 4, we develop a new parallel algorithm for the LLL. The basic idea of this
algorithm is to select a random resampling table and then precompute all possible resamplingpaths compatible with it. Surprisingly, this larger collection, which in a sense represents all possible
trajectories of the Resampling Algorithm, can still be computed relatively quickly (in approximately
O(ǫ−1 log2 n) time). Next, we find a single MIS of this larger collection, which determines the
complete set of resamplings. It is this reduction from ǫ−1 log n separate MIS algorithms to just one
that is the key to our improved run-time.
We will later analyze this parallel algorithm in terms of the Shearer criterion, but this requires
many preliminary definitions. We give a simpler statement of our new algorithm for the symmetric
LLL criterion:
Theorem 1.2. Suppose that we have a Bad-Event Checker using O(log mn) time and poly(m, n)
processors. Suppose that each bad event B has PΩ (B) ≤ p and is dependent with at most d bad
events and that epd(1 + ǫ) ≤ 1 for some ǫ > 0. Then, there is an EREW PRAM algorithm to find
a configuration avoiding B whp using Õ(ǫ−1 log(mn) log n) time and poly(m, n) processors.4
In Section 5, we derandomize this algorithm under a slightly more stringent LLL criterion. The
full statement of the result is somewhat complex, but a summary is that if epd1+ǫ < 1 then we
obtain a deterministic EREW algorithm using O(ǫ−1 log2 (mn)) time and (mn)O(1/ǫ) processors.
This is NC for constant ǫ > 0.
The following table summarizes previous and new parallel run-time bounds for the LLL. For
simplicity, we state the symmetric form of the LLL criterion, although many of these algorithms
are compatible with asymmetric LLL criteria as well. The run-time bounds are simplified for
readability, omitting terms which are negligible in typical applications.
Model
LLL criterion
Reference
Run-time
Previous results
Randomized CRCW PRAM epd(1 + ǫ) ≤ 1
[17]
ǫ−1 log2 m
Randomized EREW PRAM epd(1 + ǫ) ≤ 1
[17]
ǫ−1 log3 m
1+ǫ
Deterministic EREW PRAM
epd
≤1
[3]
ǫ−1 log3 m
2
Randomized CRCW PRAM
epd ≤ 1 − ǫ
[4]
ǫ−1 log m
Randomized EREW PRAM
epd2 ≤ 1 − ǫ
[4]
ǫ−1 log2 m
This paper
Randomized CRCW PRAM epd(1 + ǫ) ≤ 1 Resampling Algorithm ǫ−1 log2 n
Randomized EREW PRAM epd(1 + ǫ) ≤ 1 Resampling Algorithm ǫ−1 log3 n
Randomized EREW PRAM epd(1 + ǫ) ≤ 1
New algorithm
ǫ−1 log2 m
1+ǫ
Deterministic EREW PRAM
epd
≤1
New algorithm
ǫ−1 log2 m
Although the main focus of this paper is on parallel algorithms, our techniques also lead to a
new and stronger concentration result for the run-time of the sequential Resampling Algorithm.
The full statement appears in Section 6; we provide a summary here:
3Note that Kolipaka & Szegedy use m for the number of variables and n for the number of bad events, while we
do the opposite. In this paper, we have translated all of their results into our notation. The reader should be careful
to keep this in mind when reading their original paper.
4The Õ notation hides polylogarithmic factors, i.e. Õ(t) = t(log t)O(1) .
PARALLEL ALGORITHMS AND CONCENTRATION BOUNDS FOR THE LLL VIA WITNESS DAGS
5
Theorem 1.3. Suppose that the asymmetric LLL criterion is satisfied with ǫ-slack. Then whp the
P x(B)
2
Resampling Algorithm performs O ( B 1−x(B)
) + logǫ n resamplings. Alternatively, suppose that
the symmetric LLL criterion epd ≤ 1 is satisfied. Then whp the Resampling Algorithm performs
O(n + d log2 n) resamplings.
Similar concentration bounds have been shown in [14] and [1]. The main technical innovation
P
x(B)
B
1−x(B)
), whereas the new concentration
here is that prior concentration bounds have the form O(
ǫ
bounds are largely independent of ǫ (as long as it is not too small).
1.3. Stronger LLL criteria. The LLL criterion, in either its symmetric or asymmetric form, depends on only two parameters: the probabilities of the bad events, and their dependency structure.
The symmetric LLL criterion epd ≤ 1 is a very simple criterion involving these parameters, but it
is not the most powerful. In [18], Shearer gave the strongest possible criterion that can be stated in
terms of these parameters alone. This criterion is somewhat cumbersome to state and difficult to
work with technically, but it is useful theoretically because it subsumes many of the other simpler
criteria.
We note that the “lopsided” form of the LLL can be applied to this setting, in which bad events
are atomic configurations of the variables (as in a k-SAT instance), and this can be stronger than
the ordinary LLL. As shown in [10], there are forms of lopsidependency in the Moser-Tardos setting
which can even go beyond the Shearer criterion itself. However, the Parallel Resampling Algorithm
does not work in this setting; alternate, slower, parallel algorithms which can take advantage of
this lopsidependency phenomenon are given in [10], [12]. In this paper we are only concerned with
the standard (not lopsided) LLL.
To state the Shearer criterion, it will be useful to suppose that the dependency structure of our
bad events B is fixed, but the probabilities for the bad events have not been specified. We define
the independent-set polynomial Q(I, p) as
X
Y
Q(I, p) =
p(B)
(−1)|J|−|I|
I⊆J⊆B
J independent
B∈J
for any I ⊆ B. Note that Q(I, p) = 0 if I is not an independent set. This quantity plays a key role
in Shearer’s criterion for the LLL [18] and the behavior of the Resampling Algorithm. We say that
the probabilities p satisfy the Shearer criterion iff Q(∅, p) > 0 and Q(I, p) ≥ 0 for all independent
sets I ⊆ B.
Proposition 1.4 ([18]). Suppose that p satisfies the Shearer criterion. Then any probability space
with the given dependency structure and probabilities PΩ = p has a positive probability that none of
the bad events B are true.
Suppose that p do not satisfy the Shearer criterion. Then there is a probability space Ω with the
given dependency structure and probabilities PΩ = p for which, with probability one, at least one
B ∈ B is true.
Proposition 1.5 ([18]). Suppose that p(B) ≤ p′ (B) for all B ∈ B. Then, if p′ satisfies the Shearer
criterion, so does p.
One useful parameter for us will be the following:
Definition 1.6. For any bad event B, define the measure of B to be µ(B) =
Q({B},PΩ )
Q(∅,PΩ ) .
In [14], Kolipaka & Szegedy showed that if the Shearer criterion is satisfied, then the Resampling
Algorithm terminates with probability one; furthermore, the run-time of the Resampling Algorithms
can be bounded in terms of the measures µ.
Proposition 1.7 ([14]). The expected number of resamplings of any B ∈ B is at most µ(B).
6
BERNHARD HAEUPLER AND DAVID G. HARRIS
P
This leads us to define the work parameter for the LLL by W = B∈B µ(B). Roughly speaking,
the expected running time of the Resampling Algorithm is O(W ); we will later show (in Section 6)
that such a bound holds whp as well.
Although the sequential Resampling Algorithm can often work well when the Shearer criterion is
satisfied (almost) exactly, the Parallel Resampling Algorithm typically requires an additional small
slack.
Definition 1.8. We say that the Shearer criterion is satisfied with ǫ-slack, if the vector of probabilities (1 + ǫ)PΩ satisfies the Shearer criterion.
It it extremely difficult to directly show that the Shearer criterion is satisfied in a particular
instance. There are alternative criteria, which are weaker than the full Shearer criterion but much
easier to work with computationally. Perhaps the simplest is the asymmetric LLL criterion. The
connection between the Shearer criterion and the asymmetric LLL criterion was shown by Kolipaka
& Szegedy in [14].
Theorem 1.9 ([14]). Suppose that a weighting function x : B → (0, 1) satisfies
Y
∀B ∈ B
PΩ (B)(1 + ǫ) ≤ x(B)
(1 − x(A))
A∼B
A6=B
Then the Shearer criterion is satisfied with ǫ-slack, and µ(B) ≤
x(B)
1−x(B)
for all B ∈ B.
This was extended to the cluster-expansion LLL criterion of [2] by Harvey & Vondrák in [13]:
Theorem 1.10 ([13]). For any bad event B, let N (B) denote the set of bad events A with A ∼ B.
Suppose that a weighting function µ̃ : B → [0, ∞) satisfies
Y
X
µ̃(A)
∀B ∈ B
µ̃(B) ≥ PΩ (B)(1 + ǫ)
I⊆N (B) A∈I
I independent
Then the Shearer criterion is satisfied with ǫ-slack, and µ(B) ≤ µ̃(B) for all B ∈ B.
For the remainder of this paper, we will assume unless stated otherwise that our probability
space Ω satisfies the Shearer criterion with ǫ-slack. We will occasionally derive certain results for
the symmetric LLL criterion as a corollary of results on the full Shearer criterion.
2. The witness DAG and related structures
There are two key analytical tools introduced by Moser & Tardos to analyze their algorithm:
the resampling table and witness tree.
The resampling table R is a table of values R(i, t), where i ranges over the variables 1, . . . , n
and t ranges over the positive natural numbers. Each cell R(i, t) is drawn independently from the
distribution of variable Xi , that is, R(i, t) = j with probability pij , independently of all other cells.
The intent of this table is that, instead of choosing new values for the variables in “on-line” fashion,
we precompute the future values of all the variables. The first entry in the table R(i, 1), is the
initial value for the variable Xi ; on the tth resampling, we set Xi = R(i, t + 1).5
The witness tree is a structure which records the history of all variables involved in a given
resampling. Moser & Tardos give a very clear and detailed description of the process for forming
witness trees; we provide a simplified description here. Suppose that the Resampling Algorithm
resample bad events B1 , . . . , Bt in order (the algorithm has not necessarily terminated by this
point). We build a witness-tree τ̂t for the tth resampling, as follows. We place a node labeled by
5Although nominally the resampling table provides a countably infinite stream of values for each variable, in
practice we will only need to use approximately ǫ−1 log n distinct values for each variable.
PARALLEL ALGORITHMS AND CONCENTRATION BOUNDS FOR THE LLL VIA WITNESS DAGS
7
Bt at the root of the tree. We then go backwards in time for j = t − 1, . . . , 1. For each Bj , if there
is a node v ′ in the tree labeled by B ′ ∼ Bj , then we add a new node v labeled by Bj as a child
of v ′ ; if there are multiple choices of v ′ , we always select the one of greatest depth (breaking ties
arbitrarily.) If there is no such node v ′ , then we do not add any nodes to the tree for that value of
j.
2.1. The witness DAG. The witness tree τ̂t only provides an explanation for the single resampling
at time t; it may discard information about other resamplings. We now consider a related object,
the witness DAG (abbreviated WD ) that can record information about multiple resamplings, or
all of the resamplings.
A WD is a directed acyclic graph, whose nodes are labeled by bad events. For nodes v, v ′ ∈ G,
we write v ≺ v ′ if there is an edge from v to v ′ . We impose two additional requirements, which we
refer to as the comparability conditions. First, if nodes v, v ′ are labeled by B, B ′ and B ∼ B ′ , then
either v ≺ v ′ or v ′ ≺ v; second, if B 6∼ B ′ then there is no edge between v, v ′ .
We let |G| denote the number of vertices in a WD G.
It is possible that a WD can contain multiple nodes with the same label. However, because of the
comparability conditions, all such nodes are linearly ordered by ≺. Thus for any WD G and any
B ∈ B, the nodes of G labeled B can be unambiguously sorted. Accordingly, we use the notation
(B, k) for the kth node of G labeled by B. For any node v, we refer to this ordered pair (B, k)
as the extended label of v. Every node in a WD receives a distinct extended label. We emphasize
that this is a notational convenience, as an extended label of a node can be recovered from the WD
along with its original labels.
Given a full execution of the Resampling Algorithm, there is a particularly important WD which
we refer to as the Full Witness DAG Ĝ (abbreviated FWD ). We construct this as follows. Suppose
that we resample bad events B1 , . . . , Bt . Then Ĝ has vertices v1 , . . . , vt which are labeled B1 , . . . , Bt .
We place an edge from vi to vj iff i < j and Bi ∼ Bj . We emphasize that Ĝ is a random variable.
The FWD (under different terminology) was analyzed by Kolipaka & Szegedy in [14], and we will
use their results in numerous places. However, we will also consider partial WDs, which record
information about only a subset of the resamplings.
As witness trees and single-sink WDs are closely related, we will often use the notation τ for a
single-sink WD. We let Γ denote the set of all single-sink WDs, and for any B ∈ B we let Γ(B)
denote the set of single-sink WDs whose sink node is labeled B.
2.2. Compatibility conditions for witness DAGs and resampling tables. The MoserTardos proof hinged upon a method for converting an execution log into a witness tree, and
necessary conditions were given for a witness tree being produced in this fashion in terms of its
consistency with the resampling table. We will instead use these conditions as a definition of
compatibility.
Definition 2.1 (Path of a variable). Let G be a WD. For any i ∈ [n], let G[i] denote the subgraph
of G induced on all vertices v labeled by B with i ∈ SB . Because of the comparability conditions,
G[i] is linearly ordered by ≺; thus we refer to G[i] as the path of variable i.
Definition 2.2 (Configuration of v). Let G be a WD and R a resampling table. Let v ∈ G be
labeled by B. For each i ∈ SB , let yv,i denote the number of vertices w ∈ G[i] such that w ≺ v.
We now define the configuration of v by
v
XG
(i) = R(i, 1 + yv,i )
Definition 2.3 (Compatibility of WD G with resampling table R). For a WD G and a resampling
table R, we say that G is compatible with R if, for all nodes v ∈ G labeled by B ∈ B, it is the case
v . This is well-defined because X v assigns values to all the
that B is true on the configuration XG
G
variables in SB .
8
BERNHARD HAEUPLER AND DAVID G. HARRIS
We define ΓR to be the set of single-sink WDs compatible with R, and similarly for ΓR (B).
The following are key results used by Moser & Tardos to bound the running time of their
resampling algorithm:
Definition 2.4 (Weight of a WD). Let G be any WD,
Q whose nodes are labeled by bad events
B1 , . . . , Bs . We define the weight of G to be w(G) = sk=1 PΩ (Bk ).
Proposition 2.5. For a random resampling table R, any WD G has probability w(G) of being
compatible with R.
v follows the law of Ω, and so the probability that B is
Proof. For any node v ∈ G, note that XG
v
true of the configuration XG is PΩ (B). Next, note that each node v ∈ G imposes conditions on
disjoint sets of entries of R, and so these events are independent.
Proposition 2.6. Suppose we run the Resampling Algorithm, taking values for the variables from
the resampling table R. Then Ĝ is compatible with R.
Proof. Suppose there is a node v ∈ Ĝ with extended label (B, k). Thus, B must be resampled at
least k times. Suppose that the kth resampling occurs at time t. Let Y be the configuration at time
v (i) for all i ∈ S . For, the graph Ĝ must
t, just before this resampling. We claim that Y (i) = XĜ
B
contain all the resamplings involving variable i. All such nodes would be connected to vertex v (as
they overlap in variable i), and those that occur before time t are precisely those with an edge to v.
So yv,i is exactly the number of bad events up to time t that involve variable i. Thus, just before
v (i),
the resampling at time t, variable i was on its 1 + yv,i resampling. So Y (i) = R(i, 1 + yv,i ) = XĜ
as claimed.
In order for B to be resampled at time t, it must have been the case that B was true, i.e., that
v on S , necessarily B holds on configuration
B held on configuration Y . Since Y agrees with XĜ
B
v
XĜ as well. This is true for all v ∈ G and so G is compatible with R.
2.3. Prefixes of a WD. A WD records information about many resamplings. If we are only
interested in the history of a subset of its nodes, then we can form a prefix subgraph which discards
irrelevant information.
Definition 2.7 (Prefix graph). For any WD G and vertices v1 , . . . , vℓ ∈ G, let G(v1 , . . . , vℓ ) denote
the subgraph of G induced on all vertices which have a path to at least one of v1 , . . . , vℓ .
If H is a subgraph of G with H = G(v1 , . . . , vℓ ) for some v1 , . . . , vℓ ∈ G, then we say that H is
a prefix of G.
Using Definition 2.7, we can give a more compact definition of the configuration of a node:
v (i) = R(i, |G(v)[i]|).
Proposition 2.8. For any WD G and v ∈ G, we have XG
Proof. Suppose that v is labeled by B. The graph G(v)[i] contains precisely v itself and the other
nodes w ∈ G[i] with w ≺ v. So |G(v)[i]| = yv,i + 1.
Proposition 2.9. Suppose G is compatible with R and H is a prefix of G. Then H is compatible
with R.
Proof. Let H = G(v1 , . . . , vℓ ). Consider w ∈ H labeled by B. We claim that H(w) = G(w). For,
consider any u ∈ H(w). So u has a path to w in H; it also must have a path to w in G. On the
other hand, suppose u ∈ G(w), so u has a path p to w in G. As w has a path to one of v1 , . . . , vℓ ,
this implies that every vertex in the path p also has such a path. Thus, the path p is in H, and
hence u has a path in H to w, so u ∈ H(w).
Next, observe that for any i ∈ SB we have
w
w
XG
(i) = R(i, |G(w)[i]|) = R(i, |H(w)[i]|) = XH
(i)
w.
and by hypothesis, B is true on XG
PARALLEL ALGORITHMS AND CONCENTRATION BOUNDS FOR THE LLL VIA WITNESS DAGS
9
2.4. Counting witness trees and WDs. In this section, we bound the summed weights of
certain classes of WDs. In light of Proposition 2.5, this will upper-bound the expected number of
resamplings.
Proposition 2.10 ([14]). For any B ∈ B, we have
X
w(τ ) ≤ µ(B).
τ ∈Γ(B)
Proof. For any WD G with a single sink node v labeled B, we define Ij′ for non-negative integers
′
j using the following recursion. I0′ = {v}, and Ij+1
is the set of vertices in G whose out-neighbors
′
′
all lie in I0 ∪ · · · ∪ Ij . Let Ij denote the labels of the vertices in Ij′ ; so I0 = {B}.
Now observe that by the comparability conditions each set Ij is an independent set, and for each
B ′ ∈ Ij+1 there is some B ′′ ∼ B ′ , B ′′ ∈ Ij . Also, the mapping from G to I0 , . . . , Ij is injective. We
thus may sum over all such I1 , . . . , I∞ to obtain an upper bound on the weight of such WDs. In
[14] Theorem 14, this sum is shown to be Q({B}, PΩ )/Q(∅, PΩ ) (although their notation is slightly
different.)
We will now take advantage of the ǫ-slack in our probabilities.
Proposition 2.11. Given any V ⊆ B, we say that V is a dependency-clique if B ∼ B ′ for all
B, B ′ ∈ V . If V is a dependency-clique then for any ρ ∈ [0, ǫ) we have
X Q({B}, (1 + ρ)PΩ )
1+ρ
≤
.
Q(∅, (1 + ρ)PΩ )
ǫ−ρ
B∈V
Proof. Consider the probability vector p defined by
(
(1 + ǫ)PΩ (B) if B ∈ V
p(B) =
(1 + ρ)PΩ (B) if B ∈
/V
Since V is a clique, any independent set I has |I ∩ V | ≤ 1. Thus we calculate Q(∅, p) as
Y
Y
X
X
ǫ−ρ
)PΩ (A)
(1 + [A ∈ V ]
p(A) =
Q(∅, p) =
(−1)|I|
(−1)|I|
1+ρ
A∈I
I⊆B
I independent
=
X
(−1)|I| (1 + [I ∩ V 6= ∅]
X
(−1)|I|
I⊆B
I independent
=
I⊆B
I independent
= Q(∅, (1 + ρ)PΩ ) −
Y
A∈I
I⊆B
I independent
ǫ−ρ Y
)
(1 + ρ)PΩ (A)
1+ρ
(1 + ρ)PΩ (A) +
A∈I
ǫ−ρ X
1+ρ
B∈V
A∈I
X
I⊆B
I independent
I∩V ={B}
(−1)|I|
Y
(1 + ρ)PΩ (A)
A∈I
(ǫ − ρ) X
Q({B}, (1 + ρ)PΩ )
1+ρ
B∈V
We use here the Iverson notation so that [I ∩ V 6= ∅] is one if I ∩ V 6= ∅ and zero otherwise.
Note that p ≤ (1 + ǫ)PΩ and so by Propositions 1.4 and 1.5 we have Q(∅, p) > 0 and Q({B}, (1 +
ρ)PΩ ) ≥ 0 for all B ∈ V . Thus,
P
X Q({B}, (1 + ρ)PΩ )
B∈V Q({B}, (1 + ρ)PΩ )
=
P
Q(∅, (1 + ρ)PΩ )
Q(∅, p) + (ǫ−ρ)
B∈V
B∈V Q({B}, (1 + ρ)PΩ )
1+ρ
P
Q({B}, (1 + ρ)PΩ )
1+ρ
=
≤ (ǫ−ρ) B∈V
P
ǫ−ρ
Q({B}, (1 + ρ)PΩ )
1+ρ
B∈V
10
BERNHARD HAEUPLER AND DAVID G. HARRIS
Definition 2.12 (Adjusted weight). For any WD G, we define the adjusted weight with respect
to rate factor ρ by
aρ (G) = w(G)(1 + ρ)|G| .
Observe that w(G) = a0 (G).
Corollary 2.13. Suppose that V ⊆ B is a dependency-clique. Then for any ρ ∈ [0, ǫ) we have
X
1+ρ
aρ (B) ≤
.
ǫ−ρ
B∈V
Proof. Applying Proposition 2.10 using the probability vector p = (1 + ρ)PΩ gives
X X
X Q({B}, (1 + ρ)PΩ )
aρ (B) ≤
Q(∅, (1 + ρ)PΩ )
B∈V τ ∈Γ(B)
B∈V
Now apply Proposition 2.11.
Corollary 2.14. We have the bound W ≤ n/ǫ, where we recall the definition W =
Proof. We write
W =
X
B∈B
µ(B) =
P
B∈B
µ(B).
X Q({B}, PΩ )
X X Q({B}, PΩ )
≤
Q(∅, PΩ )
Q(∅, PΩ )
i∈[n] B:SB ∋i
B∈B
Now note that for any i ∈ [n], the set of bad events B with i ∈ SB forms a dependency-clique.
P
1
Ω)
Thus, applying Proposition 2.11 with ρ = 0 gives i∈SB Q({B},P
Q(∅,PΩ ) ≤ ǫ .
Corollary 2.15 ([14]). The total weight of all single-sink WDs is at most n/ǫ.
Proof. Follows immediately from Proposition 2.10 and Corollary 2.14.
Proposition 2.16. For r ≥ 1 + 1/ǫ, the expected number of single-sink WDs compatible with R
containing more than r nodes is at most enr(1 + ǫ)−r
Proof. For any ρ ∈ [0, ǫ), sum over such WDs to obtain:
X
X
X
X
aρ (τ )
w(τ )(1 + ρ)|τ | = (1 + ρ)−r
w(τ ) ≤ (1 + ρ)−r
P (τ compatible with R) =
τ ∈Γ
|τ |≥r
≤ (1 + ρ)−r
τ ∈Γ
|τ |≥r
τ ∈Γ
|τ |≥r
τ ∈Γ
|τ |≥r
X X
X
aρ (τ )
i∈[n] B:SB ∋i τ ∈Γ(B)
≤ (1 + ρ)−r n
1+ρ
ǫ−ρ
by Corollary 2.13
Now take ρ = ǫ−(1+ǫ)/r. By our condition r ≥ 1+1/ǫ we have ρ ∈ [0, ǫ) and so Proposition 2.11
nr r
applies. Hence the expected number of such WDs is thus at most (r−1)r−1
≤ enr(1 + ǫ)−r .
(1+ǫ)r
) nodes. Whp all but
Corollary 2.17. Whp, every element of ΓR contains O( log(n/ǫ)
ǫ
10 log n
R
of Γ contain at most
nodes.
ǫ
Proof. This follows immediately from Markov’s inequality and Proposition 2.16.
Corollary 2.18. Whp, all WDs compatible with R have height O( logǫ n ).
10 log n
ǫ
elements
PARALLEL ALGORITHMS AND CONCENTRATION BOUNDS FOR THE LLL VIA WITNESS DAGS
11
Proof. Suppose that there is a WD G of height T compatible with R. Then for i = 1, . . . , T there
is a single-sink WD of height i compatible with R (take the graph G(v), where v is a node of height
i.) This implies that there Ω(T ) members of ΓR of height Ω(T ). By Proposition 2.17, this implies
T = O( logǫ n ).
Corollary 2.18 leads to a better bound on the complexity of the Parallel Resampling Algorithm.
The following Proposition 2.19 is remarkable in that the complexity is phrased solely in terms of
the number of variables n and the slack ǫ, and is otherwise independent of B.
Proposition 2.19. Suppose that the Shearer criterion is satisfied with ǫ-slack. Then whp the
Parallel Resampling Algorithm terminates after O( logǫ n ) rounds.
Suppose we have a Bad-Event Checker in time T and polynomial processors. Then the Parallel Resampling Algorithm can be executed using nO(1) /ǫ processors with an expected run-time of
2
O( (log n)(Tǫ+log n) ) (in the EREW model) or O( (log n)(Tǫ +log n) ) (in the CRCW model).
Proof. An induction on i shows that if the Parallel Resampling Algorithm runs for i steps, then Ĝ
has depth i, and it is compatible with R. By Corollary 2.18 whp this implies that i = O( logǫ n ).
n
This implies that the total time needed to identify true bad events is O(iT ) ≤ O( T log
ǫ ).
We next compute the time required for MIS calculations. We only show the calculation for the
EREW model, as the CRCW bound is nearly identical. Suppose that at stage i the number of
bad events which
P are currently true is vi . Then the total time spent calculating MIS, over the full
algorithm, is ti=1 O(log2 vi ).
P
Since log x is a concave-down function of x, this sum is at most O(t log2 ( vi /t)). On the other
hand, for each bad event which is at true at eachPstage, one
P can construct a distinct corresponding
single-sink WD compatible with R. Hence, E[ vi ] ≤ τ ∈Γ w(τ ) ≤ n/ǫ. As t ≤ logǫ n , we have
P
E[t log2 ( vi /t)] ≤ ǫ−1 log3 n. This shows the bound on the time complexity of the algorithm.
The expected number of bad events which are ever true is at most the weight of all single-sink
WDs, which is at most W ≤ n/ǫ. By Markov’s inequality, whp the total number of bad events
which are ever true is bounded by nO(1) /ǫ. The processor bound follows from this observation.
3. Mutual consistency of witness DAGs
In Section 2, we have seen conditions for WDs to be compatible with a given resampling table
R. In this section, we examine when a set of WDs can be mutually consistent, in the sense that
they could all be prefixes of some (unspecified) FWD.
Definition 3.1 (Consistency of G, G′ ). Let G, G′ be WDs. We say that G is consistent with G′
is, for all variables i, either G[i] is an initial segment of G′ [i] or G′ [i] is an initial segment of G[i],
both of these as labeled graphs. (Carefully note the position of the quantifiers: If n = 2 and G[1] is
an initial segment of G′ [1] and G′ [2] is an initial segment of G[2], then G, G′ are consistent.)
Let G be any set of WDs. We say that G is pairwise consistent if G, G′ are consistent with each
other for all G, G′ ∈ G.
Proposition 3.2. Suppose H1 , H2 are prefixes of G. Then H1 is consistent with H2 .
Proof. Observe that for any w1 ≺ w2 ∈ Hj , we must have w1 ∈ Hj as well. It follows that Hj [i] is
an initial segment of G[i] for any i ∈ [n]. As both H1 [i] and H2 [i] are initial segments of G[i], one
of them must be an initial segment of the other.
Definition 3.3 (Merge of two consistent WDs). Let G, G′ be consistent WDs. Then we define the
merge G ∨ G′ as follows. If either G or G′ has a node v with an extended label (B, k), then we
create a corresponding node w ∈ G ∨ G′ labeled by B. We refer to the corresponding label of w as
(B, k).
12
BERNHARD HAEUPLER AND DAVID G. HARRIS
Now, let v1 , v2 ∈ G ∨ G′ have corresponding label (B1 , k1 ) and (B2 , k2 ). We create an edge
from v1 to v2 if either G or G′ has an edge between vertices with extended label (B1 , k1 ), (B2 , k2 )
respectively.
Note that for every vertex v ∈ G with extended label (B, k), there is a vertex in G ∨ G′ with
corresponding label (B, k); we will abuse notation slightly and use v to refer this vertex in G ∨ G′
as well.
Proposition 3.4. Let G, G′ be consistent WDs and let H = G ∨ G′ . If there is a path v1 , . . . , vℓ in
H and vℓ ∈ G, then also v1 , . . . , vℓ ∈ G.
Proof. Suppose that this path has corresponding labels (B1 , k1 ), . . . , (Bℓ , kℓ ). Let i ≤ ℓ be minimal
such that vi , . . . , vℓ are all in G. (This is well-defined as vℓ ∈ G). If i = 1 we are done.
Otherwise, we have vi ∈ G, vi−1 ∈ G′ − G. Note that Bi−1 ∼ Bi , so let j ∈ SBi−1 ∩ SBi . Note
that vi ∈ G[j], vi−1 ∈ G′ [j]. But observe that in H there is an edge from vi−1 to vi . As vi−1 ∈
/ G,
this edge must have been present in G′ . So G′ [j] contains the vertices vi−1 , vi , in that order, while
G[j] contains only the vertex vi . Thus, neither G[j] or G′ [j] can be an initial segment of the other.
This contradicts the hypothesis.
Proposition 3.5. Let G, G′ be consistent WDs and let H = G ∨ G′ . Then H is a WD and both G
and G′ are prefixes of it.
Proof. Suppose that H contains a cycle v1 , . . . , vℓ , v1 , and suppose v1 ∈ G. Then by Proposition 3.4
the cycle v1 , . . . , vℓ , v1 is present also in G, which is a contradiction.
Next, we show that the comparability conditions hold for H. Suppose that (B1 , k1 ) and (B2 , k2 )
are the corresponding labels of vertices in H, and B1 ∼ B2 . So let i ∈ SB1 ∩ SB2 . Without loss of
generality, suppose that G[i] is an initial segment of G′ [i]. So (B1 , k1 ) and (B2 , k2 ) appear in G′ [i].
Because of the comparability conditions for G′ , there is an edge in G′ on these vertices, and hence
there is an edge in H as well. On the other hand, if there is an edge in H between vertices (B1 , k1 )
and (B2 , k2 ), there must be such an edge in G or G′ as well; by the comparability conditions this
implies B1 ∼ B2 .
Finally, we claim that G = H(v1 , . . . , vℓ ) where v1 , . . . , vℓ are the vertices of G. It is clear that
G ⊆ H(v1 , . . . , vℓ ). Now, suppose w ∈ H(v1 , . . . , vℓ ). Then there is a path w, x1 , x2 , . . . , xk , v where
x1 , . . . , xk ∈ H and v ∈ G. By Proposition 3.4, this implies that w ∈ G.
Proposition 3.6. If v ∈ G ∨ G′ has corresponding label (B, k), then v also has extended label
(B, k).
Proof. Because of our rule for forming edges in G ∨ G′ , the only edges that can go to v from other
nodes labeled B, would have corresponding labels (B, ℓ) for ℓ < k. Thus, there are at most k − 1
nodes labeled B with an edge to v.
On the other hand, there must be nodes with extended label (B, k) in G or G′ ; say without loss
of generality the first. Then G must also have nodes with extended labels (B, 1), . . . , (B, k − 1).
These correspond to vertices w1 , . . . , wk−1 with corresponding labels (B, 1), . . . , (B, k − 1), all of
which have an edge to v. So there are at least k − 1 nodes labeled B with an edge to v.
Thus, G has exactly k − 1 nodes labeled B with an edge to v and hence v has extended label
(B, k).
Proposition 3.7. The operation ∨ is commutative and associative.
Proof. Commutativity is obvious from the symmetric way in which ∨ was defined. To show associativity, we can give the following symmetric characterization of H = (G1 ∨ G2 ) ∨ G3 : If G1 , G2
or G3 has a node labeled (B1 , k1 ) then so does H; there is an edge in H from (B1 , k1 ) to (B2 , k2 )
if there is such an edge in G1 , G2 or G3 .
Proposition 3.8. If G1 , G2 , G3 are pairwise consistent WDs, then G1 ∨ G2 is consistent with G3 .
PARALLEL ALGORITHMS AND CONCENTRATION BOUNDS FOR THE LLL VIA WITNESS DAGS
13
Proof. For any variable i ∈ [n], note that either G1 [i] is an initial segment of G2 [i] or vice-versa.
Also note that (G1 ∨ G2 )[i] is the longer of G1 [i] or G2 [i].
Now we claim that for any variable i, either G3 [i] is an initial segment of (G1 ∨G2 )[i] or vice-versa.
Suppose without loss of generality that G1 [i] is an initial segment of G2 [i]. Then (G1 ∨G2 )[i] = G1 [i].
By definition of consistency, either G1 [i] is an initial segment of G3 [i] or vice-versa. So (G1 ∨ G2 )[i]
is an initial segment of G3 [i] or vice-versa.
In light of Propositions 3.7 and 3.8, we can unambiguously define, for any pairwise consistent
set of WDs G = {G1 , . . . , Gℓ }, the merge
_
G = G1 ∨ G2 ∨ G3 · · · ∨ Gℓ
We can give another characterization of pairwise consistency, which is more illuminating although
less explicit:
Proposition 3.9. The WDs G1 , . . . , Gℓ are pairwise consistent iff there is some WD H such that
G1 , . . . , Gℓ are all prefixes of H.
Proof. For the forward direction: let H = G1 ∨ · · · ∨ Gl . By Proposition 3.5, each Gi is a prefix
of H. For the backward direction: by Proposition 2.10, any Gi1 , Gi2 are both prefixes of H, hence
consistent.
Proposition 3.10. Let G1 , G2 be consistent WDs and R a resampling table. Then G1 ∨ G2 is
compatible with R iff both G1 and G2 are compatible with R.
Proof. For the forward direction: let v ∈ G1 labeled by B. By Proposition 3.4, we have G1 (v) =
v =
(G1 ∨ G2 )(v). Thus for i ∈ SB we have |G1 (v)[i]| = |(G1 ∨ G2 )(v)[i]|. This implies that XG
1
v
v
v . As this is true for all v ∈ G , it
XG
. By hypothesis, B is true on XG
and hence XG
1
1 ∨G2
1 ∨G2
1
follows that G1 is compatible with R. Similarly, G2 is compatible with R.
For the backward direction: Let v ∈ G1 ∨ G2 . Suppose without loss of generality that v ∈ G1 .
v = Xv
As in the forward direction, we have XG
G1 ∨G2 ; by hypothesis B is true on the former so it is
1
true on the latter. Since this holds for all v ∈ G1 ∨ G2 , it follows that G1 ∨ G2 is compatible with
R.
4. A new parallel algorithm for the LLL
In this section, we will develop a parallel algorithm to enumerate the entire set ΓR . This will
allow us to enumerate (implicitly) all WDs compatible with R. In particular, we are able to simulate
all possible values for the FWD Ĝ, without running the Resampling Algorithm.
In a sense, both the Parallel Resampling Algorithm and our new parallel algorithm are building
up Ĝ. However, the Parallel Resampling Algorithm does this layer by layer, in an inherently
sequential way: it does not determine layer i + 1 until it has fixed a value for layer i, and resolving
each layer requires a separate MIS calculation.
Our new algorithm breaks this sequential bottleneck by exploring, in parallel, all possible values
for the computed MIS. Although this might seem like an exponential blowup, in fact we are able
to reduce this to a polynomial size by taking advantage of two phenomena: first, we can represent
the FWD in terms of single-sink WDs; second, a random resampling table drastically prunes the
space of compatible WDs.
4.1. Collectible witness DAGs. We will enumerate the set ΓR by building up its members nodeby-node. In order to do so, we must keep track of a slightly more general type of WDs, namely,
those derived by removing the root node from a single-sink WD. Such WDs have multiple sink
nodes, which are all at distance two in the dependency graph. The set of such WDs is larger than
ΓR , but still polynomially bounded.
14
BERNHARD HAEUPLER AND DAVID G. HARRIS
Definition 4.1 (Collectible WD). Suppose we are given a WD G, whose sink nodes are labeled
B1 , . . . , Bs . We say that G is collectible to B if B ∼ B1 , . . . , B ∼ Bs .
We say that G is collectible if it is collectible to some B ∈ B. Note that if G ∈ Γ(B) then G is
collectible to B. We use CWD as an abbreviation for collectible witness DAG.6
Proposition 4.2. Define
W′ =
X
B∈B
X
1
w(τ )
PΩ (B)
τ ∈Γ(B)
The expected total number of CWDs compatible with R is at most W ′ .
Proof. Suppose that G is a WD collectible to B. Then define G′ by adding to G a new sink node
labeled by B. As all the sink nodes in G are labeled by B ′ ∼ B, this G′ is a single-sink WD. Also,
w(G′ )
P (G compatible with R) = w(G) =
PΩ (B)
The total probability that there is some G compatible with R and collectible to B, is at most
the sum over all such G. Each WD G′ ∈ Γ(B) appears at most once in the sum and so
X X w(τ )
X
X
X
w(G′ )
≤
= W′
w(G) ≤
w(G) ≤
PΩ (B)
PΩ (B)
G collectible
B∈B
G collectible to B
B∈B
G collectible to B
B∈B τ ∈Γ(B)
Corollary 4.3. We have m ≤ W ′ ≤
P
µ(B)
B∈B PΩ (B) .
Proof. The upper bound
followsPfrom Proposition 2.10. For the lower bound, consider the contriP
bution to the sum B∈B PΩ1(B) τ ∈Γ(B) from the WDs τ with |τ | = 1.
The parameter W ′ , which dictates the run-time of our parallel algorithm, has a somewhat complicated behavior. For most applications of the LLL where the bad events are “balanced,” we have
W ′ ≈ m. Here are two examples of this.
Proposition 4.4. If the symmetric LLL criterion epd ≤ 1 is satisfied then W ′ ≤ me.
Proof. The asymmetric LLL criterion is satisfied by setting x(B) =
Theorem 1.9, we have µ(B) ≤ ePΩ (B) for all B ∈ B.
ePΩ (B)
1+ePΩ (B)
for all B ∈ B. By
Proposition 4.5. Let p : B → [0, 1] be a vector satisfying the two conditions:
(1) PΩ (B) ≤ p(B) for all B ∈ B
(2) p satisfies the Shearer criterion with ǫ-slack.
Then we have
(n/ǫ)
W′ ≤
minB∈B p(B)
Proof. For any WD G whose nodes are labeled B1 , . . . , Bs , define w′ (G) to be p(B1 ) · · · p(Bs ).
Now consider some τ ∈ Γ(B) which has s additional nodes labeled B1 , . . . , Bs . We have that
s
s
Y
Y
w′ (G)
w(G)
p(Bi ) =
PΩ (Bi ) ≤
=
.
PΩ (B)
p(B)
i=1
Thus, by Corollary 2.15,
X
W′ ≤
B∈B
i=1
P
X
(n/ǫ)
1
τ ∈Γ w(τ )
′
w (τ ) ≤
≤
.
p(B)
minB∈B p(B)
minB∈B p(B)
τ ∈Γ(B)
6This definition is close to the concept of partial witness trees introduced in [3].
PARALLEL ALGORITHMS AND CONCENTRATION BOUNDS FOR THE LLL VIA WITNESS DAGS
15
We note that Proposition 4.5 seems to say that W ′ becomes large when the probabilities p are
small. This seems strange, as bad events with small probability have a negligible effect. (For
instance, if a bad event has probability zero, we can simply ignore it.) A more accurate way to
read Proposition 4.5 is that W ′ becomes large only for problem instances which are close to the
Shearer bound and have small probabilities. Such instances would have some bad events with
simultaneously very low probability and very high dependency; in these cases then indeed W ′ can
become exponentially large.
4.2. Algorithmically enumerating witness DAGs. In the Moser-Tardos setting, the witness
trees were not actually part of the algorithm but were a theoretical device for analyzing it. Our new
algorithm is based on explicit enumeration of the CWDs. We will show that, for an appropriate
choice of parameter K, Algorithm 3 can be used to to enumerate all ΓR .
Algorithm 3 Enumerating witness DAGs
1: Randomly sample the resampling table R.
2: For each B ∈ B true in the initial configuration R(·, 1), create a graph with a single vertex
labeled B. We denote this initial set by F1 .
3: for k = 1, . . . , K do
4:
For each consistent pair G1 , G2 ∈ Fk , form G′ = G1 ∨G2 . If G′ is collectible and |G′ | ≤ k +1,
then add it to Fk+1 .
5:
For each G ∈ Fk which is collectible to B, create a new WD G′ by adding to G a new sink
node labeled B. If G′ is compatible with R then add it to Fk+1 .
Proposition 4.6. Let G ∈ Fk for any integer k ≥ 1. Then G is compatible with R.
Proof. We show this by induction on k. When k = 1, then G ∈ F1 is a singleton node v labeled by
v (i) = R(i, 1) for all i ∈ S , and so B is true on X v . So G is compatible with R.
B. Note that XG
B
G
Now for the induction step. First suppose G was formed by G = G1 ∨ G2 , for G1 , G2 ∈ Fk−1 . By
induction hypothesis, G1 , G2 are compatible with R. So by Proposition 3.8, G is compatible with
R. Second suppose G was formed in step (5), so by definition it must be compatible with R.
Proposition 4.7. If G is a CWD with at most k nodes and G is compatible with R, then G ∈ Fk .
Proof. We show this by induction on k. When k = 1, then G is a singleton node v labeled by B,
v (i) = R(i, 1). So B is true on the configuration R(·, 1), and so G ∈ F .
and XG
1
For the induction step, first note that if |G| < k, then by induction hypothesis G ∈ Fk−1 . So G
will be added to Fk in step (4) by taking G1 = G2 = G.
Next, suppose G has a single sink node v labeled by B. Consider the WD G′ = G − v. Note
that |G′ | = k − 1. Also, all the sink nodes in G′ must be labeled by some B ′ ∼ B (as otherwise
they would remain sink nodes in G). So G′ is collectible to B. By induction hypothesis, G′ ∈ Fk−1 .
Iteration k − 1 transforms the graph G′ ∈ Fk−1 into G (by adding a new sink node labeled by B),
and so G ∈ Fk as desired.
Finally, suppose that G has sink nodes v1 , . . . , vs labeled by B1 , . . . , Bs , where s ≥ 2 and B ∼
B1 , . . . , Bs for some B ∈ B. Let G′ = G(v1 ) and let G′′ = G(v2 , . . . , vs ). Note that G′ is missing the
vertex vs and similarly G′′ is missing the vertex v1 . So |G′ | < k, |G′′ | < k, and G′ , G′′ are collectible
to B. Thus, G′ , G′′ ∈ Fk−1 and so G = G′ ∨ G′′ is added to Fk in step (4).
Proposition 4.8. Suppose that we have a Bad-Event Checker using O(log mn) time and poly(m,n)
processors. Then there is an EREW PRAM algorithm to enumerate ΓR whp using poly(W ′ , ǫ−1 , n)
processors and Õ(ǫ−1 (log n)(log(W ′ n))) time.
16
BERNHARD HAEUPLER AND DAVID G. HARRIS
Proof. Proposition 4.7 shows that Fk contains all the CWDs compatible with R using at most
k nodes. Furthermore, by Corollary 2.17, whp every member of ΓR has at most O(ǫ−1 log(n/ǫ))
nodes. Hence, for K = Θ(ǫ−1 log(n/ǫ)), we have FK ⊇ ΓR whp.
The expected total number of CWD compatible with R is at most W ′ ; hence whp the total
number of such WDs is at most W ′ nO(1) . The Bad-Event Checker can be used to check whether
WDs are compatible with R, and routine parallel algorithms can be used to merge WDs. Thus, each
iteration of this algorithm can be implemented using (W ′ mn/ǫ)O(1) processors and O(log(W ′ mn/ǫ))
time. We obtain the stated bounds by using the fact that W ′ ≥ m.
4.3. Producing the final configuration. Now that we have generated the complete set ΓR , we
are ready to finish the algorithm by producing a satisfying assignment.
′
Define a graph G, whose nodes correspond to ΓR , with an edge between
W WDs G, G if they
are inconsistent. Let I be a maximal independent set of G, and let G = I. Finally define the
configuration X ∗ , which we refer to as the final configuration, by
X ∗ (i) = R(i, |G[i]| + 1)
for all i ∈ [n].
Proposition 4.9. The configuration X ∗ avoids B.
Proof. Suppose that B is true on X ∗ . Define the WD H by adding to G a new sink node v labeled
by B. Observe that G is a prefix of H. By Proposition 3.2 the WDs H, G are consistent.
We claim that H is compatible with R. By Proposition 2.9, G is compatible with R so this is
clear for all the vertices of H except for its sink node v. For this vertex, observe that for each
v (i) = R(i, |H[i]|) = R(i, |G[i]| + 1) = X ∗ (i). By Proposition 2.9, this implies
i ∈ SB we have XH
that H(v) is compatible with R as well.
So H(v) ∈ ΓR , and consequently H(v) is a node of G. Observe that H(v) and all the WDs
′
G ∈ I are prefixes of H. By Proposition 3.9, H(v) is consistent with all of them. As I was chosen
to be a maximal independent set, this implies that H(v) ∈ I.
By Proposition 3.5, this implies that H(v) is a prefix of G. This implies that |G[i]| ≥ |H(v)[i]|
for any variable i. But for i ∈ SB we have |H(v)[i]| = |H[i]| = |G[i]| + 1, a contradiction.
We thus obtain our faster parallel algorithm for the LLL:
Theorem 4.10. Suppose that the Shearer criterion is satisfied with ǫ-slack and that we have a BadEvent Checker using O(log mn) time and poly(m, n) processors. Then there is an EREW PRAM
algorithm to find a configuration avoiding B using Õ(ǫ−1 (log n) log(W ′ n)) time and (W ′ n/ǫ)O(1)
processors whp.
Proof. Use Proposition 4.8 to enumerate ΓR using Õ(ǫ−1 (log n)(log(W ′ n)) time and (W ′ n/ǫ)O(1)
processors. Whp, |ΓR | ≤ W nO(1) . Using Luby’s MIS algorithm, find a maximal independent set
of such WDs in time O(log2 (W n)) and (W n)O(1) processors. Finally, form the configuration X ∗
as indicated in Proposition 4.9 using O(log(W n/ǫ)) time and (W n/ǫ)O(1) processors. Using the
bound W ≤ n/ǫ gives the stated result.
Corollary 4.11. Suppose that the symmetric LLL criterion is satisfied with ǫ-slack, i.e., epd(1 +
ǫ) ≤ 1, and have a Bad-Event Checker using O(log mn) time and poly(m, n) processors. Then there
is an EREW PRAM algorithm to find a configuration avoiding B using Õ(ǫ−1 log(mn) log n) time
and (mn)O(1) processors whp.
P
Proof. By Proposition 4.4, we have W ′ ≤ me. Also, W = B∈B ep ≤ O(m/d). Since m ≤ nd, we
have W ≤ O(n). Now apply Theorem 4.10.
PARALLEL ALGORITHMS AND CONCENTRATION BOUNDS FOR THE LLL VIA WITNESS DAGS
17
4.4. A heuristic lower bound. In this section, we give some intuition as to why we believe
that the run-time of this algorithm, approximately O(ǫ−1 log2 n), is essentially optimal for LLL
algorithms which are based on the resampling paradigm. We are not able to give a formal proof,
because we do not have any fixed model of computation in mind. Also it is not clear whether our
new algorithm is based on resampling.
Suppose we are given a problem instance on n variables
whose distributions are all Bernoulli-q
√
n
bad
events, each of which is a threshold
for some parameter
q
∈
[0,
1].
The
space
B
consists
of
√
function on n variables, and all these events are completely disjoint from each other. By adjusting
the threshold and the parameter q, we can ensure that each bad event has probability p = 1 − ǫ.
The number of resamplings of each event is a geometric random variable, and it is not hard to see
that with high probability there will be some bad event B which requires Ω(ǫ−1 log n) resamplings
before it is false. Also, whenever we perform a resampling
of B, we must compute whether B is
√
currently true. This requires computing a sum of n binary variables, which itself requires time
Ω(log n). Thus, the overall running time of this algorithm must be Ω(ǫ−1 log2 n).
The reason we consider this a heuristic lower bound is that, technically, the parallel algorithm
we have given is not based on resampling. That is, there is no current “state” of the variables
which is updated as bad events are discovered. Rather, all possible resamplings are precomputed
in advance from the table R.
5. A deterministic parallel algorithm
In this section, we derandomize the algorithm of Section 4. The resulting deterministic algorithm
requires an additional slack compared to the Parallel Resampling Algorithm (which in turn requires
additional slack compared to the sequential algorithm). For the symmetric LLL setting, we require
that epd1+ǫ ≤ 1 for some (constant) ǫ > 0; this is the same criterion used by [3].
For deterministic algorithms, it is quite difficult to handle general classes of bad-events or asymmetric LLL criteria (whereas the randomized algorithms allow the bad-events to be almost arbitrary,
and converge under almost the same conditions as the Shearer criterion.) In [3] and [11], these issues
are discussed in more detail; our algorithm would also be compatible with most such extensions.
However in order to focus on the main case we restrict ourselves to the simplest scenario; we only
consider the symmetric criterion, and we assume that the set B contains m bad events, which are
explicitly represented as atomic events, that is, conjunctions of terms of the form Xi = j.
The paradigmatic example of this setting is the k-SAT problem. In this simplified setting, the
algorithm of [3] requires O(ǫ−1 log3 (mn)) time and (mn)O(1/ǫ) processors on the EREW PRAM. Our
main contribution in this section will be to obtain a faster algorithm, using just O(ǫ−1 log2 (mn))
time.7
The proof strategy behind our derandomization is similar to that of [3]: we first show a range
lemma (Lemma 5.4), allowing us to ignore WDs which have a large number of nodes. This will
show that if the resampling table R is drawn from a probability distribution which satisfies an
approximate independence condition, then the algorithm of Section 4 will have similar behavior
to the scenario in which R is drawn with full independence. Such probability distributions have
polynomial support size, so we can obtain an NC algorithm by searching them in parallel.
Definition 5.1. We say a probability space Ω′ is k-wise, δ-approximately independent, if for all
subsets of variables Xi1 , . . . , Xik , and all possible valuations j1 , . . . , jk , we have
PΩ′ (Xi1 = j1 ∧ · · · ∧ Xik = jk ) − PΩ (Xi1 = j1 ∧ · · · ∧ Xik = jk ) ≤ δ
7While randomized MIS algorithms appear to be faster in the CRCW model as compared to EREW, this does
not appear to be true for deterministic algorithms. The fastest known NC algorithms for MIS appear to require
O(log 2 |V |) time (in both models). Thus, for the deterministic algorithms, we do not distinguish between EREW and
CRCW PRAM models.
18
BERNHARD HAEUPLER AND DAVID G. HARRIS
Theorem 5.2 ([6]). There are k-wise, δ-approximately independent probability spaces which have
a support size poly(log n, 2k , 1/δ).
Our algorithm will be defined in terms of a key parameter K, which will determine later.
Lemma 5.3. Suppose that a resampling table R has the following properties:
(A1) For all τ ∈ ΓR we have |τ | ≤ K.
(A2) There are at most S CWDs compatible with R of size at most K.
Then, if we run the algorithm of Section 4 up to K steps, it will terminate with a configuration
avoiding B, using O(Klog(mnS) + log2 S) time and poly(K, S, m, n) processors.
Proof. By Proposition 4.7, Algorithm 3 enumerates all ΓR . By Proposition 4.9, the final configuration avoids B. The running time can be bounded noting that the total number of CWDs produced
at any step in the overall process is bounded by S.
Lemma 5.4. If there is a CWD G compatible with R with |G| ≥ K, then there exists a CWD G′
compatible with R of size K ≤ |G′ | ≤ 2K.
Proof. We prove this by induction on |G|. If |G| ≤ 2K then this holds immediately so suppose
|G| > 2K.
Suppose G has s > 1 sink nodes v1 , . . . , vs and is collectible to B. For each i = 1, . . . , s define
Hi = G(v1 , v2 , . . . , vi−1 , vi+1 , . . . , vs ).
Clearly G = H1 ∨ H2 ∨ · · · ∨ Hs . Hence there must exist i such that |G| > |Hi | ≥ |G|(1 − 1/s) ≥
|G|/2 ≥ K. Also, Hi is collectible to B and is a prefix of G, hence is compatible with R. So apply
the induction hypothesis to Hi .
Finally, suppose that G has a single sink node v labeled B. Then G − v is collectible to B and is
compatible with R. Since |G| > 2K we have |G − v| ≥ K and so we apply the induction hypothesis
to G − v.
Proposition 5.5. There are at most me(ed)t CWDs with t nodes.
Proof. Let Z denote the number of CWDs with t nodes. Then
X
X
X
(ed)−|G|
(ed)−|G| ≤ (ed)t
1 = (ed)t
Z=
CWD G
|G|=t
CWD G
|G|=t
CWD G
P
P
The term CWD G (ed)−|G| can be interpreted as CWD G w(G) under the probability vector
1
p(B) = ed
. This probability vector satisfies the symmetric LLL criterion, and so Proposition 4.4
P
gives CWD G (ed)−|G| ≤ me.
′
Proposition 5.6. For c, c′ sufficiently large constants, d ≥ 2, and K = c log(m/ǫ)
ǫ log d , the following
holds.
Suppose that B consists of atomic events on s variables and epd1+ǫ ≤ 1 for ǫ ∈ (0, 1). Suppose
′
that R is drawn from a probability distribution which is (m/ǫ)−c/ǫ -approximately, 2c sǫlog(m/ǫ)
-wise
log d
independent. Then with positive probability the following events both occur:
(B1) Every τ ∈ ΓR has |τ | ≤ 2K.
(B2) There are at most O(m) CWDs compatible with R.
Proof. By Lemma 5.4, a necessary condition for (B1) to fail is for some CWD G of size K ≤ |G| ≤
2K to be compatible with R. For any such G, let E be the event that G is compatible with R.
The event E is a conjunction of events corresponding to the vertices in G. Each such vertex event
′
depends on at most s variables, so in total E is an atomic event on at most |G|s ≤ 2Ks ≤ 2c sǫlog(m/ǫ)
log d
−c/ǫ
terms. So, by Definition 5.1, P (E) ≤ w(G) + (m/ǫ)
.
PARALLEL ALGORITHMS AND CONCENTRATION BOUNDS FOR THE LLL VIA WITNESS DAGS
Summing over all such G gives
X
P (G compatible with R) ≤
CWD G
K≤|G|≤2K
X
(w(G) + (m/ǫ)−c/ǫ ) ≤
CWD G
K≤|G|≤2K
X
w(G) +
X
19
(m/ǫ)−c/ǫ
CWD G
|G|≤2K
CWD G
|G|≥K
For any integer t, Proposition 5.5 shows there are at most me(ed)t total CWDs with t nodes,
andPhence their total weight is at most me(edp)t ≤ md−tǫ . So the first summand is at most
me t≥K d−tǫ ≤ O(md−Kǫ /ǫ); this is at most 0.1 for c′ sufficiently large.
Now let us fix c′ , and show that we can take c sufficiently
large. By Proposition 5.5, the total
P
t ≤ 2m(ed)2K . Hence the second
number of CWDs of size at most 2K is at most 2K
me(ed)
t=1
c′ log(m/ǫ)
summand is at most 2me(m/ǫ)−c/ǫ (ed) ǫ log d ; this is at most 0.1 for c sufficiently large.
Thus, altogether, we see that there is a probability of at most 0.2 that there exists a CWD (and
in particular a single-sink WD) compatible with R with more than 2K nodes.
Next suppose that all CWDs compatible with R have size at most 2K. Thus, the expected total
number of CWDs compatible with R is given by
X
CWD G
|G|≤2K
P (G compatible with R) ≤
X
CWD G
|G|≤2K
2K
X
m(ed)t )(m/ǫ)−c/ǫ
(w(G) + (m/ǫ)−c/ǫ ) ≤ W ′ + (
t=1
As we have already seen, the second summand is at most 0.1 for our choice of c, c′ . By Proposition 4.4, we have W ′ ≤ me. Thus, this sum is O(m). Finally apply Markov’s inequality.
Theorem 5.7. Suppose epd1+ǫ ≤ 1 for ǫ ∈ (0, 1) and every bad event is an atomic configuration on
at most s variables. Then there is a deterministic EREW PRAM algorithm to find a configuration
avoiding B, using O( s log(mn)/ logǫ d+log
2
(mn)
) time and (mn)O(
s+log d
)
ǫ log d
processors.
Proof. We begin with a few simple pre-processing steps. If there is any variable Xi which can take
on more than m values, then there must be one such value which appears in no bad events; simply
set Xi to that value. So we assume every variable can take at most m values, and so there are at
most mn possible assignments to the variables.
If ǫ < 1/(mn), then one can exhaustively test the full set of assignments using O(mn ) ≤ (mn)1/ǫ
processors. So we assume ǫ > 1/(mn). Similarly, if d = 1, then every bad event is independent,
and we can find a satisfying assignment exhaustively using O(ms ) processors. So we assume d ≥ 2.
′
-wise indeWe first construct a probability space Ω which is (mn)−c/ǫ -approximately, 2c sǫlog(mn)
log d
pendent on a ground set of size nK. We use this to form a resampling table R(i, x) for i ∈ [n]
s+log d
and x ≤ K. Theorem 5.2 ensures that Ω has a support size of (mn)O( ǫ log d ) . We subdivide our
processors so that ω ∈ Ω is explored by an independent group of processors; the total cost of this
log(mn)
). Henceforth, every element of this
allocation/subdivision step is O(log |Ω|) ≤ O( (s+logǫ d)
log d
probability space will be searched independently, with no further inter-communication (except at
the final stage of reporting a satisfying solution.)
′
Next, given a resampling table R, simulate the algorithm of Section 4 up to K = c log(mn)
ǫ log d .
′
By Lemma 5.6, for large enough constants c, c there is at least one element ω ∈ Ω for which all
single-sink WDs have size O(K) and for which the total number of CWDs compatible with R is
O(m). By Lemma 5.3, the algorithm of Section 4, applied to ω, produces a satisfying assignment
2
) time and using poly(m, n) processors.
using O(K log(mn) + log2 (mn)) = O( log (mn)
ǫ
Corollary 5.8. Suppose that there is a k-SAT instance with m clauses in which each variable
k/(1+ǫ)
appears in at most L ≤ 2 ek
clauses. Then a satisfying assignment can be deterministically
found on a EREW PRAM using O( log
2
(mn)
)
ǫ
time and (mn)O(1/ǫ) processors.
20
BERNHARD HAEUPLER AND DAVID G. HARRIS
Proof. Each bad event (a clause being false) is an atomic configuration on s = k variables. Here
p = 2−k and d = Lk = 2k/(1+ǫ) /e. Note that epd1+ǫ = e−ǫ < 1, and so the criterion of Theorem 5.7 is
2
satisfied. Since log d = Θ(s), the total processor count is (mn)O(1/ǫ) and the run-time is O( log (mn)
).
ǫ
Theorem 5.7 is essentially a “black-box” simulation of the randomized algorithm using an appropriate probability space. A more recent algorithm of [11] gives a “white-box” simulation, which
searches the probability space more efficiently; this significantly relaxes the constraint on the types
of bad events allowed. For problems where the bad events are simple (e.g. k-SAT), the algorithms
have essentially identical run-times.
6. Concentration for the number of resamplings
The expected number of resamplings for the Resampling Algorithm is at most W . Suppose we
wish to ensure that the number of resamplings is bounded with high probability, not merely in
expectation. One simple way to achieve this would be to run log n instances of the Resampling
Algorithm in parallel; this is a generic amplification technique which ensures that whp the total
number of resamplings performed will be O(W log n).
Can we avoid this extraneous factor of log n? In this section, we answer this question in the
affirmative by giving a concentration result for the number of resamplings. We show that whp the
number of resamplings will not exceed O(W ) (assuming that W is sufficiently large).
We note that a straightforward approach to show concentration would be the following: the
probability that there are T resamplings is at most the probability that there is a T -node WD
compatible with R; this can be upper-bounded by summing the weights of all such T -node WDs.
This proof strategy leads to the weaker result that number of resamplings is bounded by O(W/ǫ).
This multiplicative dependence on ǫ appears in prior concentration bounds. For instance, Kolipaka & Szegedy [14] shows that the Resampling Algorithm performs O(n2 /ǫ + n/ǫ log(1/ǫ)) resamplings with constant probability, and Achlioptas & Iliopoulos [1] shows that in the symmetric LLL
setting the Resampling Algorithm performs O(n/ǫ) resamplings whp. Such results have a large gap
when ǫ is small. Our main contribution is to remove this factor of ǫ−1 .
Proposition 6.1. Given any distinct bad
Q events B1 , . . . , Bs , the total weight of all WDs with s
sink nodes labeled B1 , . . . , Bs , is at most si=1 µ(Bi ).
Proof. We define a function F which which maps s-tuples (τ1 , . . . , τs ) ∈ Γ(B1 ) × · · · × Γ(Bs ) to
WDs G = F (τ1 , . . . , τs ) whose sink nodes are labeled B1 , . . . , Bs . The function is defined by first
forming the disjoint union of the graphs τ1 , . . . , τs . We then add an edge from a node B ∈ τi to
B ′ ∈ τj iff i < j and B ∼ B ′ .
Now, consider any WD G whose sink nodes v1 , . . . , vs are labeled B1 , . . . , Bs . For i = 1, . . . , j,
define τi recursively by
τi = G(vi ) − τ1 − · · · − τi−1
Each τi contains the sink node vi , so it is non-empty. Also, all the nodes in τi are connected to vi ,
so τi indeed has a single sink node. Finally, every node of G has a path to one of v1 , . . . , vj , so it
must in exactly one τi .
Thus, for each WD G with sink nodes labeled B1 , . . . , Bs , there exist τ1 , . . . , τs in Γ(B1 ), . . . , Γ(Bs )
respectively such that G = F (τ1 , . . . , τs ); furthermore the nodes of G are the union of the nodes of
τ1 , . . . , τs . In particular, w(G) = w(τ1 ) · · · w(τs ). Proposition 2.10 then gives
X
G has sink nodes
B1 ,...,Bs
w(G) ≤
X
τ1 ,...,τs
w(τ1 ) . . . w(τs ) =
s
Y
X
i=1 τ ∈Γ(Bi )
w(τ ) ≤
s
Y
µ(Bi )
i=1
PARALLEL ALGORITHMS AND CONCENTRATION BOUNDS FOR THE LLL VIA WITNESS DAGS
Theorem 6.2. Whp, the Resampling Algorithms performs O(W +
log2 n
ǫ )
21
resamplings.
Proof. Define Us to be the set of WDs which have s sink nodes and which are compatible with the
resampling table R; here s is a parameter to be specified later. Note that each of the s sink nodes
of a WD must receive a distinct label.
First, let us consider the expected size of Us . We use Proposition 6.1 to compute
X
X
E[|Us |] =
P (G compatible with R) ≤
w(G)
G has s sink nodes
≤
X
B1 ,...,Bs
distinct
G has s sink nodes
Ws
1 X
µ(B))s =
µ(B1 ) . . . µ(Bs ) ≤ (
s!
s!
B∈B
Now, suppose that the Resampling Algorithm runs for t time-steps, and consider the event E
2
that t ≥ c(W + logǫ n ) where c is some sufficiently large constant (to be determined). Let Ĝ be
the FWD of the resulting execution. Each resampling at time i ∈ {1, . . . , t} corresponds to some
vertex vi in Ĝ, and we define Hi = G(vi ).
Also, define A to be the set of indices i1 , . . . , is satisfying the properties that t ≥ i1 > i2 >
i3 > · · · > is−1 > is ≥ 1 and that ij ∈
/ Hi1 ∪ · · · ∪ Hij−1 for j = 1, . . . , s. For each tuple
a = (i1 , . . . , is ) ∈ A, define Ĝ(a) as Ĝ(vi1 , . . . , vis ). The condition ij ∈
/ Hi1 ∪ · · · ∪ Hij−1 ensures that
each vij is a sink node in Ĝ(a), and thus Ĝ(a) contains exactly s sink nodes. By Proposition 2.9 it
is compatible with R.
Further, we claim that Ĝ(a) 6= Ĝ(a′ ) for a 6= a′ . To see this, we note that we can recover
i1 , . . . , is uniquely from the graph Ĝ(a). For, suppose that B is the label of a sink node u of Ĝ(a);
in this case, if Ĝ(a)[u] contains k nodes labeled B, then vij must be the unique node in Ĝ with
extended label (B, k). This allows us to recover the unordered set {i1 , . . . , is }; the condition that
i1 > · · · > is allows us to recover a uniquely.
n
members of ΓR have size greater than
Define E ′ to be the (rare) event that more than 10 log
ǫ
10 log n
′
−Ω(1)
. By Proposition 2.17, P (E ) ≤ n
. Let X denote the set {i | |Ĝ(vi )| ≤ h} where
ǫ
n
10 log n
′
. Conditioned on the event E , we have |X| ≥ t − 10 log
; for c sufficiently large this
h =
ǫ
ǫ
implies |X| ≥ t/2.
Each Ĝ(a) is a distinct element of Us . Conditioning on the event E ′ , we therefore have:
X
X X
X
|Us | ≥ |A| ≥ |A ∩ X × · · · × X| =
···
1
i3 <i2
i1 ∈X i2 <i1
/ i1 ∪Hi2
i2 ∈H
/ i1 i3 ∈H
i3 ∈X
i2 ∈X
is <is−1
is ∈H
/ i1 ∪Hi2 ∪···∪His−1
is ∈X
By Proposition 6.5 (which we defer to after this proof), this expression is at least |X|−(s−1)h
≥
s
(t/2−sh)s
since |Hj | ≤ h for all j ∈ X.
s!
s
. Since E[|Us |] ≤
Hence, we have shown that E requires either event E ′ or that |Us | ≥ (t/2−sh)
s!
W s /s!, Markov’s inequality gives P (E) ≤ W s /(t/2 − sh)s + P (E ′ ).
t
Setting s = 4h
, we bound this as
Ws
= (4W/t)s ≤ 2−s = n−Ω(1)
(t/2 − sh)s
2
using the facts that t ≥ 8W and t ≥ Ω( logǫ n ).
Corollary 6.3. The Resampling Algorithm performs O(n/ǫ) resamplings whp.
Proof. By Corollary 2.14 we have W ≤ n/ǫ. Thus, by Theorem 6.2, the Resampling Algorithm
2
performs O( nǫ + logǫ n ) = O( nǫ ) resamplings whp.
22
BERNHARD HAEUPLER AND DAVID G. HARRIS
For the symmetric LLL, we can even obtain concentration without ǫ-slack.
Corollary 6.4. If the symmetric LLL criterion epd ≤ 1 is satisfied, then whp the number of
resamplings is O(n + d log2 n).
Proof. Set x(B) = d1 for all B ∈ B. Now a simple calculation shows that this satisfies the asymmetric
d−1 − 1 = Ω(1/d). Thus µ(B) ≤ x(B)/(1− x(B)) = 1/d, and so
LLL condition with slack ǫ = e( d−1
d )
W ≤ m/d. We also may observe that m ≤ nd. So, by Theorem 6.2, the total number of resamplings
is O(n + d log2 n) whp.
To finish this proof, we need to show the following simple combinatorial bound:
Proposition 6.5. Suppose that A is a finite set and suppose that for each integer j there is a set
Ij ⊆ A with |Ij | ≤ h. Define
X
X X
X
f (A, s, I) =
···
1
i1 ∈A
Then we have
i3 <i2
i2 <i1
i2 ∈A−Ii1 i3 ∈A−Ii1 −Ii2
is <is−1
is ∈A−Ii1 −Ii2 −···−Iis−1
|A| − (s − 1)h
f (A, s, I) ≥
s
P
Proof. Set t = |A|. We prove this by induction on s. When s = 1 then f (A, 1, I) = i1 ∈A 1 = t =
t−(1−1)h
as claimed.
1
We move to the induction step s > 1. Suppose that A = {a1 , . . . , at }, and suppose that we
select the value i1 = aj . Then the remaining sum over i2 , . . . , is is equal to f (A′j , s − 1, I), where
A′j = {a1 , . . . , aj−1 } − Ii1 . Summing over all j = 1, . . . , t gives:
f (A, s, I) =
t
X
f (A′j , s
j=1
≥
≥
=
− 1, I) ≥
t
X
j=(s−1)(h+1)
t
X
j=(s−1)(h+1)
j=s−1
j=(s−1)h+s
f (A′j , s − 1, I)
|A′j | − (s − 2)h
s−1
(j − 1 − h) − (s − 2)h
s−1
t−1−h(s−1)
X
t
X
by inductive hypothesis
as |A′j | ≥ j − 1 − h
j
t − (s − 1)h
=
s−1
s
and the induction is proved.
7. Acknowledgments
Thanks to Aravind Srinivasan and Navin Goyal, for helpful discussions and comments. Thanks
to the anonymous reviewers, for many helpful comments and corrections.
References
[1] Achlioptas, D., Iliopoulos, F.: Random walks that find perfect objects and the Lovász Local Lemma. Journal of
the ACM 63-3, #22 (2016)
[2] Bissacot, R., Fernandez, R., Procacci, A., Scoppola, B.: An improvement of the Lovász Local Lemma via cluster
expansion. Combinatorics, Probability and Computing 20-5, pp. 709-719 (2011)
[3] Chandrasekaran, K., Goyal, N., Haeupler, B.: Deterministic algorithms for the Lovász local lemma. SIAM Journal
on Computing 42-6, pp. 2132-2155 (2013)
PARALLEL ALGORITHMS AND CONCENTRATION BOUNDS FOR THE LLL VIA WITNESS DAGS
23
[4] Chung, K., Pettie, S., Hsin-Hao, S.: Distributed algorithms for the Lovász local lemma and graph coloring.
Distributed Computing, pp. 1-20 (2016)
[5] Erdős, P., Lovász, L.: Problems and results on 3-chromatic hypergraphs and some related questions. In A. Hajnal,
R. Rado, and V. T. Sos, eds. Infinite and Finite Sets II, pp. 607-726 (1975)
[6] Even, G., Goldreich, O., Luby, M., Nisan, N., Velickovic, B.: Efficient approximation of product distributions.
Random Structures and Algorithms, 13(1), pp. 1-16 (1998)
[7] Fischer, M., Ghaffari, M.: Sublogarithmic distributed algorithms for Lovász local lemma, and the complexity
hierarchy. arXiv:1705.04840 (2017)
[8] Ghaffari, M.: Towards an optimal distributed algorithm for maximal independent set. Proceedings of the 27th
Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pp. 270-277 (2016)
[9] Haeupler, B., Saha, B., Srinivasan, A.: New constructive aspects of the Lovász Local Lemma. Journal of the
ACM, 58-6, #28 (2011)
[10] Harris, D.: Lopsidependency in the Moser-Tardos framework: beyond the Lopsided Lovász Local Lemma. ACM
Transactions on Algorithms 13-1, #17 (2016)
[11] Harris, D.: Deterministic parallel algorithms for fooling polylogarithmic juntas and the Lovász Local Lemma.
Proceedings of the 28th ACM-SIAM Symposium on Discrete Algorithms (SODA), pp. 1188-1203 (2017)
[12] Harris, D.: Parallel algorithms for the lopsided Lovász local lemma. arxiv:17.02547 (2017)
[13] Harvey, N., Vondrák, J.: An algorithmic proof of the Lovász local lemma via resampling oracles. Proceedings of
the 56th Annual IEEE Symposium on Foundations of Computer Science (FOCS), pp. 1327-1346 (2015)
[14] Kolipaka, K., Szegedy, M.: Moser and Tardos meet Lovász. Proceedings of the 43rd Annual ACM Symposium
on Theory of Computing (STOC), pp. 235-244 (2011)
[15] Luby, M.: A simple parallel algorithm for the maximal independent set problem. SIAM Journal on Computing
15-4, pp. 1036-1053 (1996)
[16] Moser, R. : Derandomizing the Lovász Local Lemma more effectively. arXiv:0807.2120 (2008)
[17] Moser, R., Tardos, G.: A constructive proof of the general Lovász Local Lemma. Journal of the ACM 57-2, #11
(2010)
[18] Shearer, J. B.: On a problem of Spencer. Combinatorica 5, pp. 241-245 (1985)
| 8cs.DS |
Exact Weighted Minwise Hashing in Constant Time
Anshumali Shrivastava
Department of Computer Science, Rice University, Houston, TX, USA
arXiv:1602.08393v1 [cs.DS] 26 Feb 2016
Abstract
Weighted minwise hashing (WMH) is one
of the fundamental subroutine, required by
many celebrated approximation algorithms,
commonly adopted in industrial practice for
large scale-search and learning. The resource bottleneck of the algorithms is the
computation of multiple (typically a few
hundreds to thousands) independent hashes
of the data. The fastest hashing algorithm
is by Ioffe (Ioffe, 2010), which requires one
pass over the entire data vector, O(d) (d is the
number of non-zeros), for computing one
hash. However, the requirement of multiple hashes demands hundreds or thousands
passes over the data. This is very costly for
modern massive dataset.
In this work, we break this expensive barrier and show an expected constant amortized time algorithm which computes k independent and unbiased WMH in time O(k)
instead of O(dk) required by Ioffe’s method.
Moreover, our proposal only needs a few
bits (5 - 9 bits) of storage per hash value compared to around 64 bits required by the stateof-art-methodologies. Experimental evaluations, on real datasets, show that for
computing 500 WMH, our proposal can be
60000x faster than the Ioffe’s method without losing any accuracy. Our method is also
around 100x faster than approximate heuristics capitalizing on the efficient “densified”
one permutation hashing schemes (Shrivastava & Li, 2014a). Given the simplicity of our
approach and its significant advantages, we
hope that it will replace existing implementations in practice.
1. Introduction and Motivation
(Weighted) Minwise hashing (or Minwise Sampling)
(Broder, 1997; Broder et al., 1997; Manasse et al.,
ANSHUMALI @ RICE . EDU
2010) is the most popular and successful randomized hashing technique (Shrivastava & Li, 2014b),
commonly deployed in commercial big-data systems (Henzinge, 2004; Henzinger, 2006) for large scale
search and learning. Minwise hashing was originally developed for fast similarity estimation. However, quickly it was found that the same sampling (or
hashing) scheme made several applications over the
web, such as, near-duplicate detection (Broder, 1998;
Henzinger, 2006), all-pairs similarity (Bayardo et al.,
2007), record linkage (Koudas et al., 2006), temporal correlations (Chien & Immorlica, 2005), etc., efficient by drastically reducing their computational and
memory requirements.
The idea underlying Minwise hashing is also the key
breakthrough that led to the theory of Locality Sensitive Hashing (LSH) (Indyk & Motwani, 1998). In
particular, minwise sampling is a known LSH for the
Jaccard similarity (Rajaraman & Ullman). Given two
positive vectors x, y ∈ RD , x, y > 0, the (generalized)
Jaccard similarity is defined as
PD
min{xi , yi }
.
(1)
J(x, y) = PDi=1
i=1 max{xi , yi }
J(x, y) is a frequently used measure for comparing
web-documents (Broder, 1997), histograms (specially
images (Ioffe, 2010)), gene sequences (Rasheed &
Rangwala), etc. Recently, it was shown to be a
very effective kernel for large-scale non-linear learning (Li, 2015). Furthermore, it was found that WMH
leads to the best-known LSH for L1 distance (Ioffe,
2010), commonly used in computer vision, improving
over (Datar et al., 2004).
Weighted Minwise Hashing (WMH) (or Minwise
Sampling) generates randomized hash (or fingerprint) h(x), of the given data vector x ≥ 0, such that
for any pair of vectors x and y, the probability of hash
collision (or agreement of hash values) is given by,
P
min{xi , yi }
= J(x, y).
(2)
Pr(h(x) = h(y)) = P
max{xi , yi }
A notable special case is when x and y are binary (or
Exact Weighted Minwise Hashing in Constant Time
sets), i.e. xi , yi ∈ {0, 1}D . For this
case, the similarity
P
min{xi ,yi }
measure boils down to J(x, y) = P max{x
= |x∩y|
|x∪y| .
i ,yi }
Being able to generate a randomized signature, h(x),
satisfying Equation 2 is the key breakthrough behind
some of the best-known approximations algorithms
for metric labelling (Kleinberg & Tardos, 1999), metric
embedding (Charikar, 2002), mechanism design and
differential privacy (Dwork & Roth).
A typical requirement for algorithms relying on minwise hashing is to generate, some large enough, k
independent Minwise hashes (or fingerprints) of the
data vector x, i.e. compute hi (x) i ∈ {1, 2, ..., k} repeatedly with independent randomization. These independent hashes can then be used for a variety of data
mining tasks such as cheap similarity estimation, indexing for sublinear-search, kernel features for large
scale learning, etc. The bottleneck step in all these
applications is the costly computation of the multiple hashes, requiring multiple passes over the data.
The number of required hashes typically ranges from
few hundreds to several thousand (Shrivastava & Li,
2014a;c). For example, the number of hashes required
by the famous LSH algorithm is O(nρ ) which grows
with the size of the data. (Li, 2015) showed the necessity of around 4000 hashes per data vector in largescale learning with J(x, y) as kernel, making hash generation the most costly step.
Owing to the significance of WMH and its impact in
practice, there is a series of work over the last decade
trying to reduce the costly computation cost associated with hash generation (Haeupler et al., 2014).
The first groundbreaking work on Minwise hashing (Broder, 1997) computed hashes h(x) only for unweighted sets x (or binary vectors), i.e. when the
vector components xi s can only take values 0 and 1.
Later it was realized that vectors with positive integer weights, which are equivalent to weighted sets,
can be reduced to unweighted set by replicating elements in proportion to their weights (Gollapudi &
Panigrahy, 2006; Haeupler et al., 2014). This scheme
was very expensive due to blowup in the number of
elements caused by replications. Also, it cannot handle real weights. In (Haeupler et al., 2014), the authors
showed few approximate solutions based on ideas
developed in (Gollapudi & Panigrahy, 2006; Kleinberg & Tardos, 1999) to reduce these replications.
Later (Manasse et al., 2010), introduced the concept of
consistent weighted sampling (CWS), which focuses
on sampling directly from some well-tailored distribution to avoid any replication. This method, unlike
previous ones, could handle real weights exactly. Go-
ing a step further, Ioffe (Ioffe, 2010) was able to compute the exact distribution of minwise sampling leading to a scheme with worst case O(d), where d is the
number of non-zeros. This is the fastest known exact
weighted minwise sampling scheme so far. Since this
will be our main baseline, we review it in Section 2.
O(dk) for computing k independent hashes is very
expensive for modern massive datasets, especially
when k with ranges up to thousands. Recently, there
was a big success for the binary case, where using
the novel idea of “Densification” (Shrivastava & Li,
2014a;c) the computation time for unweighted minwise was brought down to O(d + k). This resulted
into over 100-1000 fold improvement. However, this
speedup was limited only to binary vectors. Moreover, the samples were not completely independent.
Capitalizing on recent advances for fast unweighted
minwise hashing, (Haeupler et al., 2014) exploited the
old idea of replication to convert weighted sets into
unweighted sets. To deal with non-integer weights,
the method samples the coordinates with probabilities proportional to leftover weights. The overall process converts the weighted minwise sampling to an
unweighted problem, however, at a cost of incurring
some bias (see Algorithm 2). This scheme is faster
than Ioffe’s scheme but, unlike other prior works on
CWS, it is not exact and leads to biased and correlated samples. Moreover, it requires strong independence (Indyk, 2001) leading to poor estimation.
All these lines of work lead to a natural question:
does there exist an unbiased and independent WMH
scheme with same property as Ioffe’s hashes but significantly faster than all existing methodologies? We
answer this question positively.
1.1. Our Contributions:
1. We provide an unbiased weighted minwise hashing scheme, where each sampling scheme takes
constant time in expectation. This improves
upon the best-known scheme in the literature by
Ioffe (Ioffe, 2010) which takes O(d), where d is the
number of non-zeros in the data vector.
2. Our hashing scheme requires much fewer bits
usually (5-9) bits instead of 64 bits (or higher) required by existing schemes, leading to around 8x
savings in space.
3. We derive our scheme from elementary first principles. Our scheme is simple and it only requires access to simple uniform random number
generator, instead of costly sampling needed by
other methods. The hashing procedure is differ-
Exact Weighted Minwise Hashing in Constant Time
ent from traditional schemes and could be of independent interest in itself.
4. Experimental evaluations on real image histograms show more than 60000x speedup over
the best known exact scheme and around 100x
times faster than biased approximate schemes
based on the recent idea of fast minwise hashing.
5. Weighted Minwise sampling is a fundamental
subroutine in many celebrated approximation algorithms. Some of the immediate consequences
of our proposal are as follows:
• We obtain an algorithmic improvement,
over the query time of LSH based algorithm, for L1 distance and Jaccard Similarity search. In particular, we reduce the
worst-case query time of (K, L) parameterized LSH algorithm (Andoni & Indyk, 2006)
from O(dKL) to mere O(KL + dL).
• We reduce the kernel feature (Rahimi &
Recht, 2007) computation time with minmax kernels (Li, 2015).
• We reduce the sketching time for fast estimation of a variety of measures, including L1
and earth mover distance (Kleinberg & Tardos, 1999; Charikar, 2002).
2. Review: Ioffe’s Algorithm and Fast
Unweighted Minwise Hashing
Algorithm 1 Ioffe’s CWS (Ioffe, 2010)
input Vector x, random seed[][]
for i = 1 to k do
for Iterate over x j s.t x j > 0 do
randomseed = seed[i][ j];
Sample ri, j , ci, j ∼ Gamma(2, 1).
Sample
βi, j ∼ Uni
f orm(0, 1)
tj =
log x j
ri, j
+ βi, j
y j = exp(ri, j (t j − βi, j ))
z j = y j ∗ exp(ri, j )
a j = ci, j /z j
end for
k∗ = arg min j a j
HashPairs[i] = (k∗ , tk∗ )
end for
RETURN HashPairs[]
We briefly review the state-of-the-art methodologies
for Weighted Minwise Hashing (WMH). Since WMH
is only defined for weighted sets, our vectors under
consideration will always be positive, i.e. every xi ≥
0. D will denote the dimensionality of the data, and
we will use d to denote the number (or the average)
of non-zeros of the vector(s) under consideration.
The fastest known scheme for exact weighted minwise hashing is based on an elegant derivation of
the exact sampling process for “Consistent Weighted
Sampling” (CWS) due to Ioffe (Ioffe, 2010). This made
computation of a single weighted minwise hash h(x)
possible in O(d). Although this is still slow for modern massive datasets, it is a good improvement over
the prior works which rely on expensive rejection
based sampling to sample from the CWS distribution (Manasse et al., 2010). Ioffe’s overall algorithm,
samples from some smartly tailored distribution to
achieve the required Equation 2. The overall procedure is summarized in Algorithm 1.
O(d) for a single hash computation is quite expensive.
Even the unweighted case of minwise hashing had
complexity O(d) per hashes, until recently. (Shrivastava & Li, 2014a;c) showed a new one permutation
based scheme for generating k near-independent unweighted minwise hashes in O(d + k) breaking the old
O(dk) barrier. However, this improvement does not
directly extend to the weighted case.
It was known that with some bias, weighted minwise sampling can be reduced to an unweighted minwise sampling using the idea of replicating weights
in proportion to their probabilities (Gollapudi & Panigrahy, 2006; Kleinberg & Tardos, 1999). Algorithm 2
describes such a procedure. A reasonable idea is then
to use the fast unweighted hashing scheme, on the
top of this biased approximation. The inside for-loop
in Algorithm 2 blows up the number of non-zeros in
the returned unweighted set. This makes the process
slower and dependent on the magnitude of weights.
Moreover, unweighted sampling requires very costly
random permutations for good accuracy (Pătraşcu &
Thorup, 2010), which is further biased due to bias incurred during approximation.
Both the Ioffe’s scheme and the biased un-weighted
approximation scheme generate big hash values requiring 32-bits or higher storage per hash value. For
reducing this to a manageable size of say 4-8 bits, a
commonly adopted practical methodology is to randomly rehash it to smaller space at the cost of loss
in accuracy (Li & König, 2011). It turns out that our
hashing scheme only generates 5-9 bits values, h(x),
satisfying Equation 2, without losing any accuracy.
Next, we present a simple hashing procedure for
weighted minwise hashing, which is unbiased and
runs in expected constant time. We will later demon-
Exact Weighted Minwise Hashing in Constant Time
strate that it can be around 60000x faster than Ioffe’s
scheme and around 100 times faster, in practice, than
the fast un-weighted approximation.
x
𝒙𝟏
𝑴𝟎 = 𝟎
Algorithm 2 Reduce to Unweighted (Haeupler et al.,
2014)
input Vector x,
y
S =φ
for Iterate over x j s.t x j > 0 do
f loorx j = bx j c
for i = 1 to f loorx j do
S = S ∪ (i, j)
end for
r = Uni f orm(0, 1)
if r ≤ x j − f loorx j then
S = S ∪ ( f loorx j + 1, j)
end if
end for
RETURN S (unweighted set)
We will denote the ith component of vector x ∈ RD by
xi . Let mi be the upper bound on the value of component xi in the given dataset. We can always assume
the mi to be an integer, otherwise we take the ceiling
dmi e as our upper bound. Define
k=1
mi = M D = M
𝒚𝟒
𝒚𝟑 = 𝟎
𝑴𝟏
𝑴𝟐
𝑴𝟑
𝑴𝟒 = 𝑴
(3)
Initialise Hashes[] to all 0s.
for i = 1 to k do
randomseed = seed[i];
while true do
r = M × Uni f orm(0, 1);
if ISGREEN(r), (check if r ∈ xred then
break;
end if
randomseed = dr ∗ 1000000e ;
Hashes[i] + +;
end while
end for
RETURN Hashes
Our sampling procedure simply draws an independent random real number between [0, M], if the random number lies in the red region we repeat and resample. We stop the process as soon as the generated
random number lies in the green region. Our hash
value for a given data vector, h(x), is simply the number of steps taken before we stop. We summarize the
procedure in Algorithm 3. More formally,
k=1
If the data is normalized, then mi = 1 and M = D.
Given a vector x, we first create a red-green map associated with it, as shown in Figure 3.1. For this, we first
take an interval [0, M] and divide it into D disjoint intervals, with ith interval being [Mi−1 , Mi ] which is of
PD
the size mi . Note that i=1
mi = M, so we can always
do that. We then create two regions, red and green.
For the ith interval [Mi−1 , Mi ], we mark the subinterval
[Mi−1 , Mi−1 +xi ] as green and the rest [Mi−1 +xi , Mi ] with
red, as shown in Figure 3.1. If xi = 0 for some i, then
the whole ith interval [Mi−1 , Mi ] is marked as red.
Formally, for a given vector x, define the green xgreen
and the red xred regions as follows
D
xgreen = ∪i=1
[Mi , Mi +xi ];
𝒚𝟐
𝑴𝟒 = 𝑴
𝑴𝟑 𝒙 𝟒
𝑴𝟐
Algorithm 3 Weighted MinHash
input Vector x, Mi ’s, k, random seed[].
3.1. Procedure
D
X
𝒙𝟑
Figure 1. Illustration of Red-Green Map of 4 dimensional
vectors x and y.
We first describe our procedure in details. We will
later talk about the correctness of the scheme, and we
will show that our proposed scheme can be implemented in expected constant time.
mi = Mi . and
𝒚𝟏
𝑴𝟎 = 𝟎
3. Our Proposal: New Hashing Scheme
i
X
𝒙𝟐
𝑴𝟏
D
xred = ∪i=1
[Mi +xi , Mi+1 ]; (4)
Definition 1 Define {ri : i = 1, 2, 3....} as a sequence of
i.i.d uniformly generated random number between [0, M].
Then we define the hash of x, h(x) as
h(x) = arg min ri , s.t. ri ∈ xgreen
i
(5)
We want our hashing scheme to be consistent (Ioffe,
2010) across different data points to guarantee Equation 2. This requires ensuring the consistency of
the random numbers in hashes (Ioffe, 2010). We
can achieve the required consistency, either by pregenerating sequence of random numbers and storing
them analogous to other hashing schemes. However,
there is an easy way to generate a sequence of random numbers by ensuring the consistency of the random seed. This does not require any storage, except
Exact Weighted Minwise Hashing in Constant Time
the starting seed. Our Algorithm 3 uses this criterion,
to ensure that the sequence of random numbers is the
same, we start with a fixed random seed for generating random numbers. If the generated random number lies in the red region, then before re-sampling, we
reset the seed of our random number generator as a
function of discarded random number. In the algorithm, we used d100000 ∗ re, where de is the ceiling operation, as a convenient way to ensure the consistency
of sequence, without any memory overhead. This
also works nicely in practice. Since we are sampling
real numbers, the probability of any repetition (or cycle) is zero. For generating k independent hashes we
just use different random seeds which are kept fixed
for the entire dataset.
3.2. Proof of Correctness
We show that the simple scheme given in Algorithms 3 actually does possess the required property,
i.e. for any pair of points x and y Equation 2 holds.
Unlike the previous works on this line (Manasse et al.,
2010; Ioffe, 2010) which requires computing the exact
distribution of associated quantities, the proof of our
proposed scheme is elementary and can be derived
from first principles. This is not surprising given the
simplicity of our procedure.
Theorem 1 For any two vectors x and y, we have
PD
min{xi , yi }
Pr h(x) = h(y) = J(x, y) = PDi=1
i=1 max{xi , yi }
(6)
(7)
For any pair of points x and y, consider the following
three events: 1) h(x) = h(y) = j, 2) h(x) > h(y) = j and
3) j = h(x) < h(y). Observe that,
h(x) = h(y) = j if and only if r j ∈ xgreen ∩ ygreen
(8)
h(x) > h(y) = j if and only if r j ∈ ygreen − xgreen
(9)
h(y) > h(x) = j if and only if r j ∈ xgreen − ygreen
(10)
Since r j is uniformly chosen, we have,
Pr h(x) = h(y)
|xgreen ∩ ygreen |
|(xgreen ∩ ygreen ) ∪ (xgreen − ygreen ) ∪ (ygreen − xgreen )|
|xgreen ∩ ygreen |
=
(11)
|xgreen ∪ ygreen |
=
D
D
|xgreen ∩ ygreen | = | ∪i=1
[Mi , Mi + xi ] ∩ ∪i=1
[Mi , Mi + yi ]|
D
X
D
= | ∪i=1
[Mi , Mi + min{xi , yi }]| =
min{xi , yi }
(12)
i=1
and
D
D
|xgreen ∪ ygreen | = | ∪i=1
[Mi , Mi + xi ] ∪ ∪i=1
[Mi , Mi + yi ]|
D
= | ∪i=1
[Mi , Mi + max{xi , yi }]| =
D
X
max{xi , yi },
(13)
i=1
Theorem 1 implies that the sampling process is exact
and we automatically have an unbiased estimator of
J(x, y), using k independently generated WMH, hi (x)s
from Algorithm 3.
k
1 X
ˆ = J(x, y);
Jˆ =
1{hi (x) =hi (y)} ; E( J)
k i=1
ˆ =
Var( J)
J(x, y)(1 − J(x, y))
,
k
(14)
(15)
where 1 is the indicator function.
3.3. Running Time Analysis and Fast
Implementation
Proof: First note that, every number between [0, M]
is random and equally likely in a random sampling.
Therefore, for a given point x, at the time we stop
we sample uniformly from the green region xgreen =
D
∪i=1
[Mi , Mi + xi ]. Consider the index j defined as,
j = min{h(x), h(y)}
The proof follows from substituting the values of
|xgreen ∩ ygreen | and |xgreen ∪ ygreen | given by:
Define
Size of green region
sx =
=
Size of red region
PD
i=1
M
xi
=
||x||1
,
M
(16)
as the effective sparsity of the vector x. Note that this
is also the probability of Pr(r ∈ xgreen ). Algorithm 3
has a while loop.
We show that the expected times the while loops runs,
which is also the expected value of h(x), is a small constant. Formally,
Theorem 2
1 − sx
1
; Var(h(x)) =
;
sx
s2x
log δ
Pr h(x) ≥
≤ δ.
log (1 − s x )
E(h(x)) =
(17)
(18)
Proof: Equation 17 follows immediately from the fact
that the number of sampling step taken before the
process stops,which is also h(x) is a geometric random
variable with p = s x . Equation 18 follows from observing that Pr(h(x) > k) = (1 − s x )k ≤ δ which implies
δ
k ≤ loglog
(1−s x ) yielding the required bound.
Exact Weighted Minwise Hashing in Constant Time
Remarks: The time to compute each hash value in expectation is the inverse of effective sparsity 1s . Therefore, for datasets with 1s << d, we can expect our
method to be much faster. For real datasets, such as
image histograms, where minwise sampling is popular(Ioffe, 2010), the value of this sparsity is of the order of 0.02-0.08 (see Section 5.2) leading to s1x ≈ 13−50.
On the other hand, the number of non-zeros is around
half million. Therefore, we can expect significant
speed-up in practice.
As a consequence of Theorem 2 we have:
Corollary 1 The expected amount of bits required to represent h(x) is small, in particular,
E(bits) ≤ − log s x ; E(bits) ≈ log
1 (1 − s x )
;
−
sx
2
(19)
Proof: The proof follows from Jensens Inequality,
E(log x) ≤ log E(x) and second order Taylor series apVar(x)
proximation of E(log x) ≈ log E(x) − 2 log
E(x)2
Remarks: Existing hashing scheme require 64 bits,
which is quite expensive. A popular approach for reducing space uses least significant bits of hashes (Li &
König, 2011; Ioffe, 2010). This tradeoff in space comes
at the cost of accuracy (Li & König, 2011). Our hashing scheme naturally requires only few bits, typically
5-9 (see Section 5.2), eliminating the need for trading
accuracy for manageable space.
We know from Theorem 2 that each hash function
computation requires a constant number of function calls to ISGREEN(r). If we can implement ISGREEN(r) in constant time, i.e O(1), then we can generate generate k independent hashes in total O(d + k)
time instead of O(dk) required by (Ioffe, 2010). Note
that O(d) is the time to read the input vector which
cannot be avoided. Once the data is loaded into the
memory, our procedure is actually O(k) for computing k hashes, for all k ≥ 1.
Before we jump into a constant time implementation,
we would like readers to note that there is a straightforward binary search algorithm for ISGREEN(r) in
log d time. We consider d intervals [Mi , Mi + xi ] for all i,
such that xi , 0. Because of the nature of the problem,
Mi−1 + xi−1 ≤ Mi ∀i. Therefore, these intervals are disjoint and sorted. Therefore, given a random number
D
r, determining if r ∈ ∪i=1
[Mi , Mi + xi ] only needs binary
search over d ranges. Thus, given that we only need
constant number of random sampling in expectation
before we stop, we already have a scheme that generates k independent hashes in total O(d + k log d) time
improving over best known O(dk) required by (Ioffe,
2010) for unbiased sampling.
Algorithm 4 ComputeHashMaps (Once per dataset)
input Mi ’s,
index =0, CompToM[0] =0
for i = 0 to D − 1 do
if i < D − 1 then
CompT oM[i + 1] = Mi + CompT oM[i]
end if
for j = 0 to Mi − 1 do
IntT oComp[index] = i
index++
end for
end for
RETURN CompToM[] and IntToComp[]
We show that with some algorithmic tricks and few
more data structures, we can implement ISGREEN(r)
in constant time O(1). We need two global precomputed hashmaps, IntToComp (Integer to Vector
Component) and CompToM (Vector Component to M
value). IntToComp is a hashmap that maps every integer between [0, M] to the associated components, i.e.,
all integers between [Mi , Mi+1 ] are mapped to i, because it is associated with ith component. CompToM
maps every component of vectors i ∈ {1, 2, 3, ..., D}
to its associated value Mi . The procedure for computing these hashmaps is straightforward and is summarized in Algorithm 4. It should be noted that these
hash-maps computation is a one time pre-processing
operation over the entire dataset having a negligible
cost. Mi ’s can be computed (estimated) while reading
the data.
Algorithm 5 ISGREEN(r)
input r, x, Hashmaps IntToComp[] and CompToM[]
from Algorithm 4.
index = dre
i = IntT oComp[index]
Mi = CompT oM[i]
if r ≤ Mi + xi then
RETURN TRUE
end if
RETURN FALSE
Using these two pre-computed hashmaps, the ISGREEN(r) methodology works as follows: We first
compute the ceiling of r, i.e. dre, then we find the
component i associated with r, i.e., r ∈ [Mi , Mi+1 ],
and the corresponding associated Mi using hashmaps
IntToComp and CompToM. Finally, we return true if
r ≤ xi + Mi otherwise we return false. The main observation is that since we ensure that all Mi ’s are Integers,
for any real number r, if r ∈ [Mi , Mi+1 ] then the same
Exact Weighted Minwise Hashing in Constant Time
holds for dre, i.e., dre ∈ [Mi , Mi+1 ]. Hence we can work
with hashmaps using dre as the key. The overall procedure of ISGREEN(r) is summarized in Algorithm 5.
over existing methods. We also need to validate our
theory that our scheme is unbiased and should be indistinguishable in accuracy with Ioffe’s method.
Note that our overall procedure is much simpler compared to Algorithm 1. We only need to generate random numbers followed by a simple condition check
using two hash lookups. Our analysis shows that
we have to repeat this only for a constant number of
times. Compare this with the scheme of Ioffe where
for every non-zero component of a vector we need to
sample two Gamma variables followed by computing several expensive transformations including exponentials. We next demonstrate the benefits of our
approach in practice.
Baselines: Ioffe’s method is the fastest known exact
method in the literature, so it serves as our natural
baseline. We also compare our method with biased
unweighted approximations (see Algorithm 2) which
capitalizes on recent success in fast unweighted minwise hashing (Shrivastava & Li, 2014a;c), we call it
Fast-WDOPH (for Fast Weighted Densified One Permutation Hashing). Fast-WDOPH needs very long
permutation, which is expensive. For efficiency, we
implemented the permutation using fast 2-universal
hashing which is also what is recommended in practice (Mitzenmacher & Vadhan, 2008).
4. Practical Issues
Datasets: Weighted Minwise sampling is commonly
used for sketching image histograms (Ioffe, 2010). We
chose two popular publicly available vision dataset
Caltech101 (Fei-Fei et al., 2007) and Oxford (Philbin
et al., 2007). We used the standard publicly available
Histogram of Oriented Gradient (HOG) codes (Dalal
& Triggs, 2005), popular in vision task, to convert images into feature vectors. In addition, we also used
random web images (Wang et al., 1999) and computed simple histograms of RGB values. We call
this dataset as Web-Images Hist. The basic statistics of these datasets is summarized in Table 5. These
datasets cover a wide range of variations in terms of
dimensionality, non-zeros and sparsity.
PD
xi
PDi=1
i=1 Mi
deterTheorem 2 shows that the quantity s x =
mines the runtime. If s x is very very small then, although the running time is constant (independent of
d or D), it can still make the algorithm unnecessarily
slow. Note that for the algorithm to work we choose
Mi to be the largest integer greater than the maximum
possible value of co-ordinate i in the given dataset. If
this integer gap is big then we unnecessarily increase
the running time. Ideally, the best running time is obtained when the maximum value, is itself an integer,
or is very close to its ceiling value. If all the values
are integers, scaling up does not matter, as it does not
change s x , but scaling down can make s x worse. IdePD
αx
ally we should scale, such that, α = arg maxα = PDi=1dαmi e
i
i=1
is maximized, where mi is the maximum value of coordinate i in the given dataset.
5. Experiments
Table 1. Basic Statistics of the Datasets Used in the Paper
Data
Web-Images
Hist
Caltech101
Oxford
non-zeros (d)
Dim (D)
737
768
95029
401879
485640
580644
Sparsity
(s)
0.081
0.024
0.086
We have proposed a new Weighted Minwise Hashing scheme which is significantly faster in theory than
existing methodologies. In this section, we would
like to assess its impact in practice. We compare
the proposed scheme with the existing state-of-artmethodologies on estimation accuracy, running time
and space. In particular, we will demonstrate that
in real high-dimensional settings, our proposal provides significant speedup and requires less memory
Table 2. Time taken in milliseconds (ms) to compute 500
hashes by different schemes. Our proposed scheme is significantly faster.
Method
Prop
Ioffe
Web-Images Hist
Caltech101
Oxford
10ms
57ms
11ms
986ms
87105ms
746120ms
FastWDOPH
57ms
268ms
959ms
5.1. Comparing Estimation Accuracy
Table 3. The range of the observed hash values, using the
proposed scheme, along with the maximum bits needed per
hash value. The mean hash values agrees with Theorem 2
Mean Values
Hash Range
Bits Needed
WebImages
Hist
11.94
[1,107]
7
Caltech101 Oxford
52.88
[1,487]
9
9.13
[1,69]
7
In this section, we perform a sanity check experiment
and compare the estimation accuracy with WMH. For
this task we take 9 pairs of vectors from our datasets
with varying level of similarities. For each of the pair
Exact Weighted Minwise Hashing in Constant Time
Sim=0.8
0.05
0.15
Sim=0.72
0.1
0.05
0
0.15
0.1
0.05
0
10
20
30
40
50
0
10
Number of Hashes
20
30
40
50
0.1
0
0.2
Sim=0.44
0.1
0
10
20
30
40
Number of Hashes
50
20
30
40
10
20
30
40
Number of Hashes
50
20
30
40
50
Number of Hashes
0.4
Proposed
Fast-WDOPH
Ioffe
0.3
Sim=0.35
0.2
0.1
0
10
Sim=0.61
0.1
50
0.4
Proposed
Fast-WDOPH
Ioffe
Average Error
Average Error
0.2
Proposed
Fast-WDOPH
Ioffe
0.2
Number of Hashes
0.3
Sim=0.56
Proposed
Fast-WDOPH
Ioffe
0
10
Number of Hashes
0.3
Average Error
0.2
0.3
Proposed
Fast-WDOPH
Ioffe
Sim=0.67
Average Error
0.1
0.2
Average Error
0.15
0.25
Proposed
Fast-WDOPH
Ioffe
Average Error
0.25
Proposed
Fast-WDOPH
Ioffe
Average Error
Average Error
0.2
Proposed
Fast-WDOPH
Ioffe
0.3
Sim=0.27
0.2
0.1
0
10
20
30
40
50
Number of Hashes
10
20
30
40
50
Number of Hashes
Figure 2. Average Errors in Jaccard Similarity Estimation with the Number of Hash Values. Estimates are averaged over
200 repetitions.
(x, y), we generate k weighted minwise hashes hi (x)
and hi (y) for i ∈ {1, 2, .., k}, using the three competing
schemes. We then compute the estimate of the Jaccard
P
similarity J(x, y) using the formula 1k ki=1 1{hi (x) =
hi (y)} (See Equation 14). We compute the errors in
the estimate as a function of k. To minimize the effect of randomization, we average the errors from 200
random repetitions with different seeds. We plot this
average error with k = {1, 2, ..., 50} in Figure 2 for
different similarity levels.
We can clearly see from the plots that the accuracy
of the proposed scheme is indistinguishable from
Ioffe’s scheme. This is not surprising because both the
schemes are unbiased and have the same theoretical
distribution. This validates Theorem 1
The accuracy of Fast-WDOPH is inferior to that of the
other two unbiased schemes and sometimes its performance is poor. This is because the weighted to unweighted reduction is biased and approximate. The
bias of this reduction depends on the vector pairs under consideration, which can be unpredictable.
5.2. Speed Comparisons
In this section, we demonstrate the actual benefits of
having a constant time and low memory approach
over existing methodologies. For this task, we compute the average time (in milliseconds) taken by the
competing algorithms to compute 500 hashes of a
given data vector for all the three datasets. Our experiments were coded in C# on Intel Xenon CPU with 256
GB RAM. Table 2 summarises the comparison. We
do not include the data loading cost in these numbers
and assume that the data is in the memory for all the
three methodologies.
We can clearly see tremendous speedup over Ioffe’s
scheme. For Web-Images dataset with mere 768 nonzeros, our scheme is 100 times faster than Ioffe’s
scheme and around 5 times faster than Fast-WDOPH
approximation. While on caltech101 and Oxford
datasets, which are high dimensional and dense
datasets, our scheme can be 1500x to 70000x faster
than Ioffe’s scheme, while it is around 5 to 100x times
faster than Fast-WDOPH scheme. Dense datasets like
Caltech101 and Oxford represent more realistic scenarios. These features are taken from real applications (Dalal & Triggs, 2005) and such level of sparsity
and dimensionality are more common in practice.
The results are not surprising because Ioffe’s scheme
is very slow O(dk). Moreover, the constant are inside bigO is also large, because of complex transformations. Therefore, for datasets with high values of
d (non-zeros) this scheme is very slow. Similar phenomena were observed in (Ioffe, 2010), that decreasing the non-zeros by ignoring non-frequent dimensions can be around 150 times faster. However, ignoring dimension looses accuracy.
5.3. Memory Comparisons
Table 3 summarizes the range of the hash values and
the maximum number of bits needed to encode these
hash values without any bias. We can clearly see
that the hash values, even for such high-dimensional
datasets, only require 7-9 bits. This is a huge saving compared to existing hashing schemes which can
require (32-64) bits (Li & König, 2011). Note, Ioffe’s
scheme generates a pair of integers as hashes which
Exact Weighted Minwise Hashing in Constant Time
has significant storage cost. Thus, our method leads
to around 5-6 times savings in space and it eliminates
the need of shrinking hashes to small bits with a loss
in the estimation accuracy. The mean values observed
(Table 3) validate the formula in Theorem 2.
6. Conclusions
We show that the existing popular weighted minwise
hashing scheme can be made simpler, significantly
faster and memory efficient. Our theoretical and empirical results show that these advantages come for
free and the hashes are indistinguishable from the unbiased scheme of Ioffe.
Experiments on real datasets show up to 60000
fold speedup in computing 500 hashes compared to
the state-of-the-art procedure. We believe that our
scheme will replace existing implementations in bigdata systems.
References
Andoni, Alexandr and Indyk, Piotr. Near-optimal
hashing algorithms for approximate nearest neighbor in high dimensions. In FOCS, pp. 459–468,
Berkeley, CA, 2006.
Bayardo, Roberto J., Ma, Yiming, and Srikant, Ramakrishnan. Scaling up all pairs similarity search.
In WWW, pp. 131–140, 2007.
Broder, Andrei Z. On the resemblance and containment of documents. In the Compression and Complexity of Sequences, pp. 21–29, Positano, Italy, 1997.
Broder, Andrei Z. Filtering near-duplicate documents. In FUN, Isola d’Elba, Italy, 1998.
Broder, Andrei Z., Glassman, Steven C., Manasse,
Mark S., and Zweig, Geoffrey. Syntactic clustering
of the web. In WWW, pp. 1157 – 1166, Santa Clara,
CA, 1997.
Charikar, Moses S. Similarity estimation techniques
from rounding algorithms. In STOC, pp. 380–388,
Montreal, Quebec, Canada, 2002.
Chien, Steve and Immorlica, Nicole. Semantic similarity between search engine queries using temporal correlation. In WWW, pp. 2–11, 2005.
Dalal, Navneet and Triggs, Bill. Histograms of oriented gradients for human detection. In Computer
Vision and Pattern Recognition, volume 1, pp. 886–
893. IEEE, 2005.
Datar, Mayur, Immorlica, Nicole, Indyk, Piotr, and
Mirrokn, Vahab S.
Locality-sensitive hashing
scheme based on p-stable distributions. In SCG, pp.
253 – 262, Brooklyn, NY, 2004.
Dwork, Cynthia and Roth, Aaron. The algorithmic
foundations of differential privacy.
Fei-Fei, Li, Fergus, Rob, and Perona, Pietro. Learning
generative visual models from few training examples: An incremental bayesian approach tested on
101 object categories. Computer Vision and Image Understanding, 106(1):59–70, 2007.
Gollapudi, Sreenivas and Panigrahy, Rina. Exploiting
asymmetry in hierarchical topic extraction. In Proceedings of the 15th ACM international conference on
Information and knowledge management, pp. 475–482.
ACM, 2006.
Haeupler, Bernhard, Manasse, Mark, and Talwar,
Kunal. Consistent weighted sampling made fast,
small, and easy. Technical report, arXiv:1410.4266,
2014.
Henzinge, Monika .R. Algorithmic challenges in web
search engines. Internet Mathematics, 1(1):115–123,
2004.
Henzinger, Monika. Finding near-duplicate web
pages: a large-scale evaluation of algorithms. In
Proceedings of the 29th annual international ACM SIGIR conference on Research and development in information retrieval, pp. 284–291. ACM, 2006.
Indyk, Piotr. A small approximately min-wise independent family of hash functions. Journal of Algorithms, 38(1):84–90, 2001.
Indyk, Piotr and Motwani, Rajeev. Approximate nearest neighbors: Towards removing the curse of dimensionality. In STOC, pp. 604–613, Dallas, TX,
1998.
Ioffe, Sergey.
Improved consistent sampling,
weighted minhash and L1 sketching. In ICDM, pp.
246–255, Sydney, AU, 2010.
Kleinberg, Jon and Tardos, Eva. Approximation algorithms for classification problems with pairwise
relationships: Metric labeling and Markov random
fields. In FOCS, pp. 14–23, New York, 1999.
Koudas, Nick, Sarawagi, Sunita, and Srivastava, Divesh. Record linkage: similarity measures and algorithms. In Proceedings of the 2006 ACM SIGMOD
international conference on Management of data, pp.
802–803. ACM, 2006.
Exact Weighted Minwise Hashing in Constant Time
Li, Ping. 0-bit consistent weighted sampling. In KDD,
2015.
Li, Ping and König, Arnd Christian. Theory and applications b-bit minwise hashing. Commun. ACM,
2011.
Manasse, Mark, McSherry, Frank, and Talwar, Kunal.
Consistent weighted sampling. Technical Report
MSR-TR-2010-73, Microsoft Research, 2010.
Mitzenmacher, Michael and Vadhan, Salil. Why simple hash functions work: exploiting the entropy in
a data stream. In Proceedings of the nineteenth annual
ACM-SIAM symposium on Discrete algorithms, pp.
746–755. Society for Industrial and Applied Mathematics, 2008.
Philbin, J., Chum, O., Isard, M., Sivic, J., and Zisserman, A. Object retrieval with large vocabularies
and fast spatial matching. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition, 2007.
Pătraşcu, Mihai and Thorup, Mikkel. On the kindependence required by linear probing and minwise independence. In ICALP, pp. 715–726, 2010.
Rahimi, Ali and Recht, Benjamin. Random features
for large-scale kernel machines. In Advances in
neural information processing systems, pp. 1177–1184,
2007.
Rajaraman, Anand and Ullman, Jeffrey. Mining of
Massive Datasets.
Rasheed, Zeehasham and Rangwala, Huzefa. Mcminh: Metagenome clustering using minwise
based hashing. SIAM.
Shrivastava, Anshumali and Li, Ping. Densifying
one permutation hashing via rotation for fast near
neighbor search. In ICML, Beijing, China, 2014a.
Shrivastava, Anshumali and Li, Ping. {In Defense of
Minhash over Simhash}. In Proceedings of the Seventeenth International Conference on Artificial Intelligence and Statistics, pp. 886–894, 2014b.
Shrivastava, Anshumali and Li, Ping. Improved densification of one permutation hashing. In UAI, Quebec, CA, 2014c.
Wang, J, Li, Jia, Chan, Desmond, and Wiederhold,
Gio. Semantics-sensitive retrieval for digital picture
libraries. D-Lib Magazine, 5(11), 1999.
| 8cs.DS |
SHEAFIFICATIONS OF THE DE RHAM-WITT COMPLEX
arXiv:1703.08868v1 [math.AC] 26 Mar 2017
VERONIKA ERTL AND LANCE EDWARD MILLER
Abstract. Recent important and powerful frameworks for the study of differential forms by Huber-Jörder and
Huber-Kebekus-Kelly based on Voevodksy’s h-topology have greatly simplified and unified many approaches. This
article builds towards the goal of putting Illusie’s de Rham-Witt complex in the same framework by exploring the hsheafification of the rational de Rham-Witt differentials. Assuming resolution of singularities in positive characteristic
one recovers a complete cohomological h-descent for all terms of the complex. We also provide unconditional hdescent for the global sections and draw the expected conclusions. The approach is to realize that a certain right
Kan extension introduced by Huber-Kebekus-Kelly takes the sheaf of rational de Rham-Witt forms to a qfh-sheaf.
As such, we state and prove many results about qfh-sheaves which are of independent interest.
1. Introduction
Among the most ubiquitous tools for studying varieties over a field are differential forms, or more formally de
Rham cohomology and its Hodge type variants. These are well-known to only be well-behaved for smooth varieties
over the complex numbers or other algebraically closed fields of characteristic 0; the key facet being Hironaka’s
resolution of singularities. Adding to the difficulty in extending these methods, even defining such forms outside the
smooth case becomes a challenge, with many variants tailored to specific considerations, e.g., reflexive differentials
for the normal setting, log differentials, the Du Bois complex etc. However, recent work of Huber–Jörder [HJ14] and
Huber–Kebekus–Kelly [HKK15] have described a successful and powerful approach unifying and conceptualizing
studies using differential forms by introducing sheaf theoretic methods utilizing various Grothendieck topologies,
notably Voevodsky’s h-topology. This approach is built on the intuition that all varieties over a field of characteristic
0 are h-locally smooth. The h-topology meets challenges in positive characteristic, where one quickly notes that
the basic Kähler differentials become zero under h-sheafification owing to the Frobenius map being an h-cover.
However, more subtle sheaf methods help to overcome this difficulty via a specific right Kan extension.
In p-adic Hodge theory, differential forms are rarely themselves the right consideration and one notes the ubiquity
of Illusie’s de Rham-Witt complex and its variants. The aim of this article is to consider these complexes and their
sheafifications under the same Grothendieck topologies considered in [HJ14] and [HKK15]. Specifically, we consider
h-descent statements for the differentials W ΩnQ = W Ωn ⊗ Q of the rational de Rham–Witt complex. An optimal
result would be a cohomological descent statement analogous to [HJ14, Cor. 6.5]. However, we can only recover it
up to an assumption of resolution of singularities in positive characteristic. In particular, we describe a proof of the
following theorem, which is well-known to experts.
Theorem. Let X be smooth variety over a perfect field of characteristic p. Assume that resolution of singularities
holds. Consider the h-sheaf and W ΩnQ,h . For all i ≥ 0 and n ≥ 0
i
(X, W ΩnQ ).
Hhi (X, W ΩnQ,h ) ∼
= HZar
Avoiding resolution of singularities presents, at present, an insurmountable challenge. Undeterred, we exploit
many of the techniques of [HKK15] to recover a descent statement for all n and i = 0 free of any assumption on
resolution of singularities. This relies strongly on considering the ( )dvr -construction, which is related to the hsheafification. In particular, for a base scheme S, denote by Sch(S) the category of separated finite type S-schemes
and Reg(S) the full subcategory of regular schemes. The restriction Sch(S) → Reg(S) has a right adjoint ( )dvr on
the category of presheaves, and this auxiliary sheaf is critical to extending the results of [HJ14] to the de Rham–Witt
setting. The biggest new result of the paper shows that this presheaf is a sheaf for the quasi-finite or qfh-topology.
Theorem. Let k be a perfect field of characteristic p > 0. The presheaf W ΩnQ,dvr is a qfh-sheaf on Sch(k).
The key point is that under expected circumstances, for sheaves such as these one may establish an h-descent.
Theorem. Fix a base scheme S and a qfh-sheaf F on Sch(S). When F is also an eh-sheaf, F dvr satisfies h-descent
on Sch(S). In particular, for a perfect field of characteristic p, the sheaves W ΩnQ,dvr on Sch(k) satisfy h-descent.
We draw from this many expected consequences, in particular recovering an analogue of [HJ14, Prop. 4.2] for
the rational de Rham-Witt differentials.
1
Page 2 of 14
Acknowledgments: This paper owes a significant intellectual debt to Annette Huber and benefited from many
helpful discussions with David Zureick-Brown, Shane Kelly, and Bhargav Bhatt. The authors would also like to
thank the University of Freiburg, and their home institutions for making key visitations possible.
2. Preliminaries and notational conventions
Throughout this work we consider Grothendeick topologies. We refer to [SP] as a general reference for the
uninitiated reader. All schemes are assumed to be separated. For a fixed noetherian base scheme S denote by
Sch(S) the category of separated schemes of finite type over S. When S = Spec R is affine, we conflate notation
writing Sch(R) for Sch(S). We denote by Reg(S) the full subcategory of Sch(S) of regular schemes, by Nor(S) the
full subcategory of Sch(S) of normal schemes, and by Var(S) the full subcategory of Sch(S) which are reduced and
irreducible. For a Grothendieck topology τ on Sch(S) we denote by Sch(S)τ the associated site and by Sh(S)τ the
associated topos.
Fix a base S. It will be critically important to distinguish various sheafifications involving complexes of sheaves on
Sch(S). Consider τ a topology on the category Sch(S). For a presheaf F on Sch(S), denote by F τ the sheafification
q
of F . Via functoriality, for a complex C of presheaves, the sheafifications of the terms can be placed into a new
q
complex. We denote by Cτ the complex with terms Cτn . We note that this is not the same thing as considering a
τ -sheaf valued in a category of complexes and often contains less information. Despite this, it carries quite a bit of
information and was the setting considered in [HJ14] which formed the inspiration for this work.
Topologies. Fix a category C. For two Grothendieck topologies τ and µ on C one calls τ finer than µ provided
every µ-cover is a τ -cover. This may be denoted µ → τ and it induces a morphism of sites Cτ → Cµ . In particular,
for any fixed base S and topology τ on Sch(S) finer than the Zariski topology, one has an associated morphism of
sites ρ : Sch(S)τ → Sch(S)Zar inducing a geometric morphism of topoi (ρ∗ , ρ∗ ) : Sh(S)τ → Sh(S)Zar . These also
have extensions to morphisms of bounded derived categories D+ (Sh(S)τ ) → D+ (Sh(S)Zar ). Thus one may consider
q
q
Rρ∗ (C ) of any complex C ∈ D+ (Sh(S)τ ).
Fix a base scheme S and two topologies τ and µ on Sch(S) with µ → τ . For X ∈ Sch(S), we say that a
τ -sheaf F satisfies µ-descent for X provided that one has an isomorphism F µ (X) ∼
= F τ (X) and that F satisfies
cohomological µ-descent for X provided Hµi (X, F µ ) ∼
= Hτi (X, F ) for all i ≥ 0.
We are primarily concerned with Voevodksy’s h-topology (cf. [Voe96]) in particular the characteristic p case.
An overview geared towards the mixed characteristic setting can also be found in [Bei12]. We review the basic
definition.
Definition 2.1. A morphism in Sch(S) is called a topological epimorphism, if it is surjective and the Zariski
topology of the target is the quotient topology of the Zariski topology of the source. It is called universal if this
property is preserved by any base change. The h-topology is the Grothendieck topology on Sch(S) with coverings
universal topological epimorphisms of finite type.
The h-topology is the coarsest topology finer than the proper and Zariski topology. It is also finer than the
étale topology and in fact it is finer than the eh-topology which is generated by étale covers and so called abstract
blow-ups.
Definition 2.2. Let X ∈ Sch(S), f : X ′ → X a proper map and Z ⊂ X a closed subscheme with preimage E ⊂ X ′ .
∼
We say (X ′ , Z) is an abstract blow-up of X, if f induces an isomorphism X ′ \E −
→ X\Z outside Z. We denote an
f
i
→ X, Z −
→ X) or more simply by (X ′ , Z).
abstract blow-up either by (X ′ −
This means all proper birational maps are eh-covers and so h-covers. Moreover, finite maps are also h-covers. As
such, Beilinson notes that in the presence of resolution of singularities, schemes are h-locally smooth [Bei14, Section
2] and in more general settings de Jong’s alteration theorem gives that schemes over discrete valuations rings are
h-locally regular.
There are several close useful relatives of the h-topology. In particular, to minor extents we will utilize the
Nisnevich topology generated by completely decomposed families of étale morphisms, the cdp or envelope-topology
generated by completely decomposed families of proper morphisms, the rh-topology generated by Zariski, and cdpcovers, and the cdh-topology generated by Nisnevich covers and rh-covers. A much more central topology for our
consideration is the quasi-finite topology, or qfh-topology generated by étale covers and finite covers. Throughout,
it will be helpful to keep in mind the following diagram of how many of these different topologies on Sch(S) relate
(c.f., [GK15]).
Voevodsky describes a particularly nice decomposition of h-covers which will play a critical role in the proof of
pi
Lemma 4.10. An h-cover {Ui −→ X} is said to be of normal form when it admits a factorization pi = s ◦ f ◦ ιi ,
2
Page 3 of 14
fin
cdp
Zar
Nis
ét
rh
cdh
eh
prop
qfh
h
Figure 1. Topologies on Sch from coarser to finer.
ι
f
s
i
→ X is the blow-up along a closed
U } is a Zariski cover, U −
→ XZ is finite surjective and XZ −
→
where {Ui −
subscheme Z of X. In relatively mild cases, all h-covers admit refinements of normal form.
pi
Theorem 2.3 (V. Voevodsky). Let {Ui −→ X} be an h-cover of an excellent, noetherian, reduced scheme. There
exists an h-cover of normal form which is a refinement of {pi }.
Review of differential forms. The work of Huber–Jörder and Huber–Kebekus–Kelly describe descent for differential forms. Fix a field k and the base scheme S = Spec(k). Denote by Ω1 the presheaf of Kähler differential forms
on Sch and for each positive integer q denote by Ωn its n-fold exterior power. Assume k has characteristic 0. For
the relationships cdh → eh → h between the topologies, we have descent Ωneh (X) ∼
= Ωncdh (X) [HJ14, Cor. 2.8] and
n
n
∼
Ωeh (X) = Ωh (X) [HJ14, Thm. 3.6] for any X ∈ Var(k). If X is in addition smooth, one has the descent isomorphisms Ωn (X) ∼
= Ωneh (X) by [Gei06, Thm. 4.7] which depends on resolution of singularities. Thus in characteristic 0
the h-sheafification of Kähler differentials agrees with the usual ones in the smooth case. For a klt-base X, one also
has Ωnh (X) ∼
= (Ωn (X))∗∗ the pushforward of the differentials on the smooth locus [HJ14, Thm. 5.4].
When k is of characteristic p > 0 the situation is more complicated. For example an h-sheafification F h of a
presheaf F can be zero. This happens when F has p-torsion. However some descent statements can be recovered
at least for the cdh-topology via an auxiliary sheaf constructed using the right adjoint of the natural restriction
Sch → Reg denoted ( )dvr . Notably, when X ∈ Reg(S) one has F dvr (X) = F (X) and when k has characteristic 0
one can show Ωndvr ∼
= Ωnh using [HJ14]. This construction works particularly well under the assumption that F is
unramified, see [HKK15, Def. 4.5].
The de Rham–Witt complex. As we will primarily be concerned with questions on the de Rham-Witt complex,
q
we set notations here. Fix a perfect field k of characteristic p > 0. For a scheme X over k, denote by W ΩX the
q
q
p-typical absolute de Rham-Witt complex of X defined by Illusie [Ill79]. Set W ΩX,Q := W ΩX ⊗ Q. It is a complex
of étale sheaves. For a fixed integer n, consider the presheaf X 7→ W ΩnQ (X) of degree n differentials of the de
Rham–Witt complex. We abuse terminology and refer to this as the sheaf of rational Witt differentials. For a
q
topology τ on Sch(k), finer than the étale topology we denote by W ΩQ,τ the complex of τ -sheaves whose n-th term
n
is W ΩQ,τ , which we call τ -Witt differentials.
3. Cohomological descent for de Rham-Witt complexes assuming resolution of singularities
Our primary interest is in the sheaf W ΩnQ , in particular in its descent properties. For smooth schemes one hopes
i
to establish a cohomological descent for the h-topology, i.e., isomorphisms Hhi (X, W ΩnQ,h ) ∼
(X, W ΩnQ,h ) for
= HZar
each n ≥ 0, i ≥ 0 and X ∈ Sm(k). This is rather easy provided one has resolution of singularities of schemes
in positive characteristic and we warm up with a proof of this which is well-known to experts. Huber and Jörder
in [HJ14, Prop. 6.1] give an h-descent for eh-sheaves of Q-vector spaces which directly applies to our context.
Assuming resolution of singularities, an argument of Geisser [Gei06, Thm. 4.2] goes through for W ΩnQ,h giving
an eh-descent for W ΩnQ,h . This suffices to establish the cohomological descent via the usual cohomological étale
descent for de Rham-Witt sheaves. The key point is that factorization of h-covers can be done through smooth
blow-ups. Throughout this section, we work with schemes over a perfect field k of characteristic p > 0. A needed
ingredient is the following well-known lemma.
f
i
→ X, Z −
→ X) be an abstract blow-up of k-schemes, and F an h-sheaf
Lemma 3.1. [Blow-up sequence] Let (X ′ −
on Sch(k). The associated blow-up sequence
· · · → Hhi (X, F ) → Hhi (X ′ , F ) ⊕ Hhi (Z, F ) → Hhi (Z ′ , F ) → · · ·
is exact.
3
Page 4 of 14
Proof. In [Gei06, Prop. 3.2], Geisser proves the analogous statement for the eh-topology in characteristic p. Notably
one does not need to assume resolution of singularities. Since every eh-cover is also an h-cover, it follows that the
statement holds also for the h-topology. Thus, the result follows by replacing h with eh in Geisser’s proof.
Theorem 3.2. Assume that resolution of singularities holds. Let X be smooth k-scheme. Consider the h-sheaf
W ΩnQ,h . For all i ≥ 0 and n ≥ 0
i
Hhi (X, W ΩnQ,h ) ∼
(X, W ΩnQ ).
= HZar
Proof. First we note that as in the proof of [HJ14, Prop. 6.1], which does not depend on the characteristic of the
i
base field, one deduces the equality Hhi (X, W ΩnQ,h ) ∼
(X, W ΩnQ,eh ).
= Heh
i
i
To obtain the equality Heh
(X, W ΩnQ,eh ) ∼
(X, W ΩnQ ) the strategy of [Gei06, Thm. 4.3] can be applied.
= Hét
Let ρ : Sch(k)eh → Sch(k)ét be the canonical morphism of sites, which induces a morphism of étale sheaves
q
q
i
W Ωn Q → Rρ∗ W Ω. Let C be the cone of this map. One shows by contradiction that Hét
(Y, C ) = 0 for any
q
i
Y ∈ Sm(k). Assume that there is 0 6= y ∈ Hét
(Y, C ) and that moreover Y ∈ Sm(k) is of smallest dimension
q
q
∗
admitting such a non-zero element. It is ρ C = 0 by definition of C , which means that there is an eh-cover U q of
q
Y , such that C U q = 0. According to [Gei06, Corollary 2.6] U q has, under resolution of singularities a refinement
{Ui → Y ′ → Y }, where {Ui → Y ′ } is an étale cover and Y ′ → Y is a composition of smooth blow-ups. For a given
blow-up Y ′ → Y along a smooth centre Z, we have by Lemma 3.1 the blow-up sequence
i
i
i
i
(Z ′ , W ΩnQ,eh ) → · · ·
(Y ′ , W ΩnQ,eh ) ⊕ Heh
(Z, W ΩnQ,eh ) → Heh
· · · → Heh
(Y, W ΩnQ,eh ) → Heh
On the other hand, a result by Gros [Gro85, IV. Théorème 1.1.9] gives the analogues sequence for étale topology
i
i
i
i
· · · → Hét
(Y, W ΩnQ ) → Hét
(Y ′ , W ΩnQ ) ⊕ Hét
(Z, W ΩnQ ) → Hét
(Z ′ , W ΩnQ ) → · · ·
The canonical map induces a morphism of long exact sequences
···
i
/ Hét
(Y, W ΩnQ )
τY
···
/ H i (Y, W Ωn )
eh
Q,eh
i
i
/ Hét
(Y ′ , W ΩnQ ) ⊕ Hét
(Z, W ΩnQ )
τY ′ τZ
/ H i (Y ′ , W Ωn ) ⊕ H i (Z, W Ωn )
eh
Q,eh
eh
Q,eh
i
/ Hét
(Z ′ , W ΩnQ )
/ ···
τZ ′
/ H i (Z ′ , W Ωn )
eh
Q,eh
/ ···
q
By minimality of Y the maps τZ and τZ ′ are isomorphisms, from which one concludes that 0 6= y Y ′ ∈ H i (Y ′ , C ).
q
′
In particular, C Y ′ is not acyclic. But then it is also not acyclic on the étale cover {Ui → Y }, a contradiction to
q
ρ∗ C = 0.
Finally, it is well known that in the case at hand étale and Zariski cohomology coincide, which together with the
above equalities leads to the desired statement.
We cannot yet establish (smooth) cohomological h-descent for W ΩnQ without assuming resolution of singularities.
The difficulty lies in the reliance on Geisser’s work. One is tempted to utilize de Jong’s alterations as a replacement
but technical questions still abound. As we have seen in the proof above, Geisser uses the refinement of an ehcover in en étale cover and a composition of smooth blow-ups. To obtain such a refinement strong resolution of
singularities is needed. If one attempts to use alterations here instead finite maps are introduced in the process,
which cannot be controlled. Instead, we utilize ideas form [HKK15] to establish h-descent statements avoiding
resolutions of singularities by showing an h-descent for an auxiliary sheaf W ΩnQ,dvr which agrees with W ΩnQ in the
regular case.
4. dvr-Witt differentials
Recall, for a base scheme S the restriction Sch(S) → Reg(S) has a right adjoint ( )dvr on the category of
presheaves, called the extension functor. This functor may be understood using schemes essentially of finite type in
the sense of [HKK15, Def. 2.5]. Let Sch(S)ess be the category of schemes essentially of finite type, which contains
Sch(S) as a subcategory, and Dvr(S) the category of schemes essentially of finite type which are regular, local, and
of dimension at most 1. A presheaf F on Sch(S) can be extended to a presheaf F ess on Sch(S)ess as explained in
[HKK15, Def./Prop. 4.9]. In many cases of interest, F and F ess coincide on Sch(S)ess . As described in [HKK15],
torsion is an important concept in the present context and will be treated in Subsection 4.3.
4
Page 5 of 14
4.1. The extension functor applied to Witt differentials. We summarize the extension functor ( )dvr and
explore results for the rational de Rham-Witt complex. For more details on the extension functor see [HKK15, Sec.
4.1]. Fix a base scheme S.
Definition 4.1. For a presheaf F on Reg(S), let F dvr be the presheaf on Sch(S) given by
F dvr (X) :=
lim
←−
F (Y )
Y ∈Reg(X)
for X ∈ Sch(S).
Huber, Kebekus and Kelly show that F dvr can be expressed in a particularly useful form. This requires the
notion of an unramified sheaf as introduced by Morel, [HKK15, Def. 4.5].
Definition 4.2. A presheaf F on Sch(S) is unramified provided it is a Zariski sheaf and satisfies the following
three conditions for all X and Y in Reg(S)
(i) the natural morphism F (X ⊔ Y ) → F (X) × F (Y ) is an isomorphism,
(ii) For U → X a dense open immersion, F (X) → F (U ) is injective,
(iii) for every open immersion U → X containing all points of codimension at most 1, F (X) → F (U ) is an
isomorphism.
If X ∈ Sch(S) and F is unramified, then [HKK15, Prop 4.14] guarantees that
F dvr (X) =
lim
←−
F ess (W ).
(4.1)
W ∈Dvr(X)
We set from now on S = Spec k, where k is a perfect field of characteristic p > 0 as we now want to apply this
technology to study descent for the de Rham-Witt sheaves W ΩnQ where n ∈ N.
Theorem 4.3. Both the presheaf W Ωn and the presheaf W ΩnQ are unramified on Reg(k).
Proof. It suffices to show that W Ωn is unramified. We have to verify the three conditions defining unramified
presheaves. As we are working with regular schemes of finite type, we can take advantage of functoriality properties
of the de Rham–Witt complex. Specifically, the structure sheaf O is an unramified presheaf. Note that when
X ∈ Reg(k), global sections of Witt differentials are given by Γ(X, W Ωn ) = W ΩnΓ(X,O) , [Ill79, I.1.10].
It is well-known, that W Ωn`is an étale sheaf, whence a Zariski sheaf. Let X, Y ∈ Reg(k). Since O is unramified,
one has by property (i) O(X Y ) = O(X) × O(Y ). With this, we compute
a
W Ωn (X
Y ) = W ΩnO(X ` Y ) = W ΩnO(X)×O(Y ) = W ΩnO(X) × W ΩnO(Y ) = W Ωn (X) × W Ωn (Y ).
Let U → X be a dense open immersion. The induced map O(X) → O(U ) is injective because O is unramified. It
follows by functoriality of the de Rham–Witt complex, that
W Ωn (X) = W ΩnO(X) → W ΩnO(U) = W Ωn (U )
is also injective. Finally, let U → X be an open immersion which contains all points of codimension at most 1.
The map O(X) → O(U ) is then an isomorphism, and by functoriality the same holds true for the induced map
W Ωn (X) → W Ωn (U ).
We call the presheaf W ΩnQ,dvr obtained by applying the extension functor to the presheaf W ΩnQ of rational Witt
differentials on Sch(k) the dvr-Witt differentials of degree n.
Corollary 4.4. Let X ∈ Sch(k). The dvr-Witt differentials of X can be written as
W ΩnQ,dvr (X) =
lim
←−
W Ωn,ess
Q (Y ).
Y ∈Dvr(X)
Proof. Theorem 4.3 shows that
W ΩnQ
is unramified. The claim then follows from equation (4.1).
4.2. dvr-Witt differentials and the h-topology. The dvr-Witt differentials enjoy may properties that will help
us to show h-descent for Witt differentials. A prime example is a result of Huber, Kebekus, and Kelly showing
useful descent properties.
Theorem 4.5. (c.f., [HKK15, Prop. 4.18] ) Let S be a noetherian scheme. If F is an unramified presheaf on
Sch(S), then F dvr is an rh-sheaf. In particular, if F is an unramified Nisnevich, respectively étale sheaf on Sch(S),
then F dvr is a cdh-sheaf, respectively eh-sheaf.
Thus for a perfect field k of positive characteristic p, the presheaf W ΩnQ,dvr is an eh-sheaf on Sch(k).
5
Page 6 of 14
Remark 4.6. By the universal property of sheafification there are canonical morphisms
W ΩnQ → W ΩnQ,rh → W ΩnQ,cdh → W ΩnQ,eh → W ΩnQ,dvr .
Moreover, if X is regular, W ΩnQ (X) = W ΩnQ,dvr (X) by [HKK15, Rem. 4.3.3] and the above composition is an
isomorphism on X.
Following an approach used in the study of regular group schemes [AHL16, Prop. C.2], we establish preliminary
descent results.
Theorem 4.7. For a perfect field k of characteristic p > 0, the canonical morphism W ΩnQ → W ΩnQ,qfh is an
isomorphism on Nor(k).
Proof. From [HKK15, Lem. 4.4] we know that the extension functor preserves sheaves. When W ΩnQ is a qfh-sheaf
on Nor(k) and therefore on Reg(k), its extension W ΩnQ,dvr is a qfh-sheaf on Sch(k). Hence, it is enough to show that
the sheaf W ΩnQ is a qfh-sheaf on Nor(k), for which we show the canonical morphism W ΩnQ → W ΩnQ,qfh induces an
isomorphism on normal schemes.
To show that W ΩnQ is a qfh-sheaf on Nor(k) we check the sheaf condition for a qfh-cover {Ui → X}i∈I of
normal schemes. Without loss of generality, we may assume that X is connected and therefore integral as it is
noetherian. By a result of Suslin–Voevodsky [SV96, Lem. 10.4] the qfh-cover above has a refinement of the form
{Vj → V → X}j∈J , where V is the normalization of X in a finite normal extension of its function field. In
particular, V → X is finite surjective and {Vj → V } is a Zariski cover.
The sheaf condition for {Vj → X} is implied by the sheaf condition for {Vj → V } and V → X separately. To
see this we argue as in [AHL16] that W ΩnQ Nor is separated for the qfh-topology. This follows from the fact that
for a dominant morphism f : X → Y the induced morphism W ΩnQ (Y ) → W ΩnQ (X) is injective, as O(Y ) → O(X)
is and using a result due to Voevodsky [Voe96, Prop. 3.1.4].
We know already that W ΩnQ satisfies the sheaf condition for a Zariski cover {Vj → V }. Hence it remains to show
the sheaf condition for a finite surjective morphism π : V → X of normal irreducible schemes.
Let Ve be the normalization of the reduction of V ×X V and consider the sequence
0 → W ΩnQ (X) → W ΩnQ (V ) → W ΩnQ (Ve )
where the rightmost map is induced by the difference of the two projections Ve → V . Our objective is to show this
sequence is exact.
We may find a finite surjective morphism π ′ : Y → V with Y normal, such that the composition factors into
a generically purely inseparable morphism followed by a generically Galois morphism π ′ π = πs πi . Since W ΩnQ is
qfh-separated for normal schemes, the sheaf condition for π ′ π follows from a verification of the sheaf condition for
πi and πs , and this on the other hand induces the sheaf condition for π.
Suppose π is generically Galois with Galois group Gal(κ(V )/κ(X)) = Γ. By definition V is the normalization
of X in κ(V ) and the action of Γ extends by functoriality from κ(V ) to V . Furthermore, the canonical map
∼
→ (π∗ OV ) is an isomorphism, which induces an isomorphism of Witt differentials. Hence we have an injection
OX −
W ΩnQ (X) ∼
= W ΩnQ (V )Γ ⊂ W ΩnQ (V ).
To show exactness at the second entry of the sequence, note that W ΩnQ is unramified and so the morphisms
W ΩnQ (V ) ֒→ W ΩnQ (κ(V )) and W ΩnQ (Ve ) ֒→ W ΩnQ (κ(Ve )) are injective. Thus the kernel of W ΩnQ (V ) → W ΩnQ (Ve ) is
contained in the kernel of W ΩnQ (κ(V )) → W ΩnQ (κ(Ve )) = W ΩnQ (κ(V ) ⊗κ(X) κ(V )). By Galois descent an element
x in the kernel of W ΩnQ (V ) → W ΩnQ (Ve ) is then Γ-invariant, i.e. x ∈ W ΩnQ (κ(V ))Γ ∼
= W ΩnQ (X).
Now we look at the case, when π is generically purely inseparable. By Zariski’s connectedness theorem we
know that since X is normal, π is actually purely inseparable, and hence its diagonal map is a surjective closed
∼
immersion. After reduction, we may assume that it is an isomorphism ∆(π) : V red −
→ (V ×X V )red = Ve . The
n e
n
e
map W ΩQ (V ) → W ΩQ (V ) being induced by the difference of the two projections V → V consequently is the zero
map, so that it remains to show that π ∗ : W ΩnQ (X) ֒→ W ΩnQ (V ) is surjective. If π is an isomorphism this is clear.
Assume therefore that π is not an isomorphism. Similar to [AHL16] we reduce this to the case where π is a relative
Frobenius as follows.
6
Page 7 of 14
Denote by X (p) the base change of X along the absolute Frobenius Fk of Spec(k). We have a commutative
diagram
FX
}
Xo
ko
W
Fk
FX/k
X (p) o
X
③③
③③
③
③③
③
}③
k
where FX is the absolute Frobenius of X, FX/k the relative Frobenius of X over k and W the canonical projection.
From the hypothesis that Fk is an automorphism it follows that X ∼
= X (p) and the relative Frobenius FX/k coincides
with the absolute Frobenius FX up to isomorphism. For a p-power q = pr we obtain a similar diagram by iteration.
By a result due to Kollár [Kol97, Prop. 6.6] there is a p-power q and a morphism π ′ : X → V (q) such that the
r
r
relative Frobenius factors FX/k
= π ′ ◦ π : X → X (q) . Thus it suffices to show that (FX/k
)∗ is surjective.
Since X as an irreducible scheme is an Fp -scheme, the Witt vector Frobenius is induced by the absolute Frobenius
∗
FX
= F : W Ωn (X) → W Ωn (X), which commutes with the Verschiebung map. In particular, F V = V F = p.
∗
∗
r
Consequently, after inverting p, F = FX
= FX/k
, as well as (FX/k
)∗ for all q = pr .
Remark 4.8. One notes from the proof that W Ωn is a qfh-sheaf after inverting only the Witt vector Frobenius.
Corollary 4.9. The extension W ΩnQ,dvr is a qfh-sheaf on Sch(k).
Proof. According to [HKK15, Lem. 4.4] the extension functor preserves sheaves. Hence the fact that W ΩnQ is a
qfh-sheaf on Reg(k) because it is one on Nor(k) implies that W ΩnQ,dvr is a qfh-sheaf on Sch(k).
We will utilize the following lemma, which is well-known to experts.
pi
Lemma 4.10. Let X ∈ Sch(k). Every h-cover {Ui −→ X}i has a refinement of the form
{Vi → V → X ′ → X}i
where {Vi → V }i is a Zariski cover, V → X ′ is a modification and X ′ → X is a finite morphism.
Proof. By de Jong’s alteration theorem [dJ96] choose a regular alteration Y → X and consider the pullback
{Ui′ → Y } of the original covering to Y . Let {Vi → V → YZ → Y }i be a refinement of normal form of {Ui′ → Y }
which exists by Voevodsky’s Theorem 2.3, meaning that{Vi → V } is a Zariski cover, V → YZ is finite surjective
and YZ → Y is the blow-up along a closed subscheme Z of Y . We obtain a refinement of the original cover
{Vi → V → YZ → Y → X}i .
The composition V → YZ → Y → X is an alteration, therefore, it factors using Stein factorization V → X ′ → X,
where V → X ′ is proper with geometrically connected fibers and X ′ → X is finite. This gives the desired
refinement.
We close this section with a descent result which we hope, it will be of independent interest.
Theorem 4.11. Fix a base scheme S and let F be an unramified étale sheaf on Reg(S). When F is also a
qfh-sheaf, then F dvr satisfies h-descent on Sch(S).
∼ F dvr . It suffices to check that F dvr agrees with its
∼ F Q,dvr and (F dvr )qfh =
Proof. By hypothesis (F dvr )eh =
h-sheafification which may be checked via Čech cohomology. That is for any X ∈ Sch(k), the natural map
F dvr (X) → lim Ȟ 0 (U, F dvr )
U
where U runs through all h-covers U → X, is an isomorphism.
Any h-cover U → X may be refined as {Wi → X ′ → X}i∈I where the first is an eh-cover and the latter is a
qfh-cover by Lemma 4.10. As F dvr is both a qfh-sheaf and an eh-sheaf, the sheaf condition holds in the intermediate
steps. More precisely, this means, that the sequences
0 → F dvr (X) → F dvr (X ′ ) → F dvr (X ′ ×X X ′ )
and
0 → F dvr (X ′ ) →
Y
i∈I
F dvr (Wi ) →
7
Y
i,j∈I
F dvr (Wi ×X ′ Wj )
(4.2)
(4.3)
Page 8 of 14
are exact. These exact sequences clearly give injections
F dvr (X) ֒→ F dvr (X ′ ) ֒→
Y
F dvr (Wi )
i∈I
as well as
F dvr (X ′ ×X X ′ ) ֒→
Y
F dvr (Wi ×X Wj ).
i,j∈I
We obtain a commutative diagram
/ F dvr (X)
0
/ F dvr (X ′ )
/
F dvr (X ′ ×X X ′ )
Q
i∈I
F dvr (Wi )
/Q
F
(W
dvr
i ×X Wj )
i,j∈I
Q
F dvr (Wi ×X ′ Wj )
i,j∈I
With a diagram chase, keeping in mind the exact sequences (4.2) and (4.3), we obtain the exact sequence
Y
Y
0 → F dvr (X) →
F dvr (Wi ) →
F dvr (Wi ×X Wj ),
i∈I
which is exactly the sheaf property for the cover {Wi → X ′ → X}i∈I . This shows the claim.
Corollary 4.12. The sheaves
W ΩnQ,dvr ,
(4.4)
i,j∈I
n > 0, on Sch(k) satisfy h-descent.
Proof. We have already seen that W ΩnQ is an unramified étale sheaf by Theorem 4.3 and by Theorem 4.9 a qfh-sheaf
as well. Thus Theorem 4.11 applies.
4.3. Topological Torsion. For a presheaf F on Sch(S), we denote as in [HKK15] tor F (X) for the sections of
F (X) which vanish on a dense open subscheme and we call a presheaf F torsion free when tor F (X) = 0. After a
general observation we look at the torsion forms for W ΩnQ over a perfect field of positive characteristic.
Lemma 4.13. An unramified presheaf F has no torsion forms on Reg(S).
Proof. Let F be an unramified sheaf on Reg(S) and X ∈ Reg(S). Suppose ω ∈ F is torsion. This means that
there is an open dense subscheme U ֒→ X such that ω U = 0. However, the second characterizing property for
unramified presheaves gives that the induced morphism F (X) → F (U ) is injective. Thus ω = 0.
This means in particular that the sheaves W Ωn and W ΩnQ have no torsion forms on Reg(k).
Theorem 4.14. Let F be an unramified presheaf and assume in addition that F dvr is an h-sheaf. Then F dvr is
torsion free.
Proof. Let X ∈ Sch(k) and ω ∈ F dvr be a torsion form, meaning that there is an open dense subscheme U ֒→ X
such that ω U = 0. We aim to show that ω = 0. By definition ω is given by a compatible system (ωY )Y , where
Y ∈ Reg(X). It suffices to show ωY = 0 for all such Y . For a morphism f : Y → X, Y ∈ Reg(X), let YU be the
preimage of U under f . The hypothesis ω U = 0 implies that ωY U := ωY Y = 0.
U
If the preimage of U in Y under f is dense, it follows as in Lemma 4.13 that ωY = 0 because Y is regular.
e → X, i.e. X
e ∈ Reg(X). The preimage X
eU of U under g is again dense
Otherwise, choose a regular alteration g : X
e
e be the pullback of Z
in X. The image of Y in X is contained in a subscheme Z where U ∩ Z is not dense. Let Z
e
along g. Furthermore, let Y be the pullback of Y along g, which is again regular. In particular, Ye → Y is as an
alteration and thus an h-cover. We obtain a commutative diagram
Ye
/Y
e
Z
/Z
e
X
/ X.
8
Page 9 of 14
e = F dvr (X)
e constitutes an entry in the system (ωY )Y . Set ω e the image of ωY
By definition, ωXe := ω Xe ∈ F (X)
Y
e
under the map F (Y ) → F (Y ). We have the following commutative diagram where we have indicated where the
sections ω, ωY , ωXe , and ωYe
ωYe ∈ F dvr (Ye ) = F (Ye ) o
O
F dvr (Y ) = F (Y ) ∋ ωY
O
e = F (X)
e o
ωXe ∈ F dvr (X)
F dvr (X) ∋ ω
e
eU is dense in X
e and ω e
Considering that g −1 (U ) = X
e = 0 because X is regular. By
X U = 0, we know that ωX
e
e
compatibility the image ωYe of ωXe under the morphism F (X) → F (Y ) is also zero. We conclude that ωY being in
the preimage of ωYe under the map F (Y ) → F (Ye ) is 0 as well, since F ndvr is by assumption an h-sheaf and Ye → Y
an h-cover.
Corollary 4.15. The sheaf W ΩnQ,dvr is torsion free on Sch(k).
Proof. As W ΩnQ,dvr is an h-sheaf Theorem 4.14 applies.
5. Properties of h-Witt differentials
The critical property about the sheaves Ωn driving the many useful statements in [HJ14] is that Ωneh satisfies
h-descent. Working in characteristic 0, one automatically gets h-descent for Ωn on smooth schemes following work
of Geisser not available in our setting. Instead, we obtain h-descent for regular schemes from the fact that W ΩnQ,dvr
is an h-sheaf.
5.1. Regular descent. As described in [HKK15, Sec. 5], to understand Ωncdh it is critical to understand the
torsion of Ωn . In our case, something similar is true: it is important to understand the torsion of W ΩnQ in order to
understand W ΩnQ,h .
Lemma 5.1. For a base scheme S, let F be a sheaf on Sch(S) which is torsion free on Reg(S). For X ∈ Sch(S)
e → X such that α∗ ω vanishes in F (X).
e
and ω ∈ tor F (X) there is an alteration α : X
e
Moreover, if for each point x ∈ X, ω x = 0, there is an h-morphism α : X → X such that α∗ ω = 0.
e → X is a
Proof. As ω is torsion, there is a Zariski open subset U ⊂ X such that the restriction ω U = 0. If α : X
e vanishes on α−1 U , which is dense open in X.
e But since X
e is regular, F (X)
e
regular alteration, then α∗ ω ∈ F (X)
∗
is torsion free by assumption and thus α ω = 0.
To prove the second claim, we may without loss of generality assume that X is integral by considering each
irreducible component of X separately and using the fact that each scheme is eh-locally and therefore h-locally
reduced. In this case, the assumption that ω x = 0 for all points x of X, means that ω is torsion on X and we can
e → X with α∗ ω = 0.
now apply the first part to obtain an h-cover α : X
Definition 5.2. Let τ be a Grothendieck topology on Sch(S). For a presheaf F on Reg(S), we say that it satisfies
τ -descent for regular schemes, if F τ (X) = F (X) for X ∈ Reg(k). Equivalently we say that F satisfies regular
τ -descent.
Example 5.3. When k is a field of characteristic 0, F = Ωn satisfies regular h-descent by [HJ14, Thm. 3.6]. Over
any field k, any sheaf F for which F dvr which satisfies τ -descent automatically satisfies regular τ -descent.
We are ready to give the main descent theorem for the Witt differentials.
Theorem 5.4. Let k be a perfect field of positive characteristic p. The sheaves W ΩnQ , n > 0, on Sch(k) satisfy
regular h-descent. In particular, when X is regular, we have isomorphisms W ΩnQ,h (X) ∼
= W ΩnQ (X).
Proof. Since W ΩnQ,dvr is an h-sheaf, we have by the universal property of sheafification a sequence of canonical
morphisms
W ΩnQ → W ΩnQ,h → W ΩnQ,dvr .
(5.1)
The first step is to show, that the second map is a monomorphism. This follows a similar argument to [HKK15,
Prop. 5.12]. Let X ∈ Sch(k) and let ω be a section in the kernel of W ΩnQ,h (X) → W ΩnQ,dvr (X). Choose an h-cover
f : X ′ → X, such that f ∗ ω is in the image of W ΩnQ (X ′ ) → W ΩnQ,h (X ′ ). This way, we obtain an element ω ′ in the
kernel of W ΩnQ (X ′ ) → W ΩnQ,dvr (X ′ ), and have to show that it vanishes on some h-cover of X ′ . By Lemma 5.1, it is
9
Page 10 of 14
enough to show that ω ′ is trivial on every point x ∈ X ′ . For such a point, let {x} be its closure in X ′ with reduced
scheme structure, and let V be its regular locus. In particular, V ∈ Reg(X ′ ) and its generic point is isomorphic to
x. Since ω ′ is trivial in W ΩnQ,dvr (X ′ ) it vanishes by definition of the extension functor on every regular scheme over
X ′ , specifically on V as well on its generic point and therefore on x.
For the second step, let X ∈ Reg(k). Applying the global section functor with X to the composition of maps (5.1)
we obtain a composition of natural maps W ΩnQ (X) → W ΩnQ,h (X) ֒→ W ΩnQ,dvr (X), where the second morphism
is injective by what we have just seen. By Remark 4.6, W ΩnQ,dvr (X) = W ΩnQ (X) and the above composition is
actually an isomorphism, which induces that the second map is surjective, and therefore also an isomorphism, and
it follows right away that the same is true for the first map.
Remark 5.5. In the case when n = 0 there are stronger h-descent results known. Specifically, the sheaf W OQ is
an h-sheaf on Sch(k) by [BBE07, Thm. 2.4]. In fact, one even has cohomological descent for W O for any finitely
presented Fp -scheme after just inverting the Witt vector Frobenius [BS16, Thm. 5.40]. This is unlike the situation
for Ω0 = O in characteristic 0, where one has such a result only for semi-normal schemes [HJ14, Prop. 4.5].
It is well known to experts that h-sheaves can be understood as projective limits. The intuition is that sheaves
which satisfy regular h-descent are h-locally smooth. Here a detailed proof of this fact is given and applied to Witt
differentials.
Lemma 5.6. Let F be a sheaf on the Zariski site of Sch(S) which satisfies regular τ -descent. Suppose X ∈ Sch(S).
Assume there is a regular τ -cover X ′ → X and a regular τ -cover X ′′ → X ′ ×X X ′ . The sequence
0 → F τ (X) → F (X ′ ) → F (X ′′ )
is exact.
Proof. From the sheaf property, we have an exact sequence
0 → F τ (X) → F (X ′ ) → F (X ′ ×X X ′ )
for any τ -cover X ′ → X. For a τ -cover X ′′ → X ′ ×X X ′ , the induced map F τ (X ′ ×X X ′ ) ֒→ F (X ′′ ) is again
injective by the sheaf condition. Together, this gives an exact sequence 0 → F τ (X) → F τ (X ′ ) → F τ (X ′′ ). In the
case when X ′ and X ′′ are regular, by regular τ -descent F τ (X ′ ) = F (X ′ ) and F τ (X ′′ ) = F (X ′′ ) thus the exact
sequence becomes 0 → F τ (X) → F (X ′ ) → F (X ′′ ).
One can now give the analogue of [HJ14, Thm. 1] in suitable generality.
Theorem 5.7. Fix a base scheme S. For X ∈ Sch(S) and an unramified Zariski sheaf F on Reg(S) satisfying
regular h-descent, we have an isomorphism
Y ′❆
❆❆ ′
Y
❆❆f
∗
ψ
∼
❆
′
F h (X) = lim F (Y ) = (αf ) ∈
F (Y )
∀ψ :
⇒
ψ
α
=
α
❆
f
f
❆
←−
Y ∈Reg(X)
f : Y →X
/X
Y
Y regular
f
Proof. The proof follows the same steps as [HJ14, Cor. 3.9]. There is a natural injection F h (X) ֒→ limY ∈Reg(X) F (Y )
←−
as a section β ∈ F h (X) determines uniquely a compatible family βf := f ∗ β ∈ F h (Y ) with f : Y → X and Y
regular. By regular h-descent, it follows βf ∈ F h (Y ) = F (Y ). The theorem follows by demonstrating an injection
limY ∈Reg(X) F (Y ) ֒→ F h (X).
←−
We choose smooth h-covers g : X ′ → X and h : X ′′ → X ′ ×X X ′ , denote by i : X ′ ×X X ′ → X the canonical
map and by pr1,2 X ′ ×X X ′ → X ′ the projections, fitting in the diagram
X ′′ ❏
❏❏❏
❏❏h❏
❏❏❏
%
pr
′ 2
/ X′
X ′ ×X X
❏❏
❏❏ i
❏❏
pr1
g
❏❏
❏❏
$
g
/ X.
X′
For a compatible family (αf )f ∈ limY ∈Reg(X) F (Y ) we thus have
←−
(pr1 ◦h)∗ αg = αi◦h = (pr2 ◦h)∗ αg .
10
Page 11 of 14
Consequently β ∈ F h (X) exists such that αg = g ∗ β. This gives us a morphism limY ∈Reg(X) F (Y ) → F h (X). It
←−
remains to check that this morphism is injective.
To check this, it suffices to show that for any morphism f : Y → X with Y regular f ∗ (β) = αf holds. Let
p : Y ′ → Y ×X X ′ be a smooth h-cover, denote j : Y ×X X ′ → X the canonical map and prY : Y ×X X ′ → Y
respectively prX ′ : Y ×X X ′ → X ′ the projections, which again fit in a diagram
Y′ ■
■■
■■p
■■
■■
$
pr
′ X′
/ X′
Y ×X X
■■
■■ j
■■
prY
■■ g
■$
f
/ X.
Y
We therefore have equalities
(prY ◦p)∗ αf = αj◦p = (prX ′ ◦p)∗ αg = (prX ′ ◦p)∗ g ∗ β = (prY ◦p)∗ f ∗ β.
As F is unramified, the map (prY ◦p)∗ : F (Y ) → F (Y ) is injective, and we can conclude that αf = f ∗ β. This
shows that the map limY ∈Reg(X) F (Y ) → F h (X) is in fact injective and the theorem follows.
←−
Applying this general framework to Witt differentials results in the following corollaries. In particular, we will see
that dvr-Witt differentials and h-Witt differentials coincide on Sch(k) for a perfect field k of positive characteristic
p.
Corollary 5.8. Given an object X in Sch(k), and regular h-covers X ′ → X and X ′′ → X ′ ×X X ′ , the following
sequence is exact
(5.2)
0 → W ΩnQ,h (X) → W ΩnQ (X ′ ) → W ΩnQ (X ′′ ).
Proof. It is enough to apply Lemma 5.6 with F = W ΩnQ . This lemma applies as Theorem 5.4 gives the required
regular h-descent.
Corollary 5.9. For X ∈ Sch(k) and each n > 0,
W ΩnQ,h (X) ∼
=
lim
←−
W ΩnQ (Y ).
Y ∈Reg(X)
Proof. For fixed n > 0 the sheaf W ΩnQ is unramified by Theorem 4.3 and satisfies regular h-descent by Theorem 5.4.
Now apply Theorem 5.7 with F = W ΩnQ .
Corollary 5.10. For a Zariski sheaf F on Reg(S), which satisfies regular h-descent one has
F dvr (X) = lim F (Y ) ∼
= F h (X).
←−
Y ∈Reg(X)
In particular, for X ∈ Sch(k) and n > 0 there is an isomorphism
W ΩnQ,h (X) ∼
= W ΩnQ,dvr (X).
Thus one may identify the sheaves W ΩnQ,h and W ΩnQ,dvr on Sch(k).
Proof. The first equality follows from the definition of the extension functor and the second isomorphism by Theorem 5.7. The other claims follow as W ΩnQ satisfies regular h descent by Theorem 5.4.
5.2. Fundamental properties and some examples. After now establishing regular h-descent, we investigate
some properties of h-Witt differentials in the spirit of [HJ14, Prop. 4.2]. We note that we cannot show a direct
h-descent for eh-sheaves, thus our techniques differ mildly. Throughout this section, fix a perfect k of characteristic
p > 0. For an h-sheaf F and object X in Sch(k), we utilize the notation F X for the push-forward of F to the
Zariski-site of X.
[n]
Theorem 5.11. Let X ∈ Sch(k). Set W ΩQ,X := j∗ W ΩnQ,X reg where j : X reg → X is the inclusion of the regular
locus. The sheaf W ΩnQ,h |X satisfies the following:
(i) The sheaf W Ωnh,Q |X is a quasi-coherent torsion free sheaf of W OQ,X -modules,
(ii) If X is regular, then
W ΩnQ,h X = W ΩnQ,X .
11
Page 12 of 14
(iii) Let r : Xred → X be the reduction. Then
r∗ W ΩnQ,h
Xred
= W ΩnQ,h
X
.
(iv) If X is reduced, there is a natural inclusion
W ΩnQ,X /torsion ⊂ W ΩnQ,h
X
.
(v) Let X be normal and j : X reg → X the inclusion of the regular locus. There is a natural inclusion
[n]
W ΩnQ,h ⊂ W ΩQ,X .
(vi) For n > dim(X), W ΩnQ,h
X
= 0.
Proof. As before we choose smooth h-covers g : X ′ → X and h : X ′′ → X ′ ×X X ′ , denote i : X ′ ×X X ′ → X the
canonical map and e
h = i ◦ h. Without loss of generality, we assume that g and e
h are proper. Hence they are quasicompact and quasi-separated. It follows that the direct image sheaves g∗ W ΩnQ and e
h∗ W ΩnQ are quasi-coherent,
because this is the case for W ΩnQ . According to the exact sequence (5.2) W Ωnh,Q |X is the kernel of a morphism
g ∗ W Ωn → e
h∗ W Ωn of quasi-coherent sheaves, thus in itself quasi-coherent. Torsion freeness for W Ωn
= W Ωn
Q
Q
Q,dvr
Q,h
was already shown in Section 4.3. This completes (i).
Item (ii) follows immediately from regular h-descent in Theorem 5.4. Item (iii) is true as the inclusion of the
reduced subscheme is an h-cover.
For (iv) consider the natural map W ΩnQ,X → W ΩnQ,h X . For reason that W ΩnQ,h X is torsion free by (i), we may
quotient out torsion in the domain to get W ΩnQ,X /torsion → W ΩnQ,h X . Let j : X reg ֒→ X be the regular locus.
By the definition of torsion freeness we have W ΩnQ,h X ⊂ j∗ W ΩnQ,h X reg = j∗ W ΩnQ,X reg , where the last equality
holds because of (ii). Note that j∗ W ΩnQ,X reg is torsion free and that it agrees with W ΩnQ,X /torsion on X reg . Hence
W ΩnQ,X /torsion ⊂ j∗ W ΩnQ,X reg and the diagram
/ j∗ W Ωn reg
W ΩnQ,h X
Q,X
O
6
♥♥♥
♥
♥
♥♥♥
) ♥♥♥♥
W ΩnQ,X /torsion
commutes, so that the vertical map is injective as well.
Now we consider X to be normal and let again j : X reg ֒→ X be its regular locus. Because W ΩnQ,h X is torsion
free, the restriction W ΩnQ,h (X) ֒→ W ΩnQ,h (X reg ) = W ΩnQ (X reg ) is an injection. By locality this is equally true for
all open subschemes of X, so that we conclude (v).
For (vi), assume without loss of generality that X is reduced, and by torsion freeness restrict to an open subset
where it is smooth. The vanishing of W ΩnQ,h X for n > dim(X) then follows from the vanishing of W ΩnQ for smooth
schemes.
We conclude by addressing expected results in mildly singular cases. We first record some facts, that will serve as
useful tools. Recall that an abstract blow-up of a scheme X is the data (X ′ , Z) where X ′ is a scheme, f : X ′ → X
∼
is a proper map, Z ⊂ X a closed subscheme with preimage E ⊂ X ′ , and f induces an isomorphism X ′ \E −
→ X\Z
outside Z. The eh-topology is generated by abstract blow-ups and étale coverings. As such, any abstract blow up
is an h-cover as well and h-sheaves in some sense commute with these.
Lemma 5.12. Let (X ′ , Z) be an abstract blow-up of X ∈ Sch(k). For an h-sheaf F the induced square
F (X)
/ F (X ′ )
F (Z)
/ F (E)
is Cartesian.
Proof. The proof does not differ from the one in characteristic 0 given in [HJ14, Prop. 3.3]. For a more detailed
description see also the first part of the proof of [BS16, Prop. 2.8], where a more general situation is discussed.
Thus one has an exact sequence
0 → F (X) → F (X ′ ) ⊕ F (Z) → F (E).
12
(5.3)
Page 13 of 14
Lemma 5.13. Let X be normal and irreducible, L/K(X) a Galois extension of the function field of X with Galois
groups G and Y a normalization of X in L. Any étale sheaf F with eh-descent has h-descent for f : Y → X. More
precisely,
∼ F h (Y )G .
F h (X) =
Proof. This is well-known to experts and can be proven with essentially the same argument as in [HJ14, Lem. 2.10].
In the case when X is smooth and f : Y → X étale, the statement holds because F is an étale sheaf and thus
satisfies Galois descent. The general case follows by restricting to a smooth open subscheme of X where f is étale
and perform the same diagram chase as in loc. cit..
We next recover via Theorem 5.11 results analogous to [HJ14, Prop. 4.9] and [HJ14, Prop. 4.10] concerning
normal crossing varieties and group quotients.
Theorem 5.14. For X ∈ Sch(k) with normal crossings,
W ΩnQ,h
X
= W ΩnQ,X /torsion.
Proof. The proof is a direct translation of [HJ14, Prop. 4.9] with their [HJ14, Prop. 4.2] replaced by Proposition 5.11.
We repeat the argument for convenience of the reader.
By Proposition 5.11(iv) there is an inclusion
W ΩnQ,X /torsion ⊂ W ΩnQ,h
X
because X is reduced. Working étale locally, one may assume that X is a union of smooth hyperplane sections and
use induction with respect to the number of irreducible components c(X) of X.
If c(X) 6 1 then we are in the smooth case, and the statement is just Proposition 5.11(ii). Thus let c(X) > 1.
Choose an irreducible component Z ⊂ X, let X ′ = X\Z and E the inverse image of Z under the map X ′ → X.
Then (X ′ , Z) is an abstract blow-up of X and by Lemma 5.12 there is an exact sequence analogous to (5.3)
0 → W ΩnQ,h
X
→ W ΩnQ,h
X′
⊕ W ΩnQ,h
Z
→ W ΩnQ,h
E
.
As in the classical case, we see that the pull-back of torsion free Witt differentials is again torsion free. Hence,
the above sequence fits into a commutative diagram
0
/ W Ωn
Q,h
O
/ W Ωn
Q,h
X
?
W ΩnQ,X /torsion
/ W Ωn
Q,X ′
X′
⊕ W ΩnQ,h
Z
/torsion ⊕ W ΩnQ,Z /torsion
/ W Ωn
Q,h
O
E
?
/ W Ωn /torsion
Q,E
where the first and last vertical map are inclusions by (iv) of Proposition 5.11 and the middle vertical map is an
isomorphism by (ii).
By local calculations the second line in the diagram is also exact. Thus a diagram chase shows that the first
vertical map is an isomorphism as well.
Next we look at quotient singularities, meaning singular varieties that can be expressed as quotients of a smooth
variety by a finite group action. Although the associated projection is not a modification it is of course an alteration.
Corollary 5.15. Let X ∈ Reg(k) be quasi-projective with an operation of a finite group G. One has
W ΩnQ,h
X/G
∼
= (W ΩnQ,X )G .
Proof. The projection X → X/G is ramified with Galois group G. Hence, the equivalence follows immediately from
Galois descent in Lemma 5.13 and the fact that W ΩnQ is an étale sheaf with eh-descent.
Remark 5.16. Considering that X/G is normal there is according to Proposition 5.11(v) an inclusion
W ΩnQ,h
[n]
X/G
⊂ W ΩQ,X/G .
In analogy to the classical case, one expects this to be an isomorphism.
13
Page 14 of 14
References
[AHL16] Giuseppe Ancona, Annette Huber and Simon Pepin Lehalleur: On the relative motive of a commutative group scheme.
Algebraic Geometry 3 (2), (2016).
[Bei12] Alexander Beilinson: p-adic periods and derived de Rham cohomology. J. of the AMS 25, no. 3, 715–732, (2012).
[Bei14] Alexander Beilinson: On the crystalline period map. Cambridge J. Mat. 1, no. 1, 1–51, (2014).
[BBE07] Pierre Berhtelot, Spencer Bloch and Hélène Esnault: On Witt vector cohomology for singular varieties. Compos.
Math., 143(2), 363–392, (2007).
[BS16] Bhargav Bhatt and Peter Scholze: Projectivity of the Witt vector affine Grassmannian. Invent. Math., to appear,
arXiv:1507.06490v2, (2016).
[dJ96] Aise Johan de Jong: Smoothness, semi-stability and alterations. Publ. Math. IHÉS 83: 51–93, (1996).
[GK15] Ofer Gabber and Shane Kelly: Points in algebraic geometry J. Pure Appl. Algebra, 219, 10, 4667–4680, (2015).
[Gei06] Thomas Geisser: Arithmetic cohomology over finite field and special values of ζ-functions. Duke Math. J., 133(1): 27–57, 05,
(2006).
[Gro85] Michel Gros: Classes de Chern et classes de cycles en cohomologie de de Rham–Witt logarithmique. Mém. Soc. Math. France
2e série, 21: 1–87, (1985).
[HJ14] Annette Huber and Clemens Jörder : Differential forms in the h-topology, Algebr. Geom. 1, no. 4: 449–478, (2014).
[HKK15] Annette Huber, Stefan Kebekus and Shane Kelly: Differential forms in positive characteristic avoiding resolution of
sngularities. to appear (2017).
[Ill79] Luc Illusie: Complexe de de Rham–Witt et cohomologie cristalline. Ann. Sci. de l’E.N.S. 4e série, tome 12, no 4: 501–661,
(1979).
[Kol97] János Kollár: Quotient spaces modulo algebraic groups. Ann. of math. (2), 145(1): 33–79, (1997).
[SP] The Stacks Project Authors: Stacks Project.
[SV96] Andrei Suslin and Vladimir Voevodsky: Singular homology of abstractalgebraic varieties. Invent. Math., 123(1): 61–94,
(1996).
[Voe96] Vladimir Voevodsky: Homology of schemes. Selecta Math. (N.S.), 2(1): 111–153, (1996).
14
| 0math.AC |
A LEVEL TRAINING SET WITH BOTH A COMPUTERBASED CONTROL AND A COMPACT CONTROLLER
Hayati Mamur1, Ismail Atacak2, Fatih Korkmaz3 and M. R. Amin Bhuiyan1
1
Department of Electrical and Electronics Engineering, Faculty of Engineering,
Manisa Celal Bayar University, 45100, Manisa, Turkey
hayati.mamur@cbu.edu.tr
2
3
Department of Computer Engineering, Faculty of Technology,
Gazi University, 06100, Teknikokullar, Ankara, Turkey
Department of Electrical and Electronics Engineering, Faculty of Engineering,
Cankiri Karatekin University, 18100, Cankiri, Turkey
ABSTRACT
In engineering education, the combination with theoretical education and practical education is an
essential problem. The taught knowledge can be quickly forgotten without an experimental application. In
addition, the theoretical knowledge’s cannot be easily associated applications by students when they start
working in industry. To eliminate these problems, a number of education tools have been developed in
engineering education. This article presents a modelling, simulation and practice study of a newly
designed liquid level training set developed for the control engineering students to simulate, examine and
analyze theoretically and experimentally the controllers widely used in the control of many industrial
processes. The newly designed training set combines two control structures, which are a computer-based
control and a digital signal processing-based control. The set displays the results related to experiments
in real time as well as. These features have made it a suitable laboratory component on which the
students can both simulate and test the performance of liquid level control systems by using theoretical
different control structures.
KEYWORDS
Computer aided engineering, Computer-based control, DSP-based control, Engineering education,
Process control
1. INTRODUCTION
Today, control systems with a wide usage area ranging from mathematics to processes have
become more important with increased industrialization [1], [2]. In industrial processes, many
physical variables such as pressure, temperature, level and flow need to be automatically
controlled in order to keep their values at a desired level [3], [4]. In this context, automatic
control systems are engineering courses that are being taught in different departments of
faculties to teach the basic concepts and theorems of control systems [5].
In general, automatic control systems are grouped in two categories: open loop control system
and closed loop system [6], [7]. The most important element of the control system is a
controller. Personal computers (PCs) connected to data acquisition (DAQ) cards [8]–[11],
compact proportional-integral-derivative (PID) control devices [12], programmable logic
controllers (PLCs) [13], [14], microcontrollers (μCs) [6], [15], [16] and digital signal processors
(DSPs) [17] are the most commonly used controllers in practically all industrial control
applications. These controllers interconnect all parts of the physical and non-physical of a
system.
Setting of controller parameters, according to a process, has a vital importance in terms of all
system performance [18, 19]. Various techniques are employed for setting of these parameters
[20]–[22]. Each process needs to be specifically set the controller parameters, but this may not
always provide the desired performance on process [23]. Therefore, the setting of controller
parameters and the understanding how the parameters have an effect on a process are important
for engineering students.
Practicability of control methods and parameter settings through some simulation software and
their remote controllability provides convenience to engineering students in the learning of
automatic control systems [9], [24]. Moreover, the simulation software’s eliminate the need of
an experiment set required for engineering courses [25]. However, this type teaching and
learning makes difficult the adaptation of engineering students to actual systems. The
designating and executing of control methods, controllers and parameters and the observing of
the effects to the performance on an actual system in laboratory environment ensure to be more
permanent of the learned knowledge for engineering students [26].
For this purpose, Vahidi et al. [25] conducted a study on power transformers by using
MATLAB/Simulink environment for incoming engineers in industry and graduate students.
Tekin et al. [27] established an internet-based laboratory environment and then performed an
induction motor application controlled by a DSP. Stefanovic et al. [11] setup a distance learning
laboratory based on LabVIEW for control engineering education. Khairurrijal, et al. [16]
developed a temperature control learning set with proportional controller for the education of
undergraduate students. They received highly positive feedbacks from the students in the survey
conducted on the temperature control learning set. Kayisli et al. [7] developed an education tool
with a proportional-integral controller by using MATLAB/Simulink environment for DC-DC
converter circuits. Ekici [10] conducted a study on the computer-assisted training method for
power analysis courses. Dandil [17] realized a simulation and application study on an induction
motor control for undergraduate students. A study conducted by Shyr [26] revealed that the
project-based mechatronics learning is more effective in the learning of undergraduate students.
In this study, a liquid level control training set was designed and implemented for control
engineering courses. Thanks to the training set, students can both examine the differences
between the computer-based control structure with a visual interface and the DSP-based control
structure and observe the effects on process performance of the changes in the controller
parameters. Thus, two control devices were combined by the training set. Also, both theoretical
education and practical education were applied on the training set.
2. OBTAINING OF MATHEMATICAL MODEL OF THE TRAINING SET
Before proceeding to the implementation stage, the knowing of how to exhibit a dynamic
behaviour of liquid level control systems is vital in terms of the determining of a number of
parameters related to the system in the design stage. The dynamic behaviour of a system can be
observed and analyzed through some simulation studies after the obtaining of the mathematical
model to the system [28]. As in many physical systems, liquid level control systems also exhibit
a non-linear dynamic behaviour due to the inherent characteristics [29]. Nevertheless, the
theories and theorems in control systems, the most of which can be only applied to linear
systems, necessitates to be modelled nonlinear systems as a linear system. When considering
small changes in the level control system, the system can be modelled with linear differential
equations [30]–[32]. The schematic diagram of the first order liquid level control system
controlled is shown in Fig. 1.
Figure 1. A schematic diagram of the liquid level control system
This system consists of a tank with an inlet (control) valve and outlet (load) vane and represents
a single input single output (SISO) control system [5]. In the system, while the outflow liquid
from the tank is manually controlled through the load vane, the inflow liquid into the tank is
adjusted by a proportional valve. Normally, the outflow liquid from the tank is a load which is
needed by process and continuously changes due to reasons beyond control. Therefore, the
inflow liquid into the tank represents a manipulated variable (MV) depending on the liquid
level. The outflow liquid from the tank refers to a load or a disturbance [32]. The top of the
liquid tank is open and it has a cylindrical structure. The dimensions and some calculations of
the liquid level control tank are given in Table 1.
Table 1. The dimensions of the tank.
Properties
Values
The height of the tank, h (m)
The diameter of the tank, d (m)
The cross sectional area of the tank, A (m2)
The capacitance of the tank, C (m3/m)
The volume of the tank, V (m3)
The maximum liquid flow, Qo (l/s)
The maximum liquid flow, Qo (m3/s)
Resistance, R (s/m2)
1
0.15
0.5063
0.5063
0.0176
0.5
0.0005
2000
The liquid level control system has been modeled taking into account the change in the liquid
level, which results from the difference between the inlet flow rate and outlet flow rate of the
liquid in the tank. This system can be considered as a simple circuit including a capacity (C)
and a resistance (R). By supposing that all initial conditions of the system are equal to zero, the
obtained Laplace transform is given by:
RCsH (s) H (s) RQi (s) .
(1)
If it is considered that height (h) and inflow (qi) are the output and input of the system,
respectively, the transfer function is expressed as follows:
H (s)
R
,
Qi ( s ) sRC 1
(2)
where, RC (τ) is the time constant of the system (s). If outflow (q0) is taken as the output and
the input is the same (qi), then the transfer function of the system is defined by:
Qo ( s )
1
.
Qi ( s ) RCs 1
(3)
The characteristic equation of the obtained transfer function, simply the denominator of this
function, is first order. Therefore, the dynamic behavior of the system is defined in form of time
constant. When the calculated R and C values in Table 1 is substituted in (3), the transfer
function of the liquid level control system is achieved by:
G(s)
Qo ( s )
1
.
Qi ( s ) 1012 .6s 1
(4)
The MATLAB/Simulink model of the whole system, which consists of the mathematical model
derived for the liquid level control system and the PID controller with the tuned gain
parameters, is given in Fig. 2a. In Fig. 2b, the step response of the system is shown.
(a)
(b)
Figure 2. (a) MATLAB-Simulink model of the liquid level control system with PID
controller and (b) Step response of the system
3. EXPERIMENTAL SETUP
In Fig. 3, the experimental setup and block diagram of the liquid level training set are
presented. The equipment list also is given in Table 2.
In the system, a NC200 process control device is used as a compact PID controller that is
developed for educational purposes by Endcon Company. This device has the setup properties
of the fundamental control methods, including on-off control, P control, PD control, PI control
and PID control. Also, it has an analog input (AI) of 4-20 mA and an analog output (AO) of 10
V DC. While the AI of the device is connected to a Foxboro differential pressure transmitter
which measures the process variable, its AO is applied to a Siemens Acvatix SQS65
proportional vane which controls the process variable. The control signals produced by the
control methods in the device are sent to the proportional vane through the AO channel.
Another controller used in control of the liquid level system is a computer-based SCADA
control system with DAQ card. In this type control, the level variable in the system is measured
by the Foxboro differential pressure transmitter as in the system controlled by the compact PID
controller, and then the output of the transmitter is applied to the AI channel of ATS A-4011L
DAQ card. The control data that comes from the SCADA on the PC is sent to the Siemens
Acvatix SQS65 proportional vane by the AO channel of ATS A-4021L DAQ card. The
communication between the DAQ cards and the SCADA software on the PC is done by ATS
A-4520L DAQ RS232-RS484 converter The RS232 communication protocol is employed in
the data communication from the PC to the DAQ cards. Besides, the RS485 communication
protocol is used to transfer the data from the DAQ cards to the PC. The computer-based PID
control algorithm is carried out over the DAQ SCADA program developed by the authors.
(a)
(b)
(c)
Figure 3. The liquid level control training set (a) the front view, (b) the back view and (c)
the block diagram
Part no
Table 2. The equipment list and their properties
Name
Properties
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
Compact PID device
DAQ
SCADA
RS232 communication
On/off power switch
Tank
Reservoir
Load vane
Flow adjustment vane
Centrifugal pump
Proportional vane
Level transmitter
Power supply
System on/off switch
Flow transmitter
ENDCON NC200
ATS A-4520L, 4011L, 4080, 4021L
Genie
5 m cable
10 A
Fiber glass, R = 15 cm, h = 125 cm
40×30×30 cm
Manuel
Manuel
General Electric, 1~ 220 V AC, ½ HP
Siemens Acvatix SQS65, 11/4”
Foxboro, in/out: 0-6.6 kPa/4-20 mA
Meanweal, NES-100-24
16 A
ARF4 HH, turbine type
A reservoir tank that is made of plastic in 40×30×30 cm dimensions at the bottom side of the
set is utilized for the liquid reserve. A cylindrical tank that is made of transparent fiber-glass,
which has a diameter of 15 cm, a height of 125 cm and a thickness of 4 mm, is employed for
the control of the liquid level in the tank. The tope side of the tank is left open to the
atmospheric pressure. The liquid level in the tank is sensed through Foxboro differential
pressure transmitter. The transmitter has a measurement range of 0-6.6 kPa. The high pressure
input of the transmitter PH is connected to the bottom side of the cylindrical tank (PH = Patm +
Pliquid). Also, the low pressure input of the transmitter PL is left open to the atmospheric pressure
(PL = Patm). Since both the top of the tank and the low pressure input of the transmitter are left
open to the atmospheric pressure, the transmitter output changes depending on the height of the
liquid level in the tank (Pdifference = PH – PL = Pliquid). Water is used as the process variable to be
controlled in the system. The specific gravity of water is about 998.2 kg/m3 at 1 atm and 20 °C.
The laboratory temperature, in which there exists the experimental set, is about 20 °C. The
specific gravity of water has stayed at about the same value during the experiments. In order to
find out the base pressure of water in the cylindrical tank, the following equation is used
P .g.h .
(5)
Where P is the base pressure of the water (kPa), ρ is the water density (kg/m3), g is the
acceleration of gravity (9.8 m/s2) and h is the height of the water level (m). The height of the
maximum water level calculated from (5) is about 0.7 m when considering that the transmitter
measures a maximum pressure of 6.6 kPa (6,600 Pa). It means that the base pressure of the
water is 0 kPa when the water level in the tank is equal to 0 cm, and this pressure is 6.6 kPa
when the water level in the tank is equal to 70 cm. In the installation phase of the set, the output
of the transmitter has been respectively adjusted 4 mA and 20 mA through a precision
potentiometer on the transmitter when the tank is empty and full. Where the water density and
the acceleration of gravity are constant and only variable is the water level. The Siemens
Acvatix SQS65 proportional vane used for regulating the amount of the water from the
reservoir is controlled by two different control output of 0-10 V DC, which belongs to the
NC200 process control device and the SCADA control system with DAQ card. The vane
completely closes the water input when 0 V DC is applied, and also completely opens the water
input when 10 V DC is applied. The response time of the vane, the time required for passing
from full open position to fully closed position or vice versa, is 25 s. The circulation of water is
provided by a one phase AC induction motor of ½ HP manufactured by General Electric
Company. In the system, the load change is manually realized by a flow control vane. In
addition to the flow control vane, two manual flow adjustment vanes, which allow the inputoutput load changes in the system, are used to test the tuned controller parameters.
4. SOFTWARE
In the liquid level control system, the NC200 process control device that has been utilized as a
compact PID controller has a DSP with 16 bits and contains on-off, P, PD, PI, PID controller
algorithms as a standby code. There is a panel. User is able to tune the gain parameters
connected with the control methods after selecting the controller type over the panel. Thus, the
NC200 process control device has no need for additional software.
The software of the SCADA system with the DAQ card has been realized by the Advantech
VisiDAQ software which is a Windows-based data acquisition, control, analysis and
presentation development package. The feedback control algorithm with PID controller is
executed on this software developed for the computer-based control. The block diagram of the
developed software and preview image of the DAQ-SCADA window are given in Fig. 4a. In
the diagram, while the level information from the A-4011L card is taken by the AI1 block, the
output command is sent by the AO1 block to the proportional vane over the A-4021L card The
data from the flow transmitter, which is only utilized in monitoring the flow rate of the liquid,
is taken by the CTFQ1 block. NTCL1, SPIN1, SPIN2 and SPIN3 tags are used to enter the
values of set point, Kp, Ki and Kd, respectively. The PID1 block includes the PID control
algorithm which is required to obtain the control signals. The output position of the system is
regulated by PRG2 and PRG3 user programs. Preview image of the DAQ-SCADA window
which appears on the computer screen when running the software is shown in Fig. 4b.
(a)
(b)
Figure 4. (a) The block diagram of the DAQ-SCADA software and (b) preview image
of the DAQ-SCADA window
5. RESULTS AND DISCUSSION
In this section, we are going to separately present the experimental results and evaluations of
the liquid level control training set which has been controlled by both the NC200 process
control device and the computer-based SCADA system with the DAQ card.
5.1. Experimental results and evaluations of the training set controlled by the
NC200 process control device
As the compact controller, the control of the liquid level control system has been carried out by
using on-off, P, PD, PI and PID controllers over the NC200 process control device.
For the system controlled by the on-off controller, the process variable and the controller
hysteresis range have been set the values of 70% (height of 49 cm) and 10% (7 cm),
respectively. When the level value exceeds 49 cm, the controller completely closes the inflow
liquid into the tank by sending 0 V DC to the proportional vane. On the other hand, when the
level value of the tank drops below 42 cm (49 cm – 7 cm = 42 cm), the controller entirely opens
the inflow liquid into the tank by sending 10 V DC to the proportional vane and allows filling
the tank with liquid. Besides these values, we also have tested the system for the different
values of set point and hysteresis range.
Once the flow adjustment vane, which is manually controlled and regulates the inflow liquid
into the tank, is turned the highest value, the liquid level reaches 8 cm above of the set value (49
cm + 8 cm = 58 cm). This is due to the response time which occurs when the proportional vane
passes from full open position to fully closed position. This problem in the on-off control can be
tackled with using a solenoid vane instead of a proportional vane. However, solenoid vanes are
not used in the systems controlled by PID controllers since these controllers adjust the output
variable to the set value by producing an output proportional to the error. Therefore, a
proportional vane has been used in the designed liquid level control training set.
When the liquid level of the operating process is dropped to 42 cm, which is the lower limit of
the hysteresis range, a decrease has been seen up to about 8 cm (42 cm – 8 cm = 34 cm). This
situation has been also tested for the adjusted different set point and hysteresis values.
In order to control the system with P, PD, PI and PID controller methods, according to the Z-N
method, firstly, Kp, Ki and Kd values have been set 1, 9999 and 0, respectively. The NC200
process control device has been taken the PID mode. Then, the process variable has been set
70%. The controller gain parameters in the device have been tuned by the Z-N method. Then,
the training set has been operated with the adjusted set point value and Kp has been decreased by
small increments until a sustained oscillation with fixed amplitude is obtained. This oscillation
is reached when Kp is 80 and the gain value has been saved as Ku. The oscillation period Pu at
this moment has been measured as approximately 36 s. The measurement procedure of the
oscillation period is based on the average of the fifty periods counted by a chronometer. The
highest value and lowest value of these periods have been obtained as 39 s and 34 s,
respectively. The differences among the period values have arisen from the reading errors and
the changes of the process variable.
After determining Ku and Pu values, the final values of Kp, Ki and Kd for each controller have
been calculated employing the equations in Ziegler and Nichols [33], as shown in Table 3. The
experimental results of the system, which is controlled by the P, PD, PI and PID controllers with
the calculated gain parameters, are shown in Fig. 5.
Table 3. The Controller gain parameters tuned by the Z-N Method for the NC200 process
control device (Ku = 80, Pu = 36 s)
Control type
Kp
Ki
Kd
P
40
PD
36
162
PI
36
~1*
PID
48
~3*
216
*
: 1.20 and 2.66 values could not be entered to the controller. Instead of them,
1 and 3 values have been entered.
The output response curve of the P controlled system is given in Fig. 5a. For Kp = 40 value of
the P controller, the system has come to the steady state at 72 s and the steady state error has
been obtained as 8.16%.
Figure 5. Experimental results of the liquid level system controlled by the NC200 process
control device (a) P, (b) PD, (c) PI and (d) PID.
In Fig. 5b, the output response curve of the PD controlled system is shown. For Kp = 36 and Kd
= 162 values of the PD controller, although the system has reached the steady state at 42 s, the
steady state error has not been eliminated again. The steady state error has been 10.2% for the
mentioned gain parameters. As understood from the obtained results, the adding to the P
control structure of D control has only allowed the system to come to the steady state more
quickly.
In Fig. 5c, the output response curve of the PI controlled system is indicated. The operating
system with Kp = 36 and Ki = 1 has reached the steady state at 52 s. and the steady state error
has been zeroized. The deviations that occur until reaching the set value have been fairly
reduced by the PI control method, when comparing to the P and PD control methods.
In Fig. 5d, the output response curve of the system controlled by the PID controller that is
widely used in the systems with slow response time is shown. When the system is operated with
the gain values calculated for the PID controller in Table 3, the settling time and the steady
state error has been obtained as 37 s. and 0%, respectively. The deviations preceding the steady
state of the system have been much less than the other three control methods.
In addition to these analyses, the liquid level control system has been also tested by different
gain values of the controllers, but it is observed that the system has had better output response
with the gain parameter values.
5.2. Experimental results and evaluations of the training set controlled by the
computer-based SCADA system
A SCADA system with DAQ card developed by authors has been used in control of the system,
as a computer-based controller. In the system, the PID controller gain parameters has been
determined by fine tuning to the gain parameters, which have been previously obtained by the
Z-N method, on the SCADA screen given in Fig. 4b. According to the mentioned procedure,
the tuned optimum controller gains have been obtained as Kp = 0.80, Ki = 1.27 and Kd = 0.11 for
the set value of 40% of the process variable. On the SCADA screen, the chart at the bottom of
the window shows the change of the process variable versus time and this change is also
monitored by a bar graph which is located at the right side of the chart. The vertical and
horizontal axes on the chart are scaled as percent (%) and minute, respectively. The maximum
deviation of the process variable from the set value, which occurs before the system comes to
the steady state, has been measured as 15% in negative direction. The system has reached the
steady state at 4 min and the steady state error has been obtained as 0%. The transient response
to the parameter changes of the system has been tested over the set point, the output load and
the input load. Fig. 6 shows the transient responses to these changes of the liquid level control
system which is controlled by the computer-based SCADA system.
Fig. 6a, b, and c depict the transient response curves of the system for the case when the set
point is changed from the current value to the higher value. The changes in the set point has
been fulfilled for three different cases; namely from 40% to 50%, from 50% to 60% and from
60% to 80%. The transient responses to these changes of the system have been obtained as
about 1 min, 50 s, and 4 min, respectively. The steady state errors measured for three cases have
been 0 %. When the set value is dropped from 80% to 45%, a deviation of 3% has occurred in
the output response of the system in accordance with the new set value, as understood from the
Fig. 6d. The system has corrected this deviation at about 2.40 min and zeroized the steady state
error.
In Fig. 6e, the transient response of the system which operates in the set value of 45% is given
for the case when the output load is decreased by the manual vane. In that case, the deviation of
2% from the set value in the process variable has been corrected by the transient response of
about 2 min and the steady state error has been zeroized.
For the case that the output load is abruptly dropped while the system is working with the last
steady state conditions in Fig. 6e, the measured transient response is illustrated in Fig. 6f. In
that case, a deviation of 23% has happened in the process variable and this has been corrected at
about 5 min. The steady state error has been obtained as 0%.
In Fig. 6g, the transient response is shown for the case that the input load is increased the
highest value by the manual controlled flow adjustment vane while the system is operating
without changing the set value and the output load. In that case, a deviation of 5% from the set
point has been taken place in the process variable and this has been corrected at about 2 min.
However, the steady state error has been obtained as 2%.
In Fig. 6h, the transient response is given for the case that the set value of the system operating
with the last steady state conditions in Fig. 6g is changed from 45% to 60%. The deviation
arisen from the change has been corrected at about 40 s. The steady state error has been
occurred as 2%, as in the previous case.
In Fig. 6i, the transient response is shown for the case that the set value of the system is changed
from 60% to 70% while the input load is being kept the highest value. In that case, the process
variable has settled to the set value at about 50 s and the steady state error has stayed like the
previous one.
In Fig. 6j, for the case that the input load of the system working with the last steady state
conditions in Fig. 6i is manually decreased by the flow adjustment vane, the transient response
is given. In that case, the transient response of the system to the deviation of 10% occurring in
the process variable has been about 2 min and the steady state error has been zeroized.
(a)
(b)
(c)
(d)
(e)
(f)
(g)
(h)
(i)
(j)
(k)
(l)
Figure 6. Transient responses of the liquid level control system controlled by the computerbased SCADA system for Kp = 0.80, Ki = 1.27 and Kd = 0.11
In Fig. 6k, the transient response is illustrated for the cases that the set value of the system is
changed from 70% to 80%, from 80% to 90% and from 90% to 95%, consecutively. In the case
of the change from 70% to 80% of the set value, the transient response of the system is obtained
as about t1 = 1.20 min and the steady state error has been 2%. When the set value is changed
from 80% to 90%, the process variable has come the steady state in a very short time of about t2
= 5 s along with the zero steady state error. The system has reached the steady state in a long
time about t3 = 6.20 min along with the zero steady state error when the set value is changed
from 90% to 95%.
The last curve in Fig. 6l shows the transient response of the system for the case that the set point
is dropped by a large change from 95% to 50%. In that case, the steady state error has been
eliminated and the deviation of 3% resulted from the change has been corrected at about 4 min.
As a result, the fundamental control methods commonly used in automatic control systems
could be practically and easily applied on the training set. The system analyzes could be realized
in detail over the DAQ-SCADA interface, on which the parameters related to the system and
controller can be easily tuned and the experimental results can be graphically displayed.
Furthermore, the presence of a compact PID controller device in the training set would provide
significant advantages in terms of experiment interruptions both when the comparative analyzes
related to the system are carried out and when a problem in the PC is happened.
6. FEEDBACK FROM STUDENTS
The assessment of the designed liquid level control training set developed for engineering
education courses has been conducted on seventeen students who take the automatic control
system course. Firstly, basic control methods have been described to the students and then
fundamental knowledge’s have been provided about system controls. After that, the
mathematical model of the liquid level control system has been derived and the simulation
studies have been performed in MATLAB/Simulink environment by using this model. The
application study of the liquid level control system has been started by the training set after the
theory and simulation related to the system were completed.
In the first session of the application study, the NC200 process control device has been activated
and then the effects on the performance of the liquid level control system of on-off, P, PD, PI
and PID controllers have been experimentally investigated and analyzed for the cases where
both the system was worked under different conditions and the gain parameters of the
controllers were changed. In the second session of the application study, the performance
analyses of the liquid level control system controlled by the computer-based SCADA system
have been realized under the same working conditions and parameter changes.
At the end of this training, a form that contains the following six questions has been directed to
the sample student group:
a. Before studying with the liquid level control training set, the theory of the fundamental
feedback control methods in the automatic control systems, including on-off, P, PD, PI
and PID controllers, has been understood.
b. After learning the theory of the controllers, the realization of their applications has
greatly enhanced the understandability of the subject.
c. The main differences between the compact PID controller and the computer-based
SCADA system have been clearly comprehended.
d. If I encounter with a similar process in industry, then I can easily adjust the controller
gain parameters.
e. In the computer-based control, the having a flexible and visual screen of the DAQSCADA system has provided us an opportunity to comprehend and evaluate the
experimental results in a good way.
After the theory, the practices based on both the compact PID control and the DAQ-SCADA
systems are unnecessary and it is sufficient to apply only one of the control structures.
(a)
(b)
(c)
(d)
(e)
(f)
Figure 7. The survey on the automatic control students for the liquid level control training set.
The students have been requested to mark one of the following answers according to the degree
of agreement with the statement: (1) "strongly disagree", (2) "disagree", (3) "neither agree nor
disagree", (4) "agree" and (5) "strongly agree". The received feedbacks from the six questions
are given in detail in Fig. 7. The survey on the automatic control students has shown that the
liquid level control training set would be both increase the learning ability of the students and
allow them to express to be confidently worked in industry.
7. CONCLUSION
In this study, a liquid level control training set, on which the students could separately both
theoretically and experimentally apply the feedback control algorithms, have been designed for
control engineering courses. The practical application of this set realized on a prototype by both
a computer-based control and a DSP-based control. The results of the experimental studies
performed on the training set clearly shown the following acquirements: (1) the fundamental
control algorithms described in the control engineering courses could be successfully applied to
the control of a process, (2) thanks to the newly developed DAQ-SCADA software containing a
visual and flexible interface, a process could be controlled by the computer-based control
structures and analyzed under the different operating conditions in detail, (3) the comparative
analyses between the computer-based control and the DSP-based control structures could be
realized. It can be concluded from the feedbacks received from the students that the training set
provides an important contribution to the understanding and evaluating of the algorithm and
controller structures used in control of a process. With these features, the developed set is a
useful and beneficial instrument for the control engineering courses.
REFERENCES
[1].
C. M. Ionescu, E. Fabregas, S. M. Cristescu, S. Dormido, and R. De Keyser, “A remote laboratory
as an innovative educational tool for practicing control engineering concepts,” IEEE T. . Educ.,
vol. 56, no. 4, pp. 436-442, Nov. 2013. doi: 10.1109/TE.2013.2249516
[2].
Y. Linn, “An ultra low cost wireless communications laboratory for education and research,” IEEE
T. . Educ., vol. 55, no. 2, pp. 169-179, May 2012. doi: 10.1109/TE.2011.2158318
[3].
W. Maebashi, K. Ito, and M. Iwasaki “Practical feedback controller design for micro-displacement
positioning in table drive systems,” Automatika, vol. 54, no. 1, pp. 9–18, 2013. doi:
10.7305/automatika.54-1.310
[4].
S. Vijayalakshmi, and T. S. R. Raja. “Development of robust discrete controller for double
frequency buck converter,” Automatika, vol. 60, no. 1, pp. 303-317, 2015. doi:
10.7305/automatika.2015.12.647
[5].
K. Ogata, (1990). Modern Control Engineering, 2nd Ed., Prentice Hall, Englewood Cliffs, New
Jersey, (pp.1, 143–146).
[6].
A. Lozano-Nieto, “Microcontroller-based experiments for a control systems course in electrical
engineering technology,” Int. J. Elect. Enging. Educ., vol. 51, no. 1, pp. 43–53, Jan. 2014.
doi:10.7227/IJEEE.51.1.4
[7].
K. Kayisli, S. Tuncer, and M. Poyraz, “An educational tool for fundamental DC–DC converter
circuits and active power factor correction applications,” Comput. Appl. Eng. Educ., vol. 21, no. 1,
pp. 113–134, Mar. 2013. doi:10.1002/cae.20455
[8].
C. Dey, R. K. Mudi, and D. Simhachalam, “A simple nonlinear PD controller for integrating
processes,” ISA T., vol. 53, no. 1, pp. 162–172, Jan. 2014. doi:10.1016/j.isatra.2013.09.011
[9].
J. J. Fuertes, M. Domínguez, M. A. Prada, S. Alonso, and A. Morán, “A virtual laboratory of d.c.
motors for learning control theory,” Int. J. Elect. Enging. Educ., vol. 50, no. 2, pp. 172–187, Apr.
2013. doi:10.7227/IJEEE.50.2.6
[10]. S. Ekici, “Computer-aided power system fault analysis,” Comput. Appl. Eng. Educ., vol. 19, no.
1, pp. 66–70, Mar. 2011. doi:10.1002/cae.20290
[11]. M. Stefanovic, V. Cvijetkovic, M. Matijevic, and V. Simic, “A LabVIEW-based remote
laboratory experiments for control engineering education,” Comput. Appl. Eng. Educ., vol. 19, no.
3, pp. 538–549, Sep. 2011. doi:10.1002/cae.20334
[12]. I. S. Isa, B. C. C. Meng, Z. Saad, and N. A. Fauzi, “Comparative study of PID controlled modes
on automatic water level measurement system,” Proc. 7th International Colloquium on Signal
Processing and its Applications (CSPA), Penang, Malaysia, 2011.
[13]. J. Carrasco, W. P. Heath, M. C. Rodriguez Liñan, R. Alli-Oke, O. A. R. Abdel Kerim, and S.
Rodriguez Gutierrez, “Themed project case study: quadruple tanks control with PLCs,” Int. J.
Elect. Enging. Educ., vol. 50, no. 3, pp. 279–292, Jul. 2013. doi:10.7227/IJEEE.50.3.8
[14]. R. Yenitepe, “Design and implementation of a SCADA-controlled MTMPS as a mechatronics
training unit,” Comput. Appl. Eng. Educ., vol. 20, no. 2, pp. 247–254, Jun. 2012.
doi:10.1002/cae.20391
[15]. B. Najjari, S. M. Barakati, A. Mohammadi, M. J. Futohi, and M. Bostanian, “Position control of
an electro-pneumatic system based on PWM technique and FLC,” ISA T., vol. 53, no. 2, pp. 647–
657, Mar. 2014. doi:10.1016/j.isatra.2013.12.023
[16]. Khairurrijal, M. Abdullah, and M. Budiman, “Home-made PIC 16F877 microcontroller-based
temperature control system for learning automatic control,” Comput. Appl. Eng. Educ., vol. 19,
no. 1, pp. 10–17, Mar. 2011. doi:10.1002/cae.20283
[17]. B. Dandil, “An integrated electrical drive control laboratory: Speed control of induction motors,”
Comput. Appl. Eng. Educ., vol. 20, no. 3, pp. 410–418, Sep. 2012. doi:10.1002/cae.20407
[18]. N. Silva, H. Bispo, R. Brito, and J. Manzi, “Joint confidence region for the tuning parameters of
the PID controller,” Proc. IFAC Conference on Advances in PID Control (PID’12), Brescia, Italy,
2012.
[19]. B. Kumar, and R. Dhiman, “Optimization of PID controller for liquid level tank system using
intelligent techniques,” Can. J. Elect. Comput. E., vol. 2, no. 11, pp. 531–535, Dec. 2011.
http://journal.ieee.ca/index.html
[20]. D. Pavković, S. Polak, and D. Zorc, “PID controller auto-tuning based on process step response
and damping optimum criterion,” ISA T., vol. 53, no. 1, pp. 85-96, Jan. 2014.
doi:10.1016/j.isatra.2013.08.01
[21]. Y. M. Zhao, W. F. Xie, and X. W. Tu, “Performance-based parameter tuning method of modeldriven PID control systems,” ISA T., vol. 51, no. 3, pp. 393–399, May 2012.
doi:10.1016/j.isatra.2012.02.005
[22]. A. Leosirikul, D. Chilin, J. Liu, J. F. Davis, and P. D. Christofides, “Monitoring and retuning of
low-level PID control loops,” Chem. Eng. Sci., vol. 69, no. 1, pp. 287–295, Feb. 2012.
doi:10.1016/j.ces.2011.10.046
[23]. R. Zhang, S. Wu, and F. Gao, “Improved PI controller based on predictive functional control for
liquid level regulation in a coke fractionation tower,” J. Process Contr., vol. 24, no. 3, pp. 125–
132, Mar. 2014. doi:10.1016/j.jprocont.2014.01.004
[24]. M. Huba, and M. Simunek, “Modular approach to teaching PID control,” IEEE T. Ind. Electron.,
vol. 54, no. 6, pp. 3112–3121, Dec. 2007. doi:10.1109/TIE.2007.907024
[25]. B. Vahidi, S. A. Agheli, and S. Jazebi, “Teaching short-circuit withstand test on power
transformers to M.Sc. students and junior engineers using MATLAB-SIMULINK,” Comput.
Appl. Eng. Educ., vol. 20, no. 3, pp. 484–492, Sep. 2012. doi:10.1002/cae.20416
[26]. W. J. Shyr, “Teaching mechatronics: An innovative group project-based approach,” Comput.
Appl. Eng. Educ., vol. 20, no. 1, pp. 93–102, Mar. 2012. doi:10.1002/cae.20377
[27]. A. Tekin, F. Ata, and M. Gökbulut, “Remote control laboratory for DSP-controlled induction
motor drives,” Comput. Appl. Eng. Educ., vol. 20, no. 4, pp. 702–712, Dec. 2012.
doi:10.1002/cae.20440
[28]. M. I. B. Dzolkafle, “Implementation of PID controller for controlling the liquid level of the
coupled tank system,” PhD Thesis, Universiti Malaysia Pahang, 2008.
[29]. S. Bennet, “Development of the PID controller,” IEEE Contr. Syst., vol. 13, no. 6, pp. 58–62,
Dec. 1993. doi: 10.1109/37.248006
[30]. K. H. Ang, G. Chong, and Y. Li, “PID control system analysis, design, and technology,” IEEE T.
Contr. Syst. T., vol. 13, no. 4, pp. 559–576, Jul. 2005. doi:10.1109/TCST.2005.847331
[31]. A. Kumar, and A. Narain, “Designing of controller based on artificial neural network for liquid
level system,” UACEE International Journal of Artificial Intelligence and Neural Networks –
IJAINN 3, vol. 3, no. 3, pp. 79–88, Sep. 2013. http://seekdl.org/upload/files/
20130905_011602.pdf
[32]. Z. A. A. Muhasain, and S. A. A. Muhasain, “Analysis and design of controller for level process
control without sensor,” Nahrain University, College of Engineering Journal (NUCEJ), vol. 13, no.
1, pp. 84–97, Jan. 2010. http://www.iasj.net/iasj?func=fulltext&aId=28978
[33]. J. G. Ziegler, and N. B. Nichols, “Optimum settings for automatic controllers,” Trans. ASME, vol.
64, no. 11, pp. 759-768, 1942. http://chem.engr.utc.edu/Student-files/x2008-Fa/435-Blue/1942paper.pdf
Authors
Hayati Mamur received the B.S. degree from Department of Electrical Education, Gazi
University, in 1996, has graduated the M.S. and Ph.D. degrees from Department of
Electronics and Computer Education, Gazi University, Ankara, Turkey, in 2005 and 2013,
respectively. He is working as an Asst. Prof. at the Department of Electrical and
Electronics Engineering, Faculty of Engineering, Manisa Celal Bayar University, Manisa,
Turkey. His current research interests include micro wind turbines, permanent magnet
generators, convertors, invertors, micro grid systems and thermoelectric modules.
Ismail Atacak received the B.S., M.S. and Ph.D. degrees from Department of Electronics
and Computer Education, Gazi University, in 1994, 1998, and 2005, respectively. From
2007 to 2012, he worked as an Assistance Professor at the Department of Electronics and
Computer Education, Faculty of Technical Education, Ankara, Turkey. He is currently
working as an Assistance Professor at the Department of Computer Engineering, Faculty of
Engineering, Ankara, Turkey. His research interests include power systems, artificial
intelligent based algorithms, optimization based algorithms and Engineering Education.
Fatih Korkmaz was born in Kırıkkale, Turkey, in 1977. He received the B.S., M.S., and
Ph.D. degrees in in electrical education, from University of Gazi, Ankara, Turkey,
respectively in 2000, 2004 and 2011. Since 2012, he is working as an Asst. Prof. Dr. at the
Department of Electrical and Electronics Engineering, Faculty of Engineering, Cankiri
Karatekin University, Cankiri, Turkey. His current research field includes Electric
Machines Drives and Control Systems.
Mohammad Ruhul Amin Bhuiyan received the B.Sc., M.Sc. degrees in Applied Physics
and Electronic Engineering from Rajshahi University, Bangladesh in 1994 and 1995,
respectively. Ph.D. degree in Applied Physics, Electronics and Communication
Engineering from Islamic University, Bangladesh in 2008. Now at present he is working as
a visiting scientist at the Department of Electrical and Electronics Engineering, Faculty of
Engineering, Manisa Celal Bayar University, Manisa, Turkey.
| 3cs.SY |
Proofs as Relational Invariants of Synthesized Execution Grammars
CALEB VOSS, Georgia Institute of Technology
DAVID HEATH, Georgia Institute of Technology
WILLIAM HARRIS, Georgia Institute of Technology
arXiv:1710.03357v1 [cs.PL] 10 Oct 2017
The automatic verification of programs that maintain unbounded low-level data structures is a critical and open problem. Analyzers
and verifiers developed in previous work can synthesize invariants that only describe data structures of heavily restricted forms, or
require an analyst to provide predicates over program data and structure that are used in a synthesized proof of correctness.
In this work, we introduce a novel automatic safety verifier of programs that maintain low-level data structures, named Lttp. Lttp
synthesizes proofs of program safety represented as a grammar of a given program’s control paths, annotated with invariants that
relate program state at distinct points within its path of execution. Lttp synthesizes such proofs completely automatically, using a
novel inductive-synthesis algorithm.
We have implemented Lttp as a verifier for JVM bytecode and applied it to verify the safety of a collection of verification
benchmarks. Our results demonstrate that Lttp can be applied to automatically verify the safety of programs that are beyond the
scope of previously-developed verifiers.
1
INTRODUCTION
Automatically verifying that a given program satisfies a desired safety property is a fundamental problems of program
verification. Recent work has seen the development of powerful program verifiers that operate automatically [9, 22–
25, 38]. Such verifiers can often determine if practical programs satisfy properties concerning their control flow and
facts over a bounded collection of data values [5].
However, verifying the safety of programs that maintain unbounded low-level data structures remains an open
problem. A significant body of previous work has developed shape analyzers [17, 26] that, given a program P, synthesize
invariants of the reachable heaps of P represented in a particular shape domain, such as three-valued logical structures [35,
41, 45] or separation-logic formulas [12, 15, 42, 53]; the invariants synthesized by such analyzers can potentially imply
facts about program states that establish that P is safe. Another body of work has developed automatic program
verifiers [1, 4, 10, 14, 16, 19, 28, 29, 31, 34, 40, 43] and decision procedures [39, 46] that directly attempt to determine if
P is safe by attempting to synthesize sufficient invariants in such domains.
Unfortunately, all such approaches suffer from at least one of several critical limitations. In particular, they either
are only able to represent invariants over heaps of restricted forms [4, 14, 16, 28, 29, 31, 34] or require an analyst to
manually provide abstractions [1, 12, 14, 15, 17, 19, 26, 31, 39–43, 45, 46, 53] or candidate inductive invariants [10, 28].
The main contribution of this work is an automatic verifier, named Lttp, that attempts to determine the safety of a
program that may maintain low-level data structures. Lttp satisfies two key features that distinguish it from previous
approaches. First, it can potentially prove the safety of programs by establishing inductive invariants that relate multiple
low-level structures maintained by a program, each of which need not necessarily have a pre-specified shape, such as
© 2017 Copyright held by the owner/author(s). Manuscript submitted to ACM
1
list or tree. Second, Lttp performs synthesis of these invariants completely automatically, without requiring an analyst
to provide predicates over values or data-structure shapes from which to attempt to synthesize a proof.
We designed Lttp to satisfy both of the above features by developing two key technical insights. The first insight is
that the safety of a program that maintains low-level data structures can often be established by invariants that relate
only values bound to variables (i.e., local data) at multiple control points. Each proof structure synthesized by Lttp is an
annotated path grammar of the program—i.e., a graph grammar in which the yield of each derivation is a program
path. Each non-terminal A of the grammar is annotated with a formula that relates only the local values of a tuple of
program points generated by each application of A. Such a formula is in the combination of a theory that axiomatizes
the program’s data operations with the theory of uninterpreted functions.
The second insight is that proofs as relational invariants of graph grammars can potentially be synthesized automatically, using an inductive synthesizer [2, 47–50]. The inductive synthesizer implemented by Lttp, given a program
P, iteratively maintains a set of enumerated control paths of P. In each iteration, Lttp synthesizes a candidate graph
grammar G from the structure of P and enumerated paths by reduction to constraint solving. Lttp attempts to synthesize
relational invariants of G by reduction to logic programming [9, 44]. If a logic-programming solver determines that no
such invariants exist, then Lttp collects from the solver an unexplored control path which it uses to refine the proposed
grammar in subsequent iterations.
We have developed an implementation of Lttp that verifies programs represented in Java Virtual Machine (JVM)
bytecode. We evaluated Lttp by using it to attempt to verify the safety of a collection of challenging problems for shape
analyses and verifiers, some of which are adapted from those presented in the SV-COMP verifier competition [20].
The results demonstrate that Lttp is powerful enough to express non-local invariants, combining information
separated in time by unboundedly many execution steps and separated in heap space by unboundedly many field
accesses. Without manual assistance, Lttp can prove unary and binary properties over lists, reachability properties,
and even correlative properties between two disjoint data structures.
In summary, the design of Lttp builds on and contributes to multiple topics in program analysis and verification.
First, Lttp constitutes a shape verifier with strengths that are distinct from all shape verifiers presented in previous work,
though it does not necessarily subsume such verifiers, as discussed in §5.4. Second, Lttp demonstrates that techniques
from relational verification [6–8] which previously have been applied to prove properties of multiple programs by
relating states across multiple runs, can be applied to prove shape properties by relating states within program runs.
Third, Lttp demonstrates that techniques from inductive synthesis, which have previously been applied to synthesize
correct programs as completions of incomplete programs, can also be used to synthesize alternate representations of
complete programs that are amenable to verification.
The rest of this paper is organized as follows. §2 illustrates the operation of Lttp by example. §3 reviews previous
work on which Lttp is based, §4 describes the proofs synthesized by Lttp, and §5 describes Lttp’s inductive-synthesis
algorithm. §6 describes our empirical evaluation of Lttp. §7 compares Lttp to related work, and §8 concludes.
2
OVERVIEW
This section illustrates Lttp by example. §2.1 introduces an example program as a verification problem. §2.2 introduces
the proof of buildInspect synthesized by Lttp. §2.3 describes how Lttp synthesizes the proof automatically.
2.1 buildInspect: maintaining a low-level queue
2
1
2
3
4
5
6
7
8
9
10
11
12
class Element { Element next; }
public static void buildInspect(int num) {
Element head = new Element();
Element tail = head;
for (int i = 0; i < num; i++) {
Element tmp = new Element();
tail.next = tmp;
tail = tmp; }
Element elt = head;
for (int i = 0; i < num; i++)
elt = elt.next;
assert elt == tail; }
Fig. 1. buildInspect: constructs a queue with num elements and traverses it from head to tail.
a:
3
b:
5
i:
10
c:
7
d:
8
j:
11
e:
5
k:
10
f:
7
g:
8
l:
11
h:
5
m:
10
n:
12
Fig. 2. A control path of buildInspect containing two iterations of the loops, including data-dependence edges.
Figure 1 contains a program, named buildInspect, that maintains a low-level queue. The queue is represented as a
linked list of Element objects (line 1). buildInspect first constructs a single Element object and binds it to variables
that store both the queue’s head (line 3) and tail (line 4). buildInspect then iteratively creates new Elements and
adds them to the end of the queue (lines 5—8). buildInspect then iterates over the elements in the queue, storing
each element in elt (lines 10—11). Finally, buildInspect asserts that, on completion of the loop in lines 10—11, elt
stores the tail element of the queue (line 12).
Each execution of buildInspect satisfies the assertion at line 12. The key invariant for the loop over lines 5—8
establishes that the Element object stored in elt reaches the object stored in tail over some number i of dereferences
of the next pointer. The key invariant for the loop over lines 10—11 establishes that the Element object stored in elt
reaches the Element object stored in tail over num − i dereferences of the next pointer, where n is the length of the
queue. Unfortunately, it is challenging to design a verifier that can express such invariants and infer them automatically
from a given program without supplied candidate invariants.
2.2
A proof of buildInspect’s safety as relational invariants of a path grammar
A key insight behind our approach is that proofs of the safety of a program P can be represented as a graph grammar
that generates control paths of P, annotated with invariants that relate states at different path points when they co-occur
in the same grammar rule. In this section, we give a grammar GBI of buildInspect’s control paths (§2.2.1) and
relational invariants of GBI that represent a proof of the safety of buildInspect (§2.2.2).
2.2.1 A grammar of buildInspect’s control paths. Figure 2 contains the control path of buildInspect that
includes two iterations of the loop at lines 5—8 and two iterations of the loop at lines 10—11. In Figure 2, each node is
an instance of a control location, annotated with its line number. The line number of each node indicates the program
has reached, but not executed, the instruction on that line. Control steps are depicted as solid edges. Each point that
stores an object is connected to the point that loads the stored object by a dashed line.
A graph grammar GBI that generates the control paths of buildInspect is depicted in Figure 3. GBI contains two
relations, buildInspect and StLd, with buildInspect the starting relation. GBI consists of three clauses, one for
buildInspect and two for StLd. Each clause is depicted as a hypergraph. The nodes represent program control points
and are labeled with an alphabetic character for reference and with the corresponding line number in the program. The
edges consist of control-dependence edges, depicted as solid edges; the dashed edges depict data dependencies, and are
3
StLd[0]
buildInspect[0]
a:3
b:5
0
1
StLd
d:
10
2
3
StLd[1]
c:9
a:5
b:9
a:5
1
3
0
0
0
2
e:
12
c:
10
d:
12
b:7
c:8
d:5
StLd
2
f:
10
0 1
g:
11
h:
10
2 3
e:9
1
3
i:
12
Fig. 3. A graph grammar that generates buildInspect’s control paths and has relational invariants that prove the safety of
buildInspect. The grammar contains three clauses, each labeled with their head relation (either buildInspect or StLd).
described in §2.2.2. Each (ordered) hyperedge h is represented as a box labeled with a relation, and with the nodes in h
labeled with their index in h. The hyperedge that each clause defines is represented implicitly, with the nodes in the
hyperedge dashed and labeled with their index in the hyperedge.
The starting relation of GBI , buildInspect, has a single clause, buildInspect[0], that generates fragments of a
control path that model the initial step of buildInspect and the step of buildInspect from the exit of the loop at
lines 5—9 to the entry of the loop at lines 10—11.
The relation StLd simultaneously derives a pair of paths, corresponding to the two loops of the program. GBI
contains two clauses for StLd. Clause StLd[0] generates a pair of loop exit steps. Clause StLd[1] generates both a step
through the loop at lines 5—9 that allocates a new Element object and stores it at the next field of a bound object, and
a step through the loop at lines 10—11 that loads an Element object. Clause StLd[1] recursively includes an instance
of the StLd relation, which can be further expanded in order to generate the rest of each loop.
The control path in Figure 2 is generated by GBI . In particular, it is generated by applying the following sequences
of rules: buildInspect[0], StLd[1], StLd[1], StLd[0].
It should be noted that the grammar shown in Figure 3 is a simplification of the grammar required to prove the
safety of buildInspect. In particular, the depicted grammar only allows expansions of StLd such that the two loops
iterate the same number of times. It is true that every actual execution has this property, but this fact is not known a
priori. Therefore, the grammar must allow paths with different numbers of iterations in each loop so that the underlying
model checker can discover such an invariant through counterexamples.
2.2.2 Relational invariants of buildInspect’s path grammar. A second key observation that motivates our approach
is that a proof of the safety of a program P can be represented as some grammar G of the paths of P paired with
relational invariants over the location instances in the interface of each of the relations of G. In particular, a proof
of the safety of buildInspect can be represented as the path grammar GBI (given in §2.2.1), paired with relational
invariants over instances of control locations in the interface of buildInspect and StLd that establish that at the 0th
interface node of buildInspect, elt = tail.
The relational invariant for StLd establishes that if tail at index 0 is equal to elt at index 2 and if the point in the
path at which the next field of tail at index 1 was stored is the point at which the next field of elt at index 2 was
stored, then tail at index 1 is elt at index 3. It also establishes that i at index 0 is equivalent to i at index 2, indicating
4
Loop0[0,1]
3
Loop0
Loop1[0,1]
4
5
0
6
7
8
5
Loop0
Loop1
buildInspect[0]
0
5
9
10
Loop1
0
10
11
0
10
12
10
′ that generates ccbuildInspect’s control paths, but does not have relational invariants that
Fig. 4. An alternative graph grammar G BI
prove the safety of buildInspect.
that the loops will iterate the same number of times. The relational invariant for StLd can be expressed as the following
formula, where integer subscripts indicate interface node indices and next indicates a function which maps each object
to the path point at which it was last updated:
(tail0 = elt2 ∧ next1 (tail0 ) = next2 (elt2 ) =⇒ tail1 = elt3 ) ∧ i0 = i2
(1)
Formula 1 entails that each run of buildInspect satisfies its assertion at line 12. Furthermore, it can be proved to
hold over all tuples of path points in all paths generated by GBI , by induction on the derivations of GBI . In particular,
(1) in clause StLd[0], the semantic constraints of the instructions on generated control steps entail Formula 1 with the
variables at index 0, 1, 2, and 3 replaced with variables that represent state at points a, b, c, and d respectively. (2) In
clause StLd[1], semantic constraints of the instructions on generated control steps, combined with Formula 1 with the
variables at index 0, 1, 2, and 3 replaced with variables that model state at points d, e, h, and i, entail Formula 1 with
variables at index 0, 1, 2, and 3 replaced with variables that represent state at path points a, e, f , and i.
2.3
Proving safety of buildInspect automatically
Lttp, given a program P with an error location L, attempts to prove that L is unreachabale in P by synthesizing a path
grammar that has relational invariants that prove that L is unreachable in P. To prove that buildInspect always
satisfies the assertion at line 12, Lttp synthesizes the path grammar of buildInspect given in §2.2.1 and the relational
invariants of the grammar given in §2.2.2.
Lttp, given program P and error location L, attempts to synthesize a proof that L is unreachable in P as relational
invariants by performing an inductive synthesis algorithm [2, 21, 32, 51]. The algorithm maintains an initially empty
set of enumerated control paths of P.
In each iteration, Lttp synthesizes a graph grammar G that derives all control paths of P and that admits relational
invariants proving the safety of all paths enumerated. As the set of enumerated paths is initially empty, the first grpah
grammar synthesized by Lttp is likely a simple grammar whose structure directly corresponds to the control-flow
′ given in Figure 3, in which each path is derived left-recursively.
graph of buildInspect, such as the grammar G BI
′ given in Figure 3, in which each relation derives a single path leftE.g., it could synthesize the graph grammar G BI
recursively. As a heuristic, Lttp could initially synthesize path grammars that use right-recursive definitions, the
program’s syntactic structure, or other sub-structures of the program’s control-flow graph, such as components of its
Bourdoncle decomposition [11]. However, no such fixed grammar can serve as a proof of correctness in general, and
must be iteratively restructured.
5
After Lttp synthesizes a graph grammar G, it attempts to determine if G has relational invariants that represent a
proof of the safety of P. Lttp reduces this problem to solving a system S of Constrained Horn Clauses (see §3.2.2), and
runs a solver on the generated system as a black box. If the solver determines that S has a solution, then the solution
contains relational invariants of G that prove safety of P.
Otherwise, if the solver provides a counter-derivation D of S that has no solution, then D defines a control path
′ , Lttp determines that some
that cannot be proved safe by relational invariants of G. For example, when given G BI
′ , such as the one corresponding to the control path p given in Figure 2, has no solution. Such a
derivation of G BI
derivation has no solution because nodes of the path that are the sites of matching stores and loads do not co-occur in
′ .
the clauses of G BI
Lttp then inspects the control path of the unsolvable derivation and determines if it is truly unsafe. If so, then Lttp
determines that P is unsafe. Otherwise, as in the case of p, Lttp determines that the path is safe, adds it to the set of
′ that
enumerated paths, and recurses. In Lttp’s next iteration, it will synthesize a path grammar distinct from G BI
admits relation invariants proving the safety of p. Lttp eventually synthesizes G BI , which proves safety for all paths.
3
BACKGROUND
In this section, we review foundations on which Lttp is built. In §3.1, we define the low-level imperative language
targeted by Lttp. In §3.2, we review concepts from formal logic and logic programming used in the design of Lttp.
3.1
A target imperative language
In this section, we define the structure (§3.1.1) and semantics (§3.1.2) of Lttp’s target language.
3.1.1 Structure. A program is a set of instructions, each labeled with a control location and a target control location.
The space of all control locations is denoted Locs, with distinguished initial location and final location LI , LF ∈ Locs.
The finite spaces of value variables and object variables are denoted DVars and OVars, respectively; their union is
denoted Vars = DVars ∪ OVars. The finite spaces of data fields and object fields are denoted DFields and OFields,
respectively. All spaces of mutually disjoint. Lttp can be applied to programs in languages with infinite spaces of
variables and fields, because it only must consider the finite sets of variables and fields that occur in a given program.
We assume that the spaces of variables and fields themselves are finite in order to simplify the presentation of Lttp.
The space of data instructions is denoted InstrsV , and includes standard operations of Boolean and linear integer
arithmetic. Let p, q ∈ OVars, x ∈ DVars, f ∈ DFields, and g ∈ OFields be arbitrary elements. y:=p->f is a data load
and p->f:=y is a data store; q:=p->g is an object load and p->g:=q is an object store. The spaces of all data loads, data
stores, object loads, and object stores are denoted FieldsD , StoresD , LdsO , and StoresO , respectively. p:=new() is
an allocation, x:=isNil(p) is a nil test, and 1:=p=q is an object-equality test; the space of all allocations, nil tests, and
object-equality tests are denoted Allocs, NilTests, and ObjEqs, respectively. The space of all instructions is denoted
Instrs = InstrsV ∪ FieldsD ∪ StoresD ∪ LdsO ∪ StoresO ∪ Allocs ∪ NilTests ∪ ObjEqs
A pre-location, instruction, and branch-target location is a labeled instruction; i.e., the space of labeled instructions
is denoted LblInstrs = Locs × Instrs × Locs. For each labeled instruction i ∈ LblInstrs, the pre-location,
instruction, and post-location of i are denoted PreLoc[i], Instr[i], and BrTgt[i], respectively.
6
Instruction
i ∈ InstrsV
x:=p->f
p->f:=x
q:=p->f
p->f:=q
x:=p=q
x:=isNil(p)
p:=new()
Updated
Components
Updated Value
σD
σD
HD
σO
HO
σD
σD
σO
HD
HO
σ D′ ∈ CtxsV , σ D →V σ D′
i
σD [
x 7→H (σO (p), f)
]
HD [
(σO (p), f) 7→σ D (x)
]
q 7→H D (σ D (p), f)
]
σO [
HO [
(σO (p), f) 7→σO (q)
]
σD [
x 7→σO (p) = σO (q)
]
σD [
x 7→σO (p) = nil
]
σO [
p 7→o < Dom(HO ) = Dom(H D )]
H D [{o} × DFields 7→0
]
HO [{o} × OFields 7→nil
]
Table 1. Resulting post-state of instructions from, for σ D ∈ CtxsV , σO ∈ CtxsO , H D ∈ HeapsD , and HO ∈ HeapsO , the pre-state
((σ D , σO ), (H D , HO )). In each instruction, p, q ∈ OVars, x ∈ DVars, d ∈ DFields, and f ∈ OFields. In the entry for p:=new(),
for each map m, Dom(m) denotes the domain of m.
A program is a set of labeled instructions. I.e., the space of programs is Lang = P(LblInstrs), where for any set
S, P(S) denotes the powerset of S. For each P ∈ Lang and all L, L ′ ∈ Locs, there may be at most one i ∈ LblInstrs
such that PreLoc[i] = L and BrTgt[i] = L ′ . In such a case, i is denoted alternatively as Instr[P](L, L ′ ).
3.1.2 Semantics. A run of a program P is a sequence of states generated by a sequence of labeled instructions in
which adjacent instructions have matching target locations and pre-locations. The space of data is Data = B ∪ Z, the
space of non-nullable objects is a countably-infinite space Objs ′ , and the space of objects is Objs = Objs ′ ∪ {nil}.
The space of local data states is CtxsV = DVars → Data, the space of local object states is CtxsO = OVars → Objs,
and the space of local states is Ctxs = CtxsV × CtxsO . The space of data heaps is HeapsD = Objs ′ × DFields → Data,
the space of object heaps is HeapsO = Objs ′ × OFields → Objs, and the space of heaps is Heaps = HeapsD × HeapsO .
The space of states is States = Ctxs × Heaps.
For each i ∈ InstrsV , there is a transition relation →V ⊆ CtxsV × CtxsV . The transition relation of a value
i
instruction need not be total: thus, labeled instructions can implement control branches using instructions that act as
assume instructions. The transition relation of instructions is defined based on their structure. The transition relation
of each instruction, defined over an arbitrary pre-state, is given in Table 1.
For each labeled instruction i ∈ LblInstrs, the transition relation of i is the transition relation of the instruction
of i; i.e., →i =→Instr[i] . A program state σ may not be the source of any entry in the transition relation if it binds an
object variable p to nil and executes an instruction that attempts to load from or store to nil. For the remainder of this
paper, we do not consider a stuck state to be an error state; programs can be transformed so that stuck states are error
states in our formulation.
For each finite space N , E ⊆ N × N such that (N , E) is a sequential graph and each λ : N → Locs, (N , E, λ) is
a control path. For each P ∈ Lang, if (N , E, λ) is such that for all path edges (n, n ′ ) ∈ E, there exists an instruction
Instr[P](λ(n), λ(n ′ )), then (N , E, λ) is a control path of P. The space of control paths of P is denoted Paths[P]. For each
p ∈ Paths[P], the nodes of p are denoted Nodes[p].
A run of a program P is a control path p of P and a map from each node of p to a state such that states associated
with adjacent nodes satisfy the transition relation of a corresponding instruction of P.
7
Definition 1. For P ∈ Lang, finite space N , E ⊆ N × N , and λ : N → Locs such that (N , E, λ) ∈ Paths[P], let
σ : N → States be such that for all n, n ′ ∈ N with (n, n ′ ) ∈ E, σ (n) →Instr[P](λ(n), λ(n ′ )) σ (n ′ ). Then for p = (N , E, λ),
(p, σ ) is a run of p in P.
For p ∈ Paths[P], the runs of p in P are denoted Runs[p]. For r ∈ Runs[p], p is denoted alternatively as DerPathr .
If Runs[p] is empty, then p is infeasible. The runs of P are the runs of all paths of P; i.e., the runs of P are denoted
Ð
Runs[P] = p ∈Paths[P] Runs[p]. If Runs[P] is empty, then P is safe. The core problem addressed in this work is, given
program P, to determine if P is safe.
3.2
Constraint solving and Constrained Horn Clauses
3.2.1 Formal logic. The quantifier-free fragment of the theory of the combinations of linear arithmetic and uninterpreted functions is denoted EufLia. For each space of logical variables X , the space of EufLia formulas over X is
denoted Forms[X ]. For each formula φ ∈ Forms[X ], the set of variables that occur in φ (i.e., the vocabulary of φ) is
denoted V (φ). For formulas φ 0 , . . . , φ n , φ ∈ Forms[X ], the fact that φ 0 , . . . , φ n entail φ is denoted φ 0 , . . . , φ n |= φ. The
models of variables X are denoted Models[X ]. The fact that a model m satisfies a formula φ is denoted m ⊢ φ. Lttp uses
a decision procedure for EufLia, named IsSat. We assume that for each i ∈ InstrsV , Lttp may access some formula
SymRelV [i] ∈ Forms[EufLia].
3.2.2 Constrained Horn Clauses. Constrained Horn Clauses are a class of logic-programming problems that formulate
problems in program verification [9, 18, 22, 44].
Structure. A Constrained Horn Clause is a body, consisting of uninterpreted relational predicates applied to logical
variables and a constraint, and a head application. Relational predicates are symbols associated with arities.
Definition 2. For each space of symbols R and function a : R → N, (R, a) is a space of relational predicates.
An application is a relational-predicate symbol paired with a sequence of logical variables of length matching its
arity. To simplify the presentation, all objects defined in the remainder of the section are defined over a fixed space of
predicate symbols R in which each symbol has arity k ∈ N, and a fixed space of logical variables X .
Definition 3. For relational predicate R ∈ R and sequence of variables Y ∈ X ∗ such that |Y | = k, (R, Y ) is an
application of R over X .
The space of applications of symbols in R over X is denoted Apps[R, X ]. For each application A ∈ Apps[R, X ],
the predicate symbol and argument sequence of A are denoted Rel[A] and Args[A] respectively. A clause is a set of
applications, a constraint over logical variables, and a head relational predicate.
Definition 4. For B ∈ Apps[R, X ]∗ , φ ∈ Forms[X ], and H ∈ R, (B, φ, H ) is a Constrained Horn Clause.
The space of Constrained Horn Clauses over R and X is denoted CHC R,X . For each C ∈ CHC R,X , the body of
applications, constraint, and head of C are denoted Apps[C], Ctr[C], and Head[C], respectively. A set of Constrained
Horn Clauses and a query relational-predicate symbol is a system of Constrained Horn Clauses; i.e., the space of systems
of Constrained Horn Clauses is denoted CHCs R,X = P(CHC R,X ) × R.
8
Models. A model of S is a collection of logical models that certify that S does not have a solution. A derivation
of a CHC system S is a tree D labeled with relational predicates of S such that all children with a common parent
p are labeled with relational symbols of applications in the body of a common clause of S with p as its head. For
C ⊆ CHC R,X , let each relational predicate that is not the head of any clause of C be a ground relational predicate of C.
Definition 5. For C ⊆ CHC R,X and Q ∈ R, E ⊆ N ∗ × N × C, and λ E : E → C be such that (N , E) is an directed
hypertree and there is some labeling function λ N : N → R such that: (1) for r ∈ N the root of (N , E), λ N (r ) = Q;
(2) for each n ∈ N a leaf of (N , E), λ N (n) is a ground relational predicate of C; (3) for all n 0 , . . . , nk , n ∈ N with
e = ([n 0 , . . . , nk ], n) ∈ E, it holds that (a) λ N (n) = Head[λ E (e)] and (b) for each 0 ≤ i ≤ k, λ N (ni ) = Rel[Apps[C]i ].
Then (N , E, λ E ) is a derivation of (C, Q).
For S ∈ CHCs R,X , the space of derivations of S is denoted Ders(S). For each D ∈ Ders(S), the nodes and hyperedges
of D are denoted Nodes[D] and HypEdges(D).
A model of a CHC system S is a derivation D of S and a model indexed on nodes of D that satisfies the clauses that
label the edges of D.
Definition 6. For S ∈ CHCs R,X , finite space N , E ⊆ N ∗ × N , λ E : E → Clauses[S] such that (N , E, λ E ) ∈ Ders(S),
let m be an EufLia model and let i : N → Models[X ] be such that for all n, n 0 , . . . , nk ∈ N and each C ∈ Clauses[S] with
e = ([n 0 , . . . , nk ], n) ∈ E, it holds that for C = λ E (e), (1) m, i(n) ⊢ Ctr[C]; (2) for each 0 ≤ j ≤ k, i(m)(Args[Apps[C]j ]) =
i(n j )(Params). Then (D, m, i) is a model of S.
We denote the models of S as Models[S]. For each D ∈ Ders(S), if there is some EufLia model m and i : Nodes[D] →
Models[X ] such that (D, m, i) is a model of S, then D is feasible; otherwise, D is infeasible. If some derivation of S is
feasible, then S is feasible (otherwise, S is infeasible).
A CHC solver is a procedure that, given CHC system S, returns either the value Infeas to denote that S is infeasible,
or a model of S; several CHC solvers have been proposed in previous work [9, 44]. Lttp uses a CHC solver, named
SolveCHC, as a black box.
Solutions. A solution of a clause C is an interpretation of relational predicates such that the conjunction of interpretations of all relational predicates in the body of C and the constraint of C entail the interpretation of the head of C. A
solution of a CHC system S is a solution of each clause in S that interprets the query relational predicate of S as an
unsatisfiable formula.
Definition 7. For C ⊆ CHC R,X and Q ∈ R, let i : R → Forms[Params] be such that under each EufLia model, (1)
for each B ∈ Apps[R, X ]∗ , H ∈ R, and φ ∈ Forms[X ] such that (B, H, C) ∈ C,
{i(Rel[A])[Vars[A]]}A∈B , φ |= i(H )
(2) i(Q) |= False. Then i is a solution of (C, Q).
When a CHC solver determines that a given system is empty, the solver can synthesize a solution that certifies
emptiness. If Lttp runs SolveCHC on a system S and SolveCHC determines that S is empty, then Lttp does not
require SolveCHC to provide the generated solution. However, Lttp could be adapted to generate the CHC system that
it synthesizes, accompanied by its solution found by SolveCHC, as a proof of safety that can be independently certified.
9
4
PROOFS AS RUN GRAMMARS
In this section and §5, we present Lttp in technical detail. In this section, we give a class of proofs of program safety as
grammars of program runs annotated with invariants.
Lttp, given program P, attempts to determine whether or not P is safe by synthesizing a CHC system S such that (1)
each run of P corresponds to a model of S and (2) S has no models. Lttp attempts to certify that S has no models by
synthesizing a solution of S in the theory TD,Uif . Let P be a fixed, arbitrary program for the remainder of the section.
Let the theory TLang be TD,Uif restricted to contain the following uninterpreted function symbols. The nullary
symbols of TLang contain, for each L ∈ Locs, the symbol L. The unary symbols of TLang contain symbols Loc and
Succ; for each x ∈ DVars, the symbol x; for each p ∈ OVars, the symbol p. The binary symbols of TLang contain for
each f ∈ DFields, the symbol f; for each g ∈ OFields, the symbol g.
If for each m ∈ Models[S] with domain N , the graph (N , ≻m , Locm ) is a control path of P, then S is a run grammar
of P. The space of run grammars of P is denoted CHCsLang P. Let S ∈ CHCsLang P be a fixed, arbitrary element for
the remainder of the section.
Lttp uses a procedure DerPath that, given m ∈ Models[S], returns the control path of the run of m. DerPath is
implemented by returning the interpretations of symbols Loc and Succ in m.
Each run r defines a map for each data and object field f, from each control point p of r and object o allocated while
executing r to the control point in r at which the f field of o was last updated when r reached p. In particular, let N be
a finite space, L : N → Locs; ≻: N → N ; σ : N → States be such that ((N , ≻, Loc), σ ) ∈ Runs[P]. For each x ∈ DVars,
xr : N → Vals is such that for each n ∈ N , xr (n) = σ (n)(x), and for each p ∈ OVars, pr : N → Objs is defined similarly.
For each f ∈ DFields, let cc fr : N × Objs → N be such that for all ni , n j ∈ N and o ∈ Objs, if ni is the last node
before n j such that Instr[P](Loc(ni−1 ), ni ) ≡ p->f:=x and σ (ni )O (x) = o, then fr (n j , o) = ni . For each g ∈ OFields,
let gr : N × Objs → N be defined similarly. Both collections of symbols are called update histories.
For S ∈ CHCsLang P, let domain N combined with L, ≻, {xr | x ∈ DVars}, {pr | p ∈ OVars}, {fr | f ∈ DFields},
{fr | f ∈ OFields} as interpretations of Loc, Succ, DVars, OVars, DFields, and OFields be the model of TLang
denoted mr . If mr is a model of S, then r is a run of S. If each run of P is a run of S, then P is simulated by S.
For p ∈ Paths[P], if for each r ∈ Runs[p], mr is not a model of S, then p is refuted by S. For Q ⊆ Paths[P], if for each
p ∈ Q, it holds that p is refuted by S, then Q are refuted by S. If Paths[P] are refuted by S, then P is refuted by S.
If a program is simulated and refuted by a CHC system, then the program is safe.
Lemma 1. If P is simulated by S and P is refuted by S, then P is safe (§3.1.2).
If P is simulated by a run grammar S, then a solution of S (which certifies that S is infeasible and thus refutes P),
can be viewed as relational invariants that prove the safety of P.
Example 1. Recall GBI , the grammar partially depicted in §2.2.1, Figure 3. If we encode the runs of buildInspect
using the structure of GBI , we could simulate every run, and showing the emptiness of that grammar could show the safety
of buildInspect.
However, a safe program P may be simulated by a CHC system that does not refute it.
′ , the simplistic grammar proposed in Figure 4. If we try to encode P using this structure, we are
Example 2. Recall GBI
compelled to overapproximate by dropping state information because an unbounded separation arises between a load and
10
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
Input : P ∈ Lang.
Output : Decision as to whether P is safe.
Procedure Lttp(P)
Procedure Lttp ′ (F )
G := SynGrammar(P, F ) ;
switch SolveCHC(G) do
case Infeas: do
return True
case D ∈ Ders(G): do
p := DerPath(D) ;
if IsFeas(p) then
return False ;
else
return Lttp ′ (p ∪ {F }) ;
end
end
end
return Lttp ′ (∅)
Algorithm 1: Lttp: a safety verifier based on inductive synthesis. Lttp uses procedures SynGrammar
(§5.2, §5.3), SolveCHC (§3.2.2), DerPath (§4), and IsFeas
(§5.1.1).
1
2
3
4
5
6
7
Input : P ∈ Lang and p ∈ Paths[P].
Output : A minimal refuting neighborhood
ν : Nodes[p] → P(Nodes[p]) for p.
Procedure Deps(P, p)
ν := ν All ;
for n, n ′ ∈ Nodes[p] do
ν ′ := ν [n 7→ ν (n) \ {n ′ }] ;
if ¬IsSat(SymPath(P, p, ν ′ )) then ν := ν ′
;
end
return ν
Algorithm 2: Deps: given P ∈ Lang and infeasible
p ∈ Paths[P], returns a minimal refuting neighborhood for p. The map ν All and the formula SymPath
are defined in §5.1.1.
′ , but there are additional models not corresponding to
its matching store. Every runs of buildInspect is present in GBI
actual runs. This non-empty grammar simulates buildInspect, even though buildInspect is safe.
5
INDUCTIVE SYNTHESIS OF PROOFS USING LTTP
In this section, we describe Lttp, a verifier that attempts to prove the safety of a given program by inductively
synthesizing a run grammar as proof.
Alg. 1 contains pseudocode for Lttp. Lttp defines a procedure Lttp’ which, given infeasible paths F ⊆ Paths[P],
attempts to determine if P is safe by synthesizing a run grammar that overapproximates the runs of P and refutes F
(line 2—line 15). Lttp invokes Lttp’ on the empty set of control paths and returns the result (line 16).
Lttp’, given F , synthesizes a run grammar G that simulates P and refutes F by a procedure SynGrammar on P and
F (line 3). An implementation of SynGrammar that performs a reduction to constraint solving is described in §5.3.
Lttp’ then determines if G refutes P by running the CHC solver SolveCHC (described in §3.2.2) on G to determine if
G is infeasible. If SolveCHC determines that G is infeasible (line 5), then Lttp’ returns that P is safe.
Otherwise, SolveCHC returns a feasible derivation D of G (line 7). Lttp’ extracts from D some p ∈ Paths[P] not
refuted by G (line 8). Lttp’ then tests if p is feasible by running a procedure IsFeas on p (line 9); if IsFeas determines
that p is feasible, then Lttp’ returns that P is not safe.
Otherwise, if IsFeas returns that p is feasible, then Lttp’ recurses on F extended with p, and returns the result of the
recursion (line 12).
We now describe the implementation of each procedure used by Lttp. In §5.1, we describe an implementation of
IsFeas. In §5.2 and §5.3, we describe the two steps performed by SynGrammar.
11
5.1
Testing path feasibility using IsFeas
The procedure IsFeas, given P ∈ Lang and p ∈ Paths[P], returns whether or not p is a feasible path of P. Let finite
space N , E ⊆ N × N , and λ : N → Locs, be such that p = (N , E, λ). IsFeas generates a constraint φ ∈ Forms[TLang ]
for which each model corresponds to run of p.
5.1.1 Symbolic constraints over update histories. φ is constructed as a conjunction of clauses, each of which models
the effect of an instruction executed in a step of p on its local variables and update histories. The effect of each i ∈ P
executed in a step of p is formulated by a constraint SymRel[i] parameterized on N , along with distinguished n, n ′ ∈ N
that model state before executing i and state that immediately results from executing i.
SymRel[i](n, n ′, N ) is defined casewise by the structure of i. In many cases, SymRel[i] is defined using formulas that constrain equality of logical terms that model state. In particular, let EqDVars[n, n ′ ], EqOVars[n, n ′ ],
EqDFields[n, n ′, N ], and EqOFields[n, n ′, N ] constrain that the states at q and n ′ have equal local value states, local
object states, data timestamps, and object timestamps, respectively. I.e.,
Û
EqDVars[n, n ′ ] ≡
x(n) = x(n ′ )
x ∈DVars
Û
′
EqDFields[n, n , N ] ≡
f(n, p(n ′′ )) = f(n ′, p(n ′′ ))
f ∈DFields
p ∈OVars
n ′′ ∈N
EqOVars[n, n ′ ]
and
EqOFields[n, n ′, N ]
are defined similarly.
If i ∈ InstrsV , then SymRel[i] constrains that the value local state at n and the value local state at n ′ are in the
transition relation of i, and that their object stores are identical. I.e., SymRel[i](n, n ′, N ) is
SymRelV [i][DVars(n), DVars(n ′ )] ∧ EqOVars[n, n ′ ]∧
EqDFields[n, n ′, N ] ∧ EqOFields[n, n ′, N ]
For each i ∈ NilTests or i ∈ ObjEqs, the constraint SymRel[i] is defined similarly.
For x ∈ VarsV , p ∈ OVars, and f ∈ DFields, SymRel[p->f:=x](n, n ′, N ) constrains that at n ′ , the most recent
store to the f field of the object bound to p is n ′ . The local states and update history of all other fields are identical
between n and n ′ . I.e., SymRel[p->f:=x](n, n ′, N ) is
Û
EqDVars[n, n ′ ] ∧ EqOVars[n, n ′ ] ∧ f(n ′, p(n ′ )) = n ′ ∧
f(n ′, p(n ′′ )) = f(n, p(n ′′ ))∧
p ∈OVars
Û
′
′′
′′
n ′′ ∈N
′
g(n , p(n )) = g(n, p(n )) ∧ EqOFields[n, n , N ]
g,f ∈DFields
p ∈OVars
n ′′ ∈N
For each i ∈ StoresO , the constraint SymRel[i] is defined similarly.
Example 3. buildInspect contains an object store on line 7. The relation which models this instruction constrains
the local state after the instruction such that the most recent store for the next field of the object tail is set to tmp.
For each x ∈ DVars, p ∈ OVars and f ∈ DFields, SymRel[x:=p->f](n, n ′, N ) inspects all states bound to variables
in Q to determine if some n ′′ ∈ N is the point of the most recent store to f. If so, the value in x at n ′ is constrained to be
12
the value stored when stepping to n ′ ; otherwise, the value bound to x at n ′ is unconstrained. Let DStoreLocP ⊆ Locs be
control locations that are sources of data stores in P, and let Stored[P] : DStoreLocP → DVars map each such control
location to the data variable that holds that value stored by the instruction. Then SymRel[x:=p->f](n, n ′, N ) is
Û
f(q, p(q)) = n ′′ =⇒ x(q ′ ) = Stored[P](n ′′ )∧
n ′′ ∈N , λ(n ′′ )∈DStoreLocP
Û
y(n ′ ) = y(n) ∧ EqOVars[n, n ′ ] ∧ EqDFields[n, n ′, N ] ∧ EqOFields[n, n ′, N ]
y,x ∈DVars
For each i ∈ LdsO , the constraint SymRel[i] is defined similarly.
Example 4. buildInspect contains an object load on line 11. The relation which models this instruction constrains
the local state after the instruction such that the variable elt is set to the most recent store for the next field of the object
elt.
For x ∈ OVars, SymRel[x:=new()](n, n ′, N ) constrains that the identity of the allocated object is n ′ ; the fields of the
allocated object are initialized at n ′ .
EqDVars[n, n ′ ] ∧ x(n ′ ) = n ′ ∧
Û
y(n ′ ) = y(n) ∧ Stored(n ′ ) = 0∧
y,x ∈OVars
Û
′
f(n , x(n ′ )) = n ′ ∧
f ∈DFields∪OFields
Û
f(n ′, y(n ′ )) = f(n, y(n))
y,x ∈OVars
f ∈DFields∪OFields
For a path p = (N , λ, E) ∈ Paths[P], define the constraint SymPath(P, p, ν ) ∈ Forms[TLang ], where ν : N → P(N ),
as follows:
Û
SymRel[Instr[P](λ(n), λ(n ′ ))](n, n ′, ν (n ′ ))
(n,n ′ )∈E
Let ν All : N → P(N ) be such that for each n ∈ N , ν All (n) = N . IsFeas generates the constraint φ = SymPath(P, p, ν All ).
IsFeas returns that p is a feasible path of P if and only if IsSat (see §3.2.2) decides that φ is satisfiable.
IsFeas is a sound and complete procedure for testing path feasibility.
Lemma 2. If p is a feasible path of P, then IsFeas(P, p) = True. Otherwise, IsFeas(P, p) = False.
IsFeas could be implemented alternatively by reduction to satisfiability testing in alternative theories that can soundly
and completely model the feasibility of bounded paths, such as combinations of the theory of arrays. We have presented
IsFeas as using IsSat in order to introduce the symbolic relations on update histories for each instruction, which Lttp
also uses in queries in order to synthesize run grammar skeletons (§5.2), and as components of run grammars (§5.3).
5.2
Synthesizing a run-grammar skeleton using constraint solving
In this section and in §5.3, we describe an implementation of SynGrammar, which given P ∈ Lang and F ⊆ Paths[P],
synthesizes a run grammar G that simulates P and refutes F. SynGrammar synthesizes such a run grammar by
performing two steps, described in this section and §5.3. In the first step, SynGrammar finds the relations of G and the
sets of variables in each clause body to which relational predicates are applied. We refer to such an object as a CHC
system skeleton. SynGrammar finds a skeleton by reduction to constraint solving.
SynGrammar obtains a clause skeleton for P and F by running a procedure SynSkeleton on P and F. The implementation of SynSkeleton that we present always makes progress, in the sense that for each P and F, it synthesizes a
13
well-formed skeleton for P and F. However, the implementation is not complete, in that there are some safe programs
whose required skeletons are not expressible in this particular constraint-based approach to SynSkeleton. We have
found that the present implementation is sufficiently expressive for Lttp to be able to prove correctness of interesting
verification challenge problems (discussed in §6). We leave the design of alternative, more expressive, or more efficient
implementations of SynSkeleton for future work.
5.2.1 Run-grammar skeletons. In its first step, SynGrammar synthesizes a run-grammar skeleton, which defines the
control paths generated by run grammar constructed from completing it. We define a restricted implementation of
SynGrammar that only generates run grammars in which each non-terminal derives exactly two control subpaths of P.
In particular, let RSkel be a set of relational predicates of fixed, common arity n. Let Q be a space of variable symbols, and
let Q ′ ∈ Q ∗ be fixed sequence of n distinct variables in Q. For R ∈ RSkel , A ∈ Apps[RSkel , Q], L : Q → Locs, q, q ′ ∈ Q,
C = (R, A, L, q, q ′ ) is a clause skeleton. The space of clause skeletons is denoted ClauseSkels. A set of clause skeletons
combined with four maps from RSkel → Zn is a run-grammar skeleton. For the remainder of this section, let R, A, L, q,
q ′ , and S ⊆ ClauseSkels and F = Pre0 , Post0 , Pre1 , Post1 : RSkel → Zn be a fixed, arbitrary elements.
Synthesizing a skeleton that simulates a given program. (S, F ) simulate P if when each clause in S is extended
with a suitable constraint to form a clause, the resulting run grammar generates all control paths of P. In particular,
for φ ∈ Forms[TLang ], let ((A, φ), R) be the completion of C. Let S ′ ∈ CHCsLang such that each clause in S ′ is a
Ó
completion of a clause (R, A, L, q, q ′ ) with a constraint q ∈Q Loc(q) = L(q) ∧ q ′ = Succ(q) be the control-path grammar
of S. If for each R ′ ∈ RSkel and each model m of R, there is a control path from m(Pre0 (R ′ )) to m(Post0 (R ′ )) and a
control path from m(Pre1 (R ′ )) to m(Post1 (R ′ )), then (S, F ) is well-formed. If S ′ simulates p, then S simulates P. If all
r , r ′ ∈ Runs[S ′ ] that have the same control path have the same derivation in S ′ , then S is unambiguous.
SynSkeleton synthesizes a well-formed, unambiguous control-path grammar that simulates P by reduction to
constraint solving. In particular, Lttp generates an EufLia constraint φ P such that each model of φ P defines a set of
clause skeletons that simulate P. For each clause skeleton C, the problem of choosing a relational predicate to apply in
the body of C, the set of variables to which it is applied, the map from variables to locations, and instruction source and
destination variables can be encoded directly as an EufLia constraint.
Synthesizing a skeleton that refutes control paths. C refutes F if for each p ∈ F , the derivations of the control extension
of C simultaneously derive sufficient sets of matching loads and stores in p that a suitable extension of C (described in
§5.3) refutes p. Such sufficient sets are formulated precisely as a minimal refuting neighborhood of p.
For p ∈ F with N = Nodes[p] and ν : N → P(N ) such that SymPath[P, p, ν] is unsatisfiable, ν is a refuting
neighborhood of p. For all ν, ν ′ : N → P, if for all n ∈ N , ν ′ (n) ⊆ ν (n ′ ), then ν ′ is contained by ν ; if, in addition, ν is not
contained by ν ′ , then ν ′ is strictly contained by ν . If for each ν ′ : N → P(N ) that is strictly contained by ν , it holds
that ν ′ is not refuting neighborhood of p, then ν is a minimal refuting neighborhood of p.
Deps (Alg. 2), given P and an infeasible p ∈ Paths[P], returns a minimal refuting neighborhood of p. Deps maintains a
refuting neighborhood of p, from which it iteratively minimizes entries until it obtains a minimal refuting neighborhood.
Deps constructs ν All as an initial refuting neighborhood (line 2). For all n, n ′ ∈ N , it determines if the map ν ′ obtained
by removing n ′ from the image of n in its maintained refuting neighborhood ν (line 4) is a refuting neighborhood. To
determine this fact, Deps runs IsSat on SymPath[P, p, ν ′ ] (line 5). If ν ′ is a refuting neighborhood, then Deps updates
its maintained refuting neighborhood to be ν ′ (line 5).
14
For each p ∈ F with N = Nodes[p], SynSkeleton synthesizes a minimal refuting neighborhood of p, named
ν p : N → P(N ) by running Deps on P and p. SynSkeleton then constructs a constraint φ p in which each solution
defines (1) a set of clause skeletons, over the vocabulary interpreted as a clause skeleton in each solution of φ P , and (2)
a derivation D of p in which for each n ∈ N and each n ′ ∈ ν p (n), n and n ′ are derived in a common instance of a clause
in D.
′ , the simplistic grammar of Figure 4. As a skeleton, G ′ fails to refute even a short
Example 5. Recall once again GBI
BI
infeasible path of buildInspect. For example, if a path performs a store and a load (the transitions from line 7 to 8 and
from line 11 to 10), a refuting neighborhood must put the state at line 10 into the neighborhood of the state at line 8 or
′ to include a special case rule for this path. A better way is to construct the
vice versa. One way to do this is to modify GBI
grammar GBI from Figure 3, which refutes much large set of infeasible paths. In fact, this skeleton refutes all paths in
buildInspect.
We have described a particular implementation of SynSkeleton that we have implemented in the current version
of Lttp. The current of implementation of SynSkeleton is restricted, in that it only synthesizes skeletons of run
grammars that are linear and unambiguous. SynSkeleton finds such skeletons by reduction to constraint solving. The
key motivation for adopting such limitations was that a relatively simple but useful version of SynSkeleton could
be designed by reusing existing, heavily-optimized algorithms implemented in constraint solvers. However, Lttp
only requires that an implementation of SynSkeleton, given program P and infeasible paths F, synthesize a skeleton
encodes all paths of P, including refuting neighborhoods for the paths inF. Alternative, explicit implementations of
SynSkeleton may yield significant benefits over the current version based on constraint solving.
5.3
Synthesizing a run grammar from a skeleton
In its second step, SynGrammar completes C ⊆ Skels, a set of clause skeletons that simulates P and a refutes
F synthesized by SynSkeleton, to generate a run grammar G ∈ CHCsLang that simulates P and refutes F. Let
R ∈ RSkel , L : Q → Locs, q, q ∈ Q be such that (R, L, q, q ′ ) ∈ Skels. From (R, A, L, q, q ′ ), SynGrammar generates
φ ∈ Forms[TLang ], and includes in G the clause C = (A, φ, R).
φ is a conjunction of constraints that model the effect of i, and effects of allocations and stores performed in steps of
execution other than i (§5.3.1). The first conjunct φ 0 constrains that each instance of q ′ is a control success of each
instance of q; i.e., φ 0 ≡ q ′ = Succ(q). The second conjunct φ 1 constrains that for path points n and n ′ bound to q
and q ′ at an instance of C ′ , the state at n ′ is the result of transitioning from the state at n, under the instruction that
connections n and n ′ . I.e., φ 1 ≡ SymRel[Instr[P](L(n), L(n ′ ))](q, q ′, Q).
The third conjunct φ 2 models the effect of all instructions that connect points derived by other clause instances on
the objects in scope when the step of C executes. The construction of the constraint is described in detail below.
5.3.1 Formulating the effect of instructions derived outside of a clause instance. The space of valid runs with update
histories is partially constrained by the symbolic relations for each instruction, defined in §5.1.1. For each i ∈ Instrs,
SymRel[i] models the effect of executing i on a finite set of states. When SymRel[i] is used to determine feasibility of
an entire path, in which case the set consists of all states in the path. When SymRel[i] is used as a conjunct of a clause
constraint, the set of states consists of all states that are derived in the same instance of a clause C that derives the
states connected by i. However, it does not include the set of all states in the derived path, namely states that occur
15
exclusively in subderivations of C, or that occur in a subderivation outside of the derivation of C and are not provided
as arguments to C.
φ 3 constrains the effect of instructions executed in steps from such states on objects that may only be bound to
object variables in states derived in an application of C. If i is not an object load or allocation, then each object in scope
bound to a variable in an application of C is in scope for its child or and parent; as a result, φ ≡ True.
Otherwise, let p ∈ OVars be the object variable bound by i. Then φ 3 is a conjunction of two constraints. Let
R A ∈ RSkel and Q A ∈ Q ∗ be such that A = R A (Q A ). The first constraint, φ 30 ∈ Forms[TLang ], constrains that if the
object o bound to p at q is not in the scope of any states bound to Q A , then the update histories of o is identical across
both control subpaths derived by the derivation with head A. I.e., φ 30 is
Û
p(q) , p ′ (q ′′ ) =⇒
q ′′ ∈Q A
p′ ∈OVars
Û
f(Q A [Pre0 (R A )], p(q)) = f(Q A [Post0 (R A )], p(q))∧
f ∈OFields∪DFields
f(Q A [Pre1 (R A )], p(q)) = f(Q A [Post1 (R A )], p(q))
The second constraint, φ 31 ∈ Forms[TLang ], constrains that if o is not in an argument of R when C is applied, then the
update histories of o is identical between the end of the first control subpath that it derives and the beginning of the
second control subpath that it derives. I.e., φ 31 is
Û
p(q) , p ′ (q ′′ ) =⇒
q ′′ ∈Q ′
p′ ∈OVars
Û
f(Q ′ [Post0 (R)], p(q)) = f(Q ′ [Pre1 (R)], p(q))
f ∈OFields∪DFields
5.3.2 Key properties. The partial correctness of Lttp depends only on the fact that SynGrammar, given a program
and set of infeasible paths, synthesizes a run grammar that simulates the program.
Lemma 3. P is simulated by SynGrammar (P, F).
SynGrammar also synthesizes a run grammar that refutes F. However, this fact is primarily useful for proving that
Lttp’ progresses, in that it never collects the same control path from distinct invocations of SolveCHC.
5.4
Key features
Lttp is correct on all programs on which it terminates.
Theorem 1. If Lttp(P) = True, then P is safe, and if Lttp(P) = False, then P is not safe.
Because the problem of verifying safety of programs in Lang is undecidable, there are some programs on which
Lttp will not terminate.
One limitation of Lttp as presented above is that it can only effectively determine the safety of programs that contain
all stores of objects that they load. Defining a logic of program summaries that Lttp can both use and validate is a
conceptually challenging and critical direction for future work. Lttp is motivated by a number of practical applications
which it can be applied to. However, having a program summary logic would enable a direct and formal comparison on
the theoretical level to other approaches, particularly separation logic [42] and effectively propositional reasoning [29].
16
5.5
Discussion
The prototype of SynSkeleton, developed in §5.2, has some shortcomings which make it worthwhile to pursue a
replacement. It is deficient in the expressivity of the grammars it can return, and it is intractable on modest programs. For
an infeasible path p in the feedback set F , the existing constraint-based approach cannot enforce that every derivation of
p include a refuting neighborhood ν for p, merely that one such derivation does; hence, the expressivity of SynSkeleton
must be artificially limited to unambiguous grammars. If it were possible to additionally supply negative examples to
SynSkeleton (graphs of p not including any refuting neighborhood), then ambiguity would be permissible.
Learning a linear string grammar (a regular expression) through queries to an oracle is a solved problem due to
Angluin’s L∗ algorithm [3]. To be precise, one supposes a teaching oracle which can answer membership queries for
a secret language L and which can confirm or deny with a counterexample that a proposed DFA D has the property
L(D) = L. L∗ is an efficient algorithm for using this oracle to learn such a D through both positive and negative
examples. L∗ is guaranteed to terminate when L is regular. The parallel between this problem statement and the task of
SynSkeleton in Lttp is striking. In Lttp, a grammar C for an unknown language of graphs is sought. IsFeas(P, p, ν )
answers the query of whether the control path graph p = (N , λ, E) unioned with the data edges E Data implied by ν is a
member of this language. When a run grammar G is proposed, SolveCHC(SynGrammar(P, F )) attempts to confirm or
deny with a counterexample that G correctly describes the language.
The task of SynSkeleton is more difficult than the regular language learning task of L∗ because the unknown
language is a set of graphs, not strings. However, there is work towards L∗ -like algorithms for richer classes of
languages such as learners of context-free grammars [3, 13, 54] and multiple context-free (MCF) grammars [55]. A single
non-terminal symbol in an MCF grammar is not a hole in a single string, but a vector of r holes, where the rank r is
bounded. The rule for a non-terminal of rank r can invoke other non-terminals of rank r by producing characters at the
holes, of rank less than r by filling a hole, and of rank greater than r by splitting a hole into two adjacent holes. Again,
a striking parallel with Lttp arises: this behavior of MCF non-terminals is exactly the behavior of skeleton grammars
returned by SynSkeleton, where control edges play the role of characters. It remains an open question to determine if
an MCF learner can be adapted to include the data edges of a refuting neighborhood ν for each positive example p.
6
EVALUATION
We empirically evaluated Lttp in order to answer the following questions: (1) Can Lttp verify the safety of low-level
programs that operate on unbounded data structures where existing approaches fail? (2) Can Lttp verify the safety of
such programs efficiently?
We implemented Lttp as a verifier for programs represented in Java Virtual Machine (JVM) bytecode. The set of
benchmarks consists of some programs adapted from challenge benchmarks in the SV-COMP benchmark collection [20],
a program that cannot be verified by a competing approach [30], novel benchmarks designed to exhibit particular
capabilities of Lttp, and a program which demonstrates that Lttp is limited to properties that can be expressed in a
context free way.
In short, the results of our experiments indicate that Lttp can be applied to verify the safety of programs that
cannot be verified by existing shape analyzers or verifiers. The time required for the Lttp implementation to perform
verification varies dramatically, and we diagnose what causes intractability in some cases.
17
§6.1 describes our implementation of Lttp for JVM and our experimental procedure; §6.2 summarizes each benchmark;
§6.3 provide a qualitative evaluation of Lttp by describing the proofs that it synthesizes in order to prove the safety of
several illustrative benchmarks. §6.4 provides a quantitative evaluation of Lttp by analyzing its performance.
6.1
Implementation and experimental procedure
We implemented Lttp as a verifier for JVM bytecode. Lttp can thus be applied to verify programs written in Java, Scala,
or other languages with compilers that target the JVM. The hypothetical language, Lang, targeted by Lttp was defined
in §3.1, including classes of instructions that perform operations over objects, such as loads, stores, and allocations.
Each such instruction directly corresponds to an instruction in the JVM instruction set. In addition, the implementation
of Lttp can be applied to programs that use common operations over scalar data, such as linear arithmetic and Boolean
functions. Lttp does not currently support programs that consist of multiple procedures.
Lttp for JVM supports program specifications through library operations that implement the semantics of assume
and assert instructions. There are further operations for retrieving non-deterministic data. The semantics of Lang do
not include the concept of a NullPointerException or any other exception. Accessing the fields of a null object is
legal, undefined behavior in Lang. Lttp does not model exception-throwing runs of JVM programs. Even so, one can
verify the absence of a NullPointerException in the original program by injecting a null check before every field
access.
Lttp for JVM implements several optimizations, in comparison to the conceptual version of Lttp for Lang given
in §5. In particular, the performance of both the procedure SynGrammar for synthesizing a run grammar (§5) and
SolveCHC is heavily affected by the number of control locations of a given program. Lttp for JVM coalesces sequences
of value instructions, allocations, and control branches that are not loop back-edges a single block to reduce the number
of effective control locations. The benefit of this optimization is that SynSkeleton can represent the space of alternative
run-grammar skeletons more compactly, and SynGrammar can generate CHC systems with fewer relational predicates
and clauses. The disadvantage of this optimization is that the constraints in the generated CHC system are more
complex. In practice, the optimization results in a significant improvement in performance.
Furthermore, instead of modeling the value of each program variable at each control point using an uninterpreted
function (§4), Lttp translates the program into static single assignment form, and introduces first-order variables in
generated CHC systems that correspond to program variables. This increases the total number of first-order variables
included in a system, but allows the Lttp implementation to perform fixed-point analyses such as alias and object
liveness analysis that are relatively cheap compared to the performance cost of a CHC solver. Lttp runs such analyses
to simplify the constraints in the CHC clauses before solving.
Finally, on practical benchmarks, solving a given CHC system to obtain a counterexample is often far more expensive
than an explicit search through all possible program paths. Accordingly, before attempting to solve a given system, our
implementation of Lttp unrolls the system to enumerate all paths up to some heuristic depth and examines each one
with Alg. 2 to determine if it is a counterexample.
Lttp uses the Sawja program analysis framework [27] to process JVM bytecode input and the Z3 theorem prover [56]
to solve TD,Uif satisfiability queries (§3.2.1). Lttp uses as a CHC solver (§5) the implementation of Duality [9] that
accompanies Z3, modified to apply Z3’s aggressive formula simplifier after each iteration of its solving algorithm; The
original version of Duality performs only cheap simplifications, which often caused a problemtic explosion in the size
18
of the invariants. Lttp uses the XSB Datalog engine [52] to execute the fixed-point analyses for optimization prior to
solving. Aside from these components, the Lttp codebase consists of approximately 12,300 lines of OCaml.
We also implemented a baseline verifier for JVM bytecode which encodes programs using the theory of arrays to
model the heap. The baseline generates a CHC system with one relation for each control location and a clause for each
pair of adjacent control locations to model a step of execution. The heap of a program is modeled as a collection of
logical arrays, one for each field, that each map each object to the value stored at that field. Load and store instructions
are modeled with select and store operations of the theory of arrays, and objects are represented as integer identifiers.
In this way, the baseline reduces the problem of deciding program safety to the problem of solving a CHC system in
the theory of arrays with linear arithmetic, Auflia. The baseline verifier coalesces instructions similarly to Lttp as
described above, and attempts to solve the resulting system by running Duality.
We applied both the baseline and Lttp to a set of benchmarks to determine if Lttp could verify the safety of
programs that could not be verified using other theories that accurately model the semantics of memory operations.
For each benchmark program P, we gave P to Lttp and the baseline verifier and observed the result of, and resources
used by, each verifier to attempt to prove that P satisfies its single (without loss of generality) assert statement. For
most benchmarks, we discovered divergent behavior in the successive invariants generated by the baseline verifier so
that the solver would inevitably exceed any memory or time limits imposed. This behavior is reported as failure of the
baseline verifier.
We ran all experiments on a machine with 16 2.8 GHz processors and 16 GB of RAM. The current implementations
of Lttp and the baseline verifier execute using a single thread.
6.2
Description of benchmarks
The behavior of the thirteen benchmarks programs tested is as follows.
The benchmarks programs buildInspect, peel, unary, and binary perform similar tasks, but with increasing
complexity for Lttp. All four programs construct a singly-linked list (using a next field) from front to back, with some
property, and then proceed to scan the list from front to back, ensuring that the property holds. In buildInspect and
peel, the property to check is that the final non-null element scanned is indeed the final element that was added to the
list. In order to model the traversal of this list, Lttp must synchronize the matching iterations of the two loops, as
depicted in Figure 3. The grammar relations here each have two negative control pairs, representing control flow of the
two loops. While buildInspect has the list initialization peeled out of the fist loop, in peel Lttp must discover that
this peel is needed to match the iterations correctly. A data field is added to unary and binary. The first checks a
unary property over the entire list (that all data fields are zero), while the second checks a binary property (that the
integer data is increasing monotonically). A program like unary can be found in the SV-COMP benchmark collection.
The verification of allocator constitutes a demonstration that the encoding of unique object allocation is correct.
After an unbounded list is constructed on the heap, a final object o is allocated and subsequently compared to every
item in the list to check uniqueness.
lag2 constructs a list and scans it within the same loop. However, a given item is scanned exactly two iterations
after it is created.
An unbounded list ending with a fixed-length cycle is built by finiteCycle; it then scans this structure and fails
the assertion if the scan ever terminates.
19
Alternatively, breakCycle constructs a cycle of unbounded size and proceeds to consume it by destroying the next
pointers. Thus it asserts that the traversal terminates when it revisits the node from which it started. SV-COMP contains
a cycle-constructing program as well. However, the SV-COMP version of the program is fundamentally easier than
breakCycle because the data is never overwritten; the starting point is instead recognized by its integer data.
We adapted sameLength directly from a program given by Itzhaky et al. [30] which cannot be handled by effectively
propositional reasoning (EPR). The program simultaneously constructs two disjoint lists, and then simultaneously
scans them to check that they have the same length. This property relies on an invariant correlating the two disjoint
structures, which is not expressible by heap reachability (see §6.3).
The tree benchmark constructs a non-list data structure. Each element has two fields, a left pointer and a right
pointer. A loop non-deterministically allocates a left or right child for the last item allocated and maintains a counter
for the size of the structure. The program then follows the pointers to the end, and asserts that the same number of
nodes were traversed.
Two more SV-COMP benchmarks which we adapted are simpleSearch and uniqueItem. Our simpleSearch
constructs an unbounded list of consecutive integers of length at least 5, beginning with zero. It then expects to find the
integers 1 and 3. The original benchmark built a list of at least ten elements. Our uniqueItem constructs an unbounded
list with all data fields set to False, except for exactly one element, which has its data set to True. The program
then checks that there is one and only one such element. The original benchmark performed a more complicated
construction, wherein the unique element is inserted after the list is constructed.
The program named order constructs the first part of a list, of unbounded length, adds two distinguished list
elements a and b, then constructs the remainder of a list, of unbounded length. The scanning loop verifies that it does
not find b before finding a.
Finally, ctxSensitive behaves like lag2 in that the scanning of elements lags behind their construction in the
same loop. However, two elements are constructed and only one is scanned in each iteration. Incremental integer data
is stored with each element, and the program asserts that the data value in final element scanned is half of the data
value in the final element constructed.
6.3
Illustrative benchmark
In this section, we discuss in detail an additional benchmark and the operation of Lttp in verifying its safety.
A program with queues of equal length: sameLength. Figure 5 contains a program sameLength that builds two
queues simultaneously in a loop, then traverses the queues to check that they contain the same number of elements.
This benchmark was given in a previous study of Effectively Propositional Reasoning (EPR) in order to illustrate a class
of programs that techniques based on EPR cannot prove safe because their necessary invariants cannot be represented
in Effectively Propositional Logic (EPL) [29].
sameLength performs the following steps over an execution. sameLength first initializes two queues (lines 3—7).
In each iteration of the loop (lines 8—16, referred to as the building loop), sameLength adds an element to the tail of
each queue. sameLength non-deterministically chooses to exit the building loop It then traverses both queues, loading
one element in each iteration of a loop at lines 17—19, referred to as the traversal loop. sameLength exits the traversal
loop when it reaches the end of one of the queues. Finally, sameLength asserts that it has reached the end of both
queues (line 20).
20
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
class Element { Element next; }
public static void sameLength() {
Element head1 = new Element();
Element tail1 = head1;
Element head2 = new Element();
Element tail2 = head2;
boolean done = false;
while (!done) {
done = *;
Element tmp;
if (done) tmp = null;
else tmp = new Element();
tail1.next = tmp; tail1 = tmp;
if (done) tmp = null;
else tmp = new Element();
tail2.next = tmp; tail2 = tmp; }
while (head1 != null && head2 != null) {
head1 = head1.next;
head2 = head2.next; }
assert head1 == null && head2 == null; }
Fig. 5. sameLength: Constructs two queues in a loop where
each iteration adds one element to each queue, then traverses both in a second loop to show both queues have the
same length.
b:8
0
1
Loops
d:
17
2
3
Fig. 6. ctxSensitive: A program which Lttp cannot prove
safe.
Loops[0]
sameLength[0]
a:3
class Element { Element next; int data; }
public static void contextSensitive() {
Element head = new Element();
head.data = 0;
Element tail = head;
int count = 1;
while (*) {
Element tmp = new Element();
tmp.data = count;
tail.next = tmp; tail = tmp;
tmp = new Element();
tmp.data = count+1;
tail.next = tmp; tail = tmp;
head = head.next;
count = count + 2; }
assert head.data == count/2; }
c:
17
0
e:
20
Loops[1]
n:8
o:
17
0
2
1
3
0
c:
10
d:
12
j:
17
f:8
g:
14
h:8
0 1
Loops
2
k:
19
l:
17
2 3
i:
17
1
3
m:
20
Fig. 7. A graph grammar that generates sameLength’s control paths and data dependencies.
In order to prove that sameLength satisfies its assertion, a verifier must establish an invariant for the loop on lines
17—19 that expresses the fact that the lengths of both of the maintained queues are equal. Such an invariant cannot be
represented in EPL [29].
Figure 7 depicts a run grammar G that simulates and refutes sameLength, and is synthesized automatically by Lttp.
G contains two relational predicates, sameLength and Loops. The query relational predicate, sameLength, is the head
of one clause, sameLength[0]. The clause derives paths that step from the entry of sameLength to the entry of its
building loop (a—b), the step from the exit of the building loop to the entry of the traversal loop (c—d), and from the
exit of the traversal loop to the exit of the program (e).
The relational predicate Loops is the head of two clauses. Clause Loops[0] derives termination of both the building
and traversal loops. Clause Loops[1] derives simultaneous, data-dependent iterations of the building loop and traversal
loop. The clause derives a control path f—g in an iteration of the building loop up to its store that adds an element
to the first queue, a control path g–h from the store that adds to the first queue to a store that adds to the second
queue, a control path j—k that steps in a corresponding iteration of the traversal loop up to the load from the head of
21
Benchmark
allocator
binary
buildInspect
ctxSensitive
finiteCycle
lag2
order
peel
tree
sameLength
breakCycle
simpleSearch
unary
uniqueItem
Source
Novel
EPR
SVC
LoC
Iter
Rel
Cls
SynSkeleton
Time
SolveCHC
Time
Lttp
Base
22
29
22
21
19
24
32
25
38
28
25
33
26
36
3
7
5
–
8
4
7
8
4
2
10
6
6
4
7
13
7
–
8
9
14
10
11
11
13
12
12
11
13
19
13
–
15
13
24
16
16
17
21
18
18
17
17s
1h 13m 32s
19s
–
59s
20s
8h 19m 18s
7m 4s
41m 50s
23m 22s
5h 0m 46s
3h 19m 38s
56m 27s
33m 26s
4m 30s
5m 25s
1m 18s
–
1m 18s
6s
4h 10m 2s
7m 47s
16m 18s
7m 54s
55m 30s
5h 28m 15s
6m 41s
40m 10s
✓
✓
✓
–
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
–
–
–
–
–
✓
–
–
–
–
–
–
–
–
Table 2. Results of our evaluation of Lttp. Information about the program source, features of the learned grammar, and execution
times of Lttp are reported. The columns of the table indicate the name of the benchmark (“Benchmark”), the source of the program
(“Source”), the number of lines of source code in the program (“LoC”), the number of iterations required to learn the final grammar
(“Iter”), the final number of relations and clauses in the final grammar (“Rel”, “Cls”), the total time spent in calls to the SynSkeleton
prototype, (“SynSkeleton Time”), the total time spent in calls to Duality (“SolveCHC Time”), and whether Lttp and the baseline
verifiers ultimately solved the verification problem (“Lttp”, “Base”).
the first queue, and a control path j—l that steps from the load from the head of the first queue to a load from the
head of the second queue. Figure 7 also depicts data dependencies from stores at instances of node g to corresponding
loads at instances of k and from stores at instances of h—l. As in previous examples, we omit rules which derive
non-corresponding steps of the loops, to simplify the presentation.
Loop is empty; one solution contains an interpretation of Loop as an invariant that establishes that the building loop
terminates if and only if both queues are null in a corresponding iteration of the traversal loop:
F(head, tail) ≡ tail0 = head2 ∧ next1 (tail0 ) = next2 (head2 ) =⇒ head3 = tail1
(done1 ⇐⇒ head13 = null ∧ head23 = null) ∧ F(head1, tail1) ∧ F(head2, tail2)
(2)
sameLength illustrates the ability of Lttp to prove safety of programs for which safety depends on properties that
relate disjoint unbounded, low-level data structures. Lttp does so by synthesizing invariants over relational predicates
that derive steps of execution that both load and store from related data structures. This ability enables Lttp to prove
the safety of programs that cannot be proved safe by existing techniques based on EPL [16, 29, 31]. These approaches
can prove properties expressed in terms of heap reachability within a linked list, but cannot verify properties that
require application-specific inductive definitions [29].
6.4
Results and analysis
The results of our evaluation are contained in Table 2. Each of the programs in Table 2 was proved safe by Lttp. Only
one of the benchmarks, lag2, was proved safe by the baseline verifier, which took three seconds.
SynSkeleton converges to the correct grammar after only 4 infeasible path examples in verifying buildInspect,
but requires 7 examples for convergence on peel. The SynSkeleton time for unary and binary is an order of
22
magnitude longer than for peel, which is in turn an order of magnitude longer than for buildInspect. This illustrates
the poor scalability of the constraint-based prototype as the program complexity increases.
Each time an object is allocated in allocator, the run grammar constraints can only ensure that the object is
distinct from a bounded number of other objects. Yet o is verified to be distinct from every element of the unbounded
list, indicating that the earlier allocations were aware of the eventual occurrence of o. This foresight is achieved because
the scanning of the list produces a grammar similar to that of buildInspect. Had the list not been scanned, o could
not be constrained to be distinct from the other elements, which is a safe approximation precisely because the list is not
scanned.
lag2 stands out as the only one which is determined to be safe by the baseline verifier, despite the fact that the
benchmark uses an unbounded heap structure. The baseline determines the program is safe in just three seconds. An
informal explanation is that though the entire structure is unbounded, only a bounded subset of it is live. The distance
between a store and its matching load is bounded at exactly two iterations of the loop. This special property permits the
invariants to be easily synthesized by Duality in the theory of arrays as a formula over a bounded number of indices.
This same property also manifests in Lttp, without the theory of arrays. The trivial first SynGrammar(P, ∅) result is
almost correct: a state variable must be kept in a relation for two iterations following its use in a control edge.
finiteCycle illustrates that Lttp can prove cyclicity of the visited control locations from the cyclicity of a data
structure.
Showing correctness of breakCycle depends on the ability of Lttp to accurately model a field which is written to
and read from more than once, with an unbounded number of heap modifications in between. If a verifier could not
prove that the nulled pointers remain null after traversing the unbounded cycle, there would be no guarantee of ending
at the correct element of the cycle.
Lttp is well-suited to verify sameLength, though it is inexpressible in EPR, because Lttp can capture correlations
between control flow and data movement in the structure of the CHC system.
Verifying tree shows that Lttp is not confined to lists since the data structure to be traversed proceeds nondeterministically through left and right child pointers.
Lttp solved simpleSearch, although its CHC system was the hardest one tested, taking 5.5 hours for Duality
to find inductive invariants. Because of the lower bound on the list size, the shortest of the feedback paths given to
SynSkeleton must execute the list construction step at least 5 times, since the vacuous refuting neighborhood is all
that is required to show infeasibility shorter paths. Moreover, it is speculated that the CHC solver must also explore
unwindings of the CHC system to that depth before meaningful invariants begin to be discovered.
Lttp is able to solve uniqueItem, illustrating its ability to model a program which can leave a special item at a
non-deterministic point in an unbounded list, and not only find it again, but prove that there is exactly one such item.
That Lttp can further solve order shows yet more expressivity in reachability conditions regarding multiple elements
in a list. A verifier would fail if it were only able to express that the elements a and b were reachable from the head of
list, and not that the elements are ordered.
ctxSensitive is unsolvable by both Lttp and the baseline verifier, despite its similarity to lag2, which both
verifiers performed very well on. Lttp did not synthesize a correct grammar no matter how many feedback paths it
received. The key difference from lag2 is that the distance between a load and its matching store increases with each
iteration. The limitation of Lttp in this case is discussed in below.
23
Limitations: While Lttp can verify a large class of programs which manipulate heap data structures, there are certain
types of programs which it cannot. In particular, the current implementation of Lttp is restricted to the subset of
programs which can be refuted using a context-free grammar.
Figure 6 contains the source code for a program, ctxSensitive, for which Lttp cannot synthesize a proof.
ctxSensitive constructs and consumes a queue in a single loop where each iteration appends two new elements
and consumes only one previous element. The program checks that the last element read was the middle element
written. First, the program initializes a queue with a single element and initializes a counter, count, to 1 (lines 3—6).
The program iterates an arbitrary number of times. In each iteration, two new elements are added to the tail of the
queue. The first element added has its data field set to count and the second has its data field set to count + 1 (lines
7—13). Then, the first element of the queue is removed by moving through the head’s next field (line 14). At the end
of each iteration count is incremented by 2 (line 15). Once the loop finishes, the program asserts that the data field of
the front of the queue is equal to half of the current value in count (line 16).
The problem with this benchmark is that the distance between corresponding stores and loads widens as the program
progresses. In other words, the data dependency is sensitive to the number of iterations of the loop that have occurred.
Therefore, the path-grammar required to describe the data dependencies of ctxSensitive is not context-free.
Non-context-free grammars are a fundamental limitation of the presented approach. Extending our technique with
the capability of handling more complex grammars is a compelling direction for future work.
7
RELATED WORK
Three-valued logic analysis (TVLA) represents sets of program heaps as canonical structures [35, 41, 45]. Recent work has
introduced automatic shape analyses that represent sets of program heaps as formulas in separation logic [12, 15, 42, 53],
forest automata [26], or memory graphs [17]. Lttp can be applied to programs that maintain perform arbitrary low-level
heap operations and data operations. Such approaches can potentially infer invariants of heaps that maintain a variety
of data structures, such as lists or trees. However, each analyses synthesizes invariants using an abstraction fixed for
the run of the analysis: thus, the analysis can only potentially be effective if an analysis designer provides a sufficient
set of predicates for maintaining invariants required to prove that a program satisfies a desired property. Variants of
TVLA have been extended to automatically refine an abstraction used by applying Inductive Logic Programming (ILP).
However, previous work has established only have ILP can be applied to refine structural abstractions, not abstractions
over structure and data.
Recent work has also proposed decision procedures for separation logic [39, 46] and an automatic verifier that
represents sets of states as separation-logic formulas [1]. Such approaches require a fixed set of recursively-defined
predicates for reasoning about fixed classes of heap data structures. Lttp can verify program correctness without
requiring such predicates.
Previous work has developed automatic verifiers that implement predicate abstraction in shape logics. Such approaches can only synthesize invariants that describe which cells of a heap may reach each other over heap fields [4, 14],
can be applied only to programs that maintain particular data structures—such as linked lists [34]—or require an
analyst to provide loop invariants [10], predicates over which invariants are constructed [19, 40, 43], or heuristics that
ensure that the analysis converges [14]. Relational invariants of path grammars can prove correctness of programs
that maintain non-list data structures. Lttp can potentially synthesize such invariants without requiring an analyst to
provide predicates or loop invariants.
24
Previous work has proposed verifiers that determine if a program satisfies an assertion by inferring shape invariants
represented as formulas in effectively-propositional logics [16, 29, 31]. Such verifiers enjoy strong completeness
properties not satisfied by Lttp, but the class of invariants that they synthesize cannot express invariants over non-list
data structures or that relate multiple lists. Relational invariants over path grammars can prove the safety of programs
that maintain such data structures, and Lttp can potentially synthesize such invariants automatically.
One such verifier [31] is in fact one instance of a verification framework that could potentially be instantiated
with logics other than effectively-proposition logics. However, the general framework requires an analyst to provide
predicates over which verification is to be performed. Lttp does not require such predicates to be provided.
Further work on effectively-propositional logic describes a verifier that can verify properties of heap-paths of
data-structures that are not necessarily lists [28]. However, the verifier requires pre-conditions and post-conditions to
be given explicitly. It cannot verify programs with loops, although loops annotated with invariants could presumably
be verified by directly adapting the verification technique. In either case, Lttp is distinct from the verifier in that Lttp
can potentially prove safety of an iterative program automatically.
Previous work has described verifiers that verify that a given program satisfies a desired shape property by axiomatizing and inferring invariants in the theory of arrays. One approach attempts to infer invariants that are quantified
over array indices and use range predicates that describe the values at different ranges of indices in an array [33].
Such invariants are well-suited to inferring invariants that describe logical arrays that model arrays operated on by a
program, but not logical arrays that model a program’s heap fields.
Previous work has proposed approaches that attempt to verify a given program by synthesizing a tree-decomposition
of the heaps that it maintains [36, 37]. Lttp is similar to such approaches in that it reasons about tree-structured
artifacts that model program executions, namely the derivation trees of a grammar of program paths. Unlike previous
approaches, Lttp reasons about the derivation of control paths, rather than of a decomposition of the heap.
Techniques from relational verification [6–8] establish properties over states in multiple runs of a program, such as
robustness and information-flow security, or over states in runs of multiple programs, such as observational equivalence.
Lttp can be viewed as an instance of relational verification, in that it attempts to synthesize proofs that establish
properties over multiple states. Lttp is distinct from previous work in that it attempts to establish properties of
states within the same run in order to prove safety of low-level programs. Combining existing relational verification
techniques with Lttp in order to prove relational properties of programs that maintain low-level data structures and to
automate relational verification seems to be a promising direction for future work.
Sketch synthesizes finite programs [50], bit-streaming programs [50], and stencils [47] by iteratively synthesizing a
candidate version of a program, attempting to verify it, and using a counterexample to verification to guide the search
for the next version of the program. Synthesizers that are instances of the SyGuS framework [2] attempt to synthesize
a program accompanied with a proof of correctness, using a counterexample-guided, iterative process. Lttp is similar
to Sketch and instances of SyGuS in that it is an inductive synthesizer. Lttp is distinct from Sketch and instances of
SyGuS in that each of them, given a partially-complete program, attempt to complete the program with instances of a
language of possible syntactic combinations of atomic operations. Lttp, given a program P, attempts to synthesize the
basic structure of a logic program that simulates P, without synthesizing new atomic operations.
Previous work has reduced the problem of verifying concurrent programs that use a bounded number of threads
to solving a system of Constrained Horn Clauses (CHC’s) [22] and proposed a solver for CHC’s over the theory of
linear arithmetic. Previous work has developed automatic verifiers for programs with a single procedure [38] and
25
multiple recursive procedures [23] that have been generalized to design CHC solvers that use an interpolating theorem
prover [9, 44]. Lttp uses a CHC solver for the theory of uninterpreted functions as a black box. In principle, Lttp can
use any CHC solver; the implementation evaluated uses a solver that itself uses an interpolating theorem prover [9].
8
CONCLUSION
In this paper, we have presented a novel verifier, named Lttp, designed to verify programs that maintain low-level
data structures. The key feature of Lttp is that, given a program P, it attempts to synthesize a proof of the safety of P,
represented as a graph grammar, i.e. a CHC system, that generates the control paths of P, annotated with invariants
that relate the local values at multiple points in each path. Lttp is completely automated, using an inductive-synthesis
algorithm that synthesizes candidate path grammars by reduction to constraint solving and validates candidate grammars
by reduction to logic programming. Such proofs can establish correctness of programs that previously could only be
proven correct using manually-provided predicates over values and structure.
We have implemented Lttp as a verifier for JVM bytecode and evaluated it on a set of challenging problems for
shape verifiers. Lttp for JVM succeeded in learning a suitable grammar for thirteen of the fourteen benchmarks. Each
of the suitable grammars was given as a CHC system to the CHC solver Duality, which was able to find inductive
invariants to prove program safety.
The design of Lttp establishes that shape-verification problems can potentially be solved by applying techniques from
relational verification and inductive synthesis. Further work strengthening this connection could result in significant
development of each of the related fields.
ACKNOWLEDGMENTS
This work is supported by the National Science Foundation under awards 1526211 and 1650044.
REFERENCES
[1] A. Albarghouthi, J. Berdine, B. Cook, and Z. Kincaid. Spatial interpolants. In ESOP, 2015.
[2] R. Alur, R. Bodı́k, G. Juniwal, M. M. K. Martin, M. Raghothaman, S. A. Seshia, R. Singh, A. Solar-Lezama, E. Torlak, and A. Udupa. Syntax-guided
synthesis. In FMCAD, 2013.
[3] D. Angluin. Learning regular sets from queries and counterexamples. InfComp, 75(2):87–106, 1987.
[4] I. Balaban, A. Pnueli, and L. D. Zuck. Shape analysis by predicate abstraction. In VMCAI, 2005.
[5] T. Ball, V. Levin, and S. K. Rajamani. A decade of software model checking with SLAM. Commun. ACM, 54(7), 2011.
[6] G. Barthe, J. M. Crespo, and C. Kunz. Relational verification using product programs. In FM, 2011.
[7] G. Barthe, J. M. Crespo, and C. Kunz. Beyond 2-safety: Asymmetric product programs for relational program verification. In LFCS, 2013.
[8] N. Benton. Simple relational correctness proofs for static analyses and program transformations. In POPL, 2004.
[9] N. Bjørner, K. L. McMillan, and A. Rybalchenko. On solving universally quantified horn clauses. In SAS, 2013.
[10] C. Bouillaguet, V. Kuncak, T. Wies, K. Zee, and M. C. Rinard. Using first-order theorem provers in the jahob data structure verification system. In
VMCAI, 2007.
[11] F. Bourdoncle. Efficient chaotic iteration strategies with widenings. In Formal Methods, pages 128–141, 1993.
[12] C. Calcagno, D. Distefano, P. W. O’Hearn, and H. Yang. Compositional shape analysis by means of bi-abduction. J. ACM, 58(6), 2011.
[13] A. Clark. Distributional learning of some context-free languages with a minimally adequate teacher. In ICGI, volume 6339, pages 24–37, 2010.
[14] D. Dams and K. S. Namjoshi. Shape analysis through predicate abstraction and model checking. In VMCAI, 2003.
[15] D. Distefano, P. W. O’Hearn, and H. Yang. A local shape analysis based on separation logic. In TACAS, 2006.
[16] S. Drews and A. Albarghouthi. Effectively propositional interpolants. In CAV, 2016.
[17] K. Dudka, P. Peringer, and T. Vojnar. Byte-precise verification of low-level list manipulation. In LNCS, 2013.
[18] C. Flanagan. Automatic software model checking using CLP. In ESOP, 2003.
[19] C. Flanagan and S. Qadeer. Predicate abstraction for software verification. In POPL, 2002.
[20] Github - sosy-lab/sv-benchmarks: svcomp17. https://github.com/sosy-lab/sv-benchmarks, 2017. Accessed: 2017 July 1.
[21] S. Gulwani, S. Jha, A. Tiwari, and R. Venkatesan. Synthesis of loop-free programs. In PLDI, 2011.
26
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
[46]
[47]
[48]
[49]
[50]
[51]
[52]
[53]
[54]
[55]
[56]
A. Gupta, C. Popeea, and A. Rybalchenko. Predicate abstraction and refinement for verifying multi-threaded programs. In POPL, 2011.
M. Heizmann, J. Hoenicke, and A. Podelski. Nested interpolants. In POPL, 2010.
T. A. Henzinger, R. Jhala, R. Majumdar, and K. L. McMillan. Abstractions from proofs. In POPL, 2004.
T. A. Henzinger, R. Jhala, R. Majumdar, and G. Sutre. Lazy abstraction. In POPL, 2002.
L. Holı́k, O. L. O, A. Rogalewicz, J. Šimáček, and T. Vojnar. Fully automated shape analysis based on forest automata. In CAV, 2013.
L. Hubert, N. Barré, F. Besson, D. Demange, T. P. Jensen, V. Monfort, D. Pichardie, and T. Turpin. Sawja: Static analysis workshop for Java. In
FoVeOOS, 2010.
S. Itzhaky, A. Banerjee, N. Immerman, O. Lahav, A. Nanevski, and M. Sagiv. Modular reasoning about heap paths via effectively propositional
formulas. In POPL, 2014.
S. Itzhaky, A. Banerjee, N. Immerman, A. Nanevski, and M. Sagiv. Effectively-propositional reasoning about reachability in linked data structures.
In CAV, 2013.
S. Itzhaky, A. Banerjee, N. Immerman, A. Nanevski, and M. Sagiv. Effectively-propositional reasoning about reachability in linked data structures.
Technical report, Tel Aviv University, 2013.
S. Itzhaky, N. Bjørner, T. W. Reps, M. Sagiv, and A. V. Thakur. Property-directed shape analysis. In CAV, 2014.
S. Izhaky, S. Gulwani, N. Immerman, and M. Sagiv. A simple inductive synthesis methodology and its applications. In OOPSLA, 2010.
R. Jhala and K. L. McMillan. Array abstractions from proofs. In CAV, 2007.
S. K. Lahiri and S. Qadeer. Verifying properties of well-founded linked lists. In POPL, 2006.
A. Loginov, T. W. Reps, and S. Sagiv. Abstraction refinement via inductive learning. In CAV, 2005.
R. Manevich, J. Berdine, B. Cook, G. Ramalingam, and M. Sagiv. Shape analysis by graph decomposition. In TACAS, 2007.
R. Manevich, T. Lev-Ami, M. Sagiv, G. Ramalingam, and J. Berdine. Heap decomposition for concurrent shape analysis. In SAS, 2008.
K. L. McMillan. Lazy abstraction with interpolants. In CAV, 2006.
J. A. N. Pérez and A. Rybalchenko. Separation logic + superposition calculus = heap theorem prover. In PLDI, 2011.
Z. Rakamaric, J. D. Bingham, and A. J. Hu. An inference-rule-based decision procedure for verification of heap-manipulating programs with mutable
data and cyclic data structures. In VMCAI, 2007.
T. W. Reps, S. Sagiv, and R. Wilhelm. Static program analysis via 3-valued logic. In CAV, 2004.
J. C. Reynolds. Separation logic: A logic for shared mutable data structures. In LICS, 2002.
P. M. Rondon, M. Kawaguchi, and R. Jhala. Low-level liquid types. In POPL, 2010.
P. Rümmer, H. Hojjat, and V. Kuncak. Disjunctive interpolants for horn-clause verification. In CAV, 2013.
S. Sagiv, T. W. Reps, and R. Wilhelm. Parametric shape analysis via 3-valued logic. ACM Trans. Program. Lang. Syst., 24(3), 2002.
S. A. Seshia, S. K. Lahiri, and R. E. Bryant. A hybrid SAT-based decision procedure for separation logic with uninterpreted functions. In DAC, 2003.
A. Solar-Lezama, G. Arnold, L. Tancau, R. Bodı́k, V. A. Saraswat, and S. A. Seshia. Sketching stencils. In PLDI, 2007.
A. Solar-Lezama, C. G. Jones, and R. Bodı́k. Sketching concurrent data structures. In PLDI, 2008.
A. Solar-Lezama, R. M. Rabbah, R. Bodı́k, and K. Ebcioglu. Programming by sketching for bit-streaming programs. In PLDI, 2005.
A. Solar-Lezama, L. Tancau, R. Bodı́k, S. A. Seshia, and V. A. Saraswat. Combinatorial sketching for finite programs. In ASPLOS, 2006.
S. Srivastava, S. Gulwani, S. Chaudhuri, and J. Foster. Path-based inductive synthesis for program inversion. In PLDI, 2011.
Xsb datalog. https://http://xsb.sourceforge.net/, 2017. Accessed: 2017 July 1.
H. Yang, O. Lee, J. Berdine, C. Calcagno, B. Cook, D. Distefano, and P. W. O’Hearn. Scalable shape analysis for systems code. In CAV, 2008.
R. Yoshinaka. Towards dual approaches for learning context-free grammars based on syntactic concept lattices. In Developments in Language Theory,
pages 429–440, 2011.
R. Yoshinaka and A. Clark. Polynomial time learning of some multiple context-free languages with a minimally adequate teacher. In Formal
Grammar, pages 192–207, 2012.
Z3prover/z3 - github. https://github.com/Z3Prover/z3, 2017. Accessed: 2017 July 1.
27
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
class Element { Element next; }
enum St { W, RW, R, DONE; }
public static void lag() {
Element head = new Element();
Element tail = head;
St s = St.W;
while (s != St.DONE) {
if (s == St.W || s == St.RW) {
Element tmp = new Element();
tail.next = tmp;
tail = tmp; }
if (s == St.RW || s == St.R) { head = head.next; }
if (s == St.W) { s = St.RW; }
else if (s == St.RW && *) { s = St.R; }
else if (s == St.R) { s = St.DONE; } }
assert head == tail; }
Fig. 8. lag: Constructs and consumes a queue, with each load
occurring one iteration after the corresponding store.
c:
12
d:7
1
Loop[1]
k:
12
0
Loop
a:3
Fig. 9. order: Constructs and queue with two distinguished elements. It then consumes the queue to
ensure the distinguished elements remain in the order
they were inserted.
Loop[0]
lag[0]
b:7
class Element { Element next; }
public static void order() {
Element head = new Element();
Element tail = head;
while (*) {
Element tmp = new Element();
tail.next = tmp; tail = tmp; }
Element a = new Element();
tail.next = a; tail = a;
Element b = new Element();
tail.next = b; tail = b;
while (*) {
Element tmp = new Element();
tail.next = tmp; tail = tmp; }
boolean ordered = true;
while (head != a) {
if (head == b) ordered = false;
head = head.next; }
assert ordered; }
2
0
1
0
e:
16
g:7
2
m:
16
l:7
h:
12
1
0
f:
12
0
Loop
i:7
1
2
2
j:
16
Fig. 10. A graph grammar that generates lag’s control paths and data dependencies.
A
ADDITIONAL ILLUSTRATIVE BENCHMARKS
A program with nodes shared across clauses: lag. Figure 8 contains the source code for a program, lag, which builds and consumes
a queue. Both the loads and stores occur in a single loop, where each item that is loaded was stored on the previous iteration. The
program operates in the context of a small state machine. First, lag initializes the queue and sets itself in a write-only state (lines
4—6). Next, if lag is in a write state, then it adds a new element to the end of the list and sets the tail to this new element (lines 7—11).
If lag is in a read state, it moves the head to the next element (line 12). lag then performs state transitions (lines 13—15). lag moves
from the write-only state to the read-write state and from the read-only state to the done state after a single iteration. lag moves from
the read-write state to the read-only state after an arbitrary number of iterations. Finally, lag asserts that the head element reached by
this loop is the same as the tail (line 16).
Figure 10 depicts a graph grammar, G, which describes the control paths and data dependencies of lag. The grammar contains
two relations, lag and Loop, where lag is the query relation. G has three clauses. lag[0] generates fragments of the control path
which describe the loop entry (a—b), through the first store (b—c), and through the first state transition without performing a load
(c—d). Loop[0] generates fragments of the control path which store the next item in the queue (g—h), load the front of the queue, and
perform a state transition (h—i). In addition, Loop[0] contains a data dependency (f—i): The store from the previous loop iteration
28
Loop0[0]
order[0]
a:3
b:5
0
1
Loop0
2
h:
19
c:8
0
g:
16
2
3
0
1
d:
12
Loop1
3
1
f:
16
e:
15
Loop0[1]
o:5
p:8
i:5
j:5
0
1
0
2
3
2
q:
16
r:
16
l:
16
0
Loop1[0,1]
1
Loop0
m:
16
2
3
k:8
v:
12
0
1
1
3
0
n:
16
s:
12
Loop1
t:
12
0
1
w:
15
1
u:
15
Fig. 11. A graph grammar that generates order’s control paths and data dependencies.
provides the data for the load in this iteration. Loop[1] generates the final fragment which skips from the beginning of the loop to just
after the loop (l—m).
A noteworthy feature of G is node f. In particular, f provides a data dependency, but is not connected directly to the control path
through Loop[0]. This demonstrates the capability of graph grammars to separate data dependencies from control dependencies.
A program for which some heap operations have no data dependencies: order. Figure 9 contains the source code for a program,
order, which creates a queue with two distinguished elements, then consumes the queue until it can check these two elements remain
relatively ordered. First, order initializes the queue and adds an arbitrary number of elements to the tail (lines 3—7). Next, it inserts
two distinguished elements, a and b (lines 8—11). order continues to expand the queue with an arbitrary number of new elements
(lines 12—14). order traverses the queue until it reaches element a. If b is encountered before the end of this loop, then the queue has
become unordered (lines 15—18). Finally, order asserts that the distinguished elements remain in order (line 19).
Each execution of order satisfies the assertion. The key invariant which could prove the safety property establishes that for each
iteration of the loop on lines 12—14, traversing the next pointer reaches a before it reaches b. This invariant is difficult to express
without knowing particular facts about the relationships of linked lists. However, a graph grammar can express the idea of this
invariant by simulating the first and third program loops together.
Figure 11 depicts a graph grammar, G, which encodes order. G has three relations and five clauses. The query relation, order, has
only one clause, order[0], which contains the entry to the first program loop (a—b), the insertion of the two distinguished elements
and entry of the second loop (c—d), the entry to the third loop (e—f), a final execution of the third loop and the program assertion
(g—h), and a data link between the first distinguished element and the final load (d—h). The two clauses of Loop0, Loop0[0] and
Loop0[1], describe the simultaneous execution of the first and third program loops. These clauses are similar to StLd[0] and StLd[1]
from §2.2.1. The two clauses of Loop1, Loop1[0] and Loop1[1] describe the execution of the second program loop. For brevity, we
omit some clauses for Loop0 that iterate the two loops a differing number of times. We annotate Loop0 with an inductive relational
invariant which can prove the safety of order:
tail1 , a2 ∧ head3 = tail1 ∧ ordered3 ∧ tail0 = head2 ∧
next1 (tail0 ) = next2 (head2 ) =⇒ head3 = tail1
(3)
This invariant maintains that the flag ordered is always true at the end of third loop. In addition, it establishes that the simultaneous
iteration of the first and third loop terminates before the distinguished element a is reached. Interestingly, the second loop is irrelevant
to proving the safety property, so the relational invariant for Loop1 is simply t r ue.
Previous examples have all contained data dependencies for every store and load. order demonstrates that this is not a limitation of
Lttp. The stores performed in the second program loop have no corresponding loads. This is reflected by the lack of data dependencies
in the clauses of the relation Loop1.
29
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
class Tree { Tree left; Tree right; }
public static void tree () {
Tree root = new Tree();
Tree t = root;
int c = 0;
Tree l,r,tmp;
while (t != null) {
c++;
if (*) tmp = new Tree();
else tmp = null;
if (*) l = tmp; r = null;
else l = null; r = tmp;
t.left = l; t.right = r;
t = tmp; }
t = root;
int d = 0;
do {
d++;
l = t.left; r = t.right;
if (l != null) t = l; else t = r;
} while (l != null || r != null);
assert c == d; }
Fig. 12. tree: Constructs a tree by arbitrarily adding left or right nodes, then follows the path to ensure the length when consumed
is the same as the length of the constructed path.
A program with a tree data structure: tree. Figure 12 contains a program, named tree, that constructs a tree data structure.
tree builds the tree by non-deterministically choosing to continue to build the left or right subtree of a given node for a nondeterministically chosen number of iterations. tree then traverses the data structure and ensures that the number of steps taken to
traverse it is the same as the number of steps used to build it.
In particular, tree enters the loop in lines 7—14 (i.e., it’s building loop) by initializing the tree and initializing a counter c of the
number of nodes. Before each iteration of the loop, tree tests if the tree node bound to t is null, and if so, exits the loop (line 7). In
each iteration of the loop, tree increments c (line 8) and non-deterministically chooses whether to stop adding new nodes (lines
9—10). The new node is then stored in either the left or right field of the node bound to t (lines 11—13).
After executing the loop, tree binds t to the root of the build tree (line 15) and initializes a new counter d to 0 (line 16). When
tree executes the loop on lines 17—21 (i.e., its traversal loop), it traverses the constructed path by loading either the left or right
child of a node maintained in t, depending on which child is not null (lines 17—20). tree executes its traversal loop until it reaches
a node such that both children are null (line 21). tree then asserts that the number of nodes added is the same as the number of
nodes traversed (line 22).
tree satisfies the safety property tested by its assertion at line 22 However, the loop invariants required to prove the property must
express a complex property that relates data structure and values. Such invariants must establish that the tree contains c elements
while tree executes the loops in lines 7—14, and that c − d elements are reachable from the object in t while tree executes the loop
in lines 17—21. To our knowledge, no automated invariants can synthesize proofs in a language that can express such properties.
Figure 13 depicts a run grammar G that simulates and refutes tree. G contains two relational predicates, tree and Loops. The
query relational predicate, tree, is the head of clause tree[0]. Clause tree[0] derives a control path that steps from program entry to
the entry of tree’s building loop (a—b), a path that steps from the exit of the building loop to the entry of the traversal loop (c—d),
and a path that steps from the exit of the traversal loop to the assertion (e).
Loops is the head of three different clauses. Clause Loops[0] derives steps in which the building loop non-deterministically chooses
to exit; in the corresponding step of traversal loop, it determines that it has reached the end of the tree. Clause Loops[1] derives
simultaneous steps of both loops such that the building loop chooses to build a subtree from the left field of the maintained tree
node. f—g derives a path through the building loop, while i—j depicts the corresponding path through the traversal loop. Each path
point derived as an instance of g is a data dependence of the node derived as an instance of j in the same clause instance, depicted by
the data-dependence edge from g to j. Clause Loops[2] is similar to Loops[0], but derives corresponding steps of both loops in which
the building loop builds a tree at the right field of the maintained tree node. Similar to the presentation of buildInspect in §2, we
30
Loops[0]
tree[0]
a:3
b:7
0
1
c:
15
Loops[1]
r:7
s:
15
f:7
0
1
0
g:7
2
3
0
2
3
2
e:
22
t:
17
u:
22
o:
17
1
left Loops
Loops
d:
17
0
Loops[2]
p:
17
2
3
n:
13
l:7
1
0
3
2
q:
22
o:
17
m:
7
0
1
right Loops
p:
17
2
3
n:
13
1
3
q:
22
Fig. 13. A graphical depiction of a run grammar that simulates and refutes tree. Control and data dependencies between path
points are depicted similarly to Figure 3.
have omitted clauses that derive extraneous iterations of either the building or traversal loops. Such clauses are necessary to represent
a run grammar that simulates all control flow paths of tree, since it is not known a priori that every execution has an equal number
of iteration in each loop.
G simulates tree and is empty; one solution of G includes an interpretation of Loops as a relational invariant that establishes that
c and d have the equal values at the end of iterations on which they store to and load from the same field:
lrnull(ix 0, ix 1) ≡ lefti x 0 (ti x 1 ) = null ∧ righti x 0 (ti x 1 ) = null
c1 = d3 ∧ (lrnull(1, 0) ⇐⇒ lrnull(2, 2))
(4)
In this invariant we define a formula macro lrnull over two indices which specifies that the left and right fields of the variable t
are both null at the given indices. To our knowledge, no automatic shape verifier developed in previous work can prove the safety of
tree without being guided to use a manually-defined recursive predicate that describes the shape of trees, and relevant predicates
over data variables c and d. However, Lttp can prove the safety of tree automatically.
B
IMPLEMENTING SYNSKELETON BY REDUCTION TO CONSTRAINT SOLVING
We now formalize some concepts that will assist in proving the lemmas and explaining more precisely how the constraints used
to implement SynSkeleton work. In the following sections, let a program P ∈ Lang be given. Let C be a CHC system, and let
the space to which C belongs be denoted CHCsR, X . Assume, without loss of generality, that the arity of every R ∈ R is n. Let
Idxs = {i ∈ N | i ≤ n }. Let α (Y, i) be the ith item in the sequence Y for Y ∈ X n and i ∈ Idxs. Further, for every control location
l ∈ Locs[P], let Ll be a unary interpreted function. Further, let Ctrl, Data, and Conn be binary uninterpreted functions.
Location consistency. Suppose there exists λ̂ : R × Idxs → Locs with the following property for all clauses C = (B, φ, H ) ∈ C, all
applications (R, Y ) ∈ B ∪ {H }, and all indices i ∈ Idxs:
Û
φ |=
(l = λ̂(R, i) ⇐⇒ Ll (α (Y, i))).
l ∈Locs
Then C is location consistent. Also define the map λ̂C : X → Locs for each clause C = (B, φ, H ) ∈ C such that if φ |= Ll (x ) then
λ̂C (x ) = l , for all variables x ∈ X and locations l ∈ Locs. λ̂C is well-defined because C is location consistent.
Control consistency. Suppose C is location consistent. Let λ̂C be given for all clauses C ∈ C. Suppose there exists ÊC ⊆ X × X for
each clause C = (B, φ, H ) ∈ C with the following properties for all variables x 1, x 2 ∈ X :
(1) If there does not exist an instruction Instr[P](λ̂C (x 1 ), λ̂C (x 2 )), then φ |= ¬Ctrl(x 1, x 2 );
(2) φ |= ÊC (x 1, x 2 ) ⇐⇒ Ctrl(x 1, x 2 ).
Then C is control consistent.
31
Path consistency. Suppose C is control consistent, and let ÊC be given for all clauses C ∈ C Suppose there exist a total ordering
< R over Idxs and a map sgnR : Idxs → {+, −} for every relation R ∈ R, such that the following properties hold for all clauses
C = (B, φ, H ) ∈ C, all applications (R 1, Y1 ), (R 2, Y2 ), ∈ B ∪ {H }, all indices i, j, k ∈ Idxs, and all variables x 1, x 2, x 3 ∈ X :
(1) If (α (Y1, i), α (Y2, j)) ∈ ÊC then φ |= Conn(α (Y1, i), α (Y2, j)) and
(a) if also (R 1, Y1 ) = H then φ |= sgnR 1 (i) < 0,
(b) if also (R 1, Y1 ) ∈ B then φ |= sgnR 1 (i) > 0;
(2) If (α (Y1, i), α (Y2, j)) < ÊC and if φ |= α (Y1, i) = α (Y2, j) then sgnR 1 (i) = sgnR 2 (j);
(3) φ |= Conn(x 1, x 2 ) ∧ Conn(x 2, x 3 ) =⇒ Conn(x 1, x 3 );
(4) φ |= i < R 1 j ⇐⇒ Conn(α (Y1, i), α (Y1, j)).
Then C is path consistent.
For a path consistent C, define sets of index pairs ⇝+R and ⇝−R for each relation R ∈ R as follows. Let ⇝+R = {(i, j) | sgnR (i) >
0 ∧ i < R j ∧ k ∈ Idxs(i < R k < R j)} ⊆ Idxs × Idxs. Define ⇝−R similarly, but with sgnR (i) < 0. The sets ⇝+R and ⇝−R are the
positive and negative control pairs of R, respectively. Informally, at any node d of any derivation tree D of C, with head (R, Y ), the
control path between α (Y, i) and α (Y, j) is constructed entirely within the subtree of D rooted at d if i ⇝−R j. Similarly, the control
path is constructed entirely outside of the subtree of D rooted at d if i ⇝+R j. For convenience, define the application of ⇝+R to
variable tuple Y ∈ X n as the set of variable pairs ⇝+R (Y ) = {(α (Y, i), α (Y, j)) | i ⇝+R j, i, j ∈ Idxs} ⊆ X × X . Define ⇝−R (Y )
similarly.
Neighborliness. Let C be such that for each C ∈ C and all x 1, x 2 ∈ X ,
φ |= Data(x 1, x 2 ).
Then C is neighborly.
Completeness and Correctness. If every path p ∈ Paths[P] is induced as the control graph of some model of C, then C is complete
for P.
Let p ∈ Paths[P] be infeasible. If each derivation of C that induces p as a control graph also induces a ν a refuting neighborhood of
p, then C is correct for p. If C is complete for P and correct for all infeasible paths p ∈ Paths[P], then C is correct for P.
There is a direct correspondence between the models m ⊢ SymPath(P, p, ν All ) and runs of p. Model m corresponds to run (p, σ )
with σ defined by σ (n) = (γ n , U n ), where γ n (a) = m(a(σn )) and U n (f, x) = n ′ where m(f(tnow (σn ), x(σn ))) = m(tnow (σn ′ )) if
any such n ′ exists, for all n ∈ N , a ∈ Vars, x ∈ OVars, f ∈ Fields.
Generating a skeleton from constraint solutions. The prototype implementation of SynSkeleton is a constraint-based approach to
grammar synthesis. We encode a space G of grammars using uninterpreted functions to model the structure of relations and clauses.
With loss of generality, the prototype only considers linear grammars, in the sense that every clause has at most one relation in the
body. Lttp has been formulated for only single procedure programs, whose control flow can be described linearly, and a need for
non-linear grammars has not yet been encountered in Lttp. The prototype also requires as input a budget specifying the maximum
arity of relations and the maximum number of relations for systems in G. This configuration is encoded as the set of constants and
uninterpreted functions of an SMT query.
The following properties of every C ∈ G are achieved by SMT constraints:
(1) Location consistency: This gives rise to λ̂C : X → Locs for each clause C.
(2) Control consistency: WLOG, each clause C entails exactly one control edge, i.e., ÊC = {(a, b)}.
(3) Path consistency: The arguments of every relation are partitioned into some number of negative control pairs and some
number of single auxiliary variables. Let (R 2, φ, R 1 ) be a clause. Under path consistency, R 1 is responsible for completing the
control paths between its negative pairs, and R 2 inherits what is not explicitly completed by R 1 . Assume that R 1 constructs
the control edge (a, b) between negative pair (a ′, b ′ ). If (a, b) = (a ′, b ′ ), then this pair is completed and a, b become
32
auxiliary variables in R 2 . If b , b ′ , then only a ′ becomes auxiliary and a ⇝−R b ′ . Likewise for a , a ′ . If both a , a ′ and
2
b , b ′ , then the control pair is split in two in R 2 : a ′ ⇝−R a ⇝+R b ⇝−R b ′ . Auxiliary variables (those not in any negative
2
2
2
pair) may or may not be forwarded to R 2 . In the general case, for x ⇝+R y ⇝+R z, when y is not forwarded, x ⇝+R z.
1
1
2
(4) Completeness as a path grammar: For each clause with control edge (a, b), λ̂C (a) is uniquely determined by the head
relation. For each relation R, there is (at least) one clause with head R for each possible value of λ̂C (b). The query relation
contains the unique (WLOG) initial and final control locations as a negative control pair. By induction over the clauses, each
R completes all possible paths between every negative control pair.
(5) Unambiguity (with loss of generality): For each relation R, there is at most one clause with head R for each possible value of
λ̂C (b), guaranteeing that every control path p has a unique derivation.
(6) Correctness for a path p ∈ F : A collection of uninterpreted functions witnesses a derivation of p by the grammar and further
witnesses that the neighborhood ν = Deps(P, p) is achieved, assuming neighborliness. That this holds for all derivations of p
is trivial with unambiguity.
C
PROOFS OF LEMMAS AND THEOREMS
Lemma 1. If P is simulated by S and P is refuted by S, then P is safe (§3.1.2).
Proof. S has no model that is a run P, by the assumption that P is refuted by S. Therefore P has no run by the assumption that P
is simulated by S. Thus P is safe, by the definition of safety (§3.1.2).
□
Lemma 2. If p is a feasible path of P, then IsFeas(P, p, ν All ) = True. Otherwise, IsFeas(P, p, ν All ) = False.
Proof. Sketch. IsFeas uses the formula SymPath(P, p, ν All ), which is a conjunction of constraints which precisely model the
semantics of each instruction in p. Notably, for every load there is at most one matching store. Every load is constrained with the
value of its matching store since the neighborhood ν All allows for every control state of p to be examined in search of the unique
match. Thus, the semantics of Lang are not approximated, and SymPath(P, p, ν All ) is satisfiable exactly when p has a run.
□
Lemma 4. Let p = (N , λ, E) ∈ Paths[P] and a neighborhood ν 0 : N → P(N ) be given. SymPath(P, p, ν 0 ) |= SymPath(P, p, ν All ).
Proof. SymPath(P, p, ν ) is a conjunction of constraints modeling the instruction on each edge (u, v) ∈ E, which is SymRel[i](u, v, ν (v))
for the instruction i = Instr[P](λ(u), λ(v)). For every n ∈ N , ν 0 (n) ⊆ ν All (N ) = N . We have that SymRel[i](u, v, ν 0 (v)) |=
SymRel[i](u, v, N ), shown case-wise for every possible instruction i:
• If i ∈ InstrsV , then SymRel[i](u, v, Q ) does not depend on the set of states Q .
• If i ∈ Stores ∪ Loads ∪ Allocs, then SymRel[i](u, v, Q ) takes only a conjunction over the q ∈ Q . Since ν 0 (v) ⊆ N ,
SymRel[i](u, v, ν 0 (v)) is no stronger than SymRel[i](u, v, N ).
Since SymPath is only a conjunction of SymRel constraints, SymPath(P, p, ν 0 ) cannot be any stronger than SymPath(P, p, ν All ). □
Lemma 5. Let a control path p = (N , λ, E) ∈ Paths[P] and any neighborhood graph (N , E Data ) be given. Let ν : N → P(N ) be the
neighborhood induced by (N , E Data ). Suppose p with ν are modeled by some derivation D of a skeleton C. Let G be the run grammar
corresponding to C. Then there is a model of D in G iff there is a model of SymPath(P, p, ν ).
Proof. Let a model of D in G be given as (D, m D , i), where m D is a model of the background theory and i : Nodes[D] →
Models[X ] maps derivation nodes to models of the CHC variables X . Since C, a skeleton, is control consistent by the specification
of SynSkeleton, every control edge (u, v) ∈ E = m D (Ctrl) via the constraint φ of the clause C of some node d in D, as follows:
Let λ̂C , ÊC be given. There exists a pair (x 1, x 2 ) ∈ ÊC such that i(d )(x 1 ) = u, i(d )(x 2 ) = v and φ |= Ctrl(x 1, x 2 ). There is a
corresponding clause in G with a constraint ψ |= φ. Moreover, because ψ |= SymRel[Instr[P](λ̂C (x 1 ), λ̂C (x 2 ))](x 1, x 2, X ), and
because G is location consistent and neighborly, i(d ) ⊢ SymRel[Instr[P](λ(u), λ(v))](u, v, ν (v)). Let m be a model over language of
SymPath(P, p, ν ) such that m restricted to {u, v } ∪ ν (v) is i(d). Since this property of m holds for all control edges (u, v) ∈ E then
m ⊢ SymPath(P, p, ν ) by the definition of SymPath.
33
Now, let a model m of SymPath(P, p, ν ) be given. The process above can almost be reversed to obtain models m D and i(d ) for
each d ∈ Nodes[D], but for each clause in G with constraint ψ , care must be taken surrounding the extra condition which ψ entails,
described in §5.3.1. Because the encoding of loads and stores requires that every live object be held by some local state variable, if an
object is not passed from R, resp. to R, this object is guaranteed to not have any of its fields modified on the control paths between
any of the ⇝+R pairs, resp. ⇝−R pairs. Therefore, ψ does not overconstrain i(d ), and we have a model (D, m D , i) of G inducing p and
(N , E Data ).
□
Lemma 3. P is simulated by SynGrammar (P, F).
Proof. SynGrammar begins by creating a CHC system C = SynSkeleton(P, F ). By the specification of SynSkeleton, C is
complete for P; i.e., every path p ∈ Paths[P] is the control graph modeled by some derivation of C.
SynGrammar then constructs the run grammar G corresponding to C. By Lemma 5, for every path p ∈ Paths[P], there is a
derivation D of G iff there is a model of SymPath(P, p, ν ) for some neighborhood ν . By Lemma 4 and Lemma 2, there is a model of D
iff there is a run of p. Therefore, G simulates every path p ∈ Paths[P], and thus simulates P.
□
Theorem 1. If Lttp(P) = True, then P is safe, and if Lttp(P) = False, then P is not safe.
Proof. Lttp returns True only if SolveCHC finds a solution to G = SynGrammar(P, F ), which proves that G is empty. By
Lemma 3, G simulates P, and with Lemma 1, this proves P is safe.
Lttp returns False only if a derivation and model (D, m, i) of G is found for which IsFeas(P, p) = True, where p is the control
graph induced by (D, m, i). (And p is path of P because G is a path grammar of P.) By Lemma 2, p is feasible run of P, witnessing that
P is not safe.
□
34
| 6cs.PL |
352
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE,
VOL. 40,
NO. 2,
FEBRUARY 2018
Exploiting Feature and Class Relationships
in Video Categorization with Regularized
Deep Neural Networks
arXiv:1502.07209v2 [cs.CV] 21 Feb 2018
Yu-Gang Jiang , Zuxuan Wu, Jun Wang, Xiangyang Xue, and Shih-Fu Chang , Fellow, IEEE
1
Abstract—In this paper, we study the challenging problem of categorizing videos according to high-level semantics such as the
existence of a particular human action or a complex event. Although extensive efforts have been devoted in recent years, most existing
works combined multiple video features using simple fusion strategies and neglected the utilization of inter-class semantic
relationships. This paper proposes a novel unified framework that jointly exploits the feature relationships and the class relationships for
improved categorization performance. Specifically, these two types of relationships are estimated and utilized by imposing
regularizations in the learning process of a deep neural network (DNN). Through arming the DNN with better capability of harnessing
both the feature and the class relationships, the proposed regularized DNN (rDNN) is more suitable for modeling video semantics.
We show that rDNN produces better performance over several state-of-the-art approaches. Competitive results are reported on the
well-known Hollywood2 and Columbia Consumer Video benchmarks. In addition, to stimulate future research on large scale video
categorization, we collect and release a new benchmark dataset, called FCVID, which contains 91,223 Internet videos and 239
manually annotated categories.
Index Terms—Video categorization, deep neural networks, regularization, feature fusion, class relationships, benchmark dataset
Ç
INTRODUCTION
V
IDEOS carry very rich and complex semantics, and are
intrinsically multimodal. Techniques for recognizing
high-level semantics in diverse unconstrained videos can
be deployed in many applications, such as Internet video
search. However, it is well-known that semantic recognition
or categorization of videos is an extremely challenging task
due to the semantic gap between computable low-level
video features and the complex high-level semantics. While
significant progress has been achieved in recent years, most
state-of-the-art solutions rely on a large set of features to
recognize a class-of-interest. In order to derive a robust
fused representation that bridges the semantic gap, the
fusion process of multiple features is usually expected to
learn the cross-feature correlations, such as the relationship
of HOG and HOF features that model visual information
and their complements to acoustic descriptors. In addition
to the feature relationships, there are also certain correlations among multiple high-level semantic categories:
Y.-G. Jiang, Z. Wu, and X. Xue are with the School of Computer Science,
Fudan University, Shanghai 200433, China.
E-mail: {ygj, zxwu, xyxue}@fudan.edu.cn.
J. Wang is with the School of Computer Science and Software Engineering,
East China Normal University, Shanghai 200062, China.
E-mail: wongjun@gmail.com.
S.-F. Chang is with the Department of Electrical Engineering, Columbia
University, New York, NY 10027. E-mail: sfchang@ee.columbia.edu.
Manuscript received 28 Feb. 2015; revised 26 Dec. 2016; accepted 6 Feb. 2017.
Date of publication 15 Feb. 2017; date of current version 5 Jan. 2018.
Recommended for acceptance by I. Laptev.
For information on obtaining reprints of this article, please send e-mail to:
reprints@ieee.org, and reference the Digital Object Identifier below.
Digital Object Identifier no. 10.1109/TPAMI.2017.2670560
knowing the presence of one category may provide useful
clues for recognizing other related categories. For example, a
high score of a video clip containing “running” (“diving”)
will increase (decrease) the confidence of the video containing
“soccer”. Although there exist several works investigating
multi-feature fusion or exploiting the inter-class relationships,
as will be discussed in the next section, they mostly address
the two problems separately.
Motivated by the limitations of the existing techniques
and the increasing popularity of using Deep Neural Networks (DNN) for visual categorization, in this paper we
propose a novel unified framework that jointly learns the
feature relationships and the class relationships using a
DNN. Video categorization can also be carried out within
the same network utilizing the learned relationships.
Fig. 1 gives an overview of the proposed approach. We
first extract several popular video features, including the
popular frame-based features computed by the convolutional neural networks (CNN) [1], trajectory-based motion
descriptors and audio descriptors. The features are then
used as the inputs of a DNN, where the first two layers are
input and feature transformation layers, respectively. The
third layer is called fusion layer, where we impose regularization on the network weights to identify and utilize the feature relationships. Specifically, the regularization terms are
selected based on two natural properties of the inter-feature
relationships: correlation and diversity. The former means
that different features may share some common patterns in a
middle level representation lying between the original features and the high-level semantics (i.e., the transformed features after the second layer). The latter emphasizes the
0162-8828 ß 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
JIANG ET AL.: EXPLOITING FEATURE AND CLASS RELATIONSHIPS IN VIDEO CATEGORIZATION WITH REGULARIZED DEEP NEURAL...
353
Fig. 1. Illustration of the proposed rDNN framework for video categorization. See texts for more explanations.
unique characteristics of different features, which are the
complementary information that is likely to be useful for
identifying video semantics. Through modeling both properties using a feature correlation matrix, we impose a tracenorm regularization over the network weights to reveal the
hidden correlations and diversity of the features.
In order to discover and utilize the inter-class relationships, we impose similar regularizations on the weights of
the final output layer. This helps to identify the grouping
structures of video classes and the outlier classes. Semantic
classes within the same group share commonalities that can
be utilized as knowledge sharing for improved categorization performance, while the outlier classes should be
excluded from “negative” knowledge sharing. By executing
regularized learning of the two kinds of relationships within
the same DNN, we arrive at a unified framework, namely
regularized DNN (rDNN), which can be easily implemented.
Although the framework shown in Fig. 1 is built on the
static CNN feature and a few typical hand-crafted video features, it is feasible to use our approach for fusing any features. We also realize that, in the image categorization
domain, the CNN features are dominating state-of-the-art
approaches. The reasons of considering both the CNN feature and the hand-crafted features in this work are two-folds.
First, the hand-crafted features have been widely used for
video categorization and remain the key components of
many systems that generated recent state-of-the-art results
on tasks like human action recognition [2] and complex event
recognition [3], [4]. By using these features it is easy to make
comparisons with the traditional approaches. Second, so far,
very few existing works on neural networks based video feature extraction have demonstrated significantly better results
than the traditional hand-crafted features. Some end-to-end
learning methods only showed lower or similar results [5],
[6], and, similar to this paper, a recent work [7] reported better results by combining deep features with hand-crafted
features. Therefore, this paper considers both the deeply
learned and the hand-crafted features, and focuses on the
tasks of feature fusion and semantic categorization.
The main contribution of this paper is the proposal of the
rDNN. To the best of our knowledge, rDNN is the first
attempt to exploit both the feature and the class relationships
in the DNN pipeline for video categorization. Our formulation is designed to model the complex relationships such as
feature/class correlation and diversity. It is easy to implement and can be efficiently executed using a GPU. In addition, we introduce and release a new benchmark dataset,
called Fudan-Columbia Video Dataset (FCVID). FCVID contains 91,223 YouTube videos and 239 manually annotated
categories. It is currently one of the largest manually annotated datasets of Internet videos. Compared with some
recently released video benchmarks, FCVID covers a wide
range of categories popularly seen in Internet user-shared
videos, including events, scenes and objects. For example,
the new EventNet [8] consists of only events and its labels
are noisy as it is not manually labeled; the Sports-1M dataset [5] focuses only on sports and is also not manually
labeled; the ActivityNet [9] focuses on human actions; and,
the MPII Human Pose dataset was mainly designed for recognizing human poses [10]. We evaluate rDNN using our
new FCVID dataset, and hope that its public release could
stimulate future research around this challenging problem.
This work is based upon a conference publication [11] with
the following major extensions. First, a more comprehensive
survey of the state of the arts on video categorization is
included in the next section. Second, we provide proofs and
more discussions on how the learned relationships could help
improve the recognition performance. Third, we leverage
more powerful feature representations (i.e., the CNN features) to evaluate the generalization ability of the framework,
and implement several additional alternative methods to fully
justify the effectiveness of our approach. Finally, we introduce
a new dataset that is much larger than those popularly used in
recent literature. The rest of this paper is organized as follows.
Section 2 discusses related works, where we mainly focus on
the existing works exploiting feature or class relationships.
Section 3 elaborates the proposed rDNN approach. Extensive
experimental results are discussed in Section 4, where we also
briefly introduce the new FCVID dataset. Finally, Section 5
concludes this paper.
2
RELATED WORK
Video categorization has received significant research attention. Most approaches followed a very standard pipeline,
354
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE,
where various features are first extracted and then used
as inputs of classifiers. Many works have focused on the
design of novel features, such as the biologically inspired
pipeline [12], Spatial-Temporal Interest Points (STIP) [13],
trajectory-based descriptors [2], audio clues [14], and the
Convolutional Neural Networks based features [1], [5],
[6], [15].
In contrast to the variety of video features, Support Vector
Machines (SVM) have been the dominate classifier option for
over a decade. Recently, with the increasing popularity of
the deep learning based approaches, neural networks have
also been adopted for video classification [5], [6], [15].
Among them, probably the most well-known deep learning
based video categorization result is probably from Simonyan
and Zisserman [6], who used a two-stream CNN approach
to extract features from static frames and motion optical flow
respectively. The features were classified separately and the
predictions were then simply fused with fixed weights.
Using this pipeline, they reported similar performance to the
improved dense trajectories [2], one of the best hand-crafted
feature-based approaches. More recently, in addition to the
CNN, researchers also adopted recurrent neural networks
(RNN) to model the long-term temporal information in videos [16], [17], [18] and reported promising results.
Besides accuracy, efficiency is another important factor
that should be considered in the design of a modern video
classification system. Several recent studies investigated
this issue by proposing efficient classification methods [19],
[20] or parallel computing strategies [21], [22].
In the following we primarily discuss works on multifeature fusion and/or exploiting class relationships, which
are more closely related to this work.
2.1 Exploiting Feature Relationships
In most state-of-the-art video categorization systems, two
feature fusion strategies were popularly adopted, i.e., the
early fusion and the late fusion. Early fusion relies on the
assumption that multiple features are explicitly complementary to each other, however this assumption does not always
hold in the complex video data. Late fusion trains models
separately and then combines prediction scores. This
method cannot explore feature relationships in the categorization process as the features are processed separately. In
contrast, this work derives a fused representation by explicitly regularizing the fusion process, and the fusion process
and classification are conducted simultaneously under a unified objective. In other words, our approach intends to learn
what features are correlated and what are unique clues that
exist only in one input feature. These learned information is
used in generating the fused representation.
In both early and late fusion, fusion weights are needed
to weigh the importance of each individual feature dimension, which can be set as equal values (a.k.a. average fusion)
or learned based on cross validation. In several recent
works, multiple kernel learning (MKL) [23] was adopted to
estimate the fusion weights [24], [25]. MKL was reported to
produce better performance in some cases, but the gain was
also often observed to be insignificant [26].
Several more advanced feature fusion approaches were
proposed. In [27], Ye et al. proposed an optimization framework, called robust late fusion, which uses a shared low-rank
VOL. 40,
NO. 2,
FEBRUARY 2018
matrix to remove noises in certain feature modalities. In [28],
Jiang et al. focused on exploiting the correlations between
audio and visual features. They proposed to generate an
audio-visual joint codebook by discovering the correlations of
the two features for video classification. Jhuo et al. [29] followed a similar framework, and improved the speed of training the audio-visual codebook by replacing the segmentationbased region features with standard local features.
With the growing popularity of the DNN, a few recent
studies focused on combining multiple features in neural
networks, which are closely related to this work. A deep denoised auto-encoder was employed in [30] to learn a shared
representation based on mutimodal inputs. Similarly, a deep
Boltzmann machine was utilized in [31] to fuse visual and
textual features. Very recently, Kihyuk et al. [32] proposed to
learn a good shared representation by minimizing variation
of information, so that missing input modality can be better
predicted based on the available information. They showed
that this method outperforms [31] on several image classification benchmarks. Different from [30], [31] that fused the
features in a “free” way, in this paper we propose regularized
fusion of multiple features, which is intuitively reasonable
and empirically effective. Compared with [32], our objective
is to learn and use dimension-wise feature relationships.
Minimizing the variation of information in [32] might be
more suitable for images, but for videos, different modalities
(e.g., audio and visual) may represent very distinctive information and simply minimizing their variation is not a good
strategy to exploit the complementary information.
2.2 Exploiting Class Relationships
Many researchers have investigated class relationships,
commonly termed context, to improve classification performance. In [33], Torralba et al. discussed the importance of
context in the task of object detection in images. In [34], [35],
the class co-occurrence context was utilized to improve
object recognition accuracy. In the context of video analysis,
Naphade and Huang [36] proposed to utilize a probabilistic
graphical network to model the co-occurrence of semantic
concepts for video indexing and retrieval. Jiang et al. [37]
proposed a semantic diffusion algorithm to harness the
class relationships. Weng et al. [38] proposed a similar
domain-adaptive method that not only used the class relationships, but also explored temporal context information
of broadcast news videos. Recently, Deng et al. [39] proposed Hierarchy and Exclusion (HEX) graphs, which can
capture not only the co-occurrence class relationships, but
also mutual exclusion and subsumption. Another two
recent works [40], [41] utilized the co-occurrence statistics to
help video classification, where the co-occurrence of classes
was used more as a semantic feature representation.
Most of these approaches, however, rely on the cooccurrence statistics of the video classes, and thus cannot be
used in the cases where the classes share certain commonalities but do not explicitly co-occur in the same video. By
injecting a class relationship matrix into the learning process, our approach can automatically learn and utilize such
commonalities using a regularized DNN with a unified
objective, as will be elaborated in the following section.
Our formulation is partly inspired by recent research on
Multiple Task Learning (MTL) [42], [43]. MTL trains
JIANG ET AL.: EXPLOITING FEATURE AND CLASS RELATIONSHIPS IN VIDEO CATEGORIZATION WITH REGULARIZED DEEP NEURAL...
355
Fig. 2. Popular neural network structures: (a) Is the standard one-vs-all training scheme using a single type of feature; (b) is the popular structure
used in multi-class learning with a single type of feature; (c) is the one-vs-all training scheme using multiple types of features; and (d) processes
multiple features separately and then performs fusion using a middle layer [31].
multiple class models simultaneously and boosts the performance of a task (a classifier model) by seeking help from
other related tasks. MTL has demonstrated good results
in many applications, such as disease prediction [44], [45]
and financial stock selection [46]. Sharing certain commonalities among multiple tasks is the key idea of MTL
and several algorithms have been proposed with regularizations on the shared patterns across tasks [47], [48], [49].
These works exploited the class relationships in classification or regression problems using the conventional learning approaches, but never injected such regularizations
into the DNN.
In fact, neural network is one of the earliest MTL models [50]. See Fig. 2b for a standard network structure. In the
network, each unit of the output layer refers to a task (class)
and neurons of the hidden layers can be viewed as the
shared common features. In this paper, we show that, by
imposing explicit forms of regularizations, the class relationships can be better exploited for improved video categorization performance.
3
REGULARIZED DNN
3.1 Notations and Settings
We have a training set with a total of N video samples,
which are associated with C semantic classes. Since a
video sample may have M types of feature representations
(e.g., multiple visual and audio clues), we can use an
(M þ 1)-tuple to represent each video as
M
ðx1n ; . . . ; xm
n ; . . . ; xn ; yn Þ; n ¼ 1; . . . ; N:
xm
n
Here
represents the mth feature of the nth video sample,
and yn ¼ ½yn1 ; . . . ync ; . . . ynC > 2 BC is the associated
semantic label for the nth sample. If the nth sample belongs
to the cth semantic class, the cth element is set as ync ¼ 1,
otherwise ync ¼ 0. The objective for video classification
under the above setting is to train prediction models that
can categorize new test videos into the C semantic classes.
Simply, one can independently train one classifier for
each semantic class, where different features can be combined using either the early or the late fusion scheme.
Instead, here we propose a DNN framework with structure
regularization to perform video classification. In particular,
this regularized DNN carries out feature fusion with an additional layer, namely fusion layer, to exploit the correlation
and diversity of multiple features, as illustrated in Fig. 1.
Furthermore, we impose additional regularization on the
prediction layer to enforce knowledge sharing across different
semantic classes. With such a regularized DNN framework,
we are able to explicitly explore both types of relationships
in a uniform learning process. To address the details of the
proposed regularized DNN, below we first introduce the
background of training standard DNNs with a single type
of feature. After that, we elaborate our formulation and
explain why our proposed approach can realize the aforementioned goals.
3.2 DNN Learning with a Single Type of Feature
Inspired by the biological neural systems, DNN uses a
large number of interconnected neurons and construct
complex computational models to mimic the information
processing in neural systems. Through cascading the neurons in multiple layers, DNN exhibits strong non-linear
abstraction capacity and is able to learn arbitrary mapping from inputs to outputs as long as being given sufficient training data.
Given a DNN with L layers, we denote al1 and al as the
input and the output of the lth layer, l ¼ 1; . . . ; L, while Wl
and bl refer to the weight matrix and the bias vector of the
lth layer, respectively. With only a single type of feature, the
transition function from the (l 1)th layer to the lth layer
can be written as
al ¼
s ðWl1 al1 þ bl1 Þ l > 1;
x
l ¼ 1;
(1)
where the nonlinear sigmoid function sðÞ is defined as
sðxÞ ¼
1
:
1 þ ex
For simplicity, we can absorb bl1 into the weights coefficient Wl1 by adding an additional dimension to the feature vectors with a constant value one. Figs. 2a and 2b
show two types of four-layered neural networks using a
single feature as the input to classify data samples into C
semantic classes.
356
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE,
Typically, one can minimize the following cost function
to derive the optimal weights for each layer in the network
min
W
N
X
i¼1
‘ðfðxi Þ; yi Þ þ
L1
1 X
kWl k2F :
2 l¼1
(2)
The first part in the above cost function measures the empirical loss on the training data, which summarizes the dis^ i ¼ aL ¼
crepancy between the outputs of the network y
fðxi Þ and the ground-truth labels yi . The second part is a
regularization term preventing overfitting.
3.3 Regularization for Feature Fusion
The DNN using a single type of feature is effective in some
cases. However, for data with a variety of representations
like videos, the semantics could be carried by different feature representations. Motivated by the multisensory integration process of primary neurons in biological systems
[51], [52], we extend the basic DNN with structure regularization on an additional fusion layer to accommodate the
deep fusion process of multiple types of features. As demonstrated in Fig. 1, the fusion layer absorbs all the outputs
from the transformation layer to generate an integrated
representation as the input for the classification layer.
Accordingly, the transition equation for this fusion layer
can be written as the following:
!
M
X
m m
(3)
WE aE þ bE :
aF ¼ s
m¼1
We denote E as the index of the last layer of feature transformation and F as the index of the fusion layer (i.e., F ¼ E þ 1).
Hence, am
E represents the extracted mid-level representation
for the mth feature. From the above transition equation,
the mid-level representation is first linearly transformed by
the weight matrix Wm
E and then non-linearly mapped to generate the fused representation aF using a sigmoid function.
Note that the weights of the fusion layer, W1E ; . . . ; WM
E,
transform all the available features into a shared representation. Here the weight matrices are first vectorized into P
dimensional vectors separately with P ¼ jam
E j jaF j being the
’s
(m
¼
1;
.
.
.
;
M)
dimension
and the aF ’s
product of the am
E
dimension. To simplify the formulation, we assume the
extracted features am
E are of the same dimension. Then all the
coefficient vectors are stacked into a matrix WE 2 RP M .
Each column of WE corresponds to the weights of a single
feature with the element WE ði; jÞ given as
WE ði; jÞ ¼ WiE ðjÞ;
i ¼ 1; . . . ; M;
j ¼ 1; . . . ; P:
In order to perform feature fusion by exploring correlations and diversities simultaneously, we formulate the following regularized optimization problem to learn the
weights of the DNN
!
E X
M
L1
X
1 X
2
m 2
kWl kF þ
kWl kF
min L þ
W;C
2 l¼1 m¼1
l¼F
(4)
2
þ trðWE C1 W>
EÞ
2
s.t. C 0;
VOL. 40,
NO. 2,
FEBRUARY 2018
P
yi ; yi Þ is the empirical loss term. Different
where L ¼ N
i¼1 ‘ð^
from the standard single feature based neural network
(Equation (2)), we include one additional regularization
term in the above cost function with one more variable
C 2 RMM to model the inter-feature correlation.
Note that C is a symmetric and positive semidefinite
matrix and the last regularization term with the trace norm
can help utilize the inter-feature relationship. Similar formulations were often used in multiple task learning [43],
[53], where task relationships are explored to improve the
learning performance. Intuitively, the goal is to ensure that
the weight vectors of correlated feature dimensions should
contain similar values so that the correlated feature dimensions can contribute similarly to the fused representation.
On the one hand, if a non-diagonal entry of C is large,
updating WE by minimizing the trace norm ensures that the
weights of the corresponding feature dimensions are similar. On the other hand, if WE is fixed, minimizing the trace
norm can help learn a C with entries more consistent with
the network weights. Please see Equations (9), (10), and (11)
for a proof on the relationships between WE and C. In the
optimization stage, we adopt an alternative minimization
strategy to learn WE and C together, as will be explained in
Section 3.6. The coefficients 1 and 2 balance the contributions from different regularization terms.
3.4 Regularization for Class Knowledge Sharing
As discussed earlier, one can simply adopt the one-versus-all
strategy to independently train C classifiers for categorizing
video semantics. As illustrated in Figs. 2a and 2c, this one-vsall training scheme with a total of C four-layered neural networks can be applied for both single-feature and multifeature settings. To train a total of C neural networks separately, a sufficient amount of positive training samples are
desired for each video category. In addition, the independent
training process completely neglects the knowledge sharing
among different semantic categories. However, video semantics often share some commonality due to the strong correlations between different semantic categories, which have been
observed in previous studies [37], [54], [55]. Therefore, it is
critical to explore such a commonality by simultaneously
learning multiple video semantics, which can lead to better
learning performance [55]. Generally, the commonality
among multiple classes is represented by the parameter sharing among different prediction models [56], [57]. In addition,
it is fairly natural for DNN to perform simultaneous multiclass training. For example, as seen in Fig. 2b, by adopting a
set of C units in the output layer, a single-feature based DNN
can be easily extended to multi-class problems.
Motivated by the regularization methods adopted for
MTL [56], [57], here we present a regularized DNN that
aims at training multiple classifiers simultaneously with
deeper exploitation of the class relationships. To enforce
class knowledge sharing, we employ the following optimization problem as our learning objective
min
W;V
N
X
i¼1
‘ðfðxi Þ; yi Þ þ
L1
1 X
kWl k2F
2 l¼1
þ 2 trðWL1 V1 W>
L1 Þ:
s.t.
V 0:
(5)
JIANG ET AL.: EXPLOITING FEATURE AND CLASS RELATIONSHIPS IN VIDEO CATEGORIZATION WITH REGULARIZED DEEP NEURAL...
Although some previous MTL works explore similar regularization in the learning objective, they often assume that
the class relationships are explicitly given and are ready for
use as prior knowledge [43], [57]. In our formulation, we
tend to learn the prediction model as well as the class relationships. In particular, we adopt a convex formulation by
imposing a trace norm regularization term over the coefficients of the output layer WL1 with the class relationships
augmented as a matrix variable V 2 RCC . The constraint
V 0 indicates that the class relationship matrix is positive
semidefinite since it can be viewed as the similarity measure
of the semantic classes. The form of this regularization term
is the same with the feature regularization in Equation (4),
and minimizing it ensures the consistency between weight
correlations in WL1 and the non-diagonal values in V. The
coefficients 1 and 2 are regularization parameters that balance the contributions from different terms.
3.5 Final Objective of rDNN
Considering both objectives of feature fusion and class
knowledge sharing, we now present a unified DNN formulation that is able to explore both the feature and the class
relationships. In this joint framework, one additional layer
is employed to fuse multiple features, where the objective is
to bridge the gap between low-level features and the highlevel video semantics. Then another layer of neurons is
stacked over the fusion layer to generate the predictions,
where we impose the trace norm regularization over the
prediction models to encourage knowledge sharing across
different semantic categories. To build such a rDNN, we
incorporate both the feature regularization in Equation (4)
and the class regularization in Equation (5) to form the following objective
!
E X
M
L1
X
1 X
2
2
kWm
kWl kF
min L þ
l kF þ
W;C;V
2 l¼1 m¼1
l¼F
2
þ trðWE C1 W>
EÞ
2
3
þ trðWL1 V1 W>
L1 Þ;
2
s.t. C 0 trðCÞ ¼ 1;
V 0 trðVÞ ¼ 1;
(6)
where 1 ; 2 ; and 3 are regularization parameters. In the
above formulation, two trace-norm regularization terms are
tailored for the fusion of multiple features and the exploitation of the class relationships, respectively. In addition, we
impose two additional constraints trðCÞ ¼ 1 and trðVÞ ¼ 1
to restrict the complexity, as suggested in [43]. In the next
section, we introduce an alternating optimization strategy
to minimize the above cost function with respect to the network weights fWl gLl¼1 , the feature relationship matrix C, as
well as the class correlation matrix V.
3.6 Optimization and Analysis
For the optimization problem in Equation (6), two pairs of
variables, i.e., ðWE ; CÞ and ðWL1 ; VÞ, are coupled with
each other. Here we adopt an alternating optimization
approach to iteratively minimize the cost function with
respect to Wm
l ðl ¼ 1; . . . L; m ¼ 1; . . . ; MÞ, C and V.
357
We first consider the minimization problem over the network weight matrix Wm
l with fixed C and V. It is easy to see
that the original problem is degenerated to the following
unconstrained optimization problem
!
E X
M
L1
X
1 X
2
m 2
Lþ
kWl kF þ
kWl kF
min
Wm
2 l¼1 m¼1
l
l¼F
(7)
2
3
1
1
>
>
þ trðWE C WE Þ þ trðWL1 V WL1 Þ:
2
2
Since all the terms in the above cost function are smooth, the
gradient can be easily evaluated. Let Gm
l be the gradient
with respect to Wm
l . We have the following update equation
for the weight matrix Wm
l at the kth iteration
m
m
Wm
l ðkÞ ¼ Wl ðk 1Þ hGl ðkÞ;
(8)
where h is the step length of the gradient descent.
We then introduce the solution for minimizing the cost
function over C with other variables being fixed. The problem in Equation (6) can be rewritten as
min trðWE C1 W>
E Þ;
C
s.t.
C0
trðCÞ ¼ 1:
(9)
Before giving the analytical solution of C, we provide a brief
discussion on the connection of W and C, which explains
the capability of this regularization term in a more rigorous
way. We first rewrite the above equation to
min trðC1 W>
E WE Þ;
C
s.t.
C0
trðCÞ ¼ 1:
(10)
Denote U ¼ W> W, since C is a symmetric matrix and
trðCÞ ¼ 1 we have
trðC1 UÞ ¼ trðC1 UÞtrðCÞ;
1 1 2 1 2
¼ C2 U2 C2
F
F
1 1 1 2
2 2 2
C U C
(11)
F
1
¼ ðtrðU2 ÞÞ2 :
Adopting the Cauchy-Schwarz inequality, trðC1 UÞ
1 1
1
1
attains minimum ðtrðU2 ÞÞ2 if and only if C2 U2 ¼ aC2 .
Therefore, C is determined by matrix U, which defines the
relationships among multiple features.
We now provide the analytical solution of C as
1
C¼
2
ðW>
E WE Þ
1
2
trððW>
E WE Þ Þ
:
(12)
Similarly, we can derive the optimal solution for V as
1
V¼
2
ðW>
L1 WL1 Þ
1
2
trððW>
L1 WL1 Þ Þ
:
(13)
Note that Zhang et al. adopted a similar solution to identify
task correlations for a linear kernel based regression and
classification problem [43]. However, our method integrates
358
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE,
more complex structure regularizations in a neural network
architecture, where both the feature and the class relationships are exploited for a completely different application.
In summary, we first estimate the feature and class relationships using the weights in the neural network. The relationship matrices are then utilized in turn to refine the
network weights to improve the classification performance.
Due to the existence of analytical solutions, we are able to
learn the relationship matrices C and V in an efficient way.
Finally, the training procedure of the proposed rDNN is
summarized in Algorithm 1. In each epoch, we need to comm
pute the gradient matrix Gm
l for updating Wl , and then
update the matrices V and C. The complexity of calculating
the trace norms is the same as that of the ‘2 norm. The
update of V and C requires cubic-complexity operations
with respect to the number of features M and the number of
video classes C. In practical large scale settings, the values
of M and C are often significantly smaller than the number
of training samples. Therefore, the training cost of the proposed rDNN is very similar to that of a standard DNN. Our
empirical study further confirms the efficiency of our
method, as will be discussed later.
Algorithm 1. Training Procedure of rDNN
Require: xm
n : the representation of the mth feature for the nth
video sample;
yn : the semantic label of the nth video sample;
1
1
1: Initialize Wm
l randomly, C ¼ M IM and V ¼ C IC , where IM
and IC are identity matrices;
2: for iteration ¼ 1 to K do
3:
Back propagate the prediction error from layer L to layer
1 by evaluating the gradient Gm
l , and update the weight
matrix Wm
for
each
layer
and
each
feature as:
l
m
m
Wm
l ðkÞ ¼ Wl ðk 1Þ hGl ðkÞ;
4:
Update the feature relationship matrix C according to
Equation (12):
1
C¼
5:
2
ðW>
E WE Þ
1
2
trððW>
E WE Þ Þ
;
Update the class relationship matrix C according to
Equation (13):
1
V¼
6:
end for
4
EXPERIMENTS
2
ðW>
L1 WL1 Þ
1
2
trððW>
L1 WL1 Þ Þ
:
4.1 Experimental Setup
4.1.1 Dataset and Evaluation
We adopt three challenging datasets to evaluate the rDNN,
as described in the following.
Hollywood2 [13]. The Hollywood2 dataset is well-known
in the area of human action recognition in videos. Collected
from 69 Hollywood movies, it contains 1,707 short video
VOL. 40,
NO. 2,
FEBRUARY 2018
clips annotated according to 12 classes: answering phone,
driving car, eating, fighting, getting out of car, hand shaking, hugging, kissing, running, sitting down, sitting up and
standing up. Following [13], the dataset is split into a training set with 823 videos and a test set with 884 videos.
Columbia Consumer Videos (CCV) [58]. The CCV dataset is
a popular benchmark on Internet consumer video categorization. It contains 9,317 videos collected from YouTube
with annotations of 20 semantic categories, including
objects (e.g., “cats”), scenes (e.g., “playground”), and events
(e.g., “parade”). Since many categories are events, it
requires a joint use of multiple feature clues like visual and
audio representations to perform better categorization. The
dataset is evenly split into a training set and a test set.
Fudan-Columbia Video Dataset. Since both the Hollywood2
and the CCV datasets are small in terms of the number of
annotated classes and the number of videos, to substantially
evaluate our rDNN, we collect and release a new benchmark, named FCVID.1 This dataset contains 91,223 Internet
videos annotated manually according to 239 categories, covering a wide range of topics like social events (e.g., “tailgate
party”), procedural events (e.g., “making cake”), objects
(e.g., “panda”), scenes (e.g., “beach”), etc. We divide the
dataset evenly with 45,611 videos for training and 45,612
videos for testing. To the best of our knowledge, FCVID is
one of the largest datasets for video categorization with
accurate manual annotations. Due to space constraint,
please refer to the supplementary material, which can be
found on the Computer Society Digital Library at http://
doi.ieeecomputersociety.org/10.1109/TPAMI.2017.2670560
for more information of the dataset, including details on the
collection and annotation process, statistics, a category hierarchy, as well as other related released resources (e.g., all
the computed features used in this work).
For all the three datasets, we adopt average precision
(AP) to measure the performance of each category and
report mean AP (mAP) as the overall results of all the categories. The standard training and testing splits are adopted
with no separate validation sets.
4.1.2 Video Features
As aforementioned, we consider both deeply learned features and hand-crafted features in this work.
Static CNN Features. Recently, CNN has exhibited topnotch performance in various visual categorization tasks, particularly in the image domain [59]. We adopt a CNN model
pre-trained on the ImageNet 2012 Challenge data, which consists of 1.2 million images and 1,000 concept categories. For a
given video frame, we extract a 4,096-d feature representation
(CNN-fc7 ), which is the output of the 7th fully connected layer
as suggested in [60]. Finally, the frame-level features are averaged to generate a video-level representation.
Motion Trajectory Features [2]. The dense trajectory features [2] have been popular for several years, which have
exhibited strong performance on various video categorization datasets. Densely sampled local frame patches are first
tracked over time to generate the dense trajectories. For
each trajectory, four descriptors are computed based on
local motion pattern and the appearance around the
1. Available at: http://bigvid.fudan.edu.cn/FCVID/
JIANG ET AL.: EXPLOITING FEATURE AND CLASS RELATIONSHIPS IN VIDEO CATEGORIZATION WITH REGULARIZED DEEP NEURAL...
trajectory, including a 30-d trajectory shape descriptor,
a 96-d histogram of oriented gradients (HOG) descriptor,
a 108-d histogram of optical flow (HOF) descriptor, and a
108-d motion boundary histogram (MBH) descriptor.
Finally, each type of descriptor is quantized into a 4,000-d
bag-of-words representation, following the settings of [2].
Audio Features. The audio soundtracks contain very useful clues that can help categorize some video semantics.
Two types of video features are considered in this work.
The first one is the popular MFCCs (Mel-Frequency Cepstral Coefficients), which are computed over every 32 ms
time-window with 50 percent overlap and then quantized
into a bag-of-words representation. The second one is
called Spectrogram SIFT (sgSIFT) [61], where we transform
the 1-d soundtrack of a video into a 2-D image based on the
constant-Q spectrogram. Standard SIFT descriptors are
extracted from this spectrogram and quantized into a bagof-words representation.
4.1.3 Alternative Approaches for Comparison
To verify the effectiveness of our rDNN, we compare with
the following approaches:
1)
2)
3)
4)
5)
6)
7)
8)
9)
DNN. The same structure with the rDNN using the
same 0.5 dropout ratio, but our proposed regularization term is not imposed.
Early Fusion with Neural Networks (NN-EF). All the
features are concatenated into a long vector and then
used as the input to train a neural network for video
categorization.
Late Fusion with Neural Networks (NN-LF). A neural
network is trained using each feature representation
independently. The outputs of all the networks are
fused to obtain the final categorization results.
Early Fusion with SVM (SVM-EF). The popular x2
kernel SVM is adopted and the features are combined on the kernel level before classification.
Late Fusion with SVM (SVM-LF). An SVM classifier is
trained for each feature and prediction results are
then combined.
Multiple Kernel Learning (SVM-MKL). We perform
feature fusion with the ‘p -Norm MKL [62] by fixing
p ¼ 2. MKL is able to learn dynamic fusion weights.
For the above EF/LF approaches 1–4, we adopt
equal fusion weights.
Multimodal Deep Boltzmann Machines (M-DBM).
M-DBM is a fusion approach proposed in [31], where
multiple features are used as the inputs of the Deep
Boltzmann Machines.
Discriminative Model Fusion (DMF). DMF [63] is one
of the earliest approaches for exploiting the interclass relationships. It simply uses the outputs of an
initial classifier, e.g., a DNN in our case, as the features to train an SVM model as the second level classifier to generate the final prediction. The second
level SVM is expected to be able to learn and use the
class relationships.
Domain Adaptive Semantic Diffusion (DASD). DASD
[37] uses a graph diffusion formulation to utilize the
inter-class relationships for visual categorization.
Similar to DMF, the prediction outputs of a DNN
359
(without the regularizations) are used as the inputs
of the DASD in a post-processing refinement step.
The approach requires inputs of pre-computed class
correlations, which can be estimated based on statistics of label co-occurrences in the training data.
Notice that the pre-computed class correlations are
not needed by our rDNN, which can automatically
learn the relationships.
Among the alternative approaches, 2–7 focus on feature
fusion, while the last two focus on the use of the class relationships. All the neural networks based experiments are
conducted on a single NVIDIA Telsa K20 GPU.
4.2 Results and Discussion
We now report and discuss experimental results. In order to
understand the contributions of only exploiting the feature
and the class relationships, we first test the performance of
the rDNN by disabling the regularizations on the output
layer and the fusion layer, respectively. This also ensures to
make fair comparisons with the alternative approaches.
After that, we enable regularizations on both layers and
report results of the entire rDNN framework. With this setting, we analyze the effect of the number of training samples, and compare with recent state-of-the-art results. Last,
we discuss the computational efficiency of rDNN.
Throughout the experiments, we use 4 layers of neurons
in the rDNN. All the features are used as the input of the
first layer, which are then transformed using a hidden layer
with 256 neurons for each type of feature separately. The
transformed features are further fused with a fusion layer
containing 256 neurons, and the fused feature is finally converted to classification scores through the last layer. Note
that 4 layers are empirically found to be suitable. Using
more layers in rDNN may improve the results but would
probably require more training data.
For the key parameters, we set the learning rate of the
neural networks to 0.7, fix 1 to a small value of 3e5 in
order to prevent overfitting, and set 2 and 3 to 5e-5 for
Hollywood and CCV, and 3e-5 for FCVID. We adopt the
mini batch gradient descent with the batch size being 70 for
network training. The training will stop if it reaches the
maximal epochs or the training error stops to decrease in
the last 10 epochs (with difference less than 1e-5).
4.2.1 Exploiting Feature Relationships
We first report results by only using the fusion layer regularization in our rDNN, namely rDNN-Feature Regularization (rDNN-F). Table 1 shows the results of the individual
features, our rDNN-F, and the alternative feature fusion
methods. Among the static CNN, motion and audio features, motion is significant better than the other two on Hollywood2 but is slightly worse than the CNN feature on
CCV and FCVID. This is due to the fact that many classes in
CCV and FCVID (e.g., “baseball” and “desert”) can be recognized by viewing just one or a few discrete frames, but
categorizing the Hollywood2 human actions normally
requires a sequence of frames with detailed motion clues. In
addition, the overall performance on CCV is slightly lower
than that on the much larger FCVID. This is because CCV
has some highly correlated categories (see Fig. 4) that are
360
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE,
TABLE 1
Performance Comparison (mAP) Using Individual and
Multiple Features with Various Fusion Methods
Static CNN
Motion
Audio
DNN
NN-EF
NN-LF
SVM-EF
SVM-LF
SVM-MKL [62]
M-DBM [31]
rDNN-F
Hollywood2
CCV
FCVID
40.1%
62.4%
22.7%
64.2%
63.5%
60.2%
64.1%
62.7%
63.8%
63.9%
65.9%
66.1%
60.8%
25.9%
71.6%
70.2%
69.9%
71.7%
69.1%
71.3%
71.1%
72.9%
63.8%
62.8%
26.1%
72.1%
74.7%
73.8%
75.0%
72.1%
75.2%
74.4%
75.4%
“rDNN-F” indicates our rDNN focusing only on the exploitation of the
feature relationships.
very difficult to be separated. While FCVID also contains
similar confusing categories, the percentage of such
“difficult” cases is lower as it also has more “easy” categories, and therefore the overall performance is higher.
For the fusion of the three types of features, our rDNN-F
achieves the best performance with consistent gains over all
the compared methods. Note that, like the “DNN” baseline,
the M-DBM approach also utilizes a neural network for feature fusion, but in a free manner without explicitly enforcing
the use of the feature relationships. These results clearly
verifies the effectiveness of imposing the proposed fusion
regularization method. Notice that, since the Hollywood2
and the CCV datasets have been widely used, an absolute
mAP gain of 2 percent is generally considered as a significant improvement.
Among the alternative approaches, early fusion methods
tend to produce better results than late fusion. This is consistent with the observations of several recent works, where
early fusion is more popularly adopted [3]. The MKL is even
slightly worse than early fusion on Hollywood2 and CCV,
VOL. 40,
NO. 2,
FEBRUARY 2018
TABLE 2
Performance Comparison (mAP) with DMF and DASD,
which Focus on the Use of the Class Relationships
DNN
DMF [63]
DASD [37]
rDNN-C
rDNN-C prior
Hollywood2
CCV
FCVID
64.2%
61.8%
64.4%
65.1%
65.8%
71.6%
71.1%
71.7%
72.1%
72.5%
72.1%
72.5%
72.8%
74.4%
75.0%
“DNN” is a baseline without imposing our proposed regularization
term and “rDNN-C” indicates our rDNN utilizing only the class
relationships.
indicating that the learned weights do not generalize well to
testing data. In addition, for the contribution of the audio feature in the fusion experiments, we observe clearly improvement for the classes with strong audio clues, such as
“answering phone”. On the contrary, for classes like “sitting
down”, audio features may slightly degrade the result.
4.2.2 Exploiting Class Relationships
Next, we report results of rDNN using only the class relationships, namely rDNN-C. We compare with the DNN
baseline with no regularization, DMF and DASD. Results
are given in Table 2. rDNN-C outperforms the DNN baseline and the two alternative approaches. Both DMF and
DASD use the outputs of the DNN baseline as inputs for
context-based refinement. These results corroborate the
effectiveness of the class relationship regularization.
Note that, like many previous methods exploring class
relationships, the DASD requires pre-computed class relationships as the input, which are estimated based on the
label co-occurrences in the training data. This might be the
reason that it performs worse than the rDNN-C as the latter
automatically learns the commonalities shared among the
categories. The learning process can identify not only the
categories that co-occur, but also those sharing visual or
auditory commonalities but rarely appear together. To
Fig. 3. The learned class relationship matrix V on FCVID and example frames of a few category groups. Many of the found groups contain categories
that share visual/auditory commonalities but do not necessarily co-occur.
JIANG ET AL.: EXPLOITING FEATURE AND CLASS RELATIONSHIPS IN VIDEO CATEGORIZATION WITH REGULARIZED DEEP NEURAL...
361
TABLE 3
Performance of the Entire Framework (the Last Row)
Using Both Kinds of Relationships, in Comparison with
Single-Relationship Results and the Basic Deep Networks
with Various Numbers of Hidden Layers
DNN
DNN-6 layer
DNN-8 layer
rDNN-F
rDNN-C
rDNN
Hollywood2
CCV
FCVID
64.2%
60.1%
56.2%
65.9%
65.1%
66.9%
71.6%
68.1%
62.3%
72.9%
72.1%
73.5%
72.1%
67.3%
62.7%
75.4%
74.4%
76.0%
shown in Table 2 (the last row rDNN-C prior), we observe further improvements, which are however not very significant.
Therefore, we conclude that the automatically identified
visual/auditory commonalities are effective. In addition, to
further verify that the gain of our approach is really from
using the learned class relationships, not from our different
optimization strategy as compared with the baselines, we
simply fix V to be an identity matrix in the optimization process (i.e., all the classes are treated independently). Under
this setting, the performance drops 1.6, 0.8 and 3.0 percent on
Hollywood2, CCV and FCVID respectively.
Fig. 4. (a) The learned class relationship matrix V on CCV. (b) Confusion
matrix on CCV.
verify this, we visualize some found category groups in
Fig. 3. As discussed in Section 3, values in the matrix V can
indicate the learned relationships among the categories.
Hence, we apply the spectral clustering algorithm on V to
group the categories and provide examples of several classes having high similarities. We see that many categories
are grouped together because they share certain commonalities (e.g., “marathon” and “marchingBand”), not due to
high frequencies of co-occurrence. In addition, we further
visualize the learned matrix V on the smaller dataset CCV
in Fig. 4a. The learned correlated categories may be due to
either the shared objects, scenes or audio clues.
It is interesting to notice that, once prior knowledge of the
relations among multiple categories is available, it can be leveraged to initialize the relationship matrix V. For example, if
a category i is known to be more similar to j than to k, we
could simply set Vij > Vik in the initialization stage. As
4.2.3 Exploiting Both Kinds of Relationships
Finally, we discuss the results of the entire rDNN framework, using both the feature and the class relationships.
Table 3 presents the results of the overall framework. Overall, substantial performance gains are attained from the proposed approach. Using regularizations on both kinds of
relationships leads to clearly higher performance than
imposing the regularization on a single type of relationship.
Compared with DNN structures that only adopt dropout
to improve generalization, rDNN achieves better performance on all the datasets. We also deepened the network
structures with six and eight layers in order to learn the hidden relationships (indicated by “DNN-6 layer” and “DNN8 layer” in the table), but the results are significantly worse.
This is because more parameters are added with the additional layers, which will easily lead to over-fitting especially
when training with limited samples.
In addition, comparing the results across the three datasets, the improvement from exploiting the class relationships is more significant on FCVID. This is because FCVID
contains a much larger number of classes that share commonalities helpful for categorization. Fig. 4 further visualizes the confusion matrix of rDNN on the CCV dataset.
4.2.4 Training with Limited Samples
Regularization techniques could usually help improve the
results when training with limited samples. To better evaluate the effectiveness of the regularizers, we plot the performance with different numbers of training samples in Fig. 5.
We observe that the performance gain of rDNN is more significant when the number of training samples is small
(except the case of 10 training samples on FCVID, which are
too few to distinguish the 239 categories). Under all the settings, the rDNN requires less training data to achieve comparable results to the non-regularized version.
362
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE,
VOL. 40,
NO. 2,
FEBRUARY 2018
Fig. 5. Performance of different numbers of training samples. We plot the results of the DNN baseline without regularization, rDNN-F, rDNN-C and the
rDNN exploiting both types of relationships. The best mAP on the three datasets (the rDNN approach using all the training samples) are 66.9, 73.5
and 76.0 percent respectively.
4.2.5 Comparison with State of the Arts
We compare rDNN with several recent approaches in
Table 4. On Hollywood2, our proposed method achieves a
competitive mAP of 66.9 percent, outperforming many of
the compared approaches [2], [11], [64], [65], [66], except a
few recent results [67], [68], [69]. Most of these approaches
are based on the popular dense trajectory features and the
SVM classification with the simple early fusion method.
Note that some of them like Wang et al. [2] and Lan
et al. [67] encoded the features using the Fisher vector [70],
which has been shown to be more effective than the classical
bag-of-words representation used in our approach. The
approach by Lan et al. [67] extends upon the dense trajectories with a feature enhancement method called multi-skip
feature stacking, while Hoai et al. [68] and Fernando
et al. [69] explored prediction score distribution and temporal information respectively. Since the focus of these works
is different, further performance gain may be achieved by
combining them with rDNN.
On the CCV dataset, we obtain to-date the best performance with an mAP of 73.5 percent. Most recent works on
CCV focused on the joint use of multiple audio-visual features. Xu et al. [72] and Ye et al. [27] extended late fusion
with specially designed strategies to remove the noise of
individually trained classifiers. Jhuo et al. adopted a joint
audio-visual codebook to exploit feature relationships for
categorization [29].
TABLE 4
Comparison with State of the Arts
Hollywood2
mAP
CCV
Jain et al. [64]
Wang et al. [2]
Zhang et al. [65]
Ni et al. [66]
Wu et al. [11]
Lan et al. [67]
Hoai et al. [68]
Fernando et al. [69]
rDNN
62.5%
64.3%
50.9%
61.0%
65.7%
68.0%
73.6%
73.7%
66.9%
Kim et al. [71]
Xu et al. [72]
Ye et al. [27]
Jhuo et al. [29]
Ma et al. [73]
Liu et al. [74]
Wu et al. [11]
rDNN
mAP
56.5%
60.3%
64.0%
64.0%
63.4%
68.2%
70.6%
73.5%
4.2.6 Computational Efficiency
We discuss the computational efficiency of rDNN using the
Hollywood2 dataset. The average training time of each
epoch for NN-EF, NN-LF and rDNN are 1:540 0:02,
1:552 0:05 and 1:276 0:10, respectively, using the same
GPU-based implementation. rDNN is more efficient than
NN-EF and NN-LF as it contains less parameters to be
learned. Specifically, compared with the NN-EF, rDNN processes the features separately in the first two layers and thus
avoids the parameters needed for interacting among them.
The NN-LF requires the training of separate networks, which
is also more expensive. Note that the M-DBM method is not
compared because it requires much more time to pre-train
the network for weight initialization. For all the methods,
normally a few hundreds of epochs are needed to finish the
training (several minutes in total). After training, all the neural network methods are very fast in testing.
5
CONCLUSION
We have proposed a novel rDNN approach to exploit both
feature and class relationships in video categorization. By
imposing trace-norm based regularizations on the specially
tailored fusion layer and output layer, our rDNN can learn a
fused representation of multiple feature inputs and utilize
the commonalities shared among the semantic classes for
improved categorization performance. Extensive experiments
of action and event recognition on popular benchmarks have
shown that rDNN consistently outperforms several alternative approaches. Our rDNN is also efficient in both model
training and testing, which is very important for large scale
applications. In addition, we have introduced a new dataset,
FCVID, for large scale video categorization. We believe that
FCVID is helpful for stimulating research not only on video
categorization, but also on other related problems.
The current framework supports the use of any precomputed features. One interesting future work is to exploit
the joint learning of feature representations and classification models. For instance, the adopted CNN feature is computed based on off-the-shelf models. It would be probably
helpful if the feature extraction part could be further tuned
simultaneously with the regularized classification network.
JIANG ET AL.: EXPLOITING FEATURE AND CLASS RELATIONSHIPS IN VIDEO CATEGORIZATION WITH REGULARIZED DEEP NEURAL...
ACKNOWLEDGMENTS
This work was supported in part by two grants from NSF
China (#61622204, #61572138) and a grant from STCSM,
Shanghai, China (#16JC1420401).
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,” in Proc. Advances Neural Inf. Process. Syst., 2012, pp. 1106–1114.
H. Wang and C. Schmid, “Action recognition with improved
trajectories,” in Proc. IEEE Int. Conf. Comput. Vis., 2013, pp. 3551–
3558.
R. Aly, et al., “The AXES submissions at TrecVid 2013, ” in Proc.
NIST TRECVID Workshop, 2013.
Z.-Z. Lan, et al., “CMU-Informedia@TRECVID 2013 multimedia
event detection,” in Proc. NIST TRECVID Workshop, 2013.
A. Karpathy, G. Toderici, S. Shetty, T. Leung, R. Sukthankar, and
L. Fei-Fei , “Large-scale video classification with convolutional
neural networks,” in Proc. IEEE Conf. Comput. Vis. Pattern Recog.,
2014, pp. 1725–1732.
K. Simonyan and A. Zisserman, “Two-stream convolutional networks for action recognition in videos,” in Proc. Advances Neural
Inf. Process. Syst., 2014, pp. 568–576.
C. Feichtenhofer, A. Pinz, and A. Zisserman, “Convolutional twostream network fusion for video action recognition,” in Proc. IEEE
Conf. Comput. Vis. Pattern Recog., 2016, pp. 1933–1941.
G. Ye, Y. Li, H. Xu, D. Liu, and S.-F. Chang, “EventNet: A large
scale structured concept library for complex event detection in
video,” in Proc. 23rd ACM Int. Conf. Multimedia, 2015, pp. 471–480.
F. C. Heilbron, V. Escorcia, B. Ghanem, and J. C. Niebles,
“ActivityNet: A large-scale video benchmark for human activity
understanding,” in Proc. IEEE Conf. Comput. Vis. Pattern Recog.,
2015, pp. 961–970.
M. Andriluka, L. Pishchulin, P. Gehler, and B. Schiele, “2D human
pose estimation: New benchmark and state of the art analysis,” in
Proc. IEEE Conf. Comput. Vis. Pattern Recog., 2014, pp. 3686–3693.
Z. Wu, Y.-G. Jiang, J. Wang, J. Pu, and X. Xue, “Exploring interfeature and inter-class relationships with deep neural networks
for video classification,” in Proc. 22 ACM Int. Conf. Multimedia,
2014, pp. 167–176.
H. Jhuang, T. Serre, L. Wolf, and T. Poggio, “A biologically
inspired system for action recognition,” in Proc. IEEE Int. Conf.
Comput. Vis., 2007, pp. 1–8.
I. Laptev, M. Marszalek, C. Schmid, and B. Rozenfeld, “Learning
realistic human actions from movies,” in Proc. IEEE Conf. Comput.
Vis. Pattern Recog., 2008, pp. 1–8.
C. V. Cotton and D. P. W. Ellis, “Subband autocorrelation features
for video soundtrack classification,” in Proc. IEEE Int. Conf. Acoust.
Speech Signal Process., 2013, pp. 8663–8666.
S. Ji, W. Xu, M. Yang, and K. Yu, “3D convolutional neural networks for human action recognition,” in Proc. Int. Conf. Mach.
Learn., 2010, pp. 495–502.
N. Srivastava, E. Mansimov, and R. Salakhutdinov, “Unsupervised
learning of video representations using LSTMs,” ICML, pp. 843–
852, 2015.
J. Y.-H. Ng, M. Hausknecht, S. Vijayanarasimhan, O. Vinyals,
R. Monga, and G. Toderici, “Beyond short snippets: Deep networks for video classification,” in Proc. IEEE Conf. Comput. Vis.
Pattern Recog., 2015, pp. 4694–4702.
Z. Wu, X. Wang, Y.-G. Jiang, H. Ye, and X. Xue, “Modeling spatial-temporal clues in a hybrid deep learning framework for video
classification,” in Proc. 23rd ACM Int. Conf. Multimedia, 2015,
pp. 461–470.
S. Maji, A. C. Berg, and J. Malik, “Classification using intersection
kernel support vector machines is efficient,” in Proc. IEEE Conf.
Comput. Vis. Pattern Recog., 2008, pp. 1–8.
Y.-G. Jiang, “Super: Towards real-time event recognition in internet videos,” in Proc. 2nd ACM Int. Conf. Multimedia Retrieval, 2012,
Art. no. 7.
Y. Zou, X. Jin, Y. Li, Z. Guo, E. Wang, and B. Xiao, “Mariana: Tencent deep learning platform and its applications,” Proc. VLDB
Endowment, vol. 7, pp. 1772–1777, 2014.
O. Yadan, K. Adams, Y. Taigman, and M. Ranzato, “Multi-GPU
training of convnets,” arXiv preprint arXiv:1312.5853, vol. 9, 2013.
363
[23] F. R. Bach, G. R. Lanckriet, and M. I. Jordan, “Multiple kernel
learning, conic duality, and the SMO algorithm,” in Proc. 21st Int.
Conf. Mach. Learn., 2004, Art. no. 6.
[24] L. Cao, J. Luo, F. Liang, and T. S. Huang, “Heterogeneous feature
machines for visual recognition,” in Proc. IEEE Int. Conf. Comput.
Vis., 2009, pp. 1095–1102.
[25] P. Natarajan, et al., “Multimodal feature fusion for robust event
detection in web videos,” in Proc. IEEE Conf. Comput. Vis. Pattern
Recog., 2012, pp. 1298–1305.
[26] A. Vedaldi, V. Gulshan, M. Varma, and A. Zisserman, “Multiple
kernels for object detection,” in Proc. IEEE Int. Conf. Comput. Vis.,
2009, pp. 606–613.
[27] G. Ye, D. Liu, I.-H. Jhuo, and S.-F. Chang, “Robust late fusion with
rank minimization,” in Proc. IEEE Conf. Comput. Vis. Pattern
Recog., 2012, pp. 3021–3028.
[28] W. Jiang, C. Cotton, S.-F. Chang, D. Ellis, and A. Loui, “Short-term
audio-visual atoms for generic video concept classification,” in
Proc. 17th ACM Int. Conf. Multimedia, 2009, pp. 5–14.
[29] I.-H. Jhuo, et al., “Discovering joint audio-visual codewords for
video event detection,” Mach. Vis. Appl., vol. 25, pp. 33–47, 2014.
[30] J. Ngiam, A. Khosla, M. Kim, J. Nam, H. Lee, and A. Ng,
“Multimodal deep learning,” in Proc. Int. Conf. Mach. Learn, 2011,
pp. 689–696.
[31] N. Srivastava and R. Salakhutdinov, “Multimodal learning with
deep Boltzmann machines,” in Proc. Advances Neural Inf. Process.
Syst., 2012, pp. 2231–2239.
[32] K. Sohn, W. Shang, and H. Lee, “Improved multimodal deep
learning with variation of information,” in Proc. Advances Neural
Inf. Process. Syst., 2014, pp. 2141–2149.
[33] A. Torralba, “Contextual priming for object detection,” Int. J. Comput. Vis., vol. 53, pp. 169–191, 2003.
[34] S. Bengio, J. Dean, D. Erhan, E. Ie, Q. Le, A. Rabinovich, J. Shlens,
and Y. Singer, “Using web co-occurrence statistics for improving
image categorization,” arXiv preprint arXiv:1312.5697, 2013.
[35] A. Rabinovich, A. Vedaldi, C. Galleguillos, E. Wiewiora, and
S. Belongie, “Objects in context,” in Proc. IEEE Int. Conf. Comput.
Vis., 2007, pp. 1–8.
[36] H. R. Naphade and T. S. Huang, “A probabilistic framework for
semantic video indexing, filtering, and retrieval,” IEEE Trans.
Multimedia, vol. 3, no. 1, pp. 141–151, Mar. 2001.
[37] Y.-G. Jiang, Q. Dai, J. Wang, C.-W. Ngo, X. Xue, and S.-F. Chang,
“Fast semantic diffusion for large-scale context-based image and
video annotation,” IEEE Trans. Image Process., vol. 21, no. 6,
pp. 3080–3091, Jun. 2012.
[38] M.-F. Weng and Y.-Y. Chuang, “Cross-domain multicue fusion for
concept-based video indexing,” IEEE Trans. Pattern Anal. Mach.
Intell., vol. 34, no. 10, pp. 1927–1941, Oct. 2012.
[39] J. Deng, et al., “Large-scale object classification using label relation
graphs,” in Proc. 13th Eur. Conf. Comput. Vis., 2014, pp. 48–64.
[40] S. M. Assari, A. R. Zamir, and M. Shah, “Video classification using
semantic concept co-occurrences,” in Proc. IEEE Conf. Comput. Vis.
Pattern Recog., 2014, pp. 2529–2536.
[41] T. Mensink, E. Gavves, and C. G. M. Snoek, “COSTA: Cooccurrence statistics for zero-shot classification,” in Proc. IEEE
Conf. Comput. Vis. Pattern Recog., 2014, pp. 2441–2448.
[42] L. Jacob, F. R. Bach, and J.-P. Vert, “Clustered multi-task learning:
A convex formulation,” in Proc. Advances Neural Inf. Process. Syst.,
2008, pp. 745–752.
[43] Y. Zhang and D.-Y. Yeung, “A convex formulation for learning
task relationships in multi-task learning,” in Proc. 26th Conf.
Uncertainty Artif. Intell., 2010, pp. 733–742.
[44] D. Zhang and D. Shen, “Multi-modal multi-task learning for joint
prediction of multiple regression and classification variables in
Alzheimer’s disease,” Neuroimage, vol. 59, pp. 895–907, 2012.
[45] J. Zhou, L. Yuan, J. Liu, and J. Ye, “A multi-task learning formulation for predicting disease progression,” in Proc. 17th
ACM SIGKDD Int. Conf. Knowl. Discovery Data Mining, 2011,
pp. 814–822.
[46] J. Ghosn and Y. Bengio, “Multi-task learning for stock selection,”
in Proc. Advances Neural Inf. Process. Syst., 1997, pp. 946–952.
[47] J. Chen, J. Zhou, and J. Ye, “Integrating low-rank and groupsparse structures for robust multi-task learning,” in Proc. 17th
ACM SIGKDD Int. Conf. Knowl. Discovery Data Mining, 2011,
pp. 42–50.
[48] Z. Kang, K. Grauman, and F. Sha, “Learning with whom to share
in multi-task feature learning,” in Proc. Int. Conf. Mach. Learn.,
2011, pp. 521–528.
364
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE,
[49] J. Pu, Y.-G. Jiang, J. Wang, and X. Xue, “Multiple task learning
using iteratively reweighted least square,” in Proc. 23rd Int. Joint
Conf. Artif. Intell., 2013, pp. 1607–1613.
[50] R. Caruana, “Multitask learning,” Mach. Learn., vol. 28, pp. 41–75,
1997.
[51] T. Ohshiro, D. Angelaki, and G. DeAngelis, “A normalization
model of multisensory integration,” Nature Neuroscience, vol. 14,
pp. 775–782, 2011.
[52] B. E. Stein and T. R. Stanford, “Multisensory integration: Current
issues from the perspective of the single neuron,” Nature Rev. Neuroscience, vol. 9, pp. 255–266, 2008.
[53] H. Fei and J. Huan, “Structured feature selection and task relationship inference for multi-task learning,” Knowl. Inf. Syst., vol. 35,
pp. 345–364, 2013.
[54] G.-J. Qi, X.-S. Hua, Y. Rui, J. Tang, T. Mei, and H.-J. Zhang,
“Correlative multi-label video annotation,” in Proc. 15th ACM Int.
Conf. Multimedia, 2007, pp. 17–26.
[55] G. Ye, D. Liu, J. Wang, and S.-F. Chang, “Large-scale video hashing via structure learning,” in Proc. IEEE Int. Conf. Comput. Vis.,
2013, pp. 2272–2279.
[56] A. Argyriou, T. Evgeniou, and M. Pontil, “Convex multi-task feature learning,” Mach. Learn., vol. 73, pp. 243–272, 2008.
[57] J. Liu, S. Ji, and J. Ye, “Multi-task feature learning via efficient l 2,
1-norm minimization,” in Proc. 25th Conf. Uncertainty Artif. Intell.,
2009, pp. 339–348.
[58] Y.-G. Jiang, G. Ye, S.-F. Chang, D. Ellis, and A. C. Loui,
“Consumer video understanding: A benchmark database and an
evaluation of human and machine performance,” in Proc. 1st
ACM Int. Conf. Multimedia Retrieval, 2011, Art. no. 29.
[59] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D.
Erhan, V. Vanhoucke, and A. Rabinovich, “Going deeper with
convolutions,” in Proc. IEEE Conf. Comput. Vis. Pattern Recog.,
2015, pp. 1–9.
[60] R. Girshick, J. Donahue, T. Darrell, and J. Malik, “Rich
feature hierarchies for accurate object detection and semantic
segmentation,” in Proc. IEEE Conf. Comput. Vis. Pattern Recog.,
2014, pp. 580–587.
[61] B. Zhu, W. Li, Z. Wang, and X. Xue, “A novel audio fingerprinting
method robust to time scale modification and pitch shifting,” in
Proc. 18th ACM Int. Conf. Multimedia, 2010, pp. 987–990.
[62] M. Kloft, U. Brefeld, S. Sonnenburg, and A. Zien, “Lp-norm multiple kernel learning,” J. Mach. Learn. Res., vol. 12, pp. 953–997, 2011.
[63] J. R. Smith, M. Naphade, and A. Natsev, “Multimedia semantic
indexing using model vectors,” in Proc. IEEE Int. Conf. Multimedia
Expo, 2003, pp. 445–448.
[64] M. Jain, H. Jegou, and P. Bouthemy, “Better exploiting motion for
better action recognition,” in Proc. IEEE Conf. Comput. Vis. Pattern
Recog., 2013, pp. 2555–2562.
[65] H. Zhang, W. Zhou, C. M. Reardon, and L. E. Parker, “Simplexbased 3D spatio-temporal feature description for action recognition,” in Proc. IEEE Conf. Comput. Vis. Pattern Recog., 2014,
pp. 2067–2074.
[66] B. Ni, T. Li, and P. Moulin, “Beta process multiple kernel
learning,” in Proc. IEEE Conf. Comput. Vis. Pattern Recog., 2014,
pp. 963–970.
[67] Z. Lan, M. Lin, X. Li, A. G. Hauptmann, and B. Raj, “Beyond
Gaussian pyramid: Multi-skip feature stacking for action recognition,” in Proc. IEEE Conf. Comput. Vis. Pattern Recog., 2015,
pp. 204–212.
[68] M. Hoai and A. Zisserman, “Improving human action recognition
using score distribution and ranking,” in Proc. Asian Conf. Comput.
Vis., 2014, pp. 3–20.
[69] B. Fernando, E. Gavves, J. M. Oramas, A. Ghodrati, and T. Tuytelaars, “Modeling video evolution for action recognition,” in Proc.
IEEE Conf. Comput. Vis. Pattern Recog., 2015, pp. 5378–5387.
[70] J. S
anchez, F. Perronnin, T. Mensink, and J. Verbeek, “Image classification with the Fisher vector: Theory and practice,” Int. J. Comput. Vis., vol. 105, pp. 222–245, 2013.
[71] I. Kim, S. Oh, B. Byun, A. A. Perera, and C.-H. Lee, “Explicit performance metric optimization for fusion-based video retrieval,” in
Proc. Eur. Conf. Comput. Vis. Workshop, 2012, pp. 395–405.
[72] Z. Xu, Y. Yang, I. Tsang, N. Sebe, and A. Hauptmann, “Feature
weighting via optimal thresholding for video analysis,” in Proc.
IEEE Int. Conf. Comput. Vis., 2013, pp. 3440–3447.
[73] A. J. Ma and P. C. Yuen, “Reduced analytic dependency modeling: Robust fusion for visual recognition,” Int. J. Comput. Vis.,
vol. 109, pp. 233–251, 2014.
VOL. 40,
NO. 2,
FEBRUARY 2018
[74] D. Liu, K.-T. Lai, G. Ye, M.-S. Chen, and S.-F. Chang, “Samplespecific late fusion for visual category recognition,” in Proc. IEEE
Conf. Comput. Vis. Pattern Recog., 2013, pp. 803–810.
Yu-Gang Jiang received the PhD degree in computer science from City University of Hong Kong,
Kowloon, Hong Kong, in 2009. During 20082011, he was in the Department of Electrical
Engineering, Columbia University, New York. He
is currently a professor of computer science with
Fudan University, Shanghai, China. His research
interests include computer vision and multimedia.
Zuxuan Wu received the BEng degree in
software engineering from East China Normal
University, in 2013 and the MSc degree from
Fudan Univeristy, in 2016. He is currently working
toward the PhD degree in the Computer Science
Department, University of Maryland, College
Park. His research interests include computer
vision, multimedia retrieval, and deep learning.
Jun Wang received the PhD degree from Columbia University, New York, in 2011. Currently, he is
a professor of computer science with East China
Normal University, Shanghai, China. His research
interests include machine learning, information
retrieval, and data mining.
Xiangyang Xue received the BS, MS, and PhD
degrees in communication engineering from
Xidian University, Xi’an, China, in 1989, 1992,
and 1995, respectively. He is currently a professor of computer science with Fudan University,
Shanghai, China. His research interests include
multimedia information processing and machine
learning.
Shih-Fu Chang is the Richard Dicker professor,
director of the Digital Video and Multimedia Lab,
and senior vice dean of Engineering School,
Columbia University. His research is focused on
multimedia information retrieval, computer vision,
machine learning, and signal processing, with the
goal to turn unstructured multimedia data into
searchable information. He is a fellow of the
American Association for the Advancement of
Science (AAAS) and the IEEE.
" For more information on this or any other computing topic,
please visit our Digital Library at www.computer.org/publications/dlib.
| 1cs.CV |
arXiv:1712.00754v2 [cs.NE] 2 Mar 2018
Universal discrete-time reservoir computers with
stochastic inputs and linear readouts using
non-homogeneous state-affine systems
Lyudmila Grigoryeva1 and Juan-Pablo Ortega2,3
Abstract
A new class of non-homogeneous state-affine systems is introduced for use in reservoir computing.
Sufficient conditions are identified that guarantee first, that the associated reservoir computers with
linear readouts are causal, time-invariant, and satisfy the fading memory property and second, that
a subset of this class is universal in the category of fading memory filters with stochastic almost
surely uniformly bounded inputs. This means that any discrete-time filter that satisfies the fading
memory property with random inputs of that type can be uniformly approximated by elements in
the non-homogeneous state-affine family.
Key Words: reservoir computing, state-affine systems, SAS, echo state networks, ESN, echo state
affine systems, ESAS, machine learning, universality, fading memory property, linear training, stochastic
signal treatment.
1
Introduction
A reservoir computer (RC) [Jaeg 10, Jaeg 04, Maas 02, Maas 11, Croo 07, Vers 07, Luko 09] or a RC
system is a specific type of recurrent neural network determined by two maps, namely a reservoir
F : RN × Rn −→ RN , n, N ∈ N, and a readout map h : RN → R that under certain hypotheses
transform (or filter) an infinite discrete-time input z = (. . . , z−1 , z0 , z1 , . . .) ∈ (Rn )Z into an output
signal y ∈ RZ of the same type using the state-space transformation given by:
xt = F (xt−1 , zt ),
(1.1)
yt = h(xt ),
(1.2)
where t ∈ Z and the dimension N ∈ N of the state vectors xt ∈ RN will be referred to as the number
of virtual neurons of the system. The expressions (1.1)-(1.2) determine a nonlinear state-space system
and many of its dynamical properties (stability, controlability) have been studied for decades in the
literature from that point of view.
In supervised machine learning applications, the reservoir map is very often randomly generated
(see, for instance, the echo state networks in [Jaeg 10, Jaeg 04]) and the memoryless readout is trained
1 Department of Mathematics and Statistics.
Universität Konstanz. Box 146. D-78457 Konstanz. Germany.
Lyudmila.Grigoryeva@uni-konstanz.de
2 Universität Sankt Gallen. Faculty of Mathematics and Statistics. Bodanstrasse 6. CH-9000 Sankt Gallen. Switzerland. Juan-Pablo.Ortega@unisg.ch
3 Centre National de la Recherche Scientifique (CNRS). France.
1
Universal reservoir computers using non-homogeneous state-affine systems
2
so that the output matches a given teaching signal that we will denote by d ∈ RZ . Two important
advantages of this approach lay on the fact that they reduce the training of a dynamic task to a static
problem and, moreover, if the reservoir map is rich enough, good performances can be attained with
readouts that have a relatively simple functional form. Indeed, in many occasions it suffices to use an
affine map h : RN → R that is trained via a (eventually regularized) linear regression that minimizes
the Euclidean distance between the output y and the teaching signal d. These features circumvent
well-known difficulties in the training of generic recurrent neural networks having to do with bifurcation
phenomena that render classical gradient descent methods non-convergent [Doya 92].
There are two central questions that need to be addressed when designing a machine learning
paradigm, namely, the capacity and the universality problems. The capacity problem concerns the
estimation of the error that is going to be committed in the execution of a specific task. This estimation
can have the form of generic bounds that incorporate various architecture parameters of the system like
in [Pisi 81, Jone 92, Barr 93, Kurk 05]. In the specific context of reservoir computing, it has been the
subject of much research [Jaeg 02, Whit 04, Gang 08, Herm 10, Damb 12, Grig 15, Coui 16, Grig 16a].
The universality problem consists in showing that the set of input/output functionals that can be
generated with a specific architecture is dense in a sufficiently rich class, like the one containing, for
example, all continuous or all measurable functionals. For classical machine learning paradigms like
neural networks, this question has given rise to well-known results [Kolm 56, Arno 57, Spre 65, Spre 96,
Spre 97, Cybe 89, Horn 89, Rusc 98] that show that they can be considered as universal approximators
in a static and deterministic setup.
There is no general recipe that allows one to conclude the universality of a given supervised machine
learning paradigm. The proof strategy depends much on the specific paradigm itself and, more importantly, on the nature of the inputs and the outputs. In the context of reservoir computing there are
several situations for which universality has been established when the inputs/outputs are deterministic, that is, when dealing with real-valued curves or time series. There are two features that influence
significantly the level of mathematical sophistication that is needed to conclude universality: first, the
compactness of the time domain under consideration and second, if one works in continuous or discrete
time. In the following paragraphs we briefly review the results that have already been obtained and, in
passing, we present and put in context the contributions contained in this paper.
The compactness of the time domain is crucial because, as we will see later on, universality is
obtained as a consequence of various versions of the Stone-Weierstrass, that is invariably formulated
for functions defined on a compact metric space. When the time domain is compact, this property is
naturally inherited by the spaces relevant in the proofs. However, when it is not, it is obtained by
restricting the study to functionals that satisfy a condition introduced in [Boyd 85] and known as the
fading memory property. The distinction between continuous and discrete time inputs is justified
by the availability in the continuous setup of different tools coming from functional analysis that do not
exist for discrete time.
Reservoir computing universality for compact time domains is obtained as a corollary of classical
results in systems theory. Indeed, in the continuous time setup, it can be established [Flie 76, Suss 76] for
linear systems using polynomial readouts or with bilinear systems using linear readouts. In the discretetime setup, the situation is more convoluted when the readout is linear and required the introduction
in [Flie 80] of the so-called (homogeneous) state-affine systems (SAS) (see also [Sont 79a, Sont 79b]).
The extension of these results to continuous non-compact time intervals was carried out in [Boyd 85] for
fading memory filters using exponentially stable linear RCs with polynomial readouts and their bilinear
counterparts with linear readouts (see also [Maas 00, Maas 02, Maas 04, Maas 07]). An extension to
the non-compact discrete-time setup based on the Stone-Weierstrass theorem is, to our knowledge, not
available in the literature. This problem has only been tackled from an internal approximation point of
view, which consists in uniformly approximating the reservoir and readout maps in (1.1)-(1.2) in order to
obtain an approximation of the resulting filter; this strategy has been introduced in [Matt 92, Matt 93]
Universal reservoir computers using non-homogeneous state-affine systems
3
for absolutely summable systems. The proofs in those works were unfortunately based on an invalid
compactness assumption. Even though corrections were proposed in [Perr 96, Stub 97a], this approach
yields, in the best of cases, universality only within the reservoir filter category, while we aim at proving
that statement in the much larger category of fading memory filters.
Our paper contains the following four main contributions:
• A non-homogeneous variant of the state-affine systems in [Flie 80] is introduced and we identify
sufficient conditions that guarantee that the associated reservoir computers with linear readouts
have the echo state property, are causal, time-invariant, and satisfy the fading memory property.
• A subset of this class is characterized that is universal in the category of fading memory filters
with uniformly bounded inputs.
• These results are extended to the stochastic setup by formulating a version of this universality
result that is valid for almost surely uniformly bounded inputs. This result shows that any discretetime filter that has the fading memory property with almost surely uniformly bounded stochastic
inputs can be uniformly approximated by elements in the non-homogeneous state-affine family.
• The universal non-homogeneous state-affine family suggests a generalization of the echo state
networks introduced in [Jaeg 04] that have been very successful in many information processing
tasks. We call this generalization echo state affine systems (ESAS) and we empirically show
that they outperform echo state networks in a standard benchmark forecasting task having to do
with the chaotic time series generated by the Mackey-Glass system [Mack 77].
Despite some preexisting work on the uniform approximation in probability of stochastic systems
with finite memory [Perr 96, Perr 97, Stub 97b], the universality result in the stochastic setup is, to our
knowledge, the first of its type in the literature and opens the door to new developments in the learning of
stochastic processes and their obvious applications to forecasting [Galt 14]. Indeed, in the deterministic
setup, RC has been very successful (see for instance [Jaeg 04]) at the time of learning the attractors
of various dynamical systems which, in passing, is used for forecasting by continuing the paths of the
system in question using its synthetically learnt proxy. This approach led to several orders of magnitude
accuracy improvements with respect to most standard dynamical systems forecasting tecniques based
on Takens’ Theorem [Take 81] and we expect that should have analogous beneficial effects in the density
forecasting of stochastic processes that satisfy the hypotheses of the results that are formulated later
on in the paper.
2
Notation, definitions, and preliminary discussions
Vector and matrix notations. Polynomials. A column vector is denoted by a bold lower case
symbol like r and r> indicates its transpose. Given a vector v ∈ Rn , we denote its entries by vi , with
i ∈ {1, . . . , n}; we also write v = (vi )i∈{1,...,n} . We denote by Mn,m the space of real n × m matrices
with m, n ∈ N. When n = m, we use the symbol Mn to refer to the space of square matrices of
order n. Given a matrix A ∈ Mn,m , we denote its components by Aij and we write A = (Aij ), with
i ∈ {1, . . . , n}, j ∈ {1, . . . m}. Given a vector v ∈ Rn , the symbol kvk stands for its Euclidean norm.
For any A ∈ Mn,m , kAk2 denotes its matrix norm induced by the Euclidean norms in Rm and Rn , and
satisfies [Horn 13, Example 5.6.6] that kAk2 = σmax (A), with σmax (A) the largest singular value of A.
kAk2 is sometimes referred to as the spectral norm of A [Horn 13].
Given an element z ∈ Rn , we denote by R[z] the real-valued multivariate polynomials on z with real
coefficients. Analogously, Pol(Rn , R) will denote the set of real-valued polynomials on Rn . When z ∈ R
Universal reservoir computers using non-homogeneous state-affine systems
4
is a scalar and m, n ∈ N, we define the set Mm,n [z] of Mm,n -valued polynomials on z with coefficients
in Mm,n as
Mm,n [z] := {A0 + zA1 + z 2 A2 + · · · + z N AN | N ∈ N, A0 , A1 , A2 , . . . , AN ∈ Mm,n }.
(2.1)
Filters. The symbol (Rn )Z denotes the set of infinite real sequences of the form z = (. . . , z−1 , z0 , z1 , . . .),
zi ∈ Rn ; (Rn )Z− and (Rn )Z+ are the subspaces consisting of, respectively, left and right infinite sequences: (Rn )Z− = {z = (. . . z−2 , z−1 , z0 ) | zi ∈ Rn }, (Rn )Z+ = {z = (z0 , z1 , z2 , . . .) | zi ∈ Rn }. In most
cases we shall use in these infinite product spaces either the product topology (see [Munk 14, Chapter
2]) or the topology induced by the supremum norm kzk∞ := supn∈Z {. . . , z−1 , z0 , z1 , . . .}. The symbols
n
`∞ (Rn ) and `∞
± (R ) will be used to denote the Banach spaces formed by the elements in those infinite
product spaces that have a finite supremum norm k · k∞ .
We will refer to the maps of the type U : (Rn )Z −→ RZ as filters or operators and to those like
H : (Rn )Z −→ R (or H : (Rn )Z± −→ R) as functionals. A filter U : (Rn )Z −→ RZ is called causal
when for any two elements z, w ∈ RZ that satisfy that zτ = wτ for any τ ≤ t, for a given t ∈ Z, we have
that (U z)t = (U w)t . The filter U is called time-invariant when, for any τ ∈ Z, it commutes with the
associated time delay operator Uτ : (Rn )Z −→ (Rn )Z defined by (Uτ z)t := zt−τ , that is, Uτ ◦ U = U ◦ Uτ
(in this expression, the two time delay operators Uτ have to be understood as defined in the appropriate
sequence spaces). We recall (see for instance [Boyd 85]) that there is a bijection between causal timeinvariant filters and functionals on (Rn )Z− . Indeed, given a time-invariant filter U , we can associate
to it a functional HU : (Rn )Z− −→ R via the assignment HU (z) := U (ze )0 , where ze ∈ (Rn )Z is an
arbitrary extension of z ∈ (Rn )Z− to (Rn )Z . Conversely, for any functional H : (Rn )Z− −→ R, we can
define a time-invariant causal filter UH : (Rn )Z −→ (Rn )Z by UH (z)t := H(PZ− ◦ U−t (z)), where U−t is
the (−t)-time delay operator and PZ− : (Rn )Z −→ (Rn )Z− is the natural projection. It is easy to verify
that:
HUH
= H,
for any functional H : (Rn )Z− −→ R,
UHU
= U,
for any causal time-invariant filter
U : (Rn )Z −→ RZ .
Additionally, let H1 , H2 : (Rn )Z− −→ R and λ ∈ R, then UH1 +λH2 z = UH1 z+λUH2 z, for any z ∈ (Rn )Z .
Reservoir filters. Consider now the RC system determined by (1.1)–(1.2) and assume, additionally,
Z
that the following existence and uniqueness property holds: for each z ∈ (Rn ) there exists a unique
Z
x ∈ RN such that for each t ∈ Z, the relation (1.1) holds. This condition is known in the literature as
the echo state property [Jaeg 10, Yild 12] and has deserved much attention in the context of echo state
networks [Jaeg 04, Bueh 06, Bai 12, Wain 16, Manj 13]. We emphasize that the echo state property is
a genuine condition that is not automatically satisfied by all RC systems.
RC systems that satisfy the echo state property have a naturally associated filter. We will denote by
U F : (Rn )Z −→ (RN )Z the filter determined by the reservoir map via (1.1), that is, (U F z)t := xt ∈ RN
and by UhF : (Rn )Z −→ RZ the one determined by the entire reservoir system, that is, (UhF z)t := yt . UhF
will be called the reservoir filter associated to the RC system (1.1)–(1.2). The filters U F and UhF are
causal by construction and it can also be shown that they are necessarily time-invariant [Grig 18]. We
can hence associate to UhF a reservoir functional HhF : (Rn )Z− −→ R determined by HhF := HUhF .
Weighted norms and the fading memory property (FMP). Let w : N −→ (0, 1] be a decreasing
sequence with zero limit and D ⊂ Rn . We define the associated weighted norm k·kw on DZ− ⊂ (Rn )Z−
associated to the weighting sequence w as the map:
k · k w : D Z−
z
R+
−→
7−→ kzkw := supt∈Z− kzt w−t k,
Universal reservoir computers using non-homogeneous state-affine systems
where k · k denotes the Euclidean norm in Rn . It is worth noting that the space
n
o
n
n Z−
`∞
| kzkw < ∞ ,
w (R ) := z ∈ (R )
5
(2.2)
endowed with weighted norm k · kw forms a Banach space [Grig 18].
All along the paper, we will work with uniformly bounded families of sequences, both in the
deterministic and the stochastic setups. The two main properties of these subspaces in relation with
the weighted norms are spelled out in the following two lemmas.
Lemma 2.1 Let M > 0 and let KM be the set of uniformly bounded elements in DZ− by M , that is,
KM := z ∈ DZ− | kzt k ≤ M for all t ∈ Z− .
(2.3)
Then, for any weighting sequence w and z ∈ KM , we have that kzkw < ∞.
Additionally, let λ, ρ ∈ (0, 1) and let w, wρ , w1−ρ be the weighting sequences given by wt := λt ,
ρ
wt := λρt , wt1−ρ := λ(1−ρ)t , t ∈ N. Then, the following series are convergent and satisfy the inequalities:
∞
X
t=0
∞
X
kz−t kwt
kz−t kwt
t=0
=
=
∞
X
t=0
∞
X
1
,
1 − λρ
(2.4)
1
.
1 − λ1−ρ
(2.5)
kz−t kλt ≤ kzkw1−ρ
kz−t kλt ≤ kzkwρ
t=0
The following result is a discrete-time version of Lemma 1 in [Boyd 85] that is easily obtained by
noticing that in the discrete-time setup all functions are trivially continuous if we consider the discrete
topology for their domains and, moreover, all families of functions are equicontinuous. A proof is given
in the appendices for the sake of completeness.
Lemma 2.2 Let M > 0 and let KM be as in (2.3). Let w : N −→ (0, 1] be an arbitrary weighting
sequence. Then KM is a compact topological space when endowed with the relative topology inherited
from (DZ− , k · kw ).
Definition 2.3 Let K ⊂ DZ− and let HU : (Rn )Z− −→ R be the functional associated to the causal
and time-invariant filter U : (Rn )Z −→ (Rn )Z . We say that U has the fading memory property
(FMP) on K whenever there exists a weighting sequence w : N −→ (0, 1] such that the map HU : K ⊂
(DZ− , k · kw ) −→ R is continuous. This means that for any z ∈ K and any > 0 there exists a δ() > 0
such that for any s ∈ K that satisfies that
kz − skw = sup k(zt − st )w−t k < δ(),
then
|HU (z) − HU (s)| < .
t∈Z−
If the weighting sequence w is such that wt = λt , for some λ ∈ (0, 1) and all t ∈ N, then U is said to
have the λ-exponential fading memory property.
Remark 2.4 The fading memory property is in some occasions also related to the Lyapunov stability
of the autonomous system associated to the reservoir map. This connection has been made explicit, for
example, for discrete-time nonlinear state-space models that are affine in their inputs, and have direct
feed-through term in the output [Zang 04] or for time-delay reservoirs [Grig 16b].
Remark 2.5 Time-invariant fading memory filters always have the bounded input, bounded output
(BIBO) property. Indeed, if for simplicity we consider functionals HU that map the zero input to zero,
that is HU (0) = 0, and we want bounded outputs such that |HU (z)| < k, for a given constant k > 0,
by Definition 2.3 it suffices to consider inputs z ∈ (RN )Z− such that kzk∞ := supt∈Z− kzt k =< δ(k).
Indeed, if HU has the FMP with respect to a weighting sequence w, then kzkw ≤ kzk∞ < δ(k) and
hence |HU (z)| < k, as required.
Universal reservoir computers using non-homogeneous state-affine systems
6
The following lemma, that will be used later on in the paper, spells out how the FMP depends on
the weighting sequence used to define it.
Lemma 2.6 Let K ⊂ DZ− and let HU : (Rn )Z− −→ R be the functional associated to the causal and
time-invariant filter U : (Rn )Z −→ (Rn )Z . If HU : K ⊂ DZ− −→ R has the FMP with respect to a given
weighting sequence w, then it also has it with respect to any other weighting sequence w0 that satisfies
that
wt
< λ, for a fixed λ > 0 and for all t ∈ N.
wt0
In particular, the thesis of the lemma holds when w0 dominates w, that is when λ = 1.
It can be shown [Grig 18] that when in this lemma the set K is made of uniformly bounded sequences,
that is, K = KM , with KM as in (2.1), then if a filter has the FMP with respect to a given weighting
sequence, it necessarily has the same property with respect to any other weighting sequence.
3
Universality results in the deterministic setup
In this section we consider deterministic filters, in the sense that their inputs belong to a subset of
KM ⊂ (Rn )Z formed by uniformly bounded elements, as in the definition in (2.3).
We will formulate two different universality results. In the first one, we show that polynomial
algebras of filters generated by reservoir computers with the fading memory property that separate
points are able to approximate any fading memory filter. Such families of reservoir computers are said
to be universal. Two important consequences of this result is that the entire family of fading memory
RCs itself is universal, as well as the one containing all the linear reservoirs with polynomial readouts,
when certain spectral restrictions are imposed on the reservoir matrices. In the second result, we restrict
ourselves to reservoir computers with linear readouts and introduce the non-homogeneous state-affine
family in order to be able to obtain a similar universality statement.
The first result can be seen as a discrete-time translation of the one formulated in [Boyd 85] for
continuous-time filters while the second one is an extension to infinite time intervals of the main approximation result in [Flie 80] formulated for compact time intervals using homogeneous state-affine
systems.
3.1
Universality for fading memory RCs with non-linear readouts
The following statement is a direct consequence of the compactness result in Lemma 2.3 and the StoneWeierstrass theorem for polynomial subalgebras of real-valued functions defined on compact metric
spaces, as formulated in Theorem 7.3.1 in [Dieu 69]. See Appendix 6.4 for a detailed proof.
Theorem 3.1 Let KM ⊂ (Rn )Z− be a subset of the type defined in (2.3) and let
R := {HhFii : KM −→ R | hi ∈ C ∞ (RNi ), Fi : RNi × Rn −→ RNi , i ∈ I, Ni ∈ N}
(3.1)
be a set of reservoir filters defined on KM that have the FMP with respect to a given weighted norm
k · kw . Let A(R) be the polynomial algebra generated by R, that is, the set formed by finite products
and linear combinations of elements in R. If the algebra A(R) contains the constant functionals and
separates the points in KM , then any causal, time-invariant fading memory filter H : KM −→ R can
be uniformly approximated by elements in A(R), that is, A(R) is dense in the set (C 0 (KM ), k · kw )
of real-valued continuous functions on (KM , k · kw ). More explicitly, this implies that for any fading
memory filter H and any > 0, there exist indices {i1 , . . . , ir } ⊂ I and a polynomial p : Rr −→ R such
that
kH − HhF k∞ := sup |H(z) − HhF (z)| <
z∈KM
with
h := p(hi1 , . . . , hir )
and
F := (Fi1 , . . . , Fir ).
Universal reservoir computers using non-homogeneous state-affine systems
7
An important fact is that the polynomial algebra A(R) generated by a set formed by fading memory
reservoir filters is made of fading memory reservoir filters. Indeed, let hi ∈ C ∞ (RNi ), Fi : RNi × Rn −→
RNi , i ∈ {1, 2}, and λ ∈ R. Then, the product HhF11 · HhF22 and the linear combination HhF11 + λHhF22 filters
are such that
HhF11 · HhF22
HhF11
+
λHhF22
=
HhF ,
=
HhF0 ,
with
h := h1 · h2 ∈ C ∞ (RN1 +N2 ),
0
∞
with h := h1 + λh2 ∈ C (R
N1 +N2
(3.2)
),
(3.3)
and where
F : RN1 +N2 × Rn −→ RN1 +N2
is the map given by F ((x1 , x2 ), z) := (F1 (x1 , z), F2 (x2 , z)) .
(3.4)
We emphasize that the functionals HhF and HhF0 in (3.2) and (3.3) are well defined because if the reservoir
Z
Z
maps F1 and F2 satisfy the echo state property then so does F . Indeed, if x1 ∈ RN1 and x2 ∈ RN2
Z
are the solutions of the reservoir equation (1.1) for F1 and F2 associated to the input z ∈ (Rn ) , then
Z
so is x1 ⊕ x2 ∈ RN1 +N2 for F in (3.4).
This observation has as a consequence that the set of all the RC systems that have the echo state
property and the FMP with respect to a given weighted norm k · kw form a polynomial algebra that contains the constant functions (they can be obtained by using as readouts constant elements in C ∞ (RNi ))
and separate points (take the trivial reservoir map F (x, z) = z and use the separation property of
C ∞ (Rn )). This remark, put together with Theorem 3.1 yields the following corollary.
Corollary 3.2 Let KM ⊂ (Rn )Z− be a subset of the type defined in (2.3) and let
Rw := {HhF : KM −→ R | h ∈ C ∞ (RN ), F : RN × Rn −→ RN , N ∈ N}
(3.5)
be the set of all reservoir filters defined on KM that have the FMP with respect to a given weighted norm
k · kw . Then Rw is universal, that is, it is dense in the set (C 0 (KM ), k · kw ) of real-valued continuous
functions on (KM , k · kw ).
According to the previous corollary, reservoir filters that have the FMP are able to approximate any
time-invariant fading memory filter. We now show that actually a much smaller family of reservoirs
suffices to do that, namely, certain classes of linear reservoirs with polynomial readouts. Consider the
RC system determined by the expressions
xt = Axt−1 + czt ,
yt = h(xt ),
A ∈ MN , c ∈ MN,n ,
h ∈ R[x].
(3.6)
(3.7)
If this system has a reservoir filter associated (the next result provides a sufficient condition for this
to happen) we will denote by HhA,c : K −→ R the associated functional and we will refer to it as the
linear reservoir filter determined by A, c, and the polynomial h. These filters exhibit the following
universality property that is proved in Appendix 6.5.
Corollary 3.3 Let KM ⊂ (Rn )Z− be a subset of the type defined in (2.3) and let ∈ R be such that
1 > > 0. Consider the set L formed by all the linear reservoir systems as in (3.6)-(3.7) determined
by polynomial readouts and by matrices A ∈ MN such that σmax (A) < 1 − . Then, the elements in L
generate λρ -exponential fading memory reservoir filters, with λρ := (1 − )ρ , for any ρ ∈ (0, 1), that can
be explicitly written down using the expression:
!
∞
X
A,c
i
Hh (z) = h
A cz−i , for any z ∈ KM .
(3.8)
i=0
This family is dense in (C (KM ), k · kwρ ) with wtρ := λtρ , for any ρ ∈ (0, 1).
0
Universal reservoir computers using non-homogeneous state-affine systems
3.2
8
State-affine systems and universality for fading memory RCs with linear
readouts
As it was explained in the introduction, it is is particularly convenient to work with RCs that have a
linear readout since in that case the training reduces to the solution of a regression problem. That is
in most cases feasible when there is need, as it happens in many applications, for a high number of
neurons. This point makes relevant the identification of families of reservoirs that are universal when
the readout is restricted to be linear, which is the subject of this subsection. In order to simplify the
presentation, we restrict ourselves in this case to one-dimensional input signals, that is, all along this
section we set n = 1. The extension to the general case is straightforward even though more convoluted
to write down (see Remark 3.12).
Definition 3.4 Let N ∈ N, W ∈ RN , and let p(z) ∈ MN,N [z] and q(z) ∈ MN,1 [z] be two polynomials on
the variable z with matrix coefficients, as they were introduced in (2.1). The non-homogeneous stateaffine system (SAS) associated to p, q and W is the reservoir system determined by the state-space
transformation:
(
xt = p(zt )xt−1 + q(zt ),
(3.9)
yt = W > x t .
(3.10)
Notice that the linear reservoir equation (3.6) is a particular case of the SAS state equation (3.9)
that is obtained by taking for p and q polynomials of degree zero.
Our next result spells out a sufficient condition that guarantees that the SAS reservoir system (3.9)(3.10) has the echo state property. Moreover, it provides an explicit expression for the unique causal
and time-invariant solution associated to a given input. Recall that for any A ∈ Mn,m , kAk2 denotes
its matrix norm induced by the Euclidean norms in Rm and Rn and that kAk2 = σmax (A).
Proposition 3.5 Consider a non-homogeneous state-affine system as in (3.9)-(3.10) determined by
polynomials p, q, and a vector W, with inputs defined on I Z , I := [−1, 1]. Assume that
max kp(z)k2 = max σmax (p(z)) < 1.
z∈I
(3.11)
z∈I
Then, the reservoir system (3.9)-(3.10) has the echo state property and for each input z ∈ RZ it has a
unique causal and time-invariant solution given by
!
j−1
∞
Y
X
x =
p(zt−k ) q(zt−j ),
(3.12)
t
j=0 k=0
yt = W> xt .
(3.13)
p,q
p,q
In that situation, we will denote by UW
: I Z −→ I Z and HW
: I Z− −→ R the corresponding SAS
reservoir filter and SAS functional, respectively.
The next result presents two alternative conditions that imply the hypothesis maxz∈I kp(z)k2 < 1
in the previous proposition and that are easier to verify in practice.
Lemma 3.6 Let p(z) ∈ MN,N [z] be the polynomial given by
p(z) := A0 + zA1 + z 2 A2 + · · · + z n1 An1 ,
Consider the following three conditions:
n1 ∈ N.
Universal reservoir computers using non-homogeneous state-affine systems
9
(i) There exists a constant 0 < λ < 1, such that kAi k2 = σmax (Ai ) < λ, for any i ∈ {0, 1, . . . , n1 }, and
that at the same time satisfies that λ(n1 + 1) < 1.
(ii) Bp := kA0 k2 + kA1 k2 + · · · + kAn1 k2 < 1.
(iii) Mp := maxz∈I kp(z)k2 < 1.
Then, condition (i) implies (ii) and condition (ii) implies (iii).
We emphasize that since Proposition 3.5 was proved using condition (iii) in the previous lemma
then, any of the three conditions in that statement imply the echo state property for (3.12)-(3.13) and
the time-invariance of the corresponding solutions. The next result shows that the same situation holds
in relation with the fading memory property.
Proposition 3.7 Consider a non-homogeneous state-affine system as in (3.9)-(3.10) determined by
polynomials p, q, and a vector W, with inputs defined on I Z , I := [−1, 1]. If the polynomial p satisfies
any of the three conditions in Lemma 3.6 then the corresponding reservoir filter has the fading memory
p,q
property. More specifically, if p satisfies condition (i) in Lemma 3.6, then HW
: (I Z− , k · kwρ ) −→ R is
ρ
ρt ρt
continuous with wt := (n1 + 1) λ and ρ ∈ (0, 1) arbitrary. The same conclusion holds for conditions
(ii) and (iii) with wtρ := Bpρt and wtρ := Mpρt , respectively.
The importance of SAS in relation to the universality problem has to do with the fact that they
form a polynomial algebra which allows us, under certain conditions, to use the Stone-Weierstrass
theorem to prove a density statement. Before we show that, we observe that for any two polynomials
p1 (z) ∈ MN1 ,M1 [z] and p2 (z) ∈ MN2 ,M2 [z] given by
p1 (z)
p2 (z)
:= A10 + zA11 + z 2 A12 + · · · + z n1 A1n1 ,
:=
A20
+
zA21
+z
2
A22
+ ··· + z
n2
A2n2 ,
(3.14)
(3.15)
with n1 , n2 ∈ N, their direct sum p1 ⊕p2 (z) ∈ MN1 +N2 ,M1 +M2 [z] and their Kronecker product p1 ⊗p2 (z) ∈
MN1 ·N2 ,M1 ·M2 [z] is written as
p1 ⊕p2 (z) = A10 ⊕A20 +zA11 ⊕A21 +z 2 A12 ⊕A22 +· · ·+z n2 A1n2 ⊕A2n2 +z n2 +1 A1n2 +1 ⊕0+· · ·+z n1 A1n1 ⊕0, (3.16)
where we assumed that n2 ≤ n1 . Analogously,
p1 ⊗ p2 (z) =
n1 X
n2
X
z i+j A1i ⊗ A2j .
(3.17)
i=0 j=0
p1 ,q1
p2 ,q2
Proposition 3.8 Let D ⊂ R be an open set and let HW
, HW
: DZ− −→ R be two SAS reservoir
1
2
functionals associated to two corresponding time-invariant SAS reservoir systems that have V1 and V2 as
state spaces, respectively. Assume that the two polynomials with matrix coefficients p1 (z) ∈ MN1 ,M1 [z]
and p2 (z) ∈ MN2 ,M2 [z] satisfy that kp1 (z)k2 < 1 − and kp2 (z)k2 < 1 − for all z ∈ I := [−1, 1] and a
given 1 > > 0. Then, with the notation introduced in the expressions (3.16) and (3.17), we have that:
p1 ,q1
p2 ,q2
(i) For any λ ∈ R, the linear combination HW
+ λHW
is a SAS reservoir functional associated to
1
2
a SAS that has V1 ⊕ V2 as state space and:
p2 ,q2
p1 ⊕p2 ,q1 ⊕q2
p1 ,q1
HW
+ λHW
= HW
.
2
1
1 ⊕λW2
(3.18)
Universal reservoir computers using non-homogeneous state-affine systems
10
p1 ,q1
p2 ,q2
(ii) The product HW
· HW
is a SAS reservoir functional associated to a SAS that has V1 ⊕ V2 ⊕
1
2
(V1 ⊗ V2 ) as state space and:
p,q ⊕q ⊕(q ⊗q )
p1 ,q1
p2 ,q2
1
2
1
2
HW
· HW
= H0⊕0⊕(W
,
1
2
1 ⊗λW2 )
(3.19)
where p(z) ∈ MN12 [z], N12 := dim V1 + dim V2 + dim V1 · dim V2 , is the polynomial with matrix
coefficients whose block-matrix expression for the three summands in V1 ⊕ V2 ⊕ (V1 ⊗ V2 ) is:
p1
0
0
.
0
p2
0
p :=
(3.20)
p1 ⊗ q2 q1 ⊗ p2 p1 ⊗ p2
The expression p1 ⊗ q2 (respectively, q1 ⊗ p2 ) denotes the linear map from V1 (respectively, V2 ) to
V1 ⊗V2 that associates to any v1 ∈ V1 the element (p1 (z)v1 )⊗q2 (z) (respectively, q1 (z)⊗(p2 (z)v2 )).
The equalities (3.18) and (3.19) show that the SAS family forms a polynomial algebra.
Theorem 3.9 (Universality of SAS reservoir computers) Let I Z− ⊂ RZ− be the subset of real
uniformly bounded sequences in I := [−1, 1] as in (2.3), that is,
I Z− := {z ∈ RZ− | zt ∈ [−1, 1], for all
t ≤ 0},
p,q
and let S be the family of functionals HW
: I Z− −→ R induced by the state-affine systems defined
in (3.9)-(3.10) that satisfy that Mp := maxz∈I kp(z)k2 < 1 − and Mq := maxz∈I kq(z)k2 < 1 − . The
family S forms a polynomial subalgebra of Rwρ (as defined in (3.5)) with wtρ := (1 − )ρt and ρ ∈ (0, 1)
arbitrary, made of fading memory reservoir filters that contains the constant functions and separate
points. The subfamily S is hence dense in the set (C 0 (I Z− ), k · kwρ ) of real-valued continuous functions
on (I Z− , k · kwρ ) which implies that any causal, time-invariant fading memory filter H : I Z− −→ R
can be uniformly approximated by elements in S . More specifically, for any fading memory filter H
and any > 0, there exist a natural number N ∈ N, polynomials p(z) ∈ MN,N [z], q(z) ∈ MN,1 [z] with
Mp , Mq < 1 − , and a vector W ∈ RN such that
p,q
p,q
kH − HW
k∞ := sup |H(z) − HW
(z)| < .
z∈I Z−
Remark 3.10 As it is stated in Theorem 3.9, it is the condition (iii) in Lemma 3.6 that yields a
universal family of SAS fading memory reservoirs. As it can deduced from its proof (available in
the Appendix 6.10), the families determined by conditions (i) or (ii) in that lemma contain SAS
fading memory reservoirs but they do not form a polynomial algebra. In such cases, and according to
Theorem 3.1, it is the algebras generated by them and not the families themselves that are universal.
Remark 3.11 A continuous-time analog of the universality result that we just proved can be obtained
using the bilinear systems considered in Section 5.3 of [Boyd 85]. In discrete time, but only when the
number of time steps is finite, this universal approximation property is exhibited [Flie 80] by homogeneous state-affine systems, that is, by setting q(z) = 0 in (3.9)-(3.10).
Remark 3.12 Generalization to multidimensional signals. When the input signal is defined in
InZ , with In := [−1, 1]n , a SAS family with the same universality properties can be defined by replacing
the polynomials p and q in Definition 3.4, by polynomials of the form:
X
iN
p(z) =
z1i1 · · · zN
Ai1 ,...,iN ,
Ai1 ,...,iN ∈ MN,N ,
i1 ,...,iN ∈{0,...,r}
i1 +···+iN ≤r
q(z)
=
X
i1 ,...,iN ∈{0,...,s}
i1 +···+iN ≤s
iN
z1i1 · · · zN
Bi1 ,...,iN ,
Ai1 ,...,iN ∈ MN,1 .
Universal reservoir computers using non-homogeneous state-affine systems
11
Remark 3.13 SAS with trigonometric polynomials. An analogous construction can be carried
out using trigonometric polynomials of the type:
X
p(z) =
cos (i1 · z1 + · · · + iN · zN ) Ai1 ,...,iN ,
Ai1 ,...,iN ∈ MN,N ,
i1 ,...,iN ∈{0,...,r}
i1 +···+iN ≤r
q(z)
X
=
cos (i1 · z1 + · · · + iN · zN ) Bi1 ,...,iN ,
Ai1 ,...,iN ∈ MN,1 .
i1 ,...,iN ∈{0,...,s}
i1 +···+iN ≤s
In this case, it is easy to establish that the resulting SAS family forms a polynomial algebra by invoking
Proposition 3.8 and by reformulating the expressions (3.16) and (3.17) using the trigonometric identity
cos(θ) cos(φ) =
1
(cos(θ − φ) + cos(θ + φ)) .
2
Additionally, the corresponding SAS family includes the linear family and hence the point separation
property can be established as in the proof of Theorem 3.9 in the Appendix 6.10.
4
Reservoir universality results in the stochastic setup
This section extends the previously stated deterministic universality results to a setup in which the
reservoir inputs and outputs are stochastic, that is, the reservoir is not driven anymore by infinite
sequences but by discrete-time stochastic processes. We emphasize that we restrict our discussion to
reservoirs that are deterministic and hence the only source of randomness in the systems considered is
the stochastic nature of the input.
The stochastic setup. All along this section we work on a probability space (Ω, A, P). If a condition
defined on this probability space holds everywhere except for a set with zero measure, we will see that
the relation is true almost surely and we will abbreviate it a.s. We will denote by L∞ (Ω, Rn ) the
set of Rn -valued random variables whose Euclidean norms have a finite essential supremum or that,
equivalently, have almost surely bounded Euclidean norms. More specifically, let X : Ω −→ Rn be a
random variable and let
kXkL∞ := ess sup kX(ω)k = inf {ρ ∈ R+ | kXk < ρ almost surely} .
(4.1)
L∞ (Ω, Rn ) := {X : Ω −→ Rn random variable | kXkL∞ < ∞} .
(4.2)
ω∈Ω
We define
∞
n
It can be shown that L (Ω, R ) is a Banach space (see [Ledo 91, pages 42 and 46], [Lord 14, page 149].
Given an element X ∈ L∞ (Ω, Rn ), we denote by E [X] its expectation. The following lemma collects
some elementary results that will be needed later on and shows, in particular, that the expectation E [X]
k
as well as that of all the powers kXk of its norm are finite for all the elements X ∈ L∞ (Ω, Rn ).
Lemma 4.1 Let X ∈ L∞ (Ω, Rn ) and let C > 0 be a real number. Then:
(i) kXk ≤ kXkL∞ almost surely.
(ii) kXkL∞ ≤ C if and only if kXk ≤ C almost surely.
(iii) kXk ≤ C almost surely if and only if E kXkk ≤ C k for any k ∈ N.
(iv) The components Xi of X, i ∈ {1, . . . , n}, are such that E [Xi ] ≤ kXkL∞ .
Universal reservoir computers using non-homogeneous state-affine systems
12
The first point in this lemma explains why we will refer to the elements of L∞ (Ω, Rn ) as almost surely
bounded random variables.
Inputs and outputs. The filters that we will consider in this section have almost surely bounded
time series or discrete-time stochastic processes as inputs and outputs. Recall that a discrete-time
stochastic process is a map of the type:
z: Z×Ω
(t, ω)
−→
7−→
Rn
zt (ω),
(4.3)
such that, for each t ∈ Z, the assignment zt : Ω −→ Rn is a random variable. For each ω ∈ Ω, we
will denote by z(ω) := {zt (ω) ∈ Rn | t ∈ Z} the realization or the sample path of the process z.
Additionally, we use the symbol F(z) := {Ft (zt ) | t ∈ Z}, with Ft (zt ) := σ ({zs | s ≤ t}) the sigma
algebra generated by the process z up to time t, to refer to the filtration generated by z. The results
that follow are presented for stochastic processes indexed by Z but are equally valid for Z+ and Z− .
Recall that a map of the type (4.3) is a Rn -valued stochastic process if and only if the corresponding
Z
map z : Ω −→ (Rn ) into path space (designated with the same symbol) is a random variable when in
Z
(Rn ) we consider the product sigma algebra generated by cylinder sets [Come 06, Chapter 1]. Then,
Z
using the same prescription as in (4.1) with the supremum norm k · k∞ in (Rn ) , we can define a norm
n
k · kL∞ in the space of R -valued stochastic processes by setting
(4.4)
kzkL∞ := ess sup kz(ω)k∞ = ess sup sup {kzt (ω)k} .
ω∈Ω
ω∈Ω
t∈Z
The following lemma provides an alternative characterization of the norm k·kL∞ that will be very useful
in the proofs of the results that follow and in which the supremum and the essential supremum have
been intertwined. The last statement contained in it complements part (ii) of Lemma 4.1 for processes.
Lemma 4.2 Let z : Ω −→ (Rn ) be a stochastic process. Then,
kzkL∞ := ess sup sup {kzt (ω)k} = sup ess sup {kzt (ω)k} .
Z
ω∈Ω
t∈Z
t∈Z
(4.5)
ω∈Ω
Equivalently, using the notation in Lemma 4.1,
= sup kzt (ω)kL∞ .
kzkL∞ := sup kzt (ω)k
t∈Z
L∞
(4.6)
t∈Z
These equalities imply that for any positive real number C > 0, the process z satisfies that kzkL∞ ≤ C
if and only if kzt kL∞ ≤ C for all t ∈ Z or, equivalently, if and only if supt∈Z kzt kL∞ ≤ C.
We now consider the space L∞ Ω, (Rn )Z of processes with finite k·kL∞ norm, that is,
L∞ Ω, (Rn )Z := {z : Z × Ω −→ Rn stochastic process | kzkL∞ < ∞} .
We refer to the elements of L∞ Ω, (Rn )Z as almost surely bounded time series. The following
result shows that L∞ Ω, (Rn )Z coincides with the space of processes that take values in the Banach
space (`∞ (Rn ), k · k∞ ) and it is hence a Banach space.
Lemma 4.3 In the setup that we just introduced
L∞ Ω, (Rn )Z = L∞ (Ω, `∞ (Rn ))
and L∞ Ω, (Rn )Z , k · kL∞ is a Banach space.
(4.7)
Universal reservoir computers using non-homogeneous state-affine systems
13
Let now w be a weighting sequence and let k · kw be the associated weighted norm in (Rn ) − . If we
replace in (4.4) the `∞ norm k · k∞ by the weighted norm k · kw , we obtain a weighted norm k · kL∞
in
w
the space of processes z : Z− × Ω −→ Rn indexed by Z− as:
kzkL∞
:=
ess
sup
kz(ω)k
=
ess
sup
sup
{kz
(ω)kw
}
.
(4.8)
t
−t
w
w
Z
ω∈Ω
ω∈Ω
t∈Z
n Z−
We will denote by L∞
the space of processes with finite k·kL∞
norm. Since the space
w Ω, (R )
w
∞
n
`w (R ) introduced
in (2.2) is a Banach space [Grig 18] and a result similar
to Lemma 4.3 shows that
n Z−
n
∞
n Z−
L∞
= L∞ (Ω, `∞
(R
)
is
a
Banach space. Additionw Ω, (R )
w (R )), we can conclude that Lw Ω,
n Z−
:
Ω,
(R
)
ally, as in Lemma 4.2, we have that for any z ∈ L∞
w
(4.9)
kzkL∞ := ess sup sup {kzt (ω)kw−t } = sup ess sup {kzt (ω)kw−t } .
w
ω∈Ω
t∈Z
t∈Z
ω∈Ω
Deterministic filters in a stochastic setup. As we already pointed out, we consider filters U that
have almost surely bounded time series inputs and outputs, that is, if the inputs can take values in
Rn then U : L∞ (Ω, (Rn )Z ) −→ L∞ (Ω, RZ ). The same conventions as in the deterministic setup will be
used in the identification of the different signals, namely, z will denote the filter inputs, the symbol y is
reserved for the output, and d for the teaching signals. We emphasize that unlike in the deterministic
case, the components of all these signals are this time random variables and not points in a Euclidean
space.
The concepts of causality and time-invariance are defined as in the deterministic case by replacing equalities by almost sure equalities in the corresponding identities. The same applies to the correspondence between time-invariant filters U : L∞ (Ω, (Rn )Z ) −→ L∞ (Ω, RZ ) and functionals HU :
L∞ (Ω, (Rn )Z− ) −→ L∞ (Ω, R). We draw attention to the fact that assuming that the filters map into
almost surely bounded processes is a genuine hypothesis that will need to be verified in each specific
case considered.
We underline again that we will restrict our attention to intrinsically deterministic filters that are
obtained by presenting almost surely bounded stochastic inputs z ∈ L∞ (Ω, (Rn )Z− ) to causal and timeinvariant filters H : (Rn )Z− −→ R similar to those introduced in the previous section which, in some
cases, yields maps of the type H : L∞ (Ω, (Rn )Z− ) −→ L∞ (Ω, R) and that we will denote with the same
symbol. In that situation, the dependence on the probability space of the image (H(z))(ω) takes place
exclusively through the dependence z(ω) in the input. Causal and time-invariant deterministic filters
with stochastic inputs produce almost surely causal and time-invariant filters. Moreover, for any t ∈ Z,
the random variable UH (z)t := H(PZ− ◦ Ut (z)) = H (zt , zt−1 , zt−2 , . . .) is clearly Ft (zt )-measurable.
Given a weighting sequence w : N −→ (0, 1] and a time invariant filter U : L∞ (Ω, (Rn )Z ) −→
∞
L (Ω, RZ ) with stochastic inputs, we will say that U has the fading memory property with respect
to the weighting sequence w when the corresponding functional HU : (L∞ (Ω, (Rn )Z− ), k · kL∞
) −→
w
L∞ (Ω, R) is a continuous map.
Let M > 0 and define, using Lemma 4.2,
L∞
KM
:= z ∈ L∞ (Ω, (Rn )Z− ) | kzkL∞ ≤ M = z ∈ L∞ (Ω, (Rn )Z− ) | kzt kL∞ ≤ M, for all t ∈ Z− .
(4.10)
L∞
The sets KM
are the stochastic counterparts of the sets KM in the deterministic setup; we will say
L∞
that KM
is a set of almost surely uniformly bounded processes. A stochastic analog of Lemma
L∞
2.1 can be formulated for them with KM replaced by KM
, the norm k·k by k·kL∞ , and the weighted
norm k·kw by k·kL∞ .
The following result shows that the fading memory and the universality properties are naturally
inherited by deterministic filters with almost surely bounded inputs.
Universal reservoir computers using non-homogeneous state-affine systems
14
∞
L
Theorem 4.4 Let M > 0 and let KM and KM
be the sets of deterministic and stochastic inputs
defined in (2.3) and (4.10), respectively. The following properties hold true:
(i) Let H : (KM , k·kw ) −→ R be a causal and time-invariant filter. Then H : (KM , k·kw ) −→ R has
the fading memory property if and only if the corresponding filter with almost surely uniformly
L∞
bounded inputs has almost surely bounded outputs, that is, H : (KM
, k · k L∞
) −→ L∞ (Ω, R), and
w
it has the fading memory property.
(ii) Let T := {Hi : (KM , k·kw ) −→ R | i ∈ I} be a family of causal and time-invariant fading memory
filters. Then, T is dense in the set (C 0 (KM ), k · kw ) if and only if the corresponding family with
L∞
L∞
intputs in KM
is universal in the set of continuous maps of the type H : (KM
, k · kL∞
) −→
w
L∞ (Ω, R).
A first universality result using RC systems. The following result is a stochastic analog of
Theorem 3.1 and shows that any fading memory filter with almost surely uniformly bounded inputs
can be approximated using the elements of a polynomial algebra of reservoir filters with the same kind
of inputs, provided that it contains the constant functionals and has the separation property. We note
that, as in the deterministic case, the existence of the reservoir filter associated to a reservoir system
like (1.1)-(1.2) is guaranteed only in the presence of the echo state property. It is easy to verify that
if a reservoir system with uniformly bounded deterministic inputs satisfies the echo state property and
the FMP then so does its counterpart with almost surely uniformly bounded inputs.
∞
L
Theorem 4.5 Let M > 0 and let KM
duced in (4.10). Consider the set R
be the set of almost surely uniformly bounded processes intro-
∞
L
R := {HhFii : KM
−→ L∞ (Ω, R) | hi ∈ Pol(RNi , R), Fi : RNi × Rn −→ RNi , i ∈ I, Ni ∈ N}
(4.11)
formed by deterministic fading memory reservoir filters with respect to a given weighted norm k · kw
L∞
. Let A(R) be the polynomial algebra generated by R. If the
and driven by stochastic inputs in KM
algebra A(R) has the separation property and contains contains all the constant functionals, then any
L∞
deterministic, causal, time-invariant fading memory filter H : (KM
, k · k L∞
) −→ L∞ (Ω, R) can be
w
uniformly approximated by elements in A(R), that is, for any > 0, there exist indices {i1 , . . . , ir } ⊂ I
and a polynomial p : Rr −→ R such that
kH − HhF k∞ := sup kH(z) − HhF (z)kL∞ <
with
∞
h := p(hi1 , . . . , hir )
and
F := (Fi1 , . . . , Fir ).
L
z∈KM
In the next paragraphs we identify, as in the deterministic case, families of reservoirs that satisfy
the hypotheses of this theorem and where we will eventually impose linearity constraints on the readout
function. The following corollary to Theorem 4.5 is the stochastic analog of Corollary 3.2.
∞
L
Corollary 4.6 Let M > 0 and let KM
duced in (4.10). Let
be the set of almost surely uniformly bounded processes intro-
∞
L
Rw := {HhF : KM
−→ L∞ (Ω, R) | h ∈ Pol(RN , R), F : RN × Rn −→ RN , N ∈ N}
∞
(4.12)
L
be the set of all the reservoir filters defined on KM
that have the FMP with respect to a given weighted
L∞
norm k · kL∞
.
Then
R
is
universal,
that
is,
for
any
time-invariant fading memory filter H : (KM
,k·
w
w
∞
F
F
kL∞
)
−→
L
(Ω,
R)
and
any
>
0,
there
exists
a
reservoir
filter
H
∈
R
such
that
kH
−
H
k
w
h
h ∞ :=
w
F
supz∈KM
L∞ kH(z) − H (z)kL∞ < .
h
Universal reservoir computers using non-homogeneous state-affine systems
15
Linear reservoir computers with stochastic inputs are universal. As it was the case in the
deterministic setup, we can prove in the stochastic case that the linear RC family introduced in (3.6)(3.7) suffices to achieve universality. The proof of the following statement is a direct consequence of
Corollary 3.3 and Theorem 4.4.
∞
L
Corollary 4.7 Let M > 0 and let KM
be the set of almost surely uniformly bounded processes introduced in (4.10). Let L be the family introduced in Corollary 3.3 and formed by all the linear reservoir
filters HpA,c determined by matrices A ∈ MN such that σmax (A) < 1 − . The elements in L map
L∞
KM
into L∞ (Ω, R) and are time-invariant fading memory filters with respect to the weighted norm
L∞
k · kwρ associated to wtρ := (1 − )ρt , for any ρ ∈ (0, 1). Moreover, they are universal, that is, for any
L∞
time-invariant and causal fading memory filter H : (KM
, k · k L∞
) −→ L∞ (Ω, R) and any ε > 0, there
wρ
A,c
A,c
A,c
exists Hp ∈ L such that kH − Hp k∞ := supz∈KM
(z)kL∞ < ε.
L∞ kH(z) − Hp
Remark 4.8 The previous corollary has interesting consequences in the realm of time series analysis.
Indeed, many well-known parametric time series models consist in autoregressive relations, possibly
nonlinear, driven by independent or uncorrelated innovations. The parameter constraints that are
imposed on them in order to ensure that they have (second order) stationary solutions imply, in may
situations, that the resulting filter has the FMP. In those cases, Corollary 4.7 allows us to conclude that
when those models are driven by almost surely uniformly bounded innovations, they can be arbitrarily
well approximated by a polynomial function of a strong vector autoregressive model (VAR) of order 1.
This statement applies, for example, to any stationary ARMA [Box 76, Broc 06] or GARCH [Engl 82,
Boll 86, Fran 10] model driven by almost surely uniformly bounded innovations.
State-affine reservoir computers with almost surely uniformly bounded inputs are universal. As it was the case in the deterministic setup, non-homogeneous SAS are universal time-invariant
fading memory filters in the stochastic framework with almost surely uniformly bounded inputs. Their
advantage with respect to the families proposed in the previous corollary is that they use a linear readout which is of major importance in practical implementations. More specifically, the following result
holds true as a direct consequence of Theorem 3.9 and the equivalence stated in Theorem 4.4.
Theorem 4.9 (Universality of SAS reservoir computers with almost surely uniformly bounded
∞
inputs) Let KIL ⊂ L∞ (Ω, RZ− ) be the set of almost surely and uniformly bounded processes in the
interval I = [−1, 1], that is,
∞
KIL := z ∈ L∞ (Ω, RZ− ) | kzt kL∞ ≤ 1, for all t ∈ Z− .
∞
p,q
Let S be the family of functionals HW
: KIL −→ L∞ (Ω, R) induced by the state-affine systems defined
in (3.9)-(3.10) and that satisfy Mp := maxz∈I kp(z)k < 1 − and Mq := maxz∈I kq(z)k < 1 − ,
with I := [−1, 1]. The family S forms a polynomial subalgebra of Rwρ (as defined in (4.12)) with
wtρ := (1 − )ρt , made of fading memory reservoir filters that map into L∞ (Ω, R).
∞
Moreover, for any time-invariant and causal fading memory filter H : (KIL , k · kL∞
) −→ L∞ (Ω, R)
wρ
and any > 0, there exist a natural number N ∈ N, polynomials p(z) ∈ MN,N [z], q(z) ∈ MN,1 [z] with
Mp , Mq < 1 − , and a vector W ∈ RN such that
p,q
p,q
kH − HW
k∞ := sup kH(z) − HW
(z)kL∞ < .
∞
z∈KIL
Universal reservoir computers using non-homogeneous state-affine systems
5
16
Echo state affine systems (ESAS) and their forecasting performance
Echo state networks (ESNs) were introduced in [Jaeg 04] as one of the first families of reservoir computing architectures and have been very successful in many information processing tasks. An ESN is
determined by the following state-space transformation:
(
xt = σ (α + Axt−1 + γczt )) ,
>
yt = W xt .
(5.1)
(5.2)
In these equations z ∈ KM ⊂ (Rn ) is the input signal, c ∈ MN,n is called the input mask, α ∈ RN is
the input shift, and γ ∈ R is the input gain. The matrix A ∈ MN,N is referred to as the reservoir
matrix. The map σ in the state-space equation (5.1) is constructed by componentwise application of
a sigmoid function (like the hyperbolic tangent or the logistic function) and is called the activation
function. It can be shown (see [Jaeg 10]) that (5.1)-(5.2) induces a causal, time-invariant, and fading
memory filter whenever kAk2 = σmax (A) < 1. Moreover, the fading memory property does not hold
whenever the spectral radius ρ(A) of A exceeds one.
ESNs are implemented in practice by generating a large random reservoir matrix A and an input
mask c and by tuning the input shift, input gain, and the spectral radius of A so that the performance
in the execution of a specific task is maximized. Given specific input z ∈ KM and teaching signals
d ∈ RZ , the readout vector W is obtained by solving the ridge regularized regression problem:
!
t
1 X f>
2
2
(W xi − di ) + λkWk ,
W = arg min lim
t→∞ 2t
N
f
W∈R
i=−t
Z
where λ > 0 is the regularization constant.
The interesting universality properties associated to the non-homogeneous state-affine systems that
we proved in Section 3.2 suggest a generalization of the ESNs that we denominate echo state affine
systems (ESAS) and that are determined by the state-space equations:
(
xt = σ (p(zt )xt−1 + q(zt )) ,
>
yt = W x t .
(5.3)
(5.4)
An ESAS is hence a SAS in which an activation function has been added following the pattern of the
ESN case. ESAS are a strict generalization of ESNs: an ESAS for which the degree of the p polynomial
is zero and the degree of the q polynomial is one, is an ESN.
A forecasting exercise. A task in which ESNs have been particularly successful is the forecasting of
chaotic time series generated by the the Mackey-Glass system [Mack 77] characterized by the time-delay
differential equation:
dx
0.2 x(t − τ )
=
.
(5.5)
dt
1 + x(t − τ )10 − 0.1 x(t)
In our forecasting exercise we use for the delay τ the customary value in the forecasting literature of
τ = 17. The Lyapunov exponent of this system has been numerically estimated in [Jaeg 04], where
values around 6 · 10−3 were obtained, which shows that this system does indeed generate chaotic time
series.
Forecasting of those time series has been conducted using ESNs as in [Jaeg 04] by using a sparse (1%
connectivity) reservoir matrix A with dimension N = 1000 and with entries randomly drawn from a
Universal reservoir computers using non-homogeneous state-affine systems
17
uniform distribution in the interval [−1, 1]. The training is carried out by using as input z the sampling
with step size 1.0 of a solution of (5.5) with length 5000 and obtained from a differential equation solver
that produced an absolute accuracy of 1 · 10−16 . The teaching signal is set to d = U−1 z, that is, the
input signal shifted in time one step into the future. Forecasting is then carried out with the trained
system running autonomously by feeding the output into the input. The trained ESN is used in that
case as a proxy for the original system (5.5) that has been hence learned by (5.1)-(5.2). Using this
approach, the results reported in [Jaeg 04] improved by a factor of 700 the performances previously
documented in the literature.
We have checked if various ESAS architectures are able to improve the ESN performance in this
task. In order to do so and to make the two systems comparable, we have used ESAS of the same
dimension (N = 1000), the reservoir matrices in the p polyomial and the ESN reservoir matrix A are
sparse and have the same connectivity (1%) and, more importantly, the same hyperparameters have been
optimized. More specifically, in the ESN case we optimized the input shift and gain and we also tuned
the spectral radius of A. The same procedure was carried out in the ESAS case where, additionally, we
tuned the spectral radii of each of the reservoir matrices in the p polyomial.
The results of these forecasting exercises are depicted in Figure 1 where we plot the normalized
root mean square error (NRMSE) committed by the ESN and by several ESAS(r,s) arquitectures (r
and s stand for the degrees of the p and q polynomials, respectively) as a function of the forecasting
horizon. The NRMSE has been averaged over 2000 points. This figure shows that the chosen ESAS
architectures systematically outperform the ESN. A particularly significant improvement is exhibited
by the ESAS(1,2) that, at a forecasting horizon of 121, reduces the NRMSE of 39.18%.
6
Appendices
6.1
Proof of Lemma 2.1
Let w : N −→ (0, 1] be an arbitrary weighting sequence. Then for any z ∈ KM :
kzkw := sup kzt w−t k = sup kzt kw−t ≤ M · 1 = M < ∞.
t∈Z−
t∈Z−
Regarding the inequalities (2.4) and (2.5), notice that if wt = λt then:
∞
X
t=0
kz−t kwt =
∞
X
kz−t kλt =
t=0
∞
X
t=0
≤
∞
X
kz−t k(λ1−ρ λρ )t =
∞
X
kz−t kλ(1−ρ)t λρt
t=0
∞
n
o
n
oX
sup kz−i kλ(1−ρ)i λρt = sup kz−i kλ(1−ρ)i
λρt = kzkw1−ρ
t=0 i∈N
i∈N
t=0
1
,
1 − λρ
which proves (2.4). The proof of (2.5) is similar and follows from noticing that:
∞
X
t=0
6.2
kz−t kλ(1−ρ)t λρt ≤
∞
X
∞
X
sup kz−i kλρi λ(1−ρ)t) = sup kz−i kλρi
λ(1−ρ)t = kzkwρ
t=0 i∈N
i∈N
t=0
1
.
1 − λ1−ρ
Proof of Lemma 2.2
We recall first that by Lemma 2.1 we have that kzkw < ∞, for any z ∈ KM . Second, since (DZ− , k·kw )
is a normed space it is hence metrizable and therefore so is (KM , k·kw ) when endowed with the relative
topology (see, for instance, [Munk 14, Exercise 1, page 131]). We will then conclude the compactness
Universal reservoir computers using non-homogeneous state-affine systems
18
Figure 1: Results of the Mackey-Glass chaotic time series forecasting exercise. The different panels depict the normalized
root mean square error (NRMSE) committed by the ESN and by several ESAS(r,s) arquitectures (r and s stand
for the degrees of the p and q polynomials, respectively) as a function of the forecasting horizon. The NRMSE
has been averaged over 2000 points. The chosen ESAS architectures systematically outperform the ESN. A
particularly significant improvement is exhibited by the ESAS(1,2) that at a forecasting horizon of 121 reduces
the NRMSE of 39.18%.
Universal reservoir computers using non-homogeneous state-affine systems
19
of (KM , k·kw ) by showing that this space is sequentially compact (see, for example [Munk 14, Theorem
28.2]). We proceed by using the strategy in the proof of Lemma 1 in [Boyd 85].
n
For any n ∈ N, let KM
be the set obtained by projecting into D{−n,...,−1,0} the elements of KM ⊂
Z−
n
D . Given an element z ∈ KM , we will denote by z(n) := (z−n , . . . , z0 ) its projection into KM
.
n
n
Additionally, notice that KM = [−M, M ] is a compact (and hence sequentially compact) topological
space with the product topology.
Let {z(n)}n∈N ⊂ KM be a sequence of elements in KM . The argument that we just stated proves
k
that for any k ∈ N, there is a subset Nk ⊂ N and an element z(k) ∈ KM
such that
(k)
zt (n) − zt
max
t∈{−k,...,0}
−→ 0,
as n → ∞, n ∈ Nk .
Moreover, the sets Nk can be constructed so that N ⊃ N1 ⊃ N2 ⊃ · · · and so that z(k) extends z(l) when
k ≥ l. This implies the existence of an element z ∈ KM such that, for each k ∈ N,
max
t∈{−k,...,0}
kzt (n) − zt k −→ 0,
as n → ∞, n ∈ Nk ,
and hence there exists and increasing subsequence nk such that nk ∈ Nk and that for each k0 ,
max
t∈{−k0 ,...,0}
kzt (nk ) − zt k −→ 0,
as k −→ ∞.
(6.1)
We conclude by showing that the sequence {z(nk )}k∈N converges in (KM , k·kw ) to the element z ∈ KM .
First, given that wt −→ 0 as t −→ ∞, then for any ε > 0 there exists k0 such that wk < ε/2M , for any
k ≥ k0 . Additionally, since z(nk ), z ∈ KM for any k ∈ N, we have that
sup kzt (nk ) − zt k w−t ≤ 2M wk0 < ε.
(6.2)
t≤−k0
Now, by (6.1) there exists k1 such that for any k ≥ k1
kzt (nk ) − zt k w−t <
sup
t∈{−k0 ,...,0}
sup
kzt (nk ) − zt k < ε.
Consequently, (6.2) and (6.3) imply that for any k > max{k0 , k1 }, kz(nk ) − zkw < ε, as required.
6.3
(6.3)
t∈{−k0 ,...,0}
Proof of Lemma 2.6
Let δ w () be the epsilon-delta relation for the FMP associated to the weighting sequence w. We now
0
show that HU has the FMP with respect to w0 via the epsilon-delta relation given by δ w () := δ w ()/λ.
0
Indeed, for any > 0 and any z, s ∈ K such that kz − skw0 < δ w (), we have that
kz−skw = sup kzt −st kw−t = sup kzt −st k
t∈Z−
t∈Z−
0
w−t 0
0
w−t < λ sup kzt −st kw−t
< λkz−skw0 < λδ w () = δ w (),
0
w−t
t∈Z−
and consequently, since HU has the FMP with respect to the weighting sequence w, we can conclude
that |HU (z) − HU (s)| < . This shows that the implication
0
kz − skw0 < δ w () =⇒ |HU (z) − HU (s)| <
holds, as required.
Universal reservoir computers using non-homogeneous state-affine systems
6.4
20
Proof of Theorem 3.1
Since the elements in R have the FMP with respect to a given weighted norm k · kw , then so do those
in A(R) since polynomial combinations of continuous elements of the form HhFii : (KM , k · kw ) −→ R
are also continuous. Therefore, under that hypothesis, A(R) is a polynomial subalgebra of the algebra
(C 0 (KM ), k · kw ) of real-valued continuous functions on (KM , k · kw ). Since by hypothesis A(R) contains
the constant functionals and separates the points in KM and, by Lemma 2.2, the set (KM , k · kw ) is
compact, the Stone-Weierstrass theorem (Theorem 7.3.1 in [Dieu 69]) implies that A(R) is dense in
(C 0 (KM ), k · kw ), which concludes the proof.
6.5
Proof of Corollary 3.3
In order to show that the reservoir systems in L induce reservoir filters, we will show that they have
the echo state property by using the following lemma, whose proof can be found in [Grig 18].
Lemma 6.1 Let DN ⊂ RN and Dn ⊂ Rn and let F : DN × Dn −→ DN be a continuous reservoir map.
Suppose that F is a contraction map with contraction constant 0 < r < 1, that is:
kF (x, z) − F (y, z)k ≤ r kx − yk ,
for all x, y ∈ RN and all z ∈ Rn ,
then the corresponding reservoir system has the echo state property.
We start now by noting that the condition σmax (A) < 1 − < 1 implies that the reservoir map
F (x, z) := Ax+cz associated to (3.6) is a contracting map with constant σmax (A) which, by hypothesis,
is smaller than one. Indeed,
kF (x, z) − F (y, z)k = kA(x − y)k ≤ σmax (A) kx − yk
for all x, y ∈ RN and all z ∈ Rn .
By Lemma 6.1 we can conclude that this reservoir system has a reservoir filter associated that we now
show is explicitly given by (3.8). We start by proving that the conditions σmax (A) < 1 − < 1 and that
the elements in KM are uniformly bounded
Pn by a constant M imply that the infinite sum in (3.8) is
convergent. Let n, m ∈ N and let Sn := i=0 Ai cz−i . Now:
kSn − Sm k =
m
X
j=n+1
Ai cz−i ≤
m
X
i
kAk2 kck2 kz−i k ≤ M kck2
m
X
σmax (A)i ≤
j=n+1
j=n+1
M kck2
∞
X
σmax (A)i = M kck2
j=n+1
n+1
σmax (A)n+1
.
1 − σmax (A)
n+1
(c)σmax (A)
max (A)
The condition σmax (A) < 1− < 1 implies that M kck2 σ1−σ
= M σmax1−σ
→ 0 as n → ∞
max (A)
max (A)
N
and hence {Sn }n∈N is a Cauchy sequence in R that consequently converges.
The fact that the filter determined by the expression (3.8) is a solution of the recursions (3.6)(3.7) is a straightforward verification. In order to carry it out, it suffices to use that the filter UhA,c (z)
associated to the functional HhA,c (z) is given by
!
∞
X
A,c
i
Uh (z) = h
A czt−i ,
t
and that the time series xet defined by xet :=
i=0
P∞
i=0
Ai czt−i satisfies the recursion relation (3.6).
Universal reservoir computers using non-homogeneous state-affine systems
21
We now verify by hand that the filters UhA,c (z) are time-invariant. Let z ∈ KM and t, τ ∈ N arbitrary
and let Uτ be the corresponding time delay operator, then:
!
!
∞
∞
X
X
UhA,c ◦ Uτ (z) = UhA,c (Uτ (z)) = h
Ai cUτ (z)t−i = h
Ai czt−i−τ
(6.4)
t
t
i=0
i=0
At the same time,
Uτ ◦
UhA,c
(z) = Uτ UhA,c (z)
= UhA,c (z)
t
t
t−τ
= KM h
∞
X
!
i
A czt−τ −i
,
i=0
which coincides with (6.4) and proves the time-invariance of UhA,c .
The next step consists in showing that the elements in L are λρ -exponential fading memory filters,
with λρ := (1 − )ρ , for any ρ ∈ (0, 1), that is, L ⊂ Rwρ , with wρ : N → (0, 1] the sequence given by
wtρ := (1 − )ρt . Let k · kwρ be the associated weighted norm in KM and let z ∈ KM be an arbitrary
element. We start by noting that the continuity of the readout map h : RN → R implies that for any
ε > 0 there exists an element δ(ε) > 0 such that for any v ∈ RN that satisfies
!
∞
∞
X
X
i
i
(6.5)
A czt−i < ε.
A czt−i < δ(ε), then h(v) − h
v−
i=0
i=0
We now show that for any s ∈ KM such that
ks − zkwρ
δ(ε) 1 − (1 − )1−ρ
<
,
σmax (c)
then
HhA,c (s) − HhA,c (z) < ε.
(6.6)
Indeed,
∞
X
Ai cst−i −
i=0
∞
X
i=0
∞
X
i=0
Ai czt−i =
∞
X
Ai c(st−i − zt−i ) ≤
i=0
σmax (Ai ) kc(st−i − zt−i )k ≤
∞
X
Ai c(st−i − zt−i ) ≤
i=0
∞
X
i=0
σmax (A)i kc(st−i − zt−i )k ≤
∞
X
(1 − )i kc(st−i − zt−i )k .
i=0
If we now use (2.5) in Lemma 2.1 and the hypothesis in (6.6), we can conclude that
∞
X
i=0
(1 − )i kc(st−i − zt−i )k ≤ σmax (c)
∞
X
σmax (c)ks − zkwρ
(1 − )i k(st−i − zt−i )k ≤
< δ(ε),
1 − (1 − )1−ρ
i=0
which proves the continuity of the map HhA,c : (KM , k · kwρ ) −→ R and hence shows that HhA,c is a
λρ -exponential fading memory filter.
In order to establish the universality statement in the corollary we will proceed, as in the proof of
Theorem 3.1, by showing that L is a polynomial algebra that contains the constant functionals and
separates the points in KM and then by invoking the Stone-Weierstrass theorem using the compactness
of (C 0 (KM ), k · kwρ ).
In order to show that (L , k · kwρ ) is a polynomial algebra, notice first that if A1 , A2 ∈ MN are such
that σmax (A1 ), σmax (A1 ) < 1 − , then
σmax (A1 ⊕ A2 ) = max (σmax (A1 ), σmax (A2 )) < 1 − .
(6.7)
Universal reservoir computers using non-homogeneous state-affine systems
22
If we now take ci ∈ MNi ,ni , i ∈ {1, 2} and h1 , h2 two real-valued polynomials in N1 and N2 variables,
respectively, we have by the first part of the corollary that we just proved that the filter functionals
HhA11 ,c1 and HhA22 ,c2 are well defined. Additionally, by (3.2)-(3.3) so are the combinations HhA11 ,c1 · HhA22 ,c2
and HhA11 ,c1 + λHhA22 ,c2 that satisfy:
2 ,c1 ⊕c2
HhA11 ,c1 · HhA22 ,c2 = HhA11·h⊕A
,
2
⊕A2 ,c1 ⊕c2
HhA11 ,c1 + λHhA22 ,c2 = HhA11⊕λh
,
2
λ ∈ R.
(6.8)
Using the relations (6.8) and (6.7), we can conclude that both HhA11 ,c1 · HhA22 ,c2 and HhA11 ,c1 + λHhA22 ,c2
belong to L ⊂ Rwρ . This implies that (L , k · kwρ ) is a polynomial subalgebra of (Rwρ , k · kwρ )
Since L contains the constant functionals (just take constant readout maps h), in order to conclude
the proof, it is enough to show that the elements in L separate points in KM . Take z1 , z2 ∈ KM ⊂
Z
Z
(Rn ) − such that z1 6= z2 and let A ∈ M(n, n), with σmax (A) < 1−, and c := In . Let U A,c : (Rn ) − −→
n Z−
(R ) be the linear filter associated to A and c via the recursion (3.6). Using the preceding arguments
we have that
∞
X
Aj zt−j .
U A,c (z) t =
j=0
Z
Z
At the same time, it is easy to verify that the filter U In −A,c : (Rn ) − −→ (Rn ) − given by U In −A,c (z) t =
zt − Azt−1 , is the two-sided inverse of U A,c , that is, U A,c ◦ U In −A,c = U In −A,c ◦ U A,c = I(Rn )Z− ,
which guarantees the bijectivity of U A,c and implies in passing that U A,c (z1 ) 6= U A,c (z2 ). Since
polynomials in Rn separate points, there exists a polynomial readout map h : Rn −→ R such that
h U A,c (z1 ) 6= h U A,c (z2 ) which shows that the associated functional HhA,c is such that
HhA,c (z1 ) 6= HhA,c (z2 ) ,
6.6
as required.
Proof of Proposition 3.5
We start by noting, as we did in the proof of Corollary 3.3, that the condition (3.11) implies that the
reservoir map associated to (3.9) is a contraction and hence, by Lemma 6.1, it satisfies the echo state
property and has a well-defined associated filter.
We now prove that the condition (3.11) implies the convergence of the series in the expression (3.12).
Let K1 := maxz∈I kp(z)k2 = maxz∈I σmax (p(z)) < 1 and K2 := maxz∈I kq(z)k2 = maxz∈I σmax (q(z));
notice that K
1 and K2 are
well-defined due to the compactness of I. Let now n, m ∈ N and let
Pn Qj−1
Sn := j=0
p(z
)
q(zt−j ) ∈ RN . Then,
t−k
k=0
kSn − Sm k
=
≤
m
X
j−1
Y
j=n+1
k=0
j−1
m
X
Y
!
p(zt−k ) q(zt−j ) ≤
m
X
j−1
Y
j=n+1
k=0
kp(zt−k )k2 kq(zt−j )k ≤ K2
j=n+1 k=0
m
X
j=n+1
K K n+1
kq(zt−j )k
p(zt−k )
K1j ≤ K2
2
∞
X
j=n+1
K1j =
K2 K1n+1
.
1 − K1
2 1
The condition K1 < 1 implies that 1−K
→ 0 as n → ∞ and hence {Sn }n∈N is a Cauchy sequence
1
N
in R that consequently converges. This proves the convergence of the infinite series in (3.12) and
the causal character of the filter that it defines. The time-invariance can also be easily established by
mimicking the verification that we carried out in the proof of Corollary 3.3. We conclude by proving
Universal reservoir computers using non-homogeneous state-affine systems
23
that (3.12) is indeed a solution of (3.9):
∞
X
p(zt )xt−1 + q(zt ) = p(zt )
j−1
Y
j=0
k=0
!
p(zt−1−k ) q(zt−1−j ) + q(zt ) = q(zt ) + p(zt )q(zt−1 )
+ p(zt )p(zt−1 )q(zt−2 ) + p(zt )p(zt−1 )p(zt−2 )q(zt−3 ) + · · · =
6.7
∞
X
j−1
Y
j=0
k=0
!
p(zt−k ) q(zt−j ) = xt .
Proof of Lemma 3.6
(i) =⇒ (ii): kA0 k2 + kA1 k2 + · · · + kAn1 k2 < λ + λ + · · · + λ = λ(n1 + 1) < 1.
(ii) =⇒ (iii): kp(z)k2 = kA0 + zA1 + z 2 A2 + · · · + z n1 An1 k2 ≤ kA0 k2 + |z|kA1 k2 + |z 2 |kA2 k2 + · · · +
|z n1 |kAn1 k2 < kA0 k2 + kA1 k2 + · · · + kAn1 k2 < 1.
6.8
Proof of Proposition 3.7
We start by formulating and proving an elementary result that will be needed later on.
Lemma 6.2 Let f : U ⊂ Rn −→ Mm be a differentiable function defined on the convex set U . For any
z ∈ U denote by ∂i f (z) ∈ Mm the matrix containing the partial derivatives of the components of f with
respect to their ith-entry. Then, for any x, y ∈ U we have:
√
kf (y) − f (x)k2 ≤ nm max
sup k∂i f (z)k2 kx − yk .
(6.9)
i∈{1,...,n}
z∈U
Pn Pm
Proof. Given A = (Ai,j ) ∈ Mn,m , let kAkF := tr A> A = i=1 j=1 A2i,j be its Frobenius norm.
Recall (see Theorem 5.6.34 and Exercise 5.6.P24 in [Horn 13]) that
√
(6.10)
kAk2 ≤ kAkF ≤ r kAk2 ,
where r is the rank of A. Consider now x, y ∈ U arbitrary and let Df (z) : Rn −→ Mm be the differential
of f evaluated at z ∈ U . The convexity of U implies that the Mean Value Inequality holds (see Theorem
2.4.8 in [Abra 88]) and hence:
kf (y) − f (x)kF ≤ sup kDf ((1 − t)x + ty)k2 kx − yk .
(6.11)
t∈[0,1]
The first inequality in (6.10) and (6.11) imply that
kf (y) − f (x)k2 ≤ sup kDf (z)k2 kx − yk .
(6.12)
z∈U
At the same time, notice that by (6.10)
2
2
kDf (z)k2
≤
2
kDf (z)kF
=
n X
m
X
i=1 j=1
∂i fj2 (z) =
n
X
i=1
2
k∂i f (z)kF ≤ m
n
X
i=1
2
k∂i f (z)k2 ≤ mn
max
i∈{1,...,n}
2
k∂i f (z)k2 .
This inequality together with (6.12) imply the statement (6.9) since the maximum and the supremum
can be trivially exchanged. H
We now carry out the proof of the proposition under the hypothesis (iii) in Lemma 3.6 which
is implied by the other two. The modifications necessary to establish the result under the other two
Universal reservoir computers using non-homogeneous state-affine systems
24
hypotheses are straightforward. Consider two arbitrary elements z, s ∈ I Z− . Then, by the CauchySchwarz and Minkowski inequalities:
∞
X
p,q
p,q
|HW
(z) − HW
(s)| = W>
j−1
Y
j=0
k=0
≤ kWk
∞
X
!
p(z−k ) q(z−j ) −
j−1
Y
!
!
p(s−k ) q(s−j )
k=0
aj (z−j+1 )q(z−j ) − aj (s−j+1 )q(s−j ) ,
where
aj (z−j+1 ) :=
j=0
j−1
Y
p(z−k ). (6.13)
k=0
We now bound the right hand side of (6.13) as follows:
∞
X
aj (z−j+1 )q(z−j ) − aj (s−j+1 )q(s−j )
j=0
=
∞
X
aj (z−j+1 )q(z−j ) + aj (z−j+1 )q(s−j ) − aj (z−j+1 )q(s−j ) − aj (s−j+1 )q(s−j )
j=0
≤
∞
X
aj (z−j+1 )
j=0
kq(z−j ) − q(s−j )k + aj (z−j+1 ) − aj (s−j+1 )
2
2
kq(s−j )k
(6.14)
If Lq is a Lipschitz constant of q : I −→ RN then
aj (z−j+1 )
2
kq(z−j ) − q(s−j )k ≤ Mpj Lq |z−j − s−j | ,
(6.15)
which inserted in (6.14) and in (6.13) implies that
∞
∞
X
X
p,q
p,q
|HW
(z) − HW
(s)| ≤ kWk Lq
Mpj |z−j − s−j | +
aj (z−j+1 ) − aj (s−j+1 )
j=0
j=0
2
(6.16)
We now bound above the second summand in (6.16) using the inequality (6.9) in the statement of
Lemma 6.2 as well as the following straightforward identity:
aj (z−j+1 ) − aj (s−j+1 ) =
j−1
X
(p(s0 ) · · · p(s−(l−1) ) · p(z−l ) · p(z−(l+1) ) · · · p(z−(j−1) )
l=0
− p(s0 ) · · · p(s−(l−1) ) · p(s−l ) · p(z−(l+1) ) · · · p(z−(j−1) )).
Using this relation we can write:
aj (z−j+1 ) − aj (s−j+1 )
≤
j−1
X
≤
2
j−1
X
p(s0 ) · · · p(s−(l−1) ) · (p(z−l ) − p(s−l )) · p(z−(l+1) ) · · · p(z−(j−1) )
2
l=0
kp(s0 )k2 · · · p(s−(l−1) )
2
· kp(z−l ) − p(s−l )k2 · p(z−(l+1) )
2
· · · p(z−(j−1) )
2
l=0
j
X
√
|z−j+l − s−j+l | ,
≤ Mpj−1 N sup (kp0 (z)k2 )
z∈I
l=1
Universal reservoir computers using non-homogeneous state-affine systems
where the last inequality is a consequence of (6.9). Let Mp0 :=
aj (z−j+1 ) − aj (s−j+1 )
≤
2
√
25
N supz∈I (kp0 (z)k2 ), then
j
j
M p0 j X
Mp0 X l j−l
z−(j−l) − s−(j−l)
Mp
|z−j+l − s−j+l | =
M p Mp
Mp
Mp
l=1
l=1
Since the last term in this inequality is the Cauchy product of the series with general terms Mpn and
Mpn |z−n − s−n |, we can conclude using Merten’s Theorem [Apos 74, Theorem 8.46] that
∞
X
aj (z−j+1 ) − aj (s−j+1 )
j=0
≤
2
∞ j
Mp0 X X l j−l
z−(j−l) − s−(j−l)
M p Mp
Mp j=0
l=1
=
∞
X
1
M p0
M j |z−j − s−j | .
Mp 1 − Mp j=0 p
If we now substitute this relation in (6.16) and we use Lemma 2.1 with weighting sequences wtρ := Mpρt ,
for any ρ ∈ (0, 1), we obtain that:
p,q
p,q
|HW
(z) − HW
(s)|
≤
≤
X
∞
M p0
1
kWk Lq 1 +
M j |z−j − s−j |
Mp 1 − Mp j=0 p
1
1
M p0
kz − skwρ ,
kWk Lq 1 +
Mp 1 − M p
1 − Mp1−ρ
p,q
which proves the continuity of the map HW
: (I Z− , k · kwρ ) −→ R, as required.
6.9
Proof of Proposition 3.8
p2 ,q2
p1 ,q1
are well-defined then, by
, HW
We first recall that since by hypothesis the reservoir functionals HW
2
1
p1 ,q1
p2 ,q2
p1 ,q1
p2 ,q2
the comments that follow (3.2)-(3.3), so are HW1 + λHW2 and HW1 · HW
.
2
The proof of (i) is a straightforward verification. As to (ii), denote first by yt1 , yt2 and x1t , x2t the
outputs and the state variables, respectively, of the SAS corresponding to the two functionals that we
are considering. We note first that by (3.10):
>
yt1 · yt2 = W1> x1t · W2> x2t = (W1 ⊗ W2 ) (x1t ⊗ x2t ).
Using (3.9) it can be readily verified that the time evolution of the tensor product x1t ⊗ x2t is given by
x1t ⊗ x2t = (p1 (zt ) ⊗ p2 (zt ))(x1t−1 ⊗ x2t−1 ) + p1 (zt )x1t−1 ⊗ q2 (zt ) + q1 (zt ) ⊗ p2 (zt )x2t−1 + q1 (zt ) ⊗ q2 (zt ),
which proves (3.20) and hence (3.19).
In order to show that the reservoir functionals on the right hand side of (3.18) and (3.19) are
well-defined we prove the following lemma.
Lemma 6.3 Let p1 (z) ∈ MN1 ,M1 [z] and p2 (z) ∈ MN2 ,M2 [z] be two polynomials with matrix coefficients
and assume that they satisfy that kp1 (z)k2 < 1 − and kp2 (z)k2 < 1 − for all z ∈ I := [−1, 1] and a
given 1 > > 0. Then:
(i) k(p1 ⊕ p2 )(z)k2 < 1 − ,
(ii) k(p1 ⊗ p2 )(z)k2 < 1 − ,
Universal reservoir computers using non-homogeneous state-affine systems
26
for all z ∈ I := [−1, 1].
Proof of the lemma. Suppose that the polynomials p1 and p2 are defined on the vector spaces V1
and V2 , respectively. Let x = x1 ⊕ x2 ∈ V1 ⊕ V2 , with x1 ∈ V1 and x2 ∈ V2 . Then, in order to prove
part (i) note that
k(p1 ⊕ p2 )(z) · xk2 = k(p1 (z) · x1 , p2 (z) · x2 )k2 = kp1 (z) · x1 k2 + kp2 (z) · x2 k2
≤ kp1 (z)k22 kx1 k2 + kp2 (z)k22 kx2 k2 ≤ (1 − )2 kx1 k2 + kx2 k2 = (1 − )2 kxk2 .
This inequality implies that
(1 − )kxk
k(p1 ⊕ p2 )(z) · xk
≤ sup
= 1 − ,
kxk
kxk
x6=0
x6=0
k(p1 ⊕ p2 )(z)k2 = sup
as required.
As to the statement in part (ii):
k(p1 ⊗p2 )(z)k2 = σmax ((p1 ⊗p2 )(z)) = σmax (p1 (z))σmax (p2 (z)) = kp1 (z)k2 kp2 (z)k2 < (1−)2 < (1−). H
p1 ⊕p2 ,q1 ⊕q2
Now, the first part of this lemma and Proposition 3.5 guarantee that HW
is well-defined.
1 ⊕λW2
p,q ⊕q ⊕(q ⊗q )
1
2
1
2
The same conclusion holds for H0⊕0⊕(W
because due to the block diagonal character of (3.20)
1 ⊗λW2 )
then σmax (p(z)) = σmax ((p1 (z) ⊕ p2 (z) ⊕ (p1 ⊗ p2 ) (z)) = kp1 (z) ⊕ p2 (z) ⊕ (p1 ⊗ p2 ) (z)k2 . By parts (i)
and (ii) in Lemma 6.3 we can conclude that kp(z)k2 < 1 − for all z ∈ [−1, 1] and, again by Proposition
p,q1 ⊕q2 ⊕(q1 ⊗q2 )
3.5, the reservoir functional H0⊕0⊕(W
is well-defined.
1 ⊗λW2 )
6.10
Proof of Theorem 3.9
Note first that the hypothesis Mp < 1 − < 1 on the polynomials p associated to the elements in S
implies, by Propositions 3.5 and 3.7, that this family is made of time-invariant reservoir filters that have
the FMP with respect to weighting sequences of the form wtp := Mpρt , ρ ∈ (0, 1). Additionally, using
Lemma 2.6 and the hypothesis Mp < 1 − , for a fixed given ∈ (0, 1), we can conclude that all the
reservoir filters in S have the FMP with the common weighting sequence wtρ := (1 − )ρt , ρ ∈ (0, 1).
We now show that the elements in S form a polynomial algebra as a consequence mainly of
p2 ,q2
p1 ,q1
: I Z− −→ R are two elements in S then
, HW
Lemma 6.3 and Proposition 3.8. Indeed, if HW
2
p2 ,q2
p11 ,q1
by (3.18) and part (i) of Lemma 6.3 so is HW1 + λHW2 , for any λ ∈ R. The same conclup1 ,q1
p2 ,q2
sion holds for the product HW
· HW
because due to the block diagonal character of (3.20) then
1
2
σmax (p(z)) = σmax ((p1 (z) ⊕ p2 (z) ⊕ (p1 ⊗ p2 ) (z)) = kp1 (z) ⊕ p2 (z) ⊕ (p1 ⊗ p2 ) (z)k2 . By parts (i) and
(ii) in Lemma 6.3 we can conclude that kp(z)k2 < 1 − for all z ∈ [−1, 1]. The same fact can be stated
p1 ,q1
p2 ,q2
about q1 ⊕ q2 ⊕ (q1 ⊗ q2 ) and hence by (3.19) we obtain that the product HW
· HW
belongs to S .
1
2
Finally, we observe that the family S has the point separation property and contains all the constant
functionals. Indeed, since S includes the linear family L and we proved that given z1 , z2 ∈ KM ⊂
Z
(Rn ) − such that z1 6= z2 there exists A ∈ M(n, n), with σmax (A) < 1 − and c := In such that
U A,c (z1 ) 6= U A,c (z2 ), this implies the existence of a t ∈ Z such that (U A,c (z1 ))t 6= (U A,c (z2 ))t . The
point separation property follows from choosing any vector W ∈ RN such that W> (U A,c (z1 ))t 6=
A,c
A,c
A,c
A,c
W> (U A,c (z2 ))t , which implies that (UW
(z1 ))t 6= (UW
(z2 ))t . Therefore, UW
(z1 ) 6= UW
(z2 ) and
HU A,c (z1 ) 6= HU A,c (z2 ), as required.
W
W
All the constant functionals can be obtained by taking for p the zero polynomial and for q the
constant polynomials (q has degree zero). In that case, the state variables are a constant sequence
p,q
xt = q and the associated functional is the constant map HW
(z) = W> q, for all z ∈ KM .
The universality result follows hence from the Stone-Weierstrass Theorem and the compactness of
(I Z− , k · kwρ ) established in Lemma 2.2.
Universal reservoir computers using non-homogeneous state-affine systems
6.11
27
Proof of Lemma 4.1
(i) Let A := {ρ ∈ R+ | kXk < ρ almost surely}. It suffices to show that kXkL∞ := inf A ∈ A, which
implies that kXk ≤ kXkL∞ almost surely. Indeed, consider the sequence aj := kXkL∞ + 1/j, j ∈ N. By
the approximation property of the infimum, there exists a decreasing sequence of numbers {ρj }j∈N ⊂ A
in A satisfying kXkL∞ ≤ ρj < kXkL∞ + 1/j for all j ∈ N. Define F := {ω ∈ Ω | kX(ω)k > kXkL∞ }
and Fj := {ω ∈ Ω | kX(ω)k > ρj }. It is easy to see that Fj ⊂ Fj+1 , j ∈ N and that limj→∞ Fj = F
and, consequently, (see [Grim 01, Lemma 5, page 7]) limj→∞ P(Fj ) = P(F ). Since by construction
P(Fj ) = 0 for all j ∈ N then P(F ) = 0 necessarily, which shows that kXkL∞ ∈ A, as required.
(ii) If kXkL∞ ≤ C then by part (i), kXk ≤ kXkL∞ ≤ C almost surely. Conversely, if kXk ≤ C almost
surely, then C ∈ A = {ρ ∈ R+ | kXk < ρ almost surely}. Consequently, kXkL∞ = inf A ≤ C ∈ A, as
required.
(iii) Suppose first that kXk ≤ C almost surely and define F := {ω ∈ Ω | kX(ω)k > C}. By hypothesis,
we have that P(F ) = 0 and P(Ω \ F ) = 1. Then,
Z
Z
Z
h
i
k
k
k
k
E kXk
=
kXk dP =
kXk dP +
kXk dP
Ω
Ω\F
F
Z
Z
k
k
=
kXk dP ≤
C dP = C k P(Ω \ F ) = C k ,
Ω\F
Ω\F
h
i
k
as required. Conversely, assume that E kXk ≤ C k , for any k ∈ N, and define
Fn :=
1
ω ∈ Ω | kX(ω)k > C +
n
,
for all n ≥ 1. It is easy to see that Fn ⊂ Fn+1 and that limn→∞ Fn = F and, consequently, (see
[Grim 01, Lemma 5, page 7]) limn→∞ P(Fn ) = P(F ). Now,
Z
Z
h
i Z
k
k
k
k
k
C
≥ E kXk =
kXk dP =
kXk dP +
kXk dP
Ω
Z
k
Ω\Fn
Z
kXk dP ≥
≥
Fn
C+
Fn
1
n
Fn
k
dP =
C+
1
n
k
P(Fn ),
k
which implies that P(Fn ) ≤ C k / C + n1 for any k ∈ N and hence, by taking the limit k → ∞, we can
conclude that P(Fn ) = 0. Consequently, P(F ) = limn→∞ P(Fn ) = 0, which shows that kXk ≤ C a.s.
(iv) Since |Xi | ≤ kXk always and by part (i) kXk ≤ kXkL∞ almost surely, we can conclude that
|Xi | ≤ kXkL∞ almost surely. This implies that Xi ∈ L∞ (Ω, R) and hence the statement follows from
part (iii).
6.12
Proof of Lemma 4.2
Suppose first that
ess sup sup {kzt (ω)k} < sup ess sup {kzt (ω)k} .
ω∈Ω
t∈Z
t∈Z
(6.17)
ω∈Ω
By the approximation property of the supremum [Apos 74, Theorem 1.14], there exists t0 ∈ Z such that
ess sup sup {kzt (ω)k} < ess sup {kzt0 (ω)k} ≤ sup ess sup {kzt (ω)k} .
(6.18)
ω∈Ω
t∈Z
ω∈Ω
t∈Z
ω∈Ω
Universal reservoir computers using non-homogeneous state-affine systems
28
However, kzt0 (ω)k ≤ supt∈Z kzt (ω)k for all ω ∈ Ω and hence by part (i) in Lemma 4.1
kzt0 (ω)k ≤ sup kzt (ω)k ≤ ess sup sup {kzt (ω)k} , almost surely.
ω∈Ω
t∈Z
t∈Z
Now, by part (ii) in Lemma 4.1, this implies that
ess sup {kzt0 (ω)k} ≤ ess sup sup {kzt (ω)k} .
ω∈Ω
ω∈Ω
t∈Z
However, this expression is in contradiction with the first inequality in (6.18) and hence the assumption
(6.17) cannot be correct. This argument implies that
ess sup sup {kzt (ω)k} ≥ sup ess sup {kzt (ω)k} .
ω∈Ω
t∈Z
t∈Z
ω∈Ω
We now prove the reverse inequality, that is,
ess sup sup {kzt (ω)k} ≤ sup ess sup {kzt (ω)k} .
ω∈Ω
t∈Z
t∈Z
By part (ii) of Lemma 4.1, this inequality holds if and only if
sup {kzt (ω)k} ≤ sup ess sup {kzt (ω)k} ,
t∈Z
(6.19)
ω∈Ω
almost surely.
(6.20)
ω∈Ω
t∈Z
Now, by part (i) in Lemma 4.1, we have that kzt (ω)k ≤ ess supω∈Ω {kzt (ω)k}, almost surely and for
each fixed t ∈ Z. Let
S At ⊂ Ω be the zero-measure set
S suchthatPkzt (ω)k > ess supω∈Ω {kzt (ω)k} for all
ω ∈ At . Let A := t∈Z At . Notice that P(A) = P t∈Z At ≤ t∈Z P(At ) = 0 and hence B := Ac has
measure one and
kzt (ω)k ≤ ess sup {kzt (ω)k} , for all ω ∈ B and all t ∈ Z.
ω∈Ω
Since B has measure one, this implies that
sup {kzt (ω)} k ≤ sup ess sup {kzt (ω)k} ,
t∈Z
t∈Z
almost surely,
which is exactly what is required in (6.20) in order for (6.19) to hold.
6.13
(6.21)
ω∈Ω
Proof of Lemma 4.3
It is obvious that L∞ (Ω, `∞ (Rn )) ⊂ L∞ (Ω, `∞ (Rn )). Conversely, if z ∈ L∞ (Ω, `∞ (Rn )) then, by
definition, kzkL∞ < ∞ or, equivalently, ess supω∈Ω {supt∈Z {kzt (ω)k}} < ∞. By part (i) in Lemma 4.1,
this implies that
sup kzt (ω)k < ∞, almost surely.
(6.22)
t∈Z
Since the elements in the spaces in (4.7) are equivalence classes containing almost surely equal random
variables, we can take another representative z∗ : Ω −→ (Rn )Z for the class containing z defined as
z(ω),
when supt∈Z kzt (ω)k < ∞,
∗
z (ω) :=
0,
otherwise.
Since the processes z and z∗ differ by (6.22) only in a set of zero measure, they are equal in L∞ Ω, (Rn )Z
but, this time, z∗ ∈ L∞ (Ω, `∞ (Rn )), as required. Since (`∞ (Rn ), k · k∞ ) is a Banach space, so is
(L∞ (Ω, `∞ (Rn )) , k · kL∞ ) by the references quoted after (4.2) and the statement follows.
Universal reservoir computers using non-homogeneous state-affine systems
6.14
29
Proof of Theorem 4.4
Proof of part (i). All along this proof we will denote the elements in KM with a lower bold case
L∞
L∞
(z ∈ KM ) and those in KM
with an upper bold case (Z ∈ KM
).
We first assume that the functional H : (KM , k·kw ) −→ R has the fading memory property. This
means that H is a continuous map and since by Lemma 2.2 the space (KM , k·kw ) is compact, then so is
the image H(KM ) as a subset of the real line. This implies that there exists a finite real number L > 0
L∞
such that H(KM ) ⊂ [−L, L]. Let now Z ∈ KM
; the condition kZkL∞ ≤ M is equivalent to kZt k ≤ M ,
for all t ∈ Z− , almost surely, and hence implies that H (Z) ∈ [−L, L], almost surely or, equivalently,
L∞
that kH (Z)kL∞ ≤ L. This, in turn, implies that H(Z) ∈ L∞ (Ω, R) for any Z ∈ KM
, as required.
∞
L
∞
We now show that H : (KM
, k · k L∞
)
−→
L
(Ω,
R)
has
the
FMP.
The
FMP
hypothesis on
w
H : (KM , k·kw ) −→ R implies that for any z ∈ KM and any > 0 there exists a δ() > 0 such that for
any s ∈ KM that satisfies that
kz − skw = sup k(zt − st )w−t k < δ(),
then
|H(z) − H(s)| < .
(6.23)
t∈Z−
Moreover, since by Lemma 2.2 the space (KM , k·kw ) is compact, the Uniform Continuity Theorem
[Munk 14, Theorem 7.3] guarantees that the relation δ() does not depend on the point z ∈ KM .
L∞
We now prove the statement by showing that for any > 0 and Z ∈ KM
then kH(Z)−H(S)kL∞ < ,
∞
L
for all S ∈ KM
such that kZ − SkL∞
<
δ().
Indeed,
kZ
−
Sk
<
δ()
if
and
only if supt∈Z− kZt − St kL∞ w−t <
L∞
w
w
δ(). Given that for any l ∈ Z− we have that kZl − Sl kL∞ w−l ≤ supt∈Z− kZt − St kL∞ w−t < δ(),
part (ii) in Lemma 4.1 implies that kZl − Sl k w−l < δ() almost surely for any l ∈ Z− and hence
supt∈Z− kZt − St k w−t = kZ − Skw < δ(), almost surely. This implies, using (6.23), that |H(Z) −
H(S)| < , almost surely, which by part (ii) in Lemma 4.1 implies that kH(Z) − H(S)kL∞ < , as
required.
L∞
, k · kL∞
) −→ L∞ (Ω, R) has the fading memory property then so does
Conversely, if H : (KM
w
L∞
and kzk = kzkL∞ for the elements z ∈ KM .
H : (KM , k·kw ) −→ R because KM ⊂ KM
Proof of part (ii). We suppose first that T is dense in the set (C 0 (KM ), k · kw ) and show that the
L∞
L∞
corresponding family with intputs in KM
is universal. Let H : (KM
, k · k L∞
) −→ L∞ (Ω, R) be an
w
arbitrary causal and time-invariant FMP filter and let HS ∈ T such that supz∈KM kH(z) − HS (z)kL∞ <
. The existence of HS is ensured by the density hypothesis on T . We show that this ensures that
supZ∈KM
L∞ kH(Z) − HS (Z)kL∞ < . Indeed, this conclusion is true if kH(Z) − HS (Z)kL∞ < for any
L∞
Z ∈ KM which, by part (ii) in Lemma 4.1 is equivalent to |H(Z) − HS (Z)| < almost surely, for any
L∞
L∞
Z ∈ KM
. This condition is in turn true because as Z ∈ KM
, then kZt k ≤ M almost surely for all
t ∈ Z− and hence Z ∈ KM almost surely. Since HS approximates H for deterministic inputs, we have
that |H(Z) − HS (Z)| < almost surely, as required.
L∞
Conversely, if the family T with intputs in KM
is universal in the set of continuous maps of the
∞
L
∞
type H : (KM
, k · kL∞
)
−→
L
(Ω,
R)
we
can
easily
show that T is dense in (C 0 (KM ), k · kw ). Let
w
0
L∞
H ∈ (C (KM ), k · kw ) and let HS : (KM , k · kL∞
) −→ L∞ (Ω, R) be the element that, for a given
w
> 0, satisfies that kH − HS kL∞ = supZ∈KM
L∞ kH(Z) − HS (Z)kL∞ < . Given that, as we pointed out,
L∞
KM ⊂ KM and kzk = kzkL∞ , for the elements z ∈ KM , we have
kH − HS k = sup kH(z) − HS (z)k = sup kH(z) − HS (z)kL∞ ≤
z∈KM
6.15
z∈KM
sup kH(Z) − HS (Z)kL∞ < .
∞
L
Z∈KM
Proof of Theorem 4.5
We first notice that the polynomial algebra A(R) is, by Theorem 3.1 and the first part of Theorem
4.4, made of fading memory reservoir filters that map into L∞ (Ω, R). Using the other hypotheses in
Universal reservoir computers using non-homogeneous state-affine systems
30
the statement we can easily conclude that the family A(R) satisfies the thesis of Theorem 3.1 and it
is hence universal in the deterministic setup. The result follows from the second part of Theorem 4.4.
Acknowledgments: We thank Philipp Harms for carefully looking at early versions of this work
and for making suggestions that have significantly improved some of our results and Herbert Jaeger
and Josef Teichmann for fruitful discussions. The authors acknowledge partial financial support of
the French ANR “BIPHOPROC” project (ANR-14-OHRI-0002-02) as well as the hospitality of the
Centre Interfacultaire Bernoulli of the Ecole Polytechnique Fédérale de Lausanne during the program
“Stochastic Dynamical Models in Mathematical Finance, Econometrics, and Actuarial Sciences” that
made possible the collaboration that lead to some of the results included in this paper. LG acknowledges
partial financial support of the Graduate School of Decision Sciences and the Young Scholar Fund AFF
of the Universität Konstanz. JPO acknowledges partial financial support coming from the Research
Commission of the Universität Sankt Gallen and the Swiss National Science Foundation (grant number
200021 175801/1).
References
[Abra 88] R. Abraham, J. E. Marsden, and T. S. Ratiu. Manifolds, Tensor Analysis, and Applications.
Vol. 75, Applied Mathematical Sciences. Springer-Verlag, 1988.
[Apos 74] T. Apostol. Mathematical Analysis. Addison Wesley, second Ed., 1974.
[Arno 57] V. I. Arnold. “On functions of three variables”. Proceedings of the USSR Academy of
Sciences, Vol. 114, pp. 679–681, 1957.
[Bai 12]
Bai Zhang, D. J. Miller, and Yue Wang. “Nonlinear system modeling with random matrices: echo state networks revisited”. IEEE Transactions on Neural Networks and Learning
Systems, Vol. 23, No. 1, pp. 175–182, jan 2012.
[Barr 93]
A. Barron. “Universal approximation bounds for superpositions of a sigmoidal function”.
IEEE Transactions on Information Theory, Vol. 39, No. 3, pp. 930–945, may 1993.
[Boll 86]
T. Bollerslev. “Generalized autoregressive conditional heteroskedasticity”. Journal of Econometrics, Vol. 31, No. 3, pp. 307–327, 1986.
[Box 76]
G. E. P. Box and G. M. Jenkins. Time Series Analysis: Forecasting and Control. Holden-Day,
1976.
[Boyd 85] S. Boyd and L. Chua. “Fading memory and the problem of approximating nonlinear operators with Volterra series”. IEEE Transactions on Circuits and Systems, Vol. 32, No. 11,
pp. 1150–1161, nov 1985.
[Broc 06]
P. J. Brockwell and R. A. Davis. Time Series: Theory and Methods. Springer-Verlag, 2006.
[Bueh 06] M. Buehner and P. Young. “A tighter bound for the echo state property”. IEEE Transactions
on Neural Networks, Vol. 17, No. 3, pp. 820–824, 2006.
[Come 06] F. Comets and T. Meyre. Calcul Stochastique et Modèles de Diffusions. Dunod, Paris, 2006.
[Coui 16]
R. Couillet, G. Wainrib, H. Sevi, and H. T. Ali. “The asymptotic performance of linear echo
state neural networks”. Journal of Machine Learning Research, Vol. 17, No. 178, pp. 1–35,
2016.
Universal reservoir computers using non-homogeneous state-affine systems
31
[Croo 07] N. Crook. “Nonlinear transient computation”. Neurocomputing, Vol. 70, pp. 1167–1176,
2007.
[Cybe 89] G. Cybenko. “Approximation by superpositions of a sigmoidal function”. Mathematics of
Control, Signals, and Systems, Vol. 2, No. 4, pp. 303–314, dec 1989.
[Damb 12] J. Dambre, D. Verstraeten, B. Schrauwen, and S. Massar. “Information processing capacity
of dynamical systems”. Scientific reports, Vol. 2, No. 514, 2012.
[Dieu 69]
J. Dieudonne. Foundations of Modern Analysis. Academic Press, 1969.
[Doya 92] K. Doya. “Bifurcations in the learning of recurrent neural networks”. In: Proceedings of
IEEE International Symposium on Circuits and Systems, pp. 2777–2780, IEEE, 1992.
[Engl 82]
R. F. Engle. “Autoregressive conditional heteroscedasticity with estimates of the variance
of United Kingdom inflation”. Econometrica, Vol. 50, No. 4, pp. 987–1007, 1982.
[Flie 76]
M. Fliess. “Un outil algebrique : les series formelles non commutatives”. In: G. Marchesini
and S. K. Mitter, Eds., Mathematical Systems Theory, pp. 122–148, Springer Verlag, 1976.
[Flie 80]
M. Fliess and D. Normand-Cyrot. “Vers une approche algébrique des systèmes non linéaires
en temps discret”. In: A. Bensoussan and J. Lions, Eds., Analysis and Optimization of
Systems. Lecture Notes in Control and Information Sciences, vol. 28, Springer Berlin Heidelberg, 1980.
[Fran 10]
C. Francq and J.-M. Zakoian. GARCH Models: Structure, Statistical Inference and Financial
Applications. Wiley, 2010.
[Galt 14]
M. N. Galtier, C. Marini, G. Wainrib, and H. Jaeger. “Relative entropy minimizing noisy
non-linear neural network to approximate stochastic processes”. Neural Networks, Vol. 56,
pp. 10–21, 2014.
[Gang 08] S. Ganguli, D. Huh, and H. Sompolinsky. “Memory traces in dynamical systems.”. Proceedings of the National Academy of Sciences of the United States of America, Vol. 105, No. 48,
pp. 18970–5, dec 2008.
[Grig 15]
L. Grigoryeva, J. Henriques, L. Larger, and J.-P. Ortega. “Optimal nonlinear information processing capacity in delay-based reservoir computers”. Scientific Reports, Vol. 5,
No. 12858, pp. 1–11, 2015.
[Grig 16a] L. Grigoryeva, J. Henriques, L. Larger, and J.-P. Ortega. “Nonlinear memory capacity of
parallel time-delay reservoir computers in the processing of multidimensional signals”. Neural
Computation, Vol. 28, pp. 1411–1451, 2016.
[Grig 16b] L. Grigoryeva, J. Henriques, and J.-P. Ortega. “Reservoir computing: information processing of stationary signals”. In: Proceedings of the 19th IEEE International Conference on
Computational Science and Engineering, pp. 496–503, 2016.
[Grig 18]
L. Grigoryeva and J.-P. Ortega. “Echo state networks are universal”. Preprint, 2018.
[Grim 01] G. Grimmett and D. Stirzaker. Probability and Random Processes. Oxford University Press,
2001.
[Herm 10] M. Hermans and B. Schrauwen. “Memory in linear recurrent neural networks in continuous
time.”. Neural networks : the official journal of the International Neural Network Society,
Vol. 23, No. 3, pp. 341–55, apr 2010.
Universal reservoir computers using non-homogeneous state-affine systems
32
[Horn 13] R. A. Horn and C. R. Johnson. Matrix Analysis. Cambridge University Press, second Ed.,
2013.
[Horn 89] K. Hornik, M. Stinchcombe, and H. White. “Multilayer feedforward networks are universal
approximators”. Neural Networks, Vol. 2, No. 5, pp. 359–366, 1989.
[Jaeg 02]
H. Jaeger. “Short term memory in echo state networks”. Fraunhofer Institute for Autonomous Intelligent Systems. Technical Report., Vol. 152, 2002.
[Jaeg 04]
H. Jaeger and H. Haas. “Harnessing Nonlinearity: Predicting Chaotic Systems and Saving
Energy in Wireless Communication”. Science, Vol. 304, No. 5667, pp. 78–80, 2004.
[Jaeg 10]
H. Jaeger. “The ’echo state’ approach to analysing and training recurrent neural networks
with an erratum note”. Tech. Rep., German National Research Center for Information
Technology, 2010.
[Jone 92]
L. K. Jones. “A simple lemma on greedy approximation in hilbert space and convergence
rates for projection pursuit regression and neural network training”. The Annals of Statistics,
Vol. 20, No. 1, pp. 608–613, 1992.
[Kolm 56] A. N. Kolmogorov. “On the representation of continuous functions of several variables as
superpositions of functions of smaller number of variables”. Soviet Math. Dokl, Vol. 108,
pp. 179–182, 1956.
[Kurk 05] V. Kurkova and M. Sanguineti. “Learning with generalization capability by kernel methods
of bounded complexity”. Journal of Complexity, Vol. 21, No. 3, pp. 350–367, 2005.
[Ledo 91]
M. Ledoux and M. Talagrand. Probability in Banach Spaces. Springer-Verlag, 1991.
[Lord 14]
G. J. Lord, C. E. Powell, and T. Shardlow. An Introduction to Computational Stochastic
PDES. Cambridge University Press, 2014.
[Luko 09] M. Lukoševičius and H. Jaeger. “Reservoir computing approaches to recurrent neural network training”. Computer Science Review, Vol. 3, No. 3, pp. 127–149, 2009.
[Maas 00] W. Maass and E. D. Sontag. “Neural Systems as Nonlinear Filters”. Neural Computation,
Vol. 12, No. 8, pp. 1743–1772, aug 2000.
[Maas 02] W. Maass, T. Natschläger, and H. Markram. “Real-time computing without stable states:
a new framework for neural computation based on perturbations”. Neural Computation,
Vol. 14, pp. 2531–2560, 2002.
[Maas 04] W. Maass, T. Natschläger, and H. Markram. “Fading memory and kernel properties of
generic cortical microcircuit models”. Journal of Physiology Paris, Vol. 98, No. 4-6 SPEC.
ISS., pp. 315–330, 2004.
[Maas 07] W. Maass, P. Joshi, and E. D. Sontag. “Computational aspects of feedback in neural circuits”. PLoS Computational Biology, Vol. 3, No. 1, p. e165, 2007.
[Maas 11] W. Maass. “Liquid state machines: motivation, theory, and applications”. In: S. S. Barry
Cooper and A. Sorbi, Eds., Computability In Context: Computation and Logic in the Real
World, Chap. 8, pp. 275–296, 2011.
[Mack 77] M. C. Mackey and L. Glass. “Oscillation and chaos in physiological control systems”. Science,
Vol. 197, pp. 287–289, 1977.
Universal reservoir computers using non-homogeneous state-affine systems
33
[Manj 13] G. Manjunath and H. Jaeger. “Echo state property linked to an input: exploring a fundamental characteristic of recurrent neural networks”. Neural Computation, Vol. 25, No. 3,
pp. 671–696, 2013.
[Matt 92] M. B. Matthews. On the UniformApproximation of Nonlinear Discrete-Time FadingMemory Systems Using Neural Network Models. PhD thesis, ETH Zürich, 1992.
[Matt 93] M. B. Matthews. “Approximating nonlinear fading-memory operators using neural network
models”. Circuits, Systems, and Signal Processing, Vol. 12, No. 2, pp. 279–307, jun 1993.
[Munk 14] J. Munkres. Topology. Pearson, second Ed., 2014.
[Perr 96]
P. C. Perryman. Approximation Theory for Deterministic and Stochastic Nonlinear Systems.
PhD thesis, University of California, Irvine, 1996.
[Perr 97]
P. Perryman and A. Stubberud. “Uniform, in-probability approximation of stochastic systems”. In: Conference Record of The Thirtieth Asilomar Conference on Signals, Systems
and Computers, pp. 146–150, IEEE Comput. Soc. Press, 1997.
[Pisi 81]
G. Pisier. “Remarques sur un résultat non publié de B. Maurey”. Séminaire d’analyse
fonctionnelle École Polytechnique, pp. 1–12, 1981.
[Rusc 98] L. Rüschendorf and W. Thomsen. “Closedness of sum spaces and the generalized schrödinger
Problem”. Theory of Probability & Its Applications, Vol. 42, No. 3, pp. 483–494, jan 1998.
[Sont 79a] E. Sontag. “Realization theory of discrete-time nonlinear systems: Part I-The bounded
case”. IEEE Transactions on Circuits and Systems, Vol. 26, No. 5, pp. 342–356, may 1979.
[Sont 79b] E. D. Sontag. “Polynomial Response Maps”. In: Lecture Notes Control in Control and
Information Sciences. Vol. 13, Springer Verlag, 1979.
[Spre 65]
D. A. Sprecher. “A representation theorem for continuous functions of several variables”.
Proceedings of the American Mathematical Society, Vol. 16, No. 2, p. 200, apr 1965.
[Spre 96]
D. A. Sprecher. “A numerical implementation of Kolmogorov’s superpositions”. Neural
Networks, Vol. 9, No. 5, pp. 765–772, 1996.
[Spre 97]
D. A. Sprecher. “A numerical implementation of Kolmogorov’s superpositions II”. Neural
Networks, Vol. 10, No. 3, pp. 447–457, 1997.
[Stub 97a] A. Stubberud and P. Perryman. “Current state of system approximation for deterministic
and stochastic systems”. In: Conference Record of The Thirtieth Asilomar Conference on
Signals, Systems and Computers, pp. 141–145, IEEE Comput. Soc. Press, 1997.
[Stub 97b] A. Stubberud and P. Perryman. “State of system approximation for stochastic systems”.
In: Proceedings of 13th International Conference on Digital Signal Processing, pp. 711–714,
IEEE, 1997.
[Suss 76]
H. J. Sussmann. “Semigroup representations, bilinear approximations of input-output maps,
and generalized inputs”. In: G. Marchesini and S. K. MItter, Eds., Mathematical Systems
Theory, pp. 172–191, Springer Verlag, 1976.
[Take 81]
F. Takens. “Detecting strange attractors in turbulence”. pp. 366–381, Springer Berlin
Heidelberg, 1981.
Universal reservoir computers using non-homogeneous state-affine systems
[Vers 07]
34
D. Verstraeten, B. Schrauwen, M. D’Haene, and D. Stroobandt. “An experimental unification
of reservoir computing methods”. Neural Networks, Vol. 20, pp. 391–403, 2007.
[Wain 16] G. Wainrib and M. N. Galtier. “A local echo state property through the largest Lyapunov
exponent”. Neural Networks, Vol. 76, pp. 39–45, apr 2016.
[Whit 04] O. White, D. Lee, and H. Sompolinsky. “Short-Term Memory in Orthogonal Neural Networks”. Physical Review Letters, Vol. 92, No. 14, p. 148102, apr 2004.
[Yild 12]
I. B. Yildiz, H. Jaeger, and S. J. Kiebel. “Re-visiting the echo state property.”. Neural
networks : the official journal of the International Neural Network Society, Vol. 35, pp. 1–9,
nov 2012.
[Zang 04] G. Zang and P. A. Iglesias. “Fading memory and stability”. Journal of the Franklin Institute,
Vol. 340, No. 6-7, pp. 489–502, 2004.
| 9cs.NE |
A Two-Phase Power Allocation Scheme for CRNs
Employing NOMA
Ming Zeng, Georgios I. Tsiropoulos∗ , Animesh Yadav, Octavia A. Dobre, and Mohamed H. Ahmed
Faculty of Engineering and Applied Science, Memorial University, St. John, Canada
of Electrical and Computer Engineering, National Technical University of Athens, Greece
Email: {mzeng, animeshy, odobre, mhahmed}@mun.ca, gitsirop@mail.ntua.gr
arXiv:1709.05304v1 [cs.IT] 15 Sep 2017
∗ School
Abstract—In this paper, we consider the power allocation (PA)
problem in cognitive radio networks (CRNs) employing nonorthogonal multiple access (NOMA) technique. Specifically, we
aim to maximize the number of admitted secondary users (SUs)
and their throughput, without violating the interference tolerance
threshold of the primary users (PUs). This problem is divided
into a two-phase PA process: a) maximizing the number of
admitted SUs; b) maximizing the minimum throughput among
the admitted SUs. To address the first phase, we apply a
sequential and iterative PA algorithm, which fully exploits the
characteristics of the NOMA-based system. Following this, the
second phase is shown to be quasiconvex and is optimally solved
via the bisection method. Furthermore, we prove the existence
of a unique solution for the second phase and propose another
PA algorithm, which is also optimal and significantly reduces
the complexity in contrast with the bisection method. Simulation
results verify the effectiveness of the proposed two-phase PA
scheme.
I. I NTRODUCTION
With the proliferation of smart mobile devices, such as smart
phones, M2M, and emerging wearables, global mobile data
traffic is expected to grow to 30.6 EB per month by 2020
[1]. In order to meet the mobile data traffic requirement, a
10 Gbps peak data rate and 1 Gbps user experienced data
rate are proposed to be supported by 5G [2]. Hence, with
limited spectrum availability, enhancing spectral efficiency is
of significant importance for 5G, becoming one of its main
design requirements.
A prevailing way to address spectrum scarcity is to apply dynamic and efficient spectrum accessing techniques,
such as cognitive radio (CR) [3]–[6]. CR networks (CRNs)
are envisioned to provide more bandwidth to mobile users
through heterogeneous architectures and dynamic spectrum
access techniques. Therefore, network users are divided into
two main groups: licensed/primary users (PUs) and unlicensed/secondary users (SUs). Correspondingly, two requirements have to be satisfied in CRNs: a) the interference
introduced by the operation of SUs towards PUs should be
kept under a certain threshold, and b) the admitted SUs should
meet their minimum data rate requirement. Particularly, for
spectrum sharing, spectrum underlay or overlay techniques can
be used for designing CRNs [3], [4].
Another way is to employ non-orthogonal multiple access
(NOMA) [7]–[10], which has attracted considerable attention
recently owing to its potential to achieve superior spectral
efficiency. Unlike conventional orthogonal multiple access
(OMA), NOMA multiplexes users in the power-domain at
the transmitter side, and conducts multi-user signal separation
using successive interference cancellation (SIC) at the receiver
side. Thus, in NOMA-based systems, power allocation (PA)
is of great importance, since it not only impacts the users’
achievable data rates, but also determines their channel access.
A variety of PA strategies have been proposed so far, targeting
different aspects of PA in NOMA [11]–[14]. CR-inspired PA is
adopted in [13], where NOMA is considered as a special case
of CRNs and the user with poor channel condition (poor user)
is viewed as a PU. This way, the quality of service (QoS)
for the poor user can be strictly guaranteed. However, the
performance of the user with better channel condition may be
sacrificed since this user is served only after the poor user’s
QoS is met. To offer more flexibility in the tradeoffs between
the user fairness and system throughput, a dynamic PA scheme
is proposed in [14], which strictly guarantees the performance
gain of NOMA over OMA for both poor user and user with
better channel conditions.
In this paper, we study the two concepts of NOMA and
CRN, i.e., a CRN employing NOMA for its SUs, leading to a
further increase in spectral efficiency. In such a NOMA-based
CRN, PA for SUs not only determines the channel access of
SUs, but also affects the performance of PUs. Consequently,
the performance of the adopted PA scheme is vital, and a
full exploitation of the power domain should be achieved.
Motivated by this, our contributions can be summarized as
follows:
1) we propose a two-phase PA scheme to maximize the
number of admitted SUs and their throughput;
2) the first phase maximizes the number of admitted SUs.
NOMA has been well studied in terms of network
throughput, link quality, outage probability estimation
etc. However, the system capacity in terms of number of
admitted users has not been studied so far. The analysis
in this paper provides useful insights;
3) the second phase aims to maximize the minimum data
rate among the admitted SUs. It is worth noting that the
max-min problems are investigated in [15], [16]. [15]
considers the max-min fairness criterion under statistical
channel state information (CSI), and aims to achive
outage balancing among the users, which is different
from the user admission problem considered in this
paper. In [16], there is no QoS requirement for the
users, and thus, sum rate maximization is pursued, which
differs from our system model, in which each SU has a
minimum data rate requirement;
4) numerical simulations are conducted to verify the effectiveness of the proposed two-phase PA scheme.
The rest of the paper is organized as follows. The system
model and problem formulation are presented in Section II.
The proposed two-phase PA process is introduced in Section
III. Performance evaluation results are illustrated in Section
IV and conclusions are drawn in Section V.
Fig. 1. Spectrum sharing scheme.
II. S YSTEM M ODEL AND P ROBLEM F ORMULATION
A. System Model
A hierarchical spectrum sharing CRN is considered, where
spectrum underlay is employed. NOMA is adopted to reduce
the interference among SUs so as to further improve the
spectrum efficiency.
Fig. 1 shows the structure of the considered NOMA-based
CRN. The PUs are served by the base station (BS) in the
downlink. Meanwhile, the BS can serve the SUs simultaneously obeying the access rules (i.e., the interference from the
SUs towards each PU should be less than a certain interference
threshold). On the other hand, the admitted SUs should meet
their minimum data rate requirement, which is characterized
by the signal-to-interference-plus-noise ratio (SINR); an SU is
admitted if its SINR requirement is met.
The system model and transmission settings follow the ones
in [17], and are defined as follows. We assume that there exist
M PUs and N SUs in the network. The channel gain between
the BS and the nth SU is denoted as Gn , n ∈ {1, · · · , N },
which strongly depends on the distance between them. Likewise, we denote the channel gain from the BS to the mth PU
as gm , m ∈ {1, · · · , M }. We consider that the channel gains
are known at the BS. Without loss of generality, we arrange
the SUs in a descending order as follows:
G1 ≥ . . . ≥ Gn . . . ≥ GN .
(1)
The aggregate noise and interference from all PUs towards
the nth SU are denoted as Nn . As NOMA is employed among
SUs, SUs with better channel gains can cancel the interference
from users with lower channel gains through SIC. As a result,
the SINR of SUs can be calculated as [10], [18], [19]
γn = Pn−1
j=1
Pn Gn
Pj Gn + Nn
N
X
Pn gm ≤ Im .
(2)
Equation (4) can be further rewritten as
Ps ≤
Im
,
gm
(5)
PN
where Ps = n=1 Pn corresponds to the overall power for
all SUs. This is fulfilled if
Ps ≤ P M ,
Im
gm
(6)
where PM = min
, m ∈ {1, . . . , M }. As Ps should also
be constrained to a maximum power Pmax , i.e., Ps ≤ Pmax ,
it yields
Ps = min (PM , Pmax ).
(7)
In the following sections, Ps is assumed to be known and
directly used as the overall power constraint for SUs.
B. Problem Formulation
The problem of maximizing the number of admitted SUs
and their throughput is investigated. This problem consists of
two phases: 1) maximizing the number of admitted SUs under
the PUs’ QoS requirement and SUs’ minimum data requirement; and 2) maximizing the minimum SINR among the admitted SUs via the allocation of the remaining power. Denote
the total number of admitted SUs as L, L ∈ {0, . . . , N }, and
the corresponding SUs as a1 , a2 , . . . , aL . On this basis, the
problem in the first phase is formulated as
~
P
s.t.
(3)
(4)
n=1
max L
,
where γn and Pn denote the SINR and allocated power of the
nth SU, respectively.
Further, if the SINR threshold of the nth SU is Γn , γn needs
to satisfy
γn ≥ Γn .
In terms of the QoS requirement for PUs, denote the
maximum interference level tolerable by the mth PU by Im ,
the corresponding interference constraint can be formulated as
γn ≥ Γn , n ∈ {a1 , a2 , . . . , aL },
aL
X
Pn ≤ P s .
(8a)
(8b)
(8c)
n=a1
where P~ = [P1 . . . PN ] is the overall PA vector.
After the PA process of the first phase, SUs admission is
ascertained. We use L? and a1 , a2 , . . . , aL? to denote the
maximum number of admitted SUs and the corresponding
SUs. On this basis, the problem in the second phase is to
further increase the SINR among the L? admitted SUs to
increase system throughput, which is formulated as
Similarly, the power for the nth SU is given by
Pn = Γn
n−1
X
Pj +
j=1
max min(~γ )
γn ≥ Γn , n ∈ {a1 , a2 , . . . , aL? },
aL?
X
Pn ≤ Ps ,
(9b)
(9c)
n=a1
where P~L? = [Pa1 . . . PaL? ] and ~γ = [γa1 . . . γaL? ] are the PA
vector and the SINR of the L? admitted SUs, respectively.
We can consider the second phase from two perspectives:
a) from the angle of exploiting the remaining power after
the PA process of the first phase, with the goal of allocating
the remaining power appropriately to maximize the minimum
SINR among the L? admitted SUs; b) consider itself as
an independent problem without taking into account the PA
process of the first phase and its remaining power. In this
case, (9) can be viewed as a problem of allocating the overall
power of all SUs among the L? admitted SUs, subjecting to
each admitted SU satisfying its SINR requirement.
III. P ROPOSED T WO -P HASE PA S CHEME
In this section, the proposed two-phase PA scheme is
presented. First, we address the first phase and give a detailed
description of the corresponding PA algorithm. Following this,
the second phase is resolved.
A. SUs Admission and Initial PA Algorithm
In order to maximize the number of admitted SUs, we
employ the sequential and iterative PA scheme, which makes
full use of the characteristics of NOMA-based CRN and
allocates power to SUs in a descending order according to
their channel gains.
According to the SINR requirements. i.e., (2) and (3), we
have the following equations
Pn > Γ n
n−1
X
j=1
Pj +
(13)
(9a)
~L?
P
s.t.
Γn Nn
.
Gn
Γn Nn
, n = 1, . . . , N,
Gn
γ1 N1
.
G1
(11)
Sequentially and iteratively, since the power of the 1st SU
is known, it is used for the PA of the 2nd SU. We attain the
following equation according to (10)
P2 = Γ2 P1 +
Γ 2 N2
.
G2
Pn = min (Γn
n−1
X
j=1
Pj + Γ n
n−1
X
Nn
, Ps −
Pj ).
Gn
j=1
(14)
during each allocation, whenever Ps −
PFurthermore,
P
n−1
n−1
Nn
j=1 Pj < Γn
j=1 Pj + Γn Gn , it indicates there is not
enough power left for the nth SU to meet its SINR requirement. Consequently, the PA process terminates and the nth
SU to the N th SU receives no power to ensure that the QoS
requirements for PUs are not violated. The admitted SUs are
1st SU, . . . , (n − 1)th SU, with the allocated power given
by (13). After the PA process,
note that there exists some
Pn−1
remaining power, i.e., Ps − j=1 Pj . While this power is
not large enough to admit an extra SU, it can be further
allocated to the admitted SUs to increase their SINR, and thus,
to enhance the throughput. Particularly, according to (13), the
power required for the admission of the nth SU is even larger
than the sum of the power for all the former n − 1 SUs,
in case of Γn ≥ 1. Therefore, the remaining power indeed
randomly lies in the boundary [0, Pn ], which may help enhance
the throughput of the former n − 1 SUs significantly.
In [17], we have proven that when the SINR requirement
of each SU is the same, the above PA algorithm is optimal.
Moreover, the computational complexity is only O(N ).
B. Maximize the Minimum SINR among the Admitted SUs
(10)
Pn−1
where the only variable is j=1 Pj , since other parameters,
i.e., Γn , Nn and Gn are known to the BS. Therefore, if we
assign the power among SUs following the ascending order,
i.e., from the 1st SU to the N th SU sequentially,
Pn−1 the power
for the nth SU can be obtained easily, as j=1 Pj is already
known. Specifically, the power for the 1st SU is calculated as
P1 =
Obviously, (13) can be used for the PA of all SUs. On the
other hand, note that Ps has not been considered during the
above PA process. Therefore, during the PA for the nth SU,
we also
Pnneed to guarantee that the total power assigned to
SUs, j=1 Pj , does not exceed Ps . This is achieved by
(12)
1) Analysis via Convex Optimization: After the initial PA
process, the SUs admission is done. On this basis, the second
phase aims to maximize the minimum SINR among the
admitted SUs.
Indeed, the second phase is quasiconvex. Note that the
variables in (9) are the power values of the admitted SUs.
Accordingly, the SINR of each SU is in the form of linear
fractional function, which is a quasiconcave function. Following this, the operation of selecting the minimum value
from a set of quasiconcave function will not change its quasiconcavity. Moreover, maximizing a quasiconcave function is
equivalent to minimizing a quasiconvex function. In addition,
the constraint functions are all convex. Consequently, the objective function is a quasiconvex function. According to [20],
a general approach to quasiconvex optimization relies on the
representation of the sublevel sets of a quasiconvex function
via a family of convex inequalities. Now by introducing an
axillary variable t, (9) can be equivalently represented as
find
s.t.
P~L?
(15a)
?
γn ≥ t, n ∈ {1, . . . , L }
(15b)
γn ≥ Γn , n ∈ {1, . . . , L? }
L?
X
Pn ≤ Ps .
(15c)
(15d)
n=1
If we substitute (2) into (15), (15-b) and (15-c) can be
written respectively as
Pn Gn − t
n−1
X
Pj Gn + tNn ≥ 0,
(16)
Pj Gn + Γn Nn ≥ 0.
(17)
j=1
and
Pn Gn − Γn
n−1
X
j=1
Further, we combine (16) and (17) through λn = max(t, Γn ),
which yields
Pn Gn − λn
n−1
X
Pj Gn + λn Nn ≥ 0.
(18)
j=1
Then, (15) can be reformulated as
find P~L?
s.t.
Pn Gn − λn
(19a)
n−1
X
Pj Gn + λn Nn ≥ 0,
(19b)
j=1
?
n ∈ {1, . . . , L }
L?
X
Pn ≤ Ps .
(19c)
n=1
For any given value of t, λn has a specific value. Thus, (19)
is a convex feasibility problem, since the inequality constraint
functions are all linear. Let θ? denote the optimal value of the
quasiconvex optimization problem (10). If (19) is feasible, i.e.,
there exists P~L? satisfying (19), we have θ? ≥ t. Otherwise,
we can conclude θ? ≤ t. Therefore, we can check whether
the optimal value θ? of a quasiconvex optimization problem is
over or below a given value t by solving the convex feasibility
problem (19).
The observation above can be used as the basis of a simple
algorithm for solving the quasiconvex optimization problem
(19) via bisection, i.e., solving a convex feasibility problem
at each step. Firstly, the problem is set to be feasible, e.g.,
we start with an interval [l, u] known to contain the optimal
value θ? . Then, the convex feasibility problem is solved at
its midpoint t = (l + u)/2, to determine whether the optimal
value is in the lower or upper half of the interval, and update
the interval accordingly. This produces a new interval, which
also contains the optimal value, but has half the width of the
initial interval. The progress is repeated until the width of the
interval is small enough.
Algorithm 1 Optimal Method for Quasiconvex Optimization
1: Initialize parameters.
2:
Given l ≤ θ? , u ≥ θ? , tolerance ≥ 0, where
3:
l = min(Γn ), u = max(Ps /Nn ), n ∈ {1, . . . , L? };
4: repeat:
5:
t = (l + u)/2;
6:
Solve the convex problem (19);
7:
If (19) is feasible, u = t; else l = t;
8: Until u − l ≤
The interval [l, u] is guaranteed to contain θ? , i.e., we have
l ≤ θ? ≤ u at each step. It is obvious that l = min(Γn ), n ∈
{1, . . . , L? } can be used as the lower boundary, according to
the SINR requirement. In terms of the upper boundary, the
highest SINR the system can achieve should not exceed the
value when all the power is allocated to a single SU. In each
iteration, the interval is divided in two, i.e., bisected. Thus, the
length of the interval after k iterations is 2−k (u − l), where
u − l is the length of the initial interval. It follows that exactly
dlog2 ((u − l)/)eiterations are required before the algorithm
terminates. Each step involves solving the convex feasibility
problem (19).
2) An Analytical Solution Based on the Water-filling
Scheme: The second problem can be solved using the optimal
method for quasiconvex optimization. However, although the
number of iterations is fixed for a given threshold, it is still
computationally complex since each step requires solving the
convex feasibility problem. In this section, we firstly prove
the existence of a unique solution for this problem, and then
propose a PA algorithm based on the water-filling scheme to
obtain the optimal solution.
a) Existence of a unique solution: We assume that the
maximized minimum SINR among the admitted SUs via the
full exploitation of the remaining power is θ? . Accordingly,
it is proven by contradiction in the following paragraph that
γn = max(θ? , Γn ), n = {1, . . . , L? }, i.e., when Γn ≤
θ? , γn = θ? ; otherwise γn = Γn . Hence, once θ? is certain,
the SINR for each admitted SU is updated accordingly from
its targeted SINR. Then, based on (13), the power allocated
to the
SUs can be obtained, and their sum should satisfy
PL?
P
n=1 n = Ps . This equation only has one variable, i.e.,
θ? . Furthermore, Ps monotonically increases with θ? . Since
Ps is fixed, θ? has a unique value.
Proof: Let us assume that for the nth SU, Γn ≤ θ? , γn >
?
θ . Then, we can simply improve θ? by transferring some
power from the nth SU to other SUs whose SINR equal to
θ? . This contradicts our premise that θ? is the maximized
minimum SINR. Likewise, we can easily prove that for the
nth SU with Γn > θ? , γn should equal Γn .
b) Proposed PA algorithm based on the water-filling
scheme: The above analysis shows the existence of a unique
solution for the second problem. However, due to the operation
of comparison between the targeted SINR and the maximized
minimum SINR, i.e., max(θ? , Γn ), the function between the
overall power Ps and θ? is piecewise. Indeed, this piecewise
14
30
Number of Requesting SUs=5
Number of Requesting SUs=10
Number of Requesting SUs=15
12
Number of Requesting SUs=5
Number of Requesting SUs=10
Number of Requesting SUs=15
Target SINR
10
Achieved SINR [dB]
Number of Admitted SUs
25
8
6
20
15
10
4
2
5
10
15
20
25
Target SINR [dB]
5
5
10
15
20
25
Target SINR [dB]
Fig. 2. Number of admitted SUs vs. targeted SINR.
Fig. 3. Achieved SINR vs. targeted SINR.
function can be calculated using the water-filling scheme.
Procedure: After employing the initial PA algorithm,
the SINR of each admitted SU is satisfied. There is some
remaining power, which could be allocated to the admitted
SUs. We first allocate the remaining power to the SU with
lowest SINR. If the remaining power is large enough, the SINR
of the SU with lowest SINR would reach the value of the one
with the second lowest SINR. Then, power is assigned to the
above two SUs so that their SINRs are equivalent to the SU
with the third lowest SINR. The process repeats until there is
no power left. Note that the bisection method may be required
to obtain the value of θ? , when it lies in the middle of two
SINR thresholds of the admitted SUs.
Optimality analysis: According to the PA process, we can
conclude that the SINR of the SUs with extra power allocated
should be equal. Moreover, the method is optimal. Assuming
this is not optimal, then the optimal solution should provide
the SINR of the SUs with extra power allocated is not equal,
i.e., at least the SINR of one SU is larger than another one.
Since our objective is to maximize the minimum SINR of all
admitted SUs, we can simply reallocate some power from the
larger one to the smaller one to improve it. This conflicts with
our hypothesis, and proves the optimality of the proposed PA
scheme.
represent the lognormal shadowing, which are random Gaussian variables with zero mean and standard deviation equal to
6 dB. Additionally, system and transmission parameters e.g.,
antenna gain, carrier frequency, etc., are included in K = 103 .
Moreover, we set Ns = −120 dBm, Im = −90 dBm and
Pmax = 20 dBm, where Ns denotes the total noise and
interference from all PUs for each SU.
The performance of the proposed two-phase PA scheme is
investigated from two perspectives. First, the effectiveness of
the initial PA process is studied. Fig. 2 shows the mean number
of admitted SUs versus the targeted SINR, when the number of
requesting SUs varies, i.e., N = 5, 10 or 15. As the targeted
SINR increases, the number of admitted SUs decreases for all
three cases. However, the number of admitted SUs is quite high
for all targeted SINRs. Particularly, when the targeted SINR
is 5 dB, almost all requesting SUs are admitted for the three
cases, which proves the effectiveness of the initial PA process.
In addition, for any given targeted SINR, by comparing the
three cases, one can observe that as the number of requesting
SUs increases, the number of admitted SUs grows as well. This
can be explained by the fact that as the number of requesting
SUs increases, it is likely that more users will have better
channel gains. According to (14), the increase in channel gains
yields lower power consumption, and thus more users can be
admitted. Even when the targeted SINR reaches 25 dB, about
4.5 users are admitted when N = 15.
Following this, Fig. 3 compares the achieved SINR versus
targeted SINR, when the targeted SINRs for all users are
the same. For the case of N = 5, about 1.5 dB increment
is achieved for any given targeted SINR, which verifies the
usefulness of the second PA process. As the number of
requesting SUs grows, the growth in SINR declines in general.
This is because the remaining power is divided by more
admitted SUs. Particularly, when the targeted SINR is 5 and
IV. S IMULATION R ESULTS
Simulations are run for a NOMA-based CRN shown in
Fig. 1, with the cell radius of 500 m, and the BS located at the
centre. The numerical results are obtained through averaging
over 104 simulation runs. During each simulation, SUs and
PUs are randomly distributed in the area following uniform
distribution. More exactly, their channel gains are modelled
Hn
hm
as Gn = K · 10 10 Dn−4 , gm = K · 10 10 d−4
m , where Dn
and dm are the corresponding distances, while Hn and hm
R EFERENCES
10
Targeted SINR
9
Achieved SINR
8
SINR [dB]
7
6
5
4
3
2
1
0
1
2
3
4
5
6
7
8
Different admitted SUs
Fig. 4. Snapshot when the targeted SINR of each SU varies.
10, the SINR increment for N = 10 and 15 is less than 1.
This matches with Fig. 2, since for these two targeted SINR
values, the admitted number of SUs are much larger than that
of N = 5. However, for the remaining three targeted SINR
values, as the number of admitted SUs is almost the same
for the three cases, a similar increase is obtained. In all, an
average of 1 dB increment is obtained for the three cases,
which validates the effectiveness of the proposed second PA
process, when the targeted SINRs for all users are the same.
Lastly, we compare the achieved SINR versus the targeted
SINR in Fig. 4, when the targeted SINR of each SU varies.
Note that the requesting number of users is set to 15. Here,
we cannot average the results, as the targeted SINR of each
SU is different. Instead, a snapshot is given. Specifically, in
Fig. 4, 8 SUs are admitted, among which the SINR values
of 5 SUs are further increased, and for the other 3 SUs, they
remain unchanged. On average, there is 1 dB increment.
V. CONCLUSIONS
In this paper, PA for CRNs employing NOMA is considered.
A two-phase PA scheme is proposed to maximize the number
of admitted SUs and the throughput. Specifically, in the first
phase, we apply the sequential and iterative PA algorithm
to obtain the maximum number of admitted SUs. Following
this, the second PA algorithm maximizes the minimum SINR
among the admitted SUs. Simulation results show that the
number of admitted SUs is large under different number of
requesting SUs; and there is over 1 dB increment on average in
the SINR of the admitted SUs, which verifies the effectiveness
of the proposed two-phase PA scheme.
ACKNOWLEDGEMENT
This work has been supported in part by the Natural Sciences and Engineering Research Council of Canada (NSERC),
through its Discovery program.
[1] C. V. N. Index, “Global mobile data traffic forecast update, 2012-2017,”
Cisco white paper, 2015.
[2] N. Alliance, “5G white paper,” Next Generation Mobile Networks, White
paper, 2015.
[3] I. F. Akyildiz, W.-Y. Lee, M. C. Vuran, and S. Mohanty, “A survey
on spectrum management in cognitive radio networks,” IEEE Commun.
Mag., vol. 46, no. 4, pp. 40–48, Apr. 2008.
[4] G. I. Tsiropoulos, O. A. Dobre, M. H. Ahmed, and K. E. Baddour,
“Radio resource allocation techniques for efficient spectrum access in
cognitive radio networks,” IEEE Commun. Surv. Tuts., no. 99, pp. 82–95,
Jan. 2016.
[5] G. I. Tsiropoulos, Z. Ming, O. A. Dobre, and M. H. Ahmed, “A loadbalancing semi-matching approach for resource allocation in cognitive
radio networks,” in Proc. IEEE ICC, 2016, pp. 1–6.
[6] G. I. Tsiropoulos, A. Yadav, M. Zeng, and O. A. Dobre, “Cooperation in
5G heterogeneous networks: Advanced spectrum access and d2d assisted
communications,” IEEE Wireless Commun. Mag., vol. PP, no. 99, pp.
1–1, Jun. 2017.
[7] S. M. R. Islam, M. Zeng, and O. A. Dobre, “NOMA in 5G systems: Exciting possibilities for enhancing spectral efficiency,” IEEE
5G Tech. Focus, vol. 1, no. 2, May 2017. [Online]. Available: 307
http://5g.ieee.org/tech-focus.
[8] M. Zeng, A. Yadav, O. A. Dobre, G. I. Tsiropoulos, and H. V. Poor,
“Capacity comparison between MIMO-NOMA and MIMO-OMA with
multiple users in a cluster,” IEEE J. Select. Areas Commun., vol. PP,
no. 99, pp. 1–1, Jun. 2017.
[9] ——, “On the sum rate of MIMO-NOMA and MIMO-OMA systems,”
IEEE Wireless Commun. Lett., vol. PP, no. 99, pp. 1–1, Jun. 2017.
[10] S. M. R. Islam, N. Avazov, O. A. Dobre, and K. S. Kwak, “Powerdomain non-orthogonal multiple access (NOMA) in 5G systems: Potentials and challenges,” IEEE Commun. Surv. Tuts.,, vol. 19, no. 2, pp.
721–742, May 2017.
[11] A. Li, A. Harada, and H. Kayama, “A novel low computational complexity power assignment method for non-orthogonal multiple access
systems,” IEICE Trans. Fundamentals, vol. E97.A, no. 1, pp. 57–68,
2014.
[12] N. Otao, Y. Kishiyama, and K. Higuchi, “Performance of non-orthogonal
access with SIC in cellular downlink using proportional fair-based
resource allocation,” in Proc. IEEE ISWCS, 2012, pp. 476–480.
[13] Z. Ding, R. Schober, and H. V. Poor, “A general MIMO framework for
NOMA downlink and uplink transmission based on signal alignment,”
IEEE Trans. Wireless Commun., vol. 15, no. 6, pp. 4438–4454, Jun.
2016.
[14] Z. Yang, Z. Ding, P. Fan, and N. Al-Dhahir, “A general power allocation
scheme to guarantee quality of service in downlink and uplink NOMA
systems,” IEEE Trans. Wireless Commun., vol. 15, no. 11, pp. 7244–
7257, Nov. 2016.
[15] J. Cui, Z. Ding, and P. Fan, “A novel power allocation scheme under
outage constraints in noma systems,” IEEE Signal Process. Lett., vol. 23,
no. 9, pp. 1226–1230, Sep. 2016.
[16] S. Timotheou and I. Krikidis, “Fairness for non-orthogonal multiple
access in 5G systems,” IEEE Signal Process. Lett., vol. 22, no. 10, pp.
1647–1651, Oct. 2015.
[17] Z. Ming, G. I. Tsiropoulos, O. A. Dobre, and M. H. Ahmed, “Power
allocation for cognitive radio networks employing non-orthogonal multiple access,” in Proc. IEEE Globecom, pp. 1–5, 2016.
[18] L. Dai, B. Wang, Y. Yuan, S. Han, C. l. I, and Z. Wang, “Non-orthogonal
multiple access for 5G: Solutions, challenges, opportunities, and future
research trends,” IEEE Commun. Mag., vol. 53, no. 9, pp. 74–81, Sep.
2015.
[19] Z. Ding, Y. Liu, J. Choi, Q. Sun, M. Elkashlan, and H. V. Poor, “Application of non-orthogonal multiple access in LTE and 5G networks,”
IEEE Commun. Mag., vol. 55, no. 2, pp. 185–191, Feb. 2017.
[20] S. Boyd and L. Vandenberghe, Convex Optimization.
Cambridge
University Press, 2004.
| 7cs.IT |
Automatic Kolmogorov complexity and
normality revisited
Alexander Shen∗
arXiv:1701.09060v3 [cs.IT] 28 Aug 2017
August 29, 2017
Abstract
It is well known that normality (all factors of a given length appear in an infinite sequence with the same frequency) can be described as incompressibility via finite automata.
Still the statement and the proof of this result as given by Becher and Heiber [4] in terms of
“lossless finite-state compressors” do not follow the standard scheme of Kolmogorov complexity definition (an automaton is used for compression, not decompression). We modify
this approach to make it more similar to the traditional Kolmogorov complexity theory
(and simpler) by explicitly defining the notion of automatic Kolmogorov complexity and
using its simple properties. Other known notions (Shallit–Wang [16], Calude–Salomaa–
Roblot [8]) of description complexity related to finite automata are discussed (see the last
section).
As a byproduct, this approach provides simple proofs of classical results about normality (equivalence of definitions with aligned occurrences and all occurrences, Wall’s theorem saying that a normal number remains normal when multiplied by a rational number,
and Agafonov’s result saying that normality is preserved by automatic selection rules).
1 Introduction
What is an individual random object? When could we believe, looking at an infinite sequence
α of zeros and ones, that α was obtained by tossing a fair coin? The minimal requirement is
that zeros and ones appear “equally often” in α : both have limit frequency 1/2. Moreover, it
is natural to require that all 2k bit blocks of length k appear equally often. Sequences that have
this property are called normal (see the exact definition in Section 3; a historic account can be
found in [4, 6]).
Intuitively, a reasonable definition of an individual random sequence should require much
more than just normality; the corresponding notions are studied in the algorithmic randomness theory (see [9, 13] for the detailed exposition, [18] for a textbook and [17] for a short
survey). The most popular definition is called Martin-Löf randomness; the classical Schnorr–
Levin theorem says that this notion is equivalent to incompressibility: a sequence α is MartinLöf random if an only if prefixes of α are incompressible (do not have short descriptions). See
again [9, 13, 18, 17] for exact definitions and proofs.
LIRMM, CNRS and University of Montpellier, France. On leave from IITP RAS, Moscow. Supported by
ANR-15-CE40-0016 RaCAF grant. Part of the work was done while visiting National Research University High
School of Economics, Moscow. E-mail: sasha.shen@gmail.com or alexander.shen@lirmm.fr
∗
1
It is natural to expect that normality, being a weak randomness property, corresponds to
some weak incompressibility property. The connection between normality and finite-state computations was noticed long ago, as the title of [1] shows. This connection led to a characterization of normality as “finite-state incompressibility” (see [4]). However, the notion of
incompressibility that was used in [4] does not fit well the general framework of Kolmogorov
complexity (finite automata are considered as compressors, while in the usual definition of Kolmogorov complexity we restrict the class of allowed decompressors).
In this paper we give a definition of automatic Kolmogorov complexity that restricts the
class of allowed decompressors and is suitable for the characterization of normal sequences as
incompressible ones. This definition and its properties are considered in Section 2. In Section 3
we recall the notion of normal sequence. Then in Section 4 we provide a characterization
of normal sequences in terms of automatic Kolmogorov complexity. In Section 5 we show
how this characterization can be used to give simple proofs for classical results about normality,
including Wall’s theorem (normal numbers remain normal when multiplied by a rational factor).
In a similar way one can prove Agafonov’s result [1], but we need to modify the definition of
complexity using pairs as descriptions. This is done in Section 6. Finally, in Section 7 we
compare our definition of automatic complexity with other similar notions.
2 Automatic Kolmogorov complexity
Let us recall the definition of algorithmic (Kolmogorov) complexity. It is usually defined in the
following way: C(x), the complexity of an object x, is the minimal length of its “description”.
(We assume that both objects and descriptions are binary strings; the set of binary strings is
denoted by B∗ , where B = {0, 1}.) Of course, this definition makes sense only after we explain
which type of “descriptions” we consider, but most versions of Kolmogorov complexity can be
described according to this scheme [19]:
Definition 1. Let D ⊂ B∗ × B∗ be a binary relation; we read (p, x) ∈ D as “p is a D-description
of x”. Then complexity function CD is defined as
CD (x) = min{|p| : (p, x) ∈ D},
i.e., as the minimal length of a D-description of x.
Here |p| stands for the length of a binary string p and min(∅) = +∞, as usual. We say that
D is a description mode and CD (x) is the complexity of x with respect to the description mode D.
We get the original version of Kolmogorov complexity (“plain complexity”) if we consider
all computable functions as description modes, i.e., if we consider relations D f = {(p, f (p))}
for arbitrary computable partial functions f as description modes. Equivalently, we may say
that we consider (computably) enumerable relations D that are graphs of functions (for every p
there exists at most one x such that (p, x) ∈ D; each description describes at most one object).
Then the Kolmogorov–Solomonoff optimality theorem says that there exists an optimal D in
this class that makes CD minimal (up to an O(1) additive term). (We assume that the reader
is familiar with basic properties of Kolmogorov complexity, see, e.g., [11, 18]; for a short
introduction see also [17].)
Note that we could get a trivial CD if we take, e.g., the set of all pairs as a description mode
D (in this case all strings have complexity zero, since the empty string describes all of them).
So we should be careful and do not consider description modes where the same string describes
too many objects.
2
To define our class of descriptions, let us first recall some basic notions related to finite
automata. Let A and B be two finite alphabets. Consider a directed graph G whose edges are
labeled by pairs (a, b) of letters (from A and B respectively). We also allow pairs of the form
(a, ε ), (ε , b), and (ε , ε ) where ε is a special symbol (not in A or B) that informally means “no
letter”. For such a graph, consider all directed paths in it (no restriction on starting or final
points), and for each path concatenate all the first and all the second components of the pairs; ε
is replaced by an empty word. For each path we get some pair (u, v) where u ∈ A∗ and v ∈ B∗
(i.e., u and v are words over alphabets A and B). Consider all pairs that can be read in this
way along all paths in G. For each labeled graph G we obtain a relation (set of pairs) RG that
is a subset of A∗ × B∗ . For the purposes of this paper, we call the relations obtained in this
way “automatic”. This notion is similar to rational relations defined by transducers [5, Section
III.6]. The difference is that we do not fix initial/finite states (so every subpath of a valid path is
also valid) and that we do not allow arbitrary words as labels, only letters and ε . (This will be
important, e.g., for the statement (j) of Theorem 1.)
Definition 2. A relation R ⊂ A∗ × B∗ is automatic if there exists a labeled graph (automaton) G
such that R = RG .
Now we define automatic description modes as automatic relations where each string describes at most O(1) objects:
Definition 3. A relation D ⊂ B∗ × B∗ is an automatic description mode if
• D is automatic in the sense of Definition 2;
• D is a graph of an O(1)-valued function: there exists some constant c such that for each
p there are at most c values of x such that (p, x) ∈ D.
For every automatic description mode D we consider the corresponding complexity function CD . There is no optimal mode D that makes CD minimal (see Theorem 1 below). So,
stating some properties of complexity, we need to mention D explicitly. Moreover, for a statement that compares the complexities of different strings, we need to say something like “for
every automatic description mode D there exists another automatic description mode D′ such
that. . . ”, and then make a statement that involves both CD and CD′ . (A similar approach is
needed when we try to adapt inequalities for Kolmogorov complexity to the case of resourcebounded complexities.)
Let us first mention some basic properties of automatic description modes.
Proposition 1.
(a) The union of two automatic description modes is an automatic description mode.
(b) The composition of two automatic description modes is an automatic description mode.
(c) If D is a description mode, then {(p, x0) : (p, x) ∈ D} is a description mode (here x0 is the
binary string x with 0 appended); the same is true for x1 instead of x0.
Proof. There are two requirements for an automatic description mode: (1) the relation is automatic and (2) the number of images is bounded. The second one is obvious in all three cases.
The first one can be proven by a standard argument (see, e.g., [5, Theorem 4.4]) that we reproduce for completeness.
(a) The union of two relations RG and R′G for two automata G and G′ corresponds to an
automaton that is a disjoint union of G and G′ .
3
(b) Let S and T be automatic relations that correspond to automata K and L. Consider a new
graph that has set of vertices K × L. (Here we denote an automaton and the set of vertices of its
underlying graph by the same letter.)
• If an edge k → k′ with a label (a, ε ) exists in K, then the new graph has edges (k, l) →
(k′ , l) for all l ∈ L; all these edges have the same label (a, ε ).
• In the same way an edge l → l ′ with a label (ε , c) in L causes edges (k, l) → (k, l ′ ) in the
new graph for all k; all these edges have the same label (ε , c).
• Finally, if K has an edge k → k′ labeled (a, b) and at the same time L has an edge l → l ′
labeled (b, c), where b is the same letter, then we add an edge (k, l) → (k′ , l ′ ) labeled (a, c)
in the new graph.
Any path in the new graph is projected into two paths in K and L. Let (p, q) and (u, v) be the
pairs of words that can be read along these projected paths in K and L respectively, so (p, q) ∈ S
and (u, v) ∈ T . The construction of the graph K × L guarantees that q = u and that we read (p, v)
in the new graph along the path. So every pair (p, v) of strings that can be read in the new graph
belongs to the composition of S and T .
On the other hand, assume that (p, v) belong to the composition, i.e., there exists q such that
(p, q) can be read along some path in K and (q, v) can be read along some path in L. Then the
same word q appears in the second components in the first path and in the first components in
the second path. If we align the two paths in such a way that the letters of q appear at the same
time, we get a valid transition of the third type for each letter of q. Then we complete the path
by adding transitions inbetween the synchronized ones (interleaving them in arbitrary way); all
these transitions exist in the new graph by construction.
(c) We add an additional outgoing edge labeled (ε , 0) for each vertex of the graph; all these
edges go to a special vertex that has no outgoing edges.
Remark. Given a graph, one can check in polynomial time whether the corresponding relation
is O(1)-valued [21, Theorem 5.3, p. 777].
Now we are ready to prove the following simple result about the properties of automatic
Kolmogorov complexity functions, i.e., of functions CR where R is some automatic description
mode.
Theorem 1 (Basic properties of automatic Kolmogorov complexity).
(a) There exists an automatic description mode R such that CR (x) ≤ |x| for all strings x.
(b) For every automatic description mode R there exists some automatic description mode R′
such that CR′ (x0) ≤ CR (x) and CR′ (x1) ≤ CR (x) for all x.
(c) For every automatic description mode R there exists some automatic description mode R′
such that CR′ (x̄) ≤ CR (x), where x̄ stands for the reversed x.
(d) For every automatic description mode R there exists some constant c such that C(x) ≤
CR (x) + c. (Here C stands for the plain Kolmogorov complexity.)
(e) For every c > 0 there exists an automatic description mode R such that
CR (1n ) ≤ n/c for all n.
(f) For every automatic description mode R there exists some c > 0 such that CR (1n ) ≥ n/c − 1
for all n.
4
(g) For every two automatic description modes R1 and R2 there exists an automatic description
mode R such that CR (x) ≤ CR1 (x) and CR (x) ≤ CR2 (x) for all x.
(h) There is no optimal automatic description mode. (A mode R is called optimal in some class
if for every mode R′ in this class there exists some c such that CR (x) ≤ CR′ (x) + c for all
strings x.)
(i) For every automatic description mode R, if x′ is a substring of x, then CR (x′ ) ≤ CR (x).
(j) Moreover, CR (xy) ≥ CR (x) + CR (y) for every two strings x and y.
(k) For every automatic description mode R and for every constant ε > 0 there exists an automatic description mode R′ such that CR′ (xy) ≤ (1 + ε ) CR (x) + CR (y) for all strings x
and y.
(l) Let S be an automatic description mode. Then for every automatic description mode R there
exists an automatic description mode R′ such that CR′ (y) ≤ CR (x) for every (x, y) ∈ S.
(m) If we allow a bigger alphabet B instead of B as an alphabet for descriptions, we divide the
complexity by log |B|, up to a constant factor that can be chosen arbitrarily close to 1.
Proof. (a) Consider an identity relation as a description mode; it corresponds to an automaton
with one state.
(b) This is a direct corollary of Proposition 1, (c).
(c) The definition of an automaton is symmetric (all edges can be reversed), and the O(1)condition still holds.
(d) Let R be an automatic description mode. An automaton defines a decidable (computable)
relation, so R is decidable. Since R defines a O(1)-valued function, a Kolmogorov description
of some y that consists of its R-description x and the ordinal number of y among all strings that
are in R-relation to x, is only O(1) bits longer than x.
(e) Consider an automaton that consists of a cycle where it reads one input symbol 1 and
then produces c output symbols 1. (Since we consider the relation as an O(1)-multivalued function, we sometimes consider the first components of pairs as “input symbols” and the second
components as “output symbols”.) Recall that there is no restrictions on initial and finite states,
so this automaton produces all pairs (1k , 1l ) where (k − 1)c ≤ l ≤ (k + 1)c.
(f) Consider an arbitrary description mode, i.e., an automaton that defines some O(1)-valued
relation. Then every cycle in the automaton that produces some output letter should also produce some input letter, otherwise an empty input string corresponds to infinitely many output
strings. For any sufficiently long path in the graph we can cut away a minimal cycle, removing
at least one input letter and at most c output letters, where c is the number of states, until we get
a path of length less than c.
(g) This follows from Proposition 1, (a).
(h) This statement is a direct consequence of (e) and (f). Note that for finitely many automatic description modes there is a mode that is better than all of them, as (g) shows, but we
cannot do the same for all description modes (as was the case for Kolmogorov complexity).
(i) If R is a description mode, (p, x) belongs to R and x′ is a substring of x, then there exists
some substring p′ of p such that (p′ , x′ ) ∈ R. Indeed, we may consider the input symbols used
while producing x′ .
(j) Note that in the previous argument we can choose disjoint p′ for disjoint x′ .
(k) Informally, we modify the description mode as follows: a fixed fraction of input symbols
is used to indicate when a description of x ends and a description of y begins. More formally,
5
let R be an automatic description mode; we use the same notation R for the corresponding
automaton. Consider N + 1 copies of R (called 0-, 1-,. . . , Nth layers). The outgoing edges from
the vertices of ith layer that contain an input symbol are redirected to (i + 1)th layer (the new
state remains the same, only the layer changes, so the layer number counts the input length).
The edges with no input symbol are left unchanged (and go to ith layer as before). The edges
from the Nth layer are of two types: for each vertex x there is an edge with label (0, ε ) that
goes to the same vertex in 0th layer, and edges with labels (1, ε ) that connect each vertex of
Nth layer to all vertices of an additional copy of R (so we have N + 2 copies in total). If both x
and y can be read (as outputs) along the edges of R, then xy can be read, too (additional zeros
should be added to the input string after groups of N input symbols). We switch from x to y
using the edge that goes from Nth layer to the additional copy of R (using additional symbol 1
in the input string). The overhead in the description is one symbol per every N input symbols
used to describe x. We get the required bound, since N can be arbitrarily large.
The only thing to check is that the new automaton is O(1)-valued. Indeed, the possible
switch position (when we move to the states of the additional copy of R) is determined by the
positions of the auxiliary bits modulo N + 1: when this position modulo N + 1 is fixed, we look
for the first 1 among the auxiliary bits. This gives only a bounded factor (N + 1) for the number
of possible outputs that correspond to a given input.
(l) The composition S ◦ R is an automatic description mode due to Proposition 1.
(m) Take the composition of a given description mode R with a mode that provides block
encoding of inputs. Note that block encoding can be implemented by an automaton. There is
some overhead when |B| is not a power of 2, but the corresponding factor becomes arbitrarily
close to 1 if we use block code with large block size.
Remark. Not all these results are used in the sequel; we provide them for comparison with the
properties of the standard Kolmogorov complexity function.
3 Normal sequences and numbers
Consider an infinite bit sequence α = a0 a1 a2 . . . and some integer k ≥ 1. Split the sequence α
into k-bit blocks: α = A0 A1 . . .. For every k-bit string r consider the limit frequence of r among
the Ai , i.e. the limit of #{i : i < N and Ai = r}/N as N → ∞. This limit may exist or not; if it
exists for some k and for all r, we get a probability distribution on k-bit strings.
Definition 4. A sequence α is normal if for every number k and every string r of length k this
limit exists and is equal to 2−k .
Sometimes sequences with these properties are called strongly normal while the name “normal” is reserved for sequences that have this property for k = 1.
There is a version of the definition of normal sequences that considers all occurences of
some string r in α (while Definition 4 considers only aligned ones, whose starting point is a multiple of k). In this “non-aligned” version we require that the limit of #{i < N : αi αi+1 . . . αi+k−1 =
r}/N equals 2−k for all k and for all strings r of length k. A classical result (see, e.g., [12, Chapter 1, Section 8]) says that this is an equivalent notion, and we give below a simple proof of this
equivalence using automatic complexity. Before this proof is given, we will distinguish the two
definitions by using the name “non-aligned-normal” for the second version.
A real number is called normal if its binary expansion is normal (we ignore the integer
part). If a number has two binary expansions, like 0.0111 . . . = 0.1000 . . ., both expansions are
not normal, so this is not a problem.
6
A classical example of a normal number is the Champernowne number [7]
0.0 1 10 11 100 101 110 111 1000 1001 . . .
(the concatenation of all positive integers in binary). Let us sketch the proof of its normality (not
used in the sequel) using the non-aligned version of normality definition. All N-bit numbers in
the Champernowne sequence form a block that starts with 10N−1 and ends with 1N . Note that
every string of length k ≪ N appears in this block with probability close to 2−k , since each of
2N−1 strings (after the leading 1 for the N-bit numbers in the Champernowne sequence) appears
exactly once. The deviation is caused by the leading 1’s and also by the boundaries between the
consecutive N-bit numbers where the k-bit substrings are out of control. Still the deviation is
small since k ≪ N.
This is not enough to conclude that C is (non-aligned) normal, since the definition speaks
about frequencies in all prefixes; the prefixes that end on a boundary between two blocks are
not enough. The problem appears because the size of a block is comparable to the length of the
prefix before it. To deal with arbitrary prefixes, let us note that if we ignore two leading digits
in each number (first 10 and then 11) instead of one, the rest is periodic in the block (the block
consists of two periods). If we ignore three leading digits, the block consists of four periods,
etc. An arbitrary prefix is then close to the boundary between these sub-blocks, and the distance
can be made small compared to the total length of the prefix. (End of the proof sketch.)
The definition of normality can be given for an arbitrary alphabet (instead of the binary
one), and we get the notion of b-normality of a real number for every base b ≥ 2. It is known
that for different bases we get non-equivalent notions (a rather difficult result). The numbers
in [0, 1] that are normal for every base are called absolutely normal. Their existence can be
proved by a probabilistic argument. Indeed, for every base b, almost all reals are b-normal (the
non-normal numbers have Lebesgue measure 0); this is guaranteed by the Strong Law of Large
Numbers. Therefore the numbers that are not absolutely normal form a null set (a countable
union of the null sets for each b). The constructive version of this argument shows that there
exist computable absolutely normal numbers. This result goes back to an unpublished note of
Turing (1938, see [2]).
In the next section we prove the connection between normality and automatic complexity: a
sequence α is normal if for every automatic description mode D the corresponding complexity
CD of its prefix never becomes much smaller than the length of this prefix.
4 Normality and incompressibility
Theorem 2. A sequence α = a0 a1 a2 . . . is normal if and only if
lim inf
n→∞
CR (a0 a1 . . . an−1 )
≥1
n
for every automatic description mode R.
Proof. First, let us show that a sequence that is not normal is compressible. Assume that for
some bit sequence α and for some k the requirement for aligned k-bit blocks is not satisfied.
Using a compactness argument, we can find a sequence of lengths Ni such that for the prefixes
of these lengths the frequencies of k-bit blocks do converge to some probability distribution A
on Bk , but this distribution is not uniform. Then its Shannon entropy H(A) is less than k.
The Shannon theorem can then be used to construct a block code of average length close
to H(A), namely, of length at most H(A) + 1 (this “+1” overhead is due to rounding if the
7
frequencies are not powers of 2). Since this code can be easily converted into an automatic
description mode, it will give the desired result if H(A) < k − 1. It remains to show that it is the
case for long enough blocks.
Selecting a subsequence, we may assume without loss of generality that the limit frequencies
exist also for (aligned) 2k-bit blocks, so we get a random variable A1 A2 whose values are 2kbit blocks (and A1 and A2 are their first and second halves of length k). The variables A1
and A2 may be dependent, and their distributions may differ from the initial distribution A for
k-bit blocks. Still we know that A is the average of A1 and A2 (since A is computed for all
blocks, and A1 [resp. A2 ] corresponds to odd [resp. even] blocks). A convexity argument (the
function p 7→ −p log p used in the definition of entropy has negative second derivative) shows
that H(A) ≥ [H(A1) + H(A2 )]/2. Then
H(A1 A2 ) ≤ H(A1 ) + H(A2 ) ≤ 2H(A),
so A1 A2 has twice bigger difference between entropy and length (at least). Repeating this argument, we can find k such that the difference between length and entropy is greater than 1. This
finishes the proof in one direction.
Now we need to prove that normal sequence α is incompressible. Let R be an arbitrary
automatic description mode. Consider some k and split the sequence into k-bit blocks α =
A0 A1 A2 . . .. (Now Ai are just the blocks in α , not random variables). We will show that
lim infCR (A0 A1 . . . An−1 )/nk
cannot be much smaller than 1. More precisely, we will show that
lim inf
O(1)
CR (A0 A1 . . . An−1 )
≥ 1−
,
nk
k
where the constant in O(1) does not depend on k. This is enough, because (i) adding the last
incomplete block can only increase the complexity and the change in length is negligible, and
(ii) the value of k may be arbitrarily large.
Now let us prove this bound for some fixed k. Recall that
CR (A0 A1 . . . An−1 ) ≥ CR (A0 ) + CR (A1 ) + . . . + CR (An−1 )
and that C(x) ≤ CR (x) + O(1) for all x and some O(1)-constant that depends only on R (Theorem 1). By assumption, all k-bit strings appear with the same limit frequency among A0 , A1 ,. . . ,
An−1 . It remains to note that the average Kolmogorov complexity C(x) of all k-bit strings
is k − O(1); indeed, the fraction of k-bit strings that can be compressed by more than d bits
(C(x) < k − d) is at most 2−d , and the series ∑ d2−d (the upper bound for the average number
of bits saved by compression) has finite sum.
A small modification of this proof adapts it to the non-aligned definition of normality. Let
α be a sequence that is not normal in the non-aligned version. This means that for some k the
k-bit blocks do not have a correct limit distribution (non-aligned). These blocks can be split
into k groups according to their starting positions modulo k. In one of the groups blocks do
not have a correct limit distribution (otherwise the average distribution would be correct, too).
So we can delete some prefix (less than k symbols) of our sequence and get a sequence that is
not normal in the aligned sense. Its prefixes are compressible (as we have seen). The same is
true for the original sequence since adding a fixed finite prefix (or suffix) changes complexity
at most by O(1).
8
In the other direction: let us assume that the sequence is normal in the non-aligned sense.
The aligned frequency of some compressible-by-d-bits block (as well as any other block) can
be only k times bigger than its non-aligned frequency, which is exponentially small in d (the
number of saved bits), so we can choose the parameters to get the required bound.
Indeed, let us consider blocks of length k whose CR -complexity is smaller than k − d. Their
Kolmogorov complexity is then smaller than k − d + O(1), and the fraction of these blocks
(among all k-bit strings) is at most 2−d+O(1) . So their frequency among aligned blocks is at
most 2−d+O(1) · k. For all other blocks R-compression saves at most d bits, and for compressible
blocks it saves at most k bits, so the average number of saved bits (per k-bit block) is bounded
by
d + k2−d+O(1) · k = d + O(k2 2−d ).
We need this bound to be o(k), i.e., we need that
d
+ O(k2−d ) = o(1)
k
as k → ∞. This can be achieved, for example, if d = 2 log k.
In this way we get the following corollary:
Corollary. The aligned and non-aligned definitions of normality are equivalent.
Note also that adding/deleting a finite prefix does not change the compressibility, and, therefore, normality. (For the non-aligned version of the normality definition it is obvious anyway,
but for the aligned version it is not so easy to see directly.)
Another corollary is a result proven by Piatetski-Shapiro in [14]: if for some c and for all k
every k-bits block appears in a sequence with lim sup-frequency at most c2−k , then the sequence
is normal. Indeed, in the argument above we had a constant factor in O(k2−d ) anyway. (We can
even allow the constant c to depend on k if its growth as a function of k is not too fast.)
5 Wall’s theorem
Now we obtain a known result about normal numbers (Wall’s theorem) as an easy corollary.
Recall that a real number is normal if its binary expansion is normal. We agreed to ignore
the integer part (since it has only finitely many digits, adding it as a prefix would not matter
anyway).
Theorem 3 (Wall [20]). If p and q are rational numbers and α is normal, then α p+q is normal.
Proof. It is enough to show that multiplication and division by an integer c preserve normality (note that adding an integer preserves it by definition, since the integer part is ignored).
This fact follows from the incompressibility characterization (Theorem 2), the non-increase of
complexity under automatic O(1) mappings (Theorem 1, (l)) and the following lemma:
Lemma. Let c be an integer. Consider the relation Rc that consists of pairs of strings x and
y such that x and y have the same length and can be prefixes of the binary expansions of the
fractional parts of γ and cγ for some real γ . This relation, as well as its inverse, is contained in
an automatic description mode.
Assuming Lemma 5, we conclude that the prefixes of γ and cγ have the same automatic complexity. More precisely, for every automatic description mode R there exists another automatic
description mode R′ such that CR′ (y) ≤ CR (x) if x and y are prefixes of γ and cγ respectively.
9
This follows from Theorem 1, (l). So if γ is compressible, then cγ is also compressible. The
same is true if we consider the inverse relation; if cγ is compressible, then γ is also compressible.
It remains to prove Lemma 5. Indeed, the school division algorithm can be represented by
an automaton; integer parts can be different, but this creates O(1) different possible remainders. We have to take care of two representations of the same number (note that while dividing
0.29999 . . . by 3, we obtain only 0.09999 . . ., not 0.10000 . . .), but at most two representations
are possible and the relation between them is automatic, so we still get an automatic description
mode.
6 Pairs as descriptions and Agafonov’s theorem
In this section we derive another classical result about normal numbers, Agafonov’s theorem [1],
from the incompressibility characterization. (However, we will need to modify this characterization, see below).
Agafonov’s result is motivated by the von Mises’ approach to randomness (see, e.g., [18,
Chapter 9] for a historic account). As von Mises mentioned, a “random sequence” (he used the
German term “Kollektiv”) should remain random after using a reasonable selection rule. More
precisely, assume that there is some set S of binary strings. This set determines a “selection rule”
that selects a subsequence from every binary sequence α . The selection works as follows: we
observe a binary sequence α = a0 a1 a2 . . . and select terms an such that a0 a1 . . . an−1 ∈ S (without
reordering the selected terms). We get a subsequence; if an initial sequence is “random” (is
plausible as an outcome of a fair coin tossing), said von Mises, this subsequence should also be
random in the same sense. The Agafonov’s theorem says that for regular (automatic) selection
rules and normality as randomness this property is indeed true.
Theorem 4 (Agafonov). Let α = a0 a1 a2 . . . be a normal sequence. Let S be a regular (=recognizable by a finite automaton) set of binary strings. Consider a subsequence σ made of terms
an such that a0 a1 . . . an−1 ∈ S (in the same order as in the original sequence). Then σ is normal
or finite.
Proof. This proof adapts the arguments from [4] to our definition of compressibility. Using the
incompressibility criterion, we need to prove that if the sequence σ is compressible, then α is
compressible, too. The idea is simple: the selection rule splits α into two subsequences: the
selected terms (σ ) and the rest (the non-selected terms). We do not know anything about the
second subsequence, but we assume that σ is compressible, and want to prove that the entire
sequence α is compressible.
The key observation: knowing both subsequences (and the selection rule S, of course), we
can reconstruct the original sequence. Indeed, we know whether a0 is selected or not: it depends
on whether the empty string belongs to S or not. So we know where we should look for the first
term when reconstructing α from its two parts. Knowing a0 , we can check whether one-letter
word a0 belongs to S or not. So we know whether a1 is selected, so we know from which
subsequence it should be taken, etc.
So, we know that our sequence can be reconstructed from two its parts, and one part is
compressible. Then the entire sequence is compressible: a compressed description consist of a
compressed description of a compressible subsequence, and the trivial description of the other
one (for which we do not know whether it is compressible or not). To make this argument
precise, we need two things:
• prove that the selected subsequence has positive density (otherwise its compression gives
only a negligible improvement);
10
• modify the definition of complexity and the criterion of compressibility allowing pairs as
descriptions.
We start with the first part.
Lemma. If the selected subsequence is infinite, then it has a positive density, i.e., the lim inf of
the density of the selected terms is positive.
Proof of the lemma. Consider a deterministic finite automaton that recognizes the set S. We
denote this automaton by the same letter S. Let X be the set of states of S that appear infinitely
many times when S is applied to α . Starting from some time, the automaton is in X , and X
is strongly connected (when speaking about strong connectivity, we ignore the labeling of the
transition edges). Let us show that vertices in X have no outgoing edges that leave X . If these
edges exist, let us construct a string u that forces S to leave X when started from any vertex of
X . This will lead to a contradiction: a normal sequence has infinitely many occurences of u,
and one of them appears when S already is in X . How to construct this u? Take some q ∈ X
and construct a string u1 that forces S to leave X when started from q. Such a string u1 exists,
since X is strongly connected, so we can bring S to any vertex and then use the letter that forces
S to leave X . Now consider some other vertex q′ ∈ X . It may happen that u1 already forces S to
leave X when started from q′ . If not and S remains in X (being in some vertex v), we can find
some string u2 that forces S out of X when started at v. Then the string u1 u2 forces S to leave
X when started in any of the vertices q, q′ . Then we consider some other vertex q′′ and append
(if needed) some string u3 that forces S to leave X when started at q, q′ or q′′ (in the same way).
Doing this for all vertices of X , we get the desired u (and the desired contradiction).
So X has no outgoing edges (and therefore is a strongly connected component of S’s graph).
Now the same argument shows that there exists a string u that forces S to visit all vertices of X
when started from any vertex in X . This string u appears with positive density in α . So either
the selected subsequence is finite (if X has no accepting vertices) or the selected subsequence
has positive density (since every occurence of u means that at least one term is selected when S
visits the accepting vertex). Lemma is proven.
To finish the proof, we need to modify the notion of a description mode and consider pairs as
descriptions. Let A, B,C be three alphabets. We define the notion of automatic ternary relation
R ⊂ A∗ × B∗ × C∗ in the same way as for binary relations: now the edge labels are triples
(a, b, c), where each of the letters (or even all three) can be replaced by ε -symbol. This relation
can be considered as multivalued function of type A × B → C. If it is O(1)-valued, we call it
pair description mode, and a pair (u, v) is called a description of w if (u, v, w) ∈ R. We assume,
as before, that all the alphabets are binary (A = B = C = B), and the length of description is
measured as the sum of lengths:
CR (w) = min{|u| + |v| : (u, v, w) ∈ R}.
The automatic description modes are special cases of pair description modes (if we use only one
component of the pair as a description, and the other one is empty), but this generalization may
lead to smaller complexity function. (It would be interesting to find out how much the decrease
could be.) Still this modified version of automatic complexity can be used to characterize
normality:
Proposition 2. (a) For every pair description mode
C(x) ≤ CR (x) + c log CR (x) + c
for some c and all x, where C(x) stands for the Kolmogorov complexity of x.
11
(b) If R is a pair description mode and α = a0 a1 . . . is a normal sequence, then
lim
CR (a0 a1 . . . an−1 )
≥ 1.
n
(c) If a ternary relation R(u, v, w) is a pair description mode and a binary relation Q(u′ , u) is
an automatic description mode, then their joint
R′ (u′ , v, w) = ∃u [Q(u′ , u) and R(u, v, w)]
is a pair description mode.
(d) Let S be a regular set of binary strings (recognized by a finite automaton) used as a selection
rule. Then the relation
{(u, v, w) : u and v are strings of selected and non-selected bits when S is applied to w}
is a pair description mode.
Proof of the Proposition. (a) Fix some pair description mode R. If (u, v, w) ∈ R, the string w
is determined by the pair (u, v) and the ordinal number of x among the outputs of O(1)-valued
function for input (u, v). So the Kolmogorov complexity of w exceeds the Kolmogorov complexity of a pair (u, v) at most by O(1), and complexity of a pair (u, v) is bounded by l +O(log l)
where l is the total length of u and v.
(b) As before, we cut α into blocks of some large length k. Now the R-complexity of a block
can be smaller than its Kolmogorov complexity, and the decrease can be O(log k), but this does
not matter: for large k this change is negligible compared to k.
(c) The joint of two automatic relations is automatic, for the same reasons; the corresponding
function is O(1)-valued since for each values of u′ we have O(1) different values of u, and each
of them leads to O(1) values of w (for a fixed v).
(d) The process of splitting w into u and v is automatic for obvious reasons. The notion
of automatic relation does not distinguish between input and output, so this ternary relation is
automatic. As we have discussed, for a given u and v there exists at most one w that can be
obtained by merging u and v; to determine whether we take the next letter from u or v, we check
whether the string of symbols already added to w belongs to the selection rule S. (Now the
initial state is not fixed anymore, still we can at most O(1) values for given u and v.)
Now we can finish the proof of Agafonov’s theorem. Assume that some selection rule S is
applied to a normal sequence α and selects some its subsequence σ that is not normal. After
finitely many steps S splits a prefix a of α into sequence of selected terms s (it is a prefix of σ )
and the sequence u of non-selected terms. Then (s, u, a) belongs to the pair description mode
from part (d) of the proposition; let us denote it by U . Now recall that σ is not normal, so
Theorem 2 says that there exists some description mode Q such that CQ (s) < (1 − ε )|s| for
some ε > 0 and for infinitely many prefixes s of σ . Then use part (c) of the proposition and
consider the joint J of Q and U . The
CJ (a) ≤ CQ (s) + |u| ≤ (1 − ε )|s| + |u| ≤ |s| + |u| − ε |s|.
for infinitely many prefixes a of α that correspond to compressible prefixes s of σ . Lemma
guarantees that ε |s| is Ω(|a|), so we use Theorem 2 in the other direction and get a contradiction
with the normality of a.
12
7 Discussion
The connection between normality and finite-state computations was noticed long ago, as the
title of [1] shows; see also [15] where normality was related to martingales arising from finite
automata. This connection led to a characterization of normality as incompressibility (see [4]
for a direct proof). On the other hand, it was also clear that the notion of Kolmogorov complexity is not directly practical since it considers arbitrary algorithms as decompressors, and this
makes it non-computable. So restricted classes of decompressors are of interest, and finite-state
computations are a natural candidate for such a class.
Shallit and Wang [16] suggested to consider, for a given string x, the minimal number of
states in an automaton that accepts x but not other strings of the same length. Later Hyde and
Kjos-Hanssen [10] considered a similar notion using nondeterministic automata. The intrinsic
problem of this approach is that it is not naturally “calibrated” in the following sense: measuring
the information in bits, we would like to have about 2n objects of complexity at most n.
Another (and “calibrated”) approach was suggested by Calude, Salomaa and Roblot [8]: in
their definition a deterministic transducer maps a description string to a string to be described,
and the complexity of y is measured as the combination of the sizes of a transducer and an input
string needed to produce y (the minimum over all transducers and all input strings producing y
is taken). The size of the transducer is measured via some encoding, so the complexity function
depends on the choice of this encoding. The open question posed in [8, Section 6] asks whether
this notion of complexity can be used to characterize normality.
The incompressibility notion used in [4] provides such a characterization for a different definition. It uses deterministic transducers and requires additionally that for every output string
y and every final state s there exists at most one input string that produces y and brings the automaton into the state s. Our approach is a refinement of this one: we consider non-deterministic
automata without initial/final states and require only that decompressor is an O(1)-valued function. The proofs then become simpler, mainly for two reasons: (1) we use the comparison of the
automatic complexity and the plain Kolmogorov complexity and apply standard results about
Kolmogorov complexity; (2) we explicitly state and prove the property CR (xy) ≥ CR (x) +CR (y)
that is crucial for the proofs.
Acknowledgement
I am grateful to Veronica Becher, Olivier Carton and Paul Heiber for many discussions of their
paper [3] and the relations between incompressibility and normality, and for the permission
to use the observations made during these discussions in the current paper. I am also grateful
to my colleagues in LIRMM (ESCAPE team) and Moscow (Kolmogorov seminar, Computer
Science Department of the HSE). I am grateful to the anonymous reviewers of STACS and FCS
conferences; many of their remarks are taken into account in the current version.
13
References
[1] V.N. Agafonov. Normal sequences and finite automata, Doklady AN SSSR, 179, 255–256
(1968). See also the paper of V.N. Agafonov with the same name in the collection: Problemy
Kibernetiki (Cybernetics problems). Volume 20. Moscow: Nauka, 1968, p. 123–129.
[2] V. Becher, Turing’s normal numbers: towards randomness, How the World Computes, Proceedings of the Turing Centenary Conference and the 8th Conference in Computability in
Europe, CiE2012, Cambridge, UK, June 18–23, 2012. Lecture notes in computer science,
7318, 35–45. Springer-Verlag, 2012.
[3] V. Becher, O. Carton, P. Heiber, Finite-state independence, 12 November 2016,
https://arxiv.org/pdf/1611.03921.pdf
[4] V. Becher, P. Heiber, Normal number and finite automata, Theoretical Computer Science,
477, 109–116 (2013).
[5] J. Berstel, Transductions and Context-Free Languages Vieweg+Teubner Verlag, 1969.
ISBN: 978-3-519-02340-1. (For a revised 2006–2009 version see the author’s homepage,
http://www-igm.univ-mlv.fr/~berstel)
[6] Y. Bugeaud, Distribution modulo one and Diophantine approximation, Cambridge Tracts
in Mathematics, 193, Cambridge University Press, 2012.
[7] D. Champernowne, The construction of decimals normal in the scale of ten, Journal of the
London Mathematical Society, volume s1-8, issue 4 (October 1933; Received 19 April, read
27 April,1933), 254–260.
[8] C.S. Calude, K. Salomaa, T.K. Roblot, Finite state complexity, Theoretical Computer Science, 412, 5668–5677 (2011).
[9] R.G. Downey, D.R. Hirschfeldt, Algorithimic randomness and complexity, Springer, 2010,
ISBN 978-0-387-68441-3, xxviii+855 p.
[10] K.K. Hyde, B. Kjos-Hanssen, Nondeterministic complexity of overlap-free and almost
square-free words, The Electronic Journal of Combinatorics, 22:3, 2015.
[11] M. Li, P. Vitányi, An Introduction to Kolmogorov complexity and its applications, 3rd ed.,
Springer, 2008 (1 ed., 1993; 2 ed., 1997), 792 pp. ISBN 978-0-387-49820-1.
[12] L. Kuipers, H. Niederreiter, Uniform distribution of sequences. John Wiley & Sons, 1974.
[13] A. Nies, Computability and randomness, Oxford Logic Guides, Oxford University Press,
2009, ISBN 978-0199652600, 435 p.
[14] I.I. Piatetski-Shapiro, On the laws of distribution of the fractional parts of an exponential
function, Izvestia Akademii Nauk SSSR, Ser. Matem., 15(1), 47–52 (1951). In Russian.
[15] C. Schnorr, H. Stimm, Endliche Automaten und Zufallsfolgen, Acta Informatica, 1(4),
345–39 (1972).
[16] J. Shallit, M.-W. Wang, Automatic complexity of strings, Journal of Automata, Languages
and Combinatorics, 6:4. 537-554 (April 2001)
14
[17] A. Shen, Around Kolmogorov complexity: basic notions and results. Measures of Complexity. Festschrift for Alexey Chervonenkis, edited by V. Vovk, H. Papadoupoulos, A. Gammerman, Springer, 2015, p. 75–116, see also http://arxiv.org/abs/1504.04955
(2015)
[18] A. Shen, V.A. Uspensky, N. Vereshchagin, Kolmogorov complexity and algorithmic randomness, Moscow, MCCME, 2013 (In Russian). English version accepted for publication
by AMS, see www.lirmm.fr/~ ashen/kolmbook-eng.pdf
[19] V.A. Uspensky, A. Shen, Relations between varieties of Kolmogorov complexities, Mathematical Systems Theory, 29, 271–292 (1996).
[20] D.D. Wall, Normal numbers, Thesis, University of California, 1949.
[21] A. Weber, On the valuedness of finite transducers, Acta Informatica, 27(8), 749–780
(1990).
15
| 7cs.IT |
Published as a conference paper at ICLR 2017
D ENSITY ESTIMATION USING R EAL NVP
Laurent Dinh∗
Montreal Institute for Learning Algorithms
University of Montreal
Montreal, QC H3T1J4
arXiv:1605.08803v3 [cs.LG] 27 Feb 2017
Jascha Sohl-Dickstein
Google Brain
Samy Bengio
Google Brain
A BSTRACT
Unsupervised learning of probabilistic models is a central yet challenging problem
in machine learning. Specifically, designing models with tractable learning, sampling, inference and evaluation is crucial in solving this task. We extend the space
of such models using real-valued non-volume preserving (real NVP) transformations, a set of powerful, stably invertible, and learnable transformations, resulting
in an unsupervised learning algorithm with exact log-likelihood computation, exact
and efficient sampling, exact and efficient inference of latent variables, and an
interpretable latent space. We demonstrate its ability to model natural images
on four datasets through sampling, log-likelihood evaluation, and latent variable
manipulations.
1
Introduction
The domain of representation learning has undergone tremendous advances due to improved supervised learning techniques. However, unsupervised learning has the potential to leverage large pools of
unlabeled data, and extend these advances to modalities that are otherwise impractical or impossible.
One principled approach to unsupervised learning is generative probabilistic modeling. Not only do
generative probabilistic models have the ability to create novel content, they also have a wide range
of reconstruction related applications including inpainting [61, 46, 59], denoising [3], colorization
[71], and super-resolution [9].
As data of interest are generally high-dimensional and highly structured, the challenge in this domain
is building models that are powerful enough to capture its complexity yet still trainable. We address
this challenge by introducing real-valued non-volume preserving (real NVP) transformations, a
tractable yet expressive approach to modeling high-dimensional data.
This model can perform efficient and exact inference, sampling and log-density estimation of data
points. Moreover, the architecture presented in this paper enables exact and efficient reconstruction
of input images from the hierarchical features extracted by this model.
2
Related work
Substantial work on probabilistic generative models has focused on training models using maximum
likelihood. One class of maximum likelihood models are those described by probabilistic undirected
graphs, such as Restricted Boltzmann Machines [58] and Deep Boltzmann Machines [53]. These
models are trained by taking advantage of the conditional independence property of their bipartite
structure to allow efficient exact or approximate posterior inference on latent variables. However,
because of the intractability of the associated marginal distribution over latent variables, their
training, evaluation, and sampling procedures necessitate the use of approximations like Mean
Field inference and Markov Chain Monte Carlo, whose convergence time for such complex models
∗
Work was done when author was at Google Brain.
1
Published as a conference paper at ICLR 2017
Data space X
Inference
x ∼ p̂X
z = f (x)
Latent space Z
⇒
Generation
z ∼ pZ
⇐
x = f −1 (z)
Figure 1: Real NVP learns an invertible, stable, mapping between a data distribution p̂X and a latent
distribution pZ (typically a Gaussian). Here we show a mapping that has been learned on a toy
2-d dataset. The function f (x) maps samples x from the data distribution in the upper left into
approximate samples z from the latent distribution, in the upper right. This corresponds to exact
inference of the latent state given the data. The inverse function, f −1 (z), maps samples z from the
latent distribution in the lower right into approximate samples x from the data distribution in the
lower left. This corresponds to exact generation of samples from the model. The transformation of
grid lines in X and Z space is additionally illustrated for both f (x) and f −1 (z).
remains undetermined, often resulting in generation of highly correlated samples. Furthermore, these
approximations can often hinder their performance [7].
Directed graphical models are instead defined in terms of an ancestral sampling procedure, which is
appealing both for its conceptual and computational simplicity. They lack, however, the conditional
independence structure of undirected models, making exact and approximate posterior inference
on latent variables cumbersome [56]. Recent advances in stochastic variational inference [27]
and amortized inference [13, 43, 35, 49], allowed efficient approximate inference and learning of
deep directed graphical models by maximizing a variational lower bound on the log-likelihood [45].
In particular, the variational autoencoder algorithm [35, 49] simultaneously learns a generative
network, that maps gaussian latent variables z to samples x, and a matched approximate inference
network that maps samples x to a semantically meaningful latent representation z, by exploiting the
reparametrization trick [68]. Its success in leveraging recent advances in backpropagation [51, 39] in
deep neural networks resulted in its adoption for several applications ranging from speech synthesis
[12] to language modeling [8]. Still, the approximation in the inference process limits its ability
to learn high dimensional deep representations, motivating recent work in improving approximate
inference [42, 48, 55, 63, 10, 59, 34].
Such approximations can be avoided altogether by abstaining from using latent variables. Autoregressive models [18, 6, 37, 20] can implement this strategy while typically retaining a great deal of
flexibility. This class of algorithms tractably models the joint distribution by decomposing it into a
product of conditionals using the probability chain rule according to a fixed ordering over dimensions,
simplifying log-likelihood evaluation and sampling. Recent work in this line of research has taken
advantage of recent advances in recurrent networks [51], in particular long-short term memory [26],
and residual networks [25, 24] in order to learn state-of-the-art generative image models [61, 46] and
language models [32]. The ordering of the dimensions, although often arbitrary, can be critical to the
training of the model [66]. The sequential nature of this model limits its computational efficiency. For
example, its sampling procedure is sequential and non-parallelizable, which can become cumbersome
in applications like speech and music synthesis, or real-time rendering.. Additionally, there is no
natural latent representation associated with autoregressive models, and they have not yet been shown
to be useful for semi-supervised learning.
2
Published as a conference paper at ICLR 2017
Generative Adversarial Networks (GANs) [21] on the other hand can train any differentiable generative network by avoiding the maximum likelihood principle altogether. Instead, the generative
network is associated with a discriminator network whose task is to distinguish between samples
and real data. Rather than using an intractable log-likelihood, this discriminator network provides
the training signal in an adversarial fashion. Successfully trained GAN models [21, 15, 47] can
consistently generate sharp and realistically looking samples [38]. However, metrics that measure the
diversity in the generated samples are currently intractable [62, 22, 30]. Additionally, instability in
their training process [47] requires careful hyperparameter tuning to avoid diverging behavior.
Training such a generative network g that maps latent variable z ∼ pZ to a sample x ∼ pX does not
in theory require a discriminator network as in GANs, or approximate inference as in variational
autoencoders. Indeed, if g is bijective, it can be trained through maximum likelihood using the change
of variable formula:
−1
∂g(z)
.
(1)
pX (x) = pZ (z) det
∂z T
This formula has been discussed in several papers including the maximum likelihood formulation of
independent components analysis (ICA) [4, 28], gaussianization [14, 11] and deep density models
[5, 50, 17, 3]. As the existence proof of nonlinear ICA solutions [29] suggests, auto-regressive
models can be seen as tractable instance of maximum likelihood nonlinear ICA, where the residual
corresponds to the independent components. However, naive application of the change of variable
formula produces models which are computationally expensive and poorly conditioned, and so large
scale models of this type have not entered general use.
3
Model definition
In this paper, we will tackle the problem of learning highly nonlinear models in high-dimensional
continuous spaces through maximum likelihood. In order to optimize the log-likelihood, we introduce
a more flexible class of architectures that enables the computation of log-likelihood on continuous
data using the change of variable formula. Building on our previous work in [17], we define a
powerful class of bijective functions which enable exact and tractable density evaluation and exact
and tractable inference. Moreover, the resulting cost function does not to rely on a fixed form
reconstruction cost such as square error [38, 47], and generates sharper samples as a result. Also,
this flexibility helps us leverage recent advances in batch normalization [31] and residual networks
[24, 25] to define a very deep multi-scale architecture with multiple levels of abstraction.
3.1
Change of variable formula
Given an observed data variable x ∈ X, a simple prior probability distribution pZ on a latent variable
z ∈ Z, and a bijection f : X → Z (with g = f −1 ), the change of variable formula defines a model
distribution on X by
∂f (x)
(2)
pX (x) = pZ f (x) det
∂xT
∂f (x)
log (pX (x)) = log pZ f (x) + log det
,
(3)
∂xT
where
∂f (x)
∂xT
is the Jacobian of f at x.
Exact samples from the resulting distribution can be generated by using the inverse transform sampling
rule [16]. A sample z ∼ pZ is drawn in the latent space, and its inverse image x = f −1 (z) = g(z)
generates a sample in the original space. Computing the density on a point x is accomplished by
computing
the density of its image f (x) and multiplying by the associated Jacobian determinant
(x)
det ∂f
. See also Figure 1. Exact and efficient inference enables the accurate and fast evaluation
∂xT
of the model.
3
Published as a conference paper at ICLR 2017
y1
y1
y2
y2
t
-
+
=
=
t
s
÷
x
s
x1
x1
x2
(a) Forward propagation
x2
(b) Inverse propagation
Figure 2: Computational graphs for forward and inverse propagation. A coupling layer applies a
simple invertible transformation consisting of scaling followed by addition of a constant offset to
one part x2 of the input vector conditioned on the remaining part of the input vector x1 . Because of
its simple nature, this transformation is both easily invertible and possesses a tractable determinant.
However, the conditional nature of this transformation, captured by the functions s and t, significantly
increase the flexibility of this otherwise weak function. The forward and inverse propagation
operations have identical computational cost.
3.2
Coupling layers
Computing the Jacobian of functions with high-dimensional domain and codomain and computing
the determinants of large matrices are in general computationally very expensive. This combined
with the restriction to bijective functions makes Equation 2 appear impractical for modeling arbitrary
distributions.
As shown however in [17], by careful design of the function f , a bijective model can be learned which
is both tractable and extremely flexible. As computing the Jacobian determinant of the transformation
is crucial to effectively train using this principle, this work exploits the simple observation that the
determinant of a triangular matrix can be efficiently computed as the product of its diagonal terms.
We will build a flexible and tractable bijective function by stacking a sequence of simple bijections.
In each simple bijection, part of the input vector is updated using a function which is simple to invert,
but which depends on the remainder of the input vector in a complex way. We refer to each of these
simple bijections as an affine coupling layer. Given a D dimensional input x and d < D, the output
y of an affine coupling layer follows the equations
y1:d = x1:d
yd+1:D = xd+1:D
(4)
exp s(x1:d ) + t(x1:d ),
where s and t stand for scale and translation, and are functions from Rd 7→ RD−d , and
Hadamard product or element-wise product (see Figure 2(a)).
3.3
(5)
is the
Properties
The Jacobian of this transformation is
"
#
Id
0
∂y
= ∂yd+1:D diag exp [s (x )]
,
(6)
1:d
∂xT
∂xT
1:d
where diag exp [s (x1:d )] is the diagonal matrix whose diagonal elements correspond to the vector
exp [s (x1:d )]. Given the
hPobservationi that this Jacobian is triangular, we can efficiently compute
its determinant as exp
j s (x1:d )j . Since computing the Jacobian determinant of the coupling
layer operation does not involve computing the Jacobian of s or t, those functions can be arbitrarily
complex. We will make them deep convolutional neural networks. Note that the hidden layers of s
and t can have more features than their input and output layers.
Another interesting property of these coupling layers in the context of defining probabilistic models
is their invertibility. Indeed, computing the inverse is no more complex than the forward propagation
4
Published as a conference paper at ICLR 2017
1
2
5
6
3
4
7
8
4
8
3
7
2
6
1
5
Figure 3: Masking schemes for affine coupling layers. On the left, a spatial checkerboard pattern
mask. On the right, a channel-wise masking. The squeezing operation reduces the 4 × 4 × 1 tensor
(on the left) into a 2 × 2 × 4 tensor (on the right). Before the squeezing operation, a checkerboard
pattern is used for coupling layers while a channel-wise masking pattern is used afterward.
(see Figure 2(b)),
⇔
y1:d
yd+1:D
x1:d
xd+1:D
= x1:d
= xd+1:D
exp s(x1:d ) + t(x1:d )
= y1:d
= yd+1:D − t(y1:d )
exp − s(y1:d ) ,
(7)
(8)
meaning that sampling is as efficient as inference for this model. Note again that computing the
inverse of the coupling layer does not require computing the inverse of s or t, so these functions can
be arbitrarily complex and difficult to invert.
3.4
Masked convolution
Partitioning can be implemented using a binary mask b, and using the functional form for y,
y = b x + (1 − b)
x exp s(b x) + t(b x) .
(9)
We use two partitionings that exploit the local correlation structure of images: spatial checkerboard
patterns, and channel-wise masking (see Figure 3). The spatial checkerboard pattern mask has value
1 where the sum of spatial coordinates is odd, and 0 otherwise. The channel-wise mask b is 1 for the
first half of the channel dimensions and 0 for the second half. For the models presented here, both
s(·) and t(·) are rectified convolutional networks.
3.5
Combining coupling layers
Although coupling layers can be powerful, their forward transformation leaves some components
unchanged. This difficulty can be overcome by composing coupling layers in an alternating pattern,
such that the components that are left unchanged in one coupling layer are updated in the next (see
Figure 4(a)).
The Jacobian determinant of the resulting function remains tractable, relying on the fact that
∂(fb ◦ fa )
∂fa
∂fb
(xa ) =
(xa ) ·
xb = fa (xa )
T
T
T
∂xa
∂xa
∂xb
det(A · B) = det(A) det(B).
(10)
(11)
Similarly, its inverse can be computed easily as
(fb ◦ fa )−1 = fa−1 ◦ fb−1 .
5
(12)
Published as a conference paper at ICLR 2017
+
x
z1
+
x
=
z2
z3
z4
z3
(2)
4
(3)
f
h
(2)
f
+
x
=
(1)
3
(1)
4
z1
z2
h
h
x1
x2
x3
x4
=
(1)
f
(a) In this alternating pattern, units which remain identical in one
transformation are modified in the next.
(b) Factoring out variables.
At each step, half the variables are directly modeled as
Gaussians, while the other
half undergo further transformation.
Figure 4: Composition schemes for affine coupling layers.
3.6
Multi-scale architecture
We implement a multi-scale architecture using a squeezing operation: for each channel, it divides the
image into subsquares of shape 2 × 2 × c, then reshapes them into subsquares of shape 1 × 1 × 4c.
The squeezing operation transforms an s × s × c tensor into an 2s × 2s × 4c tensor (see Figure 3),
effectively trading spatial size for number of channels.
At each scale, we combine several operations into a sequence: we first apply three coupling layers
with alternating checkerboard masks, then perform a squeezing operation, and finally apply three
more coupling layers with alternating channel-wise masking. The channel-wise masking is chosen so
that the resulting partitioning is not redundant with the previous checkerboard masking (see Figure
3). For the final scale, we only apply four coupling layers with alternating checkerboard masks.
Propagating a D dimensional vector through all the coupling layers would be cumbersome, in terms
of computational and memory cost, and in terms of the number of parameters that would need to be
trained. For this reason we follow the design choice of [57] and factor out half of the dimensions at
regular intervals (see Equation 14). We can define this operation recursively (see Figure 4(b)),
h(0) = x
(z
(i+1)
(i+1)
,h
z
)=f
(L)
=f
(13)
(i+1)
(L)
z = (z
(i)
(h )
(L−1)
(h
(1)
,...,z
(14)
)
(L)
(15)
).
(16)
In our experiments, we use this operation for i < L. The sequence of coupling-squeezing-coupling
operations described above is performed per layer when computing f (i) (Equation 14). At each
layer, as the spatial resolution is reduced, the number of hidden layer features in s and t is doubled.
All variables which have been factored out at different scales are concatenated to obtain the final
transformed output (Equation 16).
As a consequence, the model must Gaussianize units which are factored out at a finer scale (in an
earlier layer) before those which are factored out at a coarser scale (in a later layer). This results in the
definition of intermediary levels of representation [53, 49] corresponding to more local, fine-grained
features as shown in Appendix D.
Moreover, Gaussianizing and factoring out units in earlier layers has the practical benefit of distributing the loss function throughout the network, following the philosophy similar to guiding intermediate
layers using intermediate classifiers [40]. It also reduces significantly the amount of computation and
memory used by the model, allowing us to train larger models.
6
Published as a conference paper at ICLR 2017
3.7
Batch normalization
To further improve the propagation of training signal, we use deep residual networks [24, 25] with
batch normalization [31] and weight normalization [2, 54] in s and t. As described in Appendix E
we introduce and use a novel variant of batch normalization which is based on a running average over
recent minibatches, and is thus more robust when training with very small minibatches.
We also use apply batch normalization to the whole coupling layer output. The effects of batch
normalization are easily included in the Jacobian computation, since it acts as a linear rescaling on
each dimension. That is, given the estimated batch statistics µ̃ and σ̃ 2 , the rescaling function
x − µ̃
x 7→ √
σ̃ 2 +
(17)
has a Jacobian determinant
!− 12
Y
(σ̃i2
+ )
.
(18)
i
This form of batch normalization can be seen as similar to reward normalization in deep reinforcement
learning [44, 65].
We found that the use of this technique not only allowed training with a deeper stack of coupling
layers, but also alleviated the instability problem that practitioners often encounter when training
conditional distributions with a scale parameter through a gradient-based approach.
4
4.1
Experiments
Procedure
The algorithm described in Equation 2 shows how to learn distributions on unbounded space. In
general, the data of interest have bounded magnitude. For examples, the pixel values of an image
typically lie in [0, 256]D after application of the recommended jittering procedure [64, 62]. In order to
x
reduce the impact of boundary effects, we instead model the density of logit(α+(1−α) 256
), where
α is picked here as .05. We take into account this transformation when computing log-likelihood and
bits per dimension. We also augment the CIFAR-10, CelebA and LSUN datasets during training to
also include horizontal flips of the training examples.
We train our model on four natural image datasets: CIFAR-10 [36], Imagenet [52], Large-scale Scene
Understanding (LSUN) [70], CelebFaces Attributes (CelebA) [41]. More specifically, we train on the
downsampled to 32 × 32 and 64 × 64 versions of Imagenet [46]. For the LSUN dataset, we train on
the bedroom, tower and church outdoor categories. The procedure for LSUN is the same as in [47]:
we downsample the image so that the smallest side is 96 pixels and take random crops of 64 × 64. For
CelebA, we use the same procedure as in [38]: we take an approximately central crop of 148 × 148
then resize it to 64 × 64.
We use the multi-scale architecture described in Section 3.6 and use deep convolutional residual
networks in the coupling layers with rectifier nonlinearity and skip-connections as suggested by [46].
To compute the scaling functions s, we use a hyperbolic tangent function multiplied by a learned
scale, whereas the translation function t has an affine output. Our multi-scale architecture is repeated
recursively until the input of the last recursion is a 4 × 4 × c tensor. For datasets of images of size
32 × 32, we use 4 residual blocks with 32 hidden feature maps for the first coupling layers with
checkerboard masking. Only 2 residual blocks are used for images of size 64 × 64. We use a batch
size of 64. For CIFAR-10, we use 8 residual blocks, 64 feature maps, and downscale only once. We
optimize with ADAM [33] with default hyperparameters and use an L2 regularization on the weight
scale parameters with coefficient 5 · 10−5 .
We set the prior pZ to be an isotropic unit norm Gaussian. However, any distribution could be used
for pZ , including distributions that are also learned during training, such as from an auto-regressive
model, or (with slight modifications to the training objective) a variational autoencoder.
7
Published as a conference paper at ICLR 2017
Dataset
CIFAR-10
Imagenet (32 × 32)
Imagenet (64 × 64)
LSUN (bedroom)
LSUN (tower)
LSUN (church outdoor)
CelebA
PixelRNN [46]
3.00
3.86 (3.83)
3.63 (3.57)
Real NVP
3.49
4.28 (4.26)
3.98 (3.75)
2.72 (2.70)
2.81 (2.78)
3.08 (2.94)
3.02 (2.97)
Conv DRAW [22]
< 3.59
< 4.40 (4.35)
< 4.10 (4.04)
IAF-VAE [34]
< 3.28
Table 1: Bits/dim results for CIFAR-10, Imagenet, LSUN datasets and CelebA. Test results for
CIFAR-10 and validation results for Imagenet, LSUN and CelebA (with training results in parenthesis
for reference).
Figure 5: On the left column, examples from the dataset. On the right column, samples from the
model trained on the dataset. The datasets shown in this figure are in order: CIFAR-10, Imagenet
(32 × 32), Imagenet (64 × 64), CelebA, LSUN (bedroom).
4.2
Results
We show in Table 1 that the number of bits per dimension, while not improving over the Pixel RNN
[46] baseline, is competitive with other generative methods. As we notice that our performance
increases with the number of parameters, larger models are likely to further improve performance.
For CelebA and LSUN, the bits per dimension for the validation set was decreasing throughout
training, so little overfitting is expected.
We show in Figure 5 samples generated from the model with training examples from the dataset
for comparison. As mentioned in [62, 22], maximum likelihood is a principle that values diversity
8
Published as a conference paper at ICLR 2017
Figure 6: Manifold generated from four examples in the dataset. Clockwise from top left: CelebA,
Imagenet (64 × 64), LSUN (tower), LSUN (bedroom).
over sample quality in a limited capacity setting. As a result, our model outputs sometimes highly
improbable samples as we can notice especially on CelebA. As opposed to variational autoencoders,
the samples generated from our model look not only globally coherent but also sharp. Our hypothesis
is that as opposed to these models, real NVP does not rely on fixed form reconstruction cost like an L2
norm which tends to reward capturing low frequency components more heavily than high frequency
components. Unlike autoregressive models, sampling from our model is done very efficiently as it is
parallelized over input dimensions. On Imagenet and LSUN, our model seems to have captured well
the notion of background/foreground and lighting interactions such as luminosity and consistent light
source direction for reflectance and shadows.
We also illustrate the smooth semantically consistent meaning of our latent variables. In the latent
space, we define a manifold based on four validation examples z(1) , z(2) , z(3) , z(4) , and parametrized
by two parameters φ and φ0 by,
z = cos(φ) cos(φ0 )z(1) + sin(φ0 )z(2) + sin(φ) cos(φ0 )z(3) + sin(φ0 )z(4) .
(19)
We project the resulting manifold back into the data space by computing g(z). Results are shown
Figure 6. We observe that the model seems to have organized the latent space with a notion of
meaning that goes well beyond pixel space interpolation. More visualization are shown in the
Appendix. To further test whether the latent space has a consistent semantic interpretation, we trained
a class-conditional model on CelebA, and found that the learned representation had a consistent
semantic meaning across class labels (see Appendix F).
5
Discussion and conclusion
In this paper, we have defined a class of invertible functions with tractable Jacobian determinant,
enabling exact and tractable log-likelihood evaluation, inference, and sampling. We have shown that
this class of generative model achieves competitive performances, both in terms of sample quality
and log-likelihood. Many avenues exist to further improve the functional form of the transformations,
for instance by exploiting the latest advances in dilated convolutions [69] and residual networks
architectures [60].
This paper presented a technique bridging the gap between auto-regressive models, variational
autoencoders, and generative adversarial networks. Like auto-regressive models, it allows tractable
and exact log-likelihood evaluation for training. It allows however a much more flexible functional
form, similar to that in the generative model of variational autoencoders. This allows for fast
and exact sampling from the model distribution. Like GANs, and unlike variational autoencoders,
our technique does not require the use of a fixed form reconstruction cost, and instead defines a
cost in terms of higher level features, generating sharper images. Finally, unlike both variational
9
Published as a conference paper at ICLR 2017
autoencoders and GANs, our technique is able to learn a semantically meaningful latent space which
is as high dimensional as the input space. This may make the algorithm particularly well suited to
semi-supervised learning tasks, as we hope to explore in future work.
Real NVP generative models can additionally be conditioned on additional variables (for instance
class labels) to create a structured output algorithm. More so, as the resulting class of invertible
transformations can be treated as a probability distribution in a modular way, it can also be used to
improve upon other probabilistic models like auto-regressive models and variational autoencoders.
For variational autoencoders, these transformations could be used both to enable a more flexible
reconstruction cost [38] and a more flexible stochastic inference distribution [48]. Probabilistic
models in general can also benefit from batch normalization techniques as applied in this paper.
The definition of powerful and trainable invertible functions can also benefit domains other than
generative unsupervised learning. For example, in reinforcement learning, these invertible functions
can help extend the set of functions for which an argmax operation is tractable for continuous Qlearning [23] or find representation where local linear Gaussian approximations are more appropriate
[67].
6
Acknowledgments
The authors thank the developers of Tensorflow [1]. We thank Sherry Moore, David Andersen and
Jon Shlens for their help in implementing the model. We thank Aäron van den Oord, Yann Dauphin,
Kyle Kastner, Chelsea Finn, Maithra Raghu, David Warde-Farley, Daniel Jiwoong Im and Oriol
Vinyals for fruitful discussions. Finally, we thank Ben Poole, Rafal Jozefowicz and George Dahl for
their input on a draft of the paper.
References
[1] Martın Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S
Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, et al. Tensorflow: Large-scale machine learning on
heterogeneous distributed systems. arXiv preprint arXiv:1603.04467, 2016.
[2] Vijay Badrinarayanan, Bamdev Mishra, and Roberto Cipolla. Understanding symmetries in deep networks.
arXiv preprint arXiv:1511.01029, 2015.
[3] Johannes Ballé, Valero Laparra, and Eero P Simoncelli. Density modeling of images using a generalized
normalization transformation. arXiv preprint arXiv:1511.06281, 2015.
[4] Anthony J Bell and Terrence J Sejnowski. An information-maximization approach to blind separation and
blind deconvolution. Neural computation, 7(6):1129–1159, 1995.
[5] Yoshua Bengio. Artificial neural networks and their application to sequence recognition. 1991.
[6] Yoshua Bengio and Samy Bengio. Modeling high-dimensional discrete data with multi-layer neural
networks. In NIPS, volume 99, pages 400–406, 1999.
[7] Mathias Berglund and Tapani Raiko. Stochastic gradient estimate variance in contrastive divergence and
persistent contrastive divergence. arXiv preprint arXiv:1312.6002, 2013.
[8] Samuel R Bowman, Luke Vilnis, Oriol Vinyals, Andrew M Dai, Rafal Jozefowicz, and Samy Bengio.
Generating sentences from a continuous space. arXiv preprint arXiv:1511.06349, 2015.
[9] Joan Bruna, Pablo Sprechmann, and Yann LeCun. Super-resolution with deep convolutional sufficient
statistics. arXiv preprint arXiv:1511.05666, 2015.
[10] Yuri Burda, Roger Grosse, and Ruslan Salakhutdinov. Importance weighted autoencoders. arXiv preprint
arXiv:1509.00519, 2015.
[11] Scott Shaobing Chen and Ramesh A Gopinath. Gaussianization. In Advances in Neural Information
Processing Systems, 2000.
[12] Junyoung Chung, Kyle Kastner, Laurent Dinh, Kratarth Goel, Aaron C Courville, and Yoshua Bengio. A
recurrent latent variable model for sequential data. In Advances in neural information processing systems,
pages 2962–2970, 2015.
[13] Peter Dayan, Geoffrey E Hinton, Radford M Neal, and Richard S Zemel. The helmholtz machine. Neural
computation, 7(5):889–904, 1995.
[14] Gustavo Deco and Wilfried Brauer. Higher order statistical decorrelation without information loss. In
G. Tesauro, D. S. Touretzky, and T. K. Leen, editors, Advances in Neural Information Processing Systems
7, pages 247–254. MIT Press, 1995.
[15] Emily L. Denton, Soumith Chintala, Arthur Szlam, and Rob Fergus. Deep generative image models using
a laplacian pyramid of adversarial networks. In Advances in Neural Information Processing Systems 28:
10
Published as a conference paper at ICLR 2017
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal,
Quebec, Canada, pages 1486–1494, 2015.
Luc Devroye. Sample-based non-uniform random variate generation. In Proceedings of the 18th conference
on Winter simulation, pages 260–265. ACM, 1986.
Laurent Dinh, David Krueger, and Yoshua Bengio. Nice: non-linear independent components estimation.
arXiv preprint arXiv:1410.8516, 2014.
Brendan J Frey. Graphical models for machine learning and digital communication. MIT press, 1998.
Leon A. Gatys, Alexander S. Ecker, and Matthias Bethge. Texture synthesis using convolutional neural
networks. In Advances in Neural Information Processing Systems 28: Annual Conference on Neural
Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pages 262–270,
2015.
Mathieu Germain, Karol Gregor, Iain Murray, and Hugo Larochelle. MADE: masked autoencoder for
distribution estimation. CoRR, abs/1502.03509, 2015.
Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair,
Aaron C. Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information
Processing Systems 27: Annual Conference on Neural Information Processing Systems 2014, December
8-13 2014, Montreal, Quebec, Canada, pages 2672–2680, 2014.
Karol Gregor, Frederic Besse, Danilo Jimenez Rezende, Ivo Danihelka, and Daan Wierstra. Towards
conceptual compression. arXiv preprint arXiv:1604.08772, 2016.
Shixiang Gu, Timothy Lillicrap, Ilya Sutskever, and Sergey Levine. Continuous deep q-learning with
model-based acceleration. arXiv preprint arXiv:1603.00748, 2016.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition.
CoRR, abs/1512.03385, 2015.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks.
CoRR, abs/1603.05027, 2016.
Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Computation, 9(8):1735–1780,
1997.
Matthew D Hoffman, David M Blei, Chong Wang, and John Paisley. Stochastic variational inference. The
Journal of Machine Learning Research, 14(1):1303–1347, 2013.
Aapo Hyvärinen, Juha Karhunen, and Erkki Oja. Independent component analysis, volume 46. John Wiley
& Sons, 2004.
Aapo Hyvärinen and Petteri Pajunen. Nonlinear independent component analysis: Existence and uniqueness
results. Neural Networks, 12(3):429–439, 1999.
Daniel Jiwoong Im, Chris Dongjoo Kim, Hui Jiang, and Roland Memisevic. Generating images with
recurrent adversarial networks. arXiv preprint arXiv:1602.05110, 2016.
Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing
internal covariate shift. arXiv preprint arXiv:1502.03167, 2015.
Rafal Józefowicz, Oriol Vinyals, Mike Schuster, Noam Shazeer, and Yonghui Wu. Exploring the limits of
language modeling. CoRR, abs/1602.02410, 2016.
Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint
arXiv:1412.6980, 2014.
Diederik P Kingma, Tim Salimans, and Max Welling. Improving variational inference with inverse
autoregressive flow. arXiv preprint arXiv:1606.04934, 2016.
Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114,
2013.
Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images, 2009.
Hugo Larochelle and Iain Murray. The neural autoregressive distribution estimator. In AISTATS, 2011.
Anders Boesen Lindbo Larsen, Søren Kaae Sønderby, and Ole Winther. Autoencoding beyond pixels using
a learned similarity metric. CoRR, abs/1512.09300, 2015.
Yann A LeCun, Léon Bottou, Genevieve B Orr, and Klaus-Robert Müller. Efficient backprop. In Neural
networks: Tricks of the trade, pages 9–48. Springer, 2012.
Chen-Yu Lee, Saining Xie, Patrick Gallagher, Zhengyou Zhang, and Zhuowen Tu. Deeply-supervised nets.
arXiv preprint arXiv:1409.5185, 2014.
Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild. In
Proceedings of International Conference on Computer Vision (ICCV), December 2015.
Lars Maaløe, Casper Kaae Sønderby, Søren Kaae Sønderby, and Ole Winther. Auxiliary deep generative
models. arXiv preprint arXiv:1602.05473, 2016.
Andriy Mnih and Karol Gregor. Neural variational inference and learning in belief networks. arXiv
preprint arXiv:1402.0030, 2014.
Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare,
Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through
deep reinforcement learning. Nature, 518(7540):529–533, 2015.
Radford M Neal and Geoffrey E Hinton. A view of the em algorithm that justifies incremental, sparse, and
other variants. In Learning in graphical models, pages 355–368. Springer, 1998.
11
Published as a conference paper at ICLR 2017
[46] Aaron van den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. arXiv
preprint arXiv:1601.06759, 2016.
[47] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep
convolutional generative adversarial networks. CoRR, abs/1511.06434, 2015.
[48] Danilo Jimenez Rezende and Shakir Mohamed. Variational inference with normalizing flows. arXiv
preprint arXiv:1505.05770, 2015.
[49] Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. arXiv preprint arXiv:1401.4082, 2014.
[50] Oren Rippel and Ryan Prescott Adams. High-dimensional probability estimation with deep density models.
arXiv preprint arXiv:1302.5125, 2013.
[51] David E Rumelhart, Geoffrey E Hinton, and Ronald J Williams. Learning representations by backpropagating errors. Cognitive modeling, 5(3):1, 1988.
[52] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang,
Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge.
International Journal of Computer Vision, 115(3):211–252, 2015.
[53] Ruslan Salakhutdinov and Geoffrey E Hinton. Deep boltzmann machines. In International conference on
artificial intelligence and statistics, pages 448–455, 2009.
[54] Tim Salimans and Diederik P Kingma. Weight normalization: A simple reparameterization to accelerate
training of deep neural networks. arXiv preprint arXiv:1602.07868, 2016.
[55] Tim Salimans, Diederik P Kingma, and Max Welling. Markov chain monte carlo and variational inference:
Bridging the gap. arXiv preprint arXiv:1410.6460, 2014.
[56] Lawrence K Saul, Tommi Jaakkola, and Michael I Jordan. Mean field theory for sigmoid belief networks.
Journal of artificial intelligence research, 4(1):61–76, 1996.
[57] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014.
[58] Paul Smolensky. Information processing in dynamical systems: Foundations of harmony theory. Technical
report, DTIC Document, 1986.
[59] Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised
learning using nonequilibrium thermodynamics. In Proceedings of the 32nd International Conference on
Machine Learning, ICML 2015, Lille, France, 6-11 July 2015, pages 2256–2265, 2015.
[60] Sasha Targ, Diogo Almeida, and Kevin Lyman. Resnet in resnet: Generalizing residual architectures.
CoRR, abs/1603.08029, 2016.
[61] Lucas Theis and Matthias Bethge. Generative image modeling using spatial lstms. In Advances in Neural
Information Processing Systems, pages 1918–1926, 2015.
[62] Lucas Theis, Aäron Van Den Oord, and Matthias Bethge. A note on the evaluation of generative models.
CoRR, abs/1511.01844, 2015.
[63] Dustin Tran, Rajesh Ranganath, and David M Blei. Variational gaussian process. arXiv preprint
arXiv:1511.06499, 2015.
[64] Benigno Uria, Iain Murray, and Hugo Larochelle. Rnade: The real-valued neural autoregressive densityestimator. In Advances in Neural Information Processing Systems, pages 2175–2183, 2013.
[65] Hado van Hasselt, Arthur Guez, Matteo Hessel, and David Silver. Learning functions across many orders
of magnitudes. arXiv preprint arXiv:1602.07714, 2016.
[66] Oriol Vinyals, Samy Bengio, and Manjunath Kudlur. Order matters: Sequence to sequence for sets. arXiv
preprint arXiv:1511.06391, 2015.
[67] Manuel Watter, Jost Springenberg, Joschka Boedecker, and Martin Riedmiller. Embed to control: A locally
linear latent dynamics model for control from raw images. In Advances in Neural Information Processing
Systems, pages 2728–2736, 2015.
[68] Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement
learning. Machine learning, 8(3-4):229–256, 1992.
[69] Fisher Yu and Vladlen Koltun. Multi-scale context aggregation by dilated convolutions. arXiv preprint
arXiv:1511.07122, 2015.
[70] Fisher Yu, Yinda Zhang, Shuran Song, Ari Seff, and Jianxiong Xiao. Construction of a large-scale image
dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015.
[71] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. arXiv preprint
arXiv:1603.08511, 2016.
12
Published as a conference paper at ICLR 2017
A
Samples
Figure 7: Samples from a model trained on Imagenet (64 × 64).
13
Published as a conference paper at ICLR 2017
Figure 8: Samples from a model trained on CelebA.
14
Published as a conference paper at ICLR 2017
Figure 9: Samples from a model trained on LSUN (bedroom category).
15
Published as a conference paper at ICLR 2017
Figure 10: Samples from a model trained on LSUN (church outdoor category).
16
Published as a conference paper at ICLR 2017
Figure 11: Samples from a model trained on LSUN (tower category).
17
Published as a conference paper at ICLR 2017
B
Manifold
Figure 12: Manifold from a model trained on Imagenet (64 × 64). Images with red borders are taken
from the validation set, and define the manifold. The manifold was computed as described in Equation
19, where the x-axis corresponds to φ, and the y-axis to φ0 , and where φ, φ0 ∈ {0, π4 , · · · , 7π
4 }.
18
Published as a conference paper at ICLR 2017
Figure 13: Manifold from a model trained on CelebA. Images with red borders are taken from the
training set, and define the manifold. The manifold was computed as described in Equation 19, where
the x-axis corresponds to φ, and the y-axis to φ0 , and where φ, φ0 ∈ {0, π4 , · · · , 7π
4 }.
19
Published as a conference paper at ICLR 2017
Figure 14: Manifold from a model trained on LSUN (bedroom category). Images with red borders are taken from the validation set, and define the manifold. The manifold was computed as
described in Equation 19, where the x-axis corresponds to φ, and the y-axis to φ0 , and where
φ, φ0 ∈ {0, π4 , · · · , 7π
4 }.
20
Published as a conference paper at ICLR 2017
Figure 15: Manifold from a model trained on LSUN (church outdoor category). Images with red
borders are taken from the validation set, and define the manifold. The manifold was computed
as described in Equation 19, where the x-axis corresponds to φ, and the y-axis to φ0 , and where
φ, φ0 ∈ {0, π4 , · · · , 7π
4 }.
21
Published as a conference paper at ICLR 2017
Figure 16: Manifold from a model trained on LSUN (tower category). Images with red borders are taken from the validation set, and define the manifold. The manifold was computed
as described in Equation 19, where the x-axis corresponds to φ, and the y-axis to φ0 , and where
φ, φ0 ∈ {0, π4 , · · · , 7π
4 }.
C
Extrapolation
Inspired by the texture generation work by [19, 61] and extrapolation test with DCGAN [47], we also evaluate
the statistics captured by our model by generating images twice or ten times as large as present in the dataset. As
we can observe in the following figures, our model seems to successfully create a “texture” representation of
the dataset while maintaining a spatial smoothness through the image. Our convolutional architecture is only
aware of the position of considered pixel through edge effects in convolutions, therefore our model is similar to
a stationary process. This also explains why these samples are more consistent in LSUN, where the training data
was obtained using random crops.
22
Published as a conference paper at ICLR 2017
(a) ×2
(b) ×10
Figure 17: We generate samples a factor bigger than the training set image size on Imagenet (64×64).
23
Published as a conference paper at ICLR 2017
(a) ×2
(b) ×10
Figure 18: We generate samples a factor bigger than the training set image size on CelebA.
24
Published as a conference paper at ICLR 2017
(a) ×2
(b) ×10
Figure 19: We generate samples a factor bigger than the training set image size on LSUN (bedroom
category).
25
Published as a conference paper at ICLR 2017
(a) ×2
(b) ×10
Figure 20: We generate samples a factor bigger than the training set image size on LSUN (church
outdoor category).
26
Published as a conference paper at ICLR 2017
(a) ×2
(b) ×10
Figure 21: We generate samples a factor bigger than the training set image size on LSUN (tower
category).
27
Published as a conference paper at ICLR 2017
D
Latent variables semantic
As in [22], we further try to grasp the semantic of our learned layers latent variables by doing ablation tests. We
infer the latent variables and resample the lowest levels of latent variables from a standard gaussian, increasing
the highest level affected by this resampling. As we can see in the following figures, the semantic of our
latent space seems to be more on a graphic level rather than higher level concept. Although the heavy use of
convolution improves learning by exploiting image prior knowledge, it is also likely to be responsible for this
limitation.
Figure 22: Conceptual compression from a model trained on Imagenet (64 × 64). The leftmost
column represent the original image, the subsequent columns were obtained by storing higher level
latent variables and resampling the others, storing less and less as we go right. From left to right:
100%, 50%, 25%, 12.5% and 6.25% of the latent variables are kept.
Figure 23: Conceptual compression from a model trained on CelebA. The leftmost column represent
the original image, the subsequent columns were obtained by storing higher level latent variables
and resampling the others, storing less and less as we go right. From left to right: 100%, 50%, 25%,
12.5% and 6.25% of the latent variables are kept.
28
Published as a conference paper at ICLR 2017
Figure 24: Conceptual compression from a model trained on LSUN (bedroom category). The leftmost
column represent the original image, the subsequent columns were obtained by storing higher level
latent variables and resampling the others, storing less and less as we go right. From left to right:
100%, 50%, 25%, 12.5% and 6.25% of the latent variables are kept.
Figure 25: Conceptual compression from a model trained on LSUN (church outdoor category). The
leftmost column represent the original image, the subsequent columns were obtained by storing
higher level latent variables and resampling the others, storing less and less as we go right. From left
to right: 100%, 50%, 25%, 12.5% and 6.25% of the latent variables are kept.
29
Published as a conference paper at ICLR 2017
Figure 26: Conceptual compression from a model trained on LSUN (tower category). The leftmost
column represent the original image, the subsequent columns were obtained by storing higher level
latent variables and resampling the others, storing less and less as we go right. From left to right:
100%, 50%, 25%, 12.5% and 6.25% of the latent variables are kept.
E
Batch normalization
We further experimented with batch normalization by using a weighted average of a moving average of the layer
statistics µ̃t , σ̃t2 and the current batch batch statistics µ̂t , σ̂t2 ,
µ̃t+1 = ρµ̃t + (1 − ρ)µ̂t
(20)
2
σ̃t+1
(21)
=
ρσ̃t2
+ (1 −
ρ)σ̂t2 ,
2
where ρ is the momentum. When using µ̃t+1 , σ̃t+1
, we only propagate gradient through the current batch
statistics µ̂t , σ̂t2 . We observe that using this lag helps the model train with very small minibatches.
We used batch normalization with a moving average for our results on CIFAR-10.
F
Attribute change
Additionally, we exploit the attribute information y in CelebA to build a conditional model, i.e. the invertible
function f from image to latent variable uses the labels in y to define its parameters. In order to observe the
information stored in the latent variables, we choose to encode a batch of images x with their original attribute y
0
and decode them using a new set of attributes
y , build by shuffling the original attributes inside the batch. We
obtain the new images x0 = g f (x; y); y 0 .
We observe that, although the faces are changed as to respect the new attributes, several properties remain
unchanged like position and background.
30
Published as a conference paper at ICLR 2017
Figure 27: Examples x from the CelebA dataset.
31
Published as a conference paper at ICLR 2017
Figure 28: From a model trained on pairs of images and attributes from the CelebA dataset, we encode
a batch of images with their original attributes before decoding them with a new set of attributes. We
notice that the new images often share similar characteristics with those in Fig 27, including position
and background.
32
| 9cs.NE |
Published as a conference paper at ICLR 2018
S KIP RNN: L EARNING TO S KIP S TATE U PDATES IN
R ECURRENT N EURAL N ETWORKS
†
‡
§
†
Γ
arXiv:1708.06834v3 [cs.AI] 5 Feb 2018
Vı́ctor Campos∗ , Brendan Jou , Xavier Giró-i-Nieto , Jordi Torres , Shih-Fu Chang
†
Barcelona Supercomputing Center, ‡Google Inc,
§
Universitat Politècnica de Catalunya, ΓColumbia University
{victor.campos, jordi.torres}@bsc.es, bjou@google.com,
xavier.giro@upc.edu, shih.fu.chang@columbia.edu
A BSTRACT
Recurrent Neural Networks (RNNs) continue to show outstanding performance
in sequence modeling tasks. However, training RNNs on long sequences often
face challenges like slow inference, vanishing gradients and difficulty in capturing
long term dependencies. In backpropagation through time settings, these issues
are tightly coupled with the large, sequential computational graph resulting from
unfolding the RNN in time. We introduce the Skip RNN model which extends
existing RNN models by learning to skip state updates and shortens the effective
size of the computational graph. This model can also be encouraged to perform
fewer state updates through a budget constraint. We evaluate the proposed model
on various tasks and show how it can reduce the number of required RNN updates
while preserving, and sometimes even improving, the performance of the baseline
RNN models. Source code is publicly available at https://imatge-upc.
github.io/skiprnn-2017-telecombcn/.
1
I NTRODUCTION
Recurrent Neural Networks (RNNs) have become the standard approach for practitioners when addressing machine learning tasks involving sequential data. Such success has been enabled by the
appearance of larger datasets, more powerful computing resources and improved architectures and
training algorithms. Gated units, such as the Long Short-Term Memory (Hochreiter & Schmidhuber, 1997) (LSTM) and the Gated Recurrent Unit (Cho et al., 2014) (GRU), were designed to deal
with the vanishing gradients problem commonly found in RNNs (Bengio et al., 1994). These architectures have been popularized, in part, due to their impressive results on a variety of tasks in
machine translation (Bahdanau et al., 2015), language modeling (Zaremba et al., 2015) and speech
recognition (Graves et al., 2013).
Some of the main challenges of RNNs are in their training and deployment when dealing with
long sequences, due to their inherently sequential behaviour. These challenges include throughput
degradation, slower convergence during training and memory leakage, even for gated architectures
(Neil et al., 2016). Sequence shortening techniques, which can be seen as a sort of conditional
computation (Bengio et al., 2013; Bengio, 2013; Davis & Arel, 2013) in time, can alleviate these
issues. The most common approaches, such as cropping discrete signals or reducing the sampling
rate in continuous signals, are based on heuristics and can be suboptimal. In contrast, we propose
a model that is able to learn which samples (i.e., elements in the input sequence) need to be used
in order to solve the target task. Consider a video understanding task as an example: scenes with
large motion may benefit from high frame rates, whereas only a few frames are needed to capture
the semantics of a mostly static scene.
The main contribution of this work is a novel modification for existing RNN architectures that allows them to skip state updates, decreasing the number of sequential operations performed, without
requiring any additional supervision signal. This model, called Skip RNN, adaptively determines
whether the state needs to be updated or copied to the next time step. We show how the network can
∗
Work done while Vı́ctor Campos was a visiting scholar at Columbia University.
1
Published as a conference paper at ICLR 2018
be encouraged to perform fewer state updates by adding a penalization term during training, allowing us to train models under different computation budgets. The proposed modification can generally
be integrated with any RNN and we show, in this paper, implementations with well-known RNNs,
namely LSTM and GRU. The resulting models show promising results on a series of sequence modeling tasks. In particular, we evaluate the proposed Skip RNN architecture on six sequence learning
problems: an adding task, sine wave frequency discrimination, digit classification, sentiment analysis in movie reviews, action classification in video, and temporal action localization in video1 .
2
R ELATED WORK
Conditional computation has been shown to gradually increase model capacity without proportional
increases in computational cost by exploiting certain computation paths for each input (Bengio et al.,
2013; Liu & Deng, 2017; Almahairi et al., 2016; McGill & Perona, 2017; Shazeer et al., 2017). This
idea has been extended in the temporal domain, such as in learning how many times an input needs
to be ”pondered” before moving to the next one (Graves, 2016) or designing RNN architectures
whose number of layers depend on the input data (Chung et al., 2017). Other works have addressed
time-dependent computation in RNNs by updating only a fraction of the hidden states based on the
current hidden state and input (Jernite et al., 2017), or following periodic patterns (Koutnik et al.,
2014; Neil et al., 2016). However, due to the inherently sequential nature of RNNs and the parallel
computation capabilities of modern hardware, reducing the size of the matrices involved in the
computations performed at each time step generally has not accelerated inference as dramatically
as hoped. The proposed Skip RNN model can be seen as form of conditional computation in time,
where the computation associated to the RNN updates may or may not be executed at every time
step. This idea is related to the UPDATE and COPY operations in hierarchical multiscale RNNs
(Chung et al., 2017), but applied to the whole stack of RNN layers at the same time. This difference
is key to allowing our approach to skip input samples, effectively reducing sequential computation
and shielding the hidden state over longer time lags. Learning whether to update or copy the hidden
state through time steps can be seen as a learnable Zoneout mask (Krueger et al., 2017) which is
shared between all the units in the hidden state. Similarly, it can be interpreted as an input-dependent
recurrent version of stochastic depth (Huang et al., 2016).
Selecting parts of the input signal is similar in spirit to the hard attention mechanisms that have
been applied to image regions (Mnih et al., 2014), where only some patches of the input image
are attended in order to generate captions (Xu et al., 2015) or detect objects (Ba et al., 2014). Our
model can be understood as generating a hard temporal attention mask on-the-fly given previously
seen samples, deciding which time steps should be attended and operating on a subset of input samples. Subsampling input sequences has been explored for visual storylines generation (Sigurdsson
et al., 2016b), although jointly optimizing the RNN weights and the subsampling mechanism is often
computationally infeasible and they use an Expectation-Maximization algorithm instead. Similar research has been conducted for video analysis tasks, discovering minimally needed evidence for event
recognition (Bhattacharya et al., 2014) and training agents that decide which frames need to be observed in order to localize actions in time (Yeung et al., 2016; Su & Grauman, 2016). Motivated by
the advantages of training recurrent models on shorter subsequences, efforts have been conducted
on learning differentiable subsampling mechanisms (Raffel & Lawson, 2017), although the computational complexity of the proposed method precludes its application to long input sequences.
In contrast, our proposed method can be trained with backpropagation and does not degrade the
complexity of the baseline RNNs.
Accelerating inference in RNNs is difficult due to their inherently sequential nature, leading to the
design of Quasi-Recurrent Neural Networks (Bradbury et al., 2017) and Simple Recurrent Units
(Lei & Zhang, 2017), which relax the temporal dependency between consecutive steps. With the
goal of speeding up RNN inference, LSTM-Jump (Yu et al., 2017) augments an LSTM cell with
a classification layer that will decide how many steps to jump between RNN updates. Despite its
promising results on text tasks, the model needs to be trained with REINFORCE (Williams, 1992),
which requires defining a reasonable reward signal. Determining these rewards are non-trivial and
may not necessarily generalize across tasks. Moreover, the number of tokens read between jumps,
1
Experiments on sine wave frequency discrimination, sentiment analysis in movie reviews and action classification in video are reported in the appendix due to space limitations.
2
Published as a conference paper at ICLR 2018
the maximum jump distance and the number of jumps allowed all need to be chosen in advance.
These hyperparameters define a reduced set of subsequences that the model can sample, instead of
allowing the network to learn any arbitrary sampling scheme. Unlike LSTM-Jump, our proposed
approach is differentiable, thus not requiring any modifications to the loss function and simplifying
the optimization process, and is not limited to a predefined set of sample selection patterns.
3
M ODEL D ESCRIPTION
An RNN takes an input sequence x = (x1 , . . . , xT ) and generates a state sequence s = (s1 , . . . , sT )
by iteratively applying a parametric state transition model S from t = 1 to T :
st = S(st−1 , xt )
(1)
We augment the network with a binary state update gate, ut ∈ {0, 1}, selecting whether the state of
the RNN will be updated (ut = 1) or copied from the previous time step (ut = 0). At every time
step t, the probability ũt+1 ∈ [0, 1] of performing a state update at t + 1 is emitted. The resulting
architecture is depicted in Figure 1 and can be characterized as follows:
ut = fbinarize (ũt )
st = ut · S(st−1 , xt ) + (1 − ut ) · st−1
∆ũt = σ(Wp st + bp )
ũt+1 = ut · ∆ũt + (1 − ut ) · (ũt + min(∆ũt , 1 − ũt ))
(2)
(3)
(4)
(5)
where Wp is a weights vector, bp is a scalar bias, σ is the sigmoid function and fbinarize : [0, 1] →
{0, 1} binarizes the input value. Should the network be composed of several layers, some columns
of Wp can be fixed to 0 so that ∆ũt depends only on the states of a subset of layers (see Section
4.3 for an example with two layers). We implement fbinarize as a deterministic step function ut =
round(ũt ), although a stochastic sampling from a Bernoulli distribution ut ∼ Bernoulli(ũt ) would
be possible as well.
The model formulation encodes the observation that the likelihood of requesting a new input to
update the state increases with the number of consecutively skipped samples. Whenever a state
update is omitted, the pre-activation of the state update gate for the following time step, ũt+1 , is
incremented by ∆ũt . On the other hand, if a state update is performed, the accumulated value is
flushed and ũt+1 = ∆ũt .
The number of skipped time steps can be computed ahead of time. For the particular formulation
used in this work, where fbinarize is implemented by means of a rounding function, the number of
skipped samples after performing a state update at time step t is given by:
Nskip (t) = min{n : n · ∆ũt ≥ 0.5} − 1
(6)
+
where n ∈ Z . This enables more efficient implementations where no computation at all is performed whenever ut = 0. These computational savings are possible because ∆ũt = σ(Wp st +bp ) =
σ(Wp st−1 + bp ) = ∆ũt−1 when ut = 0 and there is no need to evaluate it again, as depicted in
Figure 1d.
There are several advantages in reducing the number of RNN updates. From the computational
standpoint, fewer updates translates into fewer required sequential operations to process an input
signal, leading to faster inference and reduced energy consumption. Unlike some other models that
aim to reduce the average number of operations per step (Neil et al., 2016; Jernite et al., 2017),
ours enables skipping steps completely. Replacing RNN updates with copy operations increases
the memory of the network and its ability to model long term dependencies even for gated units,
since the exponential memory decay observed in LSTM and GRU (Neil et al., 2016) is alleviated.
During training, gradients are propagated through fewer updating time steps, providing faster convergence in some tasks involving long sequences. Moreover, the proposed model is orthogonal to
recent advances in RNNs and could be used in conjunction with such techniques, e.g. normalization
(Cooijmans et al., 2017; Ba et al., 2016), regularization (Zaremba et al., 2015; Krueger et al., 2017),
variable computation (Jernite et al., 2017; Neil et al., 2016) or even external memory (Graves et al.,
2014; Weston et al., 2014).
3
Published as a conference paper at ICLR 2018
+
0
ũt
1
fbinarize
ũt
ũt+1
ũt+1
ut
Δũt
Δũt
σ
σ
st-1
0
S
st
1
st-1
xt
st
S
xt
(a)
ũt
(b)
ũt+1
+
ũt
+
Δũt-1
Δũt
ũt+1
Δũt
σ
st-1
st
st-1
xt
st
xt
(c)
(d)
Figure 1: Model architecture of the proposed Skip RNN. (a) Complete Skip RNN architecture,
where the computation graph at time step t is conditioned on ut . (b) Architecture when the state
is updated, i.e. ut = 1. (c) Architecture when the update step is skipped and the previous state is
copied, i.e. ut = 0. (d) In practice, redundant computation is avoided by propagating ∆ũt between
time steps when ut = 0.
3.1
E RROR GRADIENTS
The whole model is differentiable except for fbinarize , which outputs binary values. A common
method for optimizing functions involving discrete variables is REINFORCE (Williams, 1992), although several estimators have been proposed for the particular case of neurons with binary outputs
(Bengio et al., 2013). We select the straight-through estimator (Hinton, 2012; Bengio et al., 2013),
which consists of approximating the step function by the identity when computing gradients during
the backward pass:
∂fbinarize (x)
=1
∂x
(7)
This yields a biased estimator that has proven more efficient than other unbiased but high-variance
estimators such as REINFORCE (Bengio et al., 2013) and has been successfully applied in different
works (Courbariaux et al., 2016; Chung et al., 2017). By using the straight-through estimator as
the backward pass for fbinarize , all the model parameters can be trained to minimize the target loss
function with standard backpropagation and without defining any additional supervision or reward
signal.
3.2
L IMITING COMPUTATION
The Skip RNN is able to learn when to update or copy the state without explicit information about
which samples are useful to solve the task at hand. However, a different operating point on the
trade-off between performance and number of processed samples may be required depending on
the application, e.g. one may be willing to sacrifice a few accuracy points in order to run faster on
machines with a low computational power, or to reduce energy impact on portable devices. The
4
Published as a conference paper at ICLR 2018
proposed model can be encouraged to perform fewer state updates through additional loss terms, a
common practice in neural networks with dynamically allocated computation (Liu & Deng, 2017;
McGill & Perona, 2017; Graves, 2016; Jernite et al., 2017). In particular, we consider a cost per
sample condition
Lbudget = λ ·
T
X
ut ,
(8)
t=1
where Lbudget is the cost associated to a single sequence, λ is the cost per sample and T is the
sequence length. This formulation bears a similarity to weight decay regularization, where the
network is encouraged to slowly converge toward a solution where the norm of the weights is small.
Similarly, in this case the network is encouraged to converge toward a solution where fewer state
updates are required.
Although the above budget formulation is extensively studied in our experiments, other budget loss
terms can be used depending on the application. For instance, a specific number of samples may
be encouraged by applying a L1 or L2 loss between the target value and the number of updates per
PT
sequence, t=1 ut .
4
E XPERIMENTS
In the following section, we investigate the advantages of adding this state skipping to two common
RNN architectures, LSTM and GRU, for a variety of tasks. In addition to the evaluation metric for
each task, we report the number of RNN state updates (i.e., the number of elements in the input
sequence used by the model) and the number of floating point operations (FLOPs) as measures of
the computational load for each model. Since skipping an RNN update results in ignoring its corresponding input, we will refer to the number of updates and the number of used samples (i.e. elements
in a sequence) interchangeably. With the goal of studying the effect of skipping state updates on the
learning capability of the networks, we also introduce a baseline which skips a state update with
probability pskip . We tune the skipping probability to obtain models that perform a similar number
of state updates to the Skip RNN models.
Training is performed with Adam (Kingma & Ba, 2014), learning rate of 10−4 , β1 = 0.9, β2 =
0.999 and = 10−8 on batches of 256. Gradient clipping (Pascanu et al., 2013) with a threshold of
1 is applied to all trainable variables. Bias bp in Equation 4 is initialized to 1, so that all samples are
used at the beginning of training2 . The initial hidden state s0 is learned during training, whereas ũ0
is set to a constant value of 1 in order to force the first update at t = 1.
Experiments are implemented with TensorFlow3 and run on a single NVIDIA K80 GPU.
4.1
A DDING TASK
We revisit one of the original LSTM tasks (Hochreiter & Schmidhuber, 1997), where the network is
given a sequence of (value, marker) tuples. The desired output is the addition of only two values that
are marked with a 1, whereas those marked with a 0 need to be ignored. We follow the experimental
setup in Neil et al. (2016), where the first marker is randomly placed among the first 10% of samples
(drawn with uniform probability) and the second one is placed among the last half of samples (drawn
with uniform probability). This marker distribution yields sequences where at least 40% of the
samples are distractors and provide no useful information at all. However, it is worth noting that in
this task the risk of missing a marker is very large as compared to the benefits of working on shorter
subsequences.
2
In practice, forcing the network to use all samples at the beginning of training improves its robustness
against random initializations of its weights and increases the reproducibility of the presented experiments. A
similar behavior was observed in other augmented RNN architectures such as Neural Stacks (Grefenstette et al.,
2015).
3
https://www.tensorflow.org
5
Published as a conference paper at ICLR 2018
Model
Task solved
State updates
Inference FLOPs
LSTM
LSTM (pskip = 0.2)
LSTM (pskip = 0.5)
Skip LSTM, λ = 0
Skip LSTM, λ = 10−5
Yes
No
No
Yes
Yes
100.0% ± 0.0%
80.0% ± 0.1%
50.1% ± 0.1%
81.1% ± 3.6%
53.9% ± 2.1%
2.46 × 106
1.97 × 106
1.23 × 106
2.00 × 106
1.33 × 106
GRU
GRU (pskip = 0.02)
GRU (pskip = 0.5)
Skip GRU, λ = 0
Skip GRU, λ = 10−5
Yes
No
No
Yes
Yes
100.0% ± 0.0%
98.0% ± 0.0%
49.9% ± 0.6%
97.9% ± 3.2%
50.7% ± 2.6%
1.85 × 106
1.81 × 106
9.25 × 105
1.81 × 106
9.40 × 105
Table 1: Results for the adding task, displayed as mean ± std over four different runs. The task
is considered to be solved if the MSE is at least two orders of magnitude below the variance of the
output distribution.
We train RNN models with 110 units each on sequences of length 50, where the values are uniformly
drawn from U(−0.5, 0.5). The final RNN state is fed to a fully connected layer that regresses the
scalar output. The model is trained to minimize the Mean Squared Error (MSE) between the output
and the ground truth. We consider that a model is able to solve the task when its MSE on a held-out
set of examples is at least two orders of magnitude below the variance of the output distribution.
This criterion is a stricter version of the one followed by Hochreiter & Schmidhuber (1997).
While all models learn to solve the task, results in Table 1 show that Skip RNN models are able to
do so with roughly half of the updates of their corresponding counterparts. We observed that the
models using fewer updates never miss any marker, since the penalization in terms of MSE would
be very large (see Section B.1 for examples). This is confirmed by the poor performance of the
baselines that randomly skip state updates, which are not able to solve the tasks even when the
skipping probability is low. Skip RNN models learn to skip most of the samples in the 40% of the
sequence where there are no markers. Moreover, most updates are skipped once the second marker
is found, since all the relevant information in the sequence has already been seen. This last pattern
provides evidence that the proposed models effectively learn whether to update or copy the hidden
state based on the input sequence, as opposed to learning biases in the dataset only. As a downside,
Skip RNN models show some difficulties skipping a large number of updates at once, probably due
to the cumulative nature of ũt .
4.2
MNIST C LASSIFICATION FROM A S EQUENCE OF P IXELS
The MNIST handwritten digits classification benchmark (LeCun et al., 1998) is traditionally addressed with Convolutional Neural Networks (CNNs) that efficiently exploit spatial dependencies
through weight sharing. By flattening the 28 × 28 images into 784-d vectors, however, it can be
reformulated as a challenging task for RNNs where long term dependencies need to be leveraged
(Le et al., 2015b). We follow the standard data split and set aside 5,000 training samples for validation purposes. After processing all pixels with an RNN with 110 units, the last hidden state is fed
into a linear classifier predicting the digit class. All models are trained for 600 epochs to minimize
cross-entropy loss.
Table 2 summarizes classification results on the test set after 600 epochs of training. Skip RNNs are
not only able to solve the task using fewer updates than their counterparts, but also show a lower
variation among runs and train faster (see Figure 2). We hypothesize that skipping updates make
the Skip RNNs work on shorter subsequences, simplifying the optimization process and allowing
the networks to capture long term dependencies more easily. A similar behavior was observed
for Phased LSTM, where increasing the sparsity of cell updates accelerates training for very long
sequences (Neil et al., 2016). However, the drop in performance observed in the models where the
state updates are skipped randomly suggests that learning which samples to use is a key component
in the performance of Skip RNN.
6
Published as a conference paper at ICLR 2018
Model
Accuracy
State updates
Inference FLOPs
LSTM
LSTM (pskip = 0.5)
Skip LSTM, λ = 10−4
0.910 ± 0.045
0.893 ± 0.003
0.973 ± 0.002
784.00 ± 0.00
392.03 ± 0.05
379.38 ± 33.09
3.83 × 107
1.91 × 107
1.86 × 107
GRU
GRU (pskip = 0.5)
Skip GRU, λ = 10−4
0.968 ± 0.013
0.912 ± 0.004
0.976 ± 0.003
784.00 ± 0.00
391.86 ± 0.14
392.62 ± 26.48
2.87 × 107
1.44 × 107
1.44 × 107
TANH-RNN (Le et al., 2015a)
iRNN (Le et al., 2015a)
uRNN (Arjovsky et al., 2016)
sTANH-RNN (Zhang et al., 2016)
LSTM (Cooijmans et al., 2017)
BN-LSTM (Cooijmans et al., 2017)
0.350
0.970
0.951
0.981
0.989
0.990
784.00
784.00
784.00
784.00
784.00
784.00
-
Table 2: Accuracy, used samples and average FLOPs per sequence at inference on the test set of
MNIST after 600 epochs of training. Results are displayed as mean ± std over four different runs.
Accuracy (%)
100
80
60
40
GRU
Skip GRU, = 10
20
0
100
200
300
Epochs
400
500
4
600
Figure 2: Accuracy evolution during training on the validation set of MNIST. The Skip GRU exhibits
lower variance and faster convergence than the baseline GRU. A similar behavior is observed for
LSTM and Skip LSTM, but omitted for clarity. Shading shows maximum and minimum over 4 runs,
while dark lines indicate the mean.
The performance of RNN models on this task can be boosted through techniques like recurrent batch
normalization (Cooijmans et al., 2017) or recurrent skip coefficients (Zhang et al., 2016). Cooijmans
et al. (2017) show how an LSTM with specific weight initialization schemes for improved gradient
flow (Le et al., 2015a; Arjovsky et al., 2016) can reach accuracy rates of up to 0.989%. Note that
these techniques are orthogonal to skipping state updates and Skip RNN models could benefit from
them as well.
Sequences of pixels can be reshaped back into 2D images, allowing to visualize the samples used
by the RNNs as a sort of hard visual attention model (Xu et al., 2015). Examples such as the ones
depicted in Figure 3 show how the model learns to skip pixels that are not discriminative, such as
the padding regions in the top and bottom of images. Similarly to the qualitative results for the
adding task (Section 4.1), attended samples vary depending on the particular input being given to
the network.
4.3
T EMPORAL ACTION LOCALIZATION ON C HARADES
One popular approach to video analysis tasks today is to extract frame-level features with a CNN
and modeling temporal dynamics with an RNN (Donahue et al., 2015; Yue-Hei Ng et al., 2015).
Videos are commonly recorded at high sampling rates, generating long sequences with strong tem7
Published as a conference paper at ICLR 2018
Figure 3: Sample usage examples for the Skip LSTM with λ = 10−4 on the test set of MNIST. Red
pixels are used, whereas blue ones are skipped.
poral redundancies that are challenging for RNNs. Moreover, processing frames with a CNN is
computationally expensive and may become prohibitive for high frame rates. These issues have
been alleviated in previous works by using short clips (Donahue et al., 2015) or by downsampling
the original data in order to cover long temporal spans without increasing the sequence length excessively (Yue-Hei Ng et al., 2015). Instead of addressing the long sequence problem at the input
data level, we let the network learn which frames need to be used.
Charades (Sigurdsson et al., 2016a) is a dataset containing 9,848 videos annotated for 157 action
classes in a per-frame fashion. Frames are encoded using fc7 features from the RGB stream of a
Two-Stream CNN provided by the organizers of the challenge4 , extracted at 6 fps. The encoded
frames are fed into two stacked RNN layers with 256 units each and the hidden state in the last RNN
layer is used to compute the update probability for the Skip RNN models. Since each frame may
be annotated with zero or more classes, the networks are trained to minimize element-wise binary
cross-entropy at every time step. Unlike the previous sequence tagging tasks, this setup allows us to
evaluate the performance of Skip RNN on a task where the output is a sequence as well.
Evaluation is performed following the setup by Sigurdsson et al. (2016c), but evaluating on 100
equally spaced frames instead of 25, and results are reported in Table 3. It is surprising that the
GRU baselines that randomly skip state updates perform on par with their Skip GRU counterparts
for low skipping probabilities. We hypothesize several reasons for this behavior, which was not
observed in previous experiments: (1) there is a supervision signal at every time step and the inputs
and (2) outputs are strongly correlated in consecutive frames. On the other hand, Skip RNN models
clearly outperform the random methods when fewer updates are allowed. Note that this setup is far
more challenging because of the longer time spans between updates, so properly distributing the
state updates along the sequence is key to the performance of the models. Interestingly, Skip RNN
models learn which frames need to be attended from RGB data and without having access to explicit
motion information.
Skip GRU tends to perform fewer state updates than Skip LSTM when the cost per sample is low or
none. This behavior is the opposite of the one observed in the adding task (Section 4.1), which may
be related to the observation that determining the best performing gated unit depends on the task at
hand Chung et al. (2014). Indeed, GRU models consistently outperform LSTM ones on this task.
This mismatch in the number of used samples is not observed for large values of λ, as both Skip
LSTM and Skip GRU converge to a comparable number of used samples.
A previous work reports better action localization performance by integrating RGB and optical flow
information as an input to an LSTM, reaching 9.60% mAP (Sigurdsson et al., 2016c). This boost
in performance comes at the cost of roughly doubling the number of FLOPs and memory footprint
of the CNN encoder, plus requiring the extraction of flow information during a preprocessing step.
Interestingly, our model learns which frames need to be attended from RGB data and without having
access to explicit motion information.
4
http://vuchallenge.org/charades.html
8
Published as a conference paper at ICLR 2018
Model
mAP (%)
State updates
Inference FLOPs
LSTM
LSTM (pskip = 0.75)
LSTM (pskip = 0.90)
Skip LSTM, λ = 0
Skip LSTM, λ = 10−4
Skip LSTM, λ = 10−3
Skip LSTM, λ = 10−2
8.40
8.11
7.21
8.32
8.61
8.32
7.86
172.9 ± 47.4
43.3 ± 13.2
17.2 ± 6.1
172.9 ± 47.4
172.9 ± 47.4
41.9 ± 11.3
17.4 ± 4.4
2.65 × 1012
6.63 × 1011
2.65 × 1011
2.65 × 1012
2.65 × 1012
6.41 × 1011
2.66 × 1011
GRU
GRU (pskip = 0.10)
GRU (pskip = 0.40)
GRU (pskip = 0.70)
GRU (pskip = 0.90)
Skip GRU, λ = 0
Skip GRU, λ = 10−4
Skip GRU, λ = 10−3
Skip GRU, λ = 10−2
8.70
8.94
8.81
8.42
7.09
8.94
8.76
8.68
7.95
172.9 ± 47.4
155.6 ± 42.9
103.6 ± 29.3
51.9 ± 15.4
17.3 ± 6.3
159.9 ± 46.9
100.8 ± 28.1
54.2 ± 16.2
18.4 ± 5.1
2.65 × 1012
2.39 × 1012
1.06 × 1012
7.95 × 1011
2.65 × 1011
2.45 × 1012
1.54 × 1012
8.29 × 1011
2.82 × 1011
Table 3: Mean Average Precision (mAP), used samples and average FLOPs per sequence at inference on the validation set of Charades. The number of state updates is displayed as mean ± std
over all the videos in the validation set.
5
C ONCLUSION
We presented Skip RNNs as an extension to existing recurrent architectures enabling them to skip
state updates thereby reducing the number of sequential operations in the computation graph. Unlike other approaches, all parameters in Skip RNN are trained with backpropagation. Experiments
conducted with LSTMs and GRUs showed that Skip RNNs can match or in some cases even outperform the baseline models while relaxing their computational requirements. Skip RNNs provide
faster and more stable training for long sequences and complex models, owing to gradients being
backpropagated through fewer time steps resulting in a simpler optimization task. Moreover, the
introduced computational savings are better suited for modern hardware than those methods that
reduce the amount of computation required at each time step (Koutnik et al., 2014; Neil et al., 2016;
Chung et al., 2017).
ACKNOWLEDGMENTS
This work was partially supported by the Spanish Ministry of Economy and Competitivity and the
European Regional Development Fund (ERDF) under contracts TEC2016-75976-R and TIN201565316-P, by the BSC-CNS Severo Ochoa program SEV-2015-0493, and grant 2014-SGR-1051 by
the Catalan Government. Vı́ctor Campos was supported by Obra Social “la Caixa” through La
Caixa-Severo Ochoa International Doctoral Fellowship program. We would also like to thank the
technical support team at the Barcelona Supercomputing Center.
R EFERENCES
Amjad Almahairi, Nicolas Ballas, Tim Cooijmans, Yin Zheng, Hugo Larochelle, and Aaron
Courville. Dynamic capacity networks. In ICML, 2016.
Martin Arjovsky, Amar Shah, and Yoshua Bengio. Unitary evolution recurrent neural networks. In
ICML, 2016.
Jimmy Ba, Volodymyr Mnih, and Koray Kavukcuoglu. Multiple object recognition with visual
attention. arXiv preprint arXiv:1412.7755, 2014.
Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint
arXiv:1607.06450, 2016.
9
Published as a conference paper at ICLR 2018
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly
learning to align and translate. In ICLR, 2015.
Yoshua Bengio. Deep learning of representations: Looking forward. In SLSP, 2013.
Yoshua Bengio, Patrice Simard, and Paolo Frasconi. Learning long-term dependencies with gradient
descent is difficult. IEEE Transactions on Neural Networks, 1994.
Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients
through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013.
Subhabrata Bhattacharya, Felix X Yu, and Shih-Fu Chang. Minimally needed evidence for complex
event recognition in unconstrained videos. In ICMR, 2014.
James Bradbury, Stephen Merity, Caiming Xiong, and Richard Socher. Quasi-recurrent neural networks. In ICLR, 2017.
Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics
dataset. In CVPR, 2017.
Kyunghyun Cho, Bart Van Merriënboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder
for statistical machine translation. In EMNLP, 2014.
Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of
gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014.
Junyoung Chung, Sungjin Ahn, and Yoshua Bengio. Hierarchical multiscale recurrent neural networks. In ICLR, 2017.
Tim Cooijmans, Nicolas Ballas, César Laurent, Çağlar Gülçehre, and Aaron Courville. Recurrent
batch normalization. In ICLR, 2017.
Matthieu Courbariaux, Itay Hubara, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. Binarized
neural networks: Training deep neural networks with weights and activations constrained to+ 1
or-1. arXiv preprint arXiv:1602.02830, 2016.
Andrew Davis and Itamar Arel. Low-rank approximations for conditional feedforward computation
in deep neural networks. arXiv preprint arXiv:1312.4461, 2013.
Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale
hierarchical image database. In CVPR, 2009.
Jeffrey Donahue, Lisa Anne Hendricks, Sergio Guadarrama, Marcus Rohrbach, Subhashini Venugopalan, Kate Saenko, and Trevor Darrell. Long-term recurrent convolutional networks for visual
recognition and description. In CVPR, 2015.
Alex Graves. Adaptive computation time for recurrent neural networks.
arXiv:1603.08983, 2016.
arXiv preprint
Alex Graves, Abdel-rahman Mohamed, and Geoffrey Hinton. Speech recognition with deep recurrent neural networks. In ICASSP, 2013.
Alex Graves, Greg Wayne, and Ivo Danihelka.
arXiv:1410.5401, 2014.
Neural turing machines.
arXiv preprint
Edward Grefenstette, Karl Moritz Hermann, Mustafa Suleyman, and Phil Blunsom. Learning to
transduce with unbounded memory. In NIPS, 2015.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016.
Geoffrey Hinton. Neural networks for machine learning. Coursera video lectures, 2012.
Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 1997.
10
Published as a conference paper at ICLR 2018
Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with
stochastic depth. In ECCV, 2016.
Yacine Jernite, Edouard Grave, Armand Joulin, and Tomas Mikolov. Variable computation in recurrent neural networks. In ICLR, 2017.
Yoon Kim. Convolutional neural networks for sentence classification. In EMNLP, 2014.
Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint
arXiv:1412.6980, 2014.
Jan Koutnik, Klaus Greff, Faustino Gomez, and Juergen Schmidhuber. A clockwork rnn. In ICML,
2014.
David Krueger, Tegan Maharaj, János Kramár, Mohammad Pezeshki, Nicolas Ballas, Nan Rosemary Ke, Anirudh Goyal, Yoshua Bengio, Hugo Larochelle, Aaron Courville, et al. Zoneout:
Regularizing rnns by randomly preserving hidden activations. In ICLR, 2017.
Quoc V Le, Navdeep Jaitly, and Geoffrey E Hinton. A simple way to initialize recurrent networks
of rectified linear units. arXiv preprint arXiv:1504.00941, 2015a.
Quoc V Le, Navdeep Jaitly, and Geoffrey E Hinton. A simple way to initialize recurrent networks
of rectified linear units. arXiv preprint arXiv:1504.00941, 2015b.
Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to
document recognition. Proceedings of the IEEE, 1998.
Tao Lei and Yu Zhang. Training rnns as fast as cnns. arXiv preprint arXiv:1709.02755, 2017.
Lanlan Liu and Jia Deng. Dynamic deep neural networks: Optimizing accuracy-efficiency trade-offs
by selective execution. arXiv preprint arXiv:1701.00299, 2017.
Shayne Longpre, Sabeek Pradhan, Caiming Xiong, and Richard Socher. A way out of the odyssey:
Analyzing and combining recent insights for lstms. arXiv preprint arXiv:1611.05104, 2016.
Andrew L Maas, Raymond E Daly, Peter T Pham, Dan Huang, Andrew Y Ng, and Christopher
Potts. Learning word vectors for sentiment analysis. In ACL, 2011.
Mason McGill and Pietro Perona. Deciding how to decide: Dynamic routing in artificial neural
networks. In ICML, 2017.
Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed representations of words and phrases and their compositionality. In NIPS, 2013.
Takeru Miyato, Andrew M Dai, and Ian Goodfellow. Adversarial training methods for semisupervised text classification. In ICLR, 2017.
Volodymyr Mnih, Nicolas Heess, Alex Graves, et al. Recurrent models of visual attention. In NIPS,
2014.
Daniel Neil, Michael Pfeiffer, and Shih-Chii Liu. Phased LSTM: accelerating recurrent network
training for long or event-based sequences. In NIPS, 2016.
Razvan Pascanu, Tomas Mikolov, and Yoshua Bengio. On the difficulty of training recurrent neural
networks. In ICML, 2013.
Colin Raffel and Dieterich Lawson. Training a subsampling mechanism in expectation. In ICLR
Workshop Track, 2017.
Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and
Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In
ICLR, 2017.
Evan Shelhamer, Kate Rakelly, Judy Hoffman, and Trevor Darrell. Clockwork convnets for video
semantic segmentation. arXiv preprint arXiv:1608.03609, 2016.
11
Published as a conference paper at ICLR 2018
Gunnar Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In ECCV, 2016a.
Gunnar A Sigurdsson, Xinlei Chen, and Abhinav Gupta. Learning visual storylines with skipping
recurrent neural networks. In ECCV, 2016b.
Gunnar A Sigurdsson, Santosh Divvala, Ali Farhadi, and Abhinav Gupta. Asynchronous temporal
fields for action recognition. arXiv preprint arXiv:1612.06371, 2016c.
Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition
in videos. In NIPS, 2014.
Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions
classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012.
Yu-Chuan Su and Kristen Grauman. Leaving some stones unturned: dynamic feature prioritization
for activity detection in streaming video. In ECCV, 2016.
Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In ICCV, 2015.
Jason Weston, Sumit Chopra, and Antoine Bordes.
arXiv:1410.3916, 2014.
Memory networks.
arXiv preprint
Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement
learning. Machine learning, 1992.
Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich
Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual
attention. In ICML, 2015.
Serena Yeung, Olga Russakovsky, Greg Mori, and Li Fei-Fei. End-to-end learning of action detection from frame glimpses in videos. In CVPR, 2016.
Adams Wei Yu, Hongrae Lee, and Quoc V Le. Learning to skim text. In ACL, 2017.
Joe Yue-Hei Ng, Matthew Hausknecht, Sudheendra Vijayanarasimhan, Oriol Vinyals, Rajat Monga,
and George Toderici. Beyond short snippets: Deep networks for video classification. In CVPR,
2015.
Wojciech Zaremba, Ilya Sutskever, and Oriol Vinyals. Recurrent neural network regularization. In
ICLR, 2015.
Saizheng Zhang, Yuhuai Wu, Tong Che, Zhouhan Lin, Roland Memisevic, Ruslan R Salakhutdinov,
and Yoshua Bengio. Architectural complexity measures of recurrent neural networks. In NIPS,
2016.
12
Published as a conference paper at ICLR 2018
A
A DDITIONAL E XPERIMENTS
A.1
F REQUENCY D ISCRIMINATION TASK
In this experiment, the network is trained to classify between sinusoids whose period is in range
T ∼ U (5, 6) milliseconds and those whose period is in range T ∼ {(1, 5) ∪ (6, 100)} milliseconds
(Neil et al., 2016). Every sine wave with period T has a random phase shift drawn from U(0, T ). At
every time step, the input to the network is a single scalar representing the amplitude of the signal.
Since sinusoid are continuous signals, this tasks allows to study whether Skip RNNs converge to the
same solutions when their parameters are fixed but the sampling period is changed. We study two
different sampling periods, Ts = {0.5, 1} milliseconds, for each set of hyperparameters.
We train RNNs with 110 units each on input signals of 100 milliseconds. Batches are stratified,
containing the same number of samples for each class, yielding a 50% chance accuracy. The last
state of the RNN is fed into a 2-way classifier and trained with cross-entropy loss. We consider that a
model is able to solve the task when it achieves an accuracy over 99% on a held-out set of examples.
Table 4 summarizes results for this task. When no cost per sample is set (λ = 0), the number
of updates differ under different sampling conditions. We attribute this behavior to the potentially
large number of local minima in the cost function, since there are numerous subsampling patterns
for which the task can be successfully solved and we are not explicitly encouraging the network to
converge to a particular solution. On the other hand, when λ > 0 Skip RNN models with the same
cost per sample use roughly the same number of input samples even when the sampling frequency
is doubled. This is a desirable property, since solutions are robust to oversampled input signals.
Qualitative results can be found in Section B.2.
Model
Ts = 1ms (length 100)
Ts = 0.5ms (length 200)
Task solved
State updates
Task solved
State updates
LSTM
Skip LSTM, λ = 0
Skip LSTM, λ = 10−5
Skip LSTM, λ = 10−4
Yes
Yes
Yes
Yes
100.0 ± 0.00
55.5 ± 16.9
47.4 ± 14.1
12.7 ± 0.5
Yes
Yes
Yes
Yes
200.0 ± 0.00
147.9 ± 27.0
50.7 ± 16.8
19.9 ± 1.5
GRU
Skip GRU, λ = 0
Skip GRU, λ = 10−5
Skip GRU, λ = 10−4
Yes
Yes
Yes
Yes
100.0 ± 0.00
73.7 ± 17.9
51.9 ± 10.2
23.5 ± 6.2
Yes
Yes
Yes
Yes
200.0 ± 0.00
167.0 ± 18.3
54.2 ± 4.4
22.5 ± 2.1
Table 4: Results for the frequency discrimination task, displayed as mean ± std over four different
runs. The task is considered to be solved if the classification accuracy is over 99%. Models with
the same cost per sample (λ > 0) converge to a similar number of used samples under different
sampling conditions.
A.2
S ENTIMENT A NALYSIS ON IMDB
The IMDB dataset (Maas et al., 2011) contains 25,000 training and 25,000 testing movie reviews
annotated into two classes, positive and negative sentiment, with an approximate average length of
240 words per review. We set aside 15% of training data for validation purposes. Words are embedded into 300-d vector representations before being fed to an RNN with 128 units. The embedding
matrix is initialized using pre-trained word2vec5 embeddings (Mikolov et al., 2013) when available,
or random vectors drawn from U(−0.25, 0.25) otherwise (Kim, 2014). Dropout with rate 0.2 is
applied between the last RNN state and the classification layer in order to reduce overfitting. We
evaluate the models on sequences of length 200 and 400 by cropping longer sequences and padding
shorter ones (Yu et al., 2017).
Results on the test are reported in Table 5. In a task where it is hard to predict which input tokens will
be discriminative, the Skip RNN models are able to achieve similar accuracy rates to the baseline
5
https://code.google.com/archive/p/word2vec/
13
Published as a conference paper at ICLR 2018
models while reducing the number of required updates. These results highlight the trade-off between
accuracy and the available computational budget, since a larger cost per sample results in lower
accuracies. However, allowing the network to select which samples to use instead of cropping
sequences at a given length boosts performance, as observed for the Skip LSTM (length 400, λ =
10−4 ), which achieves a higher accuracy than the baseline LSTM (length 200) while seeing roughly
the same number of words per review. A similar behavior can be seen for the Skip RNN models with
λ = 10−3 , where allowing them to select words from longer reviews boosts classification accuracy
while using a comparable number of tokens per sequence.
In order to reduce overfitting of large models, Miyato et al. (2017) leverage additional unlabeled
data through adversarial training and achieve a state of the art accuracy of 0.941 on IMDB. For an
extended analysis on how different experimental setups affect the performance of RNNs on this task,
we refer the reader to (Longpre et al., 2016).
Model
Length 200
Length 400
Accuracy
State updates
Accuracy
State updates
LSTM
Skip LSTM, λ = 0
Skip LSTM, λ = 10−5
Skip LSTM, λ = 10−4
Skip LSTM, λ = 10−3
0.843 ± 0.003
0.844 ± 0.004
0.846 ± 0.004
0.837 ± 0.006
0.811 ± 0.007
200.00 ± 0.00
196.75 ± 5.63
197.15 ± 3.16
164.65 ± 8.67
73.85 ± 1.90
0.868 ± 0.004
0.866 ± 0.004
0.865 ± 0.001
0.862 ± 0.003
0.836 ± 0.007
400.00 ± 0.00
369.70 ± 19.35
380.62 ± 18.20
186.30 ± 25.72
84.22 ± 1.98
GRU
Skip GRU, λ = 0
Skip GRU, λ = 10−5
Skip GRU, λ = 10−4
Skip GRU, λ = 10−3
0.845 ± 0.006
0.848 ± 0.002
0.842 ± 0.005
0.834 ± 0.006
0.800 ± 0.007
200.00 ± 0.00
200.00 ± 0.00
199.25 ± 1.30
180.97 ± 8.90
106.15 ± 37.92
0.862 ± 0.003
0.866 ± 0.002
0.862 ± 0.008
0.853 ± 0.011
0.814 ± 0.005
400.00 ± 0.00
399.02 ± 1.69
398.00 ± 2.06
314.30 ± 2.82
99.12 ± 2.69
Table 5: Accuracy and used samples on the test set of IMDB for different sequence lengths. Results
are displayed as mean ± std over four different runs.
A.3
ACTION CLASSIFICATION ON UCF-101
UCF-101 (Soomro et al., 2012) is a dataset containing 13,320 trimmed videos belonging to 101
different action categories. We use 10 seconds of video sampled at 25 fps, cropping longer ones
and padding shorter examples with empty frames. Activations in the Global Average Pooling layer
from a ResNet-50 (He et al., 2016) CNN pretrained on the ImageNet dataset (Deng et al., 2009) are
used as frame-level features, which are fed into two stacked RNN layers with 512 units each. The
weights in the CNN are not tuned during training to reduce overfitting. The hidden state in the last
RNN layer is used to compute the update probability for the Skip RNN models.
We evaluate the different models on the first split of UCF-101 and report results in Table 6. Skip
RNN models do not only improve the classification accuracy with respect to the baseline, but require
very few updates to do so, possibly due to the low motion between consecutive frames resulting in
frame-level features with high temporal redundancy (Shelhamer et al., 2016). Moreover, Figure 4
shows how models performing fewer updates converge faster thanks to the gradients being preserved
during longer spans when training with backpropagation through time.
Non recurrent architectures for video action recognition that have achieved high performance on
UCF-101 comprise CNNs with spatiotemporal kernels (Tran et al., 2015) or two-stream CNNs (Simonyan & Zisserman, 2014). Carreira & Zisserman (2017) show the benefits of expanding 2D CNN
filters into 3D and pretraining on larger datasets, obtaining an accuracy of 0.845 when using RGB
data only and 0.934 when incorporating optical flow information.
14
Published as a conference paper at ICLR 2018
Model
Accuracy
State updates
Inference FLOPs
LSTM
Skip LSTM, λ = 0
Skip LSTM, λ = 10−5
Skip LSTM, λ = 10−4
0.671
0.749
0.757
0.790
250.0
138.9
24.2
7.6
9.52 × 1011
5.29 × 1011
9.21 × 1010
2.89 × 1010
GRU
Skip GRU, λ = 0
Skip GRU, λ = 10−5
Skip GRU, λ = 10−4
0.791
0.796
0.792
0.793
250.0
124.2
29.7
23.7
9.51 × 1011
4.73 × 1011
1.13 × 1011
9.02 × 1010
I3D (RGB) (Carreira & Zisserman, 2017)
Two-stream I3D (Carreira & Zisserman, 2017)
0.845
0.934
-
-
Table 6: Accuracy, used samples and average FLOPs per sequence at inference on the validation set
of UCF-101 (split 1).
Accuracy (%)
80
60
40
LSTM
Skip LSTM, = 0
Skip LSTM, = 10
Skip LSTM, = 10
20
0
0
50
100
150
Epochs
200
250
5
4
300
Figure 4: Accuracy evolution during the first 300 training epochs on the validation set of UCF-101
(split 1). Skip LSTM models converge much faster than the baseline LSTM.
15
Published as a conference paper at ICLR 2018
B
Q UALITATIVE RESULTS
This appendix contains additional qualitative results for the Skip RNN models.
B.1
A DDING TASK
marker value
1
marker value
1
0
0
10
20
30
time step
40
0
50
10
20
30
time step
40
50
0
10
20
30
time step
40
50
0
10
20
30
time step
40
50
0
10
20
30
time step
40
50
0
10
20
30
time step
40
50
marker value
1
marker value
1
0
0
0
10
20
30
time step
40
0
50
marker value
1
marker value
1
0
0
10
20
30
time step
40
0
50
marker value
1
marker value
1
0
0
10
20
30
time step
40
0
50
marker value
1
marker value
1
0
0
10
20
30
time step
40
0
50
Figure 5: Sample usage examples for the Skip GRU with λ = 10−5 on the adding task. Red dots
indicate used samples, whereas blue ones are skipped.
16
Published as a conference paper at ICLR 2018
B.2
F REQUENCY DISCRIMINATION TASK
1
amplitude
amplitude
1
0
1
0
25
50
75
100 125
sample
150
175
1
200
amplitude
amplitude
0
0
25
50
75
100 125
sample
150
175
amplitude
amplitude
0
25
50
75
100 125
sample
150
175
amplitude
amplitude
100 125
sample
150
175
200
0
25
50
75
100 125
sample
150
175
200
0
25
50
75
100 125
sample
150
175
200
0
25
50
75
100 125
sample
150
175
200
0
25
50
75
100 125
sample
150
175
200
1
0
0
25
50
75
100 125
sample
150
175
0
1
200
1
amplitude
1
amplitude
75
0
1
200
1
0
1
50
1
0
1
25
0
1
200
1
1
0
1
1
1
0
0
25
50
75
100 125
sample
150
175
0
1
200
Figure 6: Sample usage examples for the Skip LSTM with λ = 10−4 on the frequency discrimination task with Ts = 0.5ms. Red dots indicate used samples, whereas blue ones are skipped. The
network learns that using the first samples is enough to classify the frequency of the sine waves, in
contrast to a uniform downsampling that may result in aliasing.
17
| 2cs.AI |
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
1
Deep Sparse Subspace Clustering
Xi Peng, Jiashi Feng, Shijie Xiao, Jiwen Lu Senior Member, IEEE, Zhang Yi Fellow, IEEE,
Shuicheng Yan Fellow, IEEE,
arXiv:1709.08374v1 [cs.CV] 25 Sep 2017
Abstract—In this paper, we present a deep extension of Sparse Subspace Clustering, termed Deep Sparse Subspace Clustering
(DSSC). Regularized by the unit sphere distribution assumption for the learned deep features, DSSC can infer a new data affinity
matrix by simultaneously satisfying the sparsity principle of SSC and the nonlinearity given by neural networks. One of the appealing
advantages brought by DSSC is: when original real-world data do not meet the class-specific linear subspace distribution assumption,
DSSC can employ neural networks to make the assumption valid with its hierarchical nonlinear transformations. To the best of our
knowledge, this is among the first deep learning based subspace clustering methods. Extensive experiments are conducted on four
real-world datasets to show the proposed DSSC is significantly superior to 12 existing methods for subspace clustering.
Index Terms—Subspace clustering, low rank representation, spectral clustering, neural networks
F
1
I NTRODUCTION
S
UBSPACE clustering aims at simultaneously implicitly
finding out an underlying subspace to fit each group
of data points and performing clustering based on the
learned subspaces, which has attracted a lot of interest from
the computer vision and image processing community [1].
Most existing subspace clustering methods can be roughly
divided into following categories: algebraic methods [1], [2],
iterative methods [3], [4], statistical methods [5], [6], and
spectral clustering based methods [7], [8], [9], [10], [11].
Recently, a large number of spectral clustering based
methods have been proposed [12], [13], [14], [15], [16], [17],
[18], [19], [20], [21], [22], [23], [24], [25], [26], [27], [28],
[29], [30], [31], which first form an affinity matrix using the
linear reconstruction coefficients of the whole data set and
then obtain clustering results by applying spectral clustering
on the affinity matrix. Those methods differ from each
other mainly in their adopted priors on the coefficients.
For example, `1 -norm based sparse subspace clustering
(SSC) [12], [14] and its `0 -norm based variant [31], low
rank representation (LRR) [19], [20], and thresholding ridge
regression (TRR) [32], [33] build the affinity matrix using
the linear representation coefficients under the constraint
of `1 -, nuclear-, and `2 -norm, respectively. Formally, SSC,
LRR, TRR, as well as many of their variants learn the
representation coefficients to build the affinity matrix via:
min L(X − XC) + R(C),
C
•
•
•
•
•
(1)
X. Peng and Z. Yi are with College of Computer Science, Sichuan
University, Chengdu, 610065, China.
E-mail: pengx.gm@gmail.com; zhangyi@scu.edu.cn
J. Feng is with Department of Electrical and Computer Engineering at
National University of Singapore.
Shijie Xiao is with OmniVision Technologies Singapore Pte. Ltd., Singapore.
J. Lu is with Department of Automation, Tsinghua University, Beijing
100084, China.
S. Yan is with Department of Electrical and Computer Engineering at
National University of Singapore and 360 Artificial Intelligence Institute,
China.
Clustering
Data
Neural Network Deeply Sparse Coding
Results
Fig. 1. The flowchart of the proposed DSSC method. For a given
data set X = [x1 , x2 , · · · , xn ], we project them into the feature space
(M )
(M )
(M )
as H(M ) = [h1 , h2 , · · · , hn ] by using a set of hierarchical
nonlinear transformations and learn the self sparse representation of
input at the top layer of the neural network, where M denotes the
top layer of the neural network. Once the neural network converges,
we apply spectral clustering on the affinity matrix built by the obtained
representation like SSC. Noted that the proposed neural network is
based on a novel structure which simultaneously enjoys the sparsity of
SSC and the nonlinearity of neural networks.
where C ∈ Rn×n denotes the linear representation of
the input X ∈ Rd×n . Here, d denotes the dimension of
data and n is the number of data points. R(C) denotes
certain imposed structure prior over C, and the choice of
representation error function L(·) is usually dependent on
the distribution assumption of X, e.g. a typical loss function
is L(X − XC) = kX − XCkF .
Although those methods have achieved impressive performance for subspace clustering, they generally suffer from
the following limitations. First of all, those methods assume
that each sample can be linearly reconstructed by the whole
sample collection. However, in real-world cases, the data
may not be linearly represented by each other in the input
space. Therefore, performance of those methods usually
drop in practice. To address this problem, several recent
works [34], [35], [36], [37] have developed kernel-based approaches which have shown their effectiveness in subspace
clustering. However, kernel-based approaches are similar
to template-based approaches, whose performance heavily
depends on the choice of kernel functions. Moreover, the
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
approaches cannot give explicit nonlinear transformations,
causing difficulties in handling large-scale data sets.
Inspired by the remarkable success of deep learning in
various applications [38], [39], in this work, we propose a
new subspace clustering framework based on neural networks, termed deep sparse subspace clustering (DSSC). The
basic idea of DSSC (see Figure 1) is simple but effective. It
uses a neural network to project data into another space in
which SSC is valid to the nonlinear subspace case. Unlike
most existing subspace clustering methods, our method
simultaneously learns a set of hierarchical transformations
parametrized by a neural network and the reconstruction
coefficients to represent each mapped sample as a combination of others. Compared with kernel based approaches,
DSSC is a deep instead of shallow model which can explicitly
map samples from the input space into a latent space, with
parameters in the transformations learned in a data-driven
way. To the best of our knowledge, DSSC is the first deep
extension of SSC, which satisfies the sparsity principle of
SSC and meanwhile makes SSC valid to nonlinear subspace
case.
The contribution of this work is twofold. From the view
of subspace clustering, we show how to make it benefit from
the success of deep neural networks so that the nonlinear
subspace clustering could be achieved. From the view of
neural networks, we show that it is feasible to integrate
the advantages of existing subspace clustering methods
and deep learning to develop new unsupervised learning
algorithms.
Notations: throughout the paper, lower-case bold letters
represent column vectors and UPPER-CASE BOLD ONES
denote matrices. AT denotes the transpose of the matrix A
and I denotes an identity matrix.
2
R ELATED W ORKS
Subspace Clustering: The past decade saw an upsurge of
subspace clustering methods with various applications in
computer vision, e.g. motion segmentation [6], [14], [16],
[19], [21], [40], face clustering [12], [17], [18], [20], [22],
image processing [15], [31], [41], multi-view analysis [24],
and video analysis [36]. Particularly, among these works,
spectral clustering based methods have achieved state-ofthe-art results. The key of these methods is to learn a
satisfactory affinity matrix A in which Aij denotes the
similarity between the i-th and the j -th sample. Ideally,
Aij 6= 0 only if the corresponding data points xi and xj
are drawn from the same subspace. To this end, some recent
works (e.g. SSC [12], [14]) assume that any given sample
can be linearly reconstructed by other samples in the input
space. Based on the self-representation, an affinity matrix
(or called similarity graph) can be constructed and fed to
spectral clustering algorithms to obtain the final clustering
results. In practice, however, high-dimensional data (such
as face images) usually resides on the nonlinear manifold.
Unfortunately, linear reconstruction assumption may not be
satisfied in the original space and in this case the methods
may fail to capture the intrinsic nonlinearity of manifold.
To address this limitation, the kernel approach is used to
first project samples into a high-dimensional feature space
2
in which the representation of the whole data set is computed [34], [35], [36], [37]. After that, the clustering result
is achieved by performing traditional subspace clustering
methods in the kernel space. However, the kernel-based
methods behave like template-based approaches which usually require the prior knowledge on the data distribution
to choose a desirable kernel function. Clearly, such a prior
is hard to obtain in practice. Moreover, they cannot learn
an explicit nonlinear mapping functions from data set, thus
suffering from the scalability issue and the out-of-sample
problem [42], [43].
Unlike these classical subspace clustering approaches,
our method learns a set of explicit nonlinear mapping functions
from data set to map the input into another space, and
calculates the affinity matrix using the representation of the
samples in the new space.
Deep Learning: Aimed at learning high-level features
from inputs, deep learning has shown promising results
in numerous computer vision tasks in the scenario of
supervised learning, such as image classification [39]. In
contrast, less attention [44], [45], [46] has been paid to the
applications with unsupervised learning scheme. Recently,
some works [47], [48], [49], [50], [51], [52], [53] have devoted
to combining deep learning and unsupervised clustering
and shown impressive results over the traditional clustering
approaches. These methods share the same basic idea, i.e.,
using deep learning to learn a good representation and then
achieving clustering with existing clustering methods like
k-means. The major differences among them reside on the
neural network structure and the objective function.
Different from these works, our framework is based
on a new neural network instead of an existing network.
Moreover, our method focuses on subspace clustering rather
than clustering, which simultaneously learns the high-level
features from inputs and the self-representation in a joint
way, whereas these existing methods do not enjoy the
effectiveness of the self-expressive subspace clustering. We
believe that such a general framework is complementary
to existing shallow subspace clustering methods, since it
can adopt the loss functions and regularizations in these
methods. To the best of our knowledge, this is one of the
first several deep subspace clustering methods. It should be
pointed out that, our model is also significantly different
from [49] as below: 1) [49] performs like manifold learning,
which requires the data could be linearly reconstructed in
the input space and embeds the obtained sparse representation from input space into latent space. In contrast, our
model aims to solve the problem of nonlinear subspace
clustering, i.e. the data cannot linearly represented in the
input space. 2) In [49], sparse representation is used as a type
of priori, which keeps unchanged. In contrast, this work
dynamically seeks an good sparse representation to jointly
optimize our neural network. 3) The proposed method can
be regarded as a deep nonlinear extension of the wellknown SSC, which makes SSC handling nonlinear subspace
clustering possible.
3
D EEP S PARSE S UBSPACE C LUSTERING
In this section, we first briefly review SSC, and then present
the details of our deep subspace clustering method.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
3.1
3
Sparse Subspace Clustering
d×n
For a given data set X = [x1 , x2 , · · · , xn ] ∈ R
, SSC
seeks to linearly reconstruct the i-th sample xi using a
few of other samples. In other words, the representation
coefficients are expected to be sparse. To achieve this end,
the problem is formulated as below,
1
kxi − Xci k2F + γkci k1
s.t. cii = 0
(2)
2
where k·k1 denotes `1 -norm (i.e., the sum of absolute values
of all elements in a vector) that acts as a relaxation of `0 norm, and cii denotes the i-th element in ci . Specifically,
penalizing kci k1 encourages ci to be sparse, and enforcing
the constraint cii = 0 to avoid trivial solutions. To deal with
the optimization problem in (2), the alternating direction
method of multipliers (ADMM) [54], [55] is often used.
Once the sparse representation of the whole data set
is obtained by solving (2), an affinity matrix in SSC is
calculated as A = |C| + |C|T , based on which spectral
clustering is applied to give clustering results.
min
ci
3.2
Deep Subspace Clustering
In most existing subspace clustering methods including
SSC, each sample is encoded as a linear combination of
the whole data set. However, when dealing with highdimensional data which usually lie on nonlinear manifolds,
such methods may fail to capture the nonlinear structure,
thus leading to inferior results. To address this issue, we
propose a deep learning based method which maps given
samples using explicit hierarchical transformations in a neural
network, and simultaneously learns the reconstruction coefficients to represent each mapped sample as a combination of
others.
As shown in Figure 1, the neural network in our proposed framework consists of M + 1 stacked layers with M
nonlinear transformations, which takes a given sample x as
the input to the first layer. For ease presentation, we make
several definitions below. For the first layer of our neural
network, we define its input as h(0) = x ∈ Rd . Moreover,
for the subsequent layers, let
(m)
h(m) = g(W(m) h(m−1) + b(m) ) ∈ Rd
(3)
be the output of the m-th layer (in which m = 1, 2, · · · , M
indexes the layer), where g(·) is a nonlinear activation function, d(m) is the dimension of the output of the m-th layer,
(m)
(m−1)
(m)
W(m) ∈ Rd ×d
and b(m) ∈ Rd
denote the weights
and bias associated with the m-th layer, respectively. In
particular, given x as the input of the first layer, the output
at the top layer of our neural network is
h(M ) = g(W(M ) h(M −1) + b(M ) ).
(4)
In fact, if denoting the expression above as f (x), we can
(m)
observe that f (·) : Rd → Rd
is a nonlinear function
determined by the weights and biases of our neural network
(i.e., {W(m) , b(m) }M
m=1 ) as well as the choice of activation
function g(·). Furthermore, for n samples, we define H(M )
as the collection of the corresponding outputs given by our
neural network, i.e.
(M )
H(M ) = [h1
(M )
, h2
)
, · · · , h(M
n ].
(5)
With the above definitions, we present the objective
function of our method in the following form:
min
{W(m) ,b(m) }M
m=1 ,C
J = J1 + λJ2 .
(6)
where λ is a positive trade-off parameter, and {Ji }2i=1 are
defined below. Intuitively, the first term J1 is designed
to minimize the discrepancy between H(M ) and its selfexpressed representation. Moreover, it meanwhile regularizes C for some desired properties. To be specific, J1 can be
expressed in the form of
J1 = L(H(M ) − H(M ) C) + R(C) + F(C),
(7)
where F(C) takes the value of +∞ if C is not in some
feasible domains, and 0 otherwise. Note that, the form of
L(·), R(·), and F(·) may be adopted from many existing
subspace clustering works. In this paper, we aim to develop
a deep extension of SSC and thus take L(·) = k · k2F , R(·) =
k·k1 , F(C) = +∞ if diag(C) = 0 is violated, and F(C) = 0
otherwise.
The second part J2 is designed to remove an arbitrary
scaling factor in the latent space. In this work, we set
n
J2 =
1X
(M )
(M )
k(hi )T hi − 1k22 ,
4 i=1
(8)
Noticed that, without the above term, our neural network
may collapse in the trivial solutions like H(M ) = 0.
With {Ji }2i=1 detailed above, the optimization problem
of our proposed DSSC can be expressed as follows:
1
min kH(M ) − H(M ) Ck2F + γkCk1
Θ,C 2
n
λX
(M )
(M )
+
k(hi )T hi − 1k22
4 i=1
s.t. diag(C) = 0,
(9)
where Θ denotes the parametric neural network, i.e., Θ =
{W(m) , b(m) }M
m=1 .
3.3
Optimization
For ease of presentation, we first rewrite (9) as follows:
min
Θ,ci
n
X
1
i=1
2
(M )
khi
(M )
− Hi
+
ci k2F + γkci k1
λ
(M )
(M )
k(hi )T hi − 1k22 ,
4
(10)
(M )
where Hi
is a variant of H(M ) , which is obtained by
(M )
simply replacing hi
in H(M ) with 0.
Given n data points, DSSC simultaneously learns M
nonlinear mapping functions {W(m) , b(m) }M
m=1 and n
sparse codes {ci }n
by
solving
(10).
As
(10)
is
a multiplei=1
variable optimization problem, we employ an alternating
minimization algorithm by alternatively updating one of
variables while fixing the others.
(m)
Step 1: Fix ci and Hi , update Θ, (10) can be rewritten
as
1 (M )
λ
(M )
(M )
(M )
min khi −Hi ci k22 +αi + k(hi )T hi −1k22 , (11)
Θ 2
4
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
(M )
(M )
(M )
(M )
−Hj cj k22 +λk(hi )T hi −1k22
j6=i khj
P
where αi =
is a constant.
To solve (11), we adopt the stochastic sub-gradient descent (SGD) algorithm to obtain the parameters {W(m) ,
b(m) }M
m=1 . Moreover, we also enforce `2 -norm on the parameters to avoid overfitting [39], [56], where the regularization parameter is fixed as ϕ = 10−3 in all experiments.
Noticed that, (11) could also be solved with mini-batch SGD,
especially, when the data size is large. However, the minibatch SGD may give two issues. First, it introduces a new
hyper parameter (i.e., batch size), which increases human
effort for model selection. Second, the efficiency may be at
the cost of performance degradation [57].
(M )
Step 2: Fix {hi }n
i=1 and update ci by
1 (M )
(M )
min khi − Hi ci k2F + γkci k1 + βi ,
(12)
ci 2
P
(M )
(M )
where βi = j6=i 21 khj − Hj ci k22 + γkcj k1 is a constant. Note that, (12) is a standard `1 -minimization problem
faced in SSC, which can be solved by using many existing
`1 -solvers [58].
Step 1 and Step 2 are repeated until convergence.
After obtaining C, we construct a similarity graph via
A = |C| + |C|T and obtain the clustering results based on
A. The optimization procedure of DSSC is summarized in
Algorithm 1.
Algorithm 1 Deep Sparse Subspace Clustering
Input: A given data set X and the tradeoff parameters λ.
// Initialization:
(0) = X.
Initialize {W(m) , b(m) }M
m=1 , and H
for m = 1, 2 · · · , M do
Do forward propagation to get {H(m) }M
m=1 and C via solving (3)
and (12), respectively.
end
// Optimization
while not converge do
for i = 1, 2, · · · , n do
Randomly select a data point xi and let h0i = xi ,
for m = 1, 2 · · · , M do
(m)
Compute hi
via (3).
end
(m)
Compute ci using hi
via (3).
for m = M, M − 1 · · · , 1 do
Calculate the gradient using the SGD algorithm.
end
for m = 1, 2, · · · , M do
Update W(m) and b(m) with the gradient.
end
end
end
Output: {W(m) , b(m) }M
m=1 and C.
3.4
Discussions
Our approach DSSC can provide satisfactory subspace clustering performance befitting from following factors. First,
different from SSC, DSSC performs sparse coding in a deep
latent space instead of the original one and the latent space
is automatically learned in a data-driven manner. After
mapping input data into the latent space via the learned
transformation matrices, the transformed samples are more
favorable for linear reconstruction. Second, DSSC can also
be deemed as a deep kernel method which automatically
4
learns transformations in a data-driven way. Considering
the demonstrated effectiveness of kernel-based subspace
clustering approaches such as [35], [36], DSSC is wellexpected to show even better performance for subspace
clustering thanks to the representative capacity of deep
neural network.
It should be pointed out that the proposed DSSC adopts
similar neural network structure with deep metric learning
networks (DMLNs) [59], [60], [61], [62], i.e., a set of fully
connected layers to perform nonlinear transformation and
then perform specific task on the output of neural network.
The major differences among them are: 1) the objective
functions are different. Our method aims to segment different sample into different subspaces, whereas these metric learning networks aim to learn similarity function that
measures how similar or related two data points are; 2)
our DSSC is unsupervised, whereas DMLNs are supervised
approaches which require the label information to train
neural networks.
3.5
Implementation Details
In this section, we introduce the implementation details
of the used activation functions and the initialization of
{W(m) , bm }.
The activation functions can be chosen from various
forms. In our experiments, we use the tanh function which
is defined as follows:
g(z) = tanh(z) =
1 − e−2z
,
1 + e−2z
(13)
and the corresponding derivative is calculated as
g 0 (z) = tanh0 (z) = 1 − tanh2 (z).
(14)
Regarding the initializations of {W(m) , bm }, we initialize W(m) as a rectangular matrix with ones at the main
diagonal and zeros as other elements. Moreover, b(m) is
initialized as 0. Note that, the used networks could also be
initialized with an auto-encoder.
4
E XPERIMENTS
In this section, we compare our method with 12 popular
subspace clustering methods on four different real-world
data sets in terms of four clustering performance metrics.
4.1
Datasets and Experimental Settings
Data sets: Four different data sets are used in our experiments, i.e. COIL20 object images [63], the MNIST handwritten digital database [64], AR facial images [65], and the
BF0502 video face data set [66].
The COIL20 database contains 1,440 samples distributed
over 20 objects, where each image is with the size of 32 ×
32. The MNIST data set includes 60,000 handwritten digit
images of which the first 2,000 training images and the first
2,000 testing images are used in our experiments, where the
size of each image is 28 × 28.
The AR database is one of the most popular facial image
data sets for subspace clustering. In our experiments, we use
a widely-used subset of the AR database [67] which consists
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
of 1,400 undisguised faces evenly distributed over 50 males
and 50 females, where the size of each image is 165 × 120.
The BF0502 data set contains facial images detected from
the TV series “Buffy the Vampire Slayer”. Following [36], a
subset of BF0502 is used, which includes 17,337 faces in 229
tracks from 6 main casts. Each facial image is represented
as a 1,937-dimensional vectors extracted from 13 facial landmark points (e.g., the left and right corners of each eye). In
our experiments, we use the first 200 samples from each
category, thus resulting in 1,200 images in total.
For the purpose of nonlinear subspace clustering, we
use the following four types of features instead of raw data
from the COIL20, MNIST, and AR data sets in experiments,
i.e. dense scale-invariant feature transform (DSIFT) [68], the
histogram of oriented gradients (HOG) [69], local binary
pattern (LBP) [70], and local phase quantization (LPQ) [71].
The details of extracting these features are introduced as
follows:
•
•
•
•
DSIFT: We divide each image into multiple nonoverlapping patches and then densely sample SIFT
descriptors from each patch. The patch sizes of AR,
COIL20, and MNIST are set as 15×15, 8×8, and 4×4,
respectively. By concatenating these SIFT descriptors
extracted from each image, we obtain a feature vector
with the dimension of 11,264 (AR), 2,048 (COIL20),
and 6,272 (MNIST).
HOG: We first divide each image into multiple blocks
with two scales, i.e. 8 × 8 and 4 × 4 for AR, and 4 × 4
and 2 × 2 for MNIST and COIL20. Then, we extract
a 9-dimensional HOG feature from each block. By
concatenating these features for each image, the dimension of the feature vector are 13,770 (AR), 2,205
(MNIST), and 2,880 (COIL20) , respectively.
LBP: Like DSIFT, we divide each image into multiple
non-overlapping patches and then extract LBP features using 8 sampling points on a circle of radius 1.
Thus, we obtain a 59-dimensional LBP feature vector
from each patch. By concatenating the descriptors
of each image, we obtain a feature vector with the
dimension of 7,788 (COIL20) and 2,891 (MNIST).
LPQ: The patch size is set as 8 × 8 for COIL20 and
MNIST. For all the tested data sets, we set the size
of LPQ window as 3, 5, and 7. By concatenating the
features of all patches of each image, the dimension
of each feature is 12,288 for COIL20 and 6,912 for
MNIST.
For computational efficiency, we perform PCA to reduce the
feature dimension of all data sets to 300, by following the
setting in [14], [59]
Baseline Methods: We compare DSSC with 12 stateof-the-art subspace clustering methods, i.e. SSC [12], [14],
Kernel SSC (KSSC) [35], LRR [19], [20], low rank subspace clustering (LRSC) [16], Kernel LRR [36], least square
regression (LSR) [21], smooth representation (SMR) [17].
LSR has two variants which are denoted by LSR1 and
LSR2, respectively. KSSC and KLRR have also two variants
which are based on the RBF function (KSSC1 / KLRR1)
and the polynomial function (KSSC2 / KLRR2), respectively. Moreover, we have also used the deep autoencoder
(DAE) with SSC as a baseline to show the efficacy of our
5
method. More specifically, we adopt the pre-training and
fine-tuning strategy [72] to train a DAE that consists of
five layers with 300, 200, 150, 200, and 300 neurons. In the
experiments, we investigate the performance of DAE with
two popular nonlinear activation functions, i.e. the sigmoid
function (DAEg) and the saturating linear transfer function
(DAEs). After the DAE converging, we perform SSC on the
output of the third layer to obtain the clustering results.
For fair comparisons, we use the same `1 -solver (i.e. the
Homotopy method [58], [73]) to solve the `1 -minimization
problem in DSSC, SSC, and DAE. Noted that, our method
could also be compatible to other neural networks such as
convolutional neural networks (CNN). In experiments, we
adopt the fully connected network (FCN) instead of CNN
because the former has offered a desirable performance
in our experiments. Moreover, FCN is with fewer hyperparameters than CNN, which remarkably reduces the effort
to seek optimal value for hyper-parameters.
Experimental Settings: In our experiments, we adopt
cross-validation for selecting the optimal parameters for
all the tested methods [56]1 . More specifically, we equally
split each data set into two partitions and tune parameters
using one partition. With the tuned parameters, we repeat
each algorithm 10 times on the other partition and report
the achieved mean and standard deviation of the used
clustering performance metrics. In all the experiments, we
train a DSSC consisting of three layers, with 300, 200, and
150 neurons respectively. Moreover, we set λ = 10−3 /n and
the convergence threshold as 10−3 for DSSC and adopt early
stopping technique (w.r.t. the parameter τ ) to avoid overfitting by following [56], where n is the data size. Once the
network converges, we experimentally found that removing
the nonlinear functions could be helpful for following clustering step in inference phrase. Note that, we directly use
the tuned parameters γ (sparsity) and δ (tolerance) of SSC
for DSSC. If these two parameters are tuned specifically, the
performance of DSSC could be further improved.
Evaluation Criteria: Like [24], we adopt four popular
metrics to evaluate the clustering performance of our algorithm, i.e. accuracy (ACC) or called purity, normalized
mutual information (NMI), adjusted rand index (ARI), and
Fscore. Higher value of these metrics indicates better performance.
4.2
Comparison with state-of-the-art methods
In this section, we compare DSSC with 12 recently-proposed
subspace clustering methods on the COIL20 and the MNIST
data sets, where each data set is with four different features.
On COIL20: We first investigate the performance of
DSSC using the COIL20 data set. Tables 1–2 report the
results from which we can see that:
•
DSSC consistently outperforms other tested methods
in terms of all of the used performance metrics.
Regarding the four types of features, DSSC achieves
at least 1.86%, 2.09%, 0.96% and 3.52% relative improvement over the ACC of the best baseline, respectively.
1. The following parameters are tuned with the cross validation
technique: DSSC (µ, τ ), SSC (γ , δ ), KSSC (γ , δ ), DAE (γ , δ ), LRR (λ),
KLRR (λ), LRSC (λ), LSR (λ), and SMR (α, k).
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
6
TABLE 1
Clustering results on the COIL20 data set. Results in boldface are significantly better than the others, according to the t-test with a significance
level at 0.05.
Features
Methods
DSSC
SSC
KSSC1
KSSC2
DAEg
DAEs
LRR
KLRR1
KLRR2
LRSC
LSR1
LSR2
SMR
ACC
NMI
DSIFT
ARI
Fscore
Para.
ACC
NMI
HOG
ARI
Fscore
Para.
80.82±2.88
78.96±3.12
71.00±2.13
72.01±2.68
55.83±2.80
55.81±2.60
71.03±1.47
70.46±1.55
70.85±1.37
71.82±0.28
63.93±2.15
68.11±1.14
76.97±0.96
90.52±0.94
89.06±1.03
78.72±0.98
83.84±0.89
70.42±1.43
70.71±1.68
80.52±1.05
79.61±1.01
80.09±1.15
77.65±0.23
73.18±1.12
75.33±0.62
85.30±0.58
77.63±2.09
76.46±2.31
63.33±1.85
64.22±3.47
47.06±2.74
48.49±3.31
63.83±2.09
61.25±1.94
62.75±1.54
62.72±0.52
53.29±2.26
56.29±1.56
71.56±1.02
78.88±1.96
77.59±2.17
65.18±1.75
66.22±3.16
50.00±2.52
51.46±3.05
65.70±1.97
63.35±1.81
64.73±1.46
64.62±0.49
55.75±2.14
58.61±1.41
73.02±0.96
2−12 , 20
0.5, 0.2
10−3 , 18
10−3 , 18
0.5, 0.2
0.5, 0.2
5e-2
500
100
0.08
0.6
0.9
2−16 ,10−3
87.10±2.82
85.01±0.85
75.29±0.97
69.53±1.30
69.60±1.00
64.75±1.31
76.89±1.46
76.74±0.27
72.33±2.65
57.11±1.24
54.81±1.80
53.81±1.51
80.15±0.87
91.67±1.07
89.99±0.38
82.75±0.49
81.27±0.69
78.52±0.47
77.48±0.60
86.52±0.78
82.00±0.14
80.98±1.21
69.91±0.73
64.44±0.94
63.00±1.22
85.93±0.6
82.56±1.26
81.13±1.08
66.46±1.43
61.16±1.83
59.38±0.79
56.81±1.12
70.79±1.73
69.43±0.48
63.11±2.88
46.27±1.57
42.28±1.55
42.07±1.5
73.51±1.06
83.51±2.12
82.08±1.02
68.20±1.33
63.32±1.69
61.63±0.74
59.13±1.06
72.39±1.62
70.96±0.45
65.07±2.68
49.20±1.48
45.35±1.44
45.19±1.42
74.87±1.01
2−12 , 30
0.5, 0.1
10−2 , 18
10−2 , 18
0.5, 0.1
0.5, 0.1
5e-3
10
5
0.01
0.5
0.3
2−16 ,10−3
TABLE 2
Clustering results on the COIL20 data set. Results in boldface are significantly better than the others, according to the t-test with a significance
level at 0.05.
Features
Methods
DSSC
SSC
KSSC1
KSSC2
DAEg
DAEs
LRR
KLRR1
KLRR2
LRSC
LSR1
LSR2
SMR
ACC
NMI
LBP
ARI
Fscore
Para.
ACC
NMI
LPQ
ARI
Fscore
Para.
72.89±1.41
70.17±0.65
69.33±1.97
70.42±1.13
40.96±2.18
40.68±1.13
71.60±4.02
65.83±0.31
70.10±1.27
62.96±0.61
70.24±2.90
70.54±3.26
71.93±1.35
84.32±0.79
82.66±0.19
80.65±0.86
83.67±0.69
53.54±0.89
52.12±0.92
84.45±1.78
77.34±0.30
79.58±0.13
73.38±0.79
82.40±1.41
81.63±1.16
81.17±0.39
67.31±1.96
64.19±0.60
61.15±1.91
65.28±1.23
26.27±1.33
23.67±1.30
65.47±5.68
56.41±0.50
62.91±0.51
53.31±1.06
64.54±2.85
63.71±2.58
63.54±1.41
69.01±1.85
66.07±0.58
63.18±1.79
68.03±1.16
30.57±1.22
28.26±1.10
66.29±5.21
58.60±0.47
64.82±0.47
55.67±1.01
67.33±2.69
66.59±2.41
66.39±1.32
2−13 , 40
10−3 , 10−2
1, 16
1, 16
10−3 , 10−2
10−3 , 10−2
0.5
30
1000
0.04
1
0.6
2−16 , 10−3
78.12±2.09
74.60±0.81
68.49±2.38
69.24±2.33
62.19±0.90
59.64±2.46
69.00±1.09
69.43±1.46
65.33±2.48
66.38±0.50
66.97±1.68
65.25±1.55
70.56±0.57
85.38±0.77
84.21±0.49
79.28±1.27
79.52±0.93
72.04±0.54
67.44±1.06
80.31±0.88
77.34±0.53
76.41±1.13
78.73±0.58
74.42±0.62
73.81±1.29
80.68±0.41
71.35±1.34
67.69±0.83
59.06±2.37
61.07±1.72
51.51±0.75
44.90±1.81
60.12±1.51
57.01±1.02
54.22±2.11
58.81±0.97
55.48±1.52
54.34±1.65
61.68±0.49
72.87±1.25
69.35±0.79
61.23±2.21
63.17±1.62
54.15±0.72
47.98±1.65
62.29±1.41
59.24±0.96
56.69±1.94
60.99±0.91
57.74±1.43
56.66±1.56
63.68±0.45
2−12 , 60
10−3 , 0.1
0.1, 12
0.1, 12
10−3 , 0.1
10−3 , 0.1
0.1
500
100
0.08
0.2
0.3
2−16 , 10−3
TABLE 3
Clustering results on the MNIST data set. Results in boldface are significantly better than the others, according to the t-test with a significance
level at 0.05.
Features
Methods
DSSC
SSC
KSSC1
KSSC2
DAEg
DAEs
LRR
KLRR1
KLRR2
LRSC
LSR1
LSR2
SMR
•
ACC
NMI
DSIFT
ARI
Fscore
Para.
ACC
NMI
HOG
ARI
Fscore
Para.
72.65±0.00
62.45±0.00
50.90±0.00
60.80±0.00
52.55±0.00
42.28±1.19
63.20±0.00
57.05±0.00
22.63±0.89
59.30±0.00
63.50±0.00
63.55±0.00
69.15±0.00
70.42±0.00
65.75±0.00
49.75±0.00
63.96±0.00
58.36±0.00
48.70±0.56
68.34±0.00
57.97±0.00
12.55±1.37
58.84±0.00
60.39±0.00
60.53±0.00
68.90±0.00
61.80±0.00
53.75±0.00
35.28±0.00
50.81±0.00
40.98±0.00
31.70±0.31
54.11±0.00
44.63±0.00
9.02±0.46
46.90±0.00
49.02±0.00
49.14±0.00
59.17±0.00
65.79±0.00
58.81±0.00
41.80±0.00
56.26±0.00
47.48±0.00
39.85±0.22
59.48±0.00
50.70±0.00
22.98±0.14
52.37±0.00
54.29±0.00
54.39±0.00
63.40±0.00
2−13 , 20
1, 10−2
10−3 , 10
10−3 , 10
1, 10−2
1, 10−2
0.05
3000
1000
0.1
0.1
0.4
2−16 , 10−3
78.10±0.00
77.35±0.00
66.90±0.00
68.00±0.00
23.36±0.66
23.07±1.14
73.30±0.00
72.15±0.00
73.55±0.00
61.20±0.00
58.42±0.09
60.40±0.02
77.22±0.05
77.51±0.00
75.70±0.00
70.20±0.00
70.74±0.00
11.56±0.78
10.48±0.52
74.49±0.00
70.95±0.00
73.30±0.00
59.65±0.00
56.41±0.07
57.78±0.00
77.25±0.00
68.72±0.00
66.90±0.00
56.79±0.00
57.69±0.00
5.83±0.28
4.91±0.31
63.20±0.00
61.38±0.00
63.69±0.00
47.05±0.00
44.85±0.12
46.45±0.00
66.85±0.00
72.03±0.00
70.23±0.00
61.35±0.00
62.12±0.00
15.89±0.32
15.02±0.29
67.12±0.00
65.42±0.00
67.50±0.00
52.59±0.00
50.79±0.11
51.98±0.00
71.11±0.00
2−17 , 30
10, 10−2
10−2 , 12
10−2 , 12
10, 10−2
10, 10−2
0.01
30
3
0.01
0.2
0.1
2−16 , 10−3
SSC usually outperforms DAEs and DAEg, whereas
our DSSC method consistently outperforms SSC in
all the settings. This shows that it is hard to achieve
a desirable performance by simply introducing deep
learning into subspace clustering since unsupervised
deep learning is an open challenging issue [44].
On MNIST: We also investigate the performance of
DSSC by using the MNIST data set.
Tables 6–4 show the result, from which we obverse
that the ACC of DSSC with the DSIFT feature is 72.65%
which improves SSC by 10.20% and the best baseline al-
gorithm by 3.50%. With respect to the other three features,
the improvement of DSSC comparing with all the baseline
approaches is also significant, which is 1.82%, 1.02%, and
1.71% in terms of ARI. It should be pointed out that, all the
tested methods perform very stable on this data set, whose
standard deviations on these four performance metrics are
close to 0.
4.3
Deep Model vs. Shallow Models
In this section, we investigate the influence of the depth of
DSSC using the AR data set with DSIFT and HOG features.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
7
TABLE 4
Clustering results on the MNIST data set. Results in boldface are significantly better than the others, according to the t-test with a significance
level at 0.05.
Features
Methods
DSSC
SSC
KSSC1
KSSC2
DAEg
DAEs
LRR
KLRR1
KLRR2
LRSC
LSR1
LSR2
SMR
ACC
NMI
LBP
ARI
Fscore
Para.
ACC
NMI
LPQ
ARI
Fscore
Para.
61.70±0.00
59.75±0.00
58.50±0.00
57.70±0.00
36.20±0.00
32.20±0.00
55.70±0.00
54.12±0.06
53.75±0.00
42.45±0.00
53.12±0.05
52.93±0.04
49.90±0.00
54.23±0.00
53.83±0.00
53.96±0.00
54.27±0.00
27.81±0.00
24.85±0.00
45.70±0.00
50.98±0.01
50.70±0.00
35.60±0.00
45.81±0.05
45.65±0.03
44.29±0.00
44.54±0.00
43.52±0.00
41.58±0.00
41.95±0.00
16.68±0.00
14.75±0.00
37.77±0.00
37.84±0.07
37.05±0.00
23.42±0.00
35.27±0.01
34.98±0.05
32.16±0.00
50.25±0.00
49.02±0.00
47.67±0.00
47.96±0.00
25.42±0.00
23.39±0.00
44.58±0.00
44.25±0.06
43.55±0.00
31.43±0.00
41.87±0.01
41.61±0.05
39.18±0.00
2−15 , 40
0.1, 0.01
0.1, 10
0.1, 10
0.1, 0.01
0.1, 0.01
0.5
1000
300
0.03
0.1
0.1
2−14 , 10−3
65.04±0.02
62.35±0.00
44.00±0.00
54.99±0.03
30.61±0.51
34.10±0.00
52.15±0.00
55.60±0.00
56.75±0.00
53.20±0.00
52.60±0.00
53.25±0.05
48.90±0.00
54.85±0.01
53.86±0.00
34.97±0.00
51.07±0.01
22.05±0.32
22.71±0.02
50.63±0.00
51.66±0.00
51.69±0.00
42.49±0.00
46.93±0.00
47.57±0.11
43.43±0.00
46.38±0.00
44.67±0.00
23.49±0.00
36.62±0.01
12.19±0.05
12.41±0.01
37.86±0.00
38.42±0.00
38.87±0.00
32.03±0.00
34.93±0.00
35.54±0.06
30.18±0.00
52.04±0.00
50.42±0.00
31.26±0.00
43.21±0.01
22.07±0.05
22.43±0.01
44.63±0.00
44.77±0.00
45.16±0.00
39.04±0.00
41.71±0.00
42.28±0.06
37.62±0.00
2−16 , 70
1, 0.01
10−3 , 16
10−3 , 16
10, 0.01
10, 0.01
0.5
3000
1000
0.05
1
1
2−14 , 10−3
TABLE 5
Deep vs. Shallow Models on the AR data set. Results in boldface are significantly better than the others, according to the t-test with a significance
level at 0.05.
Features
Methods
ACC
NMI
DSIFT
ARI
Fscore
Para.
ACC
NMI
HOG
ARI
Fscore
Para.
DSSC(M=2)
DSSC (M=1)
SSC
KSSC1
KSSC2
DAEg
DAEs
LRR
KLRR1
KLRR2
LRSC
LSR1
LSR2
SMR
85.38±1.08
83.81±1.72
74.83±1.27
70.27±1.66
78.28±1.78
74.37±1.20
72.65±0.91
82.67±1.00
79.92±1.52
23.08±0.36
83.55±1.20
82.43±1.31
82.45±1.58
71.07±1.91
95.17±0.17
94.57±0.45
89.91±0.38
87.29±0.53
91.55±0.39
89.53±0.43
88.54±0.52
93.48±0.33
91.56±0.51
52.01±0.62
92.84±0.37
92.69±0.49
92.64±0.42
87.01±0.52
82.15±0.63
81.23±1.94
66.43±1.44
58.61±1.78
71.13±1.44
65.42±1.56
62.23±1.81
77.33±1.37
71.08±2.13
5.31±0.24
78.33±1.39
74.94±1.87
74.49±1.80
60.82±2.22
82.35±0.62
81.42±1.92
66.81±1.42
59.08±1.76
71.44±1.43
65.81±1.54
62.67±1.78
77.60±1.35
71.42±2.10
6.73±0.24
78.57±1.38
75.22±1.85
74.77±1.78
61.26±2.19
2−11 , 50
2−11 , 30
10−2 , 10−3
1, 18
1, 18
10−2 , 10−3
10−2 , 10−3
0.1
300
100
0.06
0.3
0.7
2−16 , 10−2
85.05±1.53
81.90±0.96
81.65±1.18
83.12±0.90
83.22±1.34
74.67±1.25
73.32±1.31
83.00±1.36
83.92±1.26
76.07±1.69
83.42±1.43
83.32±1.70
83.65±1.07
81.38±0.73
94.36±0.43
91.93±0.35
92.48±0.41
93.07±0.34
92.71±0.32
89.07±0.49
88.17±0.43
93.27±0.46
93.00±0.45
88.78±0.74
92.67±0.48
92.45±0.49
92.45±0.45
91.75±0.27
78.98±1.58
71.87±1.97
74.23±1.76
75.68±1.37
74.56±1.06
63.77±1.52
61.12±1.42
77.34±3.21
77.49±1.49
63.93±2.67
73.86±1.73
73.11±2.24
73.24±1.77
72.51±0.85
79.21±1.56
72.17±1.95
74.52±1.74
75.94±1.36
74.84±1.05
64.17±1.50
61.56±1.40
77.61±3.16
77.73±1.47
64.34±2.63
74.15±1.71
73.40±2.21
73.54±1.75
72.81±0.84
2−12 , 30
2−12 , 20
0.5, 10−3
10−2 , 20
10−2 , 20
0.5, 10−3
0.5, 10−3
0.01
100
5
0.02
0.8
1
2−15 , 10−2
More specifically, we report the performance of DSSC with
two hidden layers (M = 2) and one hidden layer (M = 1),
respectively. In the case of M = 1, the number of hidden
neurons is also set as 150. Note that, KSSC1 and KSSC2
can be regarded as two shallow models of SSC with one
nonlinear hidden layer.
Table 5 shows the clustering results of the methods, as
well as the tuned parameters. We observe that our DSSC
(M = 2) consistently outperform the shallow models in
terms of all of these evaluation metrics. The results also
verify our claim and motivation, i.e. our deep model DSSC
significantly benefit from deep learning.
4.4
Influence of Different Activation Functions
In this section, we investigate the influence of different
nonlinear activation functions in DSSC. The investigated
functions are sigmoid, non-saturating sigmoid (nssigmoid), and
the rectified linear unit (relu) [74]. We carry out experiment
on the BF0502 data set which contains facial images detected
from the TV series “Buffy the Vampire Slayer”. Following [36], a subset of BF0502 is used, which includes 17,337
faces in 229 tracks from 6 main casts. Each facial image is
represented as a 1,937-dimensional vectors extracted from
13 facial landmark points (e.g., the left and right corners of
each eye). In our experiments, we use the first 200 samples
from each category, thus resulting in 1,200 images in total.
TABLE 6
The influence of different activation functions of DSSC on the BF0502
database. DSSC (t), DSSC (s), DSSC (n), and DSSC (r) denote DSSC
with the tanh, sigmoid, nssigmoid, and relu function, respectively.
Methods
DSSC (t)
DSSC (s)
DSSC (n)
DSSC (r)
SSC
KSSC1
KSSC2
DAEg
DAEs
LRR
KLRR1
KLRR2
LRSC
LSR1
LSR2
SMR
Accuracy
79.50
82.67
75.08
80.08
79.50
74.50
77.83
55.50
21.67
78.17
75.33
75.00
69.17
67.50
77.00
76.00
NMI
71.02
79.01
67.72
75.60
74.83
71.99
69.89
38.16
6.07
74.89
66.60
69.32
60.60
57.53
59.91
74.69
ARI
65.11
71.69
59.17
65.67
62.37
61.95
70.65
30.69
0.85
70.57
56.83
68.35
53.28
51.36
56.27
58.09
Fscore
71.09
66.55
72.11
72.11
69.15
68.85
70.55
43.15
28.65
70.58
64.07
74.16
61.71
60.19
63.60
71.87
Para.
−13
2
,90
2−17 ,60
2−17 ,10
2−17 ,10
1,0.2
0.1,12
0.1,12
0.01
3
3
0.01
1.00
0.50
2−16 ,1e-02
From Table 6, we can observe that DSSC with different
activation functions outperforms SSC by a considerable
performance margin. With the sigmoid function, DSSC is
about 3.17%, 4.18%, 9.32%, and 2.96% higher than SSC in
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
8
11000
450
9900
400
[5]
8800
300
Loss
6600
250
5500
200
4400
150
3300
Time Cost (Second)
350
7700
50
1100
[7]
50
100
150
[8]
0
200
Epoch
Fig. 2. Convergence curve and time cost of DSSC. The left y-axis
indicates the loss at each epoch and the right one is the total time cost
taken by our method.
terms of Accuracy, NMI, ARI, and Fscore, respectively. It is
worth noting that although tanh is not the best activation
function, it is more stable than the other three activation
functions in our experiments. Thus, we use the tanh function
as the activation function for comparisons as shown in the
above sections.
4.5
[6]
100
2200
0
0
[4]
Convergence Analysis and Time Cost
In this section, we examine the convergence speed and time
cost of our DSSC on the BF0502 data set. From Figure 2, we
can see that the loss of DSSC generally keeps unchanged
after 90–100 epochs. For each epoch, DSSC takes about 2.2
seconds to obtain results on a macbook with a 2.6GHz Intel
Core i5 CPU and 8GB memory. Like other deep learning
based methods, the computational cost of DSSC can be
remarkably reduced by GPU.
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
5
C ONCLUSION
In this paper, we proposed a new deep learning based
framework for simultaneous data representation learning
and subspace clustering. Based on such deep subspace
clustering framework, we further devised a new method,
i.e. DSSC. Experimental results on the facial, object, and
handwritten digit image databases data sets show the
efficacy of DSSC in terms of four performance evaluation metrics. In the future, we plan to investigate the
performance of our proposed framework when adopting
other loss/regularization functions, and extend our proposed framework for other applications such as weaklysupervised learning.
R EFERENCES
[1]
[2]
[3]
R. Vidal, Y. Ma, and S. Sastry, “Generalized principal component
analysis (GPCA),” IEEE Trans. Pattern Anal. Mach. Intell., vol. 27,
no. 12, pp. 1945–1959, 2005. 1
J. P. Costeira and T. Kanade, “A multibody factorization method
for independently moving objects,” Int. J. Comput. Vis., vol. 29,
no. 3, pp. 159–179, 1998. 1
P. S. Bradley and O. L. Mangasarian, “k-plane clustering,” J. Global
Optim., vol. 16, no. 1, pp. 23–32, 2000. 1
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
L. Lu and R. Vidal, “Combined central and subspace clustering for
computer vision applications,” in Proc. of 23th Int. Conf. Machine
Learn., Pittsburgh, USA, Jun. 2006, pp. 593–600. 1
Y. Ma, H. Derksen, W. Hong, and J. Wright, “Segmentation of
multivariate mixed data via lossy data coding and compression,”
IEEE Trans. Pattern Anal. Mach. Intell., vol. 29, no. 9, pp. 1546–1562,
2007. 1
S. Rao, R. Tron, R. Vidal, and Y. Ma, “Motion segmentation via robust subspace separation in the presence of outlying, incomplete,
or corrupted trajectories,” in Proc. of 21th IEEE Conf. Comput. Vis.
and Pattern Recognit., Anchorage, AL, Jun. 2008, pp. 1–8. 1, 2
A. Y. Ng, M. I. Jordan, and Y. Weiss, “On spectral clustering:
Analysis and an algorithm,” in Proc. of 14th Adv. in Neural Inf.
Process. Syst., Vancouver, Canada, Dec. 2001, pp. 849–856. 1
J. B. Shi and J. Malik, “Normalized cuts and image segmentation,”
IEEE Trans. Pattern Anal. Mach. Intell., vol. 22, no. 8, pp. 888–905,
2000. 1
A. Y. Yang, J. Wright, Y. Ma, and S. S. Sastry, “Unsupervised segmentation of natural images via lossy data compression,” Comput.
Vis. Image Underst., vol. 110, no. 2, pp. 212–225, May 2008. 1
G. L. Chen and G. Lerman, “Spectral curvature clustering (scc),”
Int. J. of Comput. Vision, vol. 81, no. 3, pp. 317–330, 2009. 1
F. Nie, Z. Zeng, T. I. W., D. Xu, and C. Zhang, “Spectral embedded
clustering: A framework for in-sample and out-of-sample spectral
clustering,” IEEE Trans. Neural. Netw., vol. 22, no. 11, pp. 1796–
1808, 2011. 1
E. Elhamifar and R. Vidal, “Sparse subspace clustering,” in Proc. of
22th IEEE Conf. Comput. Vis. and Pattern Recognit., Miami, FL, Jun.
2009, pp. 2790–2797. 1, 2, 5
——, “Sparse manifold clustering and embedding,” in Proc. of 24th
Adv. in Neural Inf. Process. Syst., Granada, Spain, Dec. 2011, pp. 55–
63. 1
——, “Sparse subspace clustering: Algorithm, theory, and applications,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 35, no. 11, pp.
2765–2781, 2013. 1, 2, 5
J. Feng, Z. Lin, H. Xu, and S. Yan, “Robust subspace segmentation
with block-diagonal prior,” in Proc. of 27th IEEE Conf. Comput. Vis.
and Pattern Recognit., Columbus, OH, Jun. 2014, pp. 3818–3825. 1,
2
P. Favaro, R. Vidal, and A. Ravichandran, “A closed form solution
to robust subspace estimation and clustering,” in Proc. of 24th IEEE
Conf. Comput. Vis. and Pattern Recognit., Colorado Springs, CO, Jun.
2011, pp. 1801–1807. 1, 2, 5
H. Hu, Z. Lin, J. Feng, and J. Zhou, “Smooth representation
clustering,” in Proc. of 27th IEEE Conf. Comput. Vis. and Pattern
Recognit., Columbus, OH, Jun. 2014, pp. 3834–3841. 1, 2, 5
P. Ji, M. Salzmann, and H. Li, “Shape interaction matrix revisited
and robustified: Efficient subspace clustering with corrupted and
incomplete data,” in Proc. of 15th IEEE Conf. Comput. Vis., Santiago,
Chile, Dec. 2015. 1, 2
G. Liu, Z. Lin, S. Yan, J. Sun, Y. Yu, and Y. Ma, “Robust recovery
of subspace structures by low-rank representation,” IEEE Trans.
Pattern Anal. Mach. Intell., vol. 35, no. 1, pp. 171–184, 2013. 1, 2, 5
G. Liu, Z. Lin, and Y. Yu, “Robust subspace segmentation by lowrank representation,” in Proc. of 27th Int. Conf. Mach. Learn., Haifa,
Israel, Jun. 2010, pp. 663–670. 1, 2, 5
C. Lu, H. Min, Z. Zhao, L. Zhu, D. Huang, and S. Yan, “Robust
and efficient subspace segmentation via least squares regression,”
in Proc. of 12th Eur. Conf. Comput. Vis., Florence, Italy, Oct. 2012,
pp. 347–360. 1, 2, 5
R. Vidal and P. Favaro, “Low rank subspace clustering (LRSC),”
Pattern Recognit. Lett., vol. 43, pp. 47 – 61, 2014. 1, 2
M. Soltanolkotabi, E. Elhamifar, and E. J. Candes, “Robust subspace clustering,” Ann. Stat., vol. 42, no. 2, pp. 669–699, 2014. 1
C. Zhang, H. Fu, S. Liu, G. Liu, and X. Cao, “Low-Rank Tensor
Constrained Multiview Subspace Clustering,” in Proc. of 21th Int.
Conf. Comput. Vis. Santiago: IEEE, Dec. 2015, pp. 1582–1590. 1, 2,
5
R. He, Y. Zhang, Z. Sun, and Q. Yin, “Robust subspace clustering
with complex noise,” IEEE Trans. on Image Process., vol. 24, no. 11,
pp. 4001–4013, Nov. 2015. 1
L. Zhang, W. Zuo, and D. Zhang, “LSDT: Latent Sparse Domain
Transfer Learning for Visual Adaptation,” IEEE Trans. Image Process., vol. 25, no. 3, pp. 1177–1191, 2016. 1
C.-M. Lee and L.-F. Cheong, “Minimal basis subspace representation: A unified framework for rigid and non-rigid motion
segmentation,” Int. J. of Comput. Vision, pp. 1–25, 2016. 1
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
[28] R. Liu, Z. Lin, F. D. la Torre, and Z. Su, “Fixed-rank representation
for unsupervised visual learning,” in Proc. of 25th IEEE Conf.
Comput. Vis. and Pattern Recognit., Providence, RI, Jun. 2012, pp.
598–605. 1
[29] B. Cheng, J. Yang, S. Yan, Y. Fu, and T. Huang, “Learning with
`1 -graph for image analysis,” IEEE Trans. on Image Process., vol. 19,
no. 4, pp. 858–866, 2010. 1
[30] C. You, D. P. Robinson, and R. Vidal, “Scalable sparse subspace
clustering by orthogonal matching pursuit,” in Proc. of 29th IEEE
Conf. Comput. Vis. and Pattern Recognit., Las Vegas, NV, Jun. 2016,
pp. 3918–3927. 1
[31] Y. Yang, J. Feng, N. Jojic, J. Yang, and T. S. Huang, “L0-sparse
subspace clustering,” in Proc. of 14th Euro. Conf. Comput. Vis.,
Amsterdam, Netherlands, Oct. 2016, pp. 731–747. 1, 2
[32] X. Peng, Z. Yi, and H. Tang, “Robust subspace clustering via
thresholding ridge regression,” in Proc. of 29th AAAI Conference on
Artificial Intelligence, Austin Texas, USA, Jan. 2015, pp. 3827–3833.
1
[33] X. Peng, Z. Yu, Z. Yi, and H. Tang, “Constructing the l2-graph for
robust subspace learning and subspace clustering,” IEEE Trans.
Cybern., vol. 47, no. 4, pp. 1053–1066, Apr. 2017. 1
[34] V. Patel, H. V. Nguyen, and R. Vidal, “Latent space sparse subspace
clustering,” in Proc. of 14th IEEE Conf. Comput. Vis., Sydney, VIC,
Dec. 2013, pp. 225–232. 1, 2
[35] V. Patel and R. Vidal, “Kernel sparse subspace clustering,” in Proc.
of IEEE Int. Conf. on Image Process., Paris, Oct. 2014, pp. 2849–2853.
1, 2, 4, 5
[36] S. Xiao, M. Tan, D. Xu, and Z. Dong, “Robust kernel low-rank
representation,” IEEE Trans. Neural. Netw. Learn. Syst., vol. PP,
no. 99, pp. 1–1, 2015. 1, 2, 4, 5, 7
[37] M. Yin, Y. Guo, J. Gao, Z. He, and S. Xie, “Kernel sparse subspace
clustering on symmetric positive definite manifolds,” in Proc. of
29th IEEE Conf. Comput. Vis. and Pattern Recognit., Jun. 2016, pp.
5157–5164. 1, 2
[38] G. Hinton, L. Deng, D. Yu, G. E. Dahl, A.-r. Mohamed, N. Jaitly,
A. Senior, V. Vanhoucke, P. Nguyen, T. N. Sainath et al., “Deep
neural networks for acoustic modeling in speech recognition: The
shared views of four research groups,” Signal Processing Magazine,
IEEE, vol. 29, no. 6, pp. 82–97, 2012. 2
[39] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,” in Proc. of 25th
Adv. in Neural Inf. Process. Syst., Lake Tahoe, CA, Dec. 2012, pp.
1097–1105. 2, 4
[40] B. Poling and G. Lerman, “A new approach to two-view motion
segmentation using global dimension minimization,” Int. J. of
Comput. Vision, vol. 108, no. 3, pp. 165–185, 2014. 2
[41] C. Ding, X. He, H. Zha, M. Gu, and H. Simon, “A min-max cut
algorithm for graph partitioning and data clustering,” in Proc. of
1st IEEE Int. Conf. on Data Mining, San Jose, CA, Nov. 2001, pp.
107–114. 2
[42] X. Peng, L. Zhang, and Z. Yi, “Scalable sparse subspace clustering,” in Proc. of 26th IEEE Conf. Comput. Vis. and Pattern Recognit.,
Portland, OR, Jun. 2013, pp. 430–437. 2
[43] X. Peng, H. Tang, L. Zhang, Z. Yi, and S. Xiao, “A unified
framework for representation-based subspace clustering of out-ofsample and large-scale data,” IEEE Transactions on Neural Networks
and Learning Systems, vol. PP, no. 99, pp. 1–14, 2015. 2
[44] Y. Bengio, A. Courville, and P. Vincent, “Representation learning:
A review and new perspectives,” IEEE Trans. Pattern Anal. Mach.
Intell., vol. 35, no. 8, pp. 1798–1828, Aug. 2013. 2, 5
[45] H. Lee, R. Grosse, R. Ranganath, and A. Y. Ng, “Unsupervised
learning of hierarchical representations with convolutional deep
belief networks,” Commun. ACM, vol. 54, no. 10, pp. 95–103, Oct.
2011. 2
[46] Z. Y. Wang, Q. Ling, and T. S. Huang, “Learning deep l0 encoders,”
in Proc. of 30th AAAI Conf. Artif. Intell., Feb. 2016, pp. 2194–2200. 2
[47] P. Huang, Y. Huang, W. Wang, and L. Wang, “Deep embedding
network for clustering,” in Proc. of 22nd Int. Conf. Pattern Recognit.,
Stockholm, Sweden, Aug. 2014, pp. 1532–1537. 2
[48] J. Xie, R. Girshick, and A. Farhadi, “Unsupervised deep embedding for clustering analysis,” in Proc. of 33th Int. Conf. Mach. Learn.,
New York, Jun. 2016. 2
[49] X. Peng, S. Xiao, J. Feng, W. Yau, and Z. Yi, “Deep subspace
clustering with sparsity prior,” in Proc. of 25th Int. Joint Conf. Artif.
Intell., New York, NY, USA, Jul. 2016, pp. 1925–1931. 2
[50] Z. Wang, S. Chang, J. Zhou, M. Wang, and T. S. Huang, “Learning
a task-specific deep architecture for clustering,” in Proc. of SIAM
9
[51]
[52]
[53]
[54]
[55]
[56]
[57]
[58]
[59]
[60]
[61]
[62]
[63]
[64]
[65]
[66]
[67]
[68]
[69]
[70]
[71]
[72]
Int. Conf. on Data Mining, Miami, Florida, May 2015, pp. 369–377.
2
J. Yang, D. Parikh, and D. Batra, “Joint unsupervised learning of
deep representations and image clusters,” in Proc. of 29th IEEE
Conf. Comput. Vis. and Pattern Recognit., 2016. 2
Y. Chen, L. Zhang, and Z. Yi, “Subspace clustering using a lowrank constrained autoencoder,” Information Sciences, 2017. 2
P. Ji, T. Zhang, H. Li, M. Salzmann, and I. Reid, “Deep subspace
clustering networks,” arXiv preprint arXiv:1709.02508, 2017. 2
S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein, “Distributed
optimization and statistical learning via the alternating direction
method of multipliers,” Found. Trends Mach. Learn., vol. 3, no. 1,
pp. 1–122, Jan. 2011. 3
Z. Lin, R. Liu, and Z. Su, “Linearized alternating direction method
with adaptive penalty for low-rank representation,” in Proc. of 24th
Adv. in Neural Inf. Process. Syst., Grendada, Spain, Dec. 2011, pp.
612–620. 3
G. Montavon, G. B. Orr, and K.-R. Müller, Eds., Neural Networks:
Tricks of the Trade, Reloaded, 2nd ed., ser. Lecture Notes in Computer
Science (LNCS). Springer, 2012, vol. 7700. 4, 5
P. Goyal, P. Dollár, R. Girshick, P. Noordhuis, L. Wesolowski,
A. Kyrola, A. Tulloch, Y. Jia, and K. He, “Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour,” ArXiv e-prints, Jun.
2017. 4
A. Yang, A. Ganesh, S. Sastry, and Y. Ma, “Fast l1-minimization algorithms and an application in robust face recognition: A review,”
EECS Department, University of California, Berkeley, Tech. Rep.
UCB/EECS-2010-13, Feb. 2010. 4, 5
J. Hu, J. Lu, and Y.-P. Tan, “Discriminative deep metric learning for
face verification in the wild,” in Proc. of 27th IEEE Conf. Comput.
Vis. and Pattern Recognit., Columbus, OH, Jun. 2014, pp. 1875–1882.
4, 5
F. Schroff, D. Kalenichenko, and J. Philbin, “Facenet: A unified
embedding for face recognition and clustering,” in Proc. of 28th
IEEE Conf. Comput. Vis. and Pattern Recognit., Boston, MA, Jun.
2015, pp. 815–823. 4
J. Wang, Y. Song, T. Leung, C. Rosenberg, J. Wang, J. Philbin,
B. Chen, and Y. Wu, “Learning fine-grained image similarity with
deep ranking,” in Proc. of 27th IEEE Conf. Comput. Vis. and Pattern
Recognit., Washington, DC, USA, 2014, pp. 1386–1393. 4
D. Yi, Z. Lei, S. Liao, and S. Z. Li, “Deep metric learning for
person re-identification,” in Proc. of 22nd Int Conf. Pattern Recognit.,
Stockholm, Sweden, Aug. 2014, pp. 34–39. 4
S. A. Nene, S. K. Nayar, H. Murase et al., “Columbia object image
library (coil-20),” Technical Report CUCS-005-96, Tech. Rep., 1996.
4
Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based
learning applied to document recognition,” Proc. of IEEE, vol. 86,
no. 11, pp. 2278–2324, Nov. 1998. 4
A. Martinez, “The AR face database,” CVC Technical Report, vol. 24,
1998. 4
J. Sivic, M. Everingham, and A. Zisserman, “Who are you? learning person specific classifiers from video,” in Proc. of 22th
IEEE Conf. Comput. Vis. and Pattern Recognit., Miami, FL, Jun. 2009,
pp. 1145–1152. 4
L. Zhang, M. Yang, and X. Feng, “Sparse representation or collaborative representation: Which helps face recognition?” in Proc. of
13th IEEE Int. Conf. on Comput. Vis., Barcelona, Spain, Nov. 2011,
pp. 471–478. 4
D. Lowe, “Distinctive image features from scale-invariant keypoints,” Int. J. Comput. Vis., vol. 60, no. 2, pp. 91–110, 2004. 5
N. Dalal and B. Triggs, “Histograms of oriented gradients for
human detection,” in Proc. of 18th IEEE Conf. Comput. Vis. and
Pattern Recognit., vol. 1, San Diego, CA, Jun. 2005, pp. 886–893
vol. 1. 5
T. Ahonen, A. Hadid, and M. Pietikainen, “Face description with
local binary patterns: Application to face recognition,” IEEE Trans.
Pattern Anal. Mach. Intell., vol. 28, no. 12, pp. 2037–2041, Dec. 2006.
5
V. Ojansivu and J. Heikkilä, “Blur insensitive texture classification using local phase quantization,” in Image and signal process.
Springer, 2008, pp. 236–243. 5
G. E. Hinton and R. R. Salakhutdinov, “Reducing the dimensionality of data with neural networks,” Science, vol. 313, no. 5786, pp.
504–507, 2006. 5
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
[73] M. R. Osborne, B. Presnell, and B. A. Turlach, “A new approach
to variable selection in least squares problems,” SIAM J. Numer.
Anal., vol. 20, no. 3, pp. 389–403, 2000. 5
[74] V. Nair and G. E. Hinton, “Rectified linear units improve restricted
boltzmann machines,” in Proc. of 27th Int. Conf. Mach. Learn., Haifa,
Israel, Jun. 2010, pp. 807–814. 7
10
| 1cs.CV |
1
arXiv:1712.02167v1 [cs.CR] 6 Dec 2017
Cryptanalysis of a public key encryption scheme
based on QC-LDPC and QC-MDPC codes
Vlad Dragoi∗† and Hervé Talé Kalachi ‡
∗ Faculty of Exact Sciences, “Aurel Vlaicu” University of Arad, Romania
Email: ∗ vlad.dragoi@uav.ro
† Normandie University, France; UR, LITIS Mont-saint-aignan, France
Email: † vlad-florin.dragoi@univ-rouen.fr
‡ ERAL, University of Yaounde 1
Email: ‡ hervekalachi@gmail.com
Abstract—This letter presents a cryptanalysis of the modified
McEliece cryptosystem recently proposed by Moufek, Guenda
and Gulliver [24]. The system is based on the juxtaposition of
quasi-cyclic LDPC and quasi-cyclic MDPC codes. The idea of
our attack is to find an alternative permutation matrix together
with an equivalent LDPC code which allow the decoding of any
cipher-text with a very high probability. We also apply a recent
technique to determine weak keys [4] for this scheme. The results
show that the probability of weak keys is high enough that this
variant can be ruled out as a possible secure encryption scheme.
Index Terms—Post-quantum cryptography; McEliece cryptosystem; QC-LDPC and QC-MDPC codes.
I. I NTRODUCTION
The problem of finding a practical solution for quantum
resistant cryptography has become an urgent issue, mainly
for two reasons: the existence of a quantum polynomial time
algorithm [29] that breaks the actual RSA and ECC solutions
and the improvements of classical algorithms for the discrete
logarithm in small characteristic [2]. Among the possible
candidates for post-quantum cryptography, code-based cryptography is one of the most promising solutions [6]. It is also
one of the oldest public key encryption solutions thanks to
McEliece’s idea [21].
In the original paper, McEliece proposed to use binary
Goppa codes, which remain unbroken, to date. Other families
of algebraic codes were proposed (see [25], [31], [14] ect).
But they were successfully cryptanalyzed, mainly due to their
algebraic structure [9], [22], [32]. Probabilistic codes were also
considered: concatenated codes were analyzed in [28], [31];
convolutional codes were proposed in [17], but successfully
cryptanalyzed in [16]. Polar codes were also proposed in [30]
and cryptanalyzed in [3].
The emergence of all these attacks on several variants of the
McEliece cryptosystem shows the importance and necessity
of proposing new masking techniques. For example a code
with an efficient decoding algorithm could be masked with
the use of an arbitrary code in order to sustain the security of
the scheme. This technique was proposed for the first time by
Wieschebrink in [35] to avoid the Sidelnikov-Shestakov attack
[32]. The idea is to use the juxtaposition of a Generalized
Reed-Solomon code and a random code. This solution became
famous since it was also used in the case of Reed-Muller codes
[12]. But these two variants were broken using the square code
technique in [8], [26]. Nonetheless the idea was not totally
abandoned since Wang [34] proposed to use the juxtaposition
of a GRS code with a random code and then to multiply
the generator matrix of the resulting code with a matrix that
linearly combines the columns of the GRS and the random
code. The main motivation is to obtain a random like code in
the end.Another technique was introduced in [18] where the
authors propose to use the “u|u + v” construction with two
GRS codes. This new trend also inspired researchers to use
the juxtaposition of LDPC and MDPC codes [24]. This is the
variant that we analyze in the present article.
a) Our contribution: We describe a key recovery attack
against the modified McEliece cryptosystem based on the
juxtaposition of LDPC and MDPC codes [24]. The idea of
our attack is to find an alternative permutation matrix together
with an equivalent LDPC code which allow the decoding of
any cipher-text with a very high probability.
II. BACKGROUND
ON
C ODING T HEORY
Throughout the paper we denote by F2 the finite field with 2
elements and M k,n (F2 ) the set of k × n matrices with entries
in F2 . The Hamming weight wt(x) of a vector x ∈ F2n is
the number of non-zero coordinates of x. A binary linear
code C of length n and dimension k is a k-dimensional linear
subspace of F2n . A generator matrix of C is any k × n matrix
G ∈ M k,n (F2 ) with rows that generate C . The dual C ⊥ of C
is the n − k-dimensional linear subspace defined by
n
o
Õ
C ⊥ = z ∈ F2n : ∀c ∈ C ,
ci zi = 0 .
i
A parity check matrix of C is a generator of C ⊥ .
Definition 1. A (n, k, ω)-code is a binary linear code defined
by a k × n parity-check matrix (k < n) where each row has
weight ω. When n → ∞ we define
• An LDPC code is a (n, k, ω)-code with ω = O (1) [11].
√
• An MDPC code is a (n, k, ω)-code with ω = O
n [23].
Furthermore we will use the (n, k, ωldpc ) notation, respectively (n, k, ωmdpc ). The original decoding technique for LDPC
codes is the well-known bit flipping algorithm [11]. This
2
technique is known to provide an error-correction capability
which increases linearly with the length of the code, but decreases with the weight of the parity-checks. Therefore MDPC
codes suffer from a degradation in decoding performance,
compared with LDPC codes. A full description of the bit
flipping algorithm can be found in [7], [11], [23].
Remark 1. Since the performance and the correctness of the bit
flipping algorithm depend on the density of the parity-check
matrix, any equivalent parity-check matrix that respects the
density condition enables a correct decoding algorithm.
III. M OUFEK , G UENDA AND G ULLIVER ’ S CRYPTOSYSTEM
In [24] Moufek, Guenza and Gulliver proposed a McEliece
type cryptosystem, based on LDPC and MDPC codes. The
scheme is composed of three algorithms: key generation
KeyGen(·), encryption Encrypt(·) and decryption Decrypt(·).
a) KeyGen(n, k, t1, t2 ) = (pk, sk):
• Pick a generator matrix G1 of a (n1, k, ωldpc ) LDPC code
C1 that can correct t1 errors, and a generator matrix G2 of
a(n2, k, ωmdpc ) MDPC code denoted C2 that can correct
t2 errors.
• Pick at random S in GLk (F2 ) and an n × n permutation
matrix P, where n = n1 + n2 .
def
• Compute Gpub = SGP, where G = (G1 | G2 ) .
• Return pk = (Gpub, t1, t2 ) and sk = (S, G1, G2, P ).
b) Encrypt(m, pk) = z:
n
n
• Randomly generate e = (e1 | e2 ) ∈ F2 with e1 ∈ F2 1 ,
n2
e2 ∈ F2 , wt(e1 ) = t1 and wt(e2 ) = t2 .
• Compute z = mGpub + e.
c) Decrypt(z, sk) = m:
• Compute z ∗ = zP −1 and decode using the bit flipping
algorithm for C1 and C2 . The output is m∗ ∈ F2k .
∗ −1
• Return the message m S .
∗
Remark 2. During the
decryption,
we −1have ∗z n1=
∗
∗
∗
∗
mS (G1 | G2 )+ e1 | e2 with e1 | e2 = eP and e1 ∈ F2 .
The authors of this scheme propose to use the decoding
capability of both LDPC and MDPC codes to find e∗1 and
e∗2 . However, we emphasize that one can obtain wt(e∗1 ) > t1 or
wt(e∗2 ) > t2 . This can imply a failure in the decoding algorithm
of C1 or C2 . But both situations cannot occur simultaneously
since these would then imply that wt(eP −1 ) > t. It is also
important to remark that obtaining e∗1 or e∗2 is sufficient to
recover mS and thus m.
IV. C RYPTANALYSIS OF THE M OUFEK , G UENDA
G ULLIVER ’ S SCHEME .
AND
We propose here a key recovery attack against the previous
cryptosystem. One of the key points in our attacks is the
following proposition.
weight parity-check matrix of the MDPC code, (A | B) ∈
M k,n1 +n2 (F2 ) a full-rank matrix such that G1 At + G2 B t = 0
and A, B , 0.
Proof. The proof is obvious since we have H ′Gtpub = 0.
This proposition shows the existence of a sufficient number
of codewords with weights ωldpc in the dual of the public
code. We represent this set of codewords by the matrix
H ∗ ∈ M k ∗,n1 +n2 (F2 ) , each row of H ∗ being an element
of the set and conversely. Since these
codewords of weights
ωldpc contain the rows of Hldpc 0 P, we have rank(H ∗ ) ≥
n1 − k. One can expect to have rank(H ∗ ) > n1 − k but
our practical experiments always gave an equality, that is to
say rank(H ∗ ) = n1 − k. In the sequel we can suppose that
k ∗ = n1 − k. If k ∗ > n1 − k, one can select only n1 − k rows
of H ∗ that are linearly independent.
The attack starts with a search for codewords of weights
ωldpc .
a) Search of codewords of weights ωldpc : This step
aims to find the rows of the matrix H ∗ ∈ M k ∗,n1 +n2 (F2 ) .
Assuming without loss of generality that an adversary knows
the value of the parameter ωldpc, this step can be achieved
by applying any of the ISD variants such as, for example,
Dumer’s algorithm [10]. This issue is discussed in detail in
Section V, where a complexity analysis of our attack is given.
We emphasize that ωldpc can be easily guessed during this step
by starting with ωldpc = 1 and increasing up to the value that
satisfies the condition rank(H ∗ ) = n1 − k.
Remark 3. At the end of this step we have managed to build
the matrix H∗ ∈ M k ∗,n1 +n2 (F2 ) that generates the same code
as Hldpc | 0 P and therefore has n2 zero columns.
∗
Proposition 2. There
exists a permutation matrix P such that
H ∗ P ∗ = H1∗ | 0 with H1∗ ∈ M k ∗,n1 (F2 ) . Such a matrix can
be computed with complexity O(n) and satisfies
• Hldpc | 0 P P ∗ = Hldpc P1 | 0 ,
P1 0
• PP∗ =
.
0 P2
P1 and P2 being permutation matrices of sizes n1 and n2
respectively.
Proof. The existence of P ∗ comes
directly from Remark
3 which also provides Hldpc | 0 P P ∗ = Hldpc P1 | 0 .
∗
Furthermore, this last equality with the fact that
P1 and
PP
P
0
1
are permutation matrices imply that P P ∗ =
where
0 P2
P2 is also a permutation matrix. To finish, given the matrix
H ∗, a matrix P ∗ can be easily computed by identifying the
n2 zero columns of H ∗ .
Proposition 1. A parity check matrix of the public code is
H
© ldpc
H ′ = HP = 0
« A
ª
Hmdpc ® P,
B ¬
where Hldpc ∈ M n1 −k,n1 (F2 ) is the low weight parity-check
matrix of the LDPC code, Hmdpc ∈ M n2 −k,n2 (F2 ) the low
0
This proposition shows that a cryptanalysis is able to find an
alternative permutation matrix P ∗ together with a parity check
matrix H1∗ of an equivalent LDPC code that can correct the
same number of errors as the secret one.
In the sequel, we are
going to show that the pair P ∗, H1∗ is sufficient to decode
any cipher-text with a high probability.
3
b) Decryption with P ∗, H1∗ : We show here the way
to decrypt
with P ∗ and H ∗ . Let e∗ = eP ∗ =
any cipher-text
∗
∗
∗
∗
e1 | e2 and z = zP = z1∗ | z2∗ with e∗1, z1∗ ∈ F2n1 . So
z ∗ = mGpub + e P ∗ = mS (G1 P1 | G2 P2 ) + e∗1 | e∗2 .
TABLE I
T HE PROBABILITY Pr(wt(x1 ) > t1 ) FOR t1 = n1 /20 AND t2 = n2 /40 WITH n1 = 4k AND n2 = 2k . T HE FIRST
NAMELY
This implies that z1∗ = mSG1 P1 +e∗1 . Since SG1 P1 generates
an LDPC code with parity-check matrix H1∗, we can recover
e∗1 using the bit-flipping algorithm, assuming that wt(e∗1 ) ≤ t1 .
In the next paragraph we prove that the probability that
wt(e∗1 ) ≤ t1 is asymptotically close to 1 when the length of
the codes goes to infinity.
Theorem 1. For i ∈ {1, 2} let ni, ti be integers such that ti < ni
with n1 = γn2 and t1 > γt2, where γ ≥ 1. Let x = (x1 | x2 )
be a random vector over F2n1 +n2 with wt(x) = t1 + t2, where
xi ∈ F2ni for i ∈ {1, 2}. Then we have
n1 n2
Pr(wt(x1 ) > t1 ) < t2
Proof. First notice that we have
t1 +1 t2 −1
n1 +n2
t1 +t2
∀ 0 ≤ i ≤ t1 + t2 , Pr(wt(x1 ) = i) =
.
n1
n2
i t1 +t2 −i
.
n1 +n2
t1 +t2
(1)
Using the latter probability we obtain by a simple computation
Pr(wt(x1 ) = i)
(i + 1)(n2 − t1 − t2 + i + 1)
=
.
Pr(wt(x1 ) = i + 1)
(n1 − i)(t1 + t2 − i)
Hence for any i ≥ t1 + 1 we obtain
Pr(wt(x1 ) = i)
≥
Pr(wt(x1 ) = i + 1)
(t1 + 2)(n2 − t2 + 2)
. Replacing n1 = γn2 and t1 > γt2 in the
(n1 − t1 )t2
latter fraction we obtain that
∀ i ≥ t1 + 1 we have Pr(wt(x1 ) = i) > Pr(wt(x1 ) = i + 1)
(2)
From (2) and (1) we deduce the desired result.
When t1 and t2 are linear in the code length we obtain the
following asymptotic approximation
Corollary 1. Let n2 = n, n1 = γn and t1 = αn1, t2 = βn2 with
β < αγ ≤ 1/2. Then when n → ∞ we have
√ −n
Pr(wt(x1 ) > t1 ) < cα,β,γ n2
(γ+1)h(α+ β−α
γ+1 )−γh(α)−h(β)
,
where cα,β,γ is a constant and h is the binary entropy function.
Proof. Apply the Stirling approximation for factorials and
expand the series to obtain the result.
V. C OMPLEXITY
ANALYSIS AND NUMERICAL RESULTS .
a) Complexity analysis: The work factor of computing
an alternative private key (P ∗, H ∗ ) is given by the work
factor of the low weight codewords search algorithm plus
the computation required to find P ∗ . The first step can be
done using any of the ISD variants such as Dummer (DISD, [10]), May, Meurer and Thomae (MMT-ISD, [19]) or
Becker, Joux, May and Meurer (BJMM-ISD, [5]) or May
and Ozerov (MO-ISD [20]). These algorithms have a time
T HEOREM 1,
n1 n2 n1 +n2
ubk, t , t = t2
/ t +t . T HE THIRD ROW IS THE ASYMPTOTIC VALUE OF THE UPPER
t1 +1 t2 −1
1 2
1 2
β−α
√ −n (γ+1)h(α+ γ+1 )−γh(α)−h(β)
.
BOUND FROM C OROLLARY 1, I . E . ub k, α, β,γ = n2
ROW IS THE EXACT VALUE FOR THE PROBABILITY AND THE SECOND ROW IS THE UPPER BOUND FROM
k
log2 (Pr(wt(x
1) > t1 ))
log2 ub k, t1, t2 Thm1
log2 ub k, α, β,γ Cor1
6851
-117
8261
-141
9857
-166
-110
-133
-159
-106
-129
-155
complexity equal to O(e−ω ln(1−k/n)(1+o(1)) ), as long as ω = o(n)
when n → ∞ [33]. Hence, in our case computing the
matrix H ∗ requires a work factor asymptotically equal to
O(e−ωldpc ln(k/(n1 +n2 ))(1+o(1)) ), when n1 + n2 tends to infinity.
Regarding the complexity of computing P ∗, it requires n1 + n2
basic operations (here we consider binary additions of length
k binary vectors). Thus the complexity of our attack, denoted
by WF A (n1, n2, k, ωldpc ) is in the worst case dominated by the
cost of the best ISD variant.
b) Numerical results: We analyze the effective cost of
our attack on some practical parameters. Firstly we considered
suggested values given in [24], more precisely n1 +n2 = 16128.
As for the co-dimension we analyzed three different cases,
n1 + n2 − k ∈ {8064, 10080, 12096}. Hence the complexity of
finding the codewords of weight 15 in the dual of the public
code ( i.e. ωldpc = 15), using the BJMM-ISD variant equals
26.13, 210.97 respectively 219,20 . The computations were done
using a PariGP implementation similar to that in [27]. The
probability Pr(wt(x1 ) > t1 ) (for these parameters) is given in
Table I. These results show that the parameters proposed in
[24] are too vulnerable to be considered in practice.
One might generate a more resistant set of parameters for
the scheme, which fact we illustrate in Table II. However, it is
important to consider the weak keys approach, a recent technique introduced in [4], where the authors use the Extended
Euclidean algorithm in order to recover a private key given a
public key. We compute the probability of weak keys for the
Moufek et al. variant with reasonable parameters in Table II.
TABLE II
T HE PROPORTION OF WEAK KEYS AND THE COMPLEXITY OF OUR ATTACK AGAINST THE M OUFEK ET AL . M C E LIECE
VARIANT USING A
(4k, k, ωldpc ) LDPC CODE AND A (2k, k, ωmdpc ) MDPC CODE .
(k, ωldpc )
log WFA (4k, 2k, k, ωldpc )
Proportion of weak keys
(6851, 36)
(8261, 44)
(9857, 54)
80.9
101.6
127.4
2−7.3
2−10.6
2−14
Remark 4. In the first place we remark that the key size for the
[24] scheme is considerably greater than for similar schemes
such as [1], [23].
Notice from Table II that the odds of generating weak keys
are too big to imagine that the scheme can be protected against
this type of attack. Indeed, we find that the scheme might
be secured against the weak keys approach by increasing the
values of the ωldpc . But in order to obtain reasonable secure
parameters this solution is equivalent to replacing the LDPC
4
code with an MDPC code and thus is of no interest compared
with the MDPC McEliece variant [23].
Remark 5. It is also worth mentioning the recent reaction
attacks against the QC-MDPC scheme [13] and the QCLDPC scheme [15]. This technique can be used to recover
the structure of the MDPC code in the case of Guenda’s et al.
variant. Nonetheless we remark from Theorem 1 and Corollary
1 that the weight of x2 is likely to be much bigger than the
error capacity of the MDPC code. Hence, one might not be
able to retrieve the initial message unless it uses the LDPC
code.
However, the reaction attack remains highly interesting in
similar constructions, namely in the case of the direct sum
or Plotkin sum of LDPC and MDPC codes. Indeed, in these
cases attacking the LDPC code with our technique is not
sufficient for retrieving the initial message, and thus revealing
the structure of the MDPC code is necessary.
VI. C ONCLUSION
We have proposed a successful cryptanalysis of the
McEliece variant in [24]. Our attack exploits the structure of
the dual of the public code and its complexity is dominated
by the low weight search algorithm on this dual. The attack
is entirely based on finding the structure of the LDPC code,
regardless of the nature of the second code. As a consequence,
our result can be applied even if the MDPC code is replaced
by another code.
We notice that this variant is also vulnerable to the weak
keys approach [4], since the proportion of weak keys is not
negligible. Hence, one can consider that the McEliece variant
[24] is too vulnerable to be practical. We also emphasize that
similar constructions, such as the direct sum or Plotkin sum
of MDPC and LDPC codes can be attacked by combining our
technique with the latest reaction attacks.
R EFERENCES
[1] Marco Baldi. QC-LDPC Code-Based Cryptography 2014. Springer,
Berlin, Heidelberg.
[2] Razvan Barbulescu, Pierrick Gaudry, Antoine Joux, and Emmanuel
Thomé. A heuristic quasi-polynomial algorithm for discrete logarithm
in finite fields of small characteristic. In Nguyen, P.Q. Oswald, E.
(eds.) EUROCRYPT 2014 LNCS, vol. 8441, pp. 1–16,Springer, Berlin,
Heidelberg.
[3] Magali Bardet, Julia Chaulet, Vlad Dragoi, Ayoub Otmani, and JeanPierre Tillich. Cryptanalysis of the McEliece public key cryptosystem
based on polar codes. In Post-Quantum Cryptography 2016, LNCS,
Fukuoka, Japan. Springer International Publishing, Cham.
[4] Magali Bardet, Vlad Dragoi, Jean-Gabriel Luque, and Ayoub Otmani.
Weak keys for the quasi-cyclic MDPC public key encryption scheme.
In AFRICACRYPT 2016, Fes, Morocco, Proceedings, pages 346–367.
Springer International Publishing, Cham.
[5] Anja Becker, Antoine Joux, Alexander May, and Alexander Meurer.
Decoding random binary linear codes in 2 n/20 : How 1 + 1 = 0 improves
information set decoding. In EUROCRYPT 2012, LNCS Springer,
Berlin, Heidelberg.
[6] Daniel J. Bernstein, Johannes Buchmann, and Erik Dahmen, editors.
Post-Quantum Cryptography. Springer-Verlag, 2009.
[7] Julia Chaulet and Nicolas Sendrier. Worst case QC-MDPC decoder for
McEliece cryptosystem. In IEEE Int. Symposium on Inf. Theory, ISIT
2016, Barcelona, Spain, pages 1366–1370.
[8] Alain Couvreur, Philippe Gaborit, Valérie Gauthier-Umaña, Ayoub Otmani, and Jean-Pierre Tillich. Distinguisher-based attacks on publickey cryptosystems using Reed-Solomon codes. Des. Codes Cryptogr.,
73(2):641–666, 2014.
[9] Alain Couvreur, Irene Márquez-Corbella, and Ruud Pellikaan. A
polynomial time attack against algebraic geometry code based public key
cryptosystems. In Proc. IEEE Int. Symposium Inf. Theory - ISIT 2014,
pages 1446–1450.
[10] Ilya Dumer. On minimum distance decoding of linear codes. In Proc.
5th Joint Soviet-Swedish Int. Workshop Inform. Theory, pages 50–52,
Moscow, 1991.
[11] R. G. Gallager. Low Density Parity Check Codes. M.I.T. Press,
Cambridge, Massachusetts, 1963.
[12] Cheikh Thiecoumba Gueye and El Hadji Modou Mboup. Secure
cryptographic scheme based on modified Reed Muller codes. Int.
Journal of Security and Its Applications, 7(3):55–64, 2013.
[13] Guo Qian and Johansson Thomas and Stankovski Paul. A key recovery
attack on MDPC with CCA security using decoding errors, ASIACRYPT
2016, Vietnam, 2016, Proceedings, Part I 22, pages 789–815, Springer,
Berlin, Heidelberg.
[14] Heeralal Janwa and Oscar Moreno. McEliece public key cryptosystems
using algebraic-geometric codes. Des. Codes Cryptogr., 8(3):293–307,
1996.
[15] Tomáš Fabšič, Viliam Hromada, Paul Stankovski, Pavol Zajac, Qian
Guo, and Thomas Johansson. A reaction attack on the QC-LDPC
McEliece cryptosystem. In Post-Quantum Cryptography 2017, Utrecht,
The Netherlands, pages 51–68, Springer, Cham.
[16] Grégory Landais and Jean-Pierre Tillich. An efficient attack of a
McEliece cryptosystem variant based on convolutional codes. In P. Gaborit, editor, Post-Quantum Cryptography’13, volume 7932 of LNCS,
pages 102–117, June 2013. Springer, Berlin, Heidelberg
[17] Carl Löndahl and Thomas Johansson. A new version of McEliece PKC
based on convolutional codes. In Inf. and Communications Security,
ICICS,2012, volume 7168 of LNCS, pages 461–470. Springer, Berlin,
Heidelberg.
[18] Irene Márquez-Corbella and Jean-Pierre Tillich. Using Reed-Solomon
codes in the (u |u + v) construction and an application to cryptography.
In Proc. IEEE Int. Symposium Inf. Theory - ISIT, pages 930–934, 2016.
[19] Alexander May, Alexander Meurer, and Enrico Thomae. Decoding
random linear codes in O(20.054n ). In ASIACRYPT 2011, volume 7073
of LNCS, pages 107–124. Springer, Berlin, Heidelberg.
[20] Alexander May and Ilya Ozerov. On computing nearest neighbors with
applications to decoding of binary linear codes. In EUROCRYPT 2015,
volume 9056 of LNCS, pages 203–228. Springer, Berlin, Heidelberg.
[21] Robert J. McEliece. A Public-Key System Based on Algebraic Coding
Theory, pages 114–116. Jet Propulsion Lab, 1978. DSN Progress Report
44.
[22] Lorenz Minder and Amin Shokrollahi. Cryptanalysis of the Sidelnikov
cryptosystem. In EUROCRYPT 2007, volume 4515 of LNCS, pages
347–360, Barcelona, Spain. Springer, Berlin, Heidelberg.
[23] Rafael Misoczki, Jean-Pierre Tillich, Nicolas Sendrier, and Paulo S.
L. M. Barreto. MDPC-McEliece: New McEliece variants from moderate
density parity-check codes. In Proc. IEEE Int. Symposium Inf. Theory
- ISIT, pages 2069–2073, 2013.
[24] Hamza Moufek, Kenza Guenda, and T Aaron Gulliver. A new variant of
the McEliece cryptosystem based on QC-LDPC and QC-MDPC codes.
IEEE Communications Letters, 21(4):714–717, 2017.
[25] Harald Niederreiter. Knapsack-type cryptosystems and algebraic coding
theory. Problems of Control and Information Theory, 15(2):159–166,
1986.
[26] Ayoub Otmani and Hervé Talé Kalachi. Square code attack on a modified sidelnikov cryptosystem. In Codes, Cryptology, and Information
Security 2015, pages 173–183. Springer, Cham.
[27] Christiane Peters. Information-set decoding for linear codes over F q . In
Post-Quantum Cryptography 2010, volume 6061 of LNCS, pages 81–94.
Springer, Berlin, Heidelberg.
[28] Nicolas Sendrier. On the concatenated structure of a linear code. Appl.
Algebra Eng. Commun. Comput. (AAECC), 9(3):221–242, 1998.
[29] P.W. Shor. Algorithms for quantum computation: Discrete logarithms
and factoring. In S. Goldwasser, editor, FOCS, pages 124–134, 1994.
[30] Sujan Raj Shrestha and Young-Sik Kim. New McEliece cryptosystem
based on polar codes as a candidate for post-quantum cryptography. In
2014 14th Int. Symp. on Comm. and Inf. Tech. (ISCIT), pages 368–372.
IEEE, 2014.
[31] Vladimir Michilovich Sidelnikov. A public-key cryptosytem based on
Reed-Muller codes. Discrete Math. Appl., 4(3):191–207, 1994.
[32] Vladimir Michilovich Sidelnikov and S.O. Shestakov. On the insecurity
of cryptosystems based on generalized Reed-Solomon codes. Discrete
Math. Appl., 1(4):439–444, 1992.
5
[33] Rodolfo Canto Torres and Nicolas Sendrier. Analysis of information set
decoding for a sub-linear error weight. In Post-Quantum Cryptography
2016, Fukuoka, Japan, February 24-26, Proceedings, pages 144–161.
Springer, Cham.
[34] Yongge Wang. Quantum resistant random linear code based public key
encryption scheme rlce. In Proc. IEEE Int. Symposium Inf. Theory ISIT, pages 2519–2523, 2016.
[35] Christian Wieschebrink. Two NP-complete problems in coding theory
with an application in code based cryptography. In Proc. IEEE Int.
Symposium Inf. Theory - ISIT, pages 1733–1737, 2006.
| 7cs.IT |
ON THE DIVISOR-CLASS GROUP OF MONADIC SUBMONOIDS OF RINGS OF
INTEGER-VALUED POLYNOMIALS
arXiv:1604.03594v1 [math.AC] 12 Apr 2016
ANDREAS REINHART
Abstract. Let R be a factorial domain. In this work we investigate the connections between the
arithmetic of Int(R) (i.e., the ring of integer-valued polynomials over R) and its monadic submonoids
(i.e., monoids of the form {g ∈ Int(R) | g |Int(R) f k for some k ∈ N0 } for some nonzero f ∈ Int(R)).
Since every monadic submonoid of Int(R) is a Krull monoid it is possible to describe the arithmetic of
these monoids in terms of their divisor-class group. We give an explicit description of these divisor-class
groups in several situations and provide a few techniques that can be used to determine them. As an
application we show that there are strong connections between Int(R) and its monadic submonoids. If
R = Z or more generally if R has sufficiently many “nice” atoms, then we prove that the infinitude of the
elasticity and the tame degree of Int(R) can be explained by using the structure of monadic submonoids
of Int(R).
1. Introduction
The class of Krull monoids is among the most well-studied classes of monoids in factorization theory (see
citeGHK). It is known that the behavior of their factorizations only depends on their divisor-class group.
On the other hand, there are many examples of atomic, completely integrally closed monoids that fail to
be Krull. For instance, it is known that the ring of integer-valued polynomials Int(R) over an integral
domain R is a Krull domain if and only if R is a Krull domain and Int(R) = R[X] (see [1, Corollary
I.3.15] and [2, Corollary 2.7]). (Note that if Int(R) is a Krull domain, then R is a Krull domain, since
R• ⊆ Int(R)• is a divisor-closed submonoid.) Recently, it was shown that the ring of integer-valued
polynomials over a Krull domain satisfies a weaker property which is called monadically Krull [3, 7]. A
monoid is called monadically Krull if all its divisor-closed submonoids generated by one element (i.e.,
monadic submonoids) are Krull monoids.
The purpose of this work is to investigate monadic submonoids of rings of integer-valued polynomials over
factorial domains. Since these rings are monadically Krull it is possible to study the arithmetic of their
monadic submonoids by using their divisor-class groups. The restriction to factorial domains (instead of
Krull domains) is reasonable, since we are able to give more precise descriptions in this situation. We
pursue two goals. The first goal is a thorough description of divisor-class groups of monadic submonoids of
Int(R). We achieve this goal for monadic submonoids that are generated by polynomials with coefficients
in R. The second goal is to show that the elasticity and the tame degree of certain rings of integer-valued
polynomials are infinite. We present a proof that relies on the structure of divisor-class groups of monadic
submonoids of Int(R).
The second goal is motivated by results in the literature that were proved in the recent past. More
precisely, it is known that every nonempty finite subset of N≥2 is the set of lengths of some f ∈ Int(Z)•
(see [4]). This is a property that Int(Z) shares with Krull monoids whose divisor-class group is infinite and
where every class contains a height-one prime ideal (see [6]). The question arises whether it is possible
to describe this phenomenon in Int(Z) by using the theory of Krull monoids. So far, we were not able to
solve this problem. Therefore, we want to pursue a simpler goal and prove that the infinitude of certain
invariants (i.e., the elasticity and the tame degree) can be derived from the theory of Krull monoids.
2010 Mathematics Subject Classification. 13A15, 13F05, 13F15, 20M12.
Key words and phrases. Monadically Krull, Integer-valued polynomial, Divisor-class group.
1
2
ANDREAS REINHART
In the next section we discuss the notation that is used in this work. We recall the definitions of saturated,
divisor-closed, and monadic submonoids of a monoid, and present some of their elementary properties.
We briefly discuss a few simple facts about rings of integer-valued polynomials. Another important notion
that will be introduced is the image-content d(f ) of a nonzero integer-valued polynomial f . It is basically
a greatest common divisor of the image of f (over the base ring). This notion is of major importance in
this work.
The main purpose of the third section is to study the structure of atoms and height-one prime ideals
of monadic submonoids of Int(R). This is an important prerequisite concerning the investigation of the
divisor-class group, since it is possible to describe the structure of the divisor-class group of a Krull
monoid by using the v-product decompositions of principal ideals (generated by atoms) into height-one
prime ideals. We will specifically investigate the subset of constant atoms of a monadic submonoid. We
show that every constant atom generates a radical ideal. Moreover, we present a characterization result
for monadic submonoids where every constant atom is a prime element. We give a complete description
of the set of atoms of monadic submonoids of Int(R) that are generated by some f ∈ R[X]• . In what
follows we study the set of height-one prime ideals of monadic submonoids generated by some f ∈ R[X]•
that do not contain any constant elements. Finally, we present a result which will enable us to determine
the v-product decompositions of principal ideals into height-one prime ideals in many situations.
In the fourth section we present the first main result of this work. We show that the divisor-class group
of a monadic submonoid of Int(R) generated by some f ∈ R[X]• is torsion-free. Moreover, we present a
simple formula to calculate the torsion-free rank in this case. We proceed with a few results that hold
in a more general context. In particular, we prove a proposition which relates the P -adic exponents of
v-ideals between a Krull monoid and a saturated submonoid. It is an analogue to a well-known theorem
which connects P -adic exponents of ideals in a Dedekind domain to a subring that is also a Dedekind
domain (see [8]). Moreover, it will be useful to determine the divisor-class group of monadic submonoids
of Int(R) which are not covered by the first main theorem. We proceed by describing the set of height-one
prime ideals that contain constant elements. These results complement the achievements in Section 3,
and have several applications in the last section.
The fifth section is devoted to the construction of “more involved” examples of divisor-class groups. We
provide basically two sufficient criteria which will enable us to decompose certain divisor-class groups of
monadic submonoids into a direct product of divisor-class groups (up to an isomorphism). These criteria
will be helpful in last section of this work.
In the last section we provide a few examples and discuss several consequences of the prior sections.
Among them are a variety of counterexamples. For instance, it is shown that several characterization
results that hold for monadic submonoids generated by some f ∈ R[X]• no longer hold for arbitrary
monadic submonoids. We give non-trivial examples of divisor-class groups that are torsion groups or
torsion-free or none of the two. We prove that it is possible to find monadic submonoids of Int(Z) whose
divisor-class group is torsion-free with prescribed rank. Finally, we present the second main result of
this work. It shows that rings of integer-valued polynomials over certain factorial domains have infinite
elasticity and tame degree.
2. Notation and preliminaries
All monoids in this work are commutative, cancellative monoids. Let H be a monoid, and T ⊆ H a
submonoid. If x, y ∈ H, then we write x |H y if there is some c ∈ H with y = cx.
• We say that T ⊆ H is saturated if for all x, y ∈ T such that x |H y it follows that x |T y.
• T ⊆ H is called divisor-closed if for all x, y ∈ H with xy ∈ T we have x ∈ T .
• If E ⊆ H, then let [[E]]H denote the smallest divisor-closed submonoid of H which contains E. If
x ∈ H, then set [[x]]H = [[{x}]]H .
• We say that T ⊆ H is monadic if T = [[x]]H for some x ∈ H.
ON THE DIVISOR-CLASS GROUP OF MONADIC SUBMONOIDS
3
If E ⊆ H, then we write [[E]] instead of [[E]]H if the monoid H is the most obvious choice. Clearly,
every monadic submonoid of H is divisor-closed, and every divisor-closed submonoid of H is saturated.
Observe that if x ∈ H, then [[x]] = {y ∈ H | y |H xk for some k ∈ N0 }. A subset I ⊆ H is called an s-ideal
of H if IH = I. Let spec(H) be the set of all prime s-ideals of H. An s-ideal is called radical if it is an
intersection of prime s-ideals. By X(H) we denote the set of all height-one prime (s-)ideals of H, i.e.,
the set of all minimal nonempty prime s-ideals of H. By H × resp. A(H) we denote the set of units of H
resp. the set of atoms of H. We say that H is reduced if H × = {1}. If x, y ∈ H, then we say that x and
y are associated (we denote this by x ≃H y) if x = yε for some ε ∈ H × . It is well-known that ≃H defines
an equivalence relation on H. By Hred = {xH × | x ∈ H} we denote the set of equivalence classes of ≃H .
This set forms a monoid under the canonical multiplication. If E ⊆ H, then F ⊆ E is called a system of
representatives of E if for every x ∈ E there is a unique y ∈ F such that x ≃H y. Let L be a quotient
monoid of H. For X ⊆ L, set X −1 = {z ∈ L | zX ⊆ H} and Xv = (X −1 )−1 . A subset I ⊆ H is called
a divisorial ideal (or v-ideal) of H if Iv = I. Every divisorial ideal of H is an s-ideal of H. Let Iv (H)
denote the set of divisorial ideals of H. By Cv (H) we denote the divisor-class group (or v-class group) of
H. It measures how far (v-invertible) v-ideals are from being principal ideals. A precise definition can
be found in [5, Definition 2.1.8]. If I ∈ Iv (H), then let [I] denote the class of I in Cv (H). Note that H is
called a Krull monoid if H is a completely integrally closed Mori monoid (or equivalently, every v-ideal
of H is a finite v-product of height-one prime ideals of H). For a thorough introduction to Krull monoids
we refer to [5, Definition 2.3.1]. We say that H is monadically Krull if [[x]] is a Krull monoid for every
x ∈ H. Most of these notions can be defined analogously in the context of integral domains.
We want to recapitulate a few basic facts concerning saturated and divisor-closed submonoids of H. First
let T ⊆ H be saturated. Then H × ∩ T = T × . If H is a Krull monoid, then T is a Krull monoid. Now
let T ⊆ H be divisor-closed. Then T × = H × , and A(T ) = A(H) ∩ T .
If M is a set and l ∈ N, then a finite sequence (ai )li=1 ∈ M l , will be denoted by a (i.e., a = (ai )li=1 ).
Recall that if R is an integral domain with quotient field K and X is an indeterminate over K, then
Int(R) = {f ∈ K[X] | f (x) ∈ R for all x ∈ R} is called the ring of integer-valued polynomials over R. It is
well-known that Int(R)× = R[X]× = R× . Note that R• = R \ {0} forms a monoid under multiplication.
If we refer to a submonoid of R, then we always mean a submonoid of R• . Especially, if E ⊆ R• , then
let [[E]]R = [[E]]R• . We say that R is monadically Krull if R• is monadically Krull.
Now let R be a factorial domain, K a field of quotients of R, X an indeterminate over K, and Q a system
of representatives of A(R). For T ⊆ R, let GCDR (T ) be the set of all greatest common divisors of T (in
R), and let LCMR (T ) be the set of all least common multiples of T (in R). If f ∈ R[X]• , then we say
that f is primitive if every greatest common divisors of all coefficients of f is a unit of R. For convenience
we also allow the units of R to be primitive polynomials. If q ∈ Q, then let vq : R → N0 ∪ {∞} denote
Q
the q-adic valuation on R. Let dQ : Int(R)• → R• be defined by dQ (g) = p∈Q pmin{vp (g(c))|c∈R} for all
g
g ∈ Int(R)• . Set d = dQ . Note that d(g) ∈ GCDR ({g(c) | c ∈ R}) and d(g)
∈ Int(R) for all g ∈ Int(R)• .
It is straightforward to show that d(f k ) = d(f )k and d(af ) ≃R ad(f ) for all f ∈ Int(R)• , k ∈ N0 , and
a ∈ R• . Let n ∈ N, f ∈ (Int(R)• )n and x ∈ Nn0 \{0}. We say that x is f -irreducible if for all y, z ∈ Nn0
Qn
Qn
Qn
such that x = y + z and d( i=1 fixi ) = d( i=1 fiyi )d( i=1 fizi ) it follows that y = 0 or z = 0. It is
well-known (see [7, Theorem 5.2]) that Int(R) is monadically Krull. If f ∈ Int(R)• , then we can deduce
by [7, Theorem 3.6] and its proof that X([[f ]]), spec(H), and {u[[f ]] | u ∈ A([[f ]])} are finite sets. The
remarks in this section will be used without citation.
3. Atoms and height-one prime ideals
In this section we present a few basic preparatory results about atoms and height-one prime ideals of
monadic submonoids of Int(R). Many of the results in this section refer to monadic submonoids generated
by some “f ∈ R[X]• ”. Note that this is a rather natural assumption because it is straightforward to prove
that every monadic submonoid of Int(R) is contained in some monadic submonoid of Int(R) generated by
4
ANDREAS REINHART
some f ∈ R[X]• . (If g ∈ Int(R)• and b ∈ R• are such that bg ∈ R[X], then [[g]] ⊆ [[bg]].) The purpose of
the first result is to describe the set of “constant atoms” of monadic submonoids of Int(R). In particular,
we show that the principal ideals generated by constant atoms are radical ideals. Furthermore, we prove
that a height-one prime ideal contains at most one constant atom (up to associates).
Lemma 3.1. Let R be a factorial domain, and f ∈ Int(R)• .
1.
2.
3.
4.
If g ∈ [[f ]] and u ∈ R, then u |[[f ]] g if and only if u |R d(g).
A([[f ]]) ∩ R = [[f ]] ∩ A(R) = {u ∈ A(R) | u |R d(f )}.
If u ∈ A([[f ]]) ∩ R, then uH is a radical ideal of [[f ]].
If P ∈ X([[f ]]) and u, w ∈ P ∩ A(R), then u ≃[[f ]] w.
Proof. 1. “⇒”: Let u |[[f ]] g. There is some v ∈ [[f ]] such that g = uv. Since u ∈ R we infer that
d(g) ≃R ud(v), and thus u |R d(g). “⇐”: Let u |R d(g). We have d(g) |Int(R) g, and thus u |Int(R) g.
Since g ∈ [[f ]] it follows that u |[[f ]] g.
2. First we show that A([[f ]]) ∩ R = [[f ]] ∩ A(R). “⊆”: Let u ∈ A([[f ]]) ∩ R. Observe that u 6∈ [[f ]]× = R× .
Let x, y ∈ R be such that u = xy. It is clear that x, y ∈ [[f ]]. Therefore, x ∈ [[f ]]× = R× or y ∈ [[f ]]× = R× .
Consequently, u ∈ [[f ]] ∩ A(R). “⊇”: Let u ∈ [[f ]] ∩ A(R). We have u 6∈ R× = [[f ]]× . Let x, y ∈ [[f ]] be such
that u = xy. Observe that 0 = deg(u) = deg(x) + deg(y), and thus x, y ∈ R. Therefore, x ∈ R× = [[f ]]×
or y ∈ R× = [[f ]]× . We infer that u ∈ A([[f ]]) ∩ R.
Next we show that [[f ]] ∩ A(R) = {u ∈ A(R) | u |R d(f )}. “⊆”: Let u ∈ [[f ]] ∩ A(R). Then u |Int(R) f k
for some k ∈ N. Consequently, u |[[f ]] f k . It follows by 1 that u |R d(f k ) = d(f )k . This implies that
u |R d(f ). “⊇”: Let u ∈ A(R) be such that u |R d(f ). By 1 we have u |[[f ]] f . Therefore, u ∈ [[f ]] ∩ A(R).
3. Let u ∈ A([[f ]]) ∩ R. Let g ∈ [[f ]] and n ∈ N be such that u |[[f ]] g n . Then u |R d(g n ) = d(g)n by 1, and
thus u |R d(g). It follows by 1 that u |[[f ]] g. Therefore, uH is a radical ideal of [[f ]].
4. Assume to the contrary that there are P ∈ X([[f ]]) and u, w ∈ P ∩ A(R) such that u 6≃[[f ]] w. Let K be
a quotient field of R, X an indeterminate over K and L a quotient group of [[f ]]. Let h ∈ L be such that
uh, wh ∈ [[f ]]. We have h ∈ K[X], h ∈ uh, wh ∈ Int(R), and thus uh(z), wh(z) ∈ R for all z ∈ R. Observe
that u 6≃R w. If z ∈ R, then u |R uwh(z) = wuh(z), and thus u |R uh(z), hence h(z) ∈ R. Therefore,
h ∈ Int(R) ∩ L = [[f ]]. We infer that u−1 [[f ]] ∩ w−1 [[f ]] = [[f ]]. Consequently, [[f ]] = {u, w}v[[f ]] ⊆ P , a
contradiction.
Let R be a factorial domain, and f ∈ Int(R)• . Then A([[f ]]) ∩ R is called the set of constant atoms of
[[f ]]. Next, we characterize when every constant atom is a prime element.
Proposition 3.2. Let R be a factorial domain, and f ∈ Int(R)• . The following are equivalent:
1.
2.
3.
4.
Every P ∈ X([[f ]]) such that P ∩ R 6= ∅ is principal.
For every P ∈ X([[f ]]) such that P ∩ R 6= ∅ there is some n ∈ N such that (P n )v is principal.
Every constant atom of [[f ]] is a prime element.
d(gh) = d(g)d(h) for all g, h ∈ [[f ]].
If Cv ([[f ]]) is finite, then these conditions are satisfied.
Proof. 1. ⇒ 2.: Trivial.
2. ⇒ 3.: Let u be a constant atom of [[f ]]. Since [[f ]] is a Krull monoid there is some P p
∈ X([[f ]]) such
that u ∈ P . Some v-power of P is principal, and thus there is some x ∈ [[f ]] such that P = x[[f ]]. There
is some k ∈ N such that x |[[f ]] uk . Therefore, x ∈ R and x |R uk . Since R is factorial this implies that
p
p
x ≃R ul for some l ∈ N. Consequently, x ≃[[f ]] ul . It follows that P = ul [[f ]] = u[[f ]] = u[[f ]], hence u
is a prime element of [[f ]].
3. ⇒ 1.: Let P ∈ X([[f ]]) be such that P ∩ R 6= ∅. There is some x ∈ P ∩ R and some u ∈ A(R) such
that u |R x and u ∈ P . Observe that u is a constant atom of [[f ]], hence u is a prime element of [[f ]].
Therefore, P = u[[f ]].
ON THE DIVISOR-CLASS GROUP OF MONADIC SUBMONOIDS
5
gh
3. ⇒ 4.: Let g, h ∈ [[f ]]. It is sufficient to show that d( d(g)d(h)
) = 1. Assume to the contrary that
gh
gh
) 6= 1. Then there is some p ∈ A(R) such that p |R d( d(g)d(h)
). Obviously, p is a constant atom
d( d(g)d(h)
g
g
g
h
h
) = 1 or
of [[f ]] and p |[[f ]] d(g) d(h) . Therefore, p |[[f ]] d(g) or p |[[f ]] d(h) . This implies that p |R d( d(g)
h
) = 1, a contradiction.
p |R d( d(h)
4. ⇒ 3.: Let u be a constant atom of [[f ]] and g, h ∈ [[f ]] such that u |[[f ]] gh. Then u ∈ A(R) and
u |R d(gh) = d(g)d(h). We infer that u |R d(g) or u |R d(h). Consequently, u |[[f ]] g or u |[[f ]] h.
If Cv ([[f ]]) is finite, then for every P ∈ X([[f ]]) there is some n ∈ N such that (P n )v is principal, hence 2
is satisfied.
Now we show that elements of monadic submonoids that are generated by some f ∈ R[X]• can be
represented in form of special fractions. As a consequence, we provide a simple set of generators of the
quotient group of [[f ]]. This type of representability will turn out to be a crucial ingredient for our first
main result in Section 4.
Lemma 3.3. Let R be a factorial domain, K a quotient field of R, X an indeterminate over K and
f ∈ R[X]• . For every g ∈ [[f ]] there are some a, b ∈ [[f ]] ∩ R and some primitive h ∈ [[f ]] ∩ R[X] such that
h |R[X] f k for some k ∈ N, GCDR (a, b) = R× and g = bh
a .
Proof. There are some primitive h ∈ R[X] and some a, b ∈ R• such that GCDR (a, b) = R× and g = bh
a .
k
k
Since g ∈ [[f ]] there are some k ∈ N, z ∈ R[X]• and c ∈ R• such that bhz
=
f
.
It
follows
that
bhz
=
f
ac.
ac
Since h is primitive we infer that h |R[X] f k , and thus h ∈ [[f ]]. Observe that a |R d(bh) ≃R bd(h). We
infer that a |R d(h). Since d(h) ∈ [[f ]] we have a ∈ [[f ]]. Moreover, bh = ga ∈ [[f ]], and thus b ∈ [[f ]].
Lemma 3.4. Let R be a factorial domain, K a quotient field of R, X an indeterminate over K and
f ∈ R[X]• . Then the quotient group of [[f ]] is generated by ([[f ]] ∩ A(R[X])) ∪ [[f ]]× .
Proof. It is sufficient to show that [[f ]] ⊆ h([[f ]] ∩ A(R[X])) ∪ [[f ]]× i. Let x ∈ [[f ]]. By Lemma 3.3 there
are some h ∈ [[f ]] ∩ R[X] and a ∈ [[f ]] ∩ R such that x = ha . Since R and R[X] are factorial we infer that
h ∈ h([[f ]] ∩ A(R[X])) ∪ [[f ]]× i and a ∈ h([[f ]] ∩ A(R)) ∪ [[f ]]× i. Therefore, x ∈ h([[f ]] ∩ A(R[X])) ∪ [[f ]]× i.
Next we give a complete description of the set of atoms of monadic submonoids of Int(R)• that are
generated by some f ∈ R[X]• . A part of this description can be found in the proof of [7, Theorem 5.2].
Proposition 3.5. Let R be a factorial domain, a ∈ R• , n ∈ N and f ∈ (A(R[X]) \ R)n a sequence of
Q
pairwise non-associated
elements of R[X]. Set f = a ni=1 fi . Then {u[[f ]] | u ∈ A([[f ]])} = {u[[f ]] | u ∈
Q
n
y
fi i
yi [[f ]]
i=1 fi )
A(R), u |R d(f )} ∪ { d(Qi=1
n
| y ∈ Nn0 , y is f -irreducible}.
Proof. “⊆”: Let u ∈ A([[f ]]). Observe that u ∈ A(Int(R)). There is some k ∈ N such that u |Int(R) f k .
Case 1. u ∈ R: Clearly, u ∈ A(R). We have u |R d(f k ) = d(f )k , and thus u |R d(f ).
Case 2. u 6∈ R: There are some primitive t ∈ R[X] and some b, c ∈ R• such that GCDR[X] (bt, c) = R[X]×
bd(t)
t
t bd(t)
t
∈ Int(R) and u = d(t)
and u = bt
c . Obviously, c |R d(t), and thus d(t) , c
c . Therefore, u ≃Int(R) d(t) .
Q
There are some e ∈ R• and some s ∈ R[X] such that u es = f k . This implies that bst = ak ce ni=1 fik .
Qn
Qn
Therefore, t |R[X] i=1 fik , hence t ≃R[X] i=1 fiyi for some y ∈ Nn0 \ {0}. Observe that u ≃Int(R)
Qn
y
f i
Qi=1 i yi ,
f
d( n
i=1 i )
Qn
f
yi
n
i
and thus u[[f ]] = d(Qi=1
yi [[f ]]. We need to show that y is f -irreducible. Let α, β ∈ N0 be
n
i=1 fi )
Qn
Qn
α
β
Q
Q
Q
fi i
fi i
, Qi=1
such that y = α + β and d( ni=1 fiyi ) = d( ni=1 fiαi )d( ni=1 fiβi ). Clearly, d(Qi=1
∈ Int(R)
α
n
βi
n
f i)
and
Qn
y
fi i
Qi=1
y
n
d( i=1 fi i )
=
Qn
Qn
α
β
fi i
fi i
Qi=1
.
αi
Qi=1
n
n
d( i=1 fi ) d( i=1 fiβi )
Consequently,
Qn
α
fi i
Qi=1
α
n
d( i=1 fi i )
i=1
×
∈ Int(R)
i
or
i=1 fi
βi
i=1 fi
Qn
β
d( i=1 fi i )
d(
Qn
)
∈ Int(R)× .
This implies that α = 0 or β = 0.
“⊇”: Case 1. Let u ∈ A(R) be such that u |R d(f ). We have d(f ) ∈ [[f ]], and thus u ∈ [[f ]]. Let y, z ∈ [[f ]]
be such that u = yz. Then y, z ∈ R, hence y ∈ R× = [[f ]]× or z ∈ R× = [[f ]]× . Consequently, u ∈ A([[f ]]).
6
ANDREAS REINHART
Case 2. Let y ∈ Nn0 be f -irreducible. First we show that
Qn
y
fi i
Qi=1
y
n
d( i=1 fi i )
×
Qn
y
fi i
Qi=1
y
n
d( i=1 fi i )
∈ A(Int(R)). Let y, z ∈ Int(R)
= yz. There are some b, c, e, f ∈ R• and some primitive g, h ∈ R[X] such
Qn
yi
ch
that GCD(b, e) = R , GCD(c, f ) = R× , y = bg
i=1 fi and
e and z = f . This implies thatQg |R[X]
Qn
h |R[X] i=1 fiyi . Consequently, there are some v, w ∈ Nn0 such that g ≃R[X] ni=1 fivi and h ≃R[X]
Qn
wi
d(yz) = 1, Qit follows Q
that d(y) = d(z) = 1.
i=1 fi . Observe that e |R d(g) and f |R d(h). Since
Q
be such that
n
f
yi
n
f
vi
n
f
wi
i
i
i
Qi=1
≃R d(g) and fc ≃R d(h). We infer that d(Qi=1
≃R d(Qi=1
yi
vi
wi . This implies
n
n
n
i=1 fi )
i=1 fi ) d(
i=1 fi )
Q
Qn
n
wi
vi
that y = v + w and d( i=1 fi )d( i=1 fi ). Consequently, v = 0 or w = 0, and thus y ∈ Int(R)× or
z ∈ Int(R)× .
Qn
Qn
y
y
Qn
fi i
f i
Qi=1 i y
It is clear that i=1 fiyi ∈ [[f ]], hence d(Qi=1
∈
[[f
]].
This
implies
that
∈ [[f ]] ∩ A(Int(R)) =
yi
n
f )
d( n f i )
Therefore,
A([[f ]]).
e
b
i=1
i
i=1
i
Note that the set of atoms of an arbitrary monadic submonoid H of Int(R) can be derived from Proposition
3.5 and the fact that A(H) = A(Int(R)) ∩ H. We proceed with an important lemma which will enable
us to identify certain divisorial ideals of monadic submonoids of Int(R).
Lemma 3.6. Let R be a factorial domain, f ∈ Int(R)• and g ∈ [[f ]]. Then {0} =
6 LCMR ({ d(gh)
d(h) | h ∈
[[f ]]}) ⊆ [[f ]].
Proof. Let Q be a system of representatives of A(R) and d = dQ . We show that there is some T ⊆ R
such that for all h ∈ [[f ]], h(y) 6= 0 for all y ∈ T and min{vp (h(x)) | x ∈ R} = min{vp (h(x)) | x ∈ T }
for all p ∈ Q. Set T = {x ∈ R | f (x) 6= 0}. Let h ∈ [[f ]]. Then h |Int(R) f k for some k ∈ N. Let
y ∈ T . Then f (y) 6= 0, hence f (y)k 6= 0, and thus h(y) 6= 0. Let p ∈ Q. There is some v ∈ R such that
min{vp (h(x)) | x ∈ R} = vp (h(v)). It is straightforward to prove that there is some m ∈ N such that
vp (h(v + pl )) = vp (h(v)) for all l ∈ N≥m . Since R \ T is finite, we can find some n ∈ N≥m such that
v + pn ∈ T . This implies that min{vp (h(x)) | x ∈ R} = vp (h(v + pl )), hence min{vp (h(x)) | x ∈ R} =
min{vp (h(x)) | x ∈ T }.
Next we prove that for every p ∈ Q there is some z ∈ N0 such that vp ( d(gk)
d(k) ) ≤ z for all k ∈ [[f ]].
Without restriction let f 6∈ R× . By [7, Theorem 3.6 and Theorem 5.2] there is some finite ∅ =
6 U ⊆ A([[f ]])
such that [[f ]] = [U ∪ [[f ]]× ]. Let p ∈ Q. By Dickson’s theorem (see [5, Theorem 1.5.3]) there is some
finite ∅ 6= S ⊆ T such that Min({(vp (u(x)))u∈U | x ∈ T }) = {(vp (u(x)))u∈U | x ∈ S}. We show that
×
min{vp (l(x)) | x ∈ R} = min{vQ
p (l(x)) | x ∈ S} for all l ∈ [[f ]]. Let l ∈ [[f ]]. There are some
P η ∈ [[f ]] and
eu
U
(eu )u∈U ∈ N0 such that l = η u∈U u . Clearly, min{vp (l(x)) | x ∈ R} = vp (l(w)) = u∈U eu vp (u(w))
for some w ∈ T . By Dickson’s theorem (see [5, Theorem
1.5.3]) we canPfind some y ∈ S such that
P
vp (u(y)) ≤ vp (u(w)) for all u ∈ U . Since vp (l(y)) = u∈U eu vp (u(y)) ≤ u∈U eu vp (u(w)) = vp (l(w))
we infer that min{vp (l(x)) | x ∈ R} = vp (l(y)) = min{vp (l(x)) | x ∈ S}.
Set z = max{vp (g(x)) | x ∈ S}. Then z ∈ N0 . Let k ∈ [[f ]]. Now we prove that vp ( d(gk)
d(k) ) ≤ z. There
is some v ∈ S such that min{vp (k(x)) | x ∈ S} = vp (k(v)). We have vp ( d(gk)
d(k) ) = min{vp ((gk)(x)) |
x ∈ R} − min{vp (k(x)) | x ∈ R} = min{vp ((gk)(x)) | x ∈ S} − min{vp (k(x)) | x ∈ S} ≤ vp ((gk)(v)) −
vp (k(v)) = vp (g(v)) ≤ z.
Set P = {p ∈ Q | vp (d(f )) > 0}. Then P is finite. For every h ∈ [[f ]] it follows that {p ∈ Q | vp ( d(gh)
d(h) ) >
d(gh)
Q
0} ⊆ P . This implies that 0 6= p∈P pmax{vp ( d(h) )|h∈[[f ]]} ∈ LCMR ({ d(gh)
d(h) | h ∈ [[f ]]}).
Q
max{vp ( d(gh)
)|h∈[[f ]]}
d(h)
Note that d(f ) ∈ [[f ]]. Consequently, P ⊆ [[f ]], and thus p∈P p
∈ [[f ]]. Since least
common multiples are unique up to units it follows immediately that LCMR ({ d(gh)
d(h) | h ∈ [[f ]]}) ⊆ [[f ]].
Let R be a factorial domain, Q a system of representatives of A(R) and f ∈ Int(R)• . Then let ef,Q : [[f ]] →
d(gh)
Q
R• be defined by ef,Q (g) = p∈Q pmax{vp ( d(h) )|h∈[[f ]]} for all g ∈ [[f ]]. Observe that LCMR ({ d(gh)
d(h) | h ∈
ON THE DIVISOR-CLASS GROUP OF MONADIC SUBMONOIDS
7
[[f ]]}) = ef,Q (g)R× for all g ∈ [[f ]]. It follows from Lemma 3.6 that ef,Q is well defined and ef,Q ([[f ]]) ⊆ [[f ]].
In the following we suppose that a fixed Q is given and set ef = ef,Q .
A well-known and very basic result in ring theory is that contractions of ideals to subrings are ideals
again. In analogy, it holds that contractions of s-ideals of monoids to submonoids are s-ideals again.
The system of v-ideals has a very different behavior. In the last section of this work we show that the
contraction of a v-ideal of a Krull monoid to a monadic submonoid can fail to be a v-ideal. The next
result, however, gives a positive answer under more restrictive conditions.
Proposition 3.7. Let R be a factorial domain, K a field of quotients of R, X an indeterminate over
K, f ∈ R[X]• and g ∈ [[f ]]. Then gK[X] ∩ [[f ]] = efg(g) [[f ]] ∩ [[f ]]. In particular, gK[X] ∩ [[f ]] ∈ Iv ([[f ]]),
and if g ∈ A(K[X]), then gK[X] ∩ [[f ]] ∈ X([[f ]]).
Proof. Set e = ef (g). “⊆”: Let z ∈ gK[X] ∩ [[f ]]. Then there are some a, b ∈ R• and some h′ ∈ Int(R)
′
′ ′′
k
such that d(h′ ) = 1, GCDR (a, b) = R× and z = agh
b . Clearly, h h = bf for some k ∈ N and some
y′
h′′ ∈ Int(R). There are some primitive y ′ ∈ R[X], some y ′′ ∈ R[X] and some d ∈ R such that h′ = d(y
′)
′′
and h′′ = yd . We infer that y ′ y ′′ = d(y ′ )dbf k . Since y ′ is primitive it follows that y ′ |R[X] f k , and thus
y′
′
′
y ′ ∈ [[f ]]. Therefore, h′ = d(y
′ ) ∈ [[f ]]. Since bz = agh , we have bd(z) ≃R ad(gh ). This implies that
′
)
e
a |R d(z) and b |R d(gh′ ) = d(gh
d(h′ ) |R e. By Lemma 3.6 we have e ∈ [[f ]]. It follows that a, b ∈ [[f ]], hence
g
e
e
′
g z = b ah ∈ [[f ]]. Consequently, z ∈ e [[f ]] ∩ [[f ]]. “⊇”: Trivial.
Observe that ge is an element of the quotient group of [[f ]]. Therefore, gK[X] ∩ [[f ]] is an intersection of
fractional principal ideals of [[f ]], and thus gK[X] ∩ [[f ]] ∈ Iv ([[f ]]).
Now let g ∈ A(K[X]). Clearly, gK[X] ∈ spec(K[X]), and thus gK[X]• ∈ s-spec(K[X]• ) \ {∅}. This
implies that gK[X] ∩ [[f ]] = gK[X]• ∩ [[f ]] ∈ s-spec([[f ]]) \ {∅}. Since [[f ]] is a Krull monoid and gK[X] ∩ [[f ]]
is divisorial, we have gK[X] ∩ [[f ]] ∈ X([[f ]]).
As a consequence, we obtain a description of the set of height-one prime ideals (of monadic submonoids
generated by some f ∈ R[X]• ) that do not contain constant elements.
Corollary 3.8. Let R be a factorial domain, K a field of quotients of R, X an indeterminate over K and
f ∈ R[X]• . Then {P ∈ X([[f ]]) | P ∩ R = ∅} = {gK[X] ∩ [[f ]] | g ∈ [[f ]] ∩ A(R[X]) \ R} = {gK[X] ∩ [[f ]] |
g ∈ [[f ]] ∩ A(K[X])}. In particular, if R is a system of representatives of [[f ]] ∩ A(R[X]) \ R, then
Q : R → {P ∈ X([[f ]]) | P ∩ R = ∅} defined by Q(t) = tK[X] ∩ [[f ]] is a bijection.
Proof. Since R is factorial it follows that A(R[X]) \ R ⊆ A(K[X]).
First we prove that {P ∈ X([[f ]]) | P ∩ R = ∅} ⊆ {gK[X] ∩ [[f ]] | g ∈ [[f ]] ∩ A(R[X]) \ R}. Let
P ∈ X([[f ]]) be such that P ∩ R = ∅. There
some a ∈ [[f ]] ∩ R, some n ∈ N and some finite sequence
Qare
n
f ∈ ([[f ]] ∩ A(R[X]) \ R)n such that f = a i=1 fi . Clearly, f ∈ P . Consequently, there is some i ∈ [1, n]
such that fi ∈ P . It follows by Proposition 3.7 that fi K[X] ∩ [[f ]] ∈ X([[f ]]). By Proposition 3.7 we have
fi K[X] ∩ [[f ]] = eff(fi i ) [[f ]] ∩ [[f ]]. Let z ∈ fi K[X] ∩ [[f ]]. Then ef (fi )z ∈ fi [[f ]] ⊆ P , and thus z ∈ P . This
implies that fi K[X] ∩ [[f ]] ⊆ P , hence P = fi K[X] ∩ [[f ]].
It is obvious that {gK[X] ∩ [[f ]] | g ∈ [[f ]] ∩ A(R[X]) \ R} ⊆ {gK[X] ∩ [[f ]] | g ∈ [[f ]] ∩ A(K[X])}.
Finally we show that {gK[X] ∩ [[f ]] | g ∈ [[f ]] ∩ A(K[X])} ⊆ {P ∈ X([[f ]]) | P ∩ R = ∅}. Let g ∈
[[f ]] ∩ A(K[X]). Set P = gK[X] ∩ [[f ]]. It follows by Proposition 3.7 that P ∈ X([[f ]]). Since g 6∈ K we
have gK[X] ∩ R = {0}, hence P ∩ R = ∅.
One method of determining the divisor-class group of a Krull monoid is to identify the v-product decompositions of certain principal ideals into height-one prime ideals. Next, we present a useful tool which
can be used for that purpose.
Proposition 3.9. Let R be a factorial domain, K a field of quotients of R, X an indeterminate over
K, f ∈ R[X]• and L a quotient group of [[f ]]. For g ∈ [[f ]] set Pg = gK[X] ∩ [[f ]].
8
ANDREAS REINHART
1. (Pg Ph )v = Pgh for all g, h ∈ [[f ]].
2. If g ∈ [[f ]] ∩ A(K[X]), then vPg (x[[f ]]) = vg (x) for all x ∈ L.
3. If g ∈ [[f ]], Q ∈ X([[f ]]) and q ∈ Q ∩ A(R), then vQ (g[[f ]]) ≤ vq (ef (g)).
Proof. Let R be a system of representatives of [[f ]] ∩ A(R[X]) \ R. Note that R is finite.
1. First we show by induction that for all k ∈ N0 and t ∈ R it follows that (Ptk )v = Ptk . The assertion
is clear for k = 0. Now let k ∈ N0 and t ∈ R. Since Ptk+1 is a divisorial ideal of [[f ]] (by Proposition 3.7)
k+1
k
it follows that (Ptk+1 )v ⊆ Ptk+1 ( P
Qtk = (Pntt )v . We infer that (Pt R)v = Ptk+1 .
Q
Next we prove that P t∈R tnt = ( t∈R Pt )v for all (nt )t∈R ∈ N0 . Let (nt )t∈R ∈ NR
0 . QObserve that
P
and
P
are
v-coprime
for
all
distinct
s,
t
∈
R.
Since
[[f
]]
is
a
Krull
monoid
we
have
( t∈R Ptnt )v =
t
Ts
T
T
nt
n
Q
t
n
t∈R (Pt )v =
t∈R Pt t = ( t∈R t K[X]) ∩ [[f ]] = P t∈R tnt .
R
Now let g, h ∈
3.3 that there are some (nt )t∈R , (mQ
t )t∈R ∈ N0 such
Q [[f ]]. ntIt follows fromQLemma
nt +mt
mt
Q
that g ≃K[X] t∈R t and h ≃K[X] t∈R t . Therefore, Pgh = P t∈R tnt +mt = ( t∈R Pt
)v =
Q
Q
(( t∈R Ptnt )v ( t∈R Ptmt )v )v = (PQt∈R tnt PQt∈R tmt )v = (Pg Ph )v .
2. Let g ∈ [[f ]] ∩ A(K[X]) and x ∈ L. There are some y, z ∈ [[f ]] such that x = yz . By 1 we have
v (y)
v (y)+1
y ∈ g vg (y) K[X] ∩ [[f ]] = (Pg g )v and y 6∈ g vg (y)+1 K[X] ∩ [[f ]] = (Pg g
)v , and thus vPg (y[[f ]]) = vg (y).
Analogously, it follows that vPg (z[[f ]]) = vg (z), hence vPg (x[[f ]]) = vPg (y[[f ]])−vPg (z[[f ]]) = vg (y)−vg (z) =
vg (x).
3. Let g ∈ [[f ]], Q ∈ X([[f ]]) and q ∈ Q ∩ A(R). It is an easy consequence of Proposition 3.7, Corollary 3.8,
Q
Q
Q
v (g[[f ]])
v (g)
)v = ( t∈R Pt t )v = PQt∈R tvt (g) = Pg =
1 and 2 that ( P ∈X([[f ]]),P ∩R=∅ P vP (g[[f ]]) )v = ( t∈R Pt Pt
Q
ef (g)
e (g)
g
−1
. Therefore, ( P ∈X([[f ]]),P ∩R6=∅ P vP (g[[f ]]))v = g( fg [[f ]]∪[[f ]])v = (g[[f ]]∪
ef (g) [[f ]]∩[[f ]] = ( g [[f ]]∪[[f ]])
ef (g)[[f ]])v . We infer that vQ (g[[f ]]) = min{vQ (g[[f ]]), vQ (ef (g)[[f ]])}. Let S be a system of representatives
of [[f ]] ∩ A(R) such that q ∈ S. It follows from Lemma 3.1.3 and 3.1.4 that vQ (g[[f ]]) ≤ vQ (ef (g)[[f ]]) =
P
P
Q
vQ ( p∈S (p[[f ]])vp (ef (g)) ) = p∈S vp (ef (g))vQ (p[[f ]]) = p∈S vp (ef (g))δp,q = vq (ef (g)).
It is clear that every principal ideal of a monoid is a v-ideal. The converse is, of course, far from true. In
the last part of this section we describe when the v-ideals in Proposition 3.7 are principal.
Proposition 3.10. Let R be a factorial domain, K a field of quotients of R, X an indeterminate over
K, f ∈ R[X]• , and g ∈ [[f ]]. The following are equivalent:
1. gK[X] ∩ [[f ]] is a principal ideal of [[f ]].
g
2. gK[X] ∩ [[f ]] = d(g)
[[f ]].
3. d(gh) = d(g)d(h) for all h ∈ [[f ]].
Proof. Set e = ef (g). By Proposition 3.7 we have gK[X] ∩ [[f ]] = ge [[f ]] ∩ [[f ]].
1. ⇒ 2.: There is some a ∈ [[f ]] such that gK[X] ∩ [[f ]] = a[[f ]]. There are some h′ , h′′ ∈ [[f ]] such that
′
g
g
′′
= ah′′ . Therefore, e = d(g)h′ h′′ , and thus a = d(g)h
|R d(g),
a = ghe and d(g)
′′ . This implies that d(g)h
g
′′
×
hence h ∈ [[f ]] . Consequently, gK[X] ∩ [[f ]] = d(g) [[f ]].
2. ⇒ 3.: Let h ∈ [[f ]]. Since d(g)d(h) |R d(gh), it is sufficient to show that vp (d(gh)) ≤ vp (d(g))+vp (d(h))
for all p ∈ A(R). Let p ∈ A(R). There is some a ∈ [[f ]] such that vp ( d(ga)
d(a) ) = vp (e). Observe that
ga
d(a)pvp (e)
∈ gK[X] ∩ [[f ]] =
g
d(g) [[f ]].
Therefore, there is some k ∈ [[f ]] such that ad(g) = kpvp (e) d(a).
This implies that d(a)d(g) ≃R d(k)pvp (e) d(a). Consequently, vp (d(g)) ≥ vp (e) ≥ vp ( d(gh)
d(h) ), and thus
vp (d(gh)) ≤ vp (d(g)) + vp (d(h)).
g
g
3. ⇒ 1.: Obviously, e ≃R d(g), hence gK[X] ∩ R = ge [[f ]] ∩ [[f ]] = d(g)
[[f ]] ∩ [[f ]] = d(g)
[[f ]].
4. First main result and important preparation results
In this section we present the first of two main results of this work. It basically states that the divisor-class
group of a monadic submonoid of Int(R) generated by some f ∈ R[X]• (where R is a factorial domain)
ON THE DIVISOR-CLASS GROUP OF MONADIC SUBMONOIDS
9
is torsion-free. It is pointed out in the last section of this work that the condition “f ∈ R[X]• ” is crucial
here, since there are monadic submonoids of Int(Z) generated by some f ∈ Int(Z)• whose divisor-class
group is not torsion-free. We proceed by preparing several useful results (which might be interesting on
their own) to investigate the divisor-class group of arbitrary monadic submonoids of Int(R). We were
not able to give a complete description of the structure of the divisor-class group of arbitrary monadic
submonoids of Int(R). However, note that they are always finitely generated abelian groups.
Theorem 4.1. Let R be a factorial domain, K a field of quotients of R, X an indeterminate over K
and f ∈ R[X]• . Set r = |{P ∈ X([[f ]]) | P ∩ R 6= ∅}| − |{u[[f ]] | u ∈ A([[f ]]) ∩ R}|. Then Cv ([[f ]]) ∼
= Zr ,
|{[P ] | P ∈ X([[f ]])}| ≥ |{P ∈ X([[f ]]) | P ∩ R 6= ∅, P is not principal}|, and the following are equivalent:
1. [[f ]] is factorial.
2. Cv ([[f ]]) is finite.
3. d(gh) = d(g)d(h) for all g, h ∈ [[f ]].
Proof. There is some P ⊆ {P ∈ X([[f ]]) | P ∩ R 6= ∅} such that |{Q ∈ X([[f ]]) | u ∈ Q andQ
Q 6∈ P}| = 1
for every constant atom u of [[f ]]. Let h : ZP → Cv ([[f ]]) be defined by h((nP )P ∈P ) = [( P ∈P P nP )v ]
for all (nP )P ∈P ∈ ZP . Clearly, h is a well-defined group homomorphism. We show that h is a group
isomorphism.
First we show that h is surjective. It suffices to prove that for every P ∈ X([[f ]]) there is some (nQ )Q∈P ∈
ZP such that h((nQ )Q∈P ) = [P ]. Let P ∈ X([[f ]]).
Case 1. P ∈ P: Set nQ = δP,Q for all Q ∈ P. Then h((nQ )Q∈P ) = [P ].
Case 2. P ∩ R 6= ∅ and P 6∈ P: Clearly, there is some constant atom u of [[f ]] such
Q that u ∈ P . Set
nQ = −1 if Q ∈ P and u ∈ Q and nQ = 0 if Q ∈ P and u 6∈ Q. We have u[[f ]] = ( Q∈X([[f ]]),u∈Q Q)v by
Q
Q
Lemma 3.1.3, hence ( Q∈P,u∈Q Q−1 )v = u−1 P . Therefore, h((nQ )Q∈P ) = [( Q∈P,u∈Q Q−1 )v ] = [P ].
Case 3. P ∩ R = ∅: By Corollary 3.8 there is some g ∈ [[f ]] ∩ A(K[X]) such that P = gK[X] ∩ [[f ]]. It
follows from
QProposition 3.9 that vQ (g[[f ]]) = δP,Q for all Q ∈ X([[f ]]) such that Q ∩ R = ∅. Consequently,
g[[f ]] = (P m
∩ R 6= ∅}m . Since h
i=1 Qi )v for some m ∈ N0 and some sequence Q ∈ {Q ∈ X([[f ]]) | QQ
−1
is a group homomorphism, it follows by case 1 and case 2 that h((nQ )Q∈P ) = [( m
i=1 Qi )v ] for some
Qm
−1
P
−1
(nQ )Q∈P ∈ Z . Since ( i=1 Qi )v = g P we infer that h((nQ )Q∈P ) Q
= [P ].
Next we show that h is injective. Let (nP )P ∈P ∈ ZP be such that ( P ∈P P nP )v is principal. Let R
be a system
of representatives
of [[f ]] ∩ A(R[X]). By Lemma 3.4 there is some (nu )u∈R ∈ ZR such
Q
Q
that ( P ∈P P nP )v = u∈R unu [[f ]]. By Corollary 3.8 and Proposition 3.9 we infer that nu = 0 for all
u ∈ R \ R.
Q
Claim: If M ∈ X([[f ]]) and w ∈ M ∩ R ∩ R, then vM (( P ∈P P nP )v ) = nw . Let M ∈ X([[f ]]) and
w ∈ M ∩ R ∩ R.
by Lemma
and 3.1.4
u,w for all u ∈ R ∩ R.
Q It follows
Q 3.1.3
P that vM (u[[f ]]) = δP
nP
nu
Therefore,
v
((
P
)
)
=
v
(
u
[[f
]])
=
n
v
(u[[f
]])
=
M
v
M
P ∈P
u∈R
u∈R u M
u∈R∩R nu vM (u[[f ]]) =
P
n
δ
=
n
.
w
u∈R∩R u u,w
We have to show that nM = 0 for all M ∈ P. Let M ∈ P. Obviously, there is some w ∈ M ∩ R ∩ R.
It is clear that
Q]]) such that w ∈ Q and Q 6∈ P. It follows by the claim that
Q there is some Q ∈ X([[f
nM = vM (( P ∈P P nP )v ) = nw = vQ (( P ∈P P nP )v ) = 0.
Observe that |P| = r. This implies that Cv ([[f ]]) ∼
= Z|P| = Zr .
= ZP ∼
Set S = {P ∈ X([[f ]]) | P ∩ R 6= ∅, P is not principal}. To show that |{[P ] | P ∈ X([[f ]])}| ≥ |S| it is
sufficient to show that for all P, Q ∈ S such that h−1 ([P ]) = h−1 ([Q]) it follows that P = Q. Let P, Q ∈ S
be such that h−1 ([P ]) = h−1 ([Q]). Note that if P ∈ P, then h−1 ([P ]) = (δP,M )M∈P . Moreover, if P 6∈ P
and u ∈ P ∩ A(R), then h−1 ([P ])M = −1 if M ∈ P and u ∈ M and h−1 ([P ])M = 0 if M ∈ P and u 6∈ M .
In particular, if P 6∈ P, then (h−1 ([P ]))M 6= 1 for all M ∈ P. Therefore, we have either that P, Q ∈ P or
P, Q 6∈ P.
Case 1. P, Q ∈ P: Observe that (δP,M )M∈P = h−1 ([P ]) = h−1 ([Q]) = (δQ,M )M∈P , hence P = Q.
Case 2. P, Q 6∈ P: There are some u ∈ P ∩ A(R) and v ∈ Q ∩ A(R). Since h−1 ([P ]) = h−1 ([Q]), we infer
that for all M ∈ P, u ∈ M if and only if v ∈ M . Since P and Q are not principal and u[[f ]] and v[[f ]] are
10
ANDREAS REINHART
radical ideals of [[f ]] (by Lemma 3.1.3), there are some P ′ ∈ X([[f ]]) \ {P } and Q′ ∈ X([[f ]]) \ {Q} such that
u ∈ P ′ and v ∈ Q′ . It is immediately clear that P ′ , Q′ ∈ P. It follows that u, v ∈ P ′ , and thus u ≃[[f ]] v
by Lemma 3.1.4. The choice of P immediately implies that P = Q. Finally, we prove the equivalence.
1. ⇒ 2.: Trivial.
2. ⇒ 3.: This is an immediate consequence of Proposition 3.2.
3. ⇒ 1.: Since [[f ]] is a Krull monoid it is sufficient to show that every P ∈ X([[f ]]) is principal. Let
P ∈ X([[f ]]).
Case 1. P ∩ R 6= ∅: It follows by Proposition 3.2 that P is principal.
Case 2. P ∩ R = ∅: By Corollary 3.8 there is some g ∈ [[f ]] ∩ A(K[X]) such that P = gK[X] ∩ [[f ]].
Therefore, P is principal by Proposition 3.10.
We continue with a few result concerning the structure of height-one prime ideals.
S
Lemma 4.2. Let R be a factorial domain and f ∈ Int(R)• . Then {[[a]] | a ∈ [[f ]]} = {[[f ]] \ P ∈P P | P ⊆
X([[f ]])} is the set of divisor-closed submonoids of [[f ]].
S
Proof. Note that [[f ]] is a Krull monoid. Therefore, { P ∈P P | P ⊆ X([[f ]])} is the set of prime s-ideals
of [[f ]]. It is well known that the set of prime s-ideals of [[f ]] is finite. Consequently, the assertion follows
from [5, Lemma 2.2.1 and 2.2.3].
Lemma 4.3. Let T be a Krull monoid and H ⊆ T a saturated submonoid. Then for every P ∈ X(H)
there is some Q ∈ X(T ) such that P = Q ∩ H. In particular, X(H) is the set of minimal elements of
{Q ∩ H | Q ∈ X(T )} \ {∅}.
Proof. Let P ∈ X(H). Clearly, H is a Krull monoid, and thus P is a divisorial ideal of H. Therefore,
[5, Proposition 2.4.2.3] implies that PvT ∩ H =QP . There is some
Tn n ∈ N, some injective sequence
n
Q ∈ X(T )n and some n ∈ Nn such that PvT = i=1 (Qni i )vT = i=1 (Qni i )vT . This implies that P =
Tn
n
ni
thus P = (Qj j )vT ∩ H for some j ∈ [1, n]. Set Q = Qj . We infer that
i )vT ∩ H), and
i=1 ((Q
p
p
P = H (Qnj )vT ∩ H = T (Qnj )vT ∩ H = Q ∩ H.
Lemma 4.4. Let H be a monoid, I a finite non-empty set andT(LI )I∈I a family
of non-empty s-ideals
T
of H such that (LI ∪ LJ )v = H for all distinct I, J ∈ I. Then ( I∈I LI )v = I∈I (LI )v .
T
Q
Q
T
T
Proof. Observe that I∈I (LI )v = ( I∈I (LI )v )v = ( I∈I LI )v ⊆ ( I∈I LI )v ⊆ I∈I (LI )v .
The next result has some well-known analogue in the case of extension of Dedekind domains (e.g. see [8,
Proposition 2.10.2]).
Proposition 4.5. Let T be a Krull monoid, H ⊆ T a saturated submonoid, I ∈ Iv (H)• , and P ∈ X(H).
v (I )
Then vP (I) = max{⌈ vQQ(PvvT ) ⌉ | Q ∈ X(T ), Q ∩ H = P }.
T
v (I
)
Proof. Without restriction let I ⊆ P . Set m = max{⌈ vQQ(PvvT ) ⌉ | Q ∈ X(T ), Q ∩ H = P }. First we show
T
that vP (I) ≤ m. Set P = {Q ∩ H | Q ∈ X(T ), I ⊆ Q}. By Lemma 4.3 we have P ∈ P. For L ∈ P
T
set IL = Q∈X(T ),Q∩H=L (QvQ (IvT ) )vT ∩ H. Let M ∈ X(T ) be such that M ∩ H = P . Note that P m ⊆
⌈
vM (Iv
T
)
⌉
⌈
vM (Iv
T
)
⌉
⌈
vM (Iv
T
)
⌉
vM (Pv )⌈
vM (Iv
T
)
⌉
vM (Pv )
T
T
P vM (PvT ) ⊆ (PvT ) vM (PvT ) ⊆ ((M vM (PvT ) )vT ) vM (PvT ) ⊆ (M
)vT ⊆ (M vM (IvT ) )vT .
m
Therefore, (P )vH ⊆ (IP )vH , and thus vP ((IP )vH ) ≤ m.
Claim: For every L ∈ P there is some a ∈ N0 such that IL ⊇ La . Let L ∈ P. Set a = max{vQ (IvT ) |
T
T
Q ∈ X(T ), Q ∩ H = L}. Then IL ⊇ Q∈X(T ),Q∩H=L (QvQ (IvT ) ) ∩ H ⊇ Q∈X(T ),Q∩H=L LvQ (IvT ) = La .
it follows by the claim that
By the claim we have vP ((IL )vH ) = 0 for all L ∈ P \ {P }. Moreover,
T
(IL ∪ IM )vH = H for all distinct L, M ∈ P. Observe that IvT = Q∈X(T ),I⊆Q (QvQ (IvT ) )vT , hence I =
T
T
T
T
vQ (IvT )
)vT ∩H = L∈P IL . We infer by Lemma 4.4 that I = ( L∈P IL )vH = L∈P (IL )vH ,
Q∈X(T ),I⊆Q (Q
hence vP (I) = max{vP ((IL )vH ) | L ∈ P} = vP ((IP )vH ) ≤ m.
ON THE DIVISOR-CLASS GROUP OF MONADIC SUBMONOIDS
11
Q
Q
Next we show that m ≤ vP (I). We have I = ( L∈X(H) LvL (I) )vH ⊇ L∈X(H) LvL (I) , hence
vL (I)
Y Y
Y
Y
M vM (LvT )
=
(LvT )vL (I)
=
LvL (I)
IvT ⊇
vT
L∈X(H)
Y
Y
vT
L∈X(H)
M vM (LvT )vL (I)
L∈X(H) M∈X(T )
Y
M
vT
P
=
L∈X(H)
M∈X(T )
Y
Y
M vM (LvT )vL (I)
M∈X(T ) L∈X(H)
vM (LvT )vL (I)
vT
M∈X(T )
L∈X(H)
vT
vT
vT
=
=
.
Let Q ∈ X(T ) be such that Q ∩ H = P . Observe that vQ (LvT ) = δL,P vQ (PvT ) for all L ∈ X(H). We
P
v (I )
infer that vQ (IvT ) ≤ L∈X(H) vQ (LvT )vL (I) = vQ (PvT )vP (I), and thus vP (I) ≥ ⌈ vQQ(PvvT ) ⌉.
T
Let T be a monadic submonoid of Int(R), and let H be a monadic submonoid of Int(R) that is contained
in T . Suppose that we know how v-product decompositions of principal ideals (generated by atoms) in
T look like. Then Proposition 4.5 enables us to derive all v-product decompositions of principal ideals
(generated by atoms) in H.
Remark 4.6. Let R be a factorial domain, f ∈ Int(R)• and R a system of representatives of A([[f ]]) ∩ R.
Ṡ
Then {P ∈ X([[f ]]) | P ∩ R 6= ∅} = p∈R {P ∈ X([[f ]]) | p ∈ P }.
Proof. It is straightforward to prove the equality. That the union is a disjoint union follows from Lemma
3.1.4.
Corollary 4.7. Let R be a factorial domain, f ∈ Int(R)• , g ∈ [[f ]], I ∈ Iv ([[g]])• and P ∈ X([[g]]) such
that P ∩ R 6= ∅. Then vP (I) = max{vQ (Iv[[f ]] ) | Q ∈ X([[f ]]), Q ∩ [[g]] = P }.
Proof. Because of Proposition 4.5 it suffices to show that vQ (Pv[[f ]] ) = 1 for all Q ∈ X([[f ]]) such that
Q ∩ [[g]] = P . Let Q ∈ X([[f ]]) be such that Q ∩ [[g]] = P . By Remark 4.6 there is some u ∈ A([[g]]) ∩ P ∩ R.
Note that u ∈ A([[f ]]) ∩ Q ∩ R. It follows by Lemma 3.1.3 that u[[f ]] is a radical ideal of [[f ]]. Therefore,
u 6∈ (Q2 )v[[f ]] , hence P * (Q2 )v[[f ]] , and thus vQ (Pv[[f ]] ) = 1.
In the last part of this section we describe the structure of height-one prime ideals that contain constant
elements. We make an ad hoc definition to state the following results more easily. Let R be a factorial
domain, f ∈ Int(R)• , a ∈ [[f ]] ∩ R, and A ⊆ B ⊆ [[f ]]. We say that A is a-compatible if there is some
w ∈ [[f ]] such that a |R d(uw)
d(w) for all u ∈ A. Moreover, A is called maximal a-compatible in B if A is
maximal (with respect to inclusion) among the a-compatible subsets of B.
Lemma 4.8. Let R be a factorial domain, f ∈ Int(R)• , S a system of representatives of A([[f ]]) \ R, p
a constant atom of [[f ]], and A ⊆ [[f ]].
1. A is p-compatible if and only if (A ∪ {p})v[[f ]] ( [[f ]].
2. {P ∈ X([[f ]]) | p ∈ P } = {(Q ∪ {p})[[f ]] | Q ⊆ S, Q is maximal p-compatible in S}.
Proof. 1. Let L be a quotient group of [[f ]]. First let A be p-compatible. Then there is some w ∈ [[f ]]
w
such that p |R d(uw)
d(w) for all u ∈ A. Set z = pd(w) . Observe that z ∈ L. It follows by Lemma 3.1.1
uw
∈ [[f ]] for all u ∈ A. Therefore, z(A ∪ {p}) ⊆ [[f ]]. Suppose that z ∈ [[f ]]. Then
that zu = pd(w)
d(w)
= pd(w)
≃R d(z) ∈ R, a contradiction. Consequently, z 6∈ [[f ]]. We infer that [[f ]] ( ([[f ]] :L A ∪ {p}),
and thus (A ∪ {p})v[[f ]] ( [[f ]].
Next let (A ∪ {p})v[[f ]] ( [[f ]]. There is some z ∈ L \ [[f ]] such that z(A ∪ {p}) ⊆ [[f ]]. Set w = zp. It is clear
that w ∈ [[f ]]. Let u ∈ A. Then wu
p ∈ [[f ]], hence p |R d(wu) by Lemma 3.1.1. Since z 6∈ [[f ]] it follows by
1
p
Lemma 3.1.1 that p ∤R d(w), and thus p |R
d(wu)
d(w) .
12
ANDREAS REINHART
2. “⊆”: Let P ∈ X([[f ]]) be such that p ∈ P . Set Q = S ∩ P . It is obvious that (Q ∪ {p})[[f ]] ⊆ P . Let
x ∈ P . There is some u ∈ A([[f ]]) such that x ∈ u[[f ]].
Case 1. u ∈ R: It follows by Lemma 3.1.4 that u ≃[[f ]] p, and thus x ∈ p[[f ]] ⊆ (Q ∪ {p})[[f ]].
Case 2. u 6∈ R: There is some z ∈ S such that u ≃[[f ]] z, hence x ∈ z[[f ]] ⊆ (Q ∪ {p})[[f ]].
Consequently, P = (Q ∪ {p})[[f ]]. We have (Q ∪ {p})v[[f ]] ⊆ Pv[[f ]] = P , hence Q is p-compatible by 1. Let
Q′ ⊆ S be p-compatible such that Q ⊆ Q′ . It follows by 1 that there is some M ∈ X([[f ]]) such that
Q′ ∪ {p} ⊆ M . We infer that P = (Q ∪ {p})[[f ]] ⊆ (Q′ ∪ {p})[[f ]] ⊆ M , and thus M = P . Therefore,
Q′ ⊆ S ∩ P = Q. This shows that Q is maximal p-compatible in S.
“⊇”: Let Q ⊆ S be maximal p-compatible in S. Set P = (Q ∪ {p})[[f ]]. It is clear that p ∈ P . It
follows by 1 that P ⊆ M for some M ∈ X([[f ]]). Assume that P ( M . There is some u ∈ A([[f ]]) such
that u ∈ M \ P . It follows by Lemma 3.1.4 that u 6∈ R. Consequently, there is some w ∈ S such that
u ≃[[f ]] w. Set Q′ = Q ∪ {w}. It is clear that Q′ ⊆ M , hence Q′ is p-compatible in S by 1. It follows that
w ∈ Q′ = Q ⊆ P , and thus u ∈ P , a contradiction. We infer that P = M ∈ X([[f ]]).
Remark 4.9. Let R be a factorial domain, K a field of quotients of R, X an indeterminate over K,
f ∈ R[X]• , a ∈ [[f ]] ∩ R, and A ⊆ [[f ]]. Then A is a-compatible if and only if there is some primitive
g ∈ [[f ]] ∩ R[X] such that a |R d(ug)
d(g) for all u ∈ A.
Proof. First let A be a-compatible. Then there is some w ∈ [[f ]] such that a |R
d(uw)
d(w)
for all u ∈ A. By
Lemma 3.3 there are some b, c ∈ [[f ]] ∩ R and some primitive g ∈ [[f ]] ∩ R[X] such that w =
d(ug)
that a |R d(uw)
d(w) = d(g) for all u ∈ A. The converse is trivially satisfied.
bg
c .
We infer
5. Construction of divisor-class groups
In this section we present a few methods which can be used to construct “more complicated” divisor-class
groups. We start with a few preparatory results.
Lemma 5.1. Let R be a factorial domain, K a quotient field of R, X an indeterminate over K and
f, g ∈ R[X]• such that GCDR[X] (f, g) = R[X]× and d(rs) = d(r)d(s) for all r ∈ [[f ]] and s ∈ [[g]]. Then
[[f g]] = {xy | x ∈ [[f ]], y ∈ [[g]]}.
Proof. “⊆”: Let z ∈ [[f g]]. By Lemma 3.3 there are some a, b ∈ [[f g]] ∩ R and some primitive h ∈
[[f g]] ∩ R[X] such that h |R[X] f k g k for some k ∈ N, GCDR (a, b) = R× and z = bh
a . Since R[X] is
factorial, there are some h′ , h′′ ∈ R[X] such that h = h′ h′′ , h′ |R[X] f k and h′′ |R[X] g k . Consequently,
h′ ∈ [[f ]] and h′′ ∈ [[g]]. We infer that a |R d(h) = d(h′ )d(h′′ ), and thus there are some a′ , a′′ ∈ R
such that a = a′ a′′ , a′ |R d(h′ ) and a′′ |R d(h′′ ). Observe that a′ ∈ [[f ]] and a′′ ∈ [[g]]. Moreover,
b |R d((f g)l ) = d(f )l d(g)l for some l ∈ N, and thus there are some b′ , b′′ ∈ R such that b = b′ b′′ ,
′′ ′′
′ ′
b′ |R d(f )l and b′′ |R d(g)l . We infer that b′ ∈ [[f ]] and b′′ ∈ [[g]]. Set x = bah′ and y = b ah′′ . Note
that z = xy. Since a′ |R d(h′ ) we have x ∈ Int(R). Obviously, xa′ = b′ h′ ∈ [[f ]], and thus x ∈ [[f ]].
Analogously, it follows that y ∈ [[g]]. “⊇”: Trivial.
Lemma 5.2. Let R be a factorial domain, K a quotient field of R, X an indeterminate over K and
f, g ∈ Int(R)• such that GCDK[X] (f, g) = K[X]× and GCDR (f (x), g(x)) = R× for all but finitely many
x ∈ R. If h ∈ [[f ]], then [[f ]] ∩ [[hg]] = [[h]].
Proof. Let h ∈ [[f ]]. “⊆”: Let z ∈ [[f ]]∩[[hg]]. There is some k ∈ N such that z |Int(R) f k and z |Int(R) hk g k .
Since GCDK[X] (f, g) = K[X]× , it follows that z |K[X] hk . Therefore, zy = hk for some y ∈ K[X]. Let
w ∈ R be such that GCDR (f (w), g(w)) = R× . Then z(w) |R f (w)k and z(w) |R h(w)k g(w)k . Since
GCDR (z(w), g(w)k ) = R× we have z(w) |R h(w)k . We infer that y(x) ∈ R for all but finitely many
x ∈ R. Consequently, y ∈ Int(R), and thus z ∈ [[h]]. “⊇”: Trivial.
ON THE DIVISOR-CLASS GROUP OF MONADIC SUBMONOIDS
13
Lemma 5.3. Let R be a factorial domain, K a quotient field of R, X an indeterminate over K and
f, g ∈ Int(R)• such that GCDK[X] (f, g) = K[X]× and GCDR (f (z), g(z)) = R× for some z ∈ R. Then
for all x, x′ ∈ [[f ]] and y, y ′ ∈ [[g]] such that xy = x′ y ′ it follows that x ≃[[f ]] x′ and y ≃[[g]] y ′ .
Proof. Let x, x′ ∈ [[f ]] and y, y ′ ∈ [[g]] be such that xy = x′ y ′ . There is some k ∈ N such that x |Int(R) f k ,
x′ |Int(R) f k , y |Int(R) g k and y ′ |Int(R) g k . Since GCDK[X] (f, g) = K[X]× , we infer that GCDK[X] (x, y ′ ) =
GCDK[X] (x′ , y) = K[X]× , and thus there is some ε ∈ K[X]× = K × such that x = εx′ and y = ε−1 y ′ .
Moreover, x(z)y(z) = x′ (z)y ′ (z) and GCDR (x(z), y ′ (z)) = GCDR (x′ (z), y(z)) = R× , hence there is some
r ∈ R× such that x(z) = rx′ (z) and y(z) = r−1 y ′ (z). Since GCDR (x′ (z), y ′ (z)) = R× we have x′ (z) 6= 0
or y ′ (z) 6= 0.
Case 1. x′ (z) 6= 0: Observe that rx′ (z) = x(z) = εx′ (z), and thus ε = r.
Case 2. y ′ (z) 6= 0: Note that r−1 y ′ (z) = y(z) = ε−1 y ′ (z), hence ε = r.
In any case it follows that ε ∈ R× = [[f ]]× = [[g]]× . This implies that x ≃[[f ]] x′ and y ≃[[g]] y ′ .
Proposition 5.4. Let R be a factorial domain, K a quotient field of R, X an indeterminate over K and
f, g ∈ R[X]• such that GCDR[X] (f, g) = R[X]× , GCDR (f (x), g(x)) = R× for all but finitely many x ∈ R
and d(rs) = d(r)d(s) for all r ∈ [[f ]] and s ∈ [[g]]. If x ∈ [[f ]] and y ∈ [[g]], then Cv ([[xy]]) ∼
= Cv ([[x]])×Cv ([[y]]).
Proof. Without restriction let R be not a field. Let x ∈ [[f ]] and y ∈ [[g]]. We show that [[xy]]red ∼
=
[[x]]red × [[y]]red . Let ϕ : [[xy]]red → [[x]]red × [[y]]red be defined by ϕ(u[[xy]]× ) = (v[[x]]× , w[[y]]× ) if u ∈ [[xy]],
v ∈ [[x]] and w ∈ [[y]] are such that u = vw.
First we show that ϕ is well-defined. Let z ∈ [[xy]]red . Then there is some u ∈ [[xy]] such that z = u[[xy]]× .
By Lemmas 5.1 and 5.2 there are some v ∈ [[f ]] ∩ [[xy]] ⊆ [[f ]] ∩ [[xg]] = [[x]] and w ∈ [[g]] ∩ [[xy]] ⊆
[[g]] ∩ [[f y]] = [[y]] such that u = vw. Now let u′ ∈ [[xy]], v ′ ∈ [[x]] and w′ ∈ [[y]] be such that u′ = v ′ w′ and
z = u′ [[xy]]× . There is some ε ∈ [[xy]]× = [[x]]× = [[y]]× such that u = εu′ . Therefore, vw = εv ′ w′ , and
thus v ≃[[x]] εv ′ ≃[[x]] v ′ and w ≃[[y]] w′ by Lemma 5.3. Consequently, (v[[x]]× , w[[y]]× ) = (v ′ [[x]]× , w′ [[y]]× ).
Next we show that ϕ is an injective monoid homomorphism. It is clear that ϕ([[xy]]) = ([[x]], [[y]]).
Let z, z ′ ∈ [[xy]]red . There are some u, u′ ∈ [[xy]], v, v ′ ∈ [[x]] and w, w′ ∈ [[y]] such that z = u[[xy]]× ,
z ′ = u′ [[xy]]× , u = vw and u′ = v ′ w′ . Since uu′ = vv ′ ww′ we infer that ϕ(zz ′ ) = (vv ′ [[x]]× , ww′ [[y]]× ) =
(v[[x]]× , w[[y]]× )(v ′ [[x]]× , w′ [[y]]× ) = ϕ(z)ϕ(z ′ ). Now let ϕ(z) = ϕ(z ′ ). Then v[[x]] = v ′ [[x]] and w[[y]] = w′ [[y]],
and thus v ≃[[xy]] v ′ and w ≃[[xy]] w′ . This implies that u ≃[[xy]] u′ , hence z = z ′ .
It is clear that ϕ is surjective. We conclude by [5, Proposition 2.1.11.2] that Cv ([[xy]]) ∼
=
= Cv ([[xy]]red ) ∼
∼
Cv ([[x]]red ) × Cv ([[y]]red ) = Cv ([[x]]) × Cv ([[y]]).
By Theorem 4.1 we know that if H is a monadic submonoid generated by some f ∈ R[X]• , then d is
multiplicative on H if and only if H is factorial. The last proposition requires a less stringent form of
being multiplicative. Next we show that there is an interesting class of monadic submonoids for which d
satisfies this weak form of being multiplicative.
Lemma 5.5. Let R be a factorial domain, K a field of quotients of R, X an indeterminate over K,
a ∈ R and f, g ∈ R[X]• such that GCDR (f (a), g(a)) = R× and for all p ∈ A(R) and h ∈ A(R[X])
with (p |R f (a) and h |R[X] g) or (p |R g(a) and h |R[X] f ) it follows that p |R[X] h − h(a). Then
d(rs) = d(r)d(s) for all r ∈ [[f ]] and s ∈ [[g]].
Proof. Let r ∈ [[f ]] and s ∈ [[g]]. Let P be a system of representatives of A(R). To prove that d(rs) =
d(r)d(s), we need to show that for each p ∈ P there is some y ∈ R such that vp (r(y)) = min{vp (r(x)) |
x ∈ R} and vp (s(y)) = min{vp (s(x)) | x ∈ R}. Let p ∈ P. It is an easy consequence of Lemma 3.3
n
m
such that
that there are some b, c ∈ K • , n, m ∈ N, α ∈ Nn0 , β ∈ Nm
0 , f ∈ A(R[X]) and g ∈ A(R[X])
Qm βj
Qn
αi
r = b i=1 fi , s = c j=1 gj , fi |R[X] f for all i ∈ [1, n] and gj |R[X] g for all j ∈ [1, m].
Pn
If z ∈ RP
is such that vp (fi (z)) = 0 for all i ∈ [1, n], then vp (r(z)) = vp (b) + i=1 αi vp (fi (z)) = vp (b) ≤
n
vp (b) + i=1 αi vp (fi (v)) = vp (r(v)) for all v ∈ R, and thus vp (r(z)) = min{vp (r(x)) | x ∈ R}.
Analogously, if w ∈ R is such that vp (gj (w)) = 0 for all j ∈ [1, m], then vp (s(w)) = min{vp (s(x)) | x ∈ R}.
14
ANDREAS REINHART
Case 1. p ∤R (f g)(a): Observe that vp (fi (a)) = vp (gj (a)) = 0 for all i ∈ [1, n] and j ∈ [1, m]. Therefore,
vp (r(a)) = min{vp (r(x)) | x ∈ R} and vp (s(a)) = min{vp (s(x)) | x ∈ R}.
Case 2. p |R f (a): There is some y ∈ R such that vp (r(y)) = min{vp (r(x)) | x ∈ R}. Let j ∈ [1, m]. Note
that p |R[X] gj − gj (a). Consequently, p |R gj (y) − gj (a), and since p ∤R gj (a), we have vp (gj (y)) = 0.
This implies that vp (s(y)) = min{vp (s(x)) | x ∈ R}.
Case 3. p |R g(a): Goes along the same lines as case 2.
Proposition 5.6. Let R be a factorial domain, K a field of quotients of R, X an indeterminate over K,
a ∈ R and f, g ∈ R[X]• such that GCDR[X] (f, g) = R[X]× , GCDR (f (a), g(a)) = R× and d(rs) = d(r)d(s)
for all r ∈ [[f ]] and s ∈ [[g]]. Then Cv ([[f g]]) ∼
= Cv ([[f ]]) × Cv ([[g]]).
Proof. It follows from Lemmas 5.1 and 5.3 that [[f g]]red ∼
= [[f ]]red × [[g]]red . We conclude by [5, Proposition
2.1.11.2] that Cv ([[f g]]) ∼
= Cv ([[f g]]red ) ∼
= Cv ([[f ]]red ) × Cv ([[g]]red ) ∼
= Cv ([[f ]]) × Cv ([[g]]).
Corollary 5.7. Let R be a factorial domain, K a field of quotients of R, X an indeterminate over
K, a ∈ R and f, g ∈ R[X]• such that GCDR[X] (f, g) = R[X]× , GCDR (f (a), g(a)) = R× and for all
p ∈ A(R) and h ∈ A(R[X]) with (p |R f (a) and h |R[X] g) or (p |R g(a) and h |R[X] f ) it follows that
p |R[X] h − h(a). Then Cv ([[f g]]) ∼
= Cv ([[f ]]) × Cv ([[g]]).
Proof. This is an immediate consequence of Lemma 5.5 and Proposition 5.6.
6. Examples, important consequences and second main result
In this section we present several applications of the abstract theory. We start with a bunch of examples
that serve as counterexamples for various questions. We use the set of prime numbers as choice for the set
of representatives of A(Z). If Z is the base ring, then let all monadic submonoids be monadic submonoids
of Int(Z). Note that if H is an atomic S
monoid (e.g. H is a Krull monoid), Q is a system of representatives
of A(H), and P ∈ X(H), then P = u∈Q∩P uH. We will use this fact without citation. It was used
implicitly in the proof of Lemma 4.8. Also note that if T ⊆ H is a divisor-closed submonoid of H and
z ∈ H, then either zH ∩ T = zT or zH ∩ T = ∅.
Example 6.1. Let X be an indeterminate over Q. Set u1 = 2, u2 = 3, u3 = X, u4 = X − 1,
u u2 u
u u3 u
u5 = X − 2, u6 = u32u4 , u7 = u42u5 , u8 = u3 u64 u5 , u9 = 3 124 5 and u10 = 3 244 5 . For J ⊆ [1, 10], set
UJ = {uj | j ∈ J}. Set H = [[u3 u4 u5 ]], S = [[
u2 u
u23 u34 u25
]],
8
T = [[
u3 u34 u35
24 ]],
V = [[
u3 u34 u5
]],
6
W = [[
u3 u24 u5
12 ]],
Y = [[u3 u4 ]], and Z = [[ 32 4 ]].
• S = H \ U{1,4} H, T = H \ U{1,2,3,4,6} H, and V = H \ U{2,3,5,6,7} H.
• W = H \ U{1,2,3,4,5,6,7} H, Y = H \ U{2,5,7,8,9,10} H, and Z = H \ U{1,2,4,5,7,8,9,10}H.
• {uH | u ∈ A(H)} = {u1 H, u2 H, u3 H, u4 H, u5 H, u6 H, u7 H, u8 H, u9 H, u10 H}.
• {uS | u ∈ A(S)} = {u2 S, u3 S, u5 S, u6 S, u7 S, u8 S, u9 S, u10 S}.
• {uT | u ∈ A(T )} = {u5 T, u7 T, u8 T, u9 T, u10 T }, and {uV | u ∈ A(V )} = {u1 V, u4 V, u8 V, u9 V, u10 V }.
• {uW | u ∈ A(W )} = {u8 W, u9 W, u10 W }, and {uY | u ∈ A(Y )} = {u1 Y, u3 Y, u4 Y, u6 Y }.
• {uZ | u ∈ A(Z)} = {u3 Z, u6 Z}.
• X(H) = {U{1,3,5,6,8,9}H, U{1,3,5,7,8,9} H, U{1,4} H, U{2,3,6} H, U{2,4,6,7,9,10} H, U{2,5,7} H, U{3,6,8,9,10} H,
U{4,6,7,8,9,10} H, U{5,7,8,9,10} H}.
• X(S) = {U{2,3,6} S, U{2,5,7} S, U{2,6,7,9,10} S, U{3,5,6,8,9} S, U{3,5,7,8,9} S, U{3,6,8,9,10} S, U{5,7,8,9,10} S,
U{6,7,8,9,10} S}.
• X(T ) = {U{5,7} T, U{5,8,9} T, U{7,9,10} T, U{8,9,10} T }, and X(V ) = {U{1,4} V, U{1,8,9} V, U{4,9,10} V,
U{8,9,10} V }.
• X(W ) = {U{8,9} W, U{9,10} W }, X(Y ) = {U{1,3} Y, U{1,4} Y, U{3,6} Y, U{4,6} Y }, and X(Z) =
{u3 Z, u6 Z}.
• Cv (H) ∼
= Cv (S) ∼
= Z4 , Cv (T ) ∼
= Cv (V ) ∼
= Cv (Y ) ∼
= Z, and Cv (W ) ∼
= Z/2Z.
• Z is factorial and u3 Q[X] ∩ Z = U{3,6} Z = Z \ Z × is not divisorial.
ON THE DIVISOR-CLASS GROUP OF MONADIC SUBMONOIDS
15
min{2k+m,l,k+2m} min{k,l,m}
Proof. It is straightforward to prove that d(uk3 ul4 um
3
for all k, l, m ∈
5 ) = 2
N0 . Now one can show by a careful case analysis that {x ∈ N30 | x is (u3 , u4 , u5 )-irreducible} =
{(1, 0, 0), (0, 1, 0), (0, 0, 1), (1, 1, 0), (0, 1, 1), (1, 1, 1), (1, 2, 1), (1, 3, 1)}. It follows that {uH | u ∈ A(H)} =
{ui H | i ∈ [1, 10]}, by Proposition 3.5.
Note that each primitive g ∈ H ∩ R[X] is associated to some element of the form uk3 ul4 um
5 for some
l+1 m
d(uk
u5 )
d(u3 u24 u5 )
3 u4
, then l < min{k, m}, and
l m
d(u3 u4 u5 ) . If k, l, m ∈ N0 are such that 2 | d(uk
u
3 4 u5 )
k l m
d(u u u u )
that 2 ∤ d(uj k 3ul u4 m5) for every j ∈ {3, 5, 6, 7, 8, 9, 10}. Consequently, U{4} is maximal
3 4 5
k, l, m ∈ N0 . We have 2 |
then it is easy to show
2-compatible in A(H) \ Z. Analogously, we have U{1,3,5,6,8,9} and U{1,3,5,7,8,9} are maximal 2-compatible
in A(H) \ Z, and U{2,3,6} , U{2,4,6,7,9,10} , and U{2,5,7} are maximal 3-compatible in A(H) \ Z. Clearly, we
have u3 Q[X] ∩ H = U{3,6,8,9,10} H, u4 Q[X] ∩ H = U{4,6,7,8,9,10} H, and u5 Q[X] ∩ H = U{5,7,8,9,10} H. It
follows from Corollary 3.8, Remark 4.6, Lemma 4.8.2, and Remark 4.9 that X(H) can be expressed as
asserted.
It is easy to see that the other monoids (S, T, V, W, Y and Z) are all monadic submonoids of H. They
are, therefore, complements of unions of height-one prime ideals of H by Lemma 4.3. We show that
S = H \ U{1,4} H. The corresponding equalities for the remaining monoids can be proved in analogy. Set
A = H \ U{1,4} H. First, note that A = [[h]] for some h ∈ H by Lemma 4.3. Since h is a product of atoms
of H, and h ∈ A we have h is associated to a product of elements of U{2,3,5,6,7,8,9,10} . Consequently,
2
Q
u u2 u
u3 u4 u5 2 u3 u4 u5 3
A = [[u2 u3 10
) ( 12 ) ]] = [[u2 u3 u24 u5 3 124 5 ]] = S. It is now simple to prove the
i=5 ui ]] = [[u2 (
2
remaining statements concerning sets of atoms. It is clear that Z is factorial (since every height-one
prime ideal of Z is principal). Moreover, we have u3 Q[X] ∩ Z = u3 Q[X] ∩ H ∩ Z = U{3,6,8,9,10} H ∩ Z =
U{3,6} Z = Z \ Z × , since every non-unit of Z is a multiple of u3 or u6 . If Z \ Z × is divisorial, then Z is a
discrete valuation monoid, and hence it has only one atom up to associates, a contradiction. Therefore,
Z \ Z × is not divisorial.
It remains to show all statements about divisor-class groups. It follows from Theorem 4.1 that Cv (H) ∼
=
Z4 . We only show that Cv (W ) ∼
= Z/2Z. The other assertions follow in analogy. Let (Pi )9i=1 be the
sequence of height-one prime ideals of H in the above order (i.e., P1 = U{1,3,5,6,8,9} , P2 = U{1,3,5,7,8,9} H,
and so on). We determine the v-product decompositions of principal ideals of H generated by atoms.
Set f = u3 u4 u5 . The definition of ef can be found in section 3. It is straightforward to prove that
ef (u4 ) = ef (u6 ) = ef (u7 ) = ef (u9 ) = 6. We infer by Proposition 3.9 that u1 H = (P1 P2 P3 )v ,
u2 H = (P4 P5 P6 )v , u4 H = (P3 P5 P8 )v , u6 H = (P1 P4 P5 P7 P8 )v , u7 H = (P2 P5 P6 P8 P9 )v , and u9 H =
(P1 P2 P5 P7 P82 P9 )v . Therefore, u3 H = uu1 u4 6 H = (P12 P2 P4 P7 )v , u5 H = uu1 u4 7 H = (P12 P22 P6 P9 )v , u8 H =
u1 u9
u4 u9
2 2
2
3
u4 H = (P1 P2 P7 P8 P9 )v , and u10 H = u1 H = (P5 P7 P8 P9 )v . Set Q1 = U{8,9} W , and Q2 = U{9,10} W .
Then {P ∈ X(H) | P ∩ W = Q1 } = {P1 , P2 }, and {P ∈ X(H) | P ∩ W = Q2 } = {P5 }. We
have vP1 ((Q1 )vH ) = min{vP1 (u8 H), vP1 (u9 H)} = 1, vP2 ((Q1 )vH ) = min{vP2 (u8 H), vP2 (u9 H)} = 1,
and vP5 ((Q2 )vH ) = min{vP5 (u9 H), vP5 (u10 H)} = 1. We infer by Proposition 4.5 that vQ1 (u8 W ) =
v 2 (u8 H)
v 1 (u8 H)
⌉, ⌈ vP P((Q
⌉} = 2, vQ2 (u10 W ) = 2, and vQ1 (u9 W ) = vQ2 (u9 W ) = 1. Consequently,
max{⌈ vP P((Q
1 )v )
1 )v )
1
H
2
H
u8 W = (Q21 )vW , u9 W = (Q1 Q2 )vW , and u10 W = (Q22 )vW . This implies that [Q1 ] = [Q2 ], and thus [Q1 ]
is an element of order 2 which generates Cv (W ). Therefore, Cv (W ) ∼
= Z/2Z.
The last example shows that there is some factorial domain R and f, g, h, k ∈ Int(R)• such that k ∈ [[h]],
and
• [[f ]] satisfies the equivalent conditions in Proposition 3.2 and yet Cv ([[f ]]) is infinite.
• Cv ([[g]]) is finite, and yet [[g]] is not factorial.
• There are some I ∈ Iv ([[k]])• and P ∈ X([[k]]) such that vP (I) < max{vQ (Iv[[h]] ) | Q ∈ X([[h]]), Q ∩
[[k]] = P }.
16
ANDREAS REINHART
Observe that U{3,6,8,9,10} H ∩ Z = U{3,6} Z in the last example. We know that H is a Krull monoid,
U{3,6,8,9,10} H is a height-one prime ideal of H (and hence it is divisorial), Z is a monadic submonoid of
H, and yet U{3,6,8,9,10} H ∩ Z is not a divisorial ideal of Z.
Recall that if G is an additive abelian group, then the Davenport constant of G (denoted by D(G)) is
defined to be the supremum of all lengths of nonempty minimal zero-sum sequences of G (see [5, Definition
1.4.8]).
Lemma 6.2. Let R be a factorial domain, K a field of quotients of R, X an indeterminate over K,
p ∈ A(R), n ∈ N≥2 and f ∈ (A(R[X]) \ R)n a sequence of pairwise non-associated elements of R[X]
Qn
Qn
such that
d( i=1 firi ) = pmin{ri |i∈[1,n]} for all r ∈ Nn0 . Set H =
[[ i=1 fi ]]. Then {uH | u ∈ A(H)} =
Qn
Qn
fj
f
{pH, i=1 i H} ∪ {fi H | i ∈ [1, n]}, X(H) = {pH ∪ fi H, fi H ∪ j=1 H | i ∈ [1, n]}, Cv (H) ∼
= Zn−1 , all
p
p
proper divisor-closed submonoids of H are factorial, for every P ∈ X(H) there is some Q ∈ X(H) such
that (P Q)v is a principal ideal of H and D({[P ] | P ∈ X(H)}) ≥ n.
Proof. Set ei = (δi,j )nj=1 for each i ∈ [1, n]. Observe that {m ∈ Nn0 | m is f -irreducible} = {ei | i ∈ [1, n]}∪
Qn
Pn
fi
H} ∪ {fi H | i ∈
{ i=1 ei }. Therefore, Proposition 3.5 implies that {uH | u ∈ A(H)} = {pH, i=1
p
Qn
f
i
[1, n]}. Note that R = {p} is a system of representatives of A(H) ∩ R and S = {fi | i ∈ [1, n]} ∪ { i=1
}
p
is a system of representatives of A(H) \ R. It follows by Remark 4.6, Lemma 4.8.2, and Remark 4.9
that {P ∈ X(H) | P ∩ R 6= ∅} = {P ∈ X(H) | p ∈ P } = {pH ∪ fi H | iQ∈ [1, n]}. Moreover, we
n
fj
have {P ∈ X(H) | P ∩ R = ∅} = {fi K[X] ∩ H | i ∈ [1, n]} = {fi H ∪ j=1
H | i ∈ [1, n]} by
p
n−1
∼
Corollary 3.8. Consequently,
C
(H)
Z
by
Theorem
4.1.
For
i
∈
[1,
n],
set
S
=
v
i = H \ (pH ∪ fi H)
Q
n
fj
and Ti = H \ (fi H ∪ j=1
H). Let i ∈ [1, n]. By Lemma 4.2, Si and Ti are divisor-closed submonoids
p
of H and every proper divisor-closedQsubmonoid of H is a divisor-closed submonoid of Sj or Tj for some
n
fj
Si | k ∈ [1, n] \ {i}} and X(Ti ) = {pTi , fk Ti | k ∈ [1, n] \ {i}}
j ∈ [1, n]. Note that X(Si ) = {fk Si , j=1
p
by Lemma 4.3. This implies that Si and Ti are factorial (since all of their height-one prime ideals are
principal), and hence every proper divisor-closed submonoid of H is factorial.
Q
n
fj
It is an easy consequence of Proposition 3.9.3 that fi H = ((pH ∪ fi H)(fi H ∪ j=1
H))v for every
p
i ∈ [1, n]. Let P Q∈ X(H). Clearly, there are some i ∈ [1, n] and Q ∈ X(H) such that {P, Q} =
n
fj
H}. It follows that (P Q)v is a principal ideal of H.
{pH ∪ fi H, fi H ∪ j=1
Qn p
Note that pH = ( i=1 pH∪fi H)Qv (since pH is a radical ideal of H). Since p is an atom of H, it follows that
n
no nonempty v-subproduct of ( i=1 pH ∪ fi H)v is principal. Therefore, D({[P ] | P ∈ X(H)}) ≥ n.
Next we recall a simple irreducibility criterion similar to Eisenstein’s criterion. We include a proof for
the sake of completeness. If R is a factorial domain, K is a field of quotients of R, X is an indeterminate
over P
K, and f ∈ R[X] with deg(f ) = n ∈ N0 , then let (fi )ni=0 ∈ Rn+1 be the unique sequence such that
f = ni=0 fi X i .
Lemma 6.3. Let R be a factorial domain, K a field of quotients of R, X an indeterminate over K,
p ∈ A(R), and f ∈ R[X] \ R primitive such that n = deg(f ), p ∤R f0 , p2 ∤R fn , and p |R fi for all
i ∈ [1, n]. Then f ∈ A(R[X]).
Proof. Clearly, f ∈ R[X]• \ R[X]×. Let g, h ∈ R[X] be such that f = gh. Let l = deg(g) and m = deg(h).
Without restriction we can assume that l ≤ m. Observe that p |R gl hm , p2 ∤R gl hm , and m ≥ 1. We need
to show that g ∈ R× .
Case 1. p |R gl and p ∤R hm : We prove that p |R gl−i for all i ∈ [0, l] by induction on i. Let i ∈ [0, l]
Pi
be such that p |R gl−j for each j ∈ [0, i − 1]. It follows that p |R fl+m−i = j=0 gl−j hm+j−i , and hence
p |R gl−i hm . Consequently, p |R gl−i . We infer that p |R[X] g |R[X] f , and thus p |R f0 , a contradiction.
Case 2. p ∤R gl and p |R hm : We prove that l = 0. Assume to the contrary that l > 0. We show
by induction on i that p |R hm−i for all i ∈ [0, m]. Let i ∈ [0, m] be such that p |R hm−j for every
ON THE DIVISOR-CLASS GROUP OF MONADIC SUBMONOIDS
17
Pmin{i,l}
j ∈ [0, i − 1]. Note that p |R fl+m−i =
gl−j hm+j−i , and thus p |R gl hm−i . We infer that
j=0
p |R hm−i . Consequently, p |R[X] h |R[X] f , and hence p |R f0 , a contradiction. It follows that l = 0, and
thus g ∈ R. Since f is primitive we have g ∈ R× .
In the beginning of this section we have provided examples of monadic submonoids of Int(Z) whose
divisor-class group is a torsion group or torsion-free, but not trivial. Next we provide an example of a
monadic submonoid of Int(Z) whose divisor-class group is neither torsion-free nor a torsion group.
Example 6.4. Let X be Q
an indeterminate over Q. Set p1 = 7, p2 = 13, Q
p3 = 19, p4 = 31, p5 = 37,
7
7
p6 = 43, p7 = 67, a = i=1 pi , f = (aX + 1)(aX + 2)(aX + 3), g = i=1 (6Xf + pi ), and H =
(aX+1)(aX+2)2 (aX+3)g
[[
]]. Then Cv (H) ∼
= Z/2Z × Z6 .
12
Proof. It is straightforward to show that d((aX + 1)k (aX + 2)l (aX + 3)m ) = 2min{2k+m,l,k+2m} 3min{k,l,m}
2
(aX+3)
for all k, l, m ∈ N0 . As in Example 6.1 we obtain that Cv ([[ (aX+1)(aX+2)
]]) ∼
= Z/2Z. Moreover, one
12
Q7
min{b
|i∈[1,7]}
i
bi
7
for all b ∈ N0 . Therefore, it follows from Lemma 6.2
can show that d( i=1 (aXf + pi ) ) = p1
that Cv ([[g]]) ∼
= Z6 . It is clear that GCDZ[X] (f, g) = GCDZ[X] (f, a) = GCDZ[X] (6, a) = Z[X]× . Along the
same lines we infer that GCDZ (f (x), g(x)) = GCDZ (f (x), a) = GCDZ (6, a) = Z× for each x ∈ Z. It is
clear that aX + 1, aX + 2, aX + 3 ∈ A(Z[X]). It follows by Lemma 6.3 that 6Xf + pi ∈ A(Z[X]) for every
i ∈ [1, 7]. Consequently, it is obvious that for all p ∈ A(Z) and h ∈ A(Z[X]) with (p |Z f (0) and h |Z[X] g)
2
(aX+3)
or (p |Z g(0) and h |Z[X] f ) it follows that p |Z[X] h − h(0). Since (aX+1)(aX+2)
∈ [[f ]], we infer by
12
(aX+1)(aX+2)2 (aX+3)
∼
∼
]]) × Cv ([[g]]) = Z/2Z × Z6 .
Proposition 5.4 and Lemma 5.5 that Cv (H) = Cv ([[
12
Now we present a result which enables us to construct examples of monadic submonoids of Int(Z) whose
divisor-class group is torsion-free with prescribed rank.
Proposition 6.5. Let R be a factorial domain, K a field of quotients of R, X an indeterminate over K,
P a system of representatives of A(R), n ∈ N, a ∈ Rn , and p ∈ P n such that for all i ∈ [1, n], p1 |R ai − 1,
×
ai + pk R ∈ (R/pk R)
1, then n = |{pj + p1 R | j ∈ [1, n]}| = |R/p1 R| <
Qn for all k ∈ [2, n], and if i >
|R/pi R|. Set H = [[ i=1 (ai X − pi )]]. Then Cv (H) ∼
= Zn−1 , for every P ∈ X(H) there is some Q ∈ X(H)
such that (P Q)v is a principal ideal of H, and D({[P ] | P ∈ X(H)}) ≥ n.
Qn
min{ri |i∈[1,n]}
for all r ∈ Nn0 . Let
Proof. By Lemma 6.2 it is sufficient to show that
d( i=1 (ai X −pi )ri ) = p1
P
n
n
r ∈ N0 and q ∈ P. We need to show that min{
Pn i=1 ri vq (ai x − pi ) | x ∈ R} = δq,p1 min{r
Pn i | i ∈ [1, n]}.
Case 1. q 6= pi for all i ∈ [1, n]: Observe that i=1 ri vq (ai q−pi ) = 0. Therefore, min{ i=1 ri vq (ai x−pi ) |
x ∈ R} = 0.
Case 2. q = pj for some j ∈ [2, n]: Since n < |R/qR|, there is some y ∈ R such that q ∤R y − pk
×
for all k ∈ [1, n]. Let k ∈ [1, n]. Since ak + qR ∈ (R/qR)
Pn , there is some z ∈ R such that q |R
ak z −Py. Consequently, q ∤R ak z − pk . It follows that
i=1 ri vq (ai z − pi ) = 0. This implies that
min{ ni=1 ri vq (ai x − pi ) | x ∈ R} = 0.
Case 3. q = p1 : Let y ∈ R. Since R/qR = {pi +qR | i ∈ [1, n]}, there is some j ∈ [1, n] P
such that q |R y−pj .
n
Since q |R aj − 1, we infer that qP
|R aj y − pj . Therefore, min{ri | i ∈ [1, n]} ≤ rj ≤ i=1 ri vq (ai y − pi ),
n
hence min{ri | i ∈ [1, n]} ≤ min{ i=1 ri vq (ai x − pi ) | x ∈ R}.
There is some k ∈ [1, n] such that min{ri | i ∈ [1, n]} = rk . We show that vq (ak z − pk ) = 1 for some
z ∈ R. Since q |R ak − 1, there is some y ∈ R such that q |R ak y − pk . If q 2 ∤R ak y − pk , then set z = y.
Now suppose that q 2 |R ak y − pk . Set z = q + y. Then vq (ak z − pk ) = 1.
Let jP
∈ [1, n] \ {k}. Then q ∤R pj − pP
k . Since q |R aj z − z, we infer that vq (aj z − pj ) = 0. Consequently,
n
n
min{ i=1 ri vq (ai x − pi ) | x ∈ R} ≤ i=1 ri vq (ai z − pi ) = rk = min{ri | i ∈ [1, n]}.
The following result is a useful application of Corollary 5.7.
Proposition 6.6. Let R be a factorial domain, K a field of quotients of R, X an indeterminate over
K, and P a system of representatives of A(R). Let k ∈ N, and (Pi )ki=1 a finite sequence of finite and
18
ANDREAS REINHART
pairwise disjoint subsets of P such that for every i ∈ [1, k] there is
Qsome p ∈ Pi for which |Pi | = |{r + pR |
r ∈ Pi }| = |R/pR| < |R/qR| < ∞ for all q ∈ Pi \ {p}, and p |R a∈Sk
Pj a − 1. Set
j=1,j6=i
g=
k Y
Y
j=1
Pk
Then Cv ([[g]]) ∼
=Z
i=1
|Pi |−k
b∈Pj
Y
a∈
Sk
i=1,i6=j
a X −b .
Pi
.
Q
Q
a)X − b). It is sufficient to show by induction on
Proof. For j ∈ [1, k], set fj = b∈Pj (( a∈Sk
i=1,i6=j Pi
Pj
Qj
|P
|−j
i
∼
for every j ∈ [1, k]. It follows immediately from Proposition 6.5
j that Cv ([[ i=1 fi ]]) = Z i=1
Qj−1
that Cv ([[fj ]]) ∼
= Z|Pj |−1 for every j ∈ [1, k]. Let j ∈ [2, k]. Set f = i=1 fi and g ′ = fj . Note that
GCDR[X] (f, g ′ ) = R[X]× . We show that GCDR (f (0), g ′ (0)) = R× and for all p ∈ A(R) and h ∈ A(R[X])
such that (p |R f (0) and h |R[X] g ′ ) or (p |R g ′ (0) and h |R[X] f ) we have p |R[X] h − h(0). Observe
Q
Q
Sj−1
that f (0) ≃R b∈Sj−1 Pi b and g ′ (0) ≃R b∈Pj b. Since i=1 Pi and Pj are disjoint, it follows that
i=1
GCDR (f (0), g ′ (0)) = R× . Let p ∈ A(R) and h ∈ A(R[X]).
S
Case 1. p |R f (0) and h |R[X] g ′ : Of course, p ≃R b for some b ∈ ki=1,i6=j Pi and h ≃R[X]
Q
Q
a)X ≃R[X] h − h(0).
a)X − c for some c ∈ Pj . Therefore, p ≃R[X] b |R[X] ( a∈Sk
( a∈Sk
i=1,i6=j Pi Q
i=1,i6=j Pi
′
a)X−c
Case 2. p |R g (0) and h |R[X] f : It is clear that p ≃R b for some b ∈ Pj and h ≃R[X] ( a∈Sk
i=1,i6=l Pi
Q
for some l ∈ [1, j − 1] and some c ∈ Pl . Consequently, p ≃R[X] b |R[X] ( a∈Sk
a)X
≃
h
−
h(0).
R[X]
i=1,i6=l Pi
Pj
Pj−1
We infer by Corollary 5.7 that Cv ([[f g ′ ]]) ∼
= Z i=1 |Pi |−j .
= Z i=1 |Pi |−j+1 × Z|Pj |−1 ∼
= Cv ([[f ]]) × Cv ([[g ′ ]]) ∼
Example 6.7. Let g = (95095X + 2)(95095X + 3)(6X + 5)(6X + 7)(6X + 11)(6X + 13)(6X + 19) ∈ Z[X].
Then Cv ([[g]]) ∼
= Z5 .
Proof. This follows from Proposition 6.6 with k = 2, P1 = {2, 3}, and P2 = {5, 7, 11, 13, 19}.
There are many important invariants which can describe the structure of factorizations. Two of them that
are commonly used are the elasticity ρ(H) and the tame degree t(H) of a monoid H. For the definitions of
the elasticity and the tame degree we refer to [5, Definitions 1.4.1 and 1.6.4]. In what follows we want to
provide a class of rings of integer-valued polynomials over factorial domains where both of these invariants
are infinite. Note that if H is an atomic monoid and T ⊆ H is an atomic divisor-closed submonoid, then
ρ(T ) ≤ ρ(H) and t(T ) ≤ t(H).
Theorem 6.8. Let R be a factorial domain, K a field of quotients of R, X an indeterminate over K
and P a system of representatives of A(R). Let (Pi )i∈N be a sequence of finite subsets of P such that for
every i ∈ N there is some p ∈ Pi for which i < |Pi | = |{r + pR | r ∈ Pi }| = |R/pR| < |R/qR| < ∞ for all
q ∈ Pi \ {p}. Then ρ(Int(R)) = t(Int(R)) = ∞.
Q
Proof. For i ∈ N set Hi = [[ a∈Pi (X − a)]]. By Proposition 6.5 we infer that {[P ] | P ∈ X(Hi )} =
{[P −1 ] | P ∈ X(Hi )} and D({[P ] | P ∈ X(Hi )}) > i for all i ∈ N. It follows from [5, Theorem 3.4.10] that
ρ(Int(R)) ≥ ρ(Hi ) ≥ D({[P ] | P ∈ X(Hi )})/2 > i/2, and t(Int(R)) ≥ t(Hi ) ≥ D({[P ] | P ∈ X(Hi )}) > i
for every i ∈ N. This implies that ρ(Int(R)) = t(Int(R)) = ∞.
Acknowledgement. This work was supported by the Austrian Science Fund FWF, Project number
P26036-N26.
ON THE DIVISOR-CLASS GROUP OF MONADIC SUBMONOIDS
19
References
[1] P.-J. Cahen, J.-L. Chabert, Integer-valued polynomials, Mathematical Surveys and Monographs, 48, American Mathematical Society, Providence, RI, 1997, xx+322 pp.
[2] P.-J. Cahen, S. Gabelli, E. Houston, Mori domains of integer-valued polynomials, J. Pure Appl. Algebra 153 (2000),
no. 1, 1–15.
[3] S. Frisch, Relative polynomial closure and monadically Krull monoids of integer-valued polynomials, Multiplicative
Ideal Theory and Factorization Theory, Springer, to appear.
[4] S. Frisch, A construction of integer-valued polynomials with prescribed sets of lengths of factorizations, Monatsh. Math.
171 (2013), no. 3–4, 341–350.
[5] A. Geroldinger, F. Halter-Koch, Non-unique factorizations: Algebraic, combinatorial and analytic theory, Pure Appl.
Math., Chapman and Hall/CRC, Boca Raton, FL, 2006, xxii+700 pp.
[6] F. Kainrath, Factorization in Krull monoids with infinite class group, Colloq. Math. 80 (1999), no. 1, 23–30.
[7] A. Reinhart, On monoids and domains whose monadic submonoids are Krull, in Commutative Algebra - Recent
Advances in Commutative Rings, Integer-valued Polynomials, and Polynomial Functions, M. Fontana, S. Frisch, and
S. Glaz, eds., Springer, 2014.
[8] A. Reinhart, A Note on conductor ideals, Comm. Algebra, to appear.
Institut für Mathematik und wissenschaftliches Rechnen, Karl-Franzens-Universität, NAWI Graz, Heinrichstrasse 36, 8010 Graz, Austria
E-mail address: andreas.reinhart@uni-graz.at
| 0math.AC |
EMPIRICAL L2 -DISTANCE TEST STATISTICS FOR ERGODIC
DIFFUSIONS
arXiv:1801.04848v1 [math.PR] 15 Jan 2018
A. DE GREGORIO AND S.M. IACUS
Abstract. The aim of this paper is to introduce a new type of test statistic for simple null
hypothesis on one-dimensional ergodic diffusion processes sampled at discrete times. We deal
with a quasi-likelihood approach for stochastic differential equations (i.e. local gaussian approximation of the transition functions) and define a test statistic by means of the empirical
L2 -distance between quasi-likelihoods. We prove that the introduced test statistic is asymptotically distribution free; namely it weakly converges to a χ2 random variable. Furthermore,
we study the power under local alternatives of the parametric test. We show by the Monte
Carlo analysis that, in the small sample case, the introduced test seems to perform better than
other tests proposed in literature.
Keywords: asymptotic distribution free test, local alternatives, maximum-likelihood type estimator, discrete observations, quasi-likelihood function, stochastic differential equation.
1. Introduction
Let (Ω, F, F = (Ft )t≥0 , P ) be a filtered complete probability space. Let us consider a 1dimensional processes X = (Xt )t≥0 solution to the following stochastic differential equation
(1.1)
dXt = b(α, Xt )dt + σ(β, Xt )dWt ,
X0 = x0 ,
where x0 is a deterministic initial value. We assume that b : Θα × R → R, σ : Θβ × R → R are
Borel known functions (up to α and β) and (Wt )t≥0 is a one-dimensional standard Ft -Brownian
motion. Furthermore, α ∈ Θα ⊂ Rm1 , β ∈ Θβ ⊂ Rm2 , m1 , m2 ∈ N, are unknown parameters and
θ = (α, β) ∈ Θ := Θα × Θβ , where Θ represents a compact subset of Rm1 +m2 . We denote by
θ0 := (α0 , β0 ) the true value of θ and assume that θ0 ∈ Int(Θ).
The sample path of X is observed only at n + 1 equidistant discrete times tni , such that
n
ti − tni−1 = ∆n < ∞ for i = 1, ..., n, (with tn0 = 0). Therefore the data, denoted by (Xtni )0≤i≤n ,
are the discrete observations of the sample path of X. Let p be an integer with p ≥ 2. The
asymptotic scheme adopted in this paper is the following: T = n∆n → ∞, ∆n → 0 and n∆pn → 0
as n → ∞. This scheme is called rapidly increasing design, i.e. the number of observations grows
over time but no so fast.
This setting is useful, for instance, in the analysis of financial time series. In mathematical
finance and econometric theory, diffusion processes described by the stochastic differential equations (1.1) play a central role. Indeed, they have been used to model the behavior of stock prices,
exchange rates and interest rates. The underlying stochastic evolution of the financial assets can
be thought continuous in time, although the data are always recorded at discrete instants (e.g.
weekly, daily or each minute). For these reasons, the estimation problems for discretely observed
stochastic differential equations have been tackled by many authors with different approaches
(see, for instance, [10], [33], [11], [5], [23], [24], [2], [12], [18], [3], [6], [29], [34], [30], [26], [31], [20]).
For clustering time series arising from discrete observations of diffusion processes [7] propose a
Date: January 16, 2018.
1
2
A. DE GREGORIO AND S.M. IACUS
new dissimilarity measure based on the L1 distance between the Markov operators. The changepoint problem in the diffusion term of a stochastic differential equation has been considered in
[6] and [16]. In [15], the authors faced the estimation problem for hidden diffusion processes observed at discrete times. An adaptive Lasso-type estimator is proposed in [8]. For the simulation
and the practical implementation of the statistical inference for stochastic differential equations
see [13], [14] and [17].
We also recall that the statistical inference for continuously observed ergodic diffusions is a
well-developed research topic; on this point the reader can consult [25].
The main object of interest of the present paper is the problem of testing parametric hypotheses for diffusion processes from discrete observations. This research topic is less developed
in literature. It is well-known that for testing two simple alternative hypotheses, the NeymanPearson lemma provides a procedure based on the likelihood ratio which leads to the uniformly
most powerful test. In the other cases uniformly most powerful tests do not exist and for this
reason the research of new criteria is justified.
For discretely observed stochastic differential equations, [21] introduced and studied the asymptotic behavior of three kinds of test statistics: likelihood ratio type test statistic, Wald type
test statistic and Rao’s score type test statistic.
Another possible approach is based on the divergences. Indeed, several statistical divergence
measures (which are not necessarily a metric) and distances have been introduced in order to
decide if two probability distributions are close or far. The main goal of this metric is to make
“easy to distinguish” between a pair of distributions which are far from each other than between
those which are closer. These tools have been used for testing hypotheses in parametric models.
The reader can consult on this point, for example, [27] and [28]. For stochastic differential
equations sampled at discrete times, [9] introduced a family of test statistics (for p = 2 and
n∆2n → 0) based on empirical φ-divergences.
We consider the following hypotheses testing problem concerning the vector parameter θ
H0 : θ = θ0 ,
vs H1 : θ 6= θ0 ,
and assume that X is observed at discrete times; that is the data (Xtni )0≤i≤n are available. In
this work we study different test statistics with respect to those used in [9] and [21]. Indeed,
the purpose of this paper is to propose a methodology based on a suitable “distance” between
the approximated transition functions. This idea follows from the observation that in the case
of continuous observations of (1.1), we could define the L2 -distance between the continuous
loglikelihood. Clearly this approach is not useful in our framework and then, similarly to the
aforementioned papers, we consider the local gaussian approximation of the transition density of
the process X from Xti−1 to Xti . In other words, we resort the quasi-likelihood function introduced in [23], defined by means of an approximation with higher order correction terms to relax
the condition of convergence of ∆n to zero. Therefore, let lp,i (θ), θ ∈ Θ, be the approximated
log-transition function from Xti−1 to Xti representing the parametric model (1.1). We deal with
n
Dp,n (θ1 , θ2 ) :=
1X
[lp,i (θ1 ) − lp,i (θ2 )]2 ,
n i=1
θ1 , θ2 ∈ Θ,
which can be interpreted as the empirical L2 -distance between two loglikelihoods. If θ̂p,n is the
maximum quasi-likelihood estimator introduced in [23], we are able to prove that, under H0 , the
test statistic
Tp,n (θ̂p,n , θ0 ) := nDp,n (θ̂p,n , θ0 )
is asymptotically distribution free; i.e. it converges in distribution to a chi squared random
variable. Furthermore, we study the power function of the test under local alternatives.
EMPIRICAL L2 -DISTANCE TEST STATISTICS FOR ERGODIC DIFFUSIONS
3
The paper is organized as follows. Section 2 contains the notations and the assumptions of
the paper. The contrast function arising from the quasi-likelihood approach is briefly discussed
in Section 3. In the same section we define the maximum quasi-likelihood estimator and recall
its main asymptotic properties. In Section 4 we introduce and study a test statistic for the
hypotheses problem H0 : θ = θ0 vs H1 : θ 6= θ0 . The proposed new test statistic shares the
same asymptotic properties of the other test statistics presented in the literature. Therefore,
to justify its use in practice among its competitors, a numerical study is included in Section 5
which contains a comparison of several test statistics in the “small sample” case, i.e., when the
asymptotic conditions are not met. Our numerical analysis shows that, at least for p = 2, the
performance of T2,n is very good. The proofs are collected in Section 6.
It is worth to point out that for the sake of simplicity in this paper a 1-dimensional diffusion is
treated. Nevertheless, it is possible to extend our methodology to the multidimensional stochastic
differential equations setting.
2. Notations and assumptions
Throughout this paper, we will use the following notation.
• θ := (α, β) and α0 , β0 and θ0 denote the true values of α, β and θ respectively.
• c(β, x) = σ 2 (β, x).
• C is a positive constant. If C depends on a fixed quantity, for instance an integer k, we
may write Ck .
2
2
• ∂αh := ∂α∂ h , ∂βk := ∂β∂ k , ∂α2 h αk := ∂αh∂∂αk , h, k = 1, ..., m1 , ∂β2h βk := ∂βh∂∂βk , h, k =
2
•
•
•
•
1, ..., m2 , ∂α2 h βk := ∂α∂h ∂βk , h = 1, ..., m1 , k = 1, ..., m2 , ∂θ := (∂α , ∂β )0 , where ∂α :=
(∂α1 , ..., ∂αm1 )0 and ∂β := (∂β1 , ..., ∂βm2 )0 , ∂θ2 := [∂α2 j βk ]h=1,...,m1 ,k=1,...,m2 .
If f : Θ × R → R, we denote by fi−1 (θ) the value f (θ, Xtni−1 ); for instance c(β, Xtni−1 ) =
ci−1 (β).
For 0 ≤ i ≤ n, tni := i∆n and Gin := σ(Ws , s ≤ tni ).
The random sample is given by Xn := (Xtni )0≤i≤n and Xi := Xtni .
n
The probability law of (1.1) is denoted by Pθ and Eθi−1 [·] := Eθ [·|Gi−1
]. We set P0 := Pθ0
i−1
i−1
and E0 [·] := Eθ0 [·].
P
d
n→∞
n→∞
θ
• −→
and −→ stand for the convergence in probability and in distribution, respectively.
P
θ
• Let Fn : Θ × Rn → R and F : Θ → R; “Fn (θ, Xn ) −→
F (θ) uniformly in θ” stands for
n→∞
P
θ
sup |Fn (θ, Xn ) − F (θ)| −→
0.
n→∞
θ∈Θ
P
θ
Furthermore, if Fn (θ, Xn ) −→
0 uniformly in θ we set
n→∞
Fn (θ, Xn ) = oPθ (1).
• Let un be a R-valued sequence. We indicate by R a function Θ × R2 → R for which there
exists a constant C such that
R(θ, un , x) ≤ un C(1 + |x|)C ,
for all θ ∈ Θ, x ∈ R2 , n ∈ N.
Let us set Ri−1 (∆kn ) := R(θ, ∆kn , Xi−1 ). P
m Pn
• For a m × n matrix A, ||A||2 = tr(AA0 ) = i=1 j=1 |Aij |2 .
Let C↑k,h (R × Θ; R) be the space of all functions f such that:
(i) f (θ, x) is a R-valued function on Θ × R;
4
A. DE GREGORIO AND S.M. IACUS
(ii) f (θ, x) is continuously differentiable with respect to x up to order k ≥ 1 for all θ; these
x-derivatives up to order k are of polynomial growth in x, uniformly in θ;
(iii) f (θ, x) and all x-derivatives up to order k ≥ 1, are h ≥ 1 times continuously differentiable
with respect to θ for all x ∈ R. Moreover, these derivatives up to the h-th order with
respect to θ are of polynomial growth in x, uniformly in θ.
We need some standard assumptions on the regularity of the process X.
A1 . (Existence and Uniqueness) There exists a constant C such that
sup |b(α, x) − b(α, y)| + sup |σ(β, x) − σ(β, y)| ≤ C|x − y|.
α∈Θα
β∈Θβ
A2 . (Ergodicity) The process X is ergodic for θ = θ0 with invariant probability measure
π0 (dx). Thus
Z
Z
1 T
Pθ
f (Xt )dt −→
f (x)π0 (dx),
T →∞
T 0
where f ∈ L1 (π0 ). Furthermore, we assume that π0 admits all moments finite.
A3 . inf x,β σ(β, x) > 0.
A4 . (Moments) For all q ≥ 0 and for all θ ∈ Θ, supt E|Xt |q < ∞.
A5 . [k] (Smoothness) b ∈ C↑k,3 (Θα × R, R) and σ ∈ C↑k,3 (Θβ × R, R).
A6 . (Identifiability) If the coefficients b(α, x) = b(α0 , x) and σ(β, x) = σ(β0 , x) for all x
(π0 -almost surely), then α = α0 and β = β0 .
Let Lθ the infinitesimal generator of X with domain given by C 2 (R) (the space of the twice
continuously differentiable function on R); that is if f ∈ C 2 (R)
c(β, x) ∂ 2 f
∂f
(x) +
(x), L0 := Lθ0 .
∂x
2 ∂x2
Under the assumption A5 [2(j − 1)] we can define Ljθ := Lθ ◦ Lj−1
with domain C 2j (R) and
θ
0
Lθ =Id.
We conclude this section with some well-known examples of ergodic diffusion processes belonging to the class (1.1):
• the Ornstein-Uhlenbeck or Vasicek model is the unique solution to
Lθ f (x) := b(α, x)
(2.1)
dXt = α1 (α2 − Xt )dt + β1 dWt ,
X0 = x0 ,
where b(α1 , α2 , x) = α1 (α2 − x) and σ(β1 , x) = β1 with α1 , α2 ∈ R and β1 > 0. This
stochastic process is a Gaussian process and it is often used in finance where β1 is the
volatility, α2 is the long-run equilibrium of the model and α1 is the speed of mean
reversion. For α1 > 0 the Vasicek process is ergodic with invariant law π0 given by a
β2
Gaussian law with mean α2 and variance 2α11 . It is easy to check that all the conditions
A1 − A6 fulfill;
• the Cox-Ingersoll-Ross (CIR) process is the solution to
(2.2)
p
Xt dWt , X0 = x0 > 0,
√
where b(α1 , α2 , x) = α1 (α2 − x) and σ(β1 , x) = β1 x with α1 , α2 , β1 > 0. If 2α1 α2 > β12
the process is strictly positive, otherwise non negative. This model has a conditional
density given by the non central χ2 distribution. The CIR process is useful in the
description of short-term interest rates and admits invariant law π0 given by a Gamma
β2
distribution with shape parameter 2αβ12α2 and scale parameter 2α11 . If (2.2) is strictly
1
positive, we can prove that the above assumptions hold true.
dXt = α1 (α2 − Xt )dt + β1
EMPIRICAL L2 -DISTANCE TEST STATISTICS FOR ERGODIC DIFFUSIONS
5
3. Preliminaries on the quasi-likelihood function
We briefly recall the quasi-likelihood function introduced by [23] based on the Itô-Taylor
expansion. The main problem in the statistical analysis of the diffusion process X is that its
transition density is in general unknown and then the likelihood function is unknown as well. To
overcome this difficulty one can discretizes the sample path of X by means of Euler-Maruyama’s
scheme; namely
(3.1)
Z
tn
i
Xi − Xi−1 =
Z
tn
i
b(α, Xs )ds +
tn
i−1
tn
i−1
σ(β, Xs )dWs ' bi−1 (α)∆n + σi−1 (β)(Wtni − Wtni−1 ).
Hence (3.1) leads to consider a local-Gaussian approximation to the transition density; that is
L(Xi |Xi−1 ) ' N (bi−1 (α)∆n , ci−1 (β)∆n )
and the approximated loglikelihood function of the random sample Xn , called quasi-loglikelihood
function, becomes
n
1 X (Xi − Xi−1 − bi−1 (α)∆n )2
(3.2)
ln (θ) :=
+ log ci−1 (β) .
2 i=1
ci−1 (β)∆n
This approach suggests to consider the mean and the variance of the transition density of X;
that is
(3.3)
m(θ, Xi−1 ) := Eθ [Xi |Xi−1 ],
m2 (θ, Xi−1 ) := Eθ [(Xi − m(θ, Xi−1 ))2 |Xi−1 ],
and assume
L(Xi |Xi−1 ) ' N (m(θ, Xi−1 ), m2 (θ, Xi−1 )).
Thus we can consider as contrast function the following one
n
1 X (Xi − m(θ, Xi−1 ))2
(3.4)
+ log m2 (θ, Xi−1 ) .
2 i=1
m2 (θ, Xi−1 )
Nevertheless, (3.4) does not have a closed form because m(θ, Xi−1 ) and m2 (θ, Xi−1 ) are unknown.
Therefore we substitute in (3.4) closed approximations of m and m2 based on the Itô-Taylor
expansion.
Let f (y) := y, for l ≥ 0, under the assumption A5 [2l], we have the following approximation
(see Lemma 1, [23])
(3.5)
m(θ, Xi−1 ) = rl (∆n , Xi−1 , θ) + R(θ, ∆l+1
n , Xi−1 )
where
rl (∆n , Xi−1 , θ) :=
l
X
∆i
n
i=0
i!
Liθ f (x).
Now let us consider the function (y − rl (∆n , Xi−1 , θ))2 , which is for fixed x, y and θ a polynomial
in ∆n of degree 2l. We indicate by g ∆n ,x,θ,l (y) the sum of its first terms up to degree l; that is
Pl
g ∆n ,x,θ,l (y) = j=0 ∆jn g jx,θ (y) where
(3.6)
g 0x,θ (y) = (y − x)2
(3.7)
g 1x,θ (y) = −2(y − x)Lθ f (x)
(3.8)
g jx,θ (y) = −2(y − x)
Ljθ f (x)
+
j!
X
r,s≥1,r+s=j
Lrθ f (x) Lsθ f (x)
,
r!
s!
2 ≤ j ≤ l.
6
A. DE GREGORIO AND S.M. IACUS
Under the assumption A5 [2(l − 1)](i), we have that Lrθ g jx,θ (y) is well-defined for r + j = l and we
set
(3.9)
Γl (∆n , x, θ) :=
l
X
∆jn
n
r=0
j=0
where γj (θ, x) are the coefficients of
l−j
X
∆r
∆jn .
r!
Lrθ g jx,θ (x) :=
l
X
∆jn γj (θ, x),
j=0
Therefore by (3.6) to (3.9), we obtain, for instance,
γ0 (θ, x) = L0θ g 0x,θ (x) = 0
γ1 (θ, x) = Lθ g 0x,θ (x) = c(β, x)
L2θ g 0x,θ
(x) + Lθ g 1x,θ (x) + L0θ g 2x,θ (x)
2
1
∂
∂
c(β, x) ∂ 2
c(β, x)
=
b(α, x) c(β, x) + 2c(β, x) b(α, x) +
2
∂y
∂y
4 ∂y 2
γ2 (θ, x) =
Let
Γl (∆n , x, θ) := ∆n c(β, x)[1 + Γl (∆n , x, θ)]
Pl
where Γl (∆n , x, θ) :=
(see Lemma 2, [23])
(3.10)
∆jn γj (θ,x)
.
∆n c(β,x)
j=2
For l ≥ 0, under the assumption A5 [2l](i), we have that
m2 (θ, Xi−1 ) = ∆n ci−1 (β)[1 + Γl (∆n , Xi−1 , θ)] + R(θ, ∆l+1
n , Xi−1 ).
It seems quite natural at this point to substitute (3.5) and (3.10) into the expression (3.4).
Nevertheless, in order to avoid technical difficulties related to the control of denominator and
logarithmic we consider a further expansion in ∆n of (1 + Γl )−1 and log(1 + Γl ).
Let k0 = [p/2]. Under the assumption A5 [2k0 ](i), we define the quasi-loglikelihood function of
Xn as
(3.11)
lp,n (θ) := lp,n (θ, Xn ) :=
n
X
lp,i (θ)
i=1
where
(3.12)
k0
X
(Xi − rk0 (∆n , Xi−1 , θ))2
lp,i (θ) :=
∆jn dj (θ, Xi−1 )
1+
2∆n ci−1 (β)
j=1
k0
X
1
+
∆jn ej (θ, Xi−1 )
log ci−1 (β) +
2
j=1
and dj , resp. ej , is the coefficient of ∆jn in the Taylor expansion of (1 + Γk0 +1 (∆n , x, θ))−1 , resp.
log(1 + Γk0 +1 (∆n , x, θ)). It is not hard to show that, for example,
γ2 (θ, x)
,
c(β, x)
2
1
γ2 (θ, x)
d2 (θ, x) = −e2 (θ, x) =
− γ3 (θ, x) .
c(β, x) c(β, x)
d1 (θ, x) = −e1 (θ, x) = −
Remark 3.1. It is worth to point out that by assumptions A3 and A5 emerge that dj and ej ,
for all j ≤ k0 , are three times differentiable with respect to θ. Furthermore, all their derivatives
with respect to θ are of polynomial growth in x uniformly in θ.
EMPIRICAL L2 -DISTANCE TEST STATISTICS FOR ERGODIC DIFFUSIONS
7
The contrast function (3.11) yields to the maximum quasi-likelihood estimator θ̂p,n := (α̂p,n , β̂p,n )
defined as
(3.13)
lp,n (θ̂p,n ) = inf lp,n (θ).
θ∈Θ
Let I(θ0 ) be the Fisher information matrix at θ0 defined as follows
h,k
[Ib (θ0 )]h,k=1,...,m1
0
(3.14)
I(θ0 ) :=
,
0
[Iσh,k (θ0 )]h,k=1,...,m2
where
∂αh b ∂αk b
:=
(θ0 , x)π0 (dx),
c
Z
1
∂βh c ∂βk c
(β0 , x)π0 (dx).
Iσh,k (θ0 ) :=
2
c2
Ibh,k (θ0 )
Z
We recall an important asymptotic result which will be useful in the proof of our main theorem.
Theorem 1 ([23]). Let p be an integer and k0 = [p/2]. Under assumptions A1 to A4 , A5 [2k0 ]
and A6 , if ∆n → 0, n∆n → ∞, as n → ∞, the estimator θ̂p,n is consistent; i.e.
P
0
θ0 .
θ̂p,n −→
(3.15)
n→∞
If in addition n∆pn → 0 and θ0 ∈ Int(Θ) then
√
n∆ (α̂ − α0 )
d
√ n p,n
(3.16)
ϕ(n)−1/2 (θ̂p,n − θ0 ) =
−→ Nm1 +m2 (0, I −1 (θ0 )),
n→∞
n(β̂p,n − β0 )
where
ϕ(n) :=
1
n∆n Im1
0
0
1
I
n m2
.
Remark 3.2. We observe that l2,n does not coincide with (3.2), because (3.11) contains the
terms d1 and e1 . Nevertheless, ln also yields an asymptotical efficient estimator for θ and then
we refer to it when p = 2.
Remark 3.3. Under the same framework adopted in this paper, alternatively to θ̂p,n , [22] and
[30] proposed different types of adaptive maximum quasi-likelihood estimators. For instance, in
[30], the first type of adaptive estimator is introduced starting from the initial estimator β̃0,n is
defined by Un (β̃0,n ) = inf β∈Θβ Un (β), where
n
1 X (Xi − Xi−1 )2
Un (β) :=
+ log ci−1 (β) .
2 i=1
∆n ci−1 (β)
For p ≥ 2, k0 = [p/2] and l0 = [(p − 1)/2], the first type adaptive estimator θ̃p,n = (α̃k0 ,n , β̃l0 ,n )
is defined for k = 1, 2, ..., k0 , as follows
lp,n (α̃k,n , β̃k−1,n ) = inf lp,n (α, β̃k−1,n ),
α∈Θα
lp,n (α̃k,n , β̃k,n ) = inf lp,n (α̃k,n , β).
β∈Θβ
The maximum quasi-likelihood estimator θ̂p,n and its adaptive versions, like θ̃p,n , are asymptotically equivalent (under a minor change of the initial assumptions); i.e. they have the same
properties (3.15) and (3.16) (see [30]). In what follow we will developed a test based on θ̂p,n ;
nevertheless in light of the previous discussion, it would be possible to replace θ̂p,n with θ̃p,n .
8
A. DE GREGORIO AND S.M. IACUS
4. Test statistics
The goal of this section is to define and to analyze test statistics for the following parametric
hypotheses problem
(4.1)
vs H1 : θ 6= θ0 ,
H0 : θ = θ0 ,
concerning the stochastic differential equation (1.1). X is partially observed and therefore we
have discrete observations represented by Xn . The motivation of this research is due to the fact
that under non-simple alternative hypotheses do not exist uniformly most powerful parametric
tests. Therefore, we need proper procedure for making the right decision concerning statistical
hypothesis.
The first step consists in the introduction of a suitable measure regarding the “discrepancy”,
or the “distance”, between diffusions belonging to the parametric class (1.1). Furthermore, we
bearing in mind that as recalled in the previous section, for a general stochastic differential
equation X, the true probability transitions from Xi−1 to Xi do not exist in closed form as
well as the likelihood function. Suppose known the parameter β and assume observable the
sample path up to time T = n∆n . Let Qβ be the probability law of the process solution to
dYt = σ(β, Yt )dWt . The continuous loglikelihood of X is given by
Z
Z T
b(α, Xt )
1 T b2 (α, Xt )
dPθ
dXt −
dt.
=
log
dQβ
2 0 c(β, Xt )
0 c(β, Xt )
Thus we can consider the (squared) L2 (Qβ )-distance between the loglikelihoods log
log
dPθ2
dQβ
(4.2)
dPθ1
dQβ
and
with θ1 , θ2 ∈ Θ; that is
dPθ1
dPθ2
D(θ1 , θ2 ) := log
− log
dQβ
dQβ
2
L2 (Qβ )
2
Z
dPθ1
dPθ2
=
log
− log
dQβ .
dQβ
dQβ
Clearly for testing the hypotheses (4.1) in the framework of discretely observed stochastic differential equations, the distance (4.2) is not useful. Nevertheless, the above L2 −metric for the
continuos observations suggests to consider
n
1X
(4.3)
Dp,n (θ1 , θ2 ) :=
[lp,i (θ1 ) − lp,i (θ2 )]2 , θ1 , θ2 ∈ Θ,
n i=1
which can be interpreted as the empirical version of (4.2), where the theoretical loglikelihood
is replaced by the quasi-loglikelihood defined by (3.11). The following theorem provides the
convergence in probability of Dp,n .
Theorem 2. Let p be an integer and k0 = [p/2]. Assume A1 − A4 , A5 [2k0 ] and A6 . Under H0 ,
if ∆n → 0, n∆n → ∞, as n → ∞, we have that
P
0
Dp,n (θ, θ0 ) −→
U (β, β0 )
n→∞
uniformly in θ, where
U (β, β0 )
2
2
)
Z (
1
c(β0 , x)
c(β, x)
c(β0 , x)
c(β, x)
:=
3
− 1 + log
+2
− 1 log
π0 (dx).
4
c(β, x)
c(β0 , x)
c(β, x)
c(β0 , x)
The above result shows that Dp,n (θ, θ0 ) is not a true approximation of Dp,n (θ, θ0 ) because
R
2
it does not converge to [log(πθ (dx)/π0 (dx))] π0 (dx). Nevertheless, the function (4.3) allows
to construct the main object of interest of the paper. Let θ̂n be the maximum quasi-likelihood
EMPIRICAL L2 -DISTANCE TEST STATISTICS FOR ERGODIC DIFFUSIONS
9
estimator defined by (3.13), for testing the hypotheses (4.1) we introduce the following class of
test statistics
(4.4)
Tp,n (θ̂p,n , θ0 ) := nDp,n (θ̂p,n , θ0 ).
The first result concerns the weak convergence of Tp,n (θ̂p,n , θ0 ). We prove that Tp,n (θ̂p,n , θ0 ) is
asymptotically distribution free under H0 ; namely it weakly converges to a chi-squared random
variable with two degrees of freedom.
Theorem 3. Let p be an integer and k0 = [p/2]. Assume A1 − A4 , A5 [2k0 ] and A6 . Under H0 ,
if ∆n → 0, n∆n → ∞, n∆pn → 0, as n → ∞, we have that
d
Tp,n (θ̂p,n , θ0 ) −→ χ2m1 +m2 .
(4.5)
n→∞
Given the level α ∈ (0, 1), our criterion suggests to
reject H0 if Tp,n (θ̂p,n , θ0 ) > χ2m1 +m2 ,α ,
where χ2m1 +m2 ,α is the 1 − α quantile of the limiting random variable χ2m1 +m2 ; that is under H0
lim Pθ (Tp,n (θ̂p,n , θ0 ) > χ2m1 +m2 ,α ) = α.
n→∞
Under H1 , the power function of the proposed test are equal to the following map
θ 7→ Pθ Tp,n (θ̂p,n , θ0 ) > χ2m1 +m2 ,α
Often a way to judge the quality of sequences of tests is provided by the powers at alternatives
that become closer and closer to the null hypothesis. This justify the study of local limiting
power. Indeed, usually
the power functions of test statistic (4.4) cannot be calculated explicitly.
Nevertheless, Pθ Tp,n (θ̂p,n , θ0 ) > χ2m1 +m2 ,α can be studied and approximated under contiguous
alternatives written as
(4.6)
H1,n : θ = θ0 + ϕ(n)1/2 h,
where h ∈ Rm1 +m2 such that θ0 +ϕ(n)1/2 h ∈ Θ. In order to get a reasonable approximation of the
power function, we analyze the asymptotic law of the test statistics under the local alternatives
H1,n . We need the following assumption on the contiguity of probability measures (see [32]):
B1 . Pθ0 +ϕ(n)h is a sequence of contiguous probability measures with respect to P0 ; i.e.
limn→∞ P0 (An ) = 0 implies limn→∞ Pθ0 +ϕ(n)1/2 h (An ) = 0 for every measurable sets
An .
Remark 4.1. The assumption B1 holds if we assume A1 − A4 , A5 [2k0 ] and the conditions:
(i) there exists a constant C > 0 such that the following estimates hold
|b(α, x)| ≤ C(1 + |x|),
∂
∂
b(α, x) + |σ(β, x)| +
σ(β, x) ≤ C
∂x
∂x
for all (α, β) ∈ Θ and x ∈ R;
(ii) there exists C0 > 0 and K > 0 such that
b(α, x)x ≤ −C0 |x|2 + K
for all (α, x) ∈ Θα × R;
(iii) there exists a constant C1 > 0 such that
1
≤ σ(β, x) ≤ C1 .
C1
10
A. DE GREGORIO AND S.M. IACUS
Under the above assumptions, [12] proved the Local Asymptotic Normality (LAN) for the likelihood of the ergodic diffusions (1.1); i.e.
dPθ0 +ϕ(n)h
1
d
(Xn ) −→ h0 Nm1 +m2 (0, I(θ0 )) + h0 I(θ0 )h.
log
n→∞
dP0
2
By means of Le Cam’s first lemma (see [32]), LAN property implies the contiguity of Pθ0 +ϕ(n)h
with respect to P0 .
Now, we are able to study the asymptotic probability distribution of Tp,n under H1,n .
Theorem 4. Let p be an integer and k0 = [p/2]. Assume A1 − A4 , A5 [2k0 ], A6 and B1 fulfill.
Under the local alternative hypothesis H1,n , if ∆n → 0, n∆n → ∞, n∆pn → 0 as n → ∞, the
following weak convergence holds
(4.7)
d
Tp,n (θ̂p,n , θ0 ) −→ χ2m1 +m2 (h0 I(θ0 )h),
n→∞
where the random variable χ2l+m (h0 I(θ0 )h) is a non-central chi square random variable with l +m
degrees of freedom and non-centrality parameter h0 I(θ0 )h.
Remark 4.2. If we deal with H0 : θ = θ0 and the local alternative hypothesis H1,n , Theorem 4
leads to the following approximation of the power functions
(4.8)
Pθ Tp,n (θ̂p,n , θ0 ) > χ2m1 +m2 ,α ∼
= 1 − F χ2m1 +m2 ,α , n >> 1,
where F(·) is the cumulative function of the random variable χ2m1 +m2 (h0 I(θ0 )h).
Remark 4.3. The Generalized Quasi-Likelihood Ratio, Wald, Rao type test statistics have been
studied by [21], respectively, given by
(4.9)
Lp,n (θ̂p,n , θ0 ) := 2(lp,n (θ̂p,n ) − lp,n (θ0 ))
(4.10)
Wp,n (θ̂p,n , θ0 ) := (ϕ(n)−1/2 (θ̂p,n − θ0 ))0 Ip,n (θ̂p,n )ϕ(n)−1/2 (θ̂p,n − θ0 )
(4.11)
−1
Rp,n (θ̂p,n , θ0 ) := (ϕ(n)1/2 ∂θ lp,n (θ0 ))0 Ip,n
(θ̂p,n )ϕ(n)1/2 ∂θ lp,n (θ0 ),
where
Ip,n (θ) =
1
2
n∆n ∂α lp,n (θ)
1
√
∂ ∂ l (θ)
n ∆n β α p,n
√1
∂ ∂ l (θ)
n ∆n α β p,n
1 2
∂
n β lp,n (θ)
!
and Rp,n is well-defined if Ip,n (θ) is nonsingular. The above test statistics are asymptotically
equivalent to Tp,n ; i.e. under H0 , Lp,n , Wp,n and Rp,n weakly converge to a χ2 random variable.
Remark 4.4. In [9], the authors dealt with (for p = 2) test statistics based on an empirical
version of the true φ-divergences; i.e.
n
X
exp ln (θ)
(4.12)
2
φ
exp ln (θ0 )
i=1
where φ represents a suitable convex function and ln is given by (3.2). In the present paper,
the starting point is represented by the L2 -distance between two diffusion parametric models.
Somehow, the approach developed in this work is close to that developed by [1], where a test
based on the L2 -distance measure between the density function and its nonparametric estimator
is introduced.
EMPIRICAL L2 -DISTANCE TEST STATISTICS FOR ERGODIC DIFFUSIONS
11
Remark 4.5. From a practical point of view, since sometimes α = α0 and β = β0 have different
meanings, it is possible to deal with a stepwise procedure. For instance as p = 2, first, we test
β = β0 by means of
!
!#2
"
n
X
1
1
ci−1 (β̃0,n )
(Xi − Xi−1 )2
β
−
Tn (β̃0,n , β0 ) :=
+ log
∆n
ci−1 (β0 )
ci−1 (β̃0,n ) ci−1 (β0 )
i=1
and then, in the second step, we test α = α0 by taking into account
n
X
Tnα (α̃1,n , α0 , β̃0,n ) :=
[l2,i (α̃1,n , β̃0,n ) − l2,i (α0 , β̃0,n )]2 ,
i=1
where α̃1,n and β̃0,n are the adaptive estimators defined in the Remark 3.3.
5. Numerical analysis
Although all test statistics presented in the above and in the literature satisfy the same
asymptotic results, for small sample sizes the performance of each test statistic is determined
by the statistical model generating the data and the quality of the approximation of the quasilikelihood function. To put in evidence these effects we consider the two stochastic models
presented in Section 2, namely the Ornstein-Uhlenbeck (OU in the tables) of equation (2.1) and
the CIR model of equation (2.2). In this numerical study we consider the power of the test under
local alternatives for different test statistics:
• the φ divergence of equation (4.12) with φ(x) = 1 − x + x log(x), which is equivalent to
the approximated Kullback-Leibler divergence (see, [9]). We use the label AKL in the
tables for this approximate KL;
2
: this was proposed in [4], we name it BS in the
• the φ divergence with φ(x) = x−1
x+1
tables;
• the Generalized Quasi-Likelihood Ratio test with p = 2, see e.g., (4.9), denoted as
GQLRT in the tables;
• the Rao test statistics1 R(θ̂p,n , θ0 ) of equation (4.11), denoted as RAO in the tables;
• and the statistic Tp,n (θ̂p,n , θ0 ) proposed in this paper and defined in equation (4.4), with
p = 2, denoted as T2,n in the tables.
The sample sizes have been chosen to be equal to n = 50, 100, 250, 500, 1000 observations and
1
time horizon is set to T = n 3 , in order to satisfy the asymptotic theory. For testing θ0 against
h
the local alternatives θ0 + √n∆
for the parameters in the drift coefficient and θ0 + √hn for the
n
parameters in the diffusion coefficient, h is taken in a grid from 0 to 1, and h = 0 corresponds
to the null hypothesis H0 . For the data generating process, we consider the following statistical
models
OU: the one-dimensional Ornstein-Uhlenbeck model solution to dXt = α1 (α2 −Xt )dt+β1 dWt ,
X0 = 1, with θ0 = (α1 , α2 , β1 ) = (0.5, 0.5, 0.25);
√
CIR: the one-dimensional CIR model solution to dXt = α1 (α2 − Xt )dt + β1 Xt dWt , X0 = 1,
with θ0 = (α1 , α2 , β1 ) = (0.5, 0.5, 0.125).
In each experiments the process have been simulated at high frequency using the EulerMaruyama scheme and resampled to obtain n = 50, 100, 250, 500, 1000 observations. Remark
that, even if the Ornstein-Uhlenbeck process has a Gaussian transition density, this density is
different from the Euler-Maruyama Gaussian density for non negligible time mesh ∆n (see, [13]).
1We do not consider the Wald test of (4.10) because it was shown in [21] that it performs similarly to the Rao
test statistics.
12
A. DE GREGORIO AND S.M. IACUS
For the simulation we user the R package yuima (see, [17]). Each experiment is replicated 1000
times and from the empirical distribution of each test statistic, say Sn , we define the rejection
threshold of the test as χ̃23,0.05 , i.e. χ̃23,0.05 is the 95% quantile of the empirical distribution of
Sn , that is
0.05 = Freq(Sn (θ̂n , θ0 ) > χ̃23,0.05 ).
Similarly, we define the empirical power function of the test as
EPow(h) = Freq(Sn (θ̂n , θ0 + ϕ(n)1/2 h) > χ̃23,0.05 ),
where θ̂n is the maximum quasi-likelihood estimator defined in (3.13). The choice of using the
empirical threshold χ̃23,0.05 instead of the theoretical threshold χ23,0.05 from the χ23 distribution,
is due to the fact that otherwise the tests are non comparable. Indeed, the empirical level of
the test is not 0.05 for small sample sizes when χ23,0.05 is used as rejection threshold and, for
example, when h = 0 different choices of the test statistic produce different empirical levels of the
test. Tables 1 and 2 contain the empirical power function of each test. In these tables the bold
face font is used to put in evidence the test statistics with the highest empirical power function
EPow(h) for a given local alternative h > 0. As mentioned before, the natural benchmark test
statistics is the generalised quasi likelihood ratio test (GQLRT).
From this numerical analysis we can see several facts:
• the test statistic based on the AKL test statistics does not perform as the GQLR test
despite they are related to the same divergence; the latter being sometimes better;
• the T2,n seems to be (almost) uniformly more powerful in this experiment;
• all but RAO test seem to have a good behaviour when the alternative is sufficiently large;
• for the CIR model, the RAO test does not perform well under the alternative hypothesis
and this is probably because it requires very large T which, in our case, is at most T = 10.
For the OU Gaussian case, the performance are better and in line from those presented
in [21] for similar sample sizes.
Therefore, we can conclude that, despite all the test statistics share the same asymptotic properties, the proposed Tp,n seems to perform very well in the small sample case examined in the
above Monte Carlo experiments, at least for p = 2.
6. Proofs
In order to prove the theorems appearing in the paper, we need some preliminary results. Let
us start with the following lemmas.
Lemma 1. For k ≥ 1 and tni−1 ≤ t ≤ tni
E0i−1 [|Xt − Xi−1 |k ] ≤ Ck |t − tni−1 |k/2 (1 + |Xi−1 |)Ck .
(6.1)
If f : Θ × R → R is of polynomial growth in x uniformly in θ then
(6.2)
E0i−1 [f (θ, Xt )] ≤ Ct−tni−1 (1 + |Xi−1 |)C ,
tni−1 ≤ t ≤ tni .
Proof. See the proof of Lemma 6 in [23].
Lemma 2. For l ≥ 1
(6.3)
rl (∆n , Xi−1 , θ) = Xi−1 + ∆n bi−1 (α) + R(θ, ∆2n , Xi−1 )
(6.4)
E0i−1 [(Xi − rl (∆n , Xi−1 , θ))2 ] = ∆n ci−1 (β0 ) + R(θ, ∆2n , Xi−1 )
(6.5)
E0i−1 [(Xi − rl (∆n , Xi−1 , θ))3 ] = R(θ, ∆2n , Xi−1 )
(6.6)
E0i−1 [(Xi − rl (∆n , Xi−1 , θ))4 ] = 3∆2n c2i−1 (β0 ) + R(θ, ∆3n , Xi−1 )
(6.7)
E0i−1 [(Xi − rl (∆n , Xi−1 , θ))5 ] = R(θ, ∆3n , Xi−1 )
EMPIRICAL L2 -DISTANCE TEST STATISTICS FOR ERGODIC DIFFUSIONS
13
Table 1. Empirical power function EPow(h), for different sample sizes n and
local alternatives h. The empirical power and theoretical power is 0.05. Data
generating model: the 1-dimensional Ornstein-Uhlenbeck process.
h=0.00
h=0.01
h=0.05
h=0.10
h=0.20
h=0.30
h=0.40
h=0.50
h=0.60
h=0.70
h=0.80
h=0.90
h=1.00
AKL
0.050
0.044
0.035
0.025
0.011
0.007
0.007
0.009
0.021
0.053
0.120
0.221
0.383
n = 50
GQLRT
BS
0.050
0.050
0.048
0.046
0.032
0.041
0.029
0.033
0.031
0.042
0.054
0.069
0.108
0.147
0.216
0.269
0.359
0.448
0.527
0.591
0.670
0.736
0.794
0.852
0.882
0.910
RAO
0.050
0.053
0.057
0.064
0.078
0.096
0.121
0.138
0.146
0.149
0.150
0.148
0.145
T2,n
0.050
0.052
0.057
0.077
0.133
0.239
0.371
0.559
0.720
0.842
0.917
0.966
0.992
h=0.00
h=0.01
h=0.05
h=0.10
h=0.20
h=0.30
h=0.40
h=0.50
h=0.60
h=0.70
h=0.80
h=0.90
h=1.00
AKL
0.050
0.044
0.036
0.028
0.015
0.022
0.049
0.118
0.253
0.436
0.666
0.821
0.911
n = 250
GQLRT
BS
0.050
0.050
0.049
0.050
0.049
0.046
0.048
0.050
0.076
0.078
0.153
0.157
0.304
0.304
0.486
0.496
0.703
0.704
0.847
0.851
0.928
0.931
0.973
0.976
0.992
0.993
RAO
0.050
0.051
0.052
0.058
0.114
0.168
0.222
0.280
0.339
0.389
0.419
0.462
0.485
T2,n
0.050
0.048
0.057
0.075
0.143
0.255
0.452
0.654
0.822
0.921
0.969
0.991
1.000
h=0.00
h=0.01
h=0.05
h=0.10
h=0.20
h=0.30
h=0.40
h=0.50
h=0.60
h=0.70
h=0.80
h=0.90
h=1.00
AKL
0.050
0.046
0.038
0.035
0.061
0.122
0.219
0.426
0.655
0.821
0.930
0.978
0.994
h=0.00
h=0.01
h=0.05
h=0.10
h=0.20
h=0.30
h=0.40
h=0.50
h=0.60
h=0.70
h=0.80
h=0.90
h=1.00
AKL
0.050
0.046
0.032
0.022
0.014
0.009
0.009
0.020
0.051
0.131
0.244
0.414
0.608
n = 100
GQLRT
BS
0.050
0.050
0.047
0.046
0.035
0.035
0.029
0.030
0.038
0.042
0.089
0.083
0.159
0.163
0.283
0.291
0.465
0.472
0.644
0.659
0.789
0.801
0.883
0.893
0.937
0.944
RAO
0.050
0.050
0.050
0.058
0.082
0.101
0.128
0.155
0.183
0.199
0.213
0.221
0.225
T2,n
0.050
0.050
0.055
0.070
0.141
0.253
0.404
0.609
0.769
0.876
0.943
0.984
0.996
h=0.00
h=0.01
h=0.05
h=0.10
h=0.20
h=0.30
h=0.40
h=0.50
h=0.60
h=0.70
h=0.80
h=0.90
h=1.00
AKL
0.050
0.048
0.038
0.032
0.030
0.050
0.138
0.274
0.493
0.704
0.847
0.936
0.982
n = 500
GQLRT
BS
0.050
0.050
0.049
0.049
0.044
0.043
0.050
0.050
0.084
0.080
0.175
0.175
0.329
0.323
0.555
0.552
0.751
0.747
0.869
0.869
0.957
0.957
0.987
0.987
0.997
0.997
RAO
0.050
0.052
0.067
0.082
0.134
0.202
0.279
0.363
0.454
0.522
0.584
0.630
0.678
T2,n
0.050
0.051
0.059
0.075
0.133
0.250
0.449
0.673
0.828
0.934
0.983
0.996
0.998
n = 1000
GQLRT
BS
0.050
0.050
0.049
0.050
0.046
0.049
0.056
0.062
0.104
0.109
0.182
0.187
0.359
0.372
0.600
0.605
0.786
0.794
0.912
0.914
0.969
0.972
0.993
0.993
0.997
0.997
RAO
0.050
0.051
0.056
0.062
0.121
0.193
0.291
0.398
0.507
0.596
0.665
0.711
0.760
Tn
0.050
0.051
0.058
0.074
0.134
0.241
0.442
0.662
0.840
0.942
0.985
0.994
0.998
(6.8)
E0i−1 [(Xi − rl (∆n , Xi−1 , θ))6 ] = 5 · 3∆3n c3i−1 (β0 ) + R(θ, ∆4n , Xi−1 )
(6.9)
E0i−1 [(Xi − rl (∆n , Xi−1 , θ))7 ] = R(θ, ∆4n , Xi−1 )
(6.10)
E0i−1 [(Xi − rl (∆n , Xi−1 , θ))8 ] = 7 · 5 · 3∆4n c4i−1 (β0 ) + R(θ, ∆5n , Xi−1 )
Proof. The equalities from (6.3) to (6.6) represent the statement of Lemma 7 in [23]. By using
the same approach adopted for the proof of the aforementioned lemma, we observe that from
(6.3) to (6.6), the result (6.7) and (6.8) hold, if we are able to show that
(6.11)
E0i−1 [(Xi − Xi−1 )5 ] = R(θ, ∆3n , Xi−1 )
(6.12)
E0i−1 [(Xi − Xi−1 )6 ] = 5 · 3∆3n c3i−1 (β0 ) + R(θ, ∆4n , Xi−1 )
We only prove (6.12), because (6.11) follows by means of similar arguments. By applying the
Ito-Taylor formula (see Lemma 1, in [10]) to the function fx (y) = (y − x)6 we obtain
E0i−1 [(Xi − Xi−1 )6 ] = fXi−1 (Xi−1 ) + ∆n L0 fXi−1 (Xi−1 )
+
∆3
∆2n 2
L0 fXi−1 (Xi−1 ) + n L30 fXi−1 (Xi−1 )
2
3!
14
A. DE GREGORIO AND S.M. IACUS
Table 2. Empirical power function EPow(h), for different sample sizes n and
local alternatives h. The empirical power and theoretical power is 0.05. Data
generating model: the 1-dimensional CIR process.
h=0.00
h=0.01
h=0.05
h=0.10
h=0.20
h=0.30
h=0.40
h=0.50
h=0.60
h=0.70
h=0.80
h=0.90
h=1.00
AKL
0.050
0.041
0.025
0.009
0.013
0.044
0.186
0.475
0.760
0.913
0.981
0.997
1.000
n = 50
GQLRT
BS
0.050
0.050
0.044
0.045
0.032
0.031
0.040
0.042
0.148
0.167
0.416
0.458
0.700
0.741
0.883
0.907
0.967
0.981
0.994
0.998
1.000
1.000
1.000
1.000
1.000
1.000
RAO
0.050
0.052
0.059
0.068
0.075
0.069
0.067
0.067
0.061
0.059
0.051
0.041
0.041
T2,n
0.050
0.053
0.071
0.145
0.371
0.721
0.923
0.989
0.997
1.000
1.000
1.000
1.000
h=0.00
h=0.01
h=0.05
h=0.10
h=0.20
h=0.30
h=0.40
h=0.50
h=0.60
h=0.70
h=0.80
h=0.90
h=1.00
AKL
0.050
0.042
0.026
0.021
0.093
0.372
0.790
0.952
0.996
1.000
1.000
1.000
1.000
n = 250
GQLRT
BS
0.050
0.050
0.049
0.046
0.045
0.046
0.086
0.084
0.347
0.342
0.752
0.756
0.943
0.944
0.994
0.994
1.000
1.000
1.000
1.000
1.000
1.000
1.000
1.000
1.000
1.000
RAO
0.050
0.052
0.054
0.057
0.062
0.064
0.065
0.064
0.060
0.060
0.057
0.055
0.050
T2,n
0.050
0.050
0.071
0.144
0.505
0.864
0.977
1.000
1.000
1.000
1.000
1.000
1.000
h=0.00
h=0.01
h=0.05
h=0.10
h=0.20
h=0.30
h=0.40
h=0.50
h=0.60
h=0.70
h=0.80
h=0.90
h=1.00
Z
∆n
u1
Z
u2
Z
+
0
0
n = 100
GQLRT
BS
0.050
0.050
0.043
0.046
0.032
0.034
0.054
0.051
0.205
0.207
0.537
0.553
0.827
0.836
0.950
0.958
0.993
0.994
0.999
0.998
1.000
1.000
1.000
1.000
1.000
1.000
RAO
0.050
0.053
0.056
0.062
0.063
0.064
0.064
0.063
0.059
0.056
0.053
0.048
0.044
T2,n
0.050
0.051
0.070
0.150
0.461
0.797
0.957
0.995
1.000
1.000
1.000
1.000
1.000
h=0.00
h=0.01
h=0.05
h=0.10
h=0.20
h=0.30
h=0.40
h=0.50
h=0.60
h=0.70
h=0.80
h=0.90
h=1.00
AKL
0.050
0.043
0.030
0.032
0.180
0.598
0.898
0.992
0.998
0.999
1.000
1.000
1.000
n = 500
GQLRT
BS
0.050
0.050
0.043
0.042
0.046
0.044
0.095
0.091
0.384
0.380
0.802
0.800
0.972
0.972
0.998
0.998
0.999
0.999
1.000
1.000
1.000
1.000
1.000
1.000
1.000
1.000
RAO
0.050
0.051
0.051
0.052
0.055
0.058
0.058
0.059
0.057
0.056
0.055
0.055
0.051
T2,n
0.050
0.048
0.074
0.147
0.530
0.869
0.990
0.998
0.999
1.000
1.000
1.000
1.000
n = 1000
GQLRT
BS
0.050
0.050
0.048
0.047
0.059
0.057
0.120
0.118
0.429
0.428
0.844
0.840
0.983
0.983
0.998
0.998
1.000
1.000
1.000
1.000
1.000
1.000
1.000
1.000
1.000
1.000
AKL
0.050
0.044
0.035
0.067
0.274
0.725
0.953
0.996
1.000
1.000
1.000
1.000
1.000
Z
h=0.00
h=0.01
h=0.05
h=0.10
h=0.20
h=0.30
h=0.40
h=0.50
h=0.60
h=0.70
h=0.80
h=0.90
h=1.00
AKL
0.050
0.040
0.019
0.010
0.017
0.102
0.338
0.685
0.896
0.977
0.998
0.999
1.000
0
0
RAO
0.050
0.051
0.051
0.054
0.058
0.061
0.062
0.062
0.062
0.060
0.059
0.059
0.058
T2,n
0.050
0.050
0.079
0.144
0.527
0.886
0.989
0.999
1.000
1.000
1.000
1.000
1.000
u3
E0i−1 [L40 fXi−1 (Xtni−1 +u4 )]du1 du2 du3 du4 .
By applying (6.2), we obtain
Z
0
∆n
Z
0
u1
Z
0
u2
Z
0
u3
E0i−1 [L40 fXi−1 (Xtni−1 +u4 )]du1 du2 du3 du4 = R(θ, ∆4n , Xi−1 ).
Furthermore, by means of long and cumbersome calculations, we can show that fx (x) = L0 fx (x) =
L20 fx (x) = 0, while L30 fx (x) = 5 · 3 · 3!∆3n c3i−1 (β0 ).
Analogously to what done, from (6.3) to (6.8), the equalities (6.9) and (6.10) hold, if we are
able to show that
(6.13)
E0i−1 [(Xi − Xi−1 )7 ] = R(θ, ∆4n , Xi−1 ),
(6.14)
E0i−1 [(Xi − Xi−1 )8 ] = 7 · 5 · 3∆4n c4i−1 (β0 ) + R(θ, ∆5n , Xi−1 ).
EMPIRICAL L2 -DISTANCE TEST STATISTICS FOR ERGODIC DIFFUSIONS
15
We only prove (6.14), because (6.13) follows by means of similar arguments. The application of
the Ito-Taylor formula to the function fx (y) = (y − x)8 yields
E0i−1 [(Xi − Xi−1 )8 ] = fXi−1 (Xi−1 ) + ∆n L0 fXi−1 (Xi−1 ) +
∆2n 2
L fX (Xi−1 )
2 0 i−1
∆3n 3
∆4
L0 fXi−1 (Xi−1 ) + n L40 fXi−1 (Xi−1 )
3!
4!
Z ∆n Z u1 Z u2 Z u3 Z u4
E0i−1 [L50 fXi−1 (Xtni−1 +u5 )]du1 du2 du3 du4 du5
+
+
0
By applying (6.2), we get
Z ∆n Z u1 Z u2 Z u3 Z
0
0
0
0
0
0
u4
0
0
0
E0i−1 [L50 fXi−1 (Xtni−1 +u5 )]du1 du2 du3 du4 du5 = R(θ, ∆5n , Xi−1 ).
Furthermore, by means of long and cumbersome calculations, we can show that fx (x) = L0 fx (x) =
L20 fx (x) = L30 fx (x) = 0 while L40 fx (x) = 7 · 5 · 3 · 4!∆4n c4 (β0 , x).
Lemma 3 (Triangular arrays convegence). Let Uin and U be random variables, with Uin being
Pn
P
Gin -measurable. The two following conditions imply i=1 Uin −→ U :
n→∞
n
X
i=1
P
n
E[Uin |Gi−1
] −→ U,
n→∞
Proof. See the proof of Lemma 9 in [11].
n
X
i=1
P
n
E[(Uin )2 |Gi−1
] −→ 0
n→∞
Lemma 4. Let f : Θ × R → R be such that f (θ, x) ∈ C↑1,1 (Θ × R, R). Let us assume A1 − A6 , if
∆n → 0 and n∆n → ∞ we have that
Z
n
1X
P0
f (x, θ)π0 (dx)
fi−1 (θ) −→
n→∞
n i=1
uniformly in θ.
Proof. See the proof of Lemma 8 in [23].
Lemma 5. Let f : Θ × R → R be such that f (θ, x) ∈ C↑1,1 (Θ × R, R). Let us assume A1 − A6 , if
∆n → 0 and n∆n → ∞, as n → ∞, we have that
0,
j = 1, k = 1,
R
f
(θ,
x)c(β
,
x)π
(dx),
j
= 1, k = 2,
0
0
n
R
1 X
k P0
fi−1 (θ)(Xi − rl (∆n , Xi−1 , θ0 )) −→
f (θ, x)R(θ, 1, x)π0 (dx), j = 2, k = 3,
n→∞
n∆jn i=1
0,
j = 1, k = 4,
3 R f (θ, x)c2 (β , x)π (dx), j = 2, k = 4,
0
0
uniformly in θ.
Proof. The cases j = 1, k = 1 and j = 1, k = 2 coincide with Lemma 9 and Lemma 10 in [23]
and then we use the same approach to show that remaining convergences hold true.
By setting
1
fi−1 (θ)(Xi − rl (∆n , Xi−1 , θ0 ))3 ,
ζin (θ) :=
n∆2n
16
A. DE GREGORIO AND S.M. IACUS
we prove that the convergence holds for all θ. By taking into account Lemma 2
Z
n
1X
P0
i−1 n
f (θ, x)R(θ, 1, x)π0 (dx),
E0 [ζi (θ)] =
fi−1 (θ)R(θ, 1, Xi−1 ) −→
n→∞
n i=1
E0i−1 [(ζin (θ))2 ] =
n
1 X
P0
[5 · 3c3i−1 (β0 ) + R(θ, 1, Xi−1 )] −→
0.
2
n→∞
n ∆n i=1
Therefore by Lemma 3 we can conclude that
Z
P0
f (θ, x)R(θ, 1, x)π0 (dx),
ζin (θ) −→
n→∞
for all θ. For the uniformity of the convergence we use the same arguments adopted in the proof
of Lemma 8 in [23]. Hence, it is sufficient to prove the tightness of the sequence of random
elements
n
1 X fi−1 (θ)(Xi − rl (∆n , Xi−1 , θ0 ))3
Yn (θ) :=
n i=1
∆2n
taking values in the Banach space C(Θ) endowed with the sup-norm ||·||∞ . From the assumptions
of lemma follows that supn E0 [supθ∈Θ |∂θ Yn (θ)|] < ∞ which implies the tightness of Yn (θ) for
the criterion given by Theorem 16.5 in [19].
By setting
1
fi−1 (θ)(Xi − rl (∆n , Xi−1 , θ0 ))4 ,
ζin (θ) :=
n∆2n
we prove that the convergence holds for all θ. By taking into account Lemma 2 and Lemma 5
Z
n
1X
P0
i−1 n
2
fi−1 (θ)[3ci−1 (β0 ) + R(θ, ∆n , Xi−1 )] −→ 3 f (θ, x)c2 (β0 , x)π0 (dx),
E0 [ζi (θ)] =
n→∞
n i=1
E0i−1 [(ζin (θ))2 ] =
n
1 X
P0
[7 · 5 · 3c4i−1 (β0 ) + R(θ, ∆n , Xi−1 )] −→
0.
n→∞
n2 i=1
Therefore by Lemma 3 we get the pointwise convergence. For the uniformity of the convergence
we proceed as done above.
Before to proceed with the proofs of the main theorems of the paper, we introduce some useful
quantities coinciding with (4.2)−(4.8) appearing in [23]. We can write down
(6.15)
lp,i (θ) − lp,i (θ0 ) = ϕi,1 (θ, θ0 ) + ϕi,2 (θ, θ0 ) + ϕi,3 (θ, θ0 ) + ϕi,4 (θ, θ0 ),
where
(Xi − rk0 (∆n , Xi−1 , θ0 ))2
ϕi,1 (θ, θ0 ) :=
2∆n
(
1+
Pk0
j=1
∆jn dj (θ, Xi−1 )
ci−1 (β)
−
1+
Pk0
j=1
∆jn dj (θ0 , Xi−1 )
ci−1 (β0 )
(Xi − rk0 (∆n , Xi−1 , θ0 ))(rk0 (∆n , Xi−1 , θ0 ) − rk0 (∆n , Xi−1 , θ))
∆n ci−1 (β)
k
0
X
× 1+
∆jn dj (θ, Xi−1 ) ,
j=1
k0
2
X
(rk0 (∆n , Xi−1 , θ0 ) − rk0 (∆n , Xi−1 , θ))
1+
ϕi,3 (θ, θ0 ) :=
∆jn dj (θ, Xi−1 ) ,
2∆n ci−1 (β)
ϕi,2 (θ, θ0 ) :=
j=1
)
,
EMPIRICAL L2 -DISTANCE TEST STATISTICS FOR ERGODIC DIFFUSIONS
ϕi,4 (θ, θ0 ) :=
1
log
2
ci−1 (β)
ci−1 (β0 )
17
k
+
0
1X
∆j (ej (θ, Xi−1 ) − ej (θ0 , Xi−1 )).
2 j=1 n
Furthermore
h
h
∂αh lp,i (θ) = ηi,1
(θ) + ηi,2
(θ),
(6.16)
h = 1, 2, ..., m1 ,
where
h
ηi,1
(θ) := −(∂αh rk0 (∆n , Xi−1 , θ))(Xi − rk0 (∆n , Xi−1 , θ))
Pk0
h
ηi,2
(θ)
2
:= (Xi − rk0 (∆n , Xi−1 , θ))
j=1
o
n
Pk0
∆jn dj (θ, Xi−1 )
1 + j=1
∆n ci−1 (β)
∆jn ∂αh dj (θ, Xi−1 )
2∆n ci−1 (β)
,
k
+
0
1X
∆j ∂α ej (θ, Xi−1 ),
2 j=1 n h
and
(6.17)
k
k
k
∂βk lp,i (θ) = ξi,1
(θ) + ξi,2
(θ) + ξi,3
(θ),
k = 1, 2, ..., m2 ,
where
k0
k0
1X
2 X
(X
−
r
(∆
,
X
,
θ))
i
k0
n
i−1
k
∆jn ∂βk ej (θ, Xi−1 ),
∆jn ∂βk dj (θ, Xi−1 ) +
ξi,1
(θ) :=
2
2∆n ci−1 (β)
j=1
j=1
k0
∂ c (β)
2
X
(X
−
r
(∆
,
X
,
θ))
∂
c
(β)
β i−1
i
k0
n
i−1
βk i−1
j
k
∆
d
(θ,
X
)
+ k
,
ξi,2
(θ) := −
1
+
j
i−1
n
2∆n c2i−1 (β)
2c
i−1 (β)
j=1
n
o
Pk0
1 + j=1
∆jn dj (θ, Xi−1 )
k
ξi,3
(θ) := −(∂βk rk0 (∆n , Xi−1 , θ))(Xi − rk0 (∆n , Xi−1 , θ))
.
∆n ci−1 (β)
From (6.15) it is possible to derive
(6.18)
h,k
h,k
h,k
h,k
∂α2 h αk lp,i (θ) := δi,1
(θ) + δi,2
(θ) + δi,3
(θ) + δi,4
(θ),
h, k = 1, 2, ..., m1 ,
where
(Xi − rk0 (∆n , Xi−1 , θ0 ))2
{(∂α2 h αk d1 )i−1 (θ) + R(θ, ∆n , Xi−1 )},
2ci−1 (β)
(Xi − rk0 (∆n , Xi−1 , θ0 ))
h,k
δi,2
{−∂α2 h αk bi−1 (α) + R(θ, ∆n , Xi−1 )},
(θ) :=
ci−1 (β)
1
h,k
δi,3
(θ) := ∆n ∂α2 h αk e1 (θ, Xi−1 ),
2
2
∂αh αk bi−1 (α)(bi−1 (α) − bi−1 (α0 )) + ∂αh bi−1 (α)∂αk bi−1 (α)
h,k
+ R(θ, ∆n , Xi−1 ) ,
δi,4
(θ) := ∆n
ci−1 (β)
h,k
δi,1
(θ) :=
(6.19)
h,k
h,k
h,k
∂β2h βk lp,i (θ) := νi,1
(θ) + νi,2
(θ) + νi,3
(θ),
h, k = 1, 2, ..., m2 ,
where
(Xi − rk0 (∆n , Xi−1 , θ0 ))2
{(∂β2h βk c−1 )i−1 (β) + R(θ, ∆n , Xi−1 )},
2∆n
1
h,k
νi,2
(θ) := (Xi − rk0 (∆n , Xi−1 , θ0 ))R(θ, 1, Xi−1 )),
2
1
h,k
νi,3 (θ) := (∂β2h βk log c)i−1 (β) + R(θ, ∆n , Xi−1 )),
2
h,k
νi,1
(θ) :=
18
A. DE GREGORIO AND S.M. IACUS
and
(6.20)
∂α2 h βk lp,i (θ) := µi,1 (θ) + µi,2 (θ),
h = 1, 2, ..., m1 , k = 1, 2, ..., m2 ,
where
(Xi − rk0 (∆n , Xi−1 , θ0 ))2
R(θ, ∆n , Xi−1 ),
2∆n
(Xi − rk0 (∆n , Xi−1 , θ0 ))
µi,2 (θ) :=
R(θ, ∆n , Xi−1 ) + R(θ, ∆n , Xi−1 ).
∆n
µi,1 (θ) :=
Proof of Theorem 2. We observe that
n
4
X
1 X X
Dp,n (θ, θ0 ) =
(ϕi,k (θ, θ0 ))2 + 2
)ϕi,j (θ, θ0 )ϕi,k (θ, θ0 ) .
n
i=1
k=1
j<k
Under H0 , from Lemma 4 and Lemma 5, we derive
"
2 #
n
n
4
X
1
1
(X
−
r
(∆
,
X
,
θ
))
1
1X
i
k
n
i−1
0
0
(ϕi,1 (θ, θ0 ))2 =
−
+ R(θ, ∆n , Xi−1 )
n i=1
n i=1
4∆2n
ci−1 (β) ci−1 (β0 )
"
2 #
n
1 X (Xi − rk0 (∆n , Xi−1 , θ0 ))4
1
1
=
−
+ oP0 (1)
n i=1
4∆2n
ci−1 (β) ci−1 (β0 )
2
Z
1
1
P0 3
2
c (β0 , x)
−
π0 (dx)
−→
n→∞ 4
c(β, x) c(β0 , x)
n
n
1X
1 X (Xi − rk0 (∆n , Xi−1 , θ0 ))2
2
(ϕi,2 (θ, θ0 ))2 =
[b
(α
)
−
b
(α)]
+ oP0 (1)
i−1
0
i−1
n i=1
n i=1
c2i−1 (β0 )
P
0
−→
0
n→∞
n
n
1 X ∆2n [bi−1 (α0 ) − bi−1 (α)]4
1X
2
(ϕi,3 (θ, θ0 )) =
+ oP0 (1)
n i=1
n i=1
4c2i−1 (β)
P
0
−→
0
n→∞
2
n
n
1X
1X1
ci−1 (β)
2
(ϕi,4 (θ, θ0 )) =
log
+ oP0 (1)
n i=1
n i=1 4
ci−1 (β0 )
2
Z
c(β, x)
P0 1
−→
log
π0 (dx)
n→∞ 4
c(β0 , x)
n
n
1 X (Xi − rk0 (∆n , Xi−1 , θ0 ))2
1X
1
1
ϕi,1 (θ, θ0 )ϕi,4 (θ, θ0 ) =
−
n i=1
n i=1
4∆n
ci−1 (β) ci−1 (β0 )
ci−1 (β)
× log
+ oP0 (1)
ci−1 (β0 )
Z
1
1
c(β, x)
P0 1
−→
c(β0 , x)
−
log
π0 (dx)
n→∞ 4
c(β, x) c(β0 , x)
c(β0 , x)
n
1X
P0
ϕi,1 (θ, θ0 )ϕi,j (θ, θ0 ) −→
0, j = 2, 3,
n→∞
n i=1
EMPIRICAL L2 -DISTANCE TEST STATISTICS FOR ERGODIC DIFFUSIONS
19
n
1X
P0
0,
ϕi,2 (θ, θ0 )ϕi,j (θ, θ0 ) −→
n→∞
n i=1
j = 3, 4,
n
1X
P0
0,
ϕi,3 (θ, θ0 )ϕi,4 (θ, θ0 ) −→
n→∞
n i=1
uniformly in θ. Thus the statement of the theorem immediately follows.
Let
1
2
n∆n [∂αh αk Tp,n (θ, θ0 )]h=1,...,m1
k=1,...,m1
1
√
[∂ 2 T (θ, θ0 )]h=1,...,m1
n ∆n αh βk p,n
k=1,...,m2
(6.21)
Cp,n (θ, θ0 ) :=
√1
[∂ 2 T (θ, θ0 )]h=1,...,m1
n ∆n αh βk p,n
k=1,...,m2
1 2
n [∂βh βk Tp,n (θ, θ0 )]h=1,...,m2
k=1,...,m2
where
n
X
∂αh lp,i (θ)∂αk lp,i (θ) + [lp,i (θ) − lp,i (θ0 )]∂α2 h αk lp,i (θ) ,
(6.22)
∂α2 h αk Tp,n (θ, θ0 ) = 2
(6.23)
n
X
∂β2h βk Tp,n (θ, θ0 ) = 2
∂βh lp,i (θ)∂βk lp,i (θ) + [lp,i (θ) − lp,i (θ0 )]∂β2h βk lp,i (θ) ,
(6.24)
n
X
∂α2 h βk Tp,n (θ, θ0 ) = 2
∂αh lp,i (θ)∂βk lp,i (θ) + [lp,i (θ) − lp,i (θ0 )]∂α2 h βk lp,i (θ) .
i=1
i=1
i=1
The following proposition concerning the asymptotic behavior of Cp,n (θ, θ0 ) plays a crucial
role in the proof of Theorem 3.
Proposition 1. Under H0 , assume A1 − A6 and ∆n → 0, n∆n → ∞, as n → ∞, the following
convergences hold
P
0
Cp,n (θ0 , θ0 ) −→
2I(θ0 )
(6.25)
n→∞
and
P
0
0,
sup ||Cp,n (θ0 + θ, θ0 ) − Cp,n (θ0 , θ0 )|| −→
(6.26)
||θ||≤εn
n→∞
εn → 0.
Proof of Proposition 1. We study the uniform convergence in probability of Cp,n (θ, θ0 ). Thus we
prove that uniformly in θ
K1 (θ, θ0 ) + K2 (θ, θ0 )
0
P0
(6.27)
Cp,n (θ, θ0 ) −→ 2K(θ, θ0 ) := 2
0
K3 (θ, θ0 ) + K4 (θ, θ0 )
n→∞
where
Z
∂αh b(α, x)∂αk b(α, x)
c(β0 , x)π0 (dx),
c2 (β, x)
Z
1
c(β0 , x)
c(β0 , x)
c(β, x)
K2 (θ, θ0 ) :=
∂α2 h αk d1 (x, θ)
−1 3
+ log
− 1 π0 (dx)
4
c(β, x)
c(β, x)
c(β0 , x)
Z 2
∂αh αk b(α, x)(b(α, x) − b(α0 , x)) + ∂αh b(α, x)∂αk b(α, x)
1
+
2
c(β, x)
c(β0 , x)
c(β, x)
×
− 1 + log
π0 (dx)
c(β, x)
c(β0 , x)
Z
−∂α2 h αk b(α, x)
+
c(β, x)
K1 (θ, θ0 ) :=
20
A. DE GREGORIO AND S.M. IACUS
1
1
1
c(β0 , x)
−
R(θ, 1, x) + 2
(b(α, x) − b(α0 , x)) π0 (dx)
2 c(β0 , x) c(β, x)
c (β, x)
Z
1
c(β0 , x)∂βh c(β, x)∂βk c(β, x) 3 c(β0 , x)
1 ∂βh c(β, x)∂βk c(β, x)
K3 (θ, θ0 ) :=
−
1
+
π0 (dx),
2
c3 (β, x)
2 c(β, x)
2
c2 (β, x)
Z
1
1
1
c(β0 , x)∂β2h βk log c(β, x)
K4 (θ, θ0 ) :=
−
π0 (dx)
4
c(β, x) c(β0 , x)
Z
1
c(β, x) c(β0 , x) 2
+
log
∂
c−1 (β, x)π0 (dx)
4
c(β0 , x) c(β, x) βh βk
Z
c(β, x)
1
log
+
∂β2h βk log c(β, x)π0 (dx).
4
c(β0 , x)
×
Let us start with the analysis of the quantity
be split in two terms. From (6.16) folllows that
1
2
n∆n ∂αh αk Tp,n (θ, θ0 )
given by (6.22) which can
n
n
1 X
1 X h
h
k
k
∂α lp,i (θ)∂αk lp,i (θ) =
(η (θ) + ηi,2
(θ))(ηi,1
(θ) + ηi,2
(θ))
n∆n i=1 h
n∆n i=1 i,1
for each θ ∈ Θ. Since ∂αh rk0 (∆n , Xi−1 , θ) = ∆n ∂αh bi−1 (α) + R(θ, ∆2n , Xi−1 ), by taking into
account Lemma 5, we get
(6.28)
n
n
1 X h
1 X
k
∂αh lp,i (θ)∂αk lp,i (θ) =
η (θ)ηi,1
(θ) + oP0 (1)
n∆n i=1
n∆n i=1 i,1
n
1 X ∂αh bi−1 (α)∂αk bi−1 (α)
=
(Xi − rk0 (∆n , Xi−1 , θ))2 + oP0 (1)
n∆n i=1
c2i−1 (β)
P
0
−→
K1 (θ, θ0 )
n→∞
uniformly in θ. Now, by resorting (6.15) and (6.18), we rewrite the second term appearing in
(6.22) as follows
n
n
4 X
4
X
X
1 X
1
h,k
[lp,i (θ) − lp,i (θ0 )]∂α2 h αk lp,i (θ) =
ϕi,l (θ, θ0 )δi,j
(θ) .
n∆n i=1
n∆n i=1
j=1
l=1
By applying Lemma 1 and Lemma 5,
Z
n
1 X
P0 3
h,k
ϕi,1 (θ, θ0 )δi,1 (θ) −→
n→∞ 4
n∆n i=1
Z
n
1 X
P0 1
h,k
ϕi,1 (θ, θ0 )δi,2
(θ) −→
n→∞ 2
n∆n i=1
Z
n
1 X
P0 1
h,k
ϕi,1 (θ, θ0 )δi,3
(θ) −→
n→∞ 4
n∆n i=1
the following convergence results hold
1
c2 (β0 , x)
1
2
∂αh αk d1 (θ, x)
−
π0 (dx),
c(β, x) c(β, x) c(β0 , x)
−∂α2 h αk b(α, x)
1
1
R(θ, 1, x)π0 (dx),
−
c(β, x)
c(β, x) c(β0 , x)
c(β0 , x)
∂α2 h αk e1 (θ, x)
− 1 π0 (dx),
c(β, x)
n
1 X
h,k
ϕi,1 (θ, θ0 )δi,4
(θ)
n∆n i=1
2
Z
∂αh αk b(α, x)(b(α, x) − b(α0 , x)) + ∂αh b(α, x)∂αk b(α, x)
c(β0 , x)
P0 1
−1
π0 (dx),
−→
n→∞ 2
c(β, x)
c(β, x)
EMPIRICAL L2 -DISTANCE TEST STATISTICS FOR ERGODIC DIFFUSIONS
21
Z
n
c(β0 , x)
1 X
P0
h,k
ϕi,2 (θ, θ0 )δi,2 (θ) −→
(−∂α2 h αk b(α, x))(b(α, x) − b(α0 , x))π0 (dx),
n→∞
n∆n i=1
c2 (β, x)
Z
n
c(β, x) c(β0 , x) 2
1 X
P0 1
h,k
(θ) −→
log
ϕi,4 (θ, θ0 )δi,1
∂
d1 (θ, x)π0 (dx),
n→∞ 4
n∆n i=1
c(β0 , x) c(β, x) αh αk
Z
n
1 X
c(β, x)
P0 1
h,k
π0 (dx),
ϕi,4 (θ, θ0 )δi,3
(θ) −→
∂α2 h αk e1 (θ, x) log
n→∞ 4
n∆n i=1
c(β0 , x)
n
1 X
h,k
ϕi,4 (θ, θ0 )δi,4
(θ)
n∆n i=1
2
Z
∂αh αk b(α, x)(b(α, x) − b(α0 , x)) + ∂αk b(α, x)∂αh b(α, x)
c(β, x)
P0 1
−→
log
π0 (dx),
n→∞ 2
c(β0 , x)
c(β, x)
n
1 X
P0
h,k
0, j = 1, 3, 4,
(θ) −→
ϕi,2 (θ, θ0 )δi,j
n→∞
n∆n i=1
n
1 X
P0
h,k
0,
ϕi,3 (θ, θ0 )δi,j
(θ) −→
n→∞
n∆n i=1
j = 1, 2, 3, 4,
n
1 X
P0
h,k
ϕi,4 (θ, θ0 )δi,2
(θ) −→
0,
n→∞
n∆n i=1
uniformly in θ. Finally, since d1 (θ, x) = −e1 (θ, x), we get
n
1 X
P0
[lp,i (θ) − lp,i (θ0 )]∂α2 h αk lp,i (θ) −→
K2 (θ, θ0 ).
n→∞
n∆n i=1
(6.29)
uniformly in θ. Hence, by (6.28) and (6.29), we immediately derive
1 2
P0
∂αh αk Tp,n (θ, θ0 ) −→
2(K1 (θ, θ0 ) + K2 (θ, θ0 ))
n→∞
n∆n
(6.30)
uniformly in θ.
Now, we consider the elements of the matrix Cn,p (θ, θ0 ) given by (6.23). First, we study the
convergence probability of
n
n
1X
1X h
h
h
k
k
k
∂βh lp,i (θ)∂βk lp,i (θ) =
(ξ (θ) + ξi,2
(θ) + ξi,3
(θ))(ξi,1
(θ) + ξi,2
(θ) + ξi,3
(θ)).
n i=1
n i=1 i,1
Since ∂βh rk0 (∆n , Xi−1 , θ) = R(θ, ∆2n , Xi−1 ), from Lemma 5 and Lemma 1 we derive
n
(6.31)
n
1X
1X h
k
∂βh lp,i (θ)∂βk lp,i (θ) =
ξ (θ)ξi,2
(θ) + oP0 (1)
n i=1
n i=1 i,2
n
=
1 X ∂βh ci−1 (β)∂βk ci−1 (β)
(Xi − rk0 (∆n , Xi−1 , θ))4
n i=1
4∆2n c4i−1 (β)
n
1 X ∂βh ci−1 (β)∂βk ci−1 (β)
+
(Xi − rk0 (∆n , Xi−1 , θ))2
n i=1
2∆n c3i−1 (β)
n
+
1 X ∂βh ci−1 (β)∂βk ci−1 (β)
+ oP0 (1)
n i=1
4c2i−1 (β)
22
A. DE GREGORIO AND S.M. IACUS
P
0
−→
K3 (θ, θ0 )
n→∞
uniformly in θ. Now, by resorting (6.15) and (6.17), we rewrite the second term appearing in
(6.23) as follows
n
n
4 X
3
X
X
1X
1
h,k
[lp,i (θ) − lp,i (θ0 )]∂β2h βk lp,i (θ) =
(θ) .
ϕi,k (θ, θ0 )νi,j
n i=1
n i=1
j=1
k=1
By taking into account again Lemma 1 and Lemma 5, the following results yield
Z
n
1X
1
1
P0 1
ϕi,1 (θ, θ0 )νi,3 (θ) −→
c(β0 , x)∂β2h βk log c(β, x)
−
π0 (dx)
n→∞ 4
n i=1
c(β, x) c(β0 , x)
Z
n
1X
c(β, x) c(β0 , x) 2
P0 1
ϕi,4 (θ, θ0 )νi,1 (θ) −→
log
∂
c−1 (β, x)π0 (dx)
n→∞ 4
n i=1
c(β0 , x) c(β, x) βh βk
Z
n
1X
c(β, x)
P0 1
ϕi,4 (θ, θ0 )νi,3 (θ) −→
log
∂β2h βk log c(β, x)π0 (dx)
n→∞ 4
n i=1
c(β0 , x)
n
1X
P0
ϕi,1 (θ, θ0 )νi,j (θ) −→
0,
n→∞
n i=1
j = 1, 2,
n
1X
P0
ϕi,k (θ, θ0 )νi,j (θ) −→
0,
n→∞
n i=1
k, j = 1, 2, 3,
n
1X
P0
ϕi,4 (θ, θ0 )νi,2 (θ) −→
0,
n→∞
n i=1
uniformly in θ. Finally
n
(6.32)
1X
P0
[lp,i (θ) − lp,i (θ0 )]∂β2h βk lp,i (θ) −→
K4 (θ, θ0 )
n→∞
n i=1
uniformly in θ. Therefore, by (6.31) and (6.32), we get
(6.33)
1 2
P0
∂
Tp,n (θ, θ0 ) −→
2(K3 (θ, θ0 ) + K4 (θ, θ0 ))
n→∞
n βh βk
uniformly in θ.
Recalling the expressions (6.16), (6.17), (6.20) and (6.15), by means of similar arguments
adopted above, it is not hard to prove that
n
1 X
P0
√
∂α lp,i (θ)∂βk lp,i (θ) −→
0
n→∞
n ∆n i=1 h
and
n
1 X
P0
√
[lp,i (θ) − lp,i (θ0 )]∂α2 h βk lp,i (θ) −→
0
n→∞
n ∆n i=1
uniformly in θ. This implies that
(6.34)
uniformly in θ.
1
P0
√ ∂ 2 Tp,n (θ, θ0 ) −→
0
n→∞
n ∆n αh βk
EMPIRICAL L2 -DISTANCE TEST STATISTICS FOR ERGODIC DIFFUSIONS
23
In conclusion the results (6.30), (6.33) and (6.34) lead to the convergence (6.27). Moreover,
immediately (6.27) implies (6.25) since K(θ0 , θ0 ) = I(θ0 ). From the inequality
sup ||Cp,n (θ0 + θ, θ0 ) − Cp,n (θ0 , θ0 )||
||θ||≤εn
≤ sup ||Cp,n (θ0 + θ, θ0 ) − 2K(θ0 + θ, θ0 )|| + sup ||2K(θ0 + θ, θ0 ) − 2I(θ0 )||
||θ||≤εn
||θ||≤εn
+ ||2I(θ0 ) − Cp,n (θ0 , θ0 )||
follows (6.26). Indeed, (6.25) leads to ||2I(θ0 ) − Cp,n (θ0 , θ0 )|| −→ 0, εn → 0, while the term
n→∞
P
0
0, εn → 0, by the uniformity of the convergence
sup||θ||≤εn ||Cp,n (θ0 +θ, θ0 )−2K(θ0 +θ, θ0 )|| −→
n→∞
P
0
0, εn → 0, because
(i.e. by the result (6.27)). Furthermore, sup||θ||≤εn ||K(θ0 +θ, θ0 )−I(θ0 )|| −→
n→∞
the assumptions A3 and A5 , imply that K(θ, θ0 ) is a continuous function with respect to θ.
Now, we are able to prove Theorem 3.
Proof of Theorem 3. We adopt classical arguments. By Taylor’s formula, we have that
Tp,n (θ̂p,n , θ0 ) = Tp,n (θ0 , θ0 ) + n∂θ Tp,n (θ0 , θ0 )(θ̂p,n − θ0 )
1
+ (ϕ(n)−1/2 (θ̂n − θ0 ))0 Λp,n (θ̂p,n , θ0 ))ϕ(n)−1/2 (θ̂p,n − θ0 )
2
1
= (ϕ(n)−1/2 (θ̂n − θ0 ))0 Λp,n (θ̂p,n , θ0 )ϕ(n)−1/2 (θ̂n − θ0 )
2
where in the last step we denoted by
Z 1
Λp,n (θ̂p,n , θ0 ) := ϕ(n)1/2
(1 − u)∂θ2 Tp,n (θ0 + u(θ̂p,n − θ0 ), θ0 )duϕ(n)1/2
(6.35)
0
Z
=
0
1
1
(1 − u)[Cp,n (θ0 + u(θ̂p,n − θ0 ), θ0 ) − Cp,n (θ0 , θ0 )]du + Cp,n (θ0 , θ0 ).
2
Proposition 1 implies
(6.36)
P
0
Λp,n (θ̂p,n , θ0 ) −→
2I(θ0 ).
n→∞
By taking into account (6.35), (3.16) and (6.36), Slutsky’s theorem allows to conclude the proof.
Proof of Theorem 4. Under H1,n we have that (see Lemma 2 in [21])
d
ϕ(n)−1/2 (θ̂p,n − (θ0 + ϕ(n)h)) −→ N (0, I(θ0 )−1 ).
n→∞
Therefore, under the hypothesis H1,n
d
ϕ(n)−1/2 (θ̂p,n − θ0 ) = ϕ(n)−1/2 (θ̂p,n − θ) + h −→ N (h, I(θ0 )−1 )
n→∞
and
P
θ
Cp,n (θ̂p,n , θ0 ) −→
2I(θ0 )
n→∞
(under H1,n ).
Hence, from (6.35) we obtain the result (4.7).
Acknowlegments
We would like to thank both the referees for their comments which have greatly improved the
first version of the manuscript.
24
A. DE GREGORIO AND S.M. IACUS
Conflict of interest
On behalf of all authors, the corresponding author states that there is no conflict of interest.
References
[1] Aı̈t-Sahalia, Y. (1996) Testing continuous-time models of the spot interest rate, Review of Financial Studies,
70, 385–426.
[2] Aı̈t-Sahalia, Y. (2002). Maximum-likelihood estimation of discretely-sampled diffusions: A closed-form approximation approach, Econometrica, 70, 223–262.
[3] Aı̈t-Sahalia, Y. (2008) Closed-form likelihood expansions for multivariate diffusions, Annals of Statistics 36,
906-937.
[4] Balakrishnan, V., Sanghvi, L. D. (1968) Distance between populations on the basis of attribute data, Biometrica, 24, 859-865.
[5] Bibby, B.M., Sørensen, M. (1995) Martingale estimating functions for discretely observed diffusion processes,
Bernoulli, 1, 17-39.
[6] De Gregorio, A., Iacus, S.M. (2008) Least squares volatility change point estimation for partially observed
diffusion processes, Comm. Statist. Theory Methods, 37, 2342-2357.
[7] De Gregorio, A., Iacus, S.M. (2010) Clustering of discretely observed diffusion processes, Computational
Statistics and Data Analysis, 54, 598-606.
[8] De Gregorio, A., Iacus, S.M. (2012) Adaptive LASSO-type estimation for multivariate diffusion processes,
Econometric Theory, 28, 838-860.
[9] De Gregorio, A., Iacus, S.M. (2013) On a family of test statistics for discretely observed diffusion processes,
Journal of Multivariate Analysis, 122, 292-316.
[10] Florens-Zmirou, D. (1989) Approximate discrete-time schemes for statistics of diffusion processes, Statistics,
20, 547-557.
[11] Genon-Catalot, V., Jacod, J. (1993) On the estimation of the diffusion coefficient for multidimensional
diffusion processes, Ann. Inst. Henri Poincaré, 29, 119–151.
[12] Gobet, E. (2002) LAN property for ergodic diffusions with discrete observations, Ann. I. H. Poincaré – PR,
38, 711-737.
[13] Iacus, S.M. (2008) Simulation and Inference for Stochastic Differential Equations: with R examples, Springer
Series in Statistics, Springer NY.
[14] Iacus, S.M. (2011) Option pricing and estimation of financial models with R, John Wiley & Sons, Ltd.
[15] Iacus, S.M., Uchida, M., Yoshida, N. (2009) Parametric estimation for partially hidden diffusion processes
sampled at discrete times, Stochastic Processes and their Applications, 119, 1580-1600.
[16] Iacus, S.M., Yoshida, N. (2012) Estimation for the change point of volatility in a stochastic differential
equation, Stochastic Processes and their Applications, 122, 1068-1092.
[17] Iacus, S.M., Yoshida, N. (2017) Simulation and Inference for Stochastic Processes with YUIMA, Springer
Series in Statistics, Springer NY.
[18] Jacod, J. (2006) Parametric inference for discretely observed non-ergodic diffusions, Bernoulli, 12, 383-401.
[19] Kallenberg, O. (2001) Foundations of Modern Probability, Springer-Verlag, London.
[20] Kamatani, K., Uchida, M. (2015) Hybrid multi-step estimators for stochastic differential equations based on
sampled data, Statistical Inference for Stochastic Processes, 18, 177-204.
[21] Kitagawa, H., Uchida, M. (2014) Adaptive test statistics for ergodic diffusion processes sampled at discrete
times, Journal of Statistical Planning and Inference, 150, 84-110.
[22] Kessler, M., (1995) Estimation des parametres d’une diffusion par des contrastes corrigés, C. R. Acad. Sci.
Paris Ser. I Math., 320, 359-362. Math. 320 (1995) 359–362.
[23] Kessler, M. (1997) Estimation of an ergodic diffusion from discrete observations, Scand. J. Stat., 24, 211–229.
[24] Kessler, M., Sørensen, M. (1999) Estimating equations based on eigenfunctions for a discretely observed
diffusion process, Bernoulli, 5, 299-314.
[25] Kutoyants, Yu. A. (2004) Statistical Inference for Ergodic Diffusion Processes, Springer-Verlag, London.
[26] Li, C. (2013) Maximum-likelihood estimation for diffusion processes via closed-form density expansions,
Annals of Statistics, 41, 1350-1380.
[27] Morales, D., Pardo, L., Vajda, I. (1997) Some New Statistics for Testing Hypotheses in Parametric Models,
Journal of Multivariate Analysis, 67, 137-168.
[28] Pardo, L. (2006) Statistical Inference Based on Divergence Measures, Chapman & Hall/CRC, London.
[29] Phillips, P.C.B., Yu, J. (2009) A two-stage realized volatility approach to estimation of diffusion processes
with discrete data, Journal of Econometrics, 150. 139-150.
EMPIRICAL L2 -DISTANCE TEST STATISTICS FOR ERGODIC DIFFUSIONS
25
[30] Uchida, M., Yoshida, N. (2012) Adaptive estimation of an ergodic diffusion process based on sampled data,
Stochastic Process. Appl., 17, 181-219.
[31] Uchida, M., Yoshida, N. (2014) Adaptive Bayes type estimators of ergodic diffusion processes from discrete
observations, Statistical Inference for Stochastic Processes, 122, 2885-2924.
[32] Van der Vaart, A.W. (1998) Asymptotic Statistics, Cambridge University press,
[33] Yoshida, N. (1992) Estimation for diffusion processes from discrete observation, J. Multivariate Anal., 41,
220-242.
[34] Yoshida, N. (2011) Polynomial type large deviation inequality and its applications, Ann. Inst. Stat. Mat.,
63, 431-479.
Department of Statistical Sciences, “Sapienza” University of Rome, P.le Aldo Moro, 5 - 00185,
Rome, Italy
Department of Economics, Management and Quantitative Methods, University of Milan, Via Conservatorio 7, 20122 - Milan, Italy
E-mail address: alessandro.degregorio@uniroma1.it
E-mail address: stefano.iacus@unimi.it
| 10math.ST |
1
Born to Learn: the Inspiration, Progress, and Future
of Evolved Plastic Artificial Neural Networks
arXiv:1703.10371v2 [cs.NE] 6 Dec 2017
Andrea Soltoggio, Kenneth O. Stanley, Sebastian Risi
Abstract—Biological plastic neural networks are systems of
extraordinary computational capabilities shaped by evolution, development, and lifetime learning. The interplay of these elements
leads to the emergence of adaptive behavior and intelligence.
Inspired by such intricate natural phenomena, Evolved Plastic
Artificial Neural Networks (EPANNs) use simulated evolution
in-silico to breed plastic neural networks with a large variety
of dynamics, architectures, and plasticity rules: these artificial
systems are composed of inputs, outputs, and plastic components
that change in response to experiences in an environment. These
systems may autonomously discover novel adaptive algorithms,
and lead to hypotheses on the emergence of biological adaptation.
EPANNs have seen considerable progress over the last two
decades. Current scientific and technological advances in artificial
neural networks are now setting the conditions for radically
new approaches and results. In particular, the limitations of
hand-designed networks could be overcome by more flexible
and innovative solutions. This paper brings together a variety
of inspiring ideas that define the field of EPANNs. The main
methods and results are reviewed. Finally, new opportunities and
developments are presented.
Index Terms—Artificial Neural Networks, Lifelong Learning,
Plasticity, Evolutionary Computation.
I. I NTRODUCTION
Over the course of millions of years, evolution has led to the
emergence of innumerable biological systems, and intelligence
itself, crowned by the evolution of the human brain. Evolution,
development, and learning are the fundamental processes that
underpin biological intelligence. Thus, it is no surprise that
scientists have tried to engineer artificial systems to reproduce
such phenomena (Sanchez et al., 1996; Sipper et al., 1997;
Dawkins, 2003).
This endeavor is the domain of fields such as artificial
intelligence (AI) and artificial life (AL) (Langton, 1997) that
seek the emergence of intelligence and life from artificially
designed computation: the idea is to abstract the principles
from the medium, i.e., biological or in-silico, and utilize such
principles to devise algorithms and devices that reproduce
properties of their biological counterparts. A possible way to
design complex and intelligent systems is to simulate natural
evolution in-silico in the field of evolutionary computation
(Holland, 1975; Eiben and Smith, 2015). Sub-fields of evolutionary computation such as evolutionary robotics (Harvey
et al., 1997; Nolfi and Floreano, 2000) and neuroevolution
(Yao, 1999) specifically research algorithms that, exploiting
Department of Computer Science, Loughborough University, LE11 3TU,
Loughborough, UK, a.soltoggio@lboro.ac.uk
Department of Computer Science, University of Central Florida, Orlando,
FL, USA, kstanley@cs.ucf.edu
IT University of Copenhagen, Copenhagen, Denmark, sebr@itu.dk
artificial evolution of physical and neural models, seek more
autonomous design methods and principles towards artificial
intelligence. One assumption is that intelligence in biological
organisms relies considerably on powerful and general learning algorithms, designed by evolution, which are executed
during both development and throughout life. Thus, research
in neuroevolution, and more in general in AI, is progressively
moving towards the design, or the evolution, of lifelong online
learning neural systems (Coleman and Blair, 2012), rather than
static systems.
This paper reviews and organizes the field that studies
evolved plastic artificial neural networks, and introduces the
acronym EPANN. EPANNs are evolved because parts of
their design are determined by an evolutionary algorithm;
they are plastic because they are specifically designed to
undergo lifetime changes, e.g., in the connectivity among
neurons, at various time-scales while experiencing sensorymotor information streams. Lifetime learning is manifested
through continuous adaptation to often dynamic scenarios.
The final capabilities of such networks are determined by the
combination of evolved genetic instructions and continuous
learning that takes place as the network interacts with an
environment.
EPANNs are motivated by the promise of discovering
principles of neural adaptation, learning, and memory, and are
not limited to performance optimization on limited datasets.
This ambitious endeavor also entails a number of research
problems. One problem is the interaction of dynamics across
the evolutionary and learning dimensions, which make the
whole system difficult to design and analyze. An open question
is what are the abstractions that capture essential computational principles and enable intelligence. One further problem
is the high computational cost required to simulate even simple
models of lifetime learning and evolution. Finally, experiments
to evolve adaptation and intelligence are difficult to benchmark
because they aim to achieve loosely defined objectives such
as increasing behavioral complexity, intelligence, adaptability,
and evolvability (Miconi, 2008). For these reasons, EPANNs
explore a much larger search space and rarely compete with
machine learning algorithms specifically designed to improve
performance on well-defined and narrow problems.
The power of EPANNs, however, derives from two autonomous search processes: evolution and learning, which
arguably place them among the most advanced AI and machine
learning systems in terms of open-endedness, potential for
discovery, and human-free design. These systems rely the least
on pre-programmed instructions because they are designed to
autonomously evolve while interacting with a real or simu-
2
lated world. Plastic networks, in particular recurrent plastic
networks, are known for their computational power, which has
also been hypothesized to go beyond that of Turing Machines
(Cabessa and Siegelmann, 2014): evolution can be a valuable
tool to explore the power of those computational structures.
In recent years, progress in a number of relevant areas
has set the stage for renewed advancement of EPANNs:
neural networks, in particular deep networks, are becoming increasingly more successful and used; there has been
a remarkable increase in available computational power by
means of parallel GPU computing and dedicated hardware; a
better understanding of search, complexity, and evolutionary
computation allows for less naive approaches; and finally,
neuroscience and genetics provide us with an increasingly
large set of inspirational principles. This progress has changed
the theoretical and technological landscape in which EPANNs
first emerged, providing greater research opportunities than in
the past.
This paper aims to delineate a unique research field, firstly
defining the inspirational principles of EPANNs (Section II).
The body of research that advanced EPANNs is brought
together and described in Section III. Finally, this paper
describes research directions that are creating opportunities
and challenges (Section IV).
II. I NSPIRATION
EPANNs are strongly inspired by the biological evidence of
natural computation and intelligence (Floreano and Mattiussi,
2008; Downing, 2015). From a computational perspective, a
primary focus is to understand what are the fundamental ingredients and design principles discovered by nature that resulted
in high levels of adaptation and intelligence. Those principles
might then be abstracted and implemented in original ways insilico to achieve similar or better results to those observed in
nature. Crucially, artificial systems do not need to implement
the same constraints and limitations as biological systems
(Bullinaria, 2003). Thus, inspiration is not simple imitation.
We now know that genetic instructions determine the ultimate capabilities of a brain (Deary et al., 2009; Hopkins
et al., 2014). Different animal species manifest different levels
of skills and intelligence because of their genetic blueprint
(Schreiweis et al., 2014). The intricate structure of the brain
emerges from one single zygote cell through a developmental
process (Kolb and Gibb, 2011), which is also strongly affected
by input-output learning experiences throughout early life
(Hensch et al., 1998; Kolb and Gibb, 2011). Yet high levels
of plasticity are maintained throughout the entire lifespan
(Merzenich et al., 1984; Kiyota, 2017). These dimensions,
evolution, development and learning, also known as the phylogenetic (evolution), ontogenetic (development) and epigenetic
(learning) (POE) dimensions (Sipper et al., 1997), are essential
for the emergence of biological plastic brains.
The POE dimensions lead to a number of research questions. Can artificial intelligence systems be entirely engineered
by humans, or do they need to undergo through a less humancontrolled process such as evolution? Do intelligent systems
need to learn, or could they be born already knowing? Opinions and approaches are diverse. EPANNs assume that both
evolution and learning, if not strictly necessary, are conducive
to the emergence of artificial intelligence. While artificial
evolution is justified by the remarkable achievements of natural
evolution, the role of learning has gathered significance in
recent years. We are now more aware of the high level of brain
plasticity, and its impact on the manifestation of behaviors
and skills (LeDoux, 2003; Doidge, 2007; Grossberg, 2012).
Recent developments in machine learning (Michalski et al.,
2013; Alpaydin, 2014) and neural learning (Deng et al., 2013;
LeCun et al., 2015; Silver et al., 2016), highlight the importance of learning from large input-output data and extensive
training. Other areas of cognition such as the capabilities to
make predictions (Hawkins and Blakeslee, 2007), to establish associations (Rescorla, 2014) and to regulate behaviors
(Carver and Scheier, 2012) are also based on learning from
experience. Interestingly, skills such as reading, playing a
musical instrument, or driving a car, are mastered even if
none of those behaviors existed during evolutionary time, and
yet they are mostly unique to humans. Thus, human genetic
instructions have evolved not to learn specific tasks, but to
synthesize recipes to learn a large variety of general skills.
We can conclude that the evolutionary search of learning
mechanisms in EPANNs tackles both the long-running nature
vs. nurture debate (Moore, 2003), and a fundamental and
fast growing research area in AI. EPANNs’ inspiration, and
consequently this review, focus on evolution and learning,
and less on development because that may be seen as a form
of learning if affected by sensory-motor signals. We refer to
Stanley and Miikkulainen (2003a) for an overview of artificial
developmental theories.
Whilst the range of inspiring ideas is large and heterogeneous, the analysis in this review proposes that EPANNs build
upon the following broad principles:
• natural and artificial evolutionary processes,
• plasticity in biological neural networks,
• plasticity in artificial neural networks, and
• lifetime learning environments.
Figure 1 graphically illustrates the inspiration areas and the
particular research topics and fields that make the structure of
this review and define EPANNs.
A. Evolutionary processes
A central idea in evolutionary computation (Goldberg and
Holland, 1988) is that evolutionary processes, similar to those
that occurred in nature during the course of billions of years
(Darwin, 1859; Dobzhansky, 1970), can be simulated with
computer software. This idea led to the belief that intelligent
computer programs could emerge with little human intervention by means of evolution in-silico (Holland and Reitman,
1977; Koza, 1992; Fogel, 2006).
The emergence of evolved intelligent software, however, did
not occur as easily as initially hoped. The reasons for the slow
progress are not completely understood, but a number of problems have been identified, likely related to the simplicity of the
early implementations of evolutionary algorithms and the high
computational requirements. Current topics of investigation
focus on levels of abstraction, diversity in the population,
3
Sec. IV - NEW
DIRECTIONS
abstractions
and
representations
evolving
general
learning
incremental
and social
learning
Sec.III
PROGRESS
Evolutionary
algorithms for
EPANNs
evolving
plasticity
rules
fast
learning
evolving
learning
architectures
evolving
deep
learning
evolving
memory
EPANNs in
Evolutionary
Robotics
interaction of
evolution and
learning
Sec. II
I N S P I R AT I O N
Evolved Plastic Artificial
Neural Networks
ev
pr oluti
oc
o
es nar
se
s( y
II.A
)
fitness
objectives
open-ended
evolution
indirect
encodings
catastrophic
forgetting
behavior and
brain functions
learning &
memory
models of pathways
AI in games
machine learning
deep neural learning
reward-based
learning
neuromodulation
C O N T R I B U T I N G
plasticity in artificial
neural networks (II.C)
F I E L D S
indirectly
encoded
plasticity
adaptive behavior
structural plasticity
plasticity in biological
neural networks (II.B)
evolving
neuromodulation
cognitive problems
models of plasiticity
Hebbian plasticity
measuring
progress
control
problems
evolvability
novelty search and
diversity
GPU/
neuromorphic
hardware
A N D
ng )
rni I.D
ea ts (I
l
e
n
tim me
life viron
en
T O P I C S
Figure 1: Organization of the field of EPANNs, reflected in the structure of this paper.
selection criteria, the concepts of evolvability and scalability
(Wagner and Altenberg, 1996; Pigliucci, 2008; Lehman and
Stanley, 2013), the encoding of genetic information through
the genotype-phenotype mapping processes (Wagner and Altenberg, 1996; Hornby et al., 2002), the deception of fitness
objectives and how to avoid them (Lehman and Stanley, 2008;
Stanley and Lehman, 2015). It is also not clear yet which
stepping stones were more challenging for natural evolution
(Roff, 1993; Stanley and Lehman, 2015) in the evolutionary
path to intelligent and complex forms of life. This lack of
knowledge outlines that our understanding of evolutionary
search processes is incomplete, and thus the potential to
exploit these methods is not fully realized. Nevertheless,
artificial evolution provides a tool to search for counterintuitive and innovative solutions that emerge from adaptation
to the environment with minimal or no human intervention.
Effective evolutionary algorithms and their desirable features
for applications to EPANNs are detailed later in Section III-A.
B. Plasticity in biological neural networks
Biological neural networks demonstrate lifetime learning,
from simple reflex adaptation to the acquisition of astonishing
skills such as social behavior and learning to speak one or
more languages. Those skills are acquired by experiencing
stimuli and actions. The fact that experiences guide life-long
learning is an old and intuitive notion. Scientific evidence of
animal learning from experience was documented initially in
the works of behaviorism with scientists such as Thorndike
(1911), Pavlov (1927), Skinner (1938, 1953), and Hull (1943)
who started to test scientifically how experiences cause a
change in behavior, in particular as a result of learning associations and observable behavioral patterns (Staddon, 1983).
This approach means linking behavior to brain mechanisms
and dynamics, an idea initially entertained by Freud (KϕPPE,
1983) and later by other illustrious scientists (Hebb, 1949;
Kandel, 2007). A seminal contribution to link psychology to
physiology came from Hebb (1949), whose ubiquitous principle that neurons that fire together, wire together is relevant
to understanding both low level neural wiring and high level
psychological theories (Doidge, 2007). Much later, a Hebbiancompatible rule that regulates synaptic changes according to
the firing times of the presynaptic and postsynaptic neurons
was observed by Markram et al. (1997) and named SpikeTiming-Dependent Plasticity (STDP).
The seminal work of Kandel and Tauc (1965), and following
studies (Clark and Kandel, 1984), were the first to demonstrate
that changes in the strength of connectivity among neurons
related to behavior learning. Walters and Byrne (1983) showed
that a single neuron can perform associative learning such as
classical conditioning, a class of learning that is observed in
simple neural systems such as that of the Aplysia (Carew et al.,
1981). Plasticity driven by local neural stimuli, i.e. compatible
with the Hebb synapse (Hebb, 1949; Brown et al., 1990), is
responsible not only for fine tuning, but also for building a
working visual system in the cat’s visual cortex (Rauschecker
and Singer, 1981).
Biological plastic neural networks are also capable of
structural plasticity, which creates new pathways among neurons (Lamprecht and LeDoux, 2004; Chklovskii et al., 2004;
Russo et al., 2010): it occurs primarily during development,
but there is evidence that it continues well into adulthood
(Pascual-Leone et al., 2005). Axon growth, known to be regulated by neurotrophic nerve growth factors (Tessier-Lavigne
and Goodman, 1996), was also modeled computationally in
Roberts et al. (2014). Developmental processes and neural
plasticity are often indistinguishable (Kolb, 1989; PascualLeone et al., 2005) because the brain is highly plastic during
development. Neuroscientific advances reviewed in Damasio
(1999); LeDoux (2003); Pascual-Leone et al. (2005); Doidge
4
(2007); Draganski and May (2008) outline the importance of
structural plasticity in learning motor patterns, associations,
and ways of thinking. Both structural and functional plasticity
in biology is essential to acquiring long-lasting new skills, and
for this reason appears to be an important point of inspiration
for EPANNs.
Neuromodulation (Marder and Thirumalai, 2002; Gu, 2002;
Bailey et al., 2000) is a mechanism that mediates neural processes by means of modulatory signals. Modulatory chemicals
such as acetylcholine (ACh), norepinephrine (NE), serotonin
(5-HT) and dopamine (DA) appear to regulate a large variety
of neural functions, from arousal and behavior (Harris-Warrick
and Marder, 1991; Hasselmo and Schnell, 1994; Marder, 1996;
Katz, 1995; Katz and Frost, 1996), to pattern generation (Katz
et al., 1994), to memory consolidation (Kupfermann, 1987;
Hasselmo, 1995; Marder, 1996; Hasselmo, 1999). Learning by
reward in monkeys was linked to dopaminergic activity during
the 1990s with studies by Schultz et al. (1993, 1997); Schultz
(1998). For these reasons, neuromodulation is considered an
essential element in cognitive and behavioral processes, and
has been the topic of a considerable amount of work in
EPANNs (Section III-F).
In summary, plasticity in biological neural networks is
driven by an extraordinarily rich set of dynamics and chemicals. Relevant computational questions for EPANNs include:
(1) How does the brain structure form—driven both by genetic
instructions and neural activity—and acquire functions and
behavior? (2) What are the key plasticity mechanisms from
biology that can be applied to EPANNs? (3) Are memories,
skills, and behaviors stored in plastic synaptic connections,
in patterns of activities, or a combination of both? Whilst
neuroscience continues to provide inspiration and insight into
plasticity in biological brains, EPANNs serve the complementary objective of seeking, implementing, and verifying
designs of bio-inspired methods for adaptation, learning, and
intelligent behavior.
C. Plasticity in artificial neural networks
In EPANN experiments, evolution can be seen as a metalearning process. Established learning rules are often used as
ingredients that evolution uses to search for good parameter
configurations, efficient combinations of rules and network
topologies, new functions representing novel learning rules,
etc. EPANN experiments are suited to include the largest
possible variety of rules because of (1) the variety of possible tasks in a simulated behavioral experiment and (2) the
flexibility of evolution to combine rules with no assumptions
about their dynamics. The following gives a snapshot of the
extent and scope of various learning algorithms for ANN that
can effectively be building blocks of EPANNs.
In supervised learning, backpropagation is the most popular
learning rule used to train both shallow and deep networks
(Rumelhart et al., 1988; Widrow and Lehr, 1990; LeCun et al.,
2015) for classification or regression. Unsupervised learning
is implemented in neural networks with self-organizing maps
(SOM) (Kohonen, 1982, 1990), auto-encoders (Bourlard and
Kamp, 1988), restricted Boltzmann machines (RBM) (Hinton
and Salakhutdinov, 2006), Hebbian plasticity (Hebb, 1949;
Gerstner and Kistler, 2002a; Cooper, 2005), and various
combinations of the above (Willshaw and Dayan, 1990).
Hebbian rules, given their biological plausibility and unsupervised learning, are a particularly important inspirational
principle. Variations have been proposed to include, e.g., terms
to achieve stability (Oja, 1982; Bienenstock et al., 1982)
and various constraints (Miller and Mackay, 1994), or more
advanced update dynamics such as dual weights for fast and
slow decay (Levy and Bairaktaris, 1995; Hinton and Plaut,
1987; Bullinaria, 2009a; Soltoggio, 2015). The final learning
dynamics of those rules depend often on the topology of the
network and on the parameters of the rule. New Hebbian
rules have been recently proposed to minimize defined cost
functions (Pehlevan et al., 2015; Bahroun et al., 2017).
Often, behavior learning requires a measure of the quality
of such behaviors, implemented as positive or negative values
that represent the consequences of stimulus-action sequences.
This type of learning, also referred to as trial and error, or
operant reward learning in animal learning, was formalized
mathematically as Reinforcement Learning (RL) (Sutton and
Barto, 1998). RL in neural networks (Pennartz, 1997) makes
use of robotic or agent-environment problems with a reward
signal that affects plasticity in the network (Lin, 1993; Niv
et al., 2002; Soltoggio and Stanley, 2012; Auerbach et al.,
2014; Pugh et al., 2014). Neuromodulated plasticity (Fellous
and Linster, 1998) is often used to implement reward-learning
in neural networks (reviewed later in Section III-F). Such
a modulation of signals, or gated learning (Abbott, 1990),
allows for the regulation and learning of behaviors as shown
in numerous computational models (Baxter et al., 1999; Suri
et al., 2001; Birmingham, 2001; Alexander and Sporns, 2002;
Doya, 2002; Fujii et al., 2002; Suri, 2002; Ziemke and Thieme,
2002; Sporns and Alexander, 2003; Krichmar, 2008).
Plastic neural models are also used to demonstrate how
behavior can emerge from a particular circuitry modeled
after biological brains. Computational models of, e.g., the
basal ganglia and modulatory systems may propose plasticity mechanisms and aim to demonstrate the computational
relations among various nuclei, pathways, and learning processes (Krichmar, 2008; Vitay and Hamker, 2010; Schroll and
Hamker, 2015).
Finally, plasticity rules for spiking neural networks (Maass
and Bishop, 2001) aim to demonstrate unique learning mechanisms that emerge from spiking dynamics (Markram et al.,
1997; Izhikevich, 2006, 2007), as well as model biological
synaptic plasticity (Gerstner and Kistler, 2002b).
Plasticity in neural networks, when continuously active, was
also observed to cause catastrophic forgetting (Robins, 1995).
If learning occurs continuously, new information or skills have
the potential to overwrite previously acquired information or
skills (Abraham and Robins, 2005; Finnie and Nader, 2012).
In conclusion, plasticity is central to most approaches to
ANNs, but in the majority of cases, a careful matching and
engineering of rules, architectures and problems are necessary,
requiring considerable design effort. The variety of algorithms
also reflects the variety of problems and solutions. EPANNs
offer a unique testing tool to assess the effectiveness and
5
suitability of different models, or their combination, in a
variety of different scenarios.
process, neural plastic models, and a learning environment:
a simplified graphical representation of an EPANN setup is
given in Fig. 2.
D. Lifelong learning environments
Evolution works on organisms within an environment to
exploit its resources, find survival solutions, and adapt to
changes (Darwin, 1859). However, what makes an environment conducive to the evolution of learning and intelligence
is not clear. In particular, what are the challenges faced by
learning organisms in the natural world, and how can those be
abstracted and ported to a simulated environment?
In the early phases of AI, logic and reasoning were thought
to be key to intelligence (Cervier, 1993), so symbolic inputoutput mappings were employed as tests. Soon it became
evident that intelligence is not only symbol manipulation, but
resides also in subsymbolic problem solving abilities emerging
from the interaction of brain, body, and environment (Steels,
1993; Sims, 1994), so more complex simulators of real-life
environments were developed. A further step in complexity
is represented by high-level planning and strategies required,
e.g., when applying AI to games (Allis et al., 1994; Millington
and Funge, 2016) or articulated robotic tasks. Planning and
decision making with high bandwidth sensory-motor information flow such as those required for humanoid robots or selfdriving vehicles are current benchmarks of intelligent systems.
Finally, environments in which affective dynamics and feelings
play a role are recognized as important for human well
being (De Botton, 2016; Lee and Narayanan, 2005). Those
intelligence-testing environments are effectively the “worlds”
in which EPANNs live and evolve, and thus largely shape the
design process.
Such different testing environments have very different
features, dynamics, and goals. In some cases, a precise target
behavior exists and is known, and an error between desired
and actual behavior can be used in supervised learning. Other
times, the environment provides rewards, but the behaviors
that lead to rewards are unknown. In such cases, the objective
is to search for behavioral policies that lead to collecting
rewards. In yet other cases, it is useful to find relationships and
regularities in the environment to construct complex behaviors
and accomplish particular tasks, which can either lead to
reward, exploit resources, or guarantee survival. Finally, recent
advances in the theory of evolutionary computation (Lehman
and Stanley, 2011; Stanley and Lehman, 2015) suggest that is
not always the fittest, but at times is the novel individual or
behavior that can exploit environmental niches, thus leading
to creative evolutionary processes as those observed in nature.
Such an increasingly large range of testing environments and
multi-skill integration also calls for new evaluation metrics
(see Section IV-H).
III. P ROGRESS ON EVOLVING ARTIFICIAL PLASTIC
NEURAL NETWORKS
After outlining important features of evolutionary algorithms that operate with EPANNs (Section III-A), this section
proceeds to review studies that have evolved plastic neural
networks. Those studies generally include an evolutionary
A. Evolutionary algorithms for EPANNs
As opposed to parameter optimization (Bäck and Schwefel,
1993) in which search spaces are often of fixed dimension
and static, EPANNs evolve in dynamic search spaces in
which lifetime learning further increases the complexity of the
evolutionary search. Evolutionary algorithms (Holland, 1975;
Michalewicz, 1994) for EPANNs often require additional
advanced features to cope with the challenges of the evolution
of learning and open evolutionary design (Bentley, 1999).
Desirable properties of EA for EPANNs are listed in the
following.
1) Variable genotype length and growing complexity: In
many learning problems, the complexity and the characteristics
of the solution required to solve the problem are not known
in advance. Therefore, a desirable property of the EA is
that of increasing the length of the genotype, and thus the
information contained in it, as evolution discovers increasingly
more complex strategies and solutions.
2) Indirect genotype to phenotype encoding: In nature, phenotypes are expressions of a more compact representation, the
genetic code. Similarly, evolutionary algorithms may represent
genetic information in a compact form, which is then mapped
to a larger phenotype.
3) Expressing regularities, repetitions, and patterns: Indirect encodings are beneficial when they can use one set
of instructions in the genotype to generate more parts in
the phenotype. This may involve expressing regularities like
symmetry (e.g. symmetrical neural architectures), repetition
(e.g. neural modules), repetition with variation (similar neural
modules), and patterns (e.g. motifs in the neural architecture).
4) Effective exploration via mutation and recombination:
Genetic mutation and sexual reproduction in nature allow
for the expression of a variety of phenotypes and for the
exploration of new solutions, but seldom lead to highly unfit
individuals (Ay et al., 2007). Similarly, EAs for EPANNs
need to be able to mutate and recombine genomes without
destroying the essential properties of the solutions. As opposed
to other heuristic searches, EAs use recombination to generate
new solutions from two parents: how to effectively recombine
genetic information from two EPANN is still an open question.
5) Genetic encoding of plasticity rules: Just as neural networks need a genetic encoding to be evolved, so do plasticity
rules. EPANN algorithms require the integration of such a rule
in the genome. The encoding may be restricted to a simple
parameter search, but the field is moving progressively towards
the evolution of arbitrary and general plasticity rules. Plasticity
may also be applied to all or parts of the network, thus effectively implementing the evolution of learning architectures.
6) Diversity, survival criteria, and low selection pressure:
The variety of solutions in nature seems to suggest that
diversity is a key aspect of natural evolution. EAs for EPANNs
are likely to perform better when they can maintain diversity
in the population, both at the genotype and phenotype levels.
6
evolution via
EPANN-specific
algorithms
interaction of
evolution and
learning
learning
parameters
variable genotype
length
effective mutation/
(a) Neural
recombination
direct
mapping
Network
indirect
lifetime
learning
plasticity
functions
plastic
architectures:
acquired skills
(b) Sensors
modulatory
Figure 1: Maze Navigating Robot.
The artificial neural
network
neurons
mapping:
that representations
controls the maze navigatingregularities
robot is shown in (a). The layout
genetic
the sensors
is shown in (b). repetitions
Each arrow outside offixed
the robot’s
ofofneural
systems
patterns
body in (b) is a rangefinder sensor
that indicatesarchitectures
the distance to
instincts
&
the closest
obstacle in that direction. The robot has
four pie-slice
genetic
representations
learning
sensors
thatrules
act as a compass towards the goal, activating
when a
of
plasticity
methods
output
sensory
motor
experiences
(a) Medium Map
(b) Hard Map
navigation
In both
maps, the large circle
(e.g. mazes)
represents the starting position of the robot and the small circle
represents the goal. Cul-de-sacs in both maps that lead toward the
behavioral
foraging
goal create the potential for
deception.
diversity
Figure 2:
reward
Maze Navigation
Maps.
learning
control
Fitness-based NEAT, which will be compared to novelty
adaptation
/
search, requires
a fitness
function to reward
maze-navigating
survival/
input
self regulation
speciation
reproduction
robots. Because the objective
is to reach the goal, the fitness
modules
f is defined as the distance from the robot to the goal at the
expanding the neural networks in NEAT) to create novel bepool of genes
mapping
EPANNend of an evaluation: f = benvironment
dg , where bf is a constant
f
haviors as the simpler variants are exhausted.
bias and dg is the distance from the robot to the goal. Given
fitness / objective a maze with no deceptive obstacles, this fitness function deExperiment
fines a monotonic gradient for search to follow. The constant
A good domain for testing novelty search should have a debf ensures
will have (right)
positive allow
fitness. for plastic
Figure 2: Main ceptive
elements
of
an
EPANN
setup
in
which
simulated
evolution
(left) all
andindividuals
an environment
fitness landscape. In such a domain, the search alNEAT
with noveltyThe
search,
on the may
other include
hand, requires
networks to evolve
through
generations
and
learn
within
a
lifetime
in
the
environment.
network
diverse
gorithm following the fitness gradient may perform worse
a novelty metric to distinguish between maze-navigating
an algorithm
following novelty
gradientsneurons,
because novelements, e.g., than
modules
or architectures,
modulatory
inputs/outputs,
learning
parameters,
and
other
structures
that
robots. Defining the novelty metric requires careful conenable learning.elty cannot be deceived; it ignores fitness entirely. A comsideration because it biases the search in a fundamentally
pelling, easily-visualized domain with this property is a twodifferent way than the fitness function. The novelty metdimensional maze navigation task. A reasonable fitness
ric determines the behavior-space through which search will
function for such a domain is how close the maze navigaNiche exploration and behavioral diversity could also play a and III-A3).
Novelty
search (Lehman
andofStanley,
2008,
2011)
proceed.
It is important
that the type
behaviors
that one
tor is to the goal at the end of the evaluation. Thus, dead
hopes to distinguish
are recognized
bysurvival
the metric.of the fittest
key role for creative
design
processes.
Low
selection
pressure
was
introduced
as
an
alternative
to
the
ends that lead close to the goal are local optima to which
Thus, for the maze domain, the behavior of a navigator is
might be crucialantoobjective-based
evolve learning
in deceptive
environments
algorithm
may converge,
which makesasa a selection mechanism (III-A6). Initially, these algorithms
defined
as itstoending
position.
novelty metric
is then the
(see Section III-E).
were not devised
evolve
plasticThe
networks,
but rather
fixed
good model for deceptive problems in general.
Euclidean distance between the ending positions of two inThis paper’s experiments utilize NEAT, which has been
networks in which the final synaptic weights were encoded
dividuals. For example, two robots stuck in the same corner
7) Open-ended
evolution:
evolutionary
experiments
proven
in many Simple
control tasks
[27, 29], including
maze naviin the genome.
To operate
appear similar,
while onewith
robotEPANNs,
that simplythese
sits atalgorithms
the start
seem to quickly
optimize
parameters
and
reach
a
fitness
gation [26], the domain of the experiments in this paper. need to integrate additional genotypical instructions to evolve
position
looks
very
different
from
one
that
reaches
the goal,
plateau without further
improvement.
The maze
domain works Open-ended
as follows. A evolution
robot controlled
plasticitythough
rulesthey
(III-A5).
are both equally viable to the novelty metric.
an ANN must
from aindefinitely
starting pointmore
to an end
seeks to set up by
experiments
thatnavigate
can evolve
This
novelty
metric
rewards
the robot
for endingneuroevoluin a place
By
combining
these
features
(III-A1
- III-A7),
pointgiven
in a fixed
time. The
task is complicated
by(Taycul-de-sacs
complex solutions
sufficient
computational
power
where none
have ended
the method
of traversal
igtion of plastic
networks
aimsbefore;
to search
extremely
largeissearch
that
prevent
a
direct
route
and
that
create
local
optima
in
the
lor et al., 2016).
nored.
This
measure
reflects
that
what
is
important
is
reachspaces in fundamentally different and more creative ways
fitness landscape. The robot (figure 1) has six rangefinding a certain location (i.e. the goal) rather than the method
Many evolutionary
includetoone
moreobstacle
of those
ers that algorithms
indicate the distance
the or
nearest
and four
than traditional
heuristic searches of parameters and hyperof locomotion. Thus, although the novelty metric has no
pie-slice
radar sensors
the goal is within
desirable properties
(Fogel,
2006). that
Duefire
towhen
the complexity
of the
parameters.
knowledge of the final goal, a solution that reaches the goal
twoneuroevolution
effectors result inwas
forcesthe
that reneural network pie-slice.
design, The
the robot’s
field of
will appear novel. Furthermore, the comparison between
spectively
turn
and
propel
the
robot.
This
setup
is
similar
to
first to explore most of those extensions of standard evofitness-based
and rules
novelty-based search is fair because both
B. Evolving
plasticity
the successful maze navigating robots in NERO [26].
lutionary algorithms.
Popular algorithms include early work
scores
are
computed
only based on the distance of the final
Two maps are designed to compare the performance of Early experiments evolved
parameters
of learning rules
of Angeline et NEAT
al. (1994)
Yao andsearch
Liu (1997)
to evolve
position of the robot fromthe
other
points.
with and
fitness-based
and NEAT
with novelty
for
fixed
or
hand-designed
ANN
architectures.
Learning
rules
Finally, to confirm that novelty search is indeed
not anyfixed weights (i.e.
weights
do not2a)
change
while the
agent
search.
The that
first (figure
has deceptive
dead
ends that
are
functions
that
change
the
connection
weight
w
between
thing like random search, NEAT is also tested with a raninteracts with its
environment)
thegoal.
topology
of arbitrary
lead
the robot closeand
to the
To achieve
a higher fitness
dom fitness
assigned
to every
individualasregardless of pertwo neurons,
and are
generally
expressed
local optimum
provided
by a dead
end, the robot
neural networks,than
e.g.,therecurrent
(addressing
III-A1
and III-A4).
formance,
which
means
that
selection
is random. If the maze
travel part of the
way through(NEAT)
a more difficult
path that
Neuroevolution must
of Augmenting
Topologies
(Stanley
∆w of
= evaluations
f (x, θ) ,is recorded.
(1)
is solved, the number
requires a weaving motion. The second maze (figure 2b)
line from the goal to the center of the robot falls within the pieslice. The solid arrow indicates the robot’s heading.
and Miikkulainen, 2002) leverages on three main aspects: a
provides a more deceptive fitness landscape that requires the
is a vector of Parameters
neural signals and θ is a vector of fixed
recombination operator
intended to preserve network func- where xExperimental
search algorithm to explore areas of significantly lower fitparameters. The incoming connection weights w to a neuron
tion (addressingness
III-A4);
speciation
(addressing
III-A6);
and
before finding the global optimum (which is a network
Because NEAT with novelty search differs from origdetermine
activation value
evolution from that
small
to larger
networks (addressing III-A1). i are used
reaches
the goal).
inaltoNEAT
only the
in substituting
a novelty metric for
X
Analog Genetic Encoding (AGE) (Mattiussi and Floreano,
xi = σ
(wji · xj ) ,
(2)
2007) is a method for indirect genotype to phenotype mapping
that can be used in combination with evolution to design where xj are the activation values of presynaptic neurons that
arbitrary network topologies (III-A1 and III-A2). HyperNEAT connect to neuron i with the weights wji , and σ is a nonlinear
(Stanley et al., 2009) is an indirect representation method that function such as the sigmoid or the hyperbolic tangent. x may
combines the NEAT algorithm with compositional patterns provide local signals such as pre and postsynaptic activities,
producing networks (CPPN) (Stanley, 2007) (III-A1, III-A2 the value of the weight w, and modulatory or error signals.
7
Bengio et al. (1990, 1992) proposed the optimization of the
parameters θ of generic learning rules with gradient descent,
simulated annealing, and evolutionary search for problems
such as conditioning, boolean function mapping, and classification. Those studies are also among the first to include a
modulatory term in the learning rules. The optimization was
shown to improve the performance for those different tasks
with respect to manual parameter settings. Chalmers (1990)
evolved a learning rule that applied to every connection and
had a teaching signal. He discovered that, in 20% of the
evolutionary runs, the algorithm evolved the well-known delta
rule, or Widrow-Hoff rule (Widrow et al., 1960), used in backpropagation, thereby demonstrating the validity of evolution
as an autonomous tool to discover learning. Fontanari and
Meir (1991) used the same approach of Chalmers (1990) but
constrained the weights to binary values. Also in this case,
evolution autonomously rediscovered a hand-designed rule, the
directed drift rule by Venkatesh (1993). They also observed
that the performance on new tasks was better when the network
evolved on a larger set of tasks, possibly encouraging the
evolution of more general learning strategies.
With back-propagation of errors (Widrow and Lehr, 1990),
the input vector x of Eq. 1 requires an error signal between
each input/output pair. Rules that use only local signals have
been a more popular choice for EPANNs, though this is
already changing with the rise in effectiveness of deep learning
using back-propagation and related methods. In the simplest
form, the product of presynaptic (xj ) and postsynaptic (xi )
activities, and a learning rate η
∆w = η · xj · xi
.
(3)
is known as Hebbian plasticity (Hebb, 1949; Cooper, 2005).
More generally, any function as in Eq. 1 that uses only local
signals is considered a local plasticity rule for unsupervised
learning. Baxter (1992) evolved a network that applied the basic Hebbian rule in Eq. 3 to a subset of weights (determined by
evolution) to learn four boolean functions of one variable. The
network, called Local Binary Neural Net (LBNN), evolved
to change its weights to one of two possible values (±1),
or have fixed weights. The experiment proved that learning
can evolve when rules are optimized and applied to individual
connections.
Nolfi and Parisi (1993) evolved networks with “autoteaching” inputs, which could then provide an error signal for
the network to adjust weights during lifetime. The implication
is that error-signals do not always need to be hand-designed
but can be discovered by evolution to fit a particular problem.
A set of eight different local rules was used in Rolls and
Stringer (2000) to investigate the evolution of rules in combination with the number of synaptic connections for each neuron, different neuron classes, and other network parameters.
They found that evolution was effective in selecting from large
parameter sets to solve simple linear problems. In Maniadakis
and Trahanias (2006), co-evolution was used to evolve agents
(each being a network) that could use ten different types of
Hebbian-like learning rules for simple navigation tasks: the
authors reported that, despite the increase in the search space,
using many different learning rules results in better perfor-
mance but, understandably, a more difficult analysis of the
evolved systems. Meng et al. (2011) evolved a gene regulatory
network that in turn determined the learning parameters of
the BCM rule (Bienenstock et al., 1982), showing promising
performance in time series classification and other supervised
tasks.
One general finding from these studies is that evolution
operates well within large search spaces, particularly when
a large set of evolvable rules is used.
C. Evolving learning architectures
The interdependency of learning rules and neural architectures led to experiments in which evolution had more
freedom on the network’s design, including the architecture.
The evolution of architectures in ANNs may involve searching
an optimal number of hidden neurons, the number of layers
in a network, particular topologies or modules, the type of
connectivity, and other properties of the network architecture.
In EPANNs, evolving learning architectures implies more
specifically to discover a combination of architectures and
learning rules whose synergetic matching enables particular learning dynamics. As opposed to biological networks,
EPANNs do not have the neurophysiological constraints, e.g.,
short neural connections, sparsity, brain size limits, etc., that
impose limitations on the natural evolution of biological networks. Thus, biologically implausible artificial systems may
nevertheless be evolved in-silico (Bullinaria, 2007b, 2009b).
One seminal early study by Happel and Murre (1994) proposed the evolutionary design of modular neural networks in
which modules could perform unsupervised learning, and the
intermodule connectivity was shaped by Hebbian rules. The
finding was that the use of evolution led to enhanced learning
and better generalization capabilities in comparison to handdesigned networks. In Arifovic and Gencay (2001), the authors
used evolution to optimize the number of inputs and hidden
nodes, and allowed connections in a feedforward neural network to be trained with backpropagation. Abraham (2004) proposed a method called Meta-Learning Evolutionary Artificial
Neural Networks (MLEANN) in which evolution searches for
initial weights, neural architectures and transfer functions for a
range of supervised learning problems to be solved by evolved
networks. The evolved networks were tested in time series prediction and compared with manually designed networks. The
analysis showed that evolution consistently found networks
with better performance than the hand-designed structures.
Khan et al. (2008) proposed an evolutionary developmental
system that created an architecture that adapted with learning:
the network had a dynamic morphology in which neurons
could be inserted or deleted, and synaptic connections formed
and changed in response to stimuli. The networks were evolved
with Cartesian genetic programming and appeared to improve
their performance while playing checkers over the generations.
Downing (2007) looked at different computational models of
neurogenesis to evolve learning architectures. The proposed
evolutionary developmental system focused in particular on
abstraction levels and principles such as Neural Darwinism
(Edelman and Tononi, 2000). A combination of evolution of
8
recurrent networks with a linear learner in the output was
proposed in Schmidhuber et al. (2007), showing that the
evolved RNNs were more compact and resulted in better
learning than randomly initialized echo state networks (Jaeger
and Haas, 2004). In Khan et al. (2011b,a); Khan and Miller
(2014), the authors introduced a large number of bio-inspired
mechanisms to evolve networks with rich learning dynamics.
The idea was to use evolution to design a network that was
capable of advanced plasticity such as dendrite branch and
axon growth and shrinkage, neuron insertion and destruction,
and many others. The system was tested on the Wumpus World
(Russell and Norvig, 2013), a fairly simple problem with no
online learning required, but the purpose was to show that
evolution can design working control networks even within a
large search space of developmental dynamics.
In summary, learning mechanisms and neural architectures
are strongly interdependent, but a large set of available dynamics seem to facilitate the evolution of learning. Thus, EPANNs
become more effective precisely when manual network design
becomes less practical because of complexity and rich dynamics.
D. EPANNs in Evolutionary Robotics
Evolutionary robotics (ER) (Cliff et al., 1993; Floreano and
Mondada, 1994, 1996; Urzelai and Floreano, 2000; Floreano
and Nolfi, 2004) contributed particularly to the development
of EPANNs, providing a testbed for applied controllers in
robotics. Although ER had no specific assumptions on neural systems or plasticity (Smith, 2002), robotics experiments
suggested that neural control structures evolved with fixed
weights perform less well than those evolved with plastic
weights (Nolfi and Parisi, 1996; Floreano and Urzelai, 2001b).
Floreano and Urzelai (2001a) reported that networks evolved
faster when synaptic plasticity and neural architectures were
evolved simultaneously. In particular, plastic networks were
shown to adapt better in the transition from simulation to real
robots. The better simulation-to-hardware transition, and the
increased adaptability in changing ER environments, appeared
intuitive and supported by evidence (Nolfi and Parisi, 1996).
However, the precise nature and magnitude of the changes
from simulation to hardware is not always easy to quantify:
those studies do not clearly outline the precise principles, e.g.,
better or adaptive feedback control, minimization principles,
etc., that are discovered by evolution with plasticity to produce
those advantages. In fact, the behavioral changes required to
switch behaviors in simple ER experiments can also take place
with non-plastic recurrent neural networks because inputs can
switch the network’s activity to different attractors. These
dynamics were observed in particular in recurrent networks in
which neurons are modeled as leaky integrators with different
time constants. Such networks were called continuous time
recurrent neural networks (CTRNN) (Beer and Gallagher,
1992; Yamauchi and Beer, 1994). A study in 2003 observed
similar performance in learning when comparing plastic and
non-plastic CTRNNs (Stanley and Miikkulainen, 2003b). That
study indicates that the learning of plastic networks as described in Blynel and Floreano (2002, 2003) was at that point
still a proof-of-concept rather than a superior learning tool
because networks with fixed weights could achieve similar
performance. Nevertheless, ER maintained a focus on plasticity as demonstrated, e.g., in The Cyber Rodent Project (Doya
and Uchibe, 2005) that investigated the evolution of learning
by seeking to implement a number of features such as (1)
evolution of neural controllers, (2) learning of foraging and
mating behaviors, (3) evolution of learning architectures and
meta-parameters, (4) simultaneous learning of multiple agents
in a body, and (5) learning and evolution in a self-sustained
colony. Plasticity in the form of modulated neural activation
was used in Husbands et al. (1998) and Smith et al. (2002) with
a network that adapts its activation functions according to the
diffusion of a simulated gas spreading to the substrate of the
network. Although the robotic visual discrimination tasks did
not involve lifetime learning, the plastic networks appeared to
evolve faster than a network evolved with fixed activation functions. Similar conclusions were reached in Di Paolo (2003) and
Federici (2005). Di Paolo (2002, 2003) evolved networks with
STDP for a wheeled robot to perform positive and negative
phototaxis, depending on a conditioned stimulus, and observed
that networks with fixed weights could learn but had inferior
performance with respect to plastic networks. Federici (2005)
evolved plastic networks with STDP and an indirect encoding,
showing that plasticity helped performance even if adaptivity
was not required. Stability and evolvability of simple robotic
controllers were investigated in Hoinville et al. (2011) who
focused on EPANNs with homeostatic mechanisms.
Experiments in ER in the 1990s and early 2000s revealed
the extent, complexity, and multitude of ideas behind the
evolutionary design of learning neuro-robotics controllers.
They generally indicate that plasticity helps evolution under
a variety of conditions, even when lifetime learning is not
required, thereby promoting further interest in more specific
topics. Among those are the interaction of evolution and
learning, the evolution of neuromodulation, and the evolution
of indirectly encoded plasticity, as described in the following.
E. Interaction of evolution and learning
EPANN algorithms are influenced by the complex interaction of evolution and learning. Such interaction dynamics are
present in any system, i.e., beyond EPANNs, that seeks the
evolution of learning. Lifetime learning was first suggested to
have an impact on accelerating evolution by Baldwin (1896).
The studies in Smith (1986); Hinton and Nowlan (1987);
Boers et al. (1995); Mayley (1996); Bullinaria (2001) demonstrated the Baldwin effect with computational simulations. For
example, Mayley (1997) investigated the effects of learning
on the rate of evolution, showing that a learning process
during an individual lifetime, although costly, can improve
the speed of evolution of the species. In general, learning
can be classified in (1) learning of invariant environmental
features, typically affected by the Baldwin effect, and (2)
learning variable environmental features. A large body of work
is dedicated to study the interaction and dynamics of all three
POE dimensions, e.g. Paenke (2008); Paenke et al. (2009).
Paenke et al. (2007) investigated the influence of plasticity
9
Fitness/Reward
234
A learning strategy
is discovered by an
220 EPANN
210
200
A navigation strategy
is first evolved without
reward-learning
190
178
0
1000
2000
2500
Generations
Figure 3: Discovery of a learning strategy during evolution.
After approximately 1 000 generations, a reward signal is used
by one network to learn where is the changing location of a
stochastic reward. Figure adapted from Soltoggio et al. (2007).
This experiment suggests that the evolution of learning may
proceed through milestones without obvious gradients in the
search.
and learning on evolution under directional selection; Mery
and Burns (2010) demonstrate that behavioral plasticity affects
evolution.
In EPANNs, similarly complex interaction dynamics between evolution and learning were also observed (Nolfi and
Floreano, 1999). Stone (2007) showed that distributed neural
representations accelerate the evolution of adaptive behavior
because learning part of a skill induced the automatic acquisition of other skill components. One study (Soltoggio et al.,
2007) suggested that evolution discovers learning by discrete
stepping stones: when evolution casually discovers a weak
mechanism of learning, it is sufficient to create an evolutionary
advantage, so the neural mechanism is subsequently evolved
quickly to perfect learning: Fig. 3 shows a sudden jump in the
fitness when the agent suddenly evolves a learning strategy:
such jumps in fitness graphs are common in evolutionary
experiments in which learning is discovered from scratch,
rather than optimized, and were observed as early as in
Fontanari and Meir (1991).
When an environment changes over time, the frequency of
those changes plays a role. With timescales comparable to a
lifetime, evolution may lead to phenotypic plasticity, which is
the capacity for a genotype to express different phenotypes in
response to different environmental conditions (Lalejini and
Ofria, 2016). The frequency of environmental changes was
observed experimentally in plastic neural networks to affect
the evolution of learning (Ellefsen, 2014), revealing a complex
relationship between environmental variability and evolved
learning.
The evolution of neural learning was shown to be inherently
deceptive in Risi et al. (2010); Lehman and Miikkulainen
(2014). In Risi et al. (2009, 2010), EPANN-controlled simulated robots, evolved in a discrete T-Maze domain, revealed
that the evolutionary stepping stones towards adaptive behaviors are often not rewarded by objective-based performance measures. In other words, the necessary evolutionary
stepping stones towards achieving adaptive behavior receive
a lower fitness score than more brittle solutions with low
learning but effective behaviors. A solution to this problem
was devised in Risi et al. (2010, 2009), in which novelty
search (Lehman and Stanley, 2008, 2011) was adopted as a
substitute for performance in the fitness objective with the
aim of finding novel behaviors. Novelty search was observed
to perform significantly better in the T-Maze domain. Lehman
and Miikkulainen (2014) later showed that novelty search can
encourage the evolution of more adaptive behaviors across a
variety of different variations of the T-Maze learning tasks.
As a consequence, novelty search contributed to a philosophical change by questioning the centrality of objectivedriven search in current evolutionary algorithms (Stanley
and Lehman, 2015). By rewarding novel behaviors, novelty
search validates the importance of exploration or curiosity,
previously proposed in Schmidhuber (1991, 2006), also from
an evolutionary viewpoint. With the aim of validating the
same hypothesis, Soltoggio and Jones (2009) devised a simple
EPANN experiment in which exploration was more advantageous than exploitation in the absence of reward learning; to
do this, the reward at a particular location depleted itself if
continuously visited, so that changing location at random in
a T-maze became beneficial. Evolution discovered exploratory
behavior before discovering reward-learning, which in turn,
and surprisingly, led to an earlier evolution of reward-based
learning. Counterintuitively, this experiment suggests that a
stepping stone to evolve reward-based learning is to encourage
non-reward based exploration.
The seminal work in Bullinaria (2003, 2007a, 2009c) proposes the more general hypothesis that learning requires the
evolution of long periods of parental protection and late onset
of maturity. Similarly, Ellefsen (2013b,a) investigates sensitive
and critical periods of learning in evolved neural networks.
This fascinating hypothesis has wider implications for experiments with EPANNs, and more generally for machine learning
and AI. It is therefore foreseeable that future EPANNs will
have a protected childhood during which parental guidance
may be provided (Clutton-Brock, 1991; Klug and Bonsall,
2010; Eskridge and Hougen, 2012).
F. Evolving neuromodulation
Growing neuroscientific evidence on the role of neuromodulation (previously outlined in Section II-B) inspired the design
of experiments with neuromodulatory signals to evolve control
behavior and learning strategies. One particular case is when
neuromodulation gates plasticity. Eq. 1 can be rewritten as as
∆w = m · f (x, θ) ,
(4)
to emphasize the role of m, a modulatory signal used as
a multiplicative factor that can enhance or reduce plasticity
(Abbott, 1990). A network may produce many independent
modulatory signals m targeting different neurons or areas of
the network. Thus, modulation can vary in space and time.
Modulation may also affect other aspects of the network
dynamics, e.g., modulating activations rather than plasticity
(Krichmar, 2008). Graphically, modulation can be represented
as a different type of signal affecting various properties of the
synaptic connections of an afferent neuron i (Fig. 4).
10
wki
m
synthesis of
control signals
modulated
network
standard
neuron
j
xj
plastic
connection
equivalent but different control structures. Fig. 4 also suggests
that modulatory networks require evolution to find two essential topological structures: what signals or combination of
signals trigger modulation, and what neurons are to be targeted
by modulatory signals. In other words, a balance between
fixed and plastic architectures, or selective plasticity (DARPAL2M, 2017), is an intrinsically emergent property of evolved
modulated networks.
modulatory
connection
modulatory
neuron
k
synthesis of
learning signals
modulatory
signal
standard
neuron
i
xi
wji
Figure 4: A modulatory neuron gates plasticity of the synapses
that connect to the postsynaptic neuron. The learning is local,
but a learning signal can be created by one part of the network
and used to regulate learning elsewhere.
Evolutionary search was used to find the parameters of
a neuromodulated Hebbian learning rule in a reward-based
armed-bandit problem in Niv et al. (2002). The same problem
was used later in Soltoggio et al. (2007) to evolve arbitrary
learning architectures with a bio-inspired gene representation
method called Analog Genetic Encoding (AGE) (Mattiussi and
Floreano, 2007). In that study, evolution was used to search
both modulatory topologies and parameters of a particular
form of Eq. 4:
∆w = m · (Axi xj + Bxj + Cxi + D) ,
(5)
where the parameters A to D determined the influence of four
factors in the rule: a multiplicative Hebbian term A, a presynaptic term B, a postsynaptic term C, and pure modulatory,
or heterosynaptic, term D. Such a rule is not dissimilar from
those presented in previous studies (see Section 3.2). However,
when used in combination with modulation and a search
for network topologies, evolution seems to be particularly
effective at solving reward-based problems. Kondo (2007)
proposed an evolutionary design and behavior analysis of
neuromodulatory neural networks for mobile robot control,
validating the potential of the method.
Soltoggio et al. (2008) tested the hypothesis of whether
modulatory dynamics held an evolutionary advantage in Tmaze environments with changing reward locations. In their
algorithm, modulatory neurons were freely inserted or deleted
by random mutations, effectively allowing the evolutionary
selection mechanism to autonomously pick those networks
with advantageous computational components. After evolution, the best performing networks had modulatory neurons
regulating learning, and evolved faster than a control evolutionary experiment that could not employ modulatory neurons.
Modulatory neurons were maintained in the networks in a second phase of the experiment when genetic operators allowed
for the deletion of such neurons but not for their insertion,
thus demonstrating their essential function. In another study,
Soltoggio (2008b) suggested that evolved modulatory topologies may be essential to separate the learning circuity from
the input-output controller, effectively rediscovering an actorcritic architecture, and shortening the input-output pathways
which sped up decision processes. Soltoggio (2008a) showed
that the learning dynamics are affected by tight coupling
between rules and architectures in a search space with many
A number of further studies on the evolution of neuromodulatory dynamics confirmed the evolutionary advantages in
learning scenarios (Soltoggio, 2008c). Silva et al. (2012a) used
simulations of 2-wheel robots performing a dynamic concurrent foraging task, in which scattered food items periodically
changed their nutritive value or become poisonous, similarly to
the setup in Soltoggio and Stanley (2012). The results showed
that when neuromodulation was enabled, learning evolved
faster than when neuromodulation was not enabled, also with
multi-robot distributed systems (Silva et al., 2012b). Nogueira
et al. (2013) also reported evolutionary advantages in foraging
behavior of an autonomous virtual robot when equipped with
neuromodulated plasticity. Harrington et al. (2013) demonstrated how evolved neuromodulation applied to a gene regulatory network consistently generalized better than agents
trained with fixed parameter settings. Interestingly, Arnold
et al. (2013b) showed that neuromodulatory architectures
provided an evolutionary advantage also in reinforcement-free
environments, validating the hypothesis that plastic modulated
networks have higher evolvability in a large variety of tasks.
The evolution of social representations in neural networks
was shown to be facilitated by neuromodulatory dynamics in
Arnold et al. (2013a). An artificial life simulation environment
called Polyword (Yoder and Yaeger, 2014) helped to assess the
advantage of neuromodulated plasticity in various scenarios.
The authors found that neuromodulation may be able to
enhance or diminish foraging performance in a competitive,
dynamic environment.
Neuromodulation was evolved in Ellefsen et al. (2015) in
combination with modularity to address the problem of catastrophic forgetting. In Gustafsson (2016), networks evolved
with AGE (Mattiussi and Floreano, 2007) for video game
playing were shown to perform better with the addition of neuromodulation. In Norouzzadeh and Clune (2016), the authors
showed that neuromodulation produced forward models that
could adapt to changes significantly better than the controls.
They verified that evolution exploited variable learning rates
to perform adaptation when needed. In Velez and Clune
(2017), diffusion-based modulation, i.e., targeting entire parts
of the network, evolved to produce task-specific localized
learning and functional modularity, thus reducing the problem
of catastrophic forgetting.
The evidence in these studies suggests that neuromodulation
is a key ingredient to facilitate the evolution of learning in
EPANNs. They also indirectly suggest that neural systems
with more than one type of signal, e.g., activation and other
modulatory signals, are beneficial in the neuroevolution of
learning.
11
G. Indirectly encoded plasticity
A. Levels of abstraction and representations
An indirect genotype to phenotype mapping means that
evolution operates on a compact genotypical representation
(analogous to the DNA) that is then mapped into a fully
fledged network (analogous to a biological brain). Learning
rules may undergo a similar indirect mapping, so that compact
instructions in the genome expand to fully fledged plasticity
rules in the phenotype. One early study (Gruau and Whitley,
1993) encoded plasticity and development with a grammar
tree, and compared different learning rules on a simple
static task (parity and symmetry), demonstrating that learning
provided an evolutionary advantage in a static scenario. In
non-static contexts, and using a T-Maze domain as learning
task, Risi and Stanley (2010) showed that HyperNEAT, which
usually implements a compact encoding of weight patterns for
large-scale ANNs (Fig. 5a), can also encode patterns of local
learning rules. The approach, called adaptive HyperNEAT,
can encode arbitrary learning rules for each connection in an
evolving ANN based on a function of the ANN’s geometry
(Fig. 5b). Further flexibility was added in Risi and Stanley
(2012) to simultaneously encode the density and placement
of nodes in substrate space. The approach, called adaptive
evolvable-substrate HyperNEAT, makes it possible to indirectly encode plastic ANNs with thousands of connections
that exhibit regularities and repeating motifs. Adaptive ESHyperNEAT allows each individual synaptic connection, rather
than neuron, to be standard or modulatory, thus introducing
further design flexibility. Risi and Stanley (2014) showed how
adaptive HyperNEAT can be seeded to produce a specific
lateral connectivity pattern, thereby allowing the weights to
self-organize to form a topographic map of the input space.
The study shows that evolution can be seeded with specific
plasticity mechanisms that can facilitate the evolution of
specific types of learning.
The effect of indirectly encoded plasticity on the learning
and on the evolutionary process was investigated by Tonelli
and Mouret (2011, 2013). Using an operant conditioning task,
i.e., learning by reward, the authors showed that indirect
encodings that produced more regular neural structures also
improved the general EPANN learning abilities when compared to direct encodings. In an approach similar to adaptive
HyperNEAT, Orchard and Wang (2016) encoded the learning
rule itself as an evolving network. They named the approach
neural weights and bias update (NWB), and observed that
increasing the search space of the possible plasticity rules
created more general solutions than those based on only
Hebbian learning.
Choosing the right level of abstraction and the right representation (Bengio et al., 2013) are themes at the heart of many
problems in AI. In ANNs, low levels of abstraction are more
computationally expensive, but might be richer in dynamics.
High levels are faster to simulate, but require an intuition of
the essential dynamics that are necessary in the model.
One example that abstraction levels are still an open question derives from the comparison of two fields: deep learning
and bio-physical neural models. The performance of deep
networks (Krizhevsky et al., 2012; LeCun et al., 2015; Schmidhuber, 2015), based mainly on rate-based neurons and low
precision in the computation, appears to contrast with the lowlevel simulations of accurate bio-physical models. Examples
are the Blue Brain Project (Markram, 2006) and the Human
Brain Project (Human-Brain-Project, 2017). Although deep
learning and bio-physical models have different aims, they
both represent abstractions of neural computation. Research
in EPANNs is best placed to address the problem of levels
of abstraction because it can autonomously discover the most
appropriate level to evolve neural adaptation and intelligence.
Similarly to abstractions, representations play a critical
role. Compositional Patterns Producing Networks (CPPNs)
(Stanley, 2007), and also the previous work of Sims (1991),
demonstrated that structured phenotypes can be generated
through a function without going through the dynamic developmental process typical of multicellular organisms. Relatedly,
Hornby et al. (2002) showed that the different phenotypical
representations led to considerably different results in the
evolution of furniture. Miller (2014) discussed explicitly the
effect of abstraction levels for evolved developmental learning
networks, in particular in relation to two approaches that model
development at the neuron level or at the network level.
Finding appropriate abstractions and representations, just
as it was fundamental in the advances in deep learning to
represent input spaces and hierarchical features (Bengio et al.,
2013; Oquab et al., 2014), can also extend to representations of
internal models, learning mechanisms, and genetic encodings,
affecting the algorithms’ capabilities of evolving learning
abilities.
IV. F UTURE DIRECTIONS
The progress of EPANNs reviewed so far is based on
rapidly developing theories and technologies. In particular,
new advances in AI, machine learning, neural networks and
increased computational resources are currently creating a new
fertile research landscape, and are setting the groundwork for
new directions for EPANNs. This section presents promising
research themes that have the potential to extend and radically
change the field of EPANNs and AI as a whole.
B. Evolving general learning
One challenge in the evolution of learning is that evolved
learning may simply result in a switch among a finite set of
evolved behaviors, e.g., turning left or right in a T-Maze in a
finite sequence, which is all that evolving solutions encounter
during lifetime. A challenge for EPANNs is to acquire general
learning abilities in which the network is capable of learning problems not encountered during evolution. Mouret and
Tonelli (2014) propose the distinction between the evolution
of behavioral switches and the evolution of synaptic general
learning abilities, and suggest conditions that favor these types
of learning. General learning can be intuitively understood
as the capability to learn any association among input, internal, and output patterns, both in the spatial and temporal
dimensions, regardless of the complexity of the problem. Such
objective clearly poses practical and philosophical challenges.
12
(a) HyperNEAT
(b) Adaptive HyperNEAT
Figure 5: Example of an indirect mapping of plasticity rules from a compact genotype to a larger phenotype. (a) ANN nodes
in HyperNEAT are situated in physical space by assigning them specific coordinates. The connections between nodes are
determined by an evolved Compositional Patterns Producing Network (CPPN; Stanley (2007)), which takes as inputs the
coordinates of two ANN neurons and returns the weight between them. In the normal HyperNEAT approach (a), the CPPN is
queried once for all potential ANN connections when the agent is born. On the other hand, in adaptive HyperNEAT (b), the
CPPN is continually queried during the lifetime of the agent to determine individual connection weight changes based on the
location of neurons and additionally the activity of the presynaptic and postsynaptic neuron, and current connection weight.
Adaptive HyperNEAT is able to indirectly encode a pattern of nonlinear learning rules for each connection in the ANN (right).
Although humans are considered better at general learning
than machines, human learning skills are also specific and not
unlimited (Ormrod and Davis, 2004). Nevertheless, moving
from behavior switches to more general learning is a desirable
feature for EPANNs. Encouraging the emergence of general
learners may likely involve (1) an increased computational cost
for testing in rich environments, which include a large variety
of uncertain and stochastic scenarios with problems of various
complexity, and (2) an increased search space to explore the
evolution of complex strategies and avoid deception.
One promising direction is to evolve the ability to learn
algorithms and programs instead of simple associations. This
approach has recently shown promise in the context of deep
learning; methods such as the Neural Turing Machine (Graves
et al., 2014) can learn simple algorithms such as sort, copy,
etc. and approaches such as the Neural Programmer-Interface
(Reed and de Freitas, 2015) can learn to represent and execute
programs.
C. Incremental and social learning
An important open challenge for machine learning in general is the creation of neural systems that can continuously
integrate new knowledge and skills without forgetting what
they previously learned. A promising approach is progressive
neural networks (Rusu et al., 2016), in which a new network is created for each new task, and lateral connections
between networks allow the system to leverage previously
learned features. In the presence of time delays among stimuli,
actions and rewards, a rule called hypothesis testing plasticity
(HTP) (Soltoggio, 2015) implements fast and slow decay to
consolidate weights and suggests neural dynamics to avoid
catastrophic forgetting. A method to find the best shared
weights across multiple tasks, called elastic weight consolidation (EWC) was proposed in Kirkpatrick et al. (2017).
Plasticity rules that implement weight consolidation, given
their promise to prevent catastrophic forgetting, are likely to
become standard components in EPANNs.
Encouraging modularity (Ellefsen et al., 2015; Durr et al.,
2010) or augmenting evolving networks with a dedicated
external memory component (Lüders et al., 2016) have been
proposed recently. An evolutionary advantage is likely to
emerge for networks that can elaborate on previously learned
sub-skills during their lifetime to learn more complex tasks.
One interesting case in which incremental learning may play
a role is social learning (Best, 1999). EPANNs may learn both
from the environment and from other individuals, from scratch
or incrementally (Offerman and Sonnemans, 1998). In an
early study, McQuesten and Miikkulainen (1997) showed that
neuroevolution can benefit from parent networks teaching their
offspring through backpropagation. When social, learning may
involve imitation, language or communication, or other social
behaviors. Bullinaria (2017) proposes an EPANN framework
to simulate the evolution of culture and social learning. It
is reasonable to assume that future AI learning systems,
whether based on EPANNs or not, will acquire knowledge
through different modalities, involving direct experience with
the environment, but also social interaction, and possibly with
complex incremental learning phases.
D. Fast learning
Animal learning does not always require myriads of trials.
Humans can very quickly generalize from only a few given
examples, possibly leveraging previous experiences and a
long learning process during infancy. This type of learning,
advocated in AI and robotics systems (Thrun and Mitchell,
1995), is currently still missing in EPANNs. Inspiration for
new approaches could come from complementary learning
systems (McClelland et al., 1995; Kumaran et al., 2016) that
humans seem to possess, which include fast and slow learning
components. Additionally, approaches such as probabilistic
program induction seem to be able to learn concepts in oneshot at a human-level in some tasks (Lake et al., 2015). Fast
learning is likely to derive not just from trial-and-error, but also
from mental models that can be applied to diverse problems,
13
similarly to transfer learning (Thrun and Mitchell, 1995; Thrun
and O’Sullivan, 1996; Pan and Yang, 2010). Reusable mental
models, once learned, will allow agents to make predictions
and plan in new and uncertain scenarios with similarities
to previously learned ones. If EPANNs can discover neural
structures or learning rules that allow for generalization, an
evolutionary advantage of such a discovery will lead to its
full emergence and further optimization of such a property.
A rather different approach to accelerate learning was
proposed in Fernando et al. (2008); de Vladar and Szathmáry
(2015) and called Evolutionary Neurodynamics. According
to this theory, replication and selection might happen in a
neural system as it learns, thus mimicking an evolutionary
dynamics at the much faster time scale of a lifetime. We refer
to Fernando et al. (2012); de Vladar and Szathmáry (2015)
for an overview of the field. The appeal of this method is that
evolutionary search can be accelerated by implementing its
dynamics at both the evolution’s and life’s time-scales.
E. Evolving memory
The consequence of learning is memory, both explicit
and implicit (Anderson, 2013), and its consolidation (Dudai,
2012). For a review of computational models of memory see
Fusi (2017). EPANNs may reach solutions in which memory
evolved in different fashions, e.g., preserved as self-sustained
neural activity, encoded by connection weights modified by
plasticity rules, stored with an external memory (e.g. Neural
Turing Machine), or a combination of these approaches. Recurrent neural architectures based on long short-term memory
(LSTM) allow very complex tasks to be solved through
gradient descent training (Greff et al., 2015; Hochreiter and
Schmidhuber, 1997) and have recently shown promise when
combined with evolution (Rawal and Miikkulainen, 2016).
Neuromodulation and weight consolidation could also be used
to target areas of the network where information is stored.
Graves et al. (2014) introduced the Neural Turing Machine
(NTM), networks augmented with an external memory that
allows long-term memory storage. NTMs have shown promise
when trained through evolution (Greve et al., 2016; Lüders
et al., 2016, 2017) or gradient descent (Graves et al., 2014,
2016). The Evolvable Neural Turing Machine (ENTM) showed
good performance in solving the continuous version of the
double T-Maze navigation task (Greve et al., 2016), and
avoided catastrophic forgetting in a continual learning domain
(Lüders et al., 2016, 2017) because memory and control are
separated by design. Research in this area will reveal which
computational systems are more evolvable and how memories
will self organize and form in EPANNs.
F. EPANNs and deep learning
Deep learning has shown remarkable results in a variety
of different fields (Krizhevsky et al., 2012; Schmidhuber,
2015; LeCun et al., 2015). However, the model structures
of these networks are mostly hand-designed, include a large
number of parameters, and require extensive experiments to
discover optimal configurations. Advances in deep networks
have come from manually testing various design aspects such
as: different transfer functions, e.g., the ReLU (Krizhevsky
et al., 2012); generalization mechanisms, e.g. the dropout
(Srivastava et al., 2014); networks with different depth, e.g.
(Simonyan and Zisserman, 2014), and others (Schmidhuber,
2015). With increased computational resources, it is now
possible to search those design aspects with evolution.
Koutnı́k et al. (2014) used evolution to design a controller
that combined evolved recurrent neural networks, for the
control part, and a deep max-pooling convolutional neural
network to reduce the input dimensionality. The study does not
use evolution on the deep preprocessing networks itself, but
demonstrates nevertheless the evolutionary design of a deep
neural controller. Young et al. (2015) used an evolutionary
algorithm to optimize two parameters of a deep network: the
size (range [1,8]) and the number (range [16,126]) of the
filters in a convolutional neural network, showing that the optimized parameters could vary considerably from the standard
best-practice values. An established evolutionary computation
technique, the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) Hansen and Ostermeier (2001), was used in
Loshchilov and Hutter (2016) to optimize the parameters of
a deep network to learn to classify the MNIST dataset. The
authors reported performance close to the state-of-the-art using
30 GPU devices.
Real et al. (2017) and Miikkulainen et al. (2017) showed
that evolutionary search can be used to determine the topology,
hyperparameters and building blocks of deep networks trained
through gradient descent. The performance were shown to rival
those of hand-designed architectures in the CIFAR-10 classification task and a language modeling task (Miikkulainen et al.,
2017), while Real et al. (2017) also tested the method on the
larger CIFAR-100 dataset. Desell (2017) proposes a method
called evolutionary exploration of augmenting convolutional
topologies, inspired by NEAT (Stanley and Miikkulainen,
2002), which evolves progressively more complex unstructured convolutional neural networks using genetically specified
feature maps and filters. This approach is also able to coevolve neural network training hyperparameters. Results were
obtained using 5 500 volunteered computers at the Citizen
Science Grid who were able to evolve competitive results
on the MNIST dataset in under 12 500 trained networks in
a period of approximately ten days. Liu et al. (2017) used
evolution to search for hierarchical architecture representations
showing competitive performance on the CIFAR-10 and Imagenet databases. The Evolutionary DEep Networks (EDEN)
framework (Dufourq and Bassett, 2017) aims to generalize
deep network optimization to a variety of problems and is
interfaced with TensorFlow (Abadi et al., 2016). A number of
similar software frameworks are currently being developed.
Fernando et al. (2017) used evolution to determine a subset of pathways through a network that are trained through
backpropagation, allowing the same network to learn a variety
of different tasks. Fernando et al. (2016) were also able to
rediscover convolutional networks by means of evolution of
Differentiable Pattern Producing Networks (Stanley, 2007).
In summary, evolutionary search allows not only to tune
hyperparameters of the networks, but also to combine different
learning rules in an innovative and counter-intuitive fashion.
14
Evolutionary algorithms are particularly suitable to operate
with unsupervised and reward-based problems. EPANNs need
not be differentiable, as in gradient descent based approaches,
thereby allowing a greater variety in architectures (Greve
et al., 2016). Finally, the deception in the evolution of learning suggests that gradient descent approaches can encounter
limitations when searching for complex learning strategies,
leaving EPANNs as a compelling alternative to discover more
advanced learning models.
G. GPU implementations and neuromorphic hardware
The progress of EPANNs will crucially depend on implementations that take advantage of the increased computational
power of parallel computation with GPUs and neuromorphic
hardware (Jo et al., 2010; Monroe, 2014). Deep learning
greatly benefited from GPU-accelerated machine learning but
also standardized tools (e.g. Torch, Tensorflow, Theano, etc.)
that made it easy for anybody to download, experiment, and
extend promising deep learning models.
Howard et al. (2011, 2012, 2014) devised experiments to
evolve plastic spiking networks implemented as memristors
for simulated robotic navigation tasks. Memristive plasticity
was observed consistently to enable higher performance than
constant-weighted connections in both static and dynamic
reward scenarios. Carlson et al. (2014) used GPU implementations to evolve plastic spiking neural networks with an
evolution strategy, which resulted in an efficient and automated
parameter tuning framework.
In the context of newly emerging technologies, it is worth
noting that, just as GPUs were not developed initially for
deep learning, so novel neural computation tools and hardware
systems, not developed for EPANNs, can now be exploited to
setup the articulated processes and computationally expensive
experiments required by EPANNs.
H. Measuring progress
The number of platforms and environments for testing
the capabilities of intelligent systems is constantly growing,
e.g., the Atari or General Video Game Playing Benchmark
(GVGAI, 2017), the Project Malmo (Microsoft, 2017), or the
OpenAI Universe, “a software platform for measuring and
training an AI’s general intelligence across the world’s supply
of games, websites and other applications” (OpenAI, 2017).
Measuring performance in supervised learning problems is
straight forward because the classification error can be used
as metrics. However, EPANNs are often evolved in rewardbased, survival, or novelty-oriented environments to discover
new, unknown, or creative learning strategies or behaviors.
Thus, desired behaviors or errors are not always defined.
Moreover, the goal for EPANNs is often not to be good at
solving one particular task, but rather to test the capability
to evolve the learning required for a range of problems, to
generalize to new problems, or to recover performance after
a change in the environment. Therefore, EPANNs will require
the community to devise and accept new metrics based on one
or more objectives such as the following:
• the time (in the evolutionary scale) to evolve the learning
mechanisms in one or more scenarios;
•
•
•
•
•
•
•
•
•
the time (in the lifetime scale) for learning in one or more
scenarios;
the number of different tasks that an EPANN evolves to
solve;
a measure of the variety of skills acquired by one
EPANN;
the complexity of the tasks and/or datasets, e.g., variations
in distributions, stochasticity, etc.;
the robustness and generalization capabilities of the
learner;
the recovery time in front of high-level variations or
changes, e.g., data distribution, type of problem, stochasticity levels, etc.;
computational resources used, e.g., number of lifetime
evaluations, length of a lifetime;
size, complexity, and computational requirements of the
solution once deployed.
novelty or richness of the behavior repertoire from multiple solutions, e.g., the variety of different EPANNs and
their strategies that were designed during evolution.
Few of those metrics are currently used to benchmark
machine learning algorithms. Research in EPANNs will foster
the adoption of such criteria as wider performance metrics for
assessing lifelong learning capabilities (Thrun and Pratt, 2012;
DARPA-L2M, 2017) of evolved plastic networks.
V. C ONCLUSION
This paper first described the broad inspiration and aspirations of evolved artificial plastic neural networks (EPANNs),
which are shown to draw from large, diverse, and interdisciplinary areas. The inspirations also reveal the ambitious
and long-term research objectives of EPANNs with important
implications for artificial intelligence and biology.
EPANNs saw considerable progress in the last two decades,
which led to the discovery of a number of further research
themes, such as the design of evolutionary algorithms to
promote the evolution of learning, a better understanding of the
complex interaction dynamics between evolution and learning,
the advantages of multi-signal networks such as modulatory
networks, and the attempt to evolve learning rules, architectures, and appropriate representations of learning mechanisms.
Finally, this paper outlined that recent scientific and technical progress may enable a step change in the potential of
EPANNs. In light of current advances, a pivotal moment for
EPANNs may be at hand as the next AI tool to find flexible algorithms, capable of discovering principles for general
adaptation and intelligent systems.
ACKNOWLEDGEMENTS
We thank John Bullinaria, Kris Carlson, Jeff Clune, Travis
Desell, Keith Downing, Dean Hougen, Joel Lehman, Jeff
Krichmar, Jay McClelland, Robert Merrison-Hort, Julian
Miller, Jean-Baptiste Mouret, James Stone, Eors Szathmary,
and Joanna Turner for insightful discussions and comments
on earlier versions of this paper.
15
R EFERENCES
Martı́n Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo,
Zhifeng Chen, Craig Citro, Greg S Corrado, Andy Davis,
Jeffrey Dean, Matthieu Devin, et al. Tensorflow: Large-scale
machine learning on heterogeneous distributed systems.
arXiv preprint arXiv:1603.04467, 2016.
L. F. Abbott. Modulation of Function and Gated Learning in a
Network Memory. Proceedings of the National Academy of
Science of the United States of America, 87(23):9241–9245,
1990.
Ajith Abraham. Meta learning evolutionary artificial neural
networks. Neurocomputing, 56:1–38, 2004.
Wickliffe C Abraham and Anthony Robins.
Memory
retention–the synaptic stability versus plasticity dilemma.
Trends in neurosciences, 28(2):73–78, 2005.
William H. Alexander and Olaf Sporns. An Embodied Model
of Learning, Plasticity, and Reward. Adaptive Behavior, 10
(3-4):143–159, 2002.
Louis Victor Allis et al. Searching for solutions in games and
artificial intelligence. Ponsen & Looijen, 1994.
Ethem Alpaydin. Introduction to machine learning. MIT press,
2014.
John R Anderson. The architecture of cognition. Psychology
Press, 2013.
Peter J Angeline, Gregory M Saunders, and Jordan B Pollack.
An evolutionary algorithm that constructs recurrent neural
networks. IEEE transactions on Neural Networks, 5(1):54–
65, 1994.
Jasmina Arifovic and Ramazan Gencay. Using genetic algorithms to select architecture of a feedforward artificial
neural network. Physica A: Statistical mechanics and its
applications, 289(3):574–594, 2001.
Solvi Arnold, Reiji Suzuki, and Takaya Arita. Evolution
of social representation in neural networks. In Advances
in Artificial Life, ECAL 2013: Proceedings of the Twelfth
European Conference on the Synthesis and Simulation of
Living Systems. MIT Press, Cambridge, MA, pages 425–
430, 2013a.
Solvi Arnold, Reiji Suzuki, and Takaya Arita. Selection
for reinforcement-free learning ability as an organizing
factor in the evolution of cognition. Advances in Artificial
Intelligence, 2013:8, 2013b.
Joshua Auerbach, Chrisantha Fernando, and Dario Floreano.
Online Extreme Evolutionary Learning Machines. In Proceedings of the Fourteenth International Conference on the
Synthesis and Simulation of Living Systems, 2014.
Nihat Ay, Jessica Flack, and David C Krakauer. Robustness
and complexity co-constructed in multimodal signalling networks. Philosophical Transactions of the Royal Society of
London B: Biological Sciences, 362(1479):441–447, 2007.
Thomas Bäck and Hans-Paul Schwefel. An overview of evolutionary algorithms for parameter optimization. Evolutionary
computation, 1(1):1–23, 1993.
Y. Bahroun, E. Hunsicker, and A. Soltoggio. Building Efficient
Deep Hebbian Networks for Image Classification Tasks.
In International Conference on Artificial Neural Networks,
2017.
C. H. Bailey, M. Giustetto, Y.-Y. Huang, R. D. Hawkins, and
E. R. Kandel. Is heterosynaptic modulation essential for
stabilizing Hebbian plasticity and memory? Nature Reviews
Neuroscience, 1(1):11–20, October 2000.
J Mark Baldwin. A new factor in evolution (continued).
American naturalist, pages 536–553, 1896.
Douglas A. Baxter, Carmen C. Canavier, John W. Clark, and
John H. Byrne. Computational Model of the Serotonergic
Modulation of Sensory Neurons in Aplysia. Journal of
Neurophysiology, 82:1914–2935, 1999.
Jonathan Baxter. The Evolution of Learning Algorithms for
Artificial Neural Networks. Complex Systems, D. Green
and T.Bossomaier, Eds. Amsterdam, The Netherlands: IOS,
pages 313–326, 1992.
Randall D. Beer and John C. Gallagher. Evolving dynamical
neural networks for adaptive behavior. Adaptive Behavior,
1(1):91–122, 1992.
S. Bengio, Y. Bengio, J. Cloutier, and J. Gecsei. On the
optimization of a synaptic learning rule. In Preprints Conf.
Optimality in Artificial and Biological Neural Networks,
Univ. of Texas, Dallas, Feb 6-8, 1992, 1992.
Yoshua Bengio, Samy Bengio, and Jocelyn Cloutier. Learning a synaptic learning rule. Technical report, Université
de Montréal, Département d’informatique et de recherche
opérationnelle, 1990.
Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE
transactions on pattern analysis and machine intelligence,
35(8):1798–1828, 2013.
Peter Bentley. Evolutionary design by computers. Morgan
Kaufmann, 1999.
Michael L Best. How culture can guide evolution: An inquiry
into gene/meme enhancement and opposition. Adaptive
Behavior, 7(3-4):289–306, 1999.
L. Elie Bienenstock, Leon N. Cooper, and Paul W. Munro.
Theory for the development of neuron selectivity: orientation specificity and binocular interaction in visual cortex.
The Journal of Neuroscience, 2(1):32–48, January 1982.
John T. Birmingham. Increasing Sensor Flexibility Through
Neuromodulation. Biological Bulletin, 200:206–210, April
2001.
Jesper Blynel and Dario Floreano. Levels of Dynamics and
Adaptive Behavior in Evolutionary Neural Controllers. In
Proceedings of the seventh international conference on simulation of adaptive behavior on From animals to animats,
pages 272–281. MIT Press Cambridge, MA, USA, 2002.
Jesper Blynel and Dario Floreano. Exploring the T-Maze:
Evolving Learning-Like Robot Behaviors Using CTRNNs.
In EvoWorkshops, pages 593–604, 2003.
Egbert JW Boers, Marko V Borst, and Ida G SprinkhuizenKuyper. Evolving neural networks using the “Baldwin
effect”. In Artificial Neural Nets and Genetic Algorithms,
pages 333–336. Springer, 1995.
Hervé Bourlard and Yves Kamp. Auto-association by multilayer perceptrons and singular value decomposition. Biological cybernetics, 59(4-5):291–294, 1988.
16
Thomas H. Brown, Edward W. Kairiss, and Claude L.
Keenan. Hebbian Synapse: Biophysical Mechanisms and
Algorithms. Annual Review of Neuroscience, 13:475–511,
1990.
John A Bullinaria. Exploring the Baldwin effect in evolving
adaptable control systems. In Connectionist models of learning, development and evolution, pages 231–242. Springer,
2001.
John A Bullinaria. From biological models to the evolution of
robot control systems. Philosophical Transactions: Mathematical, Physical and Engineering Sciences, pages 2145–
2164, 2003.
John A Bullinaria. The effect of learning on life history
evolution. In Proceedings of the 9th annual conference
on Genetic and evolutionary computation, pages 222–229.
ACM, 2007a.
John A Bullinaria. Understanding the emergence of modularity
in neural systems. Cognitive science, 31(4):673–695, 2007b.
John A Bullinaria. Evolved Dual Weight Neural Architectures
to Facilitate Incremental Learning. In IJCCI, pages 427–
434, 2009a.
John A Bullinaria. The importance of neurophysiological constraints for modelling the emergence of modularity. Computational modelling in behavioural neuroscience: closing the
gap between neurophysiology and behaviour, 2:188–208,
2009b.
John A Bullinaria. Lifetime learning as a factor in life history
evolution. Artificial Life, 15(4):389–409, 2009c.
John A. Bullinaria. Imitative and Direct Learning as Interacting Factors in Life History Evolution. Artificial Life, 23,
2017.
Jeremie Cabessa and Hava T Siegelmann. The super-Turing
computational power of plastic recurrent neural networks.
International journal of neural systems, 24(08):1450029,
2014.
Thomas J. Carew, Edgar T. Walters, and Eric R. Kandel.
Classical conditioning in a simple withdrawal reflex in
Aplysia californica. The Journal of Neuroscience, 1(12):
1426–1437, December 1981.
Kristofor D Carlson, Jayram Moorkanikara Nageswaran, Nikil
Dutt, and Jeffrey L Krichmar. An efficient automated
parameter tuning framework for spiking neural networks.
Neuromorphic Engineering Systems and Applications, 8,
2014.
Charles S Carver and Michael F Scheier. Attention and selfregulation: A control-theory approach to human behavior.
Springer Science & Business Media, 2012.
D Cervier. AI: The Tumultuous Search for Artificial Intelligence, 1993.
David J Chalmers. The evolution of learning: An experiment
in genetic connectionism. In Proceedings of the 1990
connectionist models summer school, pages 81–90. San
Mateo, CA, 1990.
Dmitri B Chklovskii, BW Mel, and K Svoboda. Cortical
rewiring and information storage. Nature, 431(7010):782–
788, 2004.
Gregory A. Clark and Eric R. Kandel. Branch-specific
heterosynaptic facilitation in Aplysia siphon sensory cells.
PNAS, 81(8):2577–2581, 1984.
Dave Cliff, Phil Husbands, and Inman Harvey. Explorations
in Evolutionary Robotics. Adaptive Behavior, 2(1):73–110,
1993.
Tim H Clutton-Brock. The evolution of parental care. Princeton University Press, 1991.
Oliver J Coleman and Alan D Blair. Evolving plastic neural
networks for online learning: review and future directions.
In AI 2012: Advances in Artificial Intelligence, pages 326–
337. Springer, 2012.
Steven J. Cooper. Donald O. Hebb’s synapse and learning rule:
a history and commentary. Neuroscience and Biobehavioral
Reviews, 28(8):851–874, January 2005.
Antonio R Damasio. The feeling of what happens: Body and
emotion in the making of consciousness. Houghton Mifflin
Harcourt, 1999.
DARPA-L2M.
DARPA, Lifelong Learning Machines.
https://www.fbo.gov/spg/ODA/DARPA/CMO/
HR001117S0016/listing.html, accessed July 2017, 2017.
Charles Darwin. On the origin of species by means of natural
selection, or The preservation of favoured races in the
struggle for life. Murray, London, 1859.
Richard Dawkins. The evolution of evolvability. On growth,
form and computers, pages 239–255, 2003.
Alain De Botton. Six areas that artificial emotional intelligence
will revolutionise. http://www.wired.co.uk/article/wiredworld-2016-alain-de-botton-artificial-emotionalintelligence, 2016. Accessed: 12/11/2016.
Harold P de Vladar and Eörs Szathmáry. Neuronal boost
to evolutionary dynamics. Interface focus, 5(6):20150074,
2015.
Ian J Deary, Wendy Johnson, and Lorna M Houlihan. Genetic
foundations of human intelligence. Human genetics, 126
(1):215–232, 2009.
Li Deng, Geoffrey Hinton, and Brian Kingsbury. New types
of deep neural network learning for speech recognition and
related applications: An overview. In Acoustics, Speech
and Signal Processing (ICASSP), 2013 IEEE International
Conference on, pages 8599–8603. IEEE, 2013.
Travis Desell. Large Scale Evolution of Convolutional Neural
Networks Using Volunteer Computing. arXiv preprint
arXiv:1703.05422, 2017.
Ezequiel Di Paolo. Spike-timing dependent plasticity for
evolved robots. Adaptive Behavior, 10(3-4):243–263, 2002.
Ezequiel A. Di Paolo. Evolving spike-timing-depending
plasticity for single-trial learning in robots. Philosophical
Transactions of the Royal Society, 361(1811):2299–2319,
October 2003.
Theodosius Dobzhansky. Genetics of the evolutionary process,
volume 139. Columbia University Press, 1970.
Norman Doidge. The brain that changes itself: Stories of personal triumph from the frontiers of brain science. Penguin,
2007.
Keith L. Downing. Supplementing Evolutionary Development Systems with Abstract Models of Neurogenesis. In
17
Dirk Thiersens at al., editor, Proceeding of the Genetic
and Evolutionary Computation Conference (GECCO 2007),
pages 990–998. ACM, 2007.
Keith L Downing. Intelligence Emerging: Adaptivity and
Search in Evolving Neural Systems. MIT Press, 2015.
Kenji Doya. Metalearning and neuromodulation. Neural
Networks, 15(4-6):495–506, 2002.
Kenji Doya and Eiji Uchibe. The Cyber Rodent Project:
Exploration and Adaptive Mechanisms for Self-Preservation
and Self-Reproduction. Adaptive Behavior, 13(2):149–160,
2005.
Bogdam Draganski and Arne May. Training-induced structural
changes in the adult human brain. Behavioural brain
research, 192(1):137–142, 2008.
Yadin Dudai. The restless engram: consolidations never end.
Annual review of neuroscience, 35:227–247, 2012.
Emmanuel Dufourq and Bruce A Bassett. EDEN: Evolutionary Deep Networks for Efficient Machine Learning. arXiv
preprint arXiv:1709.09161, 2017.
Peter Durr, Dario Floreano, and Claudio Mattiussi. Genetic representation and evolvability of modular neural
controllers. IEEE Computational Intelligence Magazine, 5
(3):10–19, 2010.
Gerald M Edelman and Giulio Tononi. A universe of consciousness: How matter becomes imagination. Basic books,
2000.
Agoston E Eiben and Jim Smith. From evolutionary computation to the evolution of things. Nature, 521(7553):476–482,
2015.
Kai Olav Ellefsen. Different regulation mechanisms for
evolved critical periods in learning. In 2013 IEEE Congress
on Evolutionary Computation, pages 1193–1200. IEEE,
2013a.
Kai Olav Ellefsen. Evolved sensitive periods in learning.
Advances in Artificial Life, ECAL 2013: Proceedings of
the Twelfth European Confer- ence on the Synthesis and
Simulation of Living Systems, 12:409–416, 2013b.
Kai Olav Ellefsen. The Evolution of Learning Under Environmental Variability. Artificial Life 14: Proceedings of The
Fourteenth International Conference on the Synthesis and
Simulation of Living Systems, 2014.
Kai Olav Ellefsen, Jean-Baptiste Mouret, and Jeff Clune.
Neural Modularity Helps Organisms Evolve to Learn New
Skills without Forgetting Old Skills. PLoS Comput Biol, 11
(4):e1004128, 2015.
Brent E Eskridge and Dean F Hougen. Nurturing promotes
the evolution of learning in uncertain environments. In Development and Learning and Epigenetic Robotics (ICDL),
2012 IEEE International Conference on, pages 1–6. IEEE,
2012.
Diego Federici. Evolving Developing Spiking Neural Networks. In Proceedings of the IEEE Congress on Evolutionary Computation, CEC 2005, 2005.
Jean-Marc Fellous and Christiane Linster. Computational
Models of Neuromodulation. Neural Computation, 10:771–
805, 1998.
Chrisantha Fernando, KK Karishma, and Eörs Szathmáry.
Copying and evolution of neuronal topology. PloS one, 3
(11):e3775, 2008.
Chrisantha Fernando, Eörs Szathmáry, and Phil Husbands.
Selectionist and evolutionary approaches to brain function: a
critical appraisal. Frontiers in computational neuroscience,
6, 2012.
Chrisantha Fernando, Dylan Banarse, Malcolm Reynolds,
Frederic Besse, David Pfau, Max Jaderberg, Marc Lanctot, and Daan Wierstra. Convolution by Evolution: Differentiable Pattern Producing Networks. arXiv preprint
arXiv:1606.02580, 2016.
Chrisantha Fernando, Dylan Banarse, Charles Blundell, Yori
Zwols, David Ha, Andrei A Rusu, Alexander Pritzel,
and Daan Wierstra. PathNet: Evolution Channels Gradient Descent in Super Neural Networks. arXiv preprint
arXiv:1701.08734, 2017.
Peter SB Finnie and Karim Nader. The role of metaplasticity
mechanisms in regulating memory destabilization and reconsolidation. Neuroscience & Biobehavioral Reviews, 36
(7):1667–1707, 2012.
D. Floreano and F. Mondada. Evolution of Homing Navigation
in a Real Mobile Robot. IEEE Transactions on Systems,
Man, and Cybernetics–Part B, 26(3):396–407, 1996.
D. Floreano and J. Urzelai. Neural morphogenesis, synaptic
plasticity, and evolution. Theory in Biosciences, 240:225–
240, 2001a.
Dario Floreano and Claudio Mattiussi. Bio-inspired artificial
intelligence: theories, methods, and technologies. MIT
press, 2008.
Dario Floreano and Francesco Mondada. Automatic creation
of an autonomous agent: genetic evolution of a neuralnetwork driven robot. In Proceedings of the third international conference on Simulation of adaptive behavior :
from animals to animats 3, pages 421–430. MIT Press,
Cambridge, MA, USA, 1994.
Dario Floreano and Stefano Nolfi. Evolutionary Robotics. The
MIT Press, 2004.
Dario Floreano and Joseba Urzelai. Evolution of Plastic
Control Networks. Auton. Robots, 11(3):311–317, 2001b.
David B Fogel. Evolutionary computation: toward a new
philosophy of machine intelligence, volume 1. John Wiley
& Sons, 2006.
JF Fontanari and R Meir. Evolving a learning algorithm for
the binary perceptron. Network: Computation in Neural
Systems, 2(4):353–359, 1991.
Akinobu Fujii, Nobuhiro Saito, Kota Nakahira, and Akio
Ishiguro. Generation of an Adaptive Controller CPG for
a Quadruped Robot with Neuromodulation Mechanism. In
Proceedings of the 2002 IEEE/RSJ International Conference On Intelligent Robots and Systems, EPFL, Lausanne,
Switzerland, pages 2619–2624, 2002.
Stefano Fusi. Computational models of long term plasticity
and memory. arXiv preprint arXiv:1706.04946, 2017.
Wulfram Gerstner and Werner M Kistler. Mathematical
formulations of Hebbian learning. Biological cybernetics,
87(5-6):404–415, 2002a.
18
Wulfram Gerstner and Werner M Kistler. Spiking neuron
models: Single neurons, populations, plasticity. Cambridge
university press, 2002b.
David E Goldberg and John H Holland. Genetic algorithms
and machine learning. Machine learning, 3(2):95–99, 1988.
Alex Graves, Greg Wayne, and Ivo Danihelka. Neural Turing
machines. arXiv preprint arXiv:1410.5401, 2014.
Alex Graves, Greg Wayne, Malcolm Reynolds, Tim
Harley, Ivo Danihelka, Agnieszka Grabska-Barwińska, Sergio Gómez Colmenarejo, Edward Grefenstette, Tiago Ramalho, John Agapiou, et al. Hybrid computing using a
neural network with dynamic external memory. Nature, 538
(7626):471–476, 2016.
Klaus Greff, Rupesh Kumar Srivastava, Jan Koutnı́k, Bas R
Steunebrink, and Jürgen Schmidhuber. LSTM: A search
space odyssey. arXiv preprint arXiv:1503.04069, 2015.
Rasmus Boll Greve, Emil Juul Jacobsen, and Sebastian Risi.
Evolving Neural Turing Machines for Reward-based Learning. In Proceedings of the Genetic and Evolutionary
Computation Conference 2016, GECCO ’16, pages 117–
124, New York, NY, USA, 2016. ACM. ISBN 978-1-45034206-3.
ST Grossberg. Studies of mind and brain: Neural principles
of learning, perception, development, cognition, and motor
control, volume 70. Springer Science & Business Media,
2012.
Frederic Gruau and Darrell Whitley. Adding learning to the
cellular development of neural networks: Evolution and the
Baldwin effect. Evolutionary computation, 1(3):213–233,
1993.
Q. Gu. Neuromodulatory transmitter systems in the cortex
and their role in cortical plasticity. Neuroscience, 111(4):
815–853, 2002.
Jimmy Gustafsson. Evolving neuromodulatory topologies for
plasticity in video game playing. Master’s thesis, Blekinge
Institute of Technology, Faculty of Computing, Department
of Software Engineering, 2016.
GVGAI.
The General Video game AI Competition.
http://www.gvgai.net, 2017. Accessed: 21/01/2017.
Nikolaus Hansen and Andreas Ostermeier. Completely derandomized self-adaptation in evolution strategies. Evolutionary computation, 9(2):159–195, 2001.
Bart LM Happel and Jacob MJ Murre. Design and evolution
of modular neural network architectures. Neural networks,
7(6):985–1004, 1994.
Kyle Ira Harrington, Emmanuel Awa, Sylvain Cussat-Blanc,
and Jordan Pollack. Robot coverage control by evolved
neuromodulation. In Neural Networks (IJCNN), The 2013
International Joint Conference on, pages 1–8. IEEE, 2013.
Ronald M. Harris-Warrick and Eve Marder. Modulation of
neural networks for behavior. Annual Review of Neuroscience, 14:39–57, 1991.
Inman Harvey, Phil Husbands, Dave Cliff, Adrian Thompson,
and Nick Jakobi. Evolutionary robotics: the Sussex approach. Robotics and autonomous systems, 20(2):205–224,
1997.
Michael E. Hasselmo. Neuromodulation and cortical function:
modeling the physiological basis of behavior. Behavioural
Brain Research, 67:1–27, 1995.
Michael E. Hasselmo. Neuromodulation: acetylcholine and
memory consolidation. Trends in Cognitive Sciences, 3:
351–359, September 1999.
Michael E. Hasselmo and Eric Schnell. Laminar selectivity
of the cholinergic suppression of synaptic transmission in
rat hippocampal region CA1: computational modeling and
brain slice physiology. Journal of Neuroscience, 14(6):
3898–3914, 1994.
Jeff Hawkins and Sandra Blakeslee. On intelligence. Macmillan, 2007.
Donald Olding Hebb. The Organization of Behavior: A
Neuropsychological Theory. Wiley, New York, 1949.
Takao K Hensch, Michela Fagiolini, Nobuko Mataga,
Michael P Stryker, Steinunn Baekkeskov, and Shera F
Kash. Local GABA circuit control of experience-dependent
plasticity in developing visual cortex. Science, 282(5393):
1504–1508, 1998.
Geoffrey E. Hinton and Steven J. Nowlan. How Learning Can
Guide Evolution. Complex Systems, 1:495–502, 1987.
Geoffrey E Hinton and David C Plaut. Using fast weights to
deblur old memories. In Proceedings of the ninth annual
conference of the Cognitive Science Society, pages 177–186,
1987.
Geoffrey E Hinton and Ruslan R Salakhutdinov. Reducing the
dimensionality of data with neural networks. Science, 313
(5786):504–507, 2006.
Sepp Hochreiter and Jürgen Schmidhuber. Long Short-Term
Memory. Neural computation, 9(8):1735–1780, 1997.
Thierry Hoinville, Cecilia Tapia Siles, and Patrick Hénaff.
Flexible and multistable pattern generation by evolving
constrained plastic neurocontrollers. Adaptive Behavior, 19
(3):187–207, 2011.
John H Holland. Adaptation in natural and artificial systems:
an introductory analysis with applications to biology, control, and artificial intelligence. U Michigan Press, 1975.
John H Holland and Judith S Reitman. Cognitive systems
based on adaptive algorithms. ACM SIGART Bulletin, 63:
49–49, 1977.
William D Hopkins, Jamie L Russell, and Jennifer Schaeffer.
Chimpanzee intelligence is heritable. Current Biology, 24
(14):1649–1652, 2014.
Gregory Hornby, Jordan Pollack, et al. Creating high-level
components with a generative representation for body-brain
evolution. Artificial life, 8(3):223–246, 2002.
Gerard Howard, Ella Gale, Larry Bull, Ben de Lacy Costello,
and Andrew Adamatzky. Towards evolving spiking networks with memristive synapses. In Artificial Life (ALIFE),
2011 IEEE Symposium on, pages 14–21. IEEE, 2011.
Gerard Howard, Ella Gale, Larry Bull, Ben de Lacy Costello,
and Andy Adamatzky. Evolution of plastic learning in
spiking networks via memristive connections. IEEE Transactions on Evolutionary Computation, 16(5):711–729, 2012.
Gerard Howard, Larry Bull, Ben de Lacy Costello, Ella Gale,
and Andrew Adamatzky. Evolving spiking networks with
19
variable resistive memories. Evolutionary computation, 22
(1):79–103, 2014.
C. L. Hull. Principles of behavior. New-Your: Appleton
Century, 1943.
Human-Brain-Project. Human Brain Project. https://www.
humanbrainproject.eu/en/, accessed July 2017, 2017.
Phil Husbands, Tom Smith, Nick Jakobi, and Michael O’Shea.
Better Living Through Chemistry: Evolving GasNet for
Robots Control. Connection Science, 10:185–210, 1998.
Eugene M Izhikevich. Polychronization: computation with
spikes. Neural computation, 18(2):245–282, 2006.
Eugene M Izhikevich. Solving the distal reward problem
through linkage of STDP and dopamine signaling. Cerebral
cortex, 17(10):2443–2452, 2007.
Herbert Jaeger and Harald Haas. Harnessing nonlinearity:
Predicting chaotic systems and saving energy in wireless
communication. science, 304(5667):78–80, 2004.
Sung Hyun Jo, Ting Chang, Idongesit Ebong, Bhavitavya B
Bhadviya, Pinaki Mazumder, and Wei Lu. Nanoscale memristor device as synapse in neuromorphic systems. Nano
letters, 10(4):1297–1301, 2010.
E. R. Kandel and L. Tauc. Heterosynaptic facilitation in
neurones of the abdominal ganglion of Aplysia depilans.
The Journal of Physiology, 181:1–27, 1965.
Eric R. Kandel. In Search of Memory - The Emergence of a
New Science of Mind. WW Norton & Company, New York,
2007.
Paul S. Katz. Intrinsic and extrinsic neuromodulation of
motor circuits. Current Opinion in Neurobiology, 5:799–
808, 1995.
Paul S. Katz and William N. Frost. Intrinsic neuromodulation:
altering neuronal circuits from within. Trends in Neurosciences, 19:54–61, 1996.
Paul S. Katz, Peter A. Getting, and William N. Frost. Dynamic
neuromodulation of synaptic strength intrinsic to a central
pattern generator circuit. Nature, 367:729–731, 1994.
Gul Muhammad Khan and Julian Francis Miller. In search of
intelligence: evolving a developmental neuron capable of
learning. Connection Science, 26(4):297–333, 2014.
Gul Muhammad Khan, Julian F. Miller, and David M. Halliday. Coevolution of neuro-developmental programs that
play checkers. In Springer, editor, Proceedings of 8th
International Conference on Evolvable Systems (ICES):
From Biology to Hardware, volume 5216, pages 352–361,
2008.
Gul Muhammad Khan, Julian F. Miller, and David M. Halliday. Evolution of Cartesian Genetic Programs for Development of Learning Neural Architecture. Evolutionary
Computation, 19(3):469–523, 2011a.
Maryam Mahsal Khan, Gul Muhammad Khan, and Julian F.
Miller. Developmental Plasticity in Cartesian Genetic
Programming based Neural Networks. In ICINCO 2011
- Proceedings of the 8th International Conference on Informatics in Control, Automation and Robotics, Volume
1, Noordwijkerhout, The Netherlands, 28 - 31 July, 2011,
pages 449–458, 2011b.
James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan,
John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska,
Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and
Raia Hadsell. Overcoming catastrophic forgetting in neural
networks. Proceedings of the National Academy of Sciences,
2017.
Tomomi Kiyota. Neurogenesis and Brain Repair. In Neuroimmune Pharmacology, pages 575–597. Springer, 2017.
Hope Klug and Michael B Bonsall. Life history and the
evolution of parental care. Evolution, 64(3):823–835, 2010.
Teuvo Kohonen. Self-organized formation of topologically
correct feature maps. Biological cybernetics, 43(1):59–69,
1982.
Teuvo Kohonen. The Self-Organizing Map. Proceedings of
the IEEE, 78(9):1464–1480, 1990.
Bryan Kolb. Brain development, plasticity, and behavior.
American Psychologist, 44(9):1203, 1989.
Bryan Kolb and Robbin Gibb. Brain plasticity and behaviour
in the developing brain. Journal of the Canadian Academy
of Child & Adolescent Psychiatry, 20(4), 2011.
Toshiyuki Kondo. Evolutionary design and behaviour analysis
of neuromodulatory neural networks for mobile robots control. Applied Soft Computing, 7(1):189–202, January 2007.
Jan Koutnı́k, Jürgen Schmidhuber, and Faustino Gomez.
Evolving deep unsupervised convolutional networks for
vision-based reinforcement learning. In Proceedings of
the 2014 Annual Conference on Genetic and Evolutionary
Computation, pages 541–548. ACM, 2014.
John R Koza. Genetic programming: on the programming of
computers by means of natural selection, volume 1. MIT
press, 1992.
Jeffrey L Krichmar. The neuromodulatory system: a framework for survival and adaptive behavior in a challenging
world. Adaptive Behavior, 16(6):385–399, 2008.
Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton.
Imagenet classification with deep convolutional neural networks. In Advances in neural information processing
systems, pages 1097–1105, 2012.
Dharshan Kumaran, Demis Hassabis, and James L McClelland. What learning systems do intelligent agents need?
complementary learning systems theory updated. Trends in
Cognitive Sciences, 20(7):512–534, 2016.
Irving Kupfermann. Cellular Neurobiology: Neuromodulation.
Science, 236:863, 1987.
SIMO KϕPPE. The psychology of the neuron: Freud, Cajal
and Golgi. Scandinavian journal of psychology, 24(1):1–12,
1983.
Brenden M Lake, Ruslan Salakhutdinov, and Joshua B Tenenbaum. Human-level concept learning through probabilistic
program induction. Science, 350(6266):1332–1338, 2015.
Alexander Lalejini and Charles Ofria. The Evolutionary
Origins of Phenotypic Plasticity. In Proceedings of the
Artificial Life Conference, 2016.
Raphael Lamprecht and Joseph LeDoux. Structural plasticity
and memory. Nature Reviews Neuroscience, 5(1):45–54,
2004.
20
Christopher G Langton. Artificial life: An overview. Mit Press,
1997.
Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep
learning. Nature, 521(7553):436–444, 2015.
Joseph E LeDoux. Synaptic self: How our brains become who
we are. Penguin, 2003.
Chul Min Lee and Shrikanth S Narayanan. Toward detecting
emotions in spoken dialogs. IEEE transactions on speech
and audio processing, 13(2):293–303, 2005.
Joel Lehman and Risto Miikkulainen. Overcoming Deception in Evolution of Cognitive Behaviors. In Proceedings
of the Genetic and Evolutionary Computation Conference
(GECCO 2014), Vancouver, BC, Canada, July 2014.
Joel Lehman and Kenneth O Stanley. Exploiting OpenEndedness to Solve Problems Through the Search for Novelty. In ALIFE, pages 329–336, 2008.
Joel Lehman and Kenneth O Stanley. Abandoning objectives:
Evolution through the search for novelty alone. Evolutionary computation, 19(2):189–223, 2011.
Joel Lehman and Kenneth O Stanley. Evolvability is inevitable: increasing evolvability without the pressure to
adapt. PloS one, 8(4):e62186, 2013.
Joseph P Levy and Dimitrios Bairaktaris. Connectionist dualweight architectures. Language and Cognitive Processes,
10(3-4):265–283, 1995.
Long-Ji Lin. Reinforcement Learning for Robots Using Neural Networks. PhD thesis, School of Computer Science,
Carnegie Mellon University, 1993.
Hanxiao Liu, Karen Simonyan, Oriol Vinyals, Chrisantha
Fernando, and Koray Kavukcuoglu. Hierarchical representations for efficient architecture search. arXiv preprint
arXiv:1711.00436, 2017.
Ilya Loshchilov and Frank Hutter. CMA-ES for Hyperparameter Optimization of Deep Neural Networks. arXiv preprint
arXiv:1604.07269, 2016.
Benno Lüders, Mikkel Schläger, and Sebastian Risi. Continual
Learning through Evolvable Neural Turing Machines. In
Proceedings of the NIPS 2016 Workshop on Continual
Learning and Deep Networks (CLDL 2016), 2016.
Benno Lüders, Mikkel Schläger, Aleksandra Korach, and
Sebastian Risi. Continual and One-Shot Learning through
Neural Networks with Dynamic External Memory. In
Proceedings of the 20th European Conference on the Applications of Evolutionary Computation (EvoApplications
2017), 2017.
Wolfgang Maass and Christopher M Bishop. Pulsed neural
networks. MIT press, 2001.
Michail Maniadakis and Panos Trahanias. Modelling brain
emergent behaviours through coevolution of neural agents.
Neural Networks, 19(5):705–720, 2006.
Eve Marder. Neural modulation: Following your own rhythm.
Current Biology, 6(2):119–121, 1996.
Eve Marder and Vatsala Thirumalai. Cellular, synaptic and
network effects of neuromodulation. Neural Networks, 15:
479–493, 2002.
Henry Markram. The Blue Brain Project. Nature Reviews
Neuroscience, 7(2):153–160, 2006.
Henry Markram, Joachim Lübke, Michael Frotscher, and Bert
Sakmann. Regulation of synaptic efficacy by coincidence of
postsynaptic APs and EPSPs. Science, 275(5297):213–215,
1997.
Claudio Mattiussi and Dario Floreano. Analog genetic encoding for the evolution of circuits and networks. IEEE
Transactions on evolutionary computation, 11(5):596–607,
2007.
Giles Mayley. Landscapes, learning costs, and genetic assimilation. Evolutionary Computation, 4(3):213–234, 1996.
Giles Mayley. Guiding or hiding: Explorations into the effects
of learning on the rate of evolution. In Proceedings of the
Fourth European Conference on Artificial Life, volume 97,
pages 135–144, 1997.
James L McClelland, Bruce L McNaughton, and Randall C
O’reilly. Why there are complementary learning systems
in the hippocampus and neocortex: insights from the successes and failures of connectionist models of learning and
memory. Psychological review, 102(3):419, 1995.
Paul McQuesten and Risto Miikkulainen. Culling and Teaching in Neuro-Evolution. In ICGA, pages 760–767, 1997.
Yan Meng, Yaochu Jin, and Jun Yin. Modeling activitydependent plasticity in BCM spiking neural networks with
application to human behavior recognition. IEEE transactions on neural networks, 22(12):1952–1966, 2011.
Frederic Mery and James G Burns. Behavioural plasticity: an
interaction between evolution and experience. Evolutionary
Ecology, 24(3):571–583, 2010.
Michael M Merzenich, Randall J Nelson, Michael P Stryker,
Max S Cynader, Axel Schoppmann, and John M Zook.
Somatosensory cortical map changes following digit amputation in adult monkeys. Journal of comparative neurology,
224(4):591–605, 1984.
Zbigniew Michalewicz. GAs: What are they? In Genetic
algorithms+ data structures= evolution programs, pages
13–30. Springer, 1994.
Ryszard S Michalski, Jaime G Carbonell, and Tom M
Mitchell. Machine learning: An artificial intelligence approach. Springer Science & Business Media, 2013.
Thomas Miconi. The road to everywhere: Evolution, complexity and progress in natural and artificial systems. PhD
thesis, University of Birmingham, 2008.
Microsoft. Project malmo. https://github.com/Microsoft/
malmo, accessed Nov 2017, 2017.
Risto Miikkulainen, Jason Liang, Elliot Meyerson, Aditya
Rawal, Dan Fink, Olivier Francon, Bala Raju, Arshak
Navruzyan, Nigel Duffy, and Babak Hodjat. Evolving Deep
Neural Networks. arXiv preprint arXiv:1703.00548, 2017.
Julian F Miller. Neuro-Centric and Holocentric Approaches to
the Evolution of Developmental Neural Networks. In Growing Adaptive Machines, pages 227–249. Springer, 2014.
Kenneth D. Miller and David J. C. Mackay. The Role of
Constraints in Hebbian Learning. Neural Computation, 6:
100–126, 1994.
Ian Millington and John Funge. Artificial intelligence for
games. CRC Press, 2016.
21
Don Monroe. Neuromorphic computing gets ready for the
(really) big time. Communications of the ACM, 57(6):13–
15, 2014.
David S Moore. The Dependent Gene: The Fallacy of “Nature
vs. Nurture”. 2003.
Jean-Baptiste Mouret and Paul Tonelli. Artificial evolution of
plastic neural networks: a few key concepts. In Growing
Adaptive Machines, pages 251–261. Springer, 2014.
Yael Niv, Daphna Joel, Isaac Meilijson, and Eytan Ruppin.
Evolution of Reinforcement Learning in Uncertain Environments: A Simple Explanation for Complex Foraging
Behaviours. Adaptive Behavior, 10(1):5–24, 2002.
B Nogueira, Y Lenon, C Eduardo, Creto A Vidal, and
Joaquim B Cavalcante Neto. Evolving plastic neuromodulated networks for behavior emergence of autonomous
virtual characters. In Advances in Artificial Life, ECAL
2013: Proceedings of the Twelfth European Conference on
the Synthesis and Simulation of Living Systems. MIT Press,
Cambridge, MA, pages 577–584, 2013.
Stefano Nolfi and Dario Floreano. Learning and evolution.
Autonomous robots, 7(1):89–113, 1999.
Stefano Nolfi and Dario Floreano. Evolutionary robotics,
2000.
Stefano Nolfi and Domenico Parisi. Auto-teaching: networks
that develop their own teaching input. In Free University of
Brussels. Citeseer, 1993.
Stefano Nolfi and Domenico Parisi. Learning to adapt to
changing environments in evolving neural networks. Adaptive behavior, 5(1):75–98, 1996.
Mohammad Sadegh Norouzzadeh and Jeff Clune. Neuromodulation Improves the Evolution of Forward Models. In
Proceedings of the Genetic and Evolutionary Computation
Conference, 2016.
Theo Offerman and Joep Sonnemans. Learning by experience
and learning by imitating successful others. Journal of
economic behavior & organization, 34(4):559–575, 1998.
Erkki Oja. A Simplified Neuron Model as a Principal Component Analyzer. Journal of Mathematical Biology, 15(3):
267–273, November 1982.
OpenAI. OpenAI Universe. https://openai.com/blog/universe/,
2017. Accessed: 21/01/2017.
Maxime Oquab, Leon Bottou, Ivan Laptev, and Josef Sivic.
Learning and transferring mid-level image representations
using convolutional neural networks. In Proceedings of the
IEEE conference on computer vision and pattern recognition, pages 1717–1724, 2014.
Jeff Orchard and Lin Wang. The evolution of a generalized
neural learning rule. In Neural Networks (IJCNN), 2016
International Joint Conference on, pages 4688–4694. IEEE,
2016.
Jeanne Ellis Ormrod and Kevin M Davis. Human learning.
Merrill, 2004.
Ingo Paenke. Dynamics of Evolution and Learning. PhD thesis, Universität Karlsruhe (TH), Fakultät für Wirtschaftwsissenschaften, 2008.
Ingo Paenke, Bernhard Sendhoff, and Tadeusz J Kawecki.
Influence of plasticity and learning on evolution under
directional selection. The American Naturalist, 170(2):E47–
E58, 2007.
Ingo Paenke, T Kawecki, and Bernhard Sendhoff. The influence of learning on evolution: A mathematical framework.
Artificial Life, 15(2):227–245, 2009.
Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. IEEE Transactions on knowledge and data engineering,
22(10):1345–1359, 2010.
Alvaro Pascual-Leone, Amir Amedi, Felipe Fregni, and Lotfi B
Merabet. The plastic human brain cortex. Annu. Rev.
Neurosci., 28:377–401, 2005.
I. P. Pavlov. Conditioned reflexes. Oxford : Oxford University
Press, 1927.
Cengiz Pehlevan, Tao Hu, and Dmitri B Chklovskii. A
Hebbian/anti-Hebbian neural network for linear subspace
learning: A derivation from multidimensional scaling of
streaming data. Neural computation, 2015.
C. M. A. Pennartz. Reinforcement Learning by Hebbian
Synapses with Adaptive Threshold. Neuroscience, 81(2):
303–319, 1997.
Massimo Pigliucci. Is evolvability evolvable? Nature Reviews
Genetics, 9(1):75–82, 2008.
Justin Pugh, Andrea Soltoggio, and Kenneth O. Stanley. Realtime Hebbian Learning from Autoencoder Features for
Control Tasks. In Proceedings of the Fourteenth International Conference on the Synthesis and Simulation of Living
Systems, 2014.
Josef P. Rauschecker and Wolf Singer. The effects of early
visual experience on the cat’s visual cortex and their possible explanation by Hebb synapses. Journal of Physiology,
310:215–239, 1981.
Aditya Rawal and Risto Miikkulainen. Evolving Deep LSTMbased Memory Networks Using an Information Maximization Objective. In Proceedings of the Genetic and Evolutionary Computation Conference 2016, GECCO ’16, pages
501–508, New York, NY, USA, 2016. ACM. ISBN 978-14503-4206-3.
E. Real, S. Moore, A. Selle, S. Saxena, Y. L. Suematsu, Q. Le,
and A. Kurakin. Large-Scale Evolution of Image Classifiers.
ArXiv e-prints, March 2017.
Scott Reed and Nando de Freitas. Neural programmerinterpreters. arXiv preprint arXiv:1511.06279, 2015.
Robert A Rescorla. Pavlovian Second-Order Conditioning
(Psychology Revivals): Studies in Associative Learning.
Psychology Press, 2014.
Sebastian Risi and Kenneth O. Stanley. Indirectly Encoding
Neural Plasticity as a Pattern of Local Rules. In Proceedings
of the 11th International Conference on Simulation of
Adaptive Behavior (SAB 2010), New York, 2010. Springer.
Sebastian Risi and Kenneth O Stanley. A unified approach to
evolving plasticity and neural geometry. In Neural Networks
(IJCNN), The 2012 International Joint Conference on, pages
1–8. IEEE, 2012.
22
Sebastian Risi and Kenneth O. Stanley.
Guided SelfOrganization in Indirectly Encoded and Evolving Topographic Maps. In Proceedings of the Genetic and Evolutionary Computation Conference (GECCO-2014). New York,
NY: ACM (8 pages), 2014.
Sebastian Risi, Sandy D Vanderbleek, Charles E Hughes,
and Kenneth O Stanley. How novelty search escapes the
deceptive trap of learning to learn. In Proceedings of
the 11th Annual conference on Genetic and evolutionary
computation, pages 153–160. ACM, 2009.
Sebastian Risi, Charles E Hughes, and Kenneth O Stanley.
Evolving plastic neural networks with novelty search. Adaptive Behavior, 18(6):470–491, 2010.
Alan Roberts, Deborah Conte, Mike Hull, Robert MerrisonHort, Abul Kalam al Azad, Edgar Buhl, Roman Borisyuk,
and Stephen R Soffe. Can simple rules control development of a pioneer vertebrate neuronal network generating
behavior? Journal of Neuroscience, 34(2):608–621, 2014.
Anthony Robins. Catastrophic forgetting, rehearsal, and
pseudorehearsal. Connection Science: Journal of Neural
Computing, Artificial Intelligence and Cognitive Research,
7(123-146), 1995.
Derek Roff. Evolution of life histories: theory and analysis.
Springer Science & Business Media, 1993.
Edmund T Rolls and Simon M Stringer. On the design of
neural networks in the brain by genetic evolution. Progress
in Neurobiology, 61(6):557–579, 2000.
David E Rumelhart, Geoffrey E Hinton, and Ronald J
Williams. Learning representations by back-propagating
errors. Cognitive modeling, 5(3):1, 1988.
Stuart Russell and Peter Norvig. Artificial intelligence: a
modern approach. Pearson; 3 edition (5 Aug. 2013), 2013.
Scott J Russo, David M Dietz, Dani Dumitriu, John H Morrison, Robert C Malenka, and Eric J Nestler. The addicted
synapse: mechanisms of synaptic and structural plasticity
in nucleus accumbens. Trends in neurosciences, 33(6):267–
276, 2010.
Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan
Pascanu, and Raia Hadsell. Progressive Neural Networks.
arXiv preprint arXiv:1606.04671, 2016.
Eduardo Sanchez, Daniel Mange, Moshe Sipper, Marco
Tomassini, Andrés Pérez-Uribe, and André Stauffer. Phylogeny, ontogeny, and epigenesis: Three sources of biological inspiration for softening hardware. In International
Conference on Evolvable Systems, pages 33–54. Springer,
1996.
Jürgen Schmidhuber. Curious model-building control systems.
In Neural Networks, 1991. 1991 IEEE International Joint
Conference on, pages 1458–1463. IEEE, 1991.
Jürgen Schmidhuber. Developmental robotics, optimal artificial curiosity, creativity, music, and the fine arts. Connection
Science, 18(2):173–187, 2006.
Jürgen Schmidhuber. Deep learning in neural networks: An
overview. Neural Networks, 61:85–117, 2015.
Jürgen Schmidhuber, Daan Wierstra, Matteo Gagliolo, and
Faustino Gomez. Training recurrent networks by evolino.
Neural computation, 19(3):757–779, 2007.
Christiane Schreiweis, Ulrich Bornschein, Eric Burguière,
Cemil Kerimoglu, Sven Schreiter, Michael Dannemann,
Shubhi Goyal, Ellis Rea, Catherine A French, Rathi
Puliyadi, et al. Humanized Foxp2 accelerates learning
by enhancing transitions from declarative to procedural
performance. Proceedings of the National Academy of
Sciences, 111(39):14253–14258, 2014.
Henning Schroll and Fred H Hamker. Computational models
of basal-ganglia pathway functions: focus on functional
neuroanatomy. Frontiers in systems neuroscience, 7, 2015.
Wolfram Schultz. Predictive Reward Signal of Dopamine
Neurons. Journal of Neurophysiology, 80:1–27, 1998.
Wolfram Schultz, Paul Apicella, and Tomas Ljungberg. Responses of Monkey Dopamine Neurons to Reward and
Conditioned Stimuli during Successive Steps of Learning
a Delayed Response Task. The Journal of Neuroscience,
13:900–913, 1993.
Wolfram Schultz, Peter Dayan, and P. Read Montague. A
Neural Substrate for Prediction and Reward. Science, 275:
1593–1598, 1997.
Fernando Silva, Paulo Urbano, and Anders Lyhne Christensen.
Adaptation of robot behaviour through online evolution and
neuromodulated learning. In Ibero-American Conference on
Artificial Intelligence, pages 300–309. Springer, 2012a.
Fernando Silva, Paulo Urbano, and Anders Lyhne Christensen. Continuous adaptation of robot behaviour through
online evolution and neuromodulated learning. In Fifth
International Workshop on Evolutionary and Reinforcement
Learning for Autonomous Robot Systems (ERLARS 2012) in
conjunction with the 20th European Conference on Artificial
Intelligence (ECAI 2012). Montpellier. Citeseer, 2012b.
David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser,
Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot,
et al. Mastering the game of Go with deep neural networks
and tree search. Nature, 529(7587):484–489, 2016.
Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv
preprint arXiv:1409.1556, 2014.
Karl Sims. Artificial evolution for computer graphics, volume 25. ACM, 1991.
Karl Sims. Evolving 3D morphology and behavior by competition. Artificial life, 1(4):353–372, 1994.
M. Sipper, E. Sanchez, D. Mange, M. Tomassini, A. PerezUribe, and A. Stauffer. A phylogenetic, ontogenetic, and
epigenetic view of bio-inspired hardware system. Evolutionary Computation, IEEE Transactions on, 1(1):83–97, April
1997.
B. F. Skinner. Science and Human Behavior. New York,
MacMillan, 1953.
Burrhus Frederic Skinner. The behavior of organisms: An
experimental analysis. New York, London, D. AppletonCentury Company, Incorporated, 1938.
John Maynard Smith. When learning guides evolution. Nature,
329(6142):761–762, 1986.
23
T.M.C. Smith, P. Husbands, A. Philippides, and Michael
O’Shea. Neuronal Plasticity and Temporal Adaptivity:
GasNet Robot Control Networks. Adaptive Behavior, 10:
161–183, 2002.
Tom Smith. The evolvability of artificial neural networks for
robot control. PhD thesis, University of Sussex, October
2002.
Andrea Soltoggio. Neural Plasticity and Minimal Topologies
for Reward-based Learning Problems. In Proceeding of the
8th International Conference on Hybrid Intelligent Systems
(HIS2008), 10-12 September, Barcelona, Spain, 2008a.
Andrea Soltoggio. Neuromodulation Increases Decision Speed
in Dynamic Environments. In Proceedings of the 8th International Conference on Epigenetic Robotics, Southampton,
July 2008, 2008b.
Andrea Soltoggio. Evolutionary and Computational Advantages of Neuromodulated Plasticity. PhD thesis, School of
Computer Science, The University of Birmingham, 2008c.
Andrea Soltoggio. Short-term plasticity as cause-effect hypothesis testing in distal reward learing. Biological Cybernetics, 109(1):75–94, 2015.
Andrea Soltoggio and Ben Jones. Novelty of behaviour as a
basis for the neuro-evolution of operant reward learning. In
Proceedings of the 11th Annual conference on Genetic and
evolutionary computation, pages 169–176. ACM, 2009.
Andrea Soltoggio and Kenneth O. Stanley. From Modulated
Hebbian Plasticity to Simple Behavior Learning through
Noise and Weight Saturation. Neural Networks, 34:28–41,
October 2012.
Andrea Soltoggio, Peter Dürr, Claudio Mattiussi, and Dario
Floreano. Evolving Neuromodulatory Topologies for Reinforcement Learning-like Problems. In Proceedings of the
IEEE Congress on Evolutionary Computation, CEC 2007,
2007.
Andrea Soltoggio, John A. Bullinaria, Claudio Mattiussi, Peter Dürr, and Dario Floreano. Evolutionary Advantages
of Neuromodulated Plasticity in Dynamic, Reward-based
Scenarios. In Artificial Life XI: Proceedings of the Eleventh
International Conference on the Simulation and Synthesis
of Living Systems. MIT Press, 2008.
Olaf Sporns and William H. Alexander. Neuromodulation
in a learning robot: interactions between neural plasticity
and behavior. In Proceedings of the International Joint
Conference on Neural Networks, volume 4, pages 2789–
2794, 2003.
Nitish Srivastava, Geoffrey E Hinton, Alex Krizhevsky, Ilya
Sutskever, and Ruslan Salakhutdinov. Dropout: a simple
way to prevent neural networks from overfitting. Journal of
Machine Learning Research, 15(1):1929–1958, 2014.
J. E. R. Staddon. Adaptive Behaviour and Learning. Cambridge University Press, 1983.
Kenneth O Stanley. Compositional pattern producing networks: A novel abstraction of development. Genetic programming and evolvable machines, 8(2):131–162, 2007.
Kenneth O Stanley and Joel Lehman. Why Greatness Cannot
Be Planned. Springer Science Business Media., 10:978–3,
2015.
Kenneth O Stanley and Risto Miikkulainen. Evolving neural
networks through augmenting topologies. Evolutionary
computation, 10(2):99–127, 2002.
Kenneth O Stanley and Risto Miikkulainen. A taxonomy for
artificial embryogeny. Artificial Life, 9(2):93–130, 2003a.
Kenneth O. Stanley and Risto Miikkulainen. Evolving Adaptive Neural Networks with and without Adaptive Synapses.
In Bart Rylander, editor, Genetic and Evolutionary Computation Conference Late Breaking Papers, pages 275–282,
Chicago, USA, 12–16July 2003b.
Kenneth O Stanley, David B D’Ambrosio, and Jason Gauci.
A hypercube-based encoding for evolving large-scale neural
networks. Artificial life, 15(2):185–212, 2009.
Luc Steels. The artificial life roots of artificial intelligence.
Artificial life, 1(1 2):75–110, 1993.
James V Stone. Distributed representations accelerate evolution of adaptive behaviours. PLoS computational biology,
3(8):e147, 2007.
R. E. Suri. TD models of reward predictive responses in
dopamine neurons. Neural Networks, 15:523–533, 2002.
R. E. Suri, J. Bargas, and M. A. Arbib. Modeling functions
of striatal dopamine modulation in learning and planning.
Neuroscience, 103(1):65–85, 2001.
Richard S. Sutton and Andrew G. Barto. Reinforcement
Learning: An Introduction. MIT Press, Cambridge, MA,
USA, 1998.
Tim Taylor, Mark Bedau, Alastair Channon, David Ackley,
Wolfgang Banzhaf, Guillaume Beslon, Emily Dolson, Tom
Froese, Simon Hickinbotham, Takashi Ikegami, et al. Openended evolution: perspectives from the OEE workshop in
York. Artificial life, 2016.
Marc Tessier-Lavigne and Corey S Goodman. The molecular
biology of axon guidance. Science, 274(5290):1123, 1996.
Edward Lee Thorndike. Animal Intelligence. New York:
Macmillan, 1911.
Sebastian Thrun and Tom M Mitchell. Lifelong robot learning.
Robotics and autonomous systems, 15(1-2):25–46, 1995.
Sebastian Thrun and Joseph O’Sullivan. Discovering structure
in multiple learning tasks: The TC algorithm. In ICML,
volume 96, pages 489–497, 1996.
Sebastian Thrun and Lorien Pratt. Learning to learn. Springer
Science & Business Media, 2012.
Paul Tonelli and Jean-Baptiste Mouret. Using a map-based
encoding to evolve plastic neural networks. In Evolving and
Adaptive Intelligent Systems (EAIS), 2011 IEEE Workshop
on, pages 9–16. IEEE, 2011.
Paul Tonelli and Jean-Baptiste Mouret. On the relationships
between generative encodings, regularity, and learning abilities when evolving plastic artificial neural networks. PloS
one, 8(11):e79138, 2013.
J. Urzelai and D Floreano. Evolutionary Robotics: Coping
with Environmental Change. In Genetic and Evolutionary
Computation Conference (GECCO’2000), 2000.
Roby Velez and Jeff Clune. Diffusion-based neuromodulation can eliminate catastrophic forgetting in simple neural
networks. CoRR, abs/1705.07241, 2017.
24
Santosh S Venkatesh. Directed drift: A new linear threshold
algorithm for learning binary weights on-line. Journal of
Computer and System Sciences, 46(2):198–217, 1993.
Julien Vitay and Fred H. Hamker. A computational model of
basal ganglia and its role in memory retrieval in rewarded
visual memory tasks. Frontiers in Computational Neuroscience, 2010.
Günter P Wagner and Lee Altenberg. Perspective: complex
adaptations and the evolution of evolvability. Evolution, 50
(3):967–976, 1996.
Edgar T. Walters and John H. Byrne. Associative Conditioning
of Single Sensory Neurons Suggests a Cellular Mechanism
for Learning. Science, 219:405–408, 1983.
Bernard Widrow and Michael A Lehr. 30 years of adaptive
neural networks: perceptron, madaline, and backpropagation. Proceedings of the IEEE, 78(9):1415–1442, 1990.
Bernard Widrow, Marcian E Hoff, et al. Adaptive switching
circuits. In IRE WESCON convention record, volume 4,
pages 96–104. New York, 1960.
David Willshaw and Peter Dayan. Optimal plasticity from
matrix memories: What goes up must come down. Neural
Computation, 2(1):85–93, 1990.
Brian M. Yamauchi and Randall D. Beer. Sequential Behavior
and Learning in Evolved Dynamical Neural Networks.
Adaptive Behavior, 2(3), 1994.
Xin Yao. Evolving Artificial Neural Networks. Evolutionary Computation, IEEE Transactions on, 87(9):1423–1447,
September 1999.
Xin Yao and Yong Liu. A new evolutionary system for
evolving artificial neural networks. IEEE transactions on
neural networks, 8(3):694–713, 1997.
Jason Yoder and Larry Yaeger. Evaluating Topological Models
of Neuromodulation in Polyworld. In Artificial Life XIV,
MIT Press, Cambridge, MA, 2014.
Steven R Young, Derek C Rose, Thomas P Karnowski, SeungHwan Lim, and Robert M Patton. Optimizing deep learning
hyper-parameters through an evolutionary algorithm. In
Proceedings of the Workshop on Machine Learning in HighPerformance Computing Environments, page 4. ACM, 2015.
Tom Ziemke and Mikael Thieme.
Neuromodulation of
Reactive Sensorimotor Mappings as Short-Term Memory
Mechanism in Delayed Response Tasks. Adaptive Behavior,
10:185–199, 2002.
| 2cs.AI |
Post-Prognostics Decision for Optimizing the Commitment of Fuel
Cell Systems
Stéphane Chrétiena,1,∗, Nathalie Herrb , Jean-Marc Nicodb , Christophe Varnierb
a
arXiv:1710.08812v1 [cs.CE] 19 Oct 2017
b
National Physical Laboratory
FEMTO-ST/AS2M, Univ. Bourgogne Franche-Comté/UFC/CNRS/ENSMM,Besan con, France
Abstract
In a post-prognostics decision context, this paper addresses the problem of maximizing the
useful life of a platform composed of several parallel machines under service constraint. Application on multi-stack fuel cell systems is considered. In order to propose a solution to
the insufficient durability of fuel cells, the purpose is to define a commitment strategy by
determining at each time the contribution of each fuel cell stack to the global output so as
to satisfy the demand as long as possible. A relaxed version of the problem is introduced,
which makes it potentially solvable for very large instances. Results based on computational
experiments illustrate the efficiency of the new approach, based on the Mirror Prox algorithm, when compared with a simple method of successive projections onto the constraint
sets associated with the problem.
Keywords: Decision making, Post-prognostics decision, PHM, Fuel cell, Convex
optimization example, LATEX, template
1. Introduction and related work
In the context of the decline of fossil fuel resources, the use of fuel cells appears to be
of growing interest as a potential alternative to conventional power systems [10]. Fuel cells
can be used in many applications, such as stationary ones for domestic use, but also in
transportation and portable power applications [4]. Unfortunately, fuel cells suffer from
insufficient durability. Indeed, the current lifetime is usually between 1 500 and 3 000 hours,
whereas 5 000 hours are often required for transportation applications and up to 100 000
hours for stationary ones. Improving the performance, reliability and lifetime is therefore
an important challenge [4] in the fuel cell technology, for which techniques of Prognostics
and Health Management (PHM ) can definitely help. As recently pointed out in Jouin et al.
∗
Corresponding author
Email addresses: stephane.chretien@npl.co.uk (Stéphane Chrétien), nathalie.herr@femto-st.fr
(Nathalie Herr), jean-marc.nicod@femto-st.fr (Jean-Marc Nicod), christophe.varnier@femto-st.fr
(Christophe Varnier)
URL: https://sites.google.com/view/stephanechretien/home (Stéphane Chrétien)
Preprint submitted to Computers & Operations research
October 25, 2017
[10], research in fuel cells with a PHM viewpoint has mainly focused on data acquisition
and data processing. Less work has been done on condition assessment and diagnostics
and very few papers addressed prognostics and decision making. Works that take into
account the decision part often only proposed corrective actions (see [5] and [14]), for which
physical parameters (such as inlet and outlet gas flows, pressures and temperatures, single
cell and stacks voltages or current) are controlled in order to master each fuel cell operating
conditions as accurately as possible. These corrective actions correspond to real-time control
(from nanoseconds to seconds), necessary to compensate the natural fluctuation of fuel cells
parameters and to avoid too early irreversible degradation. At each time slot the strategy
allows to set the operating current to a value that meets the needs in power for each fuel
cell.
Decision making as addressed in this paper differs from the previous studies in many
aspects. In particular, larger time scales (hours to weeks) are considered and decision comes
within the framework of Prognostic Decision Making (PDM), which aims at optimizing
systems configuration [1]. The system considered here consists of several fuel cells, used in
parallel in order to provide a global power output. The problem we consider is to provide
the power output value for each fuel cell as a function of time, so as to meet the global power
demand. The target application considered here is based on stationary power generation for
domestic usage, also known as micro combined heat and power (micro-CHP).
In order to deliver suitable power outputs, fuel cells are used in the form of stacks,
composed of many connected single cells. Each stack is assumed to be independent of the
others, but the multi-stack fuel cell system has to deliver a given global power output. At
each time slot, the total provided power output is the sum of each output of the stacks that
are switched on. Each fuel cell stack is able to deliver an output that can vary continuously
within a given interval. The optimization problem consists in determining the appropriate
output for each fuel cell stack during the whole production horizon. Also the stacks are
not constrained to be all running at each time slot if the target output can be reached by
using only a subset of them. All the stacks may moreover not be available at all times:
if their end of life has been reached, they are not available for production. Considering
a global power output, the multi-stack system useful life depends not only on each stack
useful life, but also on both the schedule and the operating condition settings that define
the contribution of each stack over time. The same statement applies to batteries in a
health management context. Saha et al. [13] have for instance addressed the maximization
of the battery charge used while constraining the probability of a battery shut off in flight for
electric unmanned aerial vehicles. Predictions on remaining battery life are used to optimize
mission plans without exceeding the available battery charge. In a same way, we propose to
use prognostics results in the form of Remaining Useful Lives (RUL) to maximize the global
useful life of a multi-stack fuel cell system subject to service constraint.
A similar problem has been addressed in [7] and [8], where the purpose was to define
a schedule of machines that maximizes the production horizon, based on the knowledge of
each machine remaining useful life (RUL) in a PHM framework. In these studies, machine
throughputs have been assumed to take discrete values. It was shown in [8] that optimal
solutions can be found in reasonable time only for small size instances and with a very limited
2
number of machines, very few possible throughput values and short production horizons.
An other study considering this time machines whose performances can vary continuously
between two bounds has been proposed in [9]. The considered model has been built to fit
the fuel cells behavior, but the proposed resolution approach gives suboptimal solutions and
is limited to systems of reasonable size.
In order to overcome these two limitations, we propose in this paper to drastically change
the solution paradigm and to build the solutions globally on the full production horizon.
Contribution of each machine during its lifetime is considered as a whole and optimized
on the full horizon. Each machine contribution is determined via convex optimization. As
a result, our new approach allows to address large problems very quickly. Concretely, the
considered scheduling problem is addressed via optimizing a composite function subject to
several constraints due to fuel cell intrinsic characteristics. A first method is described
which performs successive projections onto the sets of constraints. The method generates
a sequence of points that can be shown to converge to the solution of the optimization
problem [2]. A second order method is developed and used to define the contribution of
each fuel cell stack to the global output over the whole production horizon. It is based on
the Mirror Prox method proposed by Nemirovski [11] as a variant of the Mirror Descent
developed by Nemirovsky and Yudin [12] to minimize a smooth convex function subject to
convex constraints. Estimation of the variable is efficient in that it depends very little on its
dimension. This is why these methods can be used to solve big optimization problems [3].
The organization of the paper is as follows: the tackled problem is first described in
Section 2, with a presentation of the application framework and the optimization problem,
followed by a mathematical formulation that complies with the proposed convex resolution
paradigm. The two proposed resolution approaches are then described in Sections 3 and 4.
Efficiency of these methods is assessed through simulation results in Section 5. Conclusion
and future work are finally given in Section 6.
2. Problem statement
2.1. Application framework
The application addressed in this paper is based on a multi-stack fuel cell system which
is supposed to meet energy requirements for domestic usage in a stationary power generation
framework. This system is supposed to be composed of m fuel cell stacks Mj (1 6 j 6 m).
All the stacks are supposed to be always supplied with raw material required for the energy
conversion. They can be used simultaneously and independently from each other.
This corresponds to a parallel machines system, in which each machine is supposed to
be able to deliver power outputs Pj that can vary continuously within a given power output
range [P minj ; P maxj ]. For each machine Mj (1 6 j 6 m), the range of available power
outputs depends on the time t and evolves as depicted in Figure 1, with the following extreme
output characteristics:
P maxj : the maximal power output which decreases with time when the machine Mj is used.
The straight line depicting its decrease has equation P maxj (t) = aj · t+P maxj (0), with
3
aj < 0 the speed of the output decline and P maxj (0) the maximal output available
at the beginning of the scheduling process. Both parameters of this equation are fuel
cells intrinsic characteristics. Useful life of each power output Pj , denoted RULj (Pj ),
is limited by this maximal power output decrease ;
P minj : the minimal power output, constant over time and associated to the maximal Remaining Useful Life: RULj (P minj ) = maxP minj 6Pj 6P maxj (RULj (Pj )) = RULmaxj .
Pj (t)
P maxj (0)
P maxj (0)
100%
P maxj (t) = aj · t + P maxj (0)
Pj
Decrease
of P maxj
P minj
0
15%
RULmaxj
RULj (Pj )
time
(End of life)
Figure 1: Evolution of available power outputs for a machine Mj
The model proposed in Figure 1 is a simplified version of the one developed in [9]. It
considers however the main characteristics of fuel cells useful for the scheduling purpose
considered here. Due to the decrease over time of the maximal power output P maxj (t), this
model is asymmetric. In a schedule defining the contribution over time of a machine, two
periods of time in which a machine is used with different outputs can then not always be
permuted (see [9] for an illustration of this property).
2.2. Optimization problem
At each time t (0 6 t 6 T ), the global outcome P tot is the sum of each stack power
contribution. During the whole production horizon, this global outcome has to reach a
given load demand σ(t). Overproduction is authorized if it allows to extend the global
system useful life (or production horizon), but should be avoided as far as possible. In
fact, storage being not considered in this study, overproduction is supposed to be lost. All
the machines are not supposed to be in use at each time if a subset of them is enough to
reach the demand. Since machines suffer from wear and tear, some machines can also be
unavailable at a certain time if their end of life (EOL) has been reached. Stop-and-start of
fuel cell stacks have moreover to be avoided as far as possible. Stopping and restarting a
fuel cell can indeed induce considerable damage and lead to premature aging [4]. Change of
power output during the use of fuel cell stacks is however still authorized.
Considering these assumptions, the point is to manage the system by defining the commitment of fuel cell stacks so as to reach the demand as long as possible. During the whole
production horizon, the purpose is then to define at each time each stack contribution to
the global power output.
4
2.3. Mathematical formulation
In this section, a mathematical formulation of the problem previously described is defined, making use of convex elements. Let fj (t) (1 6 j 6 m, 0 6 t 6 T ) be the vector
defining the evolution over time of the power output delivered by the machine Mj , with
T the length of the decision horizon. Link between this decision horizon and the solution
production horizon, denoted H, is clarified in Section 5.3. Contributions of all the machines
are gathered together in a vector F ∈ Rm(T +1) such that:
F = [f1 (0), . . . , f1 (T ), . . . , fj (t), . . . fm (0), . . . , fm (T )] .
The general idea is to minimize a convex function subject to a set of constraints. The
objective function aims at ensuring that the power demand is reached. At each time t
(0 6 t 6 T ), it is about minimizing the difference between the global power output delivered
by the set of machines and the demand σ(t). This is expressed by Equation (1), where φ
measures the error incurred by the choice of fj , ∀ 1 6 j 6 m. One possible choice of φ is
the squared loss, i.e. the squared L2 norm.
!
m
X
min φ σ(·) −
fj (·)
(1)
j=1
Constraints on each function fj relate to the definition domain of each contribution and
to the limited availability of machines. At each time t, each machine contribution is either
equal to zero or constrained between two bounds (see Equation (2)), in accordance with the
hypotheses detailed in the application framework. fj (t) = 0 means that the machine Mj is
not used at time t.
fj (t) = 0 or fminj (t) 6 fj (t) 6 fmaxj (t)
(2)
∀ 1 6 j 6 m, ∀ 0 6 t 6 T
Each contribution fj is constrained by the maximal power output decrease of the associated
machine Mj , which expresses its limited availability. Evolution over time of this maximal
power output, fmaxj (t), is a function of the use of machine Mj , fj (t). Indeed, fmaxj (t)
evolves only if Mj is used, that is, if fj (t) > 0. A first formulation is proposed in Equation (3),
with aj the speed associated to the maximal power output decrease (aj < 0).
fmaxj (t − 1) + aj if fj (t) > 0 ;
fmaxj (t) =
(3)
fmaxj (t − 1)
if fj (t) = 0
∀ 1 6 j 6 m, ∀ 1 6 t 6 T
Equations (2) and (3) being not convex, they can not be used as is within the proposed
convex programming paradigm. A second formulation of the constraints is proposed in set
of equations (4), which details the mathematical program associated to the optimization
problem. This program does not respect the real evolution over time of the maximal power
output that can be delivered by machines, but presents the advantages of being convex
5
and thus consistent with the convex resolution methods proposed in next section. In the
following, the machines behavior follows the simplified model depicted in Figure 2.
!
m
X
min φ σ(·) −
fj (·)
(4a)
j=1
s.t. fmaxj (t) 6 fmaxj (t − 1)
+µ · aj · (fj (t − 1))υ ∀ 1 6 j 6 m,
(4b)
∀ 1 6 t 6 T, with µ and υ ∈ [0, 1]
with 0 6 fj (t) 6 fmaxj (t)
(4c)
∀ 1 6 j 6 m, ∀ 0 6 t 6 T
Pj (t)
P maxj (0)
P maxj (0)
100%
fmaxj (t) = fmaxj (t − 1) + aj
Decrease
of P maxj
P minj
15%
0
RULmaxj
time
Figure 2: Simplified convex evolution of available power outputs for a machine Mj
3. Resolution based on successive projections onto the sets of constraints
The first proposed resolution approach makes use of convex projections to cope with the
scheduling problem. The general idea is to perform successive projections onto each set of
constraints previously defined in the convex program (4) in order to generate a sequence
of points that converges to the solution of the considered optimization problem. The three
projections are first detailed. The global projection scheme is then described.
3.1. Projections
Each projection allows to comply with one of the three main constraints defined by
Equations (4a), (4b) and (4c).
3.1.1. Reaching of the demand
First constraint corresponds to the reaching of the demand σ(t) at each time t (0 6 t 6
T ). The decision horizon is split into time intervals of length ∆T , with 1 6 ∆T 6 T . For
each time interval, delimited by the times tstart and tend such that tstart 6 t 6 tend, the
projection of the solution F on the demand σ is performed following the strategy detailed
hereafter:
6
Machine selection: machines Mj (1 6 j 6 m) are sorted following an ascending order
of the difference (fmaxj (t) − fj (t)). This sorting favors machines whose contributions
before the projection are the closest from their maximal reachable output fmaxj . This
allows to use first machines already started and to avoid stop-and-start of fuel cells,
which has been shown to lead to premature aging [4].
With distfmax(j) = max(fmaxj (tend) − fj (tend), 0) being the difference between the
current contribution of machine Mj and its maximal reachable contribution, the subscript of the machine Mj 0 whose contribution has to be modified first is determined by
Equation (5).
j0 =
argmin
distfmax(j)
(5)
16j6m|distfmax(j)>0
Projection method: the contribution of the selected machine Mj 0 is increased on each
time interval of length
Pm ∆T just enough to reach the demand σ. The increase being
inc = σ(tend) − j=1 fj (tend), the contribution of machine Mj 0 after the projection
is defined by Equation (6).
fj 0 (t) = min (fj 0 (tend) + inc, fmaxj 0 (tend))
∀ tstart 6 t 6 tend
(6)
The selection of a machine and the projection of its contribution on the demand are
performed on each time interval of length ∆T , until the demand is reached or until there is
no machine available anymore.
3.1.2. Evolution of the maximal output
Each maximal output reachable fmaxj is updated as a function of the contribution of
the corresponding machine, fj , following the relation defined in Equation (7).
fmaxj (t) = fmaxj (t − 1) + µ · aj · (fj (t − 1))υ
∀ 1 6 j 6 m, ∀ 1 6 t 6 T, with µ and υ ∈ [0, 1]
(7)
3.1.3. Respect of the maximal output
A projection of each machine contribution fj onto the corresponding fmaxj can be necessary to fix some possible overrun and to comply with the updated maximal reachable
output. This projection simply limits each contribution fj (t) to the corresponding fmaxj (t)
defined by the application of the previous projection.
fj (t) = fmaxj (t) if fj (t) > fmaxj (t)
∀ 1 6 j 6 m, ∀ 0 6 t 6 T
(8)
3.2. Resolution algorithm based on the projections
The proposed resolution scheme performs successive projections on the three main constraints, following the strategies previously defined. The positivity of each component of the
7
solution vector F is ensured through a positive initialization of each machine contribution
(0)
at each time (fj (t) > 0 ∀ 1 6 j 6 m, ∀ 0 6 t 6 T ).
The evolution of the vectors fmaxj being determined as a function of each fj , a combined
determination of these two elements is required and several successive launches of all the
projections are necessary to optimize the solution. The definition of a schedule that maximizes the production horizon while complying with all the constraints requires then several
iterations of the sequence composed of the three successive projections. This sequence is
iterated until a stopping condition is verified. The number of iterations necessary to obtain
a satisfying solution depends indeed strongly on the configuration of the considered problem, that is, on the number of machines, on the shape of the maximal reachable outputs
fmaxj and on the demand σ. The optimization process of the solution F is then stopped
when it does not change significantly anymore, that is, as soon as the difference between
two successive values of F (|F − F prev |) gets below a certain threshold .
Solutions obtained with this first resolution method are piecewise defined and the reached
production horizons depend strongly on the initialization of the solution vector F and on
the projection strategy. This first resolution approach does then not fully comply with the
resolution paradigm proposed in this paper, which aims to build the solutions globally on the
full production horizon. It allows however to obtain solutions very quickly, which can serve
as a point of comparison with the solutions obtained with the resolution method proposed
in next section.
4. Resolution based on a smooth penalization approach using the Mirror Prox
The second resolution method is based on a smooth penalisation approach. A Mirror
Prox method is proposed to be used to cope with the problem of minimizing the objective
function detailed previously in Equation (4a).
4.1. Principle of the Mirror-Prox
The Mirror Prox algorithm, developed by Nemirovski [11], is a variant of the Mirror
Descent algorithm, which has first been proposed by Nemirovsky and Yudin [12] for convex
programming. It has been extensively studied recently and several relationships have been
discovered between the Mirror Descent scheme and Bregman-proximal methods.
Both approaches are based on the resolution of a primal-dual saddle point problem, which
allows to take constraints into account. The purpose is to minimize a smooth convex function
under constraints. The Mirror Descent algorithm makes use of a gradient descent to find the
minimum of the considered function. The goal is to minimize the local linearization of the
function while not moving too far away from the previous point, with distances measured
via the Bregman divergence of the mirror map. A mirror function allows to transition from
the primal space, where all the constraints of the problem are defined, to the dual space.
The Mirror Prox method applies at each iteration two consecutive steps of Mirror Descent.
A very instructive description of the Mirror Prox algorithm has been proposed by Bubeck
[6]. As illustrated in Figure 3, the Mirror Prox algorithm applies a first time the Mirror
Descent to go from xt to yt+1 and then a second similar step to obtain xt+1 . This second
8
step stats again from xt , but uses the gradient of f evaluated at yt+1 (instead of xt ). These
steps are defined by Equations (9) to (12).
∇θ
∇θ(xt )
−λ∇φ(xt )
−λ∇φ(yt+1 )
∇θ(x0t+1 )
Gradient steps
Eq. (9) and (11)
0
)
Rn ∇θ(yt+1
x0t+1
(∇θ)−1
Dual space
xt
xt+1
yt+1
Projection steps
0
(10) and (12)
yEq.
t+1
D
Primal space
Figure 3: Operating principle of the Mirror Prox method, based on [6]
0
∇θ(yt+1
) = ∇θ(xt ) − λ∇φ(xt )
with λ > 0
0
0
yt+1 = PC (∇θ(yt+1
)) = argmin Dθ (x, yt+1
)
(9)
(10)
x∈C∩D
∇θ(x0t+1 ) = ∇θ(xt ) − λ∇φ(yt+1 )
with λ > 0
xt+1 = PC (∇θ(x0t+1 )) = argmin Dθ (x, x0t+1 )
(11)
(12)
x∈C∩D
4.2. Resolution algorithm based on the Mirror Prox
The proposed resolution algorithm makes use of the Mirror Prox scheme previously described. The resolution of the primal-dual problem being highly sensitive to the variation of
the different parameters associated to the model, the proposed formulation incorporates the
different constraints of the application directly into the objective function. The considered
objective function φ(F, fmax, σ), detailed in Equation (13), aims then at satisfying both the
reaching of the demand and the respect of the evolution of the maximal outputs fmaxj (t),
∀ 1 6 j 6 m and ∀ 0 6 t 6 T .
φ(F, fmax, σ) =λdem hdem (F, σ) + λslope hslope (F, fmax)
with λdem , λslope > 0
(13)
The first component of this objective function, hdem (F, σ), detailed in Equation (14),
aims at satisfying the demand σ(t) at each time t. We did not choose the L2 loss because
asymetric loss are slightly more relevant in our context. In fact, finding a solution which
does not match the demand should be more heavily penalised than a solution which matches
the demand by far. This cost function, based on the objective function previously defined in
Equation (4a), allows to heavily penalize F when the demand becomes larger than the total
output provided by the set of machines and conversely to penalize it only slightly when the
production is larger than the demand. Second part of the objective function, hslope (F, fmax),
is used to control the evolution of each fmaxj as a function of each associated contribution
9
fj (see equation (15)). This corresponds to the penalization associated to the constraint
previously defined by Equation (4b).
X
T
m
X
1
hdem (F, σ) =
exp − γ
fj (t) − σ(t)
T +1
t=0
j=1
with γ > 0
hslope (F, fmax) =
T X
m
X
(14)
exp δ fmaxj (t) − fmaxj (t − 1)
t=1 j=1
υ0
0
− µ aj (fj (t − 1))
(15)
with δ > 0, µ0 > 0, υ 0 > 1
The mirror function θ considered in the Mirror Prox formula is defined on Rm(T +1) by
Equation (16) and its gradient, which is used in the Mirror Descent steps, is expressed in
Equation (17).
θ(F ) =
T X
m
X
F ln(F )
(16)
t=0 j=1
∇θ(F ) = ln(F ) + 1
(17)
The Mirror Prox used in the proposed resolution algorithm is then of the following form
(Equations (18) to (21)):
∇θ(F (l+1) ) =∇θ(F (l) ) − λ∇F φ(F (l) , fmax, σ)
with λ > 0
(18)
Fint = exp ∇θ(F (l+1) ) − 1
(19)
∇θ(F (l+1) ) =∇θ(F (l) ) − λ∇Fint φ(F (l) , fmax, σ)
with λ > 0
(20)
F (l+1) = exp ∇θ(F (l+1) ) − 1
(21)
with the gradient of φ defined by Equations (22) to (24).
∇F φ(F (l) , fmax, σ) =λdem ∇F hdem (F, σ)
+ λslope ∇F hslope (F, fmax)
with λdem , λslope > 0
10
(22)
T
X
1
·
T +1
t=0
X
m
exp − γ
fj (t) − σ(t)
∇F hdem (F, σ) = − γ
(23)
j=1
with γ > 0
0 0
∇F hslope (F, fmax) = −δµ υ
T X
m
X
0
aj · fj (t − 1)υ −1 ·
t=0 j=1
0
υ0
exp δ fmaxj (t) − fmaxj (t − 1) − µ aj (fj (t − 1))
with δ > 0, µ0 > 0, υ 0 > 1
(24)
In order to accelerate the convergence of the method, the distance of the solution to
its fictional projection onto the demand σ(t), (F − F proj ), is proposed to be added to each
gradient step performed by the Mirror Descent steps. This allows to accelerate the evolution
of the solution towards a state which complies at best with the different constraints. The
new formulation of the Mirror Prox is then detailed in Equations (25) to (28), with the
vector F proj determined by a projection of the solution vector F onto the demand σ(t), with
gF proj = wgrad (F (l) − F proj ) and wgrad > 0.
∇θ(F (l+1) ) =∇θ(F (l) )
− λ ∇F φ(F (l) , fmax, σ) + gF proj
with λ > 0
Fint = exp ∇θ(F (l+1) ) − 1
(25)
(26)
∇θ(F (l+1) ) =∇θ(F (l) )
− λ ∇Fint φ(F (l) , fmax, σ) + gF proj
with λ > 0
F (l+1) = exp ∇θ(F (l+1) ) − 1
(27)
(28)
The use of this additional gradient step allows furthermore to define a stopping condition
for the global process, which allows to avoid the definition of a number of iterations adapted
to the method and to the optimization problem parameters. The resolution process is
stopped when the differences between all the components of the solution vector F and the
corresponding components of the vector F proj are all lower than a certain threshold. In
11
other words, the resolution process is stopped when the reaching of the demand can not
be improved anymore. After each step of the Mirror Prox, several projections onto the
constraints are moreover performed in order to guarantee the respect of these constraints
and to improve the convergence speed of the solution. An update of the maximal power
outputs reachable, fmaxj , as a function of each vector fj , allows to accelerate their evolution
even if their progression is handled by the Mirror Prox. A projection of the machines
contributions fj onto the corresponding fmaxj allows finally to comply with the maximal
power output limitation.
5. Simulation results
5.1. Problem generation
Random problem configurations have been generated using a simulator and configured
with many parameters including the number of machines in the considered multi-stack fuel
cell system, m, and intrinsic fuel cell characteristics. The latter have been defined on the
basis of fuel cell manufacturer specifications and considering a maximal lifetime RULmaxj =
1500 hours ±20% for each machine Mj (1 6 j 6 m). Each RULmaxj value is drawn following
a uniform distribution between 1200 and 1800 hours. Power output values are determined
in the same way, with P maxj (0) = 500 W ± 5% and P minj = 0.15 · P maxj (0) for each
machine Mj .
For the results presented hereafter, the power demand has been assumed to be constant
during the whole scheduling horizon: σ(t) = σ ∀ 0 6 t 6 T . Without any loss of generality,
only one demand value has then been associated to each problem configuration, but many
demands corresponding to different configurations have been tested. Many loads α have been
defined such that σ = α · P nomtot , with P nomtot the nominal total powerPoutput reachable
with the considered fuel cells system and 30% 6 α 6 90%. P nomtot = m
j=1 P nomj , with
P nomj = 0.75 · P maxj (0) the power output recommended by fuel cell manufacturers for a
nominal use of fuel cells. In the following figure, results are represented as a function of the
load α.
5.2. Resolution methods configuration
Initial values of each solution vector F has first been set to zero: fj (t) = 0 ∀ 1 6 j 6 m,
∀ 0 6 t 6 T . Quality of solutions from the point of view of the reached production horizon
globally increases with the number of iterations and stabilizes starting from a certain value.
For each resolution method proposed earlier, several iterations of the associated process are
then performed to optimize this solution. The global process is stopped when the variation
of the solution vector is not significant anymore, that is when (F − F proj ) < for the
Mirror-Prox-based algorithm (resp. (F − F prev ) < for the algorithm based on the
successive projections), with the threshold defined as a function of the demand value as
follows: = 0.1 · σ.
Tuning of the different parameters involved in the two resolution methods allows to
comply with the constraints and to adapt the shape of each fmaxj . For the evolution of
the maximal power output fmax, values have been defined as follows: µ = 0.2, υ = 0.3.
12
Values for the parameters used in the Mirror-Prox algorithm are the following: λ = 5 · 10−5 ,
λdem = λslope = 100, γ = 100, δ = 100, µ0 = 1, υ 0 = 1.
5.3. Post-processing
The main constraint of the optimization problem is the reaching of the power output demand σ(t). This constraint being tackled through the minimization of an objective function,
solutions may contain time periods during which this demand is not reached. But, with the
two resolution methods proposed in previous section, solutions are built so that the time
periods for which the power demand σ(t) is reached are gathered at the beginning of the
schedules. This is consistent with the objective to maximize the production horizon of the
set of machines. This behavior is linked with the shape of the functions fmaxj (t) representing for each machine Mj the evolution over time of the maximal power output reachable,
which limits the contribution of each machine. These functions being strictly decreasing
with the use of machines, it is in fact more likely to reach the demand at the beginning of
the scheduling process than after some time. As already mentioned, resolution algorithms
detailed previously can then be applied on overestimated horizons T , named decision horizons. The production horizon of each solution, H, is simply the maximal time during which
all the constraints are strictly satisfied. In practice, the production horizon corresponds to
the time during which the demand σ(t) is reached.
5.4. Results
Efficiency of the proposed commitment strategies defined in Sections 3 and 4 is assessed
through the comparison of reached production horizons to a theoretical upper bound. Considering a constant demand σ and a set of fuel cell stacks, this upper bound, denoted UB
and defined in Equation (29), corresponds to the theoretical maximal time during which the
demand can be reached.
P
m
0.6 · P maxj (0) · RULj (P minj )
j=1
(29)
UB =
σ
Figure 4 shows the production horizons obtained with the considered strategies normalized with the upper bound, when considering a set of 25 fuel cell stacks. This upper bound
being never reachable, results are actually better than showed. One can see that the first
strategy, performing successive convex projections, allows to reach a mean relative horizon
of around 39% of the upper bound UB and does not exceed 51, 6%. The resolution method
based on the Mirror Prox gives better results. It allows indeed to reach a mean relative
horizon of around 64.3% of UB and 74, 7% for the best case.
The two resolution approaches differentiate also themselves by the shape of the solutions
and by their computation time. Figures 5 and 6 show schedules obtained respectively with
the algorithm based on the successive projections and with the one based on the Mirror
Prox, when considering m = 3 machines. Evolution of each machine contribution and
13
Normalized horizon (%)
100
75
Resolution method
Projections
50
Mirror Prox
UB
25
0
30
40
50
60
70
80
90
load (%)
Figure 4: Normalized horizon – m = 25 machines
of the associated maximal power output reachable, fmaxj , is also shown in these figures.
One can see in Figure 6 that the algorithm based on the Mirror Prox allows to reach a
better production horizon (H = 1745 periods of time) than the one performing successive
projections onto the sets of constraints (H = 1147). The Mirror Prox-based algorithm
defines also a smooth use of machines, which complies with a continuous use, without sudden
change of output such as those that can be seen in the solutions obtained with the successive
projections (see Figure 5).
The method performing successive projections onto the sets of constraints is not efficient
in terms of reached production horizons, but allows to obtain solutions in very limited time
(less than 9 seconds1 for all the scenarios tested when considering 25 machines). The time
needed to obtain satisfying solutions with the Mirror-Prox-based algorithm is longer (5.6
minutes on average), but does not exceed 17 minutes1 for 25 machines (see Figure 7) and
35 minutes1 for 100 machines, for all the power demands.
6. Conclusion
A management of fuel cell systems has been proposed in a PHM framework. Decision
coming within the scope of Prognostic Decision Making has been addressed considering
longer time frames than those proposed so far in the literature on fuel cells. The use of
1
Simulations have been made using Matlab (Parameters:
CPU@3.30GHz×4, 15.6 Gio, 64 bits)
14
Processeur Intelr CoreT M i5-3550
convex programming has been proposed to cope with the scheduling problem of multi-stack
fuel cell systems under service constraint. A mathematical formulation of the problem has
been proposed as well as two different convex resolution methods performing a minimization
of the objective function under constraints. First one is based on successive projections onto
the sets of constraints and second one on the Mirror-Prox algorithm.
All the fuel cell properties are not observed by the solutions obtained with the proposed
approaches, but this first study is promising. It shows indeed that a global resolution on the
scale of the whole production horizon can be used to define the commitment of machines
over time with the production horizon maximization as objective.
Acknowledgements: This work has been supported by the Labex ACTION project
(contract “ANR-11-LABX-0001-01”)
References
[1] Balaban E and Alonso JJ (2012) An approach to prognostic decision making in the aerospace domain.
In: Annual Conference of the Prognostics and Health Management Society.
[2] Bauschke HH and Borwein JM (1996) On projection algorithms for solving convex feasibility problems.
Society for Industrial and Applied Mathematics (SIAM) Review 38 (3): 367–426.
[3] Beck A and Teboulle M (2003) Mirror descent and nonlinear projected subgradient methods for convex
optimization. Operations Research Letters 31: 267–175.
[4] Borup R, Meyers J, Pivovar B, Inaba M, Ota K, Ogumi Z and et al (2007) Scientific aspects of polymer
electrolyte fuel cell durability and degradation. Chemical reviews 107(10): 3904–3951.
[5] Bosco AD and Fronk MH (2000) Fuel cell flooding detection and correction. US Patent nb. 6.103.409.
[6] Bubeck S (2014) Theory of Convex Optimization for Machine Learning. arXiv preprint arXiv:1405.4980.
[7] Herr N, Nicod JM and Varnier C (2014) Prognostic decision making to extend a platform useful life
under service constraint. In: Proceedings of IEEE International Conference on Prognostics and Health
Management (PHM). Spokane, Washington.
[8] Herr N, Nicod JM and Varnier C (2014) Prognostics-based scheduling in a distributed platform: Model,
complexity and resolution. In: 2014 IEEE International Conference on Automation Science and Engineering (CASE), Taipei, Taiwan. pp. 1054–1059.
[9] Herr N, Nicod JM, Varnier C, Jardin L, Sorrentino A, Hissel D and Péra MC (2017) Decision process
to manage useful life of multi-stacks fuel cell systems under service constraint. Renewable Energy 105C:
590–600. DOI:10.1016/j.renene.2017.01.001.
[10] Jouin M, Gouriveau R, Hissel D, Péra MC and Zerhouni N (2013) Prognostics and health management
of PEMFC - State of the art and remaining challenges. Int. Journal of Hydrogen Energy 38: 15307–
15317.
[11] Nemirovski AS (2004) Prox-method with rate of convergence O(1/t) for variational inequalities with
lipschitz continuous monotone operators and smooth convex-concave saddle point problems. SIAM
Journal on Optimization 15 (1): 229–251.
[12] Nemirovsky AS and Yudin DB (1983) Problem complexity and method efficiency in optimization. Wiley
- Interscience Series in Discrete Mathematics.
[13] Saha B, Quach CC and Goebel K (2012) Optimizing battery life for electric UAVs using a bayesian
framework. In: Proc. of IEEE Aerospace Conference, Big Sky, MT, USA.
[14] Wells B and Parr RK (2004) Fuel cell system method, apparatus and scheduling. US Patent nb. 0185316
A1.
15
600
600
fmax1(t)
fmax2(t)
400
200
0
0
M2 contribution (f2(t))
Output P(t)
Output P(t)
M1 contribution (f1(t))
1000
2000
Time (hours)
600
400
200
0
0
3000
1000
2000
Time (hours)
3000
fmax (t)
3
Output P(t)
M3 contribution (f3(t))
400
200
0
0
1000
2000
Time (hours)
3000
(a) Machines contributions
600
Output P(t)
500
H = 1147
400
Demand σ
M3 contribution (f3(t))
300
M2 contribution (f3(t)+f2(t))
M1 contribution (f1(t)+f2(t)+f3(t))
200
100
0
0
500
1000
1500
Time (hours)
2000
2500
3000
(b) Schedule
Figure 5: Solution obtained with the successive projections – m = 3 machines, σ = 0.4 · P nomtot
16
600
600
fmax1(t)
fmax2(t)
400
200
0
0
M2 contribution (f2(t))
Output P(t)
Output P(t)
M1 contribution (f1(t))
1000
2000
Time (hours)
600
400
200
0
0
3000
1000
2000
Time (hours)
3000
fmax (t)
3
Output P(t)
M3 contribution (f3(t))
400
200
0
0
1000
2000
Time (hours)
3000
(a) Machines contributions
500
H = 1745
Output P(t)
400
Demand σ
M3 contribution (f3(t))
300
M2 contribution (f3(t)+f2(t))
M1 contribution (f1(t)+f2(t)+f3(t))
200
100
0
0
500
1000
1500
Time (hours)
2000
2500
3000
(b) Schedule
Figure 6: Solution obtained with the Mirror-Prox-based algorithm – m = 3 machines, σ = 0.4 · P nomtot
17
Computation time (s.)
750
Resolution method
500
Projections
Mirror Prox
250
0
30
40
50
60
70
80
90
load (%)
Figure 7: Computation times – m = 25 machines
18
| 5cs.CE |
1
arXiv:1711.05764v1 [cs.DS] 15 Nov 2017
Online Allocation with Traffic Spikes: Mixing Adversarial and
Stochastic Models
HOSSEIN ESFANDIARI, University of Maryland
NITISH KORULA, Google Research
VAHAB MIRROKNI, Google Research
Motivated by Internet advertising applications, online allocation problems have been studied extensively in
various adversarial and stochastic models. While the adversarial arrival models are too pessimistic, many
of the stochastic (such as i.i.d or random-order) arrival models do not realistically capture uncertainty in
predictions. A significant cause for such uncertainty is the presence of unpredictable traffic spikes, often due
to breaking news or similar events. To address this issue, a simultaneous approximation framework has been
proposed to develop algorithms that work well both in the adversarial and stochastic models; however, this
framework does not enable algorithms that make good use of partially accurate forecasts when making online
decisions. In this paper, we propose a robust online stochastic model that captures the nature of traffic spikes
in online advertising. In our model, in addition to the stochastic input for which we have good forecasting, an
unknown number of impressions arrive that are adversarially chosen. We design algorithms that combine a
stochastic algorithm with an online algorithm that adaptively reacts to inaccurate predictions. We provide
provable bounds for our new algorithms in this framework. We accompany our positive results with a set of
hardness results showing that our algorithms are not far from optimal in this framework. As a byproduct of
our results, we also present improved online algorithms for a slight variant of the simultaneous approximation
framework.
Additional Key Words and Phrases: Online Matching, Online Budgeted Allocation, Traffic Spikes, Online
Advertisement
ACM Reference format:
Hossein Esfandiari, Nitish Korula, and Vahab Mirrokni. 2017. Online Allocation with Traffic Spikes: Mixing
Adversarial and Stochastic Models. 1, 1, Article 1 (November 2017), 23 pages.
https://doi.org/10.1145/nnnnnnn.nnnnnnn
1
INTRODUCTION
In the past decade, online budgeted allocation problems have been studied extensively due to their
important applications in Internet Advertising. In such problems, we are given a bipartite graph
G = (X , Y , E) with a set of fixed nodes (also known as agents, or advertisers) Y , a set of online nodes
(corresponding to items or impressions) X , and a set E of edges between them. Each agent / fixed
node j ∈ Y is associated with a total weighted capacity (or budget) c j ; in the context of Internet
advertising, each agent corresponds to an advertiser with a fixed budget to spend on showing
their ad to users. The items / online nodes i ∈ X arrive one at a time, along with their incident
edges (i, j) ∈ E(G) and the weights w i, j on these edges. These online nodes correspond to search
queries, page-views, or in general, impressions of ads by users. Upon the arrival of an item i ∈ X ,
Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee
provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and
the full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored.
Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires
prior specific permission and/or a fee. Request permissions from permissions@acm.org.
© 2017 Association for Computing Machinery.
XXXX-XXXX/2017/11-ART1 $15.00
https://doi.org/10.1145/nnnnnnn.nnnnnnn
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
1:2
Hossein Esfandiari, Nitish Korula, and Vahab Mirrokni
the algorithm can assign i to at most one agent j ∈ Y where (i, j) ∈ E(G) and the total weight of
nodes assigned to j does not exceed c j . The goal is to maximize the total weight of the allocation.
This problem is known as the Budgeted Allocation or AdWords problem, and it has been studied
maxi, j w i, j
→ 0, in [Buchbinder et al. 2007; Devanur and Hayes 2009;
under the assumption that min
j cj
Mehta et al. 2007] (called the large-degree assumption). Many variants of this problem such as the
display ads problem [Feldman et al. 2009a] have been studied, and techniques to solve the budgeted
allocation problem have been generalized to solve those problems.
Traditionally, results have been developed for a worst-case arrival model in which the algorithm
does not have any prior on the arrival model of online nodes. Under this most basic online model,
known as the adversarial model, the online algorithm does not know anything about the items
or E(G) beforehand. In this model, the seminal result of Karp, Vazirani and Vazirani [Karp et al.
1990] gives an optimal online 1 − e1 -competitive algorithm to maximize the size of the matching
for unweighted graphs where w i j = 1 for each (i, j) ∈ E(G). For weighted graphs, Mehta et
al. [Buchbinder et al. 2007; Mehta et al. 2007] presented the first 1 − e1 -approximation algorithm to
maximize the total weight of the allocation for the AdWords problem.
In practical settings motivated by placement of Internet ads, the incoming traffic of page-views
may be predicted with a reasonable precision using a vast amount of historical data. Motivated
by this ability to forecast traffic patterns, various stochastic online arrival models have been
introduced. Such models include (i) the i.i.d. stochastic arrival model in which there is a (known or
unknown) distribution over the types of items, and each item that arrives is drawn i.i.d. from this
distribution [Feldman et al. 2009b; Vee et al. 2010], or (ii) the random order model [Agrawal et al.
2009; Devanur and Hayes 2009; Feldman et al. 2010], which makes the weaker assumption that
individual items and edge weights can be selected by an adversary, but that they arrive in a random
order. Several techniques have been developed to design asymptotically optimal online allocation
algorithms for these stochastic arrival models (For example, these algorithms include a set of
dual-based algorithms [Devanur and Hayes 2009; Feldman et al. 2010], and a set of primal-based
algorithms discussed later).
These algorithms for the stochastic models are useful mainly if the incoming traffic of items
(i.e. online impressions) can be predicted with high precision. In other words, such algorithms
tend to rely heavily on a precise forecast of the online traffic patterns (or if the forecast is not
explicitly provided in advance, that the pattern ‘learnt’ by the algorithm is accurate), and hence
these algorithms may not react quickly to sudden changes in traffic. In fact, the slow reaction to
such traffic spikes imposes a serious limitation in the real-world use of stochastic algorithms for
online advertising, and more generally, this is a common issue in applying stochastic optimization
techniques to online resource allocation problems (see e.g., [Wang et al. 2006]). To the best of our
knowledge, no large Internet advertising systems deploy such stochastic allocation algorithms
‘as-is’ without modifications to deal with situations where the forecasts are inaccurate. Various
techniques such as robust or control-based stochastic optimization have been described in the
literature [Ben-Tal and Nemirovski 2002; Bertsimas et al. 2004; Tan and Srikant 2011; Wang et al.
2006] to deal with this shortcoming, but they do not provide theoretical guarantees when the input
is near-adversarial.
One recent theoretical result in this direction is the simultaneous adversarial and stochastic
framework [Mirrokni et al. 2011]. The main question of this recent work is whether there exists
an algorithm which simultaneously achieves optimal approximation ratios in both the adversarial
and random-order settings. More specifically, does there exist an algorithm achieving a 1 − ε
approximation for the random-order model, and at the same time achieving a 1 − e1 -approximation
for the adversarial model? Mirrokni et al. [2011] showed that the answer to this question is positive
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
Online Allocation with Traffic Spikes: Mixing Adversarial and Stochastic Models
1:3
for unweighted bipartite graphs, but it is negative for the general budgeted allocation problem.
Further, they show that the best 1 − e1 -competitive algorithm for the adversarial model achieves a
0.76-approximation in the random-order model. Though this shows that the adversarial algorithm
has an improved competitive ratio in stochastic settings, it does not use forecast information
explicitly, and hence it can be quite far from optimal even when the forecast is perfectly accurate.
Moreover, the simultaneous approximation framework is still trying to design an algorithm that is
guaranteed to work well in extreme situations (where the input follows the forecast perfectly, or
is completely adversarial). What if the forecast is mostly, but not entirely accurate? For instance,
suppose traffic to a website largely follows the prediction, but there is a sudden spike due to a
breaking news event? Treating this as entirely adversarial input is too pessimistic.
Our Model and Results. In this paper, we propose a model of online stochastic budgeted
allocation with traffic spikes, referred to as Robust Budgeted Allocation, that goes beyond the worstcase analysis in the adversarial model, and develop algorithms that explicitly use the stochastic
information available for arrival pattern. In our model, in addition to the stochastic input for which
we have good forecasting, an unknown number of impressions arrive that are adversarially chosen.
This model is motivated by the patterns of traffic spikes in online advertising in which part of the
incoming traffic of users may be the result of a new event that we did not predict, corresponding
to a new traffic pattern. We design an algorithm that adaptively checks if the traffic forecast is
accurate, and reacts to flaws in traffic forecasting due to traffic spikes. We measure the accuracy
of the forecast in terms of a parameter λ which, roughly speaking, captures the fraction of the
value of an optimal solution that can be obtained from the stochastic input (as opposed to the
adversarially chosen impressions). In general, the competitive ratio of the algorithm will naturally
increase with λ. Furthermore, we accompany our results with a set of hardness results showing
that our provable approximation guarantees are not far from the optimal achievable bounds. Hence,
compare to the simultaneous approximation framework of [Mirrokni et al. 2011], our method
provides almost-optimal approximation guarantees for all values of λ, and not only for the extreme
cases where either the input is totally adversarial or totally stochastic.
Interestingly, our techniques also result in new approaches for the simultaneous approximation
framework of [Mirrokni et al. 2011]; though the models are slightly different (i.i.d. vs. random
order, as well as the fact that we require a possibly inaccurate forecast of traffic), our algorithm
gives improved performance for the weighted case under uncertain input compared to what was
achieved in that paper. Section 2 describes the model precisely, allowing us to formally state our
results.
Our technique is based on defining a notion of ε-closeness in distributions, and then understanding the behaviour of an online algorithm over sequences that are ε-close to a given distribution.
Most notably, we can show how to modify any online stochastic algorithm to work for online
adversarial input sequences that are ε-close to a known distribution. This technique is summarized
in the next section. We then combine such a modified stochastic algorithm with an adversarial
algorithm to guarantee robustness. Converting this idea to provable algorithms for the robust
online allocation problem requires applying several combinatorial lemmas and proving invariants
that can be converted to a factor-revealing mathematical program, which can then be analyzed
numerically and analytically to prove desirable competitive ratios.
1.1
Other Related Work
Online Stochastic Allocation. Two general techniques have been applied to get improved approximation algorithms for online stochastic allocation problems: primal-based and dual-based
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
1:4
Hossein Esfandiari, Nitish Korula, and Vahab Mirrokni
techniques. The dual-based technique is based on solving a dual linear program on a sample instance, and using this dual solution in the online decisions. This method was pioneered by Devanur
and Hayes [Devanur and Hayes 2009] for the AdWords problem and extended to more general
problems [Agrawal et al. 2009; Feldman et al. 2010; Vee et al. 2010]. It gives a 1 − ε-approximations
for the random order model if the number of items m is known to the algorithm in advance, and
m log n
O PT
w i j ≥ O( ε 3 ), where n := |Y | is the number of agents. The primal-based technique is based on
solving an offline primal instance, and applying this solution in an online manner. This method
applies the idea of power-of-two choices, and gives improved approximation algorithms for the
iid model with known distributions. This technique was initiated by [Feldman et al. 2009b] for
the online (unweighted) matching problem and has been improved [Bahmani and Kapralov 2010;
Haeupler et al. 2011; Jaillet and Lu 2013; Menshadi et al. 2011]. All the above algorithms heavily
rely on an accurate forecast of the traffic. An alternative technique that has been applied to online
stochastic allocation problems is based on optimizing a potential function at each stage of the
algorithm [Devanur et al. 2011, 2012]. This technique has been analyzed and proved to produce
asymptotically optimal results under the i.i.d. model with unknown distributions. Although this
technique does not rely on the accurate predictions as much, it does not combine stochastic and
adversarial models, and the analysis techniques used are not applicable to our robust online allocation model. For unweighted graphs, it has been recently observed that the Karp-Vazirani-Vazirani
1 − e1 -competitive algorithm for the adversarial model also achieves an improved approximation
ratio of 0.70 in the random arrival model [Karande et al. 2011; Mahdian and Yan 2011]. This holds
even without the assumption of large degrees. It is known that without this assumption, one
cannot achieve an approximation factor better than 0.82 for this problem (even in the case of i.i.d.
draws from a known distribution) [Menshadi et al. 2011]. All the above results rely on stochastic
assumptions and apply only to the random-order or the iid stochastic models.
Robust stochastic optimization. Dealing with traffic spikes and inaccuracy in forecasting the
traffic patterns is a central issue in operations research and stochastic optimization. Methods including robust or control-based stochastic optimization [Ben-Tal and Nemirovski 2002; Bertsimas et al.
2004; Tan and Srikant 2011; Wang et al. 2006] have been proposed. These techniques either try to
deal with a larger family of stochastic models at once [Ben-Tal and Nemirovski 2002; Bertsimas et al.
2004; Wang et al. 2006], try to handle a large class of demand matrices at the same time [Applegate
and Cohen 2006; Azar et al. 2003; Wang et al. 2006], or aim to design asymptotically optimal
algorithms that react more adaptively to traffic spikes [Tan and Srikant 2011]. These methods have
been applied in particular for traffic engineering [Wang et al. 2006] and inter-domain routing [Applegate and Cohen 2006; Azar et al. 2003]. Although dealing with similar issues, our approach and
results are quite different from the approaches taken in these papers. For example, none of these
previous models give theoretical guarantees in the adversarial model while preserving an improved
approximation ratio for the stochastic model. Finally, an interesting related model for combining
stochastic and online solutions for the Adwords problem is considered in [Mahdian et al. 2007],
however their approach does not give an improved approximation algorithm for the i.i.d. model.
2
2.1
PRELIMINARIES AND TECHNIQUES
Model
Let I denote a set of item ‘types’; in the Internet advertising applications, these represent queries /
ad impressions with different properties that are relevant to advertiser targeting and bidding. A
forecast F = (D, f ) has two components: A distribution D over I, together with a number f ; this is
interpreted as a prediction that f items will arrive, each of which is drawn independently from D.
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
Online Allocation with Traffic Spikes: Mixing Adversarial and Stochastic Models
1:5
In the Stochastic Budgeted Allocation problem, the input known to the algorithm in advance is a
forecast F = (D, f ), and a set of agents Y , with a capacity c j for agent j. A sequence of f items is
drawn from the distribution D and sent to the algorithm one at a time; the algorithm must allocate
these items as they appear online. The total weight of items allocated to agent j must not exceed c j ,
and the objective is to maximize the weight of the allocation. As discussed above, there has been
considerable work on near-optimal algorithms for Stochastic Budgeted Allocation [Agrawal et al.
2009; Alaei et al. 2012; Devanur and Hayes 2009; Feldman et al. 2010; Vee et al. 2010].
In this paper, we define the new Robust Budgeted Allocation problem, for which our input model is
the following: The adversary can create in advance an arbitrary forecast F = (D, f ), and a collection
of agents Y . Further, at each time step, the adversary can either create a new arbitrary item (together
with its incident edges and weights) and send it to the algorithm, or choose to send an item drawn
from D. After at least f items have been drawn from D, the adversary can either send additional
(arbitrary) items, or choose to terminate the input. The online algorithm knows in advance only
the forecast F and the agents Y , and so it knows that it will receive f items corresponding to i.i.d.
draws from D; it does not know anything about the items created by the adversary, where in the
sequence they arrive, or the total number m of items that will arrive. As usual, the competitive ratio
of the algorithm is measured by the worst-case ratio (over all inputs) of the value of its allocation
to the value of the optimal allocation on the sequence that arrives.
With the preceding description of the model, no algorithm can have a competitive ratio better
than 1 − 1/e, for the simple reason that we could set f = 0, allowing the adversary to control the
entire input. (Or even for larger f , the edge weights for the adversarial items could be considerably
larger than the weights for the forecast items in I.) We have not quantified the accuracy of the
forecast, or meaningfully limited the power of the adversary. Our goal is to design algorithms with
a competitive ratio that improves with the accuracy of the forecast. We quantify this accuracy as
follows:
Definition 2.1. For an instance I of the Robust Budgeted Allocation problem with forecast (D, f ),
let S(I ) denote the set of f ‘stochastic’ items drawn from distribution D. Let A(I ) denote the set
of n − f ‘adversarial’ items. When I is clear from context, we simply use S and A to denote the
stochastic and adversarial items respectively. We mildly abuse notation and, when clear from
context, also use I to refer to the sequence of items in an instance. For a solution Sol to an instance I ,
let ValS (Sol) denote the value obtained by Sol from allocating the items of S to agents, and ValA (Sol)
denote the value obtained by Sol from allocating the items of A.
Definition 2.2. An instance I of the Robust Budgeted Allocation problem is said to be λ-stochastic
S (Sol )]
if λ = maxSol ∈O PT (I ) { E[ValSE[Val
(Sol )+ValA (Sol )] }, where OPT (I ) is the set of all optimal solutions of I .
Note that when the forecast is completely accurate (there are no adversarial items), the instance
is 1-stochastic, and when f = 0 (all items are adversarial), the input is 0-stochastic. Though λ is
unknown to the algorithm, our goal is to design algorithms which, when restricted to λ-stochastic
instances, have good competitive ratio (ideally, increasing in λ).
2.2
Algorithms for Working with Forecasts
In this section, we consider how to solve the Stochastic Budgeted Allocation problem. Similar
problems have been applied before (see e.g. [Alaei et al. 2012]), but we describe a specific approach
below that will be a useful tool for the Robust Budgeted Allocation problem. Further, our argument
implies that this approach performs well even for an adversarial input sequence if it is sufficiently
‘close’ to the forecast.
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
1:6
Hossein Esfandiari, Nitish Korula, and Vahab Mirrokni
Roughly speaking, given a forecast F = (D, f ), if the number of items f is sufficiently large, then
we can work with an ‘expected instance’. In the expected instance, the set of items is created by
assuming each type t ∈ I arrives in proportion to P D=t . We then run any (offline) algorithm Alg on
the expected instance; when an item arrives in the real online instance, we assign it according to
the allocation given by Alg on the expected instance. If the number of items f is sufficiently large,
then only a small error is induced because we assign according to the expected instance instead of
the random realization of the forecast.
We begin by defining the notion of a sequence being ε-close to a distribution. Indeed, we show that
with high probability a ‘long’ sequence of draws from a distribution is ε-close to that distribution,
where ε is an arbitrary small constant. We next prove that if we ignore inputs which are not
ε-close to the input distribution, we lose only m1 in the competitive ratio. Finally, we show that we
can modify any online stochastic algorithm, or more strongly any offline algorithm for Budgeted
Allocation to work for online adversarial input sequences which are guaranteed to be ε-close to a
known distribution. Interestingly, this reduction loses only 4ε on the competitive ratio.
Definition 2.3. Let S = ⟨s 1 , s 2 , ..., sm ⟩ be a sequence of items and let D be a distribution over a
set of item types I. For a type t ∈ I, let P D=t be the probability that a draw from D is t. We say S
is ε-close to distribution D, if for any continuous sub-sequence Si,k = si , si+1 , . . . , sk ⊂ S and any
type t, the number of items of type t in Si,k is within the range (k − i + 1 ± εm)P D=t . If S is not
ε-close to distribution D we say it is ε-far from the distribution.
Consider that (k − i + 1)P D=t is the expected number of items of type t in a set of k − i + 1 draws
from D. When k − i + 1 is large, using the Chernoff bound we can show that the number of items of
type t is close to the expectation. On the other hand, when k − i + 1 is small, the term εm dominates
k − i + 1, and thus the number of items of type t is in the range (k − i + 1 ± εm)P D=t . Lemma 2.6
formalizes this intuition, showing that with high probability a [long enough] sequence of items
drawn from a distribution D is ε-close to D.
Definition 2.4. Given a distribution D, a sequence of f items is said to satisfy the long input
f
condition for D if log(f ) ≥ ε 2mint 15
.
∈D (P D=t )
We use the following version of the Chernoff bound in Lemma 2.6.
Proposition 2.5. Let x 1 , x 2 , . . . , xm be a set of independent boolean random variables. For X =
Ím
−δ 2
i=1 x i and µ = E[X ] we have: Pr (|X − µ | ≥ δ ) ≤ 2 exp( 3µ )
Lemma 2.6. Let S be a sequence of m items drawn from a distribution D. Assuming the long input
condition, the probability that S is ε-far from D is at most m12 .
Proof. S contains m(m−1)
subsequences. In addition, the long input condition gives us P D=t ≥
2
2m
3
15 log(m)
. This means that there are at most 15 εlog(m)
different types. Thus, we have fewer than m2
ε 2m
combinations of a type and a subsequence. We next argue that the probability that the number of
items of a fixed type t in a fixed sub sequence Si,k is out of the range (k − i + 1 ± εm)P D=t is at
3
most m25 . Applying the union bound, the probability that S is ε-far from D is at most m25 m2 = m12 ,
as desired.
For a type t and an index 1 ≤ ℓ ≤ m, let x ℓt be a random variable which is 1 if the ℓ-th item in
sequence S is of type t and is 0 otherwise. The variables x kt are independent for a fixed type t and
t be Ík x t , which is the number of items of type t in the sub sequence
different indices. Let X i,k
ℓ=i ℓ
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
Online Allocation with Traffic Spikes: Mixing Adversarial and Stochastic Models
1:7
t is (k − i + 1)P
Si,k . The expected value of X i,k
D=t , and by applying the Chernoff bound we have:
−ε 2m 2
)
3(j − i + 1)Pt =D
−ε 2m 2
)
≤ 2 exp(
2
3n 15 mε
log(m)
Pr (|X i,t j − (j − i + 1)P D=t | ≥ εm) ≤ 2 exp(
= 2 exp(−5 log(m)) =
This completes the proof of the lemma.
2
.
m5
□
In the rest of this section, we use the monotonicity and subadditivity properties of Budgeted
Allocation, stated in Lemma 2.7 and Lemma 2.8 respectively.
Lemma 2.7 (Monotonicity). Budgeted Allocation is monotone: Fixing the set of agents and their
capacities, for any sequence of items S and any sequence T ⊆ S, we have Opt(T ) ≤ Opt(S) where
Opt(S) and Opt(T ) are the values of the optimum solutions when the items that arrive are S and T
respectively.
Proof. Any feasible allocation of items of T is a feasible allocation of items of S as well. This
immediately means Opt(T ) ≤ Opt(S) as desired.
□
Lemma 2.8 (subadditivity). Budgeted Allocation is subadditive: Fixing the set of agents and
their capacities, for any sequence of items S and any sequence of items T , we have Opt(S ∪ T ) ≤
Opt(S) + Opt(T ) where Opt(X ) indicates the size of the optimum solution when the sequence of items
that arrive is X.
Proof. Fix an optimum solution Opt(S ∪ T ). The allocation of items of S in Opt(S ∪ T ) is a
feasible allocation for S. Similarly the allocation of items of T in Opt(S ∪ T ) is a feasible allocation
for T . Therefore we have Opt(S ∪ T ) ≤ Opt(S) + Opt(T ).
□
Lemma 2.6 says that w.h.p, a sequence of items drawn from a distribution D is ε-close to D. That
is, inputs which are ε-far from the input distribution are rare, but this does not immediately imply
that the total value of such rare ε-far inputs is small as well. Lemma 2.9 says that we may ignore all
inputs which are ε-far from the input distribution and only lose a small fraction in the competitive
ratio.
Lemma 2.9. Let S be a sequence of m items drawn from a distribution D, satisfying the long input
condition. Let Alg be an α-competitive algorithm for Stochastic Budgeted Allocation with forecast
(D, m). Let Algclose be an algorithm that has the same outcome as Alg when the input is ε-close to D
and 0 otherwise. Algclose is (α − m1 )-competitive.
Proof. Let Algfar be an algorithm that has the same outcome as Alg when the input is ε-far
from D and 0 otherwise. We slightly abuse notation and use Alg, Algclose , Algfar to refer both to
the algorithms and the expected values of their outcomes. By definition we have Alg = Algclose +
Algfar . Let Opt denote the expected value of an optimal solution on the sequence drawn from the
distribution. We bound the expected outcome of Algfar to compare the competitive ratio of Alg and
Algclose .
Let Sk be an input that contains k items of each type. The monotonicity of Budgeted Allocation
implies that for any sequence S of size m, Opt(Sm ) is greater than Opt(S). Moreover, by subadditivity,
Opt(Sm ) is at most m2 times Opt(S 2 ). Together, these imply that, for any sequence S of size m, we
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
1:8
Hossein Esfandiari, Nitish Korula, and Vahab Mirrokni
have Opt(S) ≤ m2 Opt(S 2 ). On the other hand, the number of items of type t in any sequence S of
size m which is ε-close to D is at least
15
15
(1 − ε)f P D=t ≥ (1 − ε) 2
P D=t ≥ (1 − ε) 2 ≥ 2.
ε mint ∈D (P D=t )
ε
Thus by monotonicity we have
of size m we have
1
P r (S is ε close to D) Optclose
Opt(S) ≤
≥ Opt(S 2 ). Therefore, for any sequence S
m
m
1
Optclose ≤ mOpt
Opt(S 2 ) ≤
2
2 (1 − m12 )
This together with Lemma 2.6 gives us
Algclose = Alg − Algfar ≥ Alg − mOpt
as desired.
1
1
1
≥ αOpt − Opt = (α − )Opt
2
m
m
m
□
Note that any algorithm Alg for Budgeted Allocation has a random outcome when the items are
drawn from a forecast, simply due to the randomness in the sequence of items. We now define a
derandomization of such algorithms: Given an algorithm Alg, a forecast F = (D, f ) and a constant
ε, algorithm DeRandεF (Alg) is defined as follows:
Let S ′ be a sequence of (1 − ε)f impressions, with (1 − ε)f P D=t impressions of type t, for each
type t. Run algorithm Alg on sequence S ′. Let Alg(S ′, t, i) be the agent to which Alg assigns the ith
impression of type t in S ′. Note that any sequence of f items which is ε-close to D contains at least
(1 − ε)f P D=t impressions of each type t. We can now describe how DeRandεF (Alg) allocates items of
a sequence S. For each type t, Algorithm DeRandεF (Alg) allocates the first (1 −ε)f · P D=t impressions
of type t in S in the same manner as Alg allocated S ′. That is, we assign the ith impression of type
t in S to Alg(S ′, t, i). After the first (1 − ε)f P D=t impressions of type t have been allocated, we do
not allocate any more items of this type. Finally, if at any time during the algorithm, we observe
that the input sequence (so far) is not ε-close to distribution D, the algorithm stops and returns
false. Otherwise, it returns true.
When it is clear from the context, we drop F and ε from the notation of DeRandεF (Alg).
Remark 2.10. Note that for any forecast F = (D, f ) and constant ε, the outcome of DeRandεF (Alg)
on any item sequence of length f that is ε-close to D is a function purely of D and ε, but not the actual
impressions in the sequence.
Theorem 2.11. Let F = (D, f ) be a distribution, and let A be an adversarial input with length
f such that A satisfies the long input condition and A is ε-close to D. Let Alg be an α-competitive
algorithm for Stochastic Budgeted Allocation. Though A is not explicitly drawn from D, DeRandεF (Alg)
is α − 2ε competitive on sequence A.
Proof. Since A is ε-close to D, DeRand(Alg) will not return false on S, and thus, the allocation
of DeRand(Alg) on S is identical to that of Alg on S ′. Let S ′′ be a sequence of (1 + ε)d impressions,
(1 + ε)f P D=t from each type t. Using the subadditivity of Budgeted Allocation, we have Opt(S ′) ≥
1−ε
′′
′′
1+ε Opt(S ) ≥ (1 − 2ε)Opt(S ).
On the other hand, using the monotonicity of Budgeted Allocation, Opt(A) ≤ Opt(S ′′). Together,
these inequalities imply that Opt(S ′) ≥ (1 − 2ε)Opt(A), which means that DeRandεF (Alg) is (α − 2ε)competitive.
□
Consider an α-competitive online algorithm (or α-approximate offline algorithm), Alg, for
Budgeted Allocation with stochastic input. By Theorem 2.11, for inputs which are ε-close to
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
Online Allocation with Traffic Spikes: Mixing Adversarial and Stochastic Models
1:9
Fig. 1. Summary of results, parametrized by λ. The blue curve denotes the upper bound on the competitive
ratio of any algorithm, the green curve is our algorithm for the unweighted case, and the red curve is our
algorithm for the weighted case.
D, DeRandεF (Alg) has a competitive ratio that is only 2ε worse than that of Alg. Moreover, Lemma
2.9 says that if we ignore all inputs which are ε-far from the input distribution, we lose only m1 on
the competitive ratio. Together with the assumption that ε ≥ m1 , we obtain the following corollary.
Corollary 2.12. Let Alg be an α-competitive algorithm for Stochastic Budgeted Allocation (or
α-approximate offline algorithm for Budgeted Allocation). Assuming the long input condition, for any
small constant ε ≥ m1 and any forecast F , DeRandεF (Alg) is an (α − 3ε)-competitive algorithm for
Stochastic Budgeted Allocation.
3
ROBUST ONLINE BUDGETED ALLOCATION
In this section, we consider the Robust Budgeted Allocation problem. As described above, the
algorithm knows in advance the set of agents Y , together with a capacity c j for each agent j ∈ Y .
Further, it knows the forecast F = (D, f ), but not how many additional items will be sent by the
adversary, nor how they are intermixed with the f items drawn from D.
Recall that a λ-stochastic instance is one where the ‘stochastic items’ (those drawn from D)
provide λ-fraction of the value of an optimal solution. (Also, note that λ is not known to the
algorithm.) As λ increases (corresponding to an increase in the accuracy of our forecast, or a smaller
traffic spike), we expect the performance of our algorithms to improve. However, we wish our
algorithms to be robust, obtaining good performance compared to an optimal offline solution even
when λ is close to 0 (corresponding to a very large traffic spike, when the typical ‘forecast’ traffic is
only a small fraction of the total).
First, in Section 3.1, we consider the unweighted Robust Budgeted Allocation problem, in which
w i j is the same for all (i, j) ∈ E. As desired, we obtain an algorithm with competitive ratio tending
to 1 as λ tends to 1, and 1 − 1/e when λ tends to 0. Then, in Section 3.2, we consider the general
weighted Robust Budgeted Allocation problem, and finally, in Section 3.3, we give upper bounds on
the competitive ratio of any algorithm. All our competitive ratios are parametrized by λ, and they
are summarized in Figure 1.
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
1:10
Hossein Esfandiari, Nitish Korula, and Vahab Mirrokni
For simplicity, throughout this section, we assume that the capacity of all the agents are the
same (normalized to 1). This assumption can be removed by dividing agents with large capacities
into multiple dummy agents. Moreover, as usual we assume maxi, j w i, j → 0 (a.k.a., large degree
assumption).
3.1
Unweighted Robust Budgeted Allocation
Our general approach will be to simulate the following idea: Suppose we first receive the f stochastic
items drawn from D. We could allocate these items optimally (assuming they are the entire input).
If this uses up most of the budgets of the agents, we automatically have a good solution, even if we
do not allocate any of the adversarial items. If some fraction of the budgets remain unused, use this
remaining capacity to allocate any adversarial items that arrive. If there is a way to allocate these
adversarial items using the remaining capacities, we obtain 1 − 1/e fraction of this value.
Unfortunately, in the real Robust Budgeted Allocation model, we do not know which items are
stochastic and which are adversarial, so we cannot perform this clean separation perfectly. Still, we
can approximate this separation as follows: Let A be an algorithm for Stochastic Budgeted Allocation.
We have two algorithms running simultaneously. The first is a slight variant of DeRandεF (A), and
the second is Balance [Kalyanasundaram and Pruhs 2000]. More precisely, we send each item to
DeRandεF (A); recall that this algorithm first checks if the input seen so far is ε-close to D. If it is, it
allocates this item according to A; otherwise, it returns false. Now, instead of returning false, we
assume that this item must have been sent by the adversary. As such, we erase this item from the
history of DeRandεF (A), and try to allocate it using Balance. That is, we ‘guess’ that all items that
are matched by DeRandεF (A) are stochastic items and all other items are adversarial.
Note that DeRandεF (A) does not match more than (1 − ε)f items, (1 − ε)f P D=t from each type t.
From Lemmas 2.6 and 2.9, we know that w.h.p., the sequence of stochastic items is ε-close to D; by
ignoring the case when it is ε-far from D, we lose at most m1 in the competitive ratio. Given the
long input assumption, 1/m < ε, and hence by losing ε in the competitive ratio, we assume that the
stochastic items are always ε-close to D. Since the sequence of stochastic items is ε-close to D, we
have at least (1 − ε)f P D=t items of type t. Therefore, the items that DeRandεF (A) leaves unallocated
are a superset of all the adversarial items. (More precisely, there may be an adversarial item of type
t that we guess is stochastic, but there must be a corresponding stochastic item of type t that we
treat as adverarial instead.)
We now complete the description of our combined algorithm: Given an allocation, let x j denote
the fraction of agent j’s budget used in this allocation. Let AlgS be a (1 − ε)-competitive algorithm
Í
for Stochastic Budgeted Allocation with the further property that it minimizes nj=1 x j2 . (In other
words, AlдS reproduces an optimal offline solution such that each item is allocated to an eligible
agent with the lowest x j .) We run DeRandεF (AlgS ) as the first of our two algorithms. Recall from
Remark 2.10 we know the exact allocation of DeRandεF (AlдS ), which is independent of the random
draws. If x j denotes the fraction of agent j’s capacity used by this allocation, for items unallocated
by DeRandεF (AlдS ), we run the Balance algorithm on the instance of Budgeted Allocation with the
adversarial items, and in which the capacity of agent j is 1 − x j .
How does this combined algorithm perform? We use S to denote the stochastic items, and A the
adversarial items. From Corollary 2.12 we know that DeRandεF (AlgS ) is 1 − O(ε) competitive against
Opt(S). We will prove in Lemma 3.12 that the optimum solution on the items of A using the remaining
capacities 1−x j is at least (1−λ −O(ε))(Opt(A∪S)−Opt(S)). Since Balance is a (1−1/e)-competitive
algorithm, the value we derive from Balance is at least (1 − e1 )(1 − λ − O(ε))(Opt(A ∪ S) − Opt(S)).
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
Online Allocation with Traffic Spikes: Mixing Adversarial and Stochastic Models
1:11
Ín
Minimize:
j=1 (y j − z j )
Ín
j=1 y j
Subject to: z j ≤ max(0, x j + y j − 1)
n
Õ
(x k + yk ) ≤ n − j
∀1 ≤ j ≤ n
∀1 ≤ j ≤ n − 1 s.t. x j < x j+1
k =j+1
x j ≤ x j+1
n
n
n
Õ
Õ
Õ
λ( x j +
yj ) ≥
xj
j=1
j=1
∀1 ≤ j ≤ n − 1
j=1
0 ≤ x j , yj , zj ≤ 1
∀1 ≤ j ≤ n
Fig. 2. Mathematical Program 2 to bound the optimum allocation of adversarial items on the remaining
capacities in uniform weight case.
Therefore, the competitive ratio of the combined algorithm is at least:
(1 − O(ε))Opt(S) + (1 − λ − O(ε))(1 − e1 )(Opt(A ∪ S) − Opt(S))
Opt(S ∪ A)
λOpt(S ∪ A) + (1 − λ)2 (1 − e1 )Opt(S ∪ A) − O(ε)Opt(S ∪ A)
=
Opt(S ∪ A)
1
= λ + (1 − λ)2 (1 − ) − O(ε)
e
where the first equality uses the definition of λ-stochastic to replace Opt(S) with λOpt(S ∪ A). This
proves the following theorem.
Theorem 3.1. Assuming the long input condition, there exists a λ +(1−λ)2 (1− e1 )−O(ε) competitive
algorithm for Robust Budgeted Allocation when the input is λ-stochastic. This algorithm does not
depend on λ.
In order to prove the key Lemma 3.12, we first write Mathematical Program 2, and show in
Lemma 3.3 that this lower bounds the ratio of the optimum allocation of the adversarial items on
the remaining capacities to the overall optimum. Next, in Lemma 3.4, we show that the solution of
this mathematical program is at least 1 − λ. Lemma 3.12 is an immediate result of combining these
two lemmas.
We now show that there exists an optimum allocation of all items such that the contribution of
stochastic items in this allocation is exactly Opt(S). We will need this for Lemma 3.3.
Lemma 3.2. For any disjoint sets of items A and S, there exist an optimum allocation of S ∪ A such
that the contribution of items of S in this allocation is exactly Opt(S).
Proof. We prove this lemma by contradiction. Assume that in all optimum allocations of S ∪ A,
the contribution of items of S is less than Opt(S). Consider an optimum allocation Opt ′(S ∪ A)
of S ∪ A such that the items of S have the maximum contribution to the allocation. Denote the
allocation of items of S in Opt ′(S ∪ A) by A′(S). By definition, A′(S) is not an optimum allocation of
S. Thus, there exists an augmenting path in A′(S) which can increase the number of assigned items
to one of the agents by exactly one, and keep the number of assigned items to all other agents the
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
1:12
Hossein Esfandiari, Nitish Korula, and Vahab Mirrokni
same. This change increases the number of assigned items in A′(S) by one, and may decrease the
number of assigned items from A by at most one. Therefore, it is an optimal solution of S ∪ A in
which items from S have more contribution, and hence gives a contradiction.
□
Lemma 3.3. The optimum allocation of the adversarial items on the remaining capacities is at least
Opt(S ∪ A) − Opt(S) times the solution of Mathematical Program 2.
Proof. Let x j be the fraction of the capacity of the agent j which is filled by AlgS . Without
loss of generality, assume that x j s are in increasing order. That is, for all 1 ≤ j ≤ n − 1, we have
x j ≤ x j+1 , which is the third constraint in Mathematical Program 2.
Consider the optimum solution of S ∪ A from Lemma 3.2. Let y j be the fraction of the capacity
of the agent j that is filled by the adversarial items in this solution. One can see that the fourth
constraint says that the contribution of the stochastic items is at most λ fraction of the total value.
(From Lemma 3.2, we could have equality in the fourth constraint, but for simplicity of analysis,
we maintain the inequality.)
Note that we want to compare the optimum solution of the adversarial items on the remaining
Í
capacities 1 − x j with the total value nj=1 y j of the adversarial items in the optimum solution of
S ∪A. For some agent j, if we have y j +x j ≤ 1, we can assign the same adversarial items to the agent
j as in the optimum allocation. On the other hand, if y j + x j ≥ 1, we can only use 1 − x j fraction of
the capacity of agent j for the adversarial items. Thus, we can always assign y j − max(0, x j + y j − 1)
fraction of the capacity of agent i to the adversarial items. This quantity is denoted by z j in the
first constraint.
By assigning adversarial items as in the optimum solution of S ∪A using the remaining capacities,
Í
we can obtain at least nj=1 y j − z j . The objective function of Mathematical Program 2 compares
this to what the optimum solution of S ∪ A can get from adversarial items.
Finally, fix an index j such that x j < x j+1 and look at all agents with index greater than j. All
stochastic items that we match to agents with index greater than j have no edge to the agents with
index less than or equal to j. (By definition of AlgS , which assigns items to the eligible agent with
Í
lowest value of x j .) Thus, in any optimum solution of S, they cover at least nk=j+1 x k of the agents
Ín
with index greater than j. Thus, this optimum solution of S ∪ A covers k=j+1 (x k +yk ) of the agents
Í
with index greater than j. Consider that we have n − j such agents, we get nk =j+1 (x k + yk ) ≤ n − j.
This is the second constraint in Mathematical program 2.
□
Lemma 3.4. For any small number δ , the solution of the mathematical program 2 is at least
1 − λ − O(δ ).
Proof. First, we show that given any solution to the Mathematical Program 2, we can construct
a related solution with a very restricted structure that only increases the objective function by O(δ ).
We will lower bound the objective value of this new solution, which will imply a lower bound on
the original competitive ratio.
Our restricted solution will satisfy Properties 3.5, 3.7, 3.8, 3.9 ,3.10, and 3.11, which we state
below. Finally, we use these properties to bound the value of the objective function by 1 − λ. This
means that the solution of MP 2 is bounded by 1 − λ − O(δ ).
Property 3.5. Without loss of generality we can assume that for any positive number δ , we have,
≤ δ.
yj
Ín 1
j=1
To prove this property given an instance of the Mathematical program 2, we provide another instance that satisfy this property and has a solution equal to the optimum solution of
the initial instance. Consider a solution Sol n =< (x 1 , . . . , x n ), (y1 , . . . , yn ), (z 1 , . . . , zn ) >, to
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
Online Allocation with Traffic Spikes: Mixing Adversarial and Stochastic Models
1:13
the MP 2 with n agents. For any integer C, one can construct an equivalent feasible solution,
′ ), (y ′ , . . . , y ′ ), (z ′ , . . . , z ′ ) > to an extended Mathematical Program equivSol Cn =< (x 1′ , . . . , xCn
1
1
Cn
Cn
alent to 2 with Cn agents as follow. For all 1 ≤ j ≤ n and all 1 ≤ k ≤ C, set
′
′
′
xC(j−1)+k
= x j , yC(j−1)+k
= y j , zC(j−1)+k
= zj .
One can easily check that if Sol n is feasible, Sol Cn is feasible as well. The value of the objective
functions in both Sol n and Sol Cn are the same. Thus, if we bound the value of the objective function
in Sol Cn , this bound holds for Sol n as well. Moreover, since we have C copies of each variable here,
Í
Ín
′
Í1
we have Cn
j=1 y j = C j=1 y j . Hence, if we set C = δ nj=1 y j , it satisfies Property 3.5. In the rest of
the proof, we assume that the solution that we are considering has Property 3.5, and for simplicity
we refer to it as a solution with n agents and variables x j s, y j s and z j s.
Note that Property 3.5 means that if we decrease the value of a constant number of z j s to 0, we
only increase the objective value by O(δ ). We later use this to provide Property 3.7.
Definition 3.6. We say an agent j is harmful if we have z j > 0. Otherwise, we say the agent is a
sink.
If an agent j is harmful, z j is positive and thus it contributes to reducing the objective value
below 1.
Property 3.7. All y j s are either 0 or 1.
Let j be the largest index such that the agent j is harmful and y j is neither 1 nor 0. If this is not
the only harmful agent with this property, there is some other index k such that agent k is harmful
and yk is neither 1 nor 0. We decrease the value of y j and z j by some ε and increase y j and zk by ε.
This change has no effect on the objective function, or the third and fourth constraints. It increases
both sides of the first constraint by ε for agent k and decreases both sides by ε for agent j, keeping
them both feasible. One can verify that this change may only decrease the left hand side of the
second constraint for indices between k and j.
We can repeat this procedure until there is just one harmful agent j with a fractional y j . We
decrease y j and z j of this agent to 0, and by property 3.5, only increase the objective value by O(δ ).
Note that harmful agents with y j = 0 now become sink agents.
We can use the same procedure to make the y j s of all sink agents either 0 or 1.
Property 3.8. If agent j is harmful, z j = x j .
For each harmful agent j, the value of z j can be as much as x j (since y j for a harmful agent is
now 1). Increasing z j s of harmful agents up to their x j s may only decrease the objective function,
and keeps all the constraints feasible.
From now on, for each harmful agent j, we keep x j and z j the same, and if we change one of
these two, we change the other one as well. This implicit change of z j keeps the constraints feasible.
However, each time we decrease x j of a harmful agent, we need to guarantee that the objective
function does not increase. This guarantee is trivial when we are increasing the value of some other
x k to match.
Property 3.9. If an agent j > 1 is a sink, we have x j−1 = x j .
Let j and h be two consecutive harmful agents with j < h. For all agents k in the range [j, h)
Íh−1
xk
=j
we replace the value of x k with the average value of xs in this range i.e. we set x k = kh−j
. This
keeps the sum of all x k s the same, and keeps all of the constraints feasible. Moreover, since x j has
the minimum value in the range [j, h), and agent j is the only harmful agent in this range, we are
only increasing the x valuess of harmful agents. This only decreases the objective function.
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
1:14
Hossein Esfandiari, Nitish Korula, and Vahab Mirrokni
Let k be the number of different x j values that we have, and let α h be the h-th smallest value
among distinct x j s. For simplicity, we assume we always have some x j = 0. Let βh be the number
of x j s which are at most α h , and set β 0 = 0. Thus, βh−1 ≤ j < βh means that x j = α h .
Property 3.10. For any 1 ≤ h ≤ k, all harmful agents in the range (β −h1 , βh], are the first agents
in the range.
Let i be a sink and let i + 1 be a harmful agent, such that x i = x i+1 . We can replace the value of
the variables for these two agents and make i harmful, instead of i + 1. It is easy to see that this
keeps the mathematical program feasible. By repeating this, for each range (β j−1 , βj], we can move
all harmful agents to be the first agents in the range.
Property 3.11. All x j s are either 0 or the same fixed number x ∗ .
We iteratively decrease the number of distinct x j s as follow.
Let ξ be some small number that we set later. It is easy to see that one of the following two
actions decreases the objective function, since the effect of one on the objective function is the
negative of the effect of the other one. The action that decreases the objective function is the one
that we do.
ξ
ξ
• For all β 1 < j ≤ β 2 decrease x j by β2 −β1 and for all β 2 < ℓ ≤ β 3 increase x ℓ by β3 −β2
• For all β 1 < j ≤ β 2 increase x j by
ξ
β 2 −β 1
and for all β 2 < ℓ ≤ β 3 decrease x ℓ by
ξ
β 3 −β 2
We can set ξ , such that the mathematical program remains feasible and one of the following
situations happen.
• x j s in the range (β 1 , β 2 ] are all 0.
• x j s in the range (β 2 , β 3 ] are all x β3 +1 .
• The second constraint of the MP is tight for j = β 2 .
In the first two cases, the number of distinct x i s is reduced by one and we are done in these cases.
Therefore, we assume that we are in the third case. Now suppose that k > 3 (that is, there are
more than 3 distinct values of x j ); later, we consider the case that k = 3. Let ξ ′ be some small number
ξ′
that we set later. For all β 2 < j ≤ βk −1 we increase x j by βk −1 −β2 and for all βk −1 < ℓ ≤ βk we
ξ′
decrease x ℓ by βk −βk −1 . Below, we show that by doing so, the objective function does not increase.
We can set ξ ′ to make the largest and second largest value of x j s become equal. This decreases the
number of distinct values of x i s.
The second constraint in MP 2 for j = βk −1 says that the fraction of harmful agents in the
range (βk −1 , βk ] is at most 1 − x n . This constraint is tight for j = β 2 . Thus, the fraction of harmful
agents in the range (β 2 , βk ] is at least 1 − x n , which means the fraction of harmful agents in the
range (β 2 , βk −1 ] is at least 1 − x n . Thus, taking some fraction of x ℓ s in the range (βk −1 , βk ] and
distributing it equally over the range (β 2 , βk −1 ] increases the sum of z j s and thus decreases the
objective function.
In the case that k = 3, we move the harmful agents in the range (β 1 , β 2 ] to some higher indices,
to make the second constraint tight. Then, we can use the above technique and decrease the value
of the only two non zero distinct x j s to become equal.
Now, given the property 3.11 we can easily bound the objective function of the mathematical
program as follow. If x ∗ ≤ λ, all z j s are at most λ, while y j s are all 1. This bounds the objective
function by 1−λ. On the other hand, if x ∗ > λ, the second constraint in MP 2 says that the number of
harmful agents is at most 1 −x ∗ fraction of the value from the stochastic items. This means that sum
λ
of z j s is at most 1 − x ∗ fraction of value of the stochastic items. Thus, it is at most (1 − x ∗ ) 1−λ
≤λ
and this bounds the objective function by 1 − λ.
□
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
Online Allocation with Traffic Spikes: Mixing Adversarial and Stochastic Models
1:15
The following lemma is an immediate result of combining Lemma 3.3 and Lemma 3.4.
Lemma 3.12. The optimum allocation of adversarial items on the remaining capacities is at least
(1 − λ − O(δ ))(Opt(S ∪ A) − Opt(S)).
3.2
Weighted Robust Budgeted Allocation
We can now describe our algorithm for general weighted case of Budgeted Allocation. As in the
unweighted case, we combine the allocations of two algorithms: One that runs on the (stochastic)
items that we guess are drawn from the forecast, and one on the items that we guess are constructed
by the adversary.
For the stochastic items, we start with a base algorithm for Stochastic Budgeted Allocation that
is not necessarily optimal. Instead of maximizing the weight of the allocation from the stochastic
items, we start with an algorithm Algpot that maximizes the potential of the allocation, as defined
below.
Definition 3.13. Let X = (x 1 , x 2 , . . . , x n ) be a vector of numbers, such that for all 1 ≤ j ≤ n, we
have 0 ≤ x j ≤ 1. We defineÍ
the potential of x j , Pot(x j ) to be x j − e (x j −1) . We define the potential of
the vector X , Pot(X ) to be nj=1 Pot(x j ).
Let x j denote the fraction of capacity c j used by the potential-maximizing allocation of Algpot .
Similarly to the unweighted case, when items arrive, we send them to DeRand(Algpot ); for those
items that are unmatched, by DeRand(Algpot ), we send them to the Balance algorithm using the
remaining capacities 1 − x j . Exactly the same argument that we provide for the unweighted case
works here to show that by losing O(ε) on the competitive ratio, we can assume that we match
all stochastic items using DeRand(Algpot ) and all adversarial items using the Balance algorithm.
We use Alg to denote this combined algorithm. In order to analyze our algorithm Alg, we need to
define another potential function based on Pot(X ).
Definition 3.14. Let X = (x 1 , x 2 , . . . , x n ) and Y = (y1 , y2 , . . . , yn ) be two vectors of numbers
between 0 and 1. For each 1 ≤ j ≤ n, we define Pot X (y j ) as follow. If x j ≤ y j , we have Pot X (y j ) =
Pot(y j ). Otherwise, we have Pot X (y j ) = Pot(x j )+(y j −x j )Pot ′(x j ), where Pot ′(.) is the first derivative
of Pot(.). Thus, for x j < y j we have:
Pot X (y j ) = x j − e x j −1 + (y j − x j )(1 − e x j −1 )
We define Pot X (Y ) to be
Ín
j=1 Pot X (y j ).
Note that the second derivative of Pot(x j ) is −e x j −1 which is always negative. Thus, Pot(x j ) is
a concave function. Notice that in the range [0, x j ], Pot X (y j ) is equal to Pot(y j ) and in the range
(x j , 1], it is the tangent line to Pot(x j ) at x j . Hence, Pot X (y j ) is a concave function as well.
Consider that, in the range (x j , 1], the function Pot X (y j ) is the degree 2 Taylor series of Pot(y j )
at point x j . In addition, the second derivative of Pot(y j ) in the range [0, 1] is lower-bounded by −1,
yielding the following lemma.
Lemma 3.15. For any constant ε and any vector of positive numbers X and any y j such that
0 ≤ y j ≤ x j + ε ≤ 1 we have |Pot X (y j ) − Pot(y j )| ≤ ε 2 .
We now have the tools to analyze Alg via means of a mathematical program: First, in Lemma
3.17, we show that the competitive ratio of our algorithm is lower bounded by the solution of the
Mathematical program 4. Next, in Lemma 3.18, we lower bound the solution of the mathematical
program; this lower bound is shown in Figure 3. Together, these lemmas allow us to prove the
following theorem:
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
1:16
Hossein Esfandiari, Nitish Korula, and Vahab Mirrokni
Fig. 3. Competitive ratio of Alд parametrized by λ.
Ín
Minimize:
j=1 x j
Í
+ (1 − e1 ) nj=1 (y j − z j )
Ín
Ín
j=1 y j + j=1 t j
Subject to: z j ≤ max(0, x j + y j − 1)
∀1 ≤ j ≤ n
(1)
(2)
Pot X (T ) ≤ Pot(X )
(3)
tj + yj ≤ 1
∀1 ≤ j ≤ n
n
n
n
Õ
Õ
Õ
λ( t j +
yj ) =
tj
(4)
j=1
j=1
0 ≤ tj , x j , yj , zj ≤ 1
(5)
j=1
∀1 ≤ j ≤ n
(6)
Fig. 4. Mathematical Program 4 to bound the competitive ratio of Alд.
Theorem 3.16. There exists an algorithm for λ-stochastic weighted budgeted allocation with competitive ratio presented in Figure 3.
Lemma 3.17. The competitive ratio of Alд is bounded by the solution of the Mathematical program
4.
Proof. For each agent j, let x j be the expected fraction of c j used by Algpot on the stochastic
E[Stochastic(Opt )]
λ
items. Consider the optimum assignment Opt λ that maximizes
. For each agent j,
E[Opt λ ]
let t j and y j respectively denote the expected fraction of c j used by stochastic items and adversarial
items in Opt λ . Therefore, the expected fraction of c j used in Opt λ is t j +y j , which is upper-bounded
by 1. This gives us the Inequality 4.
By definition of λ, the contribution of the stochastic items to Opt λ is λ fraction of the total value;
equality 5 captures this fact.
For any agent j, if we have y j ≤ 1 − x j , an offline algorithm can allocate all the same adversarial
items to agent j as Opt λ did by using the remaining 1 − x j fraction of c j . On the other hand, if
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
Online Allocation with Traffic Spikes: Mixing Adversarial and Stochastic Models
1:17
we have y j > 1 − x j , the offline algorithm can use the entire remaining capacitity 1 − x j of agent
j. Thus, an offline algorithm that assigns all the adversarial items using the 1 − x j fraction of
capacities that remain (after allocating the stochastic items according to Algpot ) only loses at most
max(0, x j + y j − 1) from agent j when compared to the allocation of adverarial items by Opt λ . We
denote max(0, x j + y j − 1) by z j , as shown in Inequality 2.
One can see that the numerator of the objective function lower-bounds the expected outcome of
Alд and the denominator is the expected outcome of Opt λ . Thus, the objective function is a lower
bound on the competitive ratio of Alд.,
It remains only to verify inequality 3. By definition of Algpot as an algorithm maximizing potential,
we know that Pot(T ) ≤ Pot(X ). However, this does not imply the inequality directly. We will show
that if the inequality does not hold, we will be able to construct a new allocation with a larger
potential, contradicting the definition of X . Suppose by way of contradiction that there exists some
T , and some positive number δ such that Pot X (T ) − Pot(X ) ≥ δ . For some arbitrary small constant
η, consider the vector ηT + (1 − η)X , which allocates η fraction of each stochastic item according to
T and 1 − η fraction of each stochastic item according to X . By concavity of Pot X (.) we have
Pot X (ηT + (1 − η)X ) ≥ ηPot X (T ) + (1 − η)Pot X (X ) = ηPot X (T ) + (1 − η)Pot(X ).
Together with the assumption that Pot X (T ) − Pot(X ) ≥ δ , this gives Pot X (ηT + (1 − η)X ) ≥
Pot(x) + ηδ . On the other hand, Lemma 3.15 implies that Pot X (ηT + (1 − η)X ) ≤ Pot(x) + nη 2 .
By setting η to be smaller than δn , we obtain a contradiction. Thus, for any solution T we have
Pot X (T ) ≤ Pot(X ) which is the inequality 3.
□
Lemma 3.18. The solution of Mathematical Program 4 is lower-bounded by the curve in Figure 3.
Proof. Again, as in the unweighted case, given any solution to Mathematical program 4, we
find a new solution with a restricted structure that only increases the objective function by O(δ ).
Eventually, we lower bound the objective value of this new solution; this provides a lower bound
on the original competitive ratio.
Our restricted solution will satisfy properties 3.19, 3.20, 3.21 and 3.22 below. Similar to the
unweighted case, we can replace each variable with with O(1/δ ) copies with the same value in
order to satisfy Property 3.19, which implies that if we change a constant number of the y j and / or
t j variables, it changes the objective function by at most O(δ ).
Property 3.19. For any positive number δ , we can create an equivalent solution such that Ín 1 y j ≤ δ
and
Ín1
j =1 t j
≤ δ.
j=1
Again as in the unweighted section, we call an agent j with positive z j a harmful agent. We say
that an agent j is a source if both z j and t j are zero. If an agent is neither harmful nor a source, we
call it a sink.
Property 3.20. All y j s and t j s are either 0 or 1.
To satisfy this property, we add some new extra agents with x . = 0, y . = 0, z . = 0, t . = 0 initially.
(Note that this requires slightly modifying the mathematical program to add constraints for the
new agents, and including them in the summations. However, it is easy to see that this does not
change the objective or feasibility of any constraint.)
Now, for any original agent j with positive y j and such that j is either a sink or a source, we
decrease y j to 0 and increase yk by the same amount for some newly added agent k. It is easy to
see that this does not affect the sum of the y variables (and hence the objective function or the
feasibility of any constraint), and that this can be done so that only a single agent has a fractional
value of y.
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
1:18
Hossein Esfandiari, Nitish Korula, and Vahab Mirrokni
Let Γ be an integer such that the sum of t j s for all sinks is in the range [Γ, Γ + 1). (Note that
sources have t j = 0.) We redistribute the t j values to be 1 on the Γ sources and sinks with the
maximum x j s, leaving at most one sink with fractional t j . Using the concavity of Pot X (T ), this
change only decreases Pot X (T ). Since all sources and sinks now have y j = 0, we still have t j +y j ≤ 1
for each agent j. (Note that some agents may change status from source to sink or vice versa during
this process.)
Now, y j and t j are either 0 or 1 for all except the harmful agents. (Further, y j is 1 only on newly
added agents, and t j is 1 only on sink agents.)
For the harmful agents, let YH and TH denote the sum of their y j and t j values respectively. From
constraint (4), we know that YH +TH is less than the total number of harmful agents. We redistribute
the y j and t j values as follows: Set y j = 1 on the ⌊YH ⌋ harmful agents with the smallest values of x j ,
and 0 otherwise; similarly, set t j = 1 on the ⌊TH ⌋ harmful agents with the largest values of x j , and
0 otherwise. In this process, we lose the fractional y j and t j values of at most one agent, possibly
increasing the competitive ratio by O(δ ). Further, note that this process leaves some of these agents
harmful (those with y j = 1), while the others (all but ⌊YH ⌋ of them) are no longer harmful; ⌊TH ⌋
of them now become sinks. There are two further effects of this move: First, as argued above for
the sinks, by setting t j = 1 on the agents with highest x j , Pot X (T ) only decreases. Second, the
redistribution of y j values affects the z j values. To satisfy constraint (2), we must decrease z j to 0
on each agent which is no longer harmful. However, we can increase the z j values on the the ⌊YH ⌋
agents which are still harmful, and it is easy to see that this only decreases the objective function
(because the increase in these z j values exceeds the decrease on the agents which are no longer
harmful).
Therefore, we satisfy Property 3.20 while increasing the objective by at most O(δ ).
For each harmful vertex i, Inequality (2) says that z j ≤ x j (since y j = 1). Increasing z j to x j only
decreases the objective function. Thus, in the rest of the proof, for any harmful vertex j we assume
z j = x j . Changing x j of a harmful vertex j implicitly affects z j as well.
Property 3.21. All harmful agents have the same x j , denoted by α 1 . All sink agents have the same
x j , denoted by α 2 .
We replace the x j of all harmful agents with their average. By concavity of the potential function,
this
Pot(X ). (Note that all harmful agents j have t j = 0.) It also does not change
Ín does notÍdecrease
n
x
and
z
,
and
thus, keeps all the constraints feasible.
j=1 j
j=1 j
Í
We also replace the x j of sinks with their average. One can rewrite Inequality 3 as nj=1 (Pot X (t j )−
Pot(x j )) ≤ 0. The left term for sink j is Pot(x j ) + (1 − x j )(1 − e x j −1 ) − Pot(x j ) = (1 − x j )(1 − e x j −1 ).
Thus, for each sink, the left term is a convex function. This means that this change does not increase
Í
Í
the left hand side of nj=1 (Pot X (t j ) − Pot(x j )) ≤ 0. This change also does not affect nj=1 x j and
Ín
j=1 z j , and hence keeps all the constraints feasible.
We use β 1 to denote the number of agents j with y j = 1; this includes harmful agents and newly
added agents. We use β 2 and β 3 to denote the number of sinks and sources respectively. We use γ
Í
to denote j x j for sources.
Property 3.22.
i is source Pot(x i )
Í
is bounded by (1 − e1 )γ − e1 β 3
This follows from the fact that Pot(x) is always less than (1 − e1 )x − e1 .
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
Online Allocation with Traffic Spikes: Mixing Adversarial and Stochastic Models
1:19
Now, using Properties 3.20, 3.21 and 3.22, one can show that the objective function is lower
bounded by
(α 1 β 1 + α 2 β 2 + γ ) + (1 − e1 )(1 − α 1 )β 1
.
(7)
β1 + β2
Í
(This is a lower bound because in the numerator, we have ( y j − z j ), which is 1 for the newly
added agents included in β 1 , and (1 − α 1 ) for the harmful agents.)
We can write Equality (5) as
λ(β 1 + β 2 ) = β 2 ,
(8)
and write Inequality (3) as
β 1 (0 − e −1 ) + β 2 (α 2 − e α 2 −1 + (1 − α 2 )(1 − e α 2 −1 )) + β 3 (0 − e −1 ) ≤
1
β 1 (α 1 − e α 1 −1 ) + β 2 (α 2 − e α 2 −1 ) + (1 − )γ + β 3 (0 − e −1 ).
e
One can drop the terms β 2 (α 2 − e α 2 −1 ) and β 3 (0 − e −1 ) from both sides of the inequality, and
rearrange as follows:
e
1
(β 1 (− − α 1 + e α 1 −1 ) + β 2 (1 − α 2 )(1 − e α 2 −1 )) ≤ γ .
e −1
e
We replace γ in the objective function (Equation 7 with the max of zero and left hand side of
the inequality. Then, we replace β 1 with 1−λ
λ β 2 using Equality (8). Now we can cancel out β 2 and
simplify the objective function as follows.
α1
1
(1 − λ) + (1 − λ)(1 − ) + λα 2 +
e
e
e
1
max(0,
((1 − λ)(− − α 1 + e α 1 −1 ) + λ(1 − α 2 )(1 − e α 2 −1 )))
e −1
e
Unfortunately, it is hard to solve this analytically to obtain a closed-form experssion for the
competitive ratio as a function of λ. When λ ≥ 0.6882, we can lower bound the competitive ratio
by the line 0.3714 + 0.4362λ. However, this is not necessarily tight. Numerically solving, the curve
lower bounding the competitive ratio as a function of λ can be seen in Figure 3
□
3.3
Hardness of Robust Budgeted Allocation
Theorem 3.23. No online algorithm for Robust Budgeted Allocation (even in the unweighted case)
has competitive ratio better than 1 − e1−λ
1−λ for λ-stochastic inputs.
Proof. Consider the following instance: We have n agents each with capacity one and n items.
Initially, all of the agents are unmarked. Each item is connected to all agents which are unmarked
upon its arrival. The first λn items are connected to all agents. After the arrival of the first λn items,
we mark the λn agents with minimum expected load. Subsequently, after assigning each item we
mark the agent with the minimum expected load among all unmarked agents.
Note that all of the first λn items are adjacent to all agents, and there is no uncertainty about
them. Thus, one can consider the first λn items as being part of the forecast F = (D, λn), where
the distribution D has a single type (items adjacent to all agents). The last (1 − λ)n vertices as
adversarially chosen items. Let v j denote the jth agent that we mark. An optimum algorithm that
knows the whole graph in advance can match the jth item to the jth agent and obtain a matching
of size n.
Now consider an arbitrary online algorithm. Let ℓij be the expected load of agent v j after assigning
the ith item and let ℓj = ℓnj be its expected load at the end of the algorithm. One can see that
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
1:20
Hossein Esfandiari, Nitish Korula, and Vahab Mirrokni
Ín
Maximize:
Subject to:
j=1 ℓ j
n
λn
Õ
ℓ j ≤ λ 2n
j=1
ℓj ≤
j−
Íj−1
ℓ
k =1 k
n−j+1
0 ≤ ℓj ≤ 1
∀ λn + 1 ≤ j ≤ n
∀1≤j ≤n
Fig. 5. Linear program to bound the competitive ratio of the best online algorithm.
ℓj = ℓjmax(λn, j) . The first λn items put a load of λn on the agents and agents v 1 to v λn are the λn
agents with lowest expected loads. Thus, we have
λn
Õ
ℓj ≤ λ2n.
j=1
In addition, for each λn < j ≤ n, the first j items put a load of at most j on the agents, and the
Í
expected load on the agents v 1 to v j−1 is kj−1
ℓ . Therefore, we can bound the expected load of
=1 k
the jth agent as follows:
Í
j − kj−1
ℓ
=1 k
.
ℓj ≤
n−j+1
Further, the expected load of an agent cannot exceed 1, and the optimum solution is a matching
in which each agent has a load of 1. Therefore, the competitive ratio of any algorithm for Robust
Budgeted Allocation on λ-stochastic inputs is bounded by Linear program 5.
First note that w.l.o.g., we can assume that in any optimal solution to LP 5, each ℓj for 1 ≤ j ≤ λn
has the same value. Consider an optimal solution in which as many constraints as possible are
tight. It is easy to see that if the constraint for ℓj is not tight, we can raise ℓj by ε and decrease ℓk by
at most ε (to maintain feasibility) for the first k > j such that ℓk > 0. By repeating this procedure
iteratively, we make all of the constraints tight. This solution is correspond to the algorithm that
matches items to their eligible agents uniformly. Hence, it allocates all λn stochastic items and at
Íξ n
1
most ξn adversarial items, for some ξ such that i=1 (1−λ)n−(i−1)
≥ 1 − λ. Notice that we have
ξn
Õ
i=1
1
≥
(1 − λ)n − (i − 1)
∫
(1−λ)n
(1−λ−ξ )n
1
1−λ
dx = log((1 − λ)n) − log((1 − λ − ξ )n) = log(
).
x
1 − (λ + ξ )
1−λ
By setting log( 1−(λ+ξ
) ) = 1 − λ and rearranging the parameters we have λ + ξ = 1 −
4
1−λ
.
e 1−λ
□
APPROXIMATING ADVERSARIAL AND STOCHASTIC BUDGETED ALLOCATION
In this section, we study a class of algorithms for Budgeted Allocation problem that provide good
approximation ratios in both stochastic and adversarial settings. We say an algorithm is (α, β)competitive if it is α-competitive in the adversarial setting and β-competitive in the stochastic
setting.
The best algorithms for the stochastic setting have a competitive ratio of 1 − ε when the input
is stochastic. However, these algorithms may have a competitive ratio close to zero when the
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
Online Allocation with Traffic Spikes: Mixing Adversarial and Stochastic Models
1:21
input is adversarial. On the other hand, the Balance algorithm has the best possible competitive
ratio of 1 − e1 when the input is adversarial, but, under some mild assumption, is 0.76-competitive
for stochastic inputs [Mirrokni et al. 2011]. Indeed, this result holds when, 1) the capacities are
Í
w
large i.e. maxi, j ci,j j ≤ ε 3 , and 2) the optimum solution is large i.e. ε −6 j maxi:opt (i)=j w i, j ≤
Opt, where Opt(i) is the agent the optimum matches to node i, and ε is a vanishingly small
w
number. In this section, we refer to maxi, j ci,j j ≤ ε 3 as the large capacities condition, and refer to
Í
max ε −6 j maxi:opt (i)=j w i, j , ε 1/3m · maxi, j w i, j ≤ Opt as the large optimum condition.
Of course, one can design algorithms with ‘intermediate’ performance: Simply use the stochastic
algorithm with probability p and Balance with probability 1 − p; this yields an algorithm with
competitive ratios of ((1 − p)(1 − e1 ) and (p + (1 − p) × 0.76) − O(ε)) in the adversarial and stochastic
settings respectively. These competitive ratios for different values of p lie on the straight line with
endpoints (1 − e1 , 0.76) and (0, 1 − ε). In this section, we use the tools that we developed in Section 2
to obtain competitive ratios those are ‘above’ this straight line.
We devise a mixed algorithm for this setting as follows: Divide the capacity of each agent /
offline vertex into two parts. The first part has 1 − p fraction of the capacity, and the second part
has the remaining p fraction of the capacity. Given a forecast F = (D, f ), at the beginning, run the
Balance algorithm on 1 − p fraction of the forecast input (that is, on the first (1 − p) · f items, using
the first (1 − p) fraction of the capacities. Then, for any 1 − ε-competitive stochastic algorithm
Alg, run Alg on the rest of the forecast input (that is, the next p f items), using the remaining p
fraction of the capacities. If at any time during the algorithm, we observe that the input sequence
is ε-far from D, or the length of the input is more than f , we can detect that (w.h.p.) we are in
the adversarial setting. We therefore flush out the items we assigned to the second part (the p
fraction) of the capacities, if any, and return to the Balance algorithm. However, we now run two
independent copies of Balance: The first using 1 − p fraction of the capacities (with some items
previously allocated by Balance), and the second using the remaining p fraction of the capacities
(beginning with an empty allocation of items to this part of the capacities). For each item arriving
online, with probability 1 − p we send it to the Balance algorithm on the first part of the capacities,
and with probability p we send it to the Balance algorithm on the second part of the capacities.
In the next two lemmas, we bound the competitive ratios of our mixed algorithm when the input
is stochastic and adversarial respectively.
Lemma 4.1. Assuming the long input, the large capacities and the large optimum conditions, if
the input is stochastic, for any constant p ∈ [0, 1] the mixed algorithm has a competitive ratio of
(1 − p) ∗ β + p − O(ε), where β is 0.76.
Proof. From Theorem 2.11, we know that by ignoring the inputs which are ε-far from the
distribution, we lose only O(ε) fraction in the competitive ratio. We use 1 − p fraction of the input
sequence on the first 1 −p fraction of the capacities. Therefore, the Balance algorithm gets (1 −p) ∗ β
fraction of the optimal solution from this part of the input, where β = 0.76 is the competitive ratio
for Balance on stochastic inputs [Mirrokni et al. 2011]. On the other hand, we use p fraction of the
input on the remaining p fraction of the capacities. Since we use a 1 − ε competitive stochastic
algorithm, we get p − O(ε) fraction of the optimal solution from this part of the input sequence. □
Lemma 4.2. Assuming the long input, the large capacities and the large optimum conditions, if
the input is adversarial, for any constant p ∈ [0, 1] the mixed algorithm has a competitive ratio of
0.76(1−p)
1/3 ).
1+0.202(1−p) − O(ε
Proof. Let S be the maximal prefix of the input before observing that the input is not ε-close
to the distribution and let A be the rest of the input. (Note that either S or A may be empty.) If
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
1:22
Hossein Esfandiari, Nitish Korula, and Vahab Mirrokni
ε
S contains at least ϵ 2/3m nodes, since S is ϵ 2/3
= ϵ 1/3 -close to the distribution and satisfies the
long input condition, we obtain at least β(1 − p − O(ϵ 1/3 ))Opt(S) from this fraction of the input
(Note that if S is smaller than (1 − p) · f , for example, we would obtain a larger fraction of Opt(S).).
On the other hand, if S contains less than ϵ 2/3m nodes, using large optimum condition, we have
Opt(S) ≤ ϵ 2/3m · max i, j w i, j ≤ ϵ 1/3Opt(A ∪ S). Thus, the competitive ratio is lower bounded by
β(1 − p)Opt(S)
β(1 − p)Opt(S)
− O(ϵ 1/3 ) ≥
− O(ϵ 1/3 ).
Opt(S ∪ A)
Opt(S) + Opt(A)
On the other hand, if we ignore ϵm nodes, the Balance algorithm on the first 1 − p fraction of
the capacities always runs on at least 1 − p fraction of the items, we always get at least (1 − e1 )(1 −
p − O(ϵ 2/3 ))Opt(S ∪ A) from this fraction of the capacities. Further, once the input stops being
ε-close to the distribution (that is, for the entire sequence A), we use the remaining p fraction of the
capacities on p fraction of A. Therefore, we always get at least (1 − e1 )pOpt(A) from this p fraction
of the capacities. Thus, the competitive ratio is lower bounded by
(1 − e1 )(1 − p − O(ϵ 2/3 ))Opt(S ∪ A) + (1 − e1 )pOpt(A)
=
Opt(S ∪ A)
(1 − e1 )pOpt(A)
1
1−
(1 − p) +
− O(ϵ 2/3 ) ≥
e
Opt(S ∪ A)
(1 − e1 )pOpt(A)
1
(1 − p) +
− O(ϵ 2/3 )
1−
e
Opt(S) + Opt(A)
where the inequality follows from the subadditivity of Budgeted Allocation.
The first of these lower bounds is decreasing in Opt(A), while the second lower bound is increasing
in Opt(A). Thus, the worst case happen when these two bounds are equal. By setting the two equal
and β = 0.76, one can achieve the desired bound.
□
The following theorem is the immediate result of combining Lemmas 4.1 and 4.2.
Theorem 4.3. Assuming the long input, the large capacities and the large optimum conditions, for
any constant p ∈ [0, 1] the mixed algorithm is (1 − p) ∗ 0.76 + p − O(ε) competitive for stochastic input
0.76(1−p)
and 1+0.202(1−p) − O(ε 1/3 ) competitive for adversarial input.
REFERENCES
Shipra Agrawal, Zizhuo Wang, and Yinyu Ye. 2009. A Dynamic Near-Optimal Algorithm for Online Linear Programming.
Computing Research Repository (2009).
Saeed Alaei, MohammadTaghi Hajiaghayi, and Vahid Liaghat. 2012. Online prophet-inequality matching with applications
to ad allocation. In Proceedings of the 13th ACM Conference on Electronic Commerce. ACM, 18–35.
David Applegate and Edith Cohen. 2006. Making routing robust to changing traffic demands: algorithms and evaluation.
IEEE/ACM Transactions on Networking (TON) 14, 6 (2006), 1193–1206.
Yossi Azar, Edith Cohen, Amos Fiat, Haim Kaplan, and Harald Racke. 2003. Optimal oblivious routing in polynomial time.
In Proceedings of the thirty-fifth annual ACM symposium on Theory of computing. ACM, 383–388.
Bahman Bahmani and Michael Kapralov. 2010. Improved bounds for online stochastic matching. In European Symposium on
Algorithms. Springer, 170–181.
Aharon Ben-Tal and Arkadi Nemirovski. 2002. Robust optimization–methodology and applications. Mathematical Programming 92, 3 (2002), 453–480.
Dimitris Bertsimas, Dessislava Pachamanova, and Melvyn Sim. 2004. Robust linear optimization under general norms.
Operations Research Letters 32, 6 (2004), 510–516.
Niv Buchbinder, Kamal Jain, and Joseph Seffi Naor. 2007. Online primal-dual algorithms for maximizing ad-auctions revenue.
In ESA. Springer, 253–264.
Nikhil Devanur and Thomas Hayes. 2009. The Adwords Problem: Online Keyword Matching with Budgeted Bidders under
Random Permutations. In EC. 71–78.
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
Online Allocation with Traffic Spikes: Mixing Adversarial and Stochastic Models
1:23
Nikhil R. Devanur, Kamal Jain, Balasubramanian Sivan, and Christopher A. Wilkens. 2011. Near optimal online algorithms
and fast approximation algorithms for resource allocation problems. In EC. ACM, 29–38.
Nikhil R Devanur, Balasubramanian Sivan, and Yossi Azar. 2012. Asymptotically optimal algorithm for stochastic adwords.
In Proceedings of the 13th ACM Conference on Electronic Commerce. ACM, 388–404.
Jon Feldman, Monika Henzinger, Nitish Korula, Vahab S. Mirrokni, and Cliff Stein. 2010. Online stochastic packing applied
to display ad allocation. In ESA. Springer, 182–194.
J. Feldman, N. Korula, V. Mirrokni, S. Muthukrishnan, and M. Pal. 2009a. Online Ad Assignment with Free Disposal. In
WINE.
Jon Feldman, Aranyak Mehta, Vahab Mirrokni, and S. Muthukrishnan. 2009b. Online Stochastic Matching: Beating 1 - 1/e.
In FOCS. 117–126.
B. Haeupler, V. Mirrokni, and M. ZadiMoghaddam. 2011. Online Stochastic Weighted Matching: Improved Approximation
Algorithms. In WINE. 170–181.
Patrick Jaillet and Xin Lu. 2013. Online stochastic matching: New algorithms with better bounds. Mathematics of Operations
Research 39, 3 (2013), 624–646.
Bala Kalyanasundaram and Kirk R. Pruhs. 2000. An optimal deterministic algorithm for online b -matching. Theoretical
Computer Science 233, 1–2 (2000), 319–325. citeseer.ist.psu.edu/8238.html
Chinmay Karande, Aranyak Mehta, and Pushkar Tripathi. 2011. Online Bipartite Matching with Unknown Distributions. In
STOC. 587–596.
R.M. Karp, U.V. Vazirani, and V.V. Vazirani. 1990. An optimal algorithm for online bipartite matching. In STOC. 352–358.
Mohammad Mahdian, Hamid Nazerzadeh, and Amin Saberi. 2007. Allocating online advertisement space with unreliable
estimates. In Proceedings of the 8th ACM conference on Electronic commerce. ACM, 288–294.
Mohammad Mahdian and Qiqi Yan. 2011. Online Bipartite Matching with Random Arrivals: A Strongly Factor Revealing LP
Approach. In STOC. 597–606.
Aranyak Mehta, Amin Saberi, Umesh Vazirani, and Vijay Vazirani. 2007. AdWords and generalized online matching. J. ACM
54, 5 (2007), 22. https://doi.org/10.1145/1284320.1284321
Vahideh H. Menshadi, Shayan Oveis Gharan, and A. Saberi. 2011. Online Stochastic Matching: Online Actions Based on
Offline Statistics. In SODA. 1285–1294.
Vahab S. Mirrokni, Shayan Oveis Gharan, and Morteza ZadiMoghaddam. 2011. Simultaneous Approximations of Stochastic
and Adversarial Budgeted Allocation Problems. In SODA. 1690–1701.
Bo Tan and R Srikant. 2011. Online advertisement, optimization and stochastic networks. In CDC-ECC. IEEE, 4504–4509.
Erik Vee, Sergei Vassilvitskii, and Jayavel Shanmugasundaram. 2010. Optimal Online Assignment with Forecasts. In EC.
109–118.
Hao Wang, Haiyong Xie, Lili Qiu, Yang Richard Yang, Yin Zhang, and Albert Greenberg. 2006. COPE: traffic engineering in
dynamic networks. ACM SIGCOMM 36, 4 (2006), 99–110.
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
| 8cs.DS |
Spectral Method and Regularized MLE Are Both
Optimal for Top-K Ranking
Yuxin Chen∗
Jianqing Fan†
arXiv:1707.09971v2 [stat.ML] 28 Sep 2017
August 2017;
Cong Ma†
Kaizheng Wang†
Revised September 2017
Abstract
This paper is concerned with the problem of top-K ranking from pairwise comparisons. Given a
collection of n items and a few pairwise binary comparisons across them, one wishes to identify the set
of K items that receive the highest ranks. To tackle this problem, we adopt the logistic parametric
model—the Bradley-Terry-Luce model, where each item is assigned a latent preference score, and where
the outcome of each pairwise comparison depends solely on the relative scores of the two items involved.
Recent works have made significant progress towards characterizing the performance (e.g. the mean
square error for estimating the scores) of several classical methods, including the spectral method and
the maximum likelihood estimator (MLE). However, where they stand regarding top-K ranking remains
unsettled.
We demonstrate that under a random sampling model, the spectral method alone, or the regularized
MLE alone, is minimax optimal in terms of the sample complexity—the number of paired comparisons
needed to ensure exact top-K identification. This is accomplished via optimal control of the entrywise
error of the score estimates. We complement our theoretical studies by numerical experiments, confirming that both methods yield low entrywise errors for estimating the underlying scores. Our theory is
established based on a novel leave-one-out trick, which proves effective for analyzing both iterative and
non-iterative optimization procedures. Along the way, we derive an elementary eigenvector perturbation
bound for probability transition matrices, which parallels the Davis-Kahan sin Θ theorem for symmetric
matrices. This further allows us to close the gap between the `2 error upper bound for the spectral
method and the minimax lower limit.
Keywords. top-K ranking, pairwise comparisons, spectral method, regularized MLE, perturbation
analysis, Erdős–Rényi random graph, leave-one-out analysis, reversible Markov chain.
1
Introduction
Imagine we have a large collection of n items, and we are given partially revealed comparisons between pairs
of items. These paired comparisons are collected in a non-adaptive fashion, and could be highly noisy and
incomplete. The aim is to aggregate these partial preferences so as to identify the K items that receive the
highest ranks. This problem, which is called top-K rank aggregation, finds applications in numerous contexts,
including web search (Dwork et al., 2001), recommendation systems (Baltrunas et al., 2010; Park et al., 2015),
sports competition (Masse, 1997), and so on. The challenge is both statistical and computational: how can
one achieve reliable top-K ranking from a minimally possible number of pairwise comparisons, while retaining
computational efficiency?
1.1
Popular approaches
To address the aforementioned challenge, many prior approaches have been put forward based on certain
statistical models. Arguably one of the most widely used parametric models is the Bradley-Terry-Luce
Author names are sorted alphabetically.
of Electrical Engineering, Princeton University, Princeton, NJ 08544, USA; Email: yuxin.chen@princeton.edu.
† Department of Operations Research and Financial Engineering, Princeton University, Princeton, NJ 08544, USA; Email:
{jqfan, congm, kaizheng}@princeton.edu.
∗ Department
1
(BTL) model (Bradley and Terry, 1952; Luce, 1959; Agresti, 2013), which assigns a latent preference score
wi∗ (1 ≤ i ≤ n) to each of the n items. The BTL model posits that: the chance of each item winning a paired
comparison is determined by the relative scores of the two items involved, or more precisely,
P {item j is preferred over item i} =
wj∗
wi∗ + wj∗
(1)
in each comparison of item i against item j. The items are repeatedly compared in pairs according to this
parametric model. The task then boils down to identifying the K items with the highest preference scores,
given these pairwise comparisons.
Among the ranking algorithms tailored to the BTL model, there are two procedures that have received
particular attention, both of which rank the items in accordance with appropriate estimates of the latent
preference scores:
1. The spectral method. By connecting the winning probability in (1) with the transition probability of
a reversible Markov chain, the spectral method attempts recovery of {wi∗ }1≤i≤n via the leading left
eigenvector of a sample transition matrix. This spectral ranking procedure, which is also known as
Rank Centrality (Negahban et al., 2017a), bears some similarity to the PageRank algorithm (Page
et al., 1999).
2. The maximum likelihood estimator (MLE). This approach proceeds by finding the score assignment
that maximizes the likelihood function (Ford, 1957). When parameterized appropriately, solving the
MLE becomes a convex program, and hence is computationally feasible. There are also important
variants of the MLE that enforce additional regularization; see Negahban et al. (2017a).
Details are postponed to Section 2.2. In addition to their remarkable practical applicability, these two
ranking paradigms are appealing in theory as well. For instance, both of them provably achieve near-optimal
mean square errors when estimating the latent preference scores (Negahban et al., 2017a; Hajek et al., 2014).
Nevertheless, the `2 error for estimating the latent scores merely serves as a “meta-metric” for the ranking
task, which does not necessarily reveal the accuracy of top-K identification. Given that the `2 loss only
reflects the estimation error in some average sense, it is certainly possible that an algorithm obtains minimal
`2 estimation loss but incurs (relatively) large errors when estimating the scores of the highest ranked items.
Interestingly, a recent work Chen and Suh (2015) demonstrates that: a careful combination of the spectral
method and the coordinate-wise MLE is optimal for top-K ranking. This leaves open the following natural
questions: where does the spectral alone, or the MLE alone, stand in top-K ranking? Are they capable of
attaining exact top-K recovery from minimal samples? These questions are poorly understood and form the
primary objectives of our study.
As we will elaborate later, the spectral method part of the preceding question has recently been explored
in an interesting paper Jang et al. (2016), for a regime where a relatively large fraction of item pairs have been
compared. However, it remains unclear how well these two algorithms can perform in a much broader—and
often more challenging—regime.
1.2
Main contributions
The central focal point of the current paper is to assess the accuracy of both the spectral method and the
regularized MLE in top-K identification. Assuming that the pairs of items being compared are randomly
selected, our paper delivers a somewhat surprising message:
Both the spectral method and the regularized MLE achieve perfect identification of top-K ranked
items under optimal sample complexity (up to some constant factor)!
It is worth emphasizing that these two algorithms work even under the sparsest possible regime, a
scenario where only an exceedingly small fraction of pairs of items have been compared. This calls for
precise control of the entrywise error—as opposed to the `2 loss—for estimating the scores. To this end,
our theory is established upon a novel leave-one-out argument, which might shed light on how to analyze
the entrywise error for more general optimization problems. As a byproduct of the analysis, we derive an
2
elementary eigenvector perturbation bound for probability transition matrices, which parallels Davis-Kahan’s
sin Θ theorem for symmetric matrices. This simple perturbation bound leads to an improved `2 error bound
for the spectral method, which matches the minimax lower limit derived in the prior literature.
Finally, we remark that both methods can be performed in time nearly proportional to reading all
samples. In fact, the near-optimal computational cost of the spectral method has already been demonstrated
in Negahban et al. (2017a). When it comes to the regularized MLE, we will show that this method—which
only exhibits a fairly weak level of strong convexity—can be computed by the gradient descent in nearly linear
time. In summary, both methods provably achieve optimal efficiency from both statistical and computational
perspectives.
1.3
Notation
Before continuing, we introduce a few notations that will be useful throughout. To begin with, vectors and
matrices are in boldface whereas scalars are not.
For any strictly positive probability vector π ∈ Rn , we define the inner product space indexed by π as a
real vector space in Rn endowed with the inner product (Brémaud, 1999)
Xn
hx, yiπ =
πi xi yi .
i=1
The corresponding vector norm and the induced matrix norm are defined respectively as
q
and
kAkπ = sup kx> Akπ .
kxkπ = hx, xiπ
kxkπ =1
If 0 < πmin ≤ πi ≤ πmax for all 1 ≤ i ≤ n, then one has the following elementary inequalities
r
r
√
√
πmin
πmax
kAk2 ≤ kAkπ ≤
kAk2 .
πmin kxk2 ≤ kxkπ ≤ πmax kxk2
and
πmax
πmin
Additionally, the standard notation f (n) = O (g(n)) or f (n) . g(n) means that there exists a constant
c > 0 such that |f (n)| ≤ c|g(n)|, f (n) = Ω (g(n)) or f (n) & g(n) means that there exists a constant c > 0
such that |f (n)| ≥ c |g(n)|, f (n) = Θ (g(n)) or f (n) g(n) means that there exist constants c1 , c2 > 0 such
(n)
that c1 |g(n)| ≤ |f (n)| ≤ c2 |g(n)|, and f (n) = o(g(n)) means that limn→∞ fg(n)
= 0.
P
>
Given a graph G with vertex set {1, 2, . . . , n} and edge set E, we denote by LG = (i,j)∈E,i>j (ei − ej ) (ei − ej )
the (unnormalized) Laplacian matrix (Chung, 1997) associated with it, where {ei }1≤i≤n are the standard
basis vectors in Rn . For a matrix A ∈ Rn×n with n real eigenvalues, we let λ1 (A) ≥ λ2 (A) ≥ · · · ≥ λn (A)
be the eigenvalues sorted in descending order.
2
2.1
Statistical models and main results
Problem setup
We begin with a formal introduction of the Bradley-Terry-Luce parametric model for binary comparisons.
Preference scores. As introduced earlier, we assume the existence of a positive latent score vector
w∗ = [w1∗ , · · · , wn∗ ]>
(2)
that comprises the underlying preference score {wi∗ }1≤i≤n assigned to each of the n items. Alternatively,
it is sometimes more convenient to reparameterize the score vector by
θ ∗ = [θ1∗ , · · · , θn∗ ]> ,
θi∗ = log wi∗ .
(3)
These scores are assumed to fall within a fixed dynamic range, namely,
wi∗ ∈ [wmin , wmax ],
or
3
θi∗ ∈ [θmin , θmax ]
(4)
for all 1 ≤ i ≤ n, where wmin > 0, wmax > 0, θmin = log wmin , and θmax = log wmax are some absolute
constants independent of n. Without loss of generality, it is assumed that
wmax ≥ w1∗ ≥ w2∗ ≥ . . . ≥ wn∗ ≥ wmin ,
(5)
meaning that items 1 through K are the desired top-K ranked items. For notational simplicity, we
define
wmax
,
(6)
κ :=
wmin
which is assumed to be a constant independent of n.
Comparison graph. Let G = (V, E) stand for a comparison graph, where the vertex set V = {1, 2, . . . , n}
represents the n items of interest. The items i and j are compared if and only if (i, j) falls within the
edge set E. Unless otherwise noted, we assume that G is drawn from the Erdős–Rényi random graph
Gn,p (Durrett, 2007), such that an edge between any pair of vertices is present independently with some
probability p. In words, p captures the fraction of item pairs being compared. In order to ensure that the
comparison graph is connected (otherwise there is no basis to rank items belonging to two disconnected
components of G), we assume throughout that
p≥
c0 log n
n
(7)
for some constant c0 > 1.
Pairwise comparisons. For each (i, j) ∈ E, we obtain L independent paired comparisons between items i
(l)
and j. Let yi,j be the outcome of the `-th comparison, which is independently drawn as
∗
1, with probability wj∗ = eθj ∗ ,
(l) ind.
θ
θ∗
wi∗ +wj∗
e i +e j
yi,j =
(8)
0, else.
(l)
(l)
By convention, we set yi,j = 1 − yj,i for all (i, j) ∈ E throughout the paper. This is also known as the
logistic pairwise comparison model, due to its strong resemblance to logistic regression (Agresti, 2013).
It is self-evident that the sufficient statistics under this model are given by
L
y := {yi,j | (i, j) ∈ E} ,
yi,j :=
1 X (l)
yi,j .
L
(9)
l=1
To simplify the notation, we shall also take
∗
∗
yi,j
wj∗
eθ j
:= ∗
=
∗ .
∗
wi + wj∗
eθi + eθj
(10)
Goal. The goal is to identify the set of top-K ranked items—that is, the set of K items that exhibit the
largest preference scores—from the pairwise comparisons y.
2.2
2.2.1
Algorithms
The spectral method: Rank Centrality
The spectral ranking algorithm, or Rank Centrality (Negahban et al., 2017a), is motivated by the connection
between the pairwise comparisons and a random walk over a directed graph. The algorithm starts by
converting the pairwise comparison data y into a transition matrix P = [Pi,j ]1≤i,j≤n in such a way that
1
if (i, j) ∈ E,
d yi,j , P
1
Pi,j = 1 − d k:(i,k)∈E yi,k , if i = j,
(11)
0,
otherwise,
4
Algorithm 1 Spectral method (Rank Centrality).
Input the comparison graph G, the sufficient statistics y, and the normalization factor d.
Define the probability transition matrix P = [Pi,j ]1≤i,j≤n as in (11).
Compute the leading left eigenvector π of P .
Output the K items that correspond to the K largest entries of π.
for some given normalization factor d, and then proceeds by computing the stationary distribution π ∈ Rn
of the Markov chain induced by P . As we shall see later, the parameter d is taken to be on the same order
of the maximum vertex degree of G while ensuring the non-negativity of P . As asserted by Negahban et al.
(2017a), π is a faithful estimate of w∗ up to some global scaling. The algorithm is summarized in Algorithm
1.
To develop some intuition regarding why this spectral algorithm gives a reasonable estimate of w∗ , it is
∗
perhaps more convenient to look at the population transition matrix P ∗ = [Pi,j
]1≤i,j≤n :
wj∗
if (i, j) ∈ E,
d1 wi∗ +wj∗ ,
∗
P
∗
w
Pi,j = 1 − 1 k:(i,k)∈E ∗ k ∗ ,
if i = j,
d
wi +wk
0,
otherwise,
which coincides with P by taking L → ∞. It can be seen that the normalized score vector
π ∗ := Pn
1
∗
i=1 wi
>
[w1∗ , w2∗ , . . . , wn∗ ]
(12)
is the stationary distribution of the Markov chain induced by P ∗ , since P ∗ and π ∗ are in detailed balance
(Brémaud, 1999), namely,
∗
∗
πi∗ Pi,j
= πj∗ Pj,i
,
∀ (i, j) .
(13)
As a result, one expects the stationary distribution of the sample version P to form a good estimate of w∗ ,
provided that the sample size is sufficiently large.
2.2.2
The regularized MLE
Under the BTL model, the negative log-likelihood function conditioned on the graph is given by (up to some
global scaling)
X
eθ j
eθi
+ (1 − yj,i ) log θi
L (θ; y) : = −
yj,i log θi
e + eθ j
e + eθ j
(i,j)∈E,i>j
X
=
−yj,i (θi − θj ) + log 1 + eθi −θj .
(14)
(i,j)∈E,i>j
The regularized MLE (Negahban et al., 2017a, Section 3.3) then amounts to solving the following convex
program
minimizeθ∈Rn
1
Lλ (θ; y) := L (θ; y) + λkθk22 ,
2
(15)
q
n
for a regularization parameter λ > 0. As will be discussed later, we shall adopt the choice λ np log
L
throughout this paper. For the sake of brevity, we let θ represent the resulting penalized maximum likelihood
estimate whenever it is clear from the context.
2.3
Main results
The most challenging part of top-K ranking is to distinguish the K-th and the (K + 1)-th items. In
fact, the score difference of these two items captures the distance between the item sets {1, . . . , K} and
5
{K + 1, . . . , n}. Unless their latent scores are sufficiently separated, the finite-sample nature of the model
would make it infeasible to distinguish these two critical items. With this consideration in mind, we define
the following separation measure
∗
w∗ − wK+1
.
(16)
∆K := K
wmax
This metric turns out to play a crucial role in determining the minimal sample complexity for perfect top-K
identification.
The main finding of this paper concerns the optimality of both the spectral method and the regularized
MLE. Recall that under the BTL model introduced before, the total number N of samples we collect
concentrates sharply around its mean, namely,
1
N = (1 + o(1))E [N ] = (1 + o(1)) n2 pL
2
(17)
occurs with high probability. Our main result is stated in terms of the sample complexity required for exact
top-K identification.
Theorem 1. Consider the pairwise comparison model specified in Section 2.1. Suppose that p >
that
c1 n log n
n2 pL
≥
2
∆2K
c0 log n
n
and
(18)
for some sufficiently large constants c0 , c1 > 0. With probability exceeding 1 − O(n−5 ), the set of top-K
ranked items can be recovered exactly by the spectral method given in Algorithmq1, and by the regularized
MLE given in (15). Here, we take d = cd np in the spectral method and λ = cλ
MLE, where cd ≥ 2 and cλ > 0 are some absolute constants.
np log n
L
in the regularized
Theorem 1 asserts that both the spectral method and the regularized MLE achieve a sample complexity
log n
on the order of n ∆
. Encouragingly, this sample complexity coincides with the minimax limit identified in
2
K
(Chen and Suh, 2015, Theorem 2) up to a constant factor:
Theorem 2 (Chen and Suh (2015)). Fix ∈ (0, 1/2), and suppose that
n2 pL ≤ 2c2
(1 − )n log n − 2
,
∆2K
(19)
4
4
). Then for any ranking procedure ψ, one can find a score vector w∗ with separation
where c2 = wmin
/(4wmax
∆K such that ψ fails to retrieve the top-K items with probability at least .
Moving beyond the sample complexity issue, we remark that the spectral method can be carried out
by means of the power method, while the regularized MLE can be computed via the standard gradient
descent (as we will elaborate in Section 4.2.1). Both procedures are computationally feasible and can be
accomplished within nearly linear time—in time proportional to reading all data. Consequently, these two
algorithms are optimal in terms of both computational complexity and sample complexity. We emphasize
that it is nontrivial to establish the nearly linear-time computational complexity of the regularized MLE,
given that the log-likelihood function (or the logistic link function) is not strongly convex. We will elaborate
on this point in Section 4.2.1.
We are now positioned to compare our results with Jang et al. (2016), which also investigates the accuracy
of the spectral method for top-K ranking. Specifically, (Jang et al., 2016, Theorem 3) establishes the
optimality of the spectral method for the relatively dense regime where
r
log n
p&
.
n
In this regime, however, the total sample size necessarily exceeds
n2 pL
n2 p p 3
≥
& n log n,
2
2
6
(20)
which rules out the possibility of achieving minimal sample complexity if ∆K is sufficiently large. For
instance, consider the case where ∆K 1, then the optimal sample size—as revealed by Theorem 1 or
(Chen and Suh, 2015, Theorem 1)—is on the order of
n log n
n log n,
∆2K
p
which is a factor of n/ log n lower than the bound in (20). In contrast, our results hold all the way down to
the sparsest possible regime where p logn n , confirming the optimality of the spectral method even for the
most challenging scenario. Furthermore, we establish that the regularized MLE shares the same optimality
guarantee as the spectral method, which was previously out of reach.
2.4
Optimal control of entrywise estimation errors
In order to establish the ranking accuracy as asserted by Theorem 1, the key is to obtain precise control of
the `∞ loss of the score estimates. Our results are as follows.
n
Theorem 3 (Entrywise error of the spectral method). Suppose that p > c0 log
for some sufficiently
n
large constant c0 > 0. Choose d = cd np for some constant cd ≥ 2 in Algorithm 1. Then the spectral estimate
π satisfies
s
kπ − π ∗ k∞
log n
(21)
.
kπ ∗ k∞
npL
with probability exceeding 1 − O(n−5 ), where π ∗ is the normalized score vector as defined in (12).
n
Theorem 4 (Entrywise error of the regularized MLE). Suppose that p > c0 log
for some sufficiently
n
q
np log n
large constant c0 > 0, and that the regularization parameter is λ = cλ
for some absolute constant
L
cλ > 0. Then the regularized MLE θ satisfies
s
∗
∗
keθ − eθ −θ 1 k∞
log n
.
(22)
∗ −θ ∗ 1
θ
npL
ke
k∞
∗
with probability exceeding 1 − O(n−5 ), where θ :=
1 > ∗
n1 θ
and eθ := [eθ1 , · · · , eθn ]> .
Theorems 3-4 indicate that if the number of comparisons associated with each item—which concentrates
around npL—exceeds log n to some sufficiently large extent, then both methods are able to achieve a small
`∞ error when estimating the scores.
Recall that the `2 estimation error of the spectral method has been characterized by Negahban et al.
(2017a) (or Theorem 6 of this paper with an improved upper bound), which obeys
s
kπ − π ∗ k2
log n
.
(23)
kπ ∗ k2
npL
with high probability. Similar theoretical guarantees have been derived for another variant of the MLE
(the constrained version) under a uniform sampling model as well (Negahban et al., 2017a). In comparison,
our results indicate that the estimation errors for both algorithms are almost evenly spread out across all
coordinates rather than being localized or clustered. Notably, the pointwise errors revealed by Theorems 3-4
immediately lead to exact top-K identification as claimed by Theorem 1.
Proof of Theorem 1. In what follows, we prove the theorem for the spectral method part. The regularized
MLE part follows from an almost identical argument and hence is omitted.
Since the spectral algorithm ranks the items in accordance with the score estimate π, it suffices to
demonstrate that
πi − πj > 0,
∀1 ≤ i ≤ K, K + 1 ≤ j ≤ n.
7
To this end, we first apply the triangle inequality to get
πj − πj∗
πi∗ − πj∗
πi − πj
|πi − πi∗ |
2kπ − π ∗ k∞
≥
−
−
≥ ∆K −
.
∗
∗
∗
∗
kπ k∞
kπ k∞
kπ k∞
kπ k∞
kπ ∗ k∞
(24)
In addition, it follows from Theorem 3 as well as our sample complexity assumption that
s
kπ − π ∗ k∞
n log n
log n
.
and
n2 pL &
.
∗
kπ k∞
npL
∆2K
These conditions taken collectively imply that
kπ − π ∗ k∞
1
< ∆K
∗
kπ k∞
2
as long as
2.5
npL∆2K
log n
is sufficiently large. Substitution into (24) reveals that πi − πj > 0, as claimed.
Heuristic arguments
We pause here to develop some heuristic explanation as to why the score estimation errors are expected to
be spread out across all entries. For simplicity, we focus on the case where p = 1 and L is sufficiently large,
so that y and P sharply concentrate around y ∗ and P ∗ , respectively. For simplicity of presentation, we
.
introduce the notation to represent asymptotic equivalence up to logarithmic factors.
We begin with the spectral algorithm. Since π and π ∗ are respectively the invariant distributions of the
Markov chains induced by P and P ∗ , we can decompose
>
>
(π − π ∗ ) = π > P − π ∗> P ∗ = (π − π ∗ ) P + π ∗> (P − P ∗ ).
|
{z
}
(25)
:=ξ
∗
∗
∗
max
When p = 1 and w
wmin 1, the entries of π (resp. the off-diagonal entries of P and P − P ) are all of the
same order and, as a result, the energy of the uncertainty term ξ is spread out. In fact, we will demonstrate
in Section 3.2 that
s
kξk∞
log n . kπ − π ∗ k2
.
,
(26)
kπ ∗ k∞
npL
kπ ∗ k2
which coincides with the optimal rate. Further, if we look at each entry of (25), then for all 1 ≤ m ≤ n,
>
∗
πm − πm
= (π − π ∗ ) P m + ξm ,
π1 − π1∗
..
∗
= (πm − πm
) Pm,m + [P1,m , · · · , Pm−1,m , 0, Pm+1,m , · · · , Pn,m ]
+ ξm .
.
∗
πn − πn
|
{z
} |
{z
}
contraction
(27)
error averaging
By construction of the transition matrix, one can easily verify that Pm,m is bounded away from 1 and
∗
Pj,m 1/n for all j 6= m. As a consequence, the identity π > P = π > allows one to treat each πm − πm
as a mixture of three effects: (i) the first term of (27) behaves as a entrywise contraction of the error; (ii)
the second term of (27) is a (nearly uniformly weighted) average of the errors over all coordinates, which
can essentially be treated as a smoothing operator applied to the error components; and (iii) the uncertainty
term ξm . Rearranging terms in (27), we are left with
∗
(1 − Pm,m ) |πm − πm
|.
8
1 Xn
|πi − πi∗ | + ξm ,
i=1
n
(28)
or equivalently,
kπ − π ∗ k∞ .
1 Xn
|πi − πi∗ | + kξk∞ .
i=1
n
(29)
Pn
There are two possibilities compatible with this bound: (1) kπ − π ∗ k∞ . n1 i=1 |πi − πi∗ |, and (2) kπ − π ∗ k∞ .
∗
k2
∗
kξk∞ . kπ−π
kπ ∗ k2 kπ k∞ by (26). In either case, the errors are fairly delocalized, revealing that
kπ − π ∗ k∞ . kπ − π ∗ k2
.
kπ ∗ k∞
kπ ∗ k2
We now move on to the regularized MLE, which follows a very similar argument. By the optimality
condition that ∇Lλ (θ) = 0, one can derive (for some η to be specified later)
θ − θ ∗ = θ − η∇Lλ (θ) − θ ∗
= θ − η∇Lλ (θ) − (θ ∗ − η∇Lλ (θ ∗ )) − η∇Lλ (θ ∗ )
| {z }
:=ζ
≈ I − η∇2 Lλ θ ∗
(θ − θ ∗ ) − ζ.
Write ∇2 Lλ (θ ∗ ) = D − A, where D and A denote respectively the diagonal and off-diagonal parts. Under
our assumptions, one can check that Dm,m n for all 1 ≤ m ≤ n and Aj,m 1 for any j 6= m. With these
notations in place, one can write the entrywise error as follows
X
∗
∗
θm − θm
= (1 − ηDm,m ) (θm − θm
)+
ηAj,m θj − θj∗ − ζm .
j:j6=m
By choosing η = c2 /n for some sufficiently small constant c2 > 0 , we get 1 − ηDm,m < 1 and ηAj,m 1/n.
Therefore, the right-hand side of the above relation also comprises a contraction term as well as an error
smoothing term, similar to (27). Carrying out the same argument as for the spectral method, we see that
the estimation errors of the regularized MLE are expected to be spread out.
2.6
Numerical experiments
It is worth noting that extensive numerical experiments have already been conducted in Negahban et al.
(2017a) to justify the practicability of both the spectral method and the regularized MLE. This section
provides some additional simulations to complement their experimental results and our theory. Throughout
the experiments, we set the number of items n to be 200, while the number of repeated comparisons L and
the edge probability p can vary with the experiments. For the tuning parameters,
q we choose d = 2dmax in
n
the spectral method where dmax is the maximum degree of the graph and λ = 2 np log
in the regularized
L
MLE, which are consistent with the configurations considered in the main theorems. All of the results are
averaged over 100 Monte Carlo simulations.
We first investigate the relative `∞ error of the spectral method and the regularized MLE when estimating
the preference scores. To this end, we generate the latent scores wi∗ (1 ≤ i ≤ n) independently and uniformly
at random over the interval [0.5, 1]. Figure 1(a) (resp. Figure 1(b)) displays the entrywise error in the score
estimation as the number of repeated comparisons L (resp. the edge probability p) varies. As can be seen
from the plots, the `∞ error of both methods gets smaller as p and L increase, confirming our results in
Theorems 3-4. Next, we show in Figure 2 the relative `∞ error and top-K ranking accuracy while fixing the
total number of samples (i.e. n2 pL). It can be seen that the performance almost does not change if the total
sample complexity n2 pL remains the same. In addition, Figure 3 illustrates the relative `∞ error and the
relative `2 error in score estimation for both methods. As we can see, the relative `∞ errors are not much
larger than the relative `2 errors (recall that n = 200), thus providing empirical evidence that the errors in
the score estimates are spread out across all entries.
Furthermore, we examine the top-K ranking accuracy of both methods. Here, we fix p = 0.25 and L = 20,
set K = 10, and let wi∗ = 1 for all 1 ≤ i ≤ K and wj∗ = 1 − ∆ for all K + 1 ≤ j ≤ n. By construction,
the score separation satisfies ∆K = ∆. Figure 4 illustrates the accuracy of both methods in identifying the
top-K ranked items. The performance of both methods improves when the score separation becomes larger,
which matches our theory in Theorem 1.
9
0.35
relative `1 estimation error
relative `1 estimation error
0.4
Spectral Method p = 0:25
Spectral Method p = 0:5
Spectral Method p = 0:75
Regularized MLE p = 0:25
Regularized MLE p = 0:5
Regularized MLE p = 0:75
0.3
Spectral Method L = 10
Spectral Method L = 30
Spectral Method L = 50
Regularized MLE L = 10
Regularized MLE L = 30
Regularized MLE L = 50
0.3
0.25
0.2
0.2
0.15
0.1
0.1
0.05
0
0
5
10
15
20
25
30
0
0.2
0.4
0.6
0.8
1
p: edge probability
L: number of repeated comparisons
(a)
(b)
Figure 1: Empirical performance of both the spectral method and the regularized MLE: (a) `∞ error vs. L,
and (b) `∞ error vs. p.
1
relative `1 estimation error
0.25
Spectral Method pL = 20
Spectral Method pL = 40
Spectral Method pL = 60
Regularized MLE pL = 20
Regularized MLE pL = 40
Regularized MLE pL = 60
0.2
0.8
0.15
0.6
0.1
0.4
0.05
0.2
0
0
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
p: edge probability
(a) relative `∞ error
(b) top-K ranking accuracy
Figure 2: Empirical performance of both the spectral method and the regularized MLE when fixing the
number of samples n2 pL: (a) relative `∞ error, and (b) top-K ranking accuracy.
0.25
`1 error, p = 0:25
`1 error, p = 0:5
`1 error, p = 0:75
`2 error, p = 0:25
`2 error, p = 0:5
`2 error, p = 0:75
0.3
relative estimation error
relative estimation error
0.4
0.2
0.1
0
5
10
15
20
25
0.2
0.15
0.1
0.05
30
L: number of repeated comparisons
`1 error, p = 0:25
`1 error, p = 0:5
`1 error, p = 0:75
`2 error, p = 0:25
`2 error, p = 0:5
`2 error, p = 0:75
5
10
15
20
25
30
L: number of repeated comparisons
(a) spectral method
(b) regularized MLE
Figure 3: Comparisons between the relative `∞ error and the relative `2 error for (a) the spectral method,
and (b) the regularized MLE.
10
top-K ranking accuracy
1
Spectral Method
Regularized MLE
0.8
0.6
0.4
0.2
0
0.1
0.2
0.3
0.4
0.5
"K : score separation
Figure 4: The top-K ranking accuracy of both the spectral method and the regularized MLE.
2.7
Other related works
The problem of ranking based on partial preferences has received much attention during the past decade.
Two types of observation models have been considered: (1) the cardinal-based model, where users provide
explicit numerical ratings of the items being compared, (2) the ordinal-based model, where users are asked to
make comparative measurements. See Ammar and Shah (2011); Shah et al. (2014) for detailed experimental
and theoretical comparisons between these two distinct models.
In terms of the ordinal-based model—and in particular, ranking from pairwise comparisons—both parametric and nonparametric models have been extensively studied. For example, Hunter (2004) examined
variants of the parametric BTL model, and established the convergence properties of the minorizationmaximization algorithm for solving the corresponding MLE problem. The BTL model was also investigated
in Borkar et al. (2016), where the authors proposed a randomized Kaczmarz algorithm to infer the latent
scores with provable `2 error guarantees. Moreover, the BTL model falls under the category of low-rank
parametric models, since the preference matrix is generated by passing a rank-2 matrix through the logistic
link function (Rajkumar and Agarwal, 2016). Additionally, the work Jiang et al. (2011); Xu et al. (2012)
proposed a least-squares type method, called HodgeRank, to estimate the full ranking, which generalizes the
simple Borda count algorithm (Borda, 1784; Ammar and Shah, 2011; Rajkumar and Agarwal, 2014). For
many of these algorithms, the sample complexities needed for perfect total ranking have been determined
by Rajkumar and Agarwal (2014), although the top-K ranking accuracy was not considered there.
Going beyond the parametric models, a recent line of works Shah et al. (2017); Shah and Wainwright
(2015); Chen et al. (2017); Pananjady et al. (2017) considered the nonparametric stochastically transitive
model, where the only model assumption is that the comparison probability matrix follows certain transitivity
rules. This type of models subsumes the BTL model as a special case. For instance, Shah and Wainwright
(2015) suggested an exceedingly simple counting-based algorithm which can reliably recover the top-K ranked
items for a wide range of models. However, the sampling paradigm considered therein is quite different from
ours in the sparse regime; for instance, their model does not come close to the setting where p is small but
L is large, which is the most challenging regime of the model adopted in our paper and Negahban et al.
(2017a); Chen and Suh (2015); Jang et al. (2016).
All of the aforementioned papers concentrate on the case where there is a single ground-truth ordering.
It would also be interesting to investigate the scenarios where different users might have different preference
scores (Negahban et al., 2017b; Lu and Negahban, 2014; Wu et al., 2015; Oh and Shah, 2014). To this
end, Negahban et al. (2017b); Lu and Negahban (2014) imposed the low-rank structure on the underlying
preference matrix and adopted the nuclear-norm relaxation approach to recover the users’ preferences. The
paper Wu et al. (2015) proposed a two-step procedure, where a clustering step was first applied to label
different users, followed by an estimation step built upon the previous clustering outcome. Additionally,
several papers explored the ranking problem for the more general Plackett-Luce model (Hajek et al., 2014;
Soufiani et al., 2013, 2014), in the presence of adaptive sampling (Braverman and Mossel, 2008; Jamieson
and Nowak, 2011; Ailon, 2012; Busa-Fekete et al., 2013; Mohajer and Suh, 2016; Heckel et al., 2016; Agarwal
et al., 2017), for the crowdsourcing scenario (Chen et al., 2013; Osting et al., 2016), and in the adversarial
11
setting (Suh et al., 2017). These are beyond the scope of the present paper.
When it comes to the technical tools, it is worth noting that the leave-one-out idea has been invoked to
analyze random designs for other high-dimensional problems as well, e.g. robust M-estimators (El Karoui
et al., 2013; El Karoui, 2017), confidence intervals for Lasso (Javanmard and Montanari, 2015), likelihood
ratio test (Sur et al., 2017). In particular, Zhong and Boumal (2017) and Abbe et al. (2017) use this technique
to precisely characterize entrywise behavior of eigenvectors of a large class of symmetric random matrices,
which improves upon the prior analysis on `∞ eigenvector perturbation (Fan et al., 2016). As a consequence,
they are able to show the sharpness of spectral methods—a type of methods based on the eigen-structure of
a certain data matrix—in many popular models. Our introduction of leave-one-out auxiliary quantities are
motivated by these papers.
Finally, we remark that the family of spectral methods has been successfully applied in numerous applications, e.g. matrix completion (Keshavan et al., 2010; Jain et al., 2013), phase retrieval (Candès et al.,
2015; Chen and Candès, 2017), graph clustering (Rohe et al., 2011; Chaudhuri et al., 2012), synchronization
and joint alignment (Singer, 2011; Chen and Candes, 2016). All of them are designed based on the eigenvectors of some symmetric matrix, or the singular vectors if the matrix of interest is asymmetric. Our paper
contributes to this growing literature by establishing a sharp eigenvector perturbation analysis framework
for an important class of asymmetric matrices—the probability transition matrices.
3
Analysis for the spectral method
This section is devoted to proving Theorem 3, which characterizes the pointwise error of the spectral estimate.
3.1
Preliminaries
Here, we gather some preliminary facts about reversible Markov chains as well as the Erdős–Rényi random
graph.
The first important result concerns the eigenvector perturbation for probability transition matrices, which
can be treated as the analogue of the celebrated Davis-Kahan sin Θ theorem (Davis and Kahan, 1970) for
asymmetric matrices. Due to its potential importance for other problems, we promote it to a theorem as
follows.
Theorem 5 (Eigenvector perturbation). Suppose that P , P̂ , and P ∗ are probability transition matrices
with stationary distributions π, π̂, π ∗ , respectively. Also, assume that P ∗ represents a reversible Markov
chain. Then it holds that
kπ − π̂kπ∗ ≤
π > (P − P̂ )
π∗
1 − max {λ2 (P ∗ ), −λn (P ∗ )} − P − P̂
.
π∗
Proof. See Appendix A.1.
Several remarks regarding Theorem 5 are in order. First, in contrast to standard perturbation results
like Davis-Kahan’s sin Θ theorem, our theorem involves three matrices in total, where P , P̂ , and P ∗ can
all be arbitrary. For example, one may choose P ∗ to be the population transition matrix, and P and P̂ as
two finite-sample versions associated with P ∗ . Second, we only impose reversibility on P ∗ , whereas P and
P̂ need not induce reversible Markov Chains. Third, Theorem 5 allows one to derive the `2 estimation error
in Negahban et al. (2017a) directly without resorting to the power method; in fact, our `2 estimation error
bound improves upon Negahban et al. (2017a) by some logarithmic factor.
Theorem 6. Suppose p ≥ c0 logn n for some sufficiently large constant c0 > 0 and d ≥ cd np for cd ≥ 2 in
Algorithm 1. With probability exceeds 1 − O(n−5 ), one has
kπ − π ∗ k2
1
.√
.
kπ ∗ k2
npL
Proof. See Appendix D.
12
Notably, Theorem 6 matches the minimax lower bound derived in (Negahban et al., 2017a, Theorem 3).
As far as we know, this is the first result to demonstrate the orderwise optimality of the spectral method
when measured by the `2 loss.
The next result is concerned with the concentration of the vertex degrees in an Erdős–Rényi random
graph.
Lemma 1 (Degree concentration). Suppose that G ∼ Gn,p . Let di be the degree of node i, dmin = min1≤i≤n di
n
for some sufficiently large constant c0 > 0, then the following event
and dmax = max1≤i≤n di . If p ≥ c0 log
n
3np
np
≤ dmin ≤ dmax ≤
(30)
A0 =
2
2
obeys
P (A0 ) ≥ 1 − O(n−10 ).
Proof. This lemma follows from the standard Chernoff bound, and hence is omitted.
Since d is chosen to be cd np for some constant cd ≥ 2, we have, by Lemma 1, that the maximum vertex
degree obeys dmax < d with high probability.
3.2
Proof outline of Theorem 3
>
>
In this subsection, we outline the proof of Theorem 3. Recall that π = [π1 , · · · , πn ] and π ∗ = [π1∗ , · · · , πn∗ ]
are the stationary distributions associated with P and P ∗ , respectively. This gives
π> P = π>
and
π ∗> P ∗ = π ∗> .
For each 1 ≤ m ≤ n, one can decompose
>
∗
∗
∗
πm − πm
= π > P·m − π ∗> P·m
= π ∗> (P·m − P·m
) + (π − π ∗ ) P·m
X
X
∗
∗
=
πj∗ Pj,m − Pj,m
+ (πm − πm
) Pm,m +
πj − πj∗ Pj,m ,
j
|
j:j6=m
{z
:=I1m
}
{z
|
:=I2m
}
where P·m denotes the m-th column of P . Then it boils down to controlling I1m , I2m and
P
j:j6=m
πj − πj∗ Pj,m .
1. Since π ∗ is deterministic while P is random, we can easily control I1m using Hoeffding’s inequality.
The bound is the following.
Lemma 2. With probability exceeding 1 − O(n−5 ), one has
r
log n ∗
m
kπ k∞ .
max |I1 | .
m
Ld
Proof. See Appendix A.2.
∗
2. Next, the term I2m behaves as a contraction of |πm − πm
|, as demonstrated below.
Lemma 3. With probability exceeding 1 − O(n−5 ), there exists some constant c > 0 such that for all
1 ≤ m ≤ n,
!
r
np
log
n
∗
|I2m | ≤ 1 −
+c
|πm − πm
|.
2(1 + κ)d
Ld
Here, we recall that κ = wmax /wmin .
Proof. See Appendix A.3.
13
3. The statistical dependency
between
π and P introduces difficulty in obtaining a sharp estimate of
P
the third term j:j6=m πj − πj∗ Pj,m . Nevertheless, the leave-one-out technique helps us decouple
the dependency and obtain effective control of this term. The key component of the analysis is the
introduction of a new probability transition matrix P (m) , which is a leave-one-out version of the original
matrix P . More precisely, P (m) replaces all of the transition probabilities involving the m-th item with
their expected values (unconditional on G); that is, for any i 6= j,
(
Pi,j ,
i 6= m, j 6= m,
(m)
Pi,j := p ∗
y
,
i
= m or j = m,
d i,j
and for any 1 ≤ i ≤ n,
(m)
Pi,i := 1 −
X
(m)
Pi,j
j:j6=i
(m)
in order to ensure that P
is a probability transition matrix. In addition, we let π (m) be the
stationary distribution of the Markov chain induced by P (m) . As to be demonstrated later, the main
advantages of introducing π (m) are two-fold: (1) the original spectral estimate π can be very well
approximated by π (m) , and (2) π (m) is statistically independent of the connectivity
of the m-th node
P
and the comparisons with regards to the m-th item. Now we further decompose j:j6=m πj − πj∗ Pj,m :
X (m)
X
X
(m)
(πj − πj∗ )Pj,m .
πj − πj∗ Pj,m =
(πj − πj )Pj,m +
j:j6=m
j:j6=m
j:j6=m
{z
|
}
:=I3m
|
{z
:=I4m
}
4. For I3m , we apply the Cauchy-Schwarz inequality to obtain that with probability at least 1 − O(n−10 ),
|I3m | ≤ kπ (m) − πk2
X
j:j6=m
2
Pj,m
12
(i)
1
≤ √ kπ (m) − πk2 ,
d
where (i) follows from the fact that Pj,m ≤ d1 for all j 6= m and dmax ≤ d on the event A0 . Consequently,
it suffices to control the `2 difference between the original spectral estimate π and its leave-one-out
version π (m) . This is accomplished in the following lemma.
Lemma 4. With probability at least 1 − O n−5 ,
√ r
16 κ log n ∗
(m)
kπ
− πk2 ≤
kπ k∞ + kπ − π ∗ k∞ ,
(31)
γ
Ld
where κ = wmax /wmin and γ = 1 − max {λ2 (P ∗ ), −λn (P ∗ )} − kP − P ∗ kπ∗ .
Proof. See Appendix A.4.
As shown in (Negahban et al., 2017a, Section 4.2), γ is lower bounded by some positive constant.
Lemma 5 (Spectral gap Negahban et al. (2017a)). Under the model specified in Section 2.1, if
p ≥ c0 logn n for some sufficiently large constant c0 > 0, then with probability at least 1 − O(n−5 ), one
has
1 dmin
1
γ := 1 − max {λ2 (P ∗ ), −λn (P ∗ )} − kP − P ∗ kπ∗ ≥ 2
≥ 2.
2κ dmax
6κ
With this spectral gap in mind, Lemma 4 reads
r
log n ∗
(m)
kπ
− πk2 .
kπ k∞ + kπ − π ∗ k∞ .
Ld
(32)
5. In order to control I4m , we exploit the statistical independence between π (m) and P·m . Specifically, we
demonstrate that:
14
Lemma 6. Suppose
that p >
least 1 − O n−10 ,
|I4m |
c0 log n
n
for some sufficiently large constant c0 > 0. With probability at
1
. √ kπ (m) − πk2 +
n
r
log n ∗
kπ k∞ +
Ld
√
np log n + log n (m)
kπ
− π ∗ k∞ .
d
Proof. See Appendix A.5.
The above bound depends on both kπ (m) − πk2 and kπ (m) − π ∗ k∞ . We can invoke Lemma 4 and the
inequality kπ (m) − π ∗ k∞ ≤ kπ (m) − πk2 + kπ − π ∗ k∞ to reach
r
√
√
np log n + log n
log n ∗
np log n + log n
1
(m)
m
√
+
kπ
− πk2 +
kπ k∞ +
kπ − π ∗ k∞
|I4 | .
d
Ld
d
n
√
√
r
1
np log n + log n
κ
log n ∗
√ +
.
+1
kπ k∞
d
γ
Ld
n
√
1
np log n + log n
+ √ +
kπ − π ∗ k∞ .
d
n
6. Finally, putting the preceding bounds together, we conclude that with high probability, for some
absolute constants c1 , c2 , c3 > 0 one has
!
r
√
√ r
np
logn
1
nplogn+logn 16 κ
logn ∗
c3
∗
−c1
|πm −πm |≤ c2 + √ + √ +c3
kπ k∞
2(1+κ)d
Ld
d
γ
Ld
n
d
√
1
c3
nplogn+logn
+ √ + √ +2c3
kπ−π ∗ k∞ .
d
n
d
simultaneously for all 1 ≤ m ≤ n. By taking the maximum over m on the left-hand side and combining
terms, we get
!
r
√
np
log n
np log n + log n
1
c3
− c1
− √ − √ − 2c3
kπ − π ∗ k∞
2(1 + κ)d
Ld
d
n
d
|
{z
}
:=α1
≤
c2 +
1
c
√ + √3 + c3
n
d
|
√
np log n + log n
d
{z
:=α2
√ r
16 κ
log n ∗
kπ k∞ .
γ
Ld
}
√
Hence, as long as np/ log n is sufficiently large, one has d np np log n > log n. This implies that
α1 > 0 and α1 α2 1, which further leads to
s
r
log
n
log n ∗
kπ − π ∗ k∞ .
kπ ∗ k∞
kπ k∞ .
Ld
npL
This finishes the proof of Theorem 3.
4
Analysis for the regularized MLE
The goal of this section is to establish the `∞ error of the regularized MLE as claimed in Theorem 4. Recall
∗
that in Theorem 4, we compare the regularized MLE θ with θ ∗ − θ 1, and hence without loss of generality
we can assume that
1> θ ∗ = 0.
(33)
15
4.1
Preliminaries and notation
Before proceeding to the proof, we gather some simple facts. To begin with, the gradient and the Hessian of
L (·; y) in (14) can be computed as
X
eθ i
(ei − ej )
(34)
∇L (θ; y) =
−yj,i + θi
e + eθ j
(i,j)∈E,i>j
and
eθ i eθ j
X
∇2 L (θ; y) =
(i,j)∈E,i>j
(eθi
+
>
2
eθj )
(ei − ej ) (ei − ej ) ,
(35)
where e1 , · · · , en stand for canonical basis vectors in Rn .
When evaluated at the truth θ ∗ , the size of the gradient can be controlled as follows.
Lemma 7. The following event
(
A2 :=
r
∗
k∇Lλ (θ ; y)k2 .
n2 p log n
L
)
.
(36)
occurs with probability exceeding 1 − O(n−10 ).
Proof. See Appendix B.1.
The following lemmas characterize the smoothness and the strong convexity of the function Lλ (·; y).
P
>
In the sequel, we denote by LG =
the (unnormalized) Laplacian matrix
(i,j)∈E,i>j (ei − ej ) (ei − ej )
(Chung, 1997) associated with G, and for any matrix A we let
λmin,⊥ (A) := µ | z > Az ≥ µkzk2 for all z with 1> z = 0 .
(37)
Lemma 8. Suppose that p >
defined in (30), one has
c0 log n
n
for some sufficiently large constant c0 > 0. Then on the event A0 as
λmax ∇2 Lλ (θ; y) ≤ λ + np,
Proof. Note that
eθi eθj
θi +eθj 2
e
)
(
∀θ ∈ Rn .
≤ 41 . It follows immediately from the Hessian in (35) that
1
1
λmax ∇2 Lλ (θ; y) ≤ λ + kLG k ≤ λ + dmax ,
4
2
where dmax is the maximum vertex degree in the graph G. In addition, on the event A0 we have dmax ≤ 2np,
which completes the proof.
Lemma 9. For all kθk∞ ≤ C where C > 0 is some absolute constant, we have
λmin,⊥ ∇2 Lλ (θ; y) ≥ λ +
e−2C
(1 + e−2C )
2 λmin,⊥
(LG ) .
Proof. The proof is straightforward and hence is omitted.
n
for some sufficiently large constant c0 > 0. Then
Lemma 10. Let G ∼ Gn,p , and suppose that p > c0 log
n
one has
1
P λmin,⊥ (LG ) ≥ np ≥ 1 − O n−10 .
2
Proof. Note that λmin,⊥ (LG ) is exactly the spectral gap of the Laplacian matrix. See (Tropp, 2015, Sec
5.3.3) for the derivation of this lemma.
16
Algorithm 2 Gradient descent for regularized MLE
Initialize θ 0 = 0.
for t = 0, 1, 2, . . . , T − 1 do
θ t+1 = θ t − ηt ∇L (θ t ; y) + λθ t ;
end for
Output θ T .
By combining Lemma 9 with Lemma 10, we reach the following result.
Corollary 1. Under the assumptions of Lemma 10, with probability exceeding 1 − O n−10 one has
λmin,⊥ ∇2 Lλ (θ; y) ≥ λ +
e−2C
2 np
2 (1 + e−2C )
simultaneously for all θ obeying kθk∞ ≤ C.
4.2
Proof outline of Theorem 4
This subsection outlines the main steps for establishing Theorem 4. Before continuing, we single out an
important fact that 1> θ = 0, which is an immediate consequence of the shift-invariance property of L (θ; y).
Fact 1. The regularized MLE θ obeys 1> θ = 0.
Proof. Suppose instead that 1> θ 6= 0. Then setting θ =
kθk22 > kθ − θ1k22
1 >
n1 θ
one has
L (θ; y) = L θ − θ1; y ,
and
where the last
identity arises since L (θ; y) is invariant under global shift. This implies that Lλ (θ; y) >
Lλ θ − θ1; y , which contradicts the optimality of θ.
The proof relies heavily on the strong convexity of the regularized negative log-likelihood function. Notably, while Lλ is a strongly convex function, its curvature might be arbitrarily close to λ for those points far
away from the optimizer. In contrast, if we restrict our attention to a set of parameters that are reasonably
close to the optimizer, a much stronger level of convexity can be exploited. In light of this observation, we
divide our analysis into two parts: (1) a coarse-level analysis demonstrating that θ is not far from θ ∗ in
an entrywise manner; (2) a fine-scale analysis that reveals the optimal performance of the regularized MLE
using refined strong convexity.
4.2.1
A coarse-level analysis
In this section, our goal is to obtain the crude bound
kθ − θ ∗ k∞ . 1.
(38)
Rather than directly resorting to the optimality condition, we adopt an algorithmic perspective to analyze
the regularized MLE θ. Specifically, we consider the standard gradient descent algorithm that is expected to
converge to the minimizer θ, and analyze the trajectory of this iterative algorithm instead. The algorithm
is stated in Algorithm 2.
Our proof can be separated into three parts. In what follows, we shall take
ηt ≡ η =
1
,
λ + np
∞
t = 0, 1, 2, · · · .
(39)
1. It is seen that the sequence {θ t }t=1 converges to the regularized MLE θ, a property that is standard
in convex optimization literature. This claim is summarized in the following lemma.
17
Lemma 11. On the event A0 as defined in (30), one has
kθ t − θk22 ≤ ct1 kθ 0 − θk22 ,
where c1 = 1 −
λ
λ+np .
Proof. The result directly follows from Lemma 8, Corollary 1, and the property of the gradient descent
algorithm (e.g. (Bubeck, 2015, Theorem 3.10)).
With this convergence property in place, if we can further show that kθ t − θ ∗ k∞ ≤ C for all t ∈ N,
then we obtain the desired result kθ − θ ∗ k∞ ≤ C.
2. Before moving to the analysis for `∞ errors, we develop an upper bound on kθ t − θ ∗ k2 .
Lemma 12. On the event A2 as defined in (36), there exists some constant C2 > 0 such that
√
t = 0, 1, 2, · · · .
kθ t − θ ∗ k2 ≤ C2 n,
Proof. See Appendix B.2.
3. We then turn to the `∞ error of θ t . To this end, we proceed by building a contraction relationship
t+1
∗
t
∗
between θm
− θm
and |θm
− θm
|. More precisely, the aim is to prove that
t+1
∗
t
∗
θm
− θm
≤ ρ θm
− θm
+ c2
for some ρ < 1 and c2 > 0, as detailed in the following lemma.
Lemma 13. On the event A0 ∩ A2 that occurs with probability at least 1 − O(n−10 ), we have
!
√
C2
e−2 3C2
t
∗
∗
t+1
∗
ηnp
θm − θm ≤ 1 − 0.3ηnp
√
2 θm − θm + 8ηnp + ηλkθ k∞ +
−2
3C
4
2
1+e
simultaneously for all 1 ≤ m ≤ n and t ∈ N. Here, C2 > 0 is the same constant as in Lemma 12.
Proof. See Appendix B.3.
Notice that θ 0 = 0 and hence kθ 0 − θ ∗ k∞ = kθ ∗ k∞ . 1. In view of Lemma 13, by picking C > 0 to
be any constant satisfying
√
e−2
1 − 0.3ηnp
1+
3C2
√
2
e−2 3C2
C + 8ηnp + ηλkθ ∗ k∞ +
C2
ηnp ≤ C,
4
(40)
one can conclude that
kθ t − θ ∗ k∞ ≤ C
=⇒
kθ t+1 − θ ∗ k∞ ≤ C.
4. By induction, the above bounds taken together guarantee that on the event A0 ∩ A2 , the iterates
satisfy kθ t − θ ∗ k∞ ≤ C throughout all iterations. In fact, this holds for all C > 0 compatible with the
condition (40), or equivalently,
8 + λ/np +
C≥
0.3
C2
4
This establishes our claim (38).
18
√
1 + e−2
e−2
3C2 2
√
3C2
1.
Remark 1 (Near-optimal computational cost). Before we move on to refine the analysis, we point out
a byproduct of the preceding analysis: the near-optimal computational complexity of the gradient descent
algorithm. Specifically, the above analysis reveals that
kθ t − θ ∗ k∞ . 1
t = 0, 1, · · · ,
provided that the starting point is θ 0 = 0. This, combined with Corollary 1, suggests a much stronger level
of convexity within a region of interest, namely,
n
o
λmin,⊥ ∇2 Lλ θ̃; y
& λ + np,
∀θ̃ ∈ B := θ̃ | kθ̃ − θ ∗ k∞ . 1 ,
where the entire trajectory of {θ t | t ≥ 0} stays within this region B. Therefore, one can immediately improve
Lemma 11 to
kθ t − θk22 ≤ ct2 kθ 0 − θk22
for some absolute constant 0 < c2 < 1. This reveals an interesting phenomenon: even though the negative
log-likelihood function is not strongly convex (due to the possibly flat curvature of the logistic function) and
the regularization parameter λ is small, the gradient descent algorithm converges geometrically fast to the
optimizer, provided that we start from θ 0 = 0. Given that the gradient update in each iteration takes linear
time—in time proportional to reading the data, the whole algorithm has nearly linear runtime.
4.2.2
A fine-scale analysis
This subsection further refines our analysis using our results in Section 4.2.1. A key ingredient of the analysis
is to introduce a leave-one-out version of the MLE that replaces all log-likelihood components involving the
m-th item with their expected values (unconditional on G); that is,
(m)
θ (m) := arg minn Lλ
θ∈R
(θ; y) ,
where
(m)
Lλ
X
(θ; y) :=
−yj,i (θi − θj ) + log 1 + eθi −θj
(i,j)∈E,i>j,i6=m,j6=m
∗
1
eθ i
θi −θm
+ λkθk22 .
+
p − θ∗
∗ (θi − θm ) + log 1 + e
i + eθ m
2
e
i:i6=m
X
(41)
As before, this auxiliary vector θ (m) serves as a reasonably good proxy for θ, while being statistically
independent of {yi,m | (i, m) ∈ E}.
Our proof proceeds with the following steps, all of which rely on the fact that θ and θ (m) live in a “good”
region near θ ∗ where strong convexity holds, guaranteed by our coarse-level analysis in Section 4.2.1.
1. Show that both θ and θ (m) are close to the truth θ ∗ when measured by the `2 loss. Specifically, we
prove that:
Lemma 14. With probability exceeding 1 − O(n−5 ), one has
s
log n
kθ − θ ∗ k2 .
,
pL
and
s
θ (m) − θ ∗
2
.
log n
,
pL
1 ≤ i ≤ m.
(42)
(43)
q
√
n
When kθ ∗ k2 n, Lemma 14 yields bound on relative `2 error kθ − θ ∗ k2 /kθ ∗ k2 . log
npL of MLE,
the same rate (up to log factors) as the one of spectral method in Theorem 6.
19
Proof. See Appendix B.4.
2. Control the `2 difference between the MLE and its leave-one-out counterpart θ (m) as follows:
n
for some sufficiently large constant c0 > 0. With probability
Lemma 15. Suppose that p > c0 log
n
exceeding 1 − O(n−10 ), there exists some constant c1 > 0 such that
s
log n 1
(m)
kθ
− θk2 ≤ c1
+ kθ − θ ∗ k∞ ,
1 ≤ m ≤ n.
npL
3
Proof. See Appendix B.5.
Notably, while kθ (m) − θk2 depends also on the pointwise error kθ − θ ∗ k∞ , we emphasize the presence
of a contraction factor (i.e. 1/3) in front of kθ − θ ∗ k∞ . This factor plays a crucial role in showing error
contraction.
3. Show that the leave-one-out estimate θ (m) is close to the truth when it comes to the m-th coordinate:
Lemma 16. With probability exceeding 1 − O(n−10 ), we have
s
log n
(m)
∗
,
1 ≤ i ≤ m.
θm − θm .
npL
Proof. See Appendix B.6.
4. Taken collectively, the above bounds imply the existence of some absolute constant c2 > 0 such that
s
log n 1
∗
(m)
∗
(m)
+ kθ − θ ∗ k∞
(44)
θm − θm ≤ θm − θm + kθ − θ k2 ≤ c2
npL
3
with high probability. Since this holds for each 1 ≤ m ≤ n, we can take the maximum over m in the
left-hand side and apply the union bound to derive
s
log n 1
kθ − θ ∗ k∞ ≤ c2
+ kθ − θ ∗ k∞ ,
(45)
npL
3
or equivalently,
∗
kθ − θ k∞
3c2
≤
2
5. It remains to show that
∗
keθ − eθ k∞
.
keθ∗ k∞
s
s
log n
.
npL
(46)
log n
.
npL
Toward this end, we observe that for each 1 ≤ m ≤ n,
∗
∗
∗
∗
|eθm − eθm |
|eθ̃m (θm − θm
)|
eθmax +kθ−θ k∞ · |θm − θm
|
=
≤
,
eθmax
eθmax
eθmax
∗
where θ̃m is between θm and θm
, and θmax is the largest entry of θ∗ . Continuing the derivation and
using (46), we arrive at
s
∗
∗
|eθm − eθm |
eθmax +kθ−θ k∞ kθ − θ ∗ k∞
log n
max
≤
.
1≤m≤n
eθmax
eθmax
npL
as long as npL/ log n is sufficiently large. This completes the proof of Theorem 4.
20
Connection with the spectral method. Before concluding this section, we remark that the proof
strategy for the regularized MLE shares a lot of similarity with that for the spectral method. For instance,
the proofs in both cases proceed by introducing appropriate leave-one-out estimates, which are exceedingly
close to the original estimates in the sense that
s
s
log
n
log n
kπ ∗ k∞
and
kθ (m) − θk2 .
.
kπ (m) − πk2 .
npL
npL
Furthermore, a crucial ingredient of the proofs is some sort of the perturbation bound. This corresponds
to the counterpart of the Davis-Kahan theorem (Theorem 5) for the spectral method, and the optimality
condition for the regularized MLE (see, e.g. (60)) under strong convexity. Both cases take the following
form: for an optimal solution z and its perturbed version z̃, one has
kz − z̃k2 .
kf (z̃)k2
,
γ
(47)
where γ represents some stability constant (spectral gap in spectral method or strong convexity parameter
in MLE). Moreover, f (·) is some vector-valued function obeying f (z) = 0 that captures the optimality of z;
more precisely, one can take
(
z > P − z, for the spectral method;
f (z) : =
∇Lλ (z),
for the regularized MLE.
We expect this form (47) to extend to a much broader class of optimization problems.
5
Discussion
In this paper, we justify the optimality of both the spectral method and the regularized MLE for top-K
rank aggregation. Our theoretical studies are by no means exhaustive, and there are numerous directions
that would be of interest for future investigations. We point out a few possibilities as follows.
• Unregularized MLE.
We have studied the optimality of the regularized MLE with the regularization
q
np log n
. Our analysis relies on the regularization term to obtain the coarse bound on
parameter λ
L
kθ − θ ∗ k∞ in Section 4.2.1. It is natural to ask whether such a regularization term is necessary or not.
This question remains open.
• More general comparison graphs. So far we have focused on a tractable but somewhat restrictive
comparison graph, namely, the Erdős–Rényi random graph. It would certainly be interesting and
important to understand the performance of both methods under a broader family of comparison
graphs, and to see which algorithms would enable optimal sample complexities under general sampling
patterns.
• Beyond pairwise comparisons. In real world applications, we often encounter the case when we are
given more than two items to compare. For instance, in online advertisement scenarios, the users might
be asked to compare three to five different advertisements. How to extend the methods and analyses
to handle more general listwise comparisons is of significant practical interest.
• Entrywise perturbation analysis for general convex optimization. The current paper provides the `∞
perturbation analysis for the regularized MLE using the leave-one-out trick as well as an inductive
argument along the algorithmic updates. We expect this analysis framework to carry over to a much
broader family of convex optimization problems, which may in turn offer a powerful tool for showing
the stability of optimization procedures in an entrywise fashion.
21
A
Proofs in Section 3
This section collects proofs of the theorems and lemmas that appear in Section 3. By Lemma 1, the event
1
3
A0 =
np ≤ dmin ≤ dmax ≤ np
2
2
happens with probability at least 1 − O(n−10 ). Throughout this section, we shall assume that we are on this
event without explicitly referring to it each time. An immediate consequence is that dmax ≤ d on this event.
A.1
Proof of Theorem 5
To begin with, we write
>
π > − π̂ > = π > P − π̂ > P̂ = π > (P − P̂ ) + (π − π̂) P̂ .
(48)
The last term of the above identity can be further decomposed as
>
>
>
(π − π̂) P̂ = (π − π̂) P ∗ + (π − π̂) (P̂ − P ∗ )
>
>
= (π − π̂) P ∗ − 1π ∗> + (π − π̂) (P̂ − P ∗ ),
(49)
>
where we have used the fact that (π − π̂) 1π ∗> = 0. Combining (48) and (49) we get
>
>
π > − π̂ > = π > (P − P̂ ) + (π − π̂) P ∗ − 1π ∗> + (π − π̂) (P̂ − P ∗ ),
which together with a little algebra gives
kπ − π̂kπ∗ ≤ π > (P − P̂ )
π∗
+ kπ − π̂kπ∗ P ∗ − 1π ∗>
kπ − π̂kπ∗ ≤
=⇒
π∗
π > (P − P̂ )
1−
kP ∗
−
1π ∗> k
π∗
+ kπ − π̂kπ∗ P̂ − P ∗
π∗
− P̂ − P ∗
π∗
.
π∗
The theorem follows by recognizing that
P ∗ − 1π ∗>
A.2
π∗
= max {λ2 (P ∗ ), −λn (P ∗ )} .
Proof of Lemma 2
Observe that
I1m =
X
∗
∗
∗
πj∗ Pj,m − Pj,m
+ πm
Pm,m − Pm,m
j:j6=m
(i)
=
(
X
∗
∗
+ πm
πj∗ Pj,m − Pj,m
!
1−
j:j6=m
=
X
X
Pm,j
j:j6=m
πj∗
+
∗
πm
Pj,m −
∗
Pj,m
!)
−
1−
X
∗
Pm,j
j:j6=m
j:j6=m
=
L
1 X X ∗
(l)
∗
∗
πj + πm
1(j,m)∈E yj,m − yj,m
,
Ld
(50)
j:j6=m l=1
where (i) follows from the fact that P and P ∗ are both probability transition matrices. By Lemma 17 and
the fact that dmax ≤ d, one can derive
L
X X
n
o
(l)
∗
∗
P |I1m | ≥ t G = P
πj∗ + πm
1(j,m)∈E yj,m − yj,m
≥ Ldt G
j:j6=m l=1
2(Ldt)2
≤ 2 exp −
.
Ldmax (2kπ ∗ k∞ )2
22
This together with the high probability event A0 indicates
n
o
−Ldt2
m
,
P |I1 | ≥ t ≤ 2 exp
2kπ ∗ k2∞
which further implies
(
r
|I1m |
P
≥4
log n ∗
kπ k∞
Ld
)
≤ 2n−8 .
The lemma is established by taking the union bound.
A.3
Proof of Lemma 3
Applying Lemma 17 to the quantity
X
∗
Pm,m − Pm,m
=−
∗
(Pm,j − Pm,j
)=−
L
1 X X
(l)
∗
,
1(j,m)∈E yj,m − yj,m
Ld
j:j6=m l=1
j:j6=m
we get
(
r
∗
P max |Pm,m − Pm,m
|≥2
m
log n
Ld
)
≤ 2n−7 .
On the other hand, we have for all 1 ≤ m ≤ n
∗
Pm,m
=1−
X
∗
Pm,j
≤1−
j:j6=m
dmin
1
np
1
·
≤1−
·
.
d
1+κ
2d 1 + κ
Combining these two pieces completes the proof.
A.4
Proof of Lemma 4
First, by the relationship between k · k2 and k · kπ∗ , we have
kπ (m) − πk2 ≤ p
1
kπ (m) − πkπ∗ ,
∗
πmin
∗
where πmin
:= mini πi∗ . Invoking Theorem 5, we obtain
kπ (m)> (P (m) − P )kπ∗
1 − max {λ2 (P ∗ ), −λn (P ∗ )} − kP − P ∗ kπ∗
1p ∗
πmax kπ (m)> (P (m) − P )k2 ,
≤
γ
kπ (m) − πkπ∗ ≤
∗
where we define γ := 1 − max {λ2 (P ∗ ), −λn (P ∗ )} − kP − P ∗ kπ∗ and πmax
:= maxi πi∗ .
(m)>
(m)
To facilitate the analysis of kπ
(P
− P )k2 , we introduce another Markov chain with transition
probability matrices P (m),G , which is also a leave-one-out version of the transition matrix P . Similar to
P (m) , P (m),G replaces all the transition probabilities involving the m-th item with their expected values
(conditional on G). Concretely, for i 6= j,
(
Pi,j ,
i 6= m, j 6= m,
(m),G
Pi,j
= 1 ∗
i = m or j = m.
d yi,j 1(i,j)∈E ,
And for each 1 ≤ i ≤ n, we define
(m),G
Pi,i
=1−
X
j:j6=i
23
(m),G
Pi,j
to make P (m),G a valid probability transition matrix. Hence by the triangle inequality, we see that
kπ (m)> (P (m) − P )k2 ≤ kπ (m)> (P − P (m),G )k2 + kπ (m)> (P (m) − P (m),G )k2 .
{z
} |
{z
}
|
:=J1m
:=J2m
The next step is then to bound J1m and J2m separately.
For J1m , similar to (50), one has
h
i (i) h
i
π (m)> (P − P (m),G )
= π (m)> (P − P ∗ )
m
m
1 X
Ld
=
L
X
(m)
πj
(l)
∗
(m)
,
1(j,m)∈E yj,m − yj,m
+ πm
j:j6=m l=1
(m),G
∗
where (i) comes from the fact that P·m
= P·m
. Recognizing that π (m) is statistically independent of
{yj,m }j6=m , by Hoeffding’s inequality (cf. Lemma 17), we get
(
)
r
log n (m)
(m)>
(m),G
(P − P
)]m ≥ 4
P [π
kπ k∞ ≤ 2n−8 .
(51)
Ld
And for j 6= m, we have
h
X (m)
i
(m),G
π (m)> P − P (m),G
=
πi
Pi,j − Pi,j
j
i
(m)
πj
(m),G
(m),G
(m)
Pj,j − Pj,j
+ πm
Pm,j − Pm,j
(m) 1 ∗
(m) 1
∗
(y
− yj,m ) 1(j,m)∈E + πm
(ym,j − ym,j
) 1(j,m)∈E
= πj
d j,m
d
(m)
∗
(m)
∗
= πj
Pj,m
− Pj,m + πm
Pm,j − Pm,j
.
=
In addition by Hoeffding’s inequality in Lemma 17, we have
max Pj,m −
j6=m
∗
Pj,m
2
≤
d
r
log n
L
with probability at least 1 − O(n−5 ). As a consequence,
( q
h
log n
4
(m)
i
k∞ ,
(m)>
(m),G
L kπ
≤ d
π
P −P
j
0,
if (j, m) ∈ E,
else.
Combining (51) and (52) yields
r
r
r
√
(i)
log n (m)
4 dmax − 1 log n (m)
log n (m)
m
kπ k∞ +
kπ k∞ ≤ 8
kπ k∞ ,
J1 ≤ 4
Ld
d
L
Ld
where (i) comes from the fact that dmax ≤ d.
Regarding J2m , we invoke the identity π ∗> (P (m) − P (m),G ) = 0 to get
>
π (m)> (P (m) − P (m),G ) = π (m) − π ∗ (P (m) − P (m),G ).
Therefore, for j 6= m we have
>
(m)
∗
(m)
(m),G
π
−π
P
−P
j
X (m)
(m)
(m),G
=
(πi − πi∗ ) Pi,j − Pi,j
i
(m)
(m)
(m),G
(m)
∗
+ (πm
− πm
) Pm,j − Pm,j
(m)
(m)
(m),G
(m)
(m),G
(m)
∗
= −(πj − πj∗ ) Pj,m − Pj,m
+ (πm
− πm
) Pm,j − Pm,j .
= (πj
(m)
(m),G
− πj∗ ) Pj,j − Pj,j
24
(52)
(m)
(m),G
Recognizing that Pj,m − Pj,m
≤
(m)
2
d
(m),G
for (j, m) ∈ E and Pj,m − Pj,m
≤
(
i
h
4
kπ (m) − π ∗ k∞ ,
≤ d2p (m)
π (m) − π ∗ (P (m) − P (m),G )
j
− π ∗ k∞ ,
d kπ
p
d
for (j, m) ∈
/ E, we have
if (j, m) ∈ E,
else.
(53)
And for j = m, it holds that
>
π (m) − π ∗
P (m) − P (m),G
m
X (m)
(m)
(m),G
(m)
∗
(m)
(m),G
+
(πj − πj∗ ) Pj,m − Pj,m
− πm
) Pm,m
− Pm,m
= (πm
(54)
j:j6=m
(m)
∗
(m)
(m),G
≤ (πm
− πm
) Pm,m
− Pm,m
X
+
(m)
(m)
(m),G
− πj∗ ) Pj,m − Pj,m
(πj
j:j6=m
=
X
(m)
(m),G
(m)
∗
(πm
− πm
) Pm,j − Pm,j
+
j:j6=m
(m)
(m),G
(m)
(πj
(m)
(m),G
− πj∗ ) Pj,m − Pj,m
.
(55)
j:j6=m
|
Given that Pm,j − Pm,j
X
{z
}
:=J3m
=
∗
ym,j
d (p
|
{z
:=J4m
}
− 1(m,j)∈E ), we have
J3m =
∗
ym,j
(m)
∗
(πm
− πm
)
(p − 1(m,j)∈E ).
d}
{z
j:j6=m |
X
(56)
(m)
:=ξj
Since kξ (m) k∞ ≤ d1 kπ (m) − π ∗ k∞ and kξ (m) k2 ≤ d1 kπ (m) − π ∗ k2 , Lemma 18 implies that
√
np log n + log n (m)
|J3m | .
kπ
− π ∗ k∞
d
with high probability. The same bound holds for J4m . Combine (53), (55) and (56) to arrive at
√ !
√
√
np
log
n
+
log
n
p
n
d
J2m .
+
+
kπ (m) − π ∗ k∞ .
d
d
d
Combining all, we deduce that
√
κ m m
(m)
(J1 +J2 )
kπ −πk2 ≤
γ
!
r
√ !
√
√
√
κ
logn (m)
nplogn+logn p n
d
(m)
∗
≤
8
kπ k∞ +C
+
+
kπ −π k∞
γ
Ld
d
d
d
!
r
r
√ !
√
√
√
κ
logn ∗
logn
nplogn+logn p n
d
≤
8
kπ k∞ +C 8
+
+
+
kπ (m) −π ∗ k∞
γ
Ld
Ld
d
d
d
√ r
(i) 8 κ
logn ∗
1
≤
kπ k∞ + kπ (m) −π ∗ k∞ ,
γ
Ld
2
where (i) holds as long as p ≥ c0 n log n for c0 sufficiently large. The triangle inequality
kπ (m) − π ∗ k∞ ≤ kπ (m) − πk2 + kπ − π ∗ k∞
yields
kπ
(m)
√ r
16 κ log n ∗
kπ k∞ + kπ − π ∗ k∞ ,
− πk2 ≤
γ
Ld
which concludes the proof.
25
(57)
A.5
Proof of Lemma 6
For ease of presentation, we define
L
ỹi,j :=
1 X (l)
ỹi,j
L
l=1
for all i 6= j, where
(
(l) ind.
ỹi,j =
1, with probability
wj∗
wi∗ +wj∗ ,
0, else.
This allows us to write y as yi,j = ỹi,j 1(i,j)∈E for all i 6= j. With this notation in place, we can obtain
!
L
X (m)
X
1
(l)
I4m =
(πj − πj∗ )
ỹj,m 1(j,m)∈E .
Ld
j:j6=m
We can further decompose
I4m
l=1
into
I4m = E [I4m | G−m , ỹ] + I4m − E [I4m | G−m , ỹ] ,
where G−m represent the graph without the m-th node, and ỹ = {ỹi,j |i 6= j} represents all the binary
outcomes.
We start with the expectation term
!
L
X (m)
1 X (l)
m
∗
ỹj,m P {(j, m) ∈ E}
E [I4 | G−m , ỹ] =
(πj − πj )
Ld
l=1
j:j6=m
√ (ii)
(i)
1
n
(m)
∗
≤ √ kπ (m) − π ∗ k2
≤ pkπ
− π k2
d
2 n
(iii)
≤
1
√ (kπ (m) − πk2 + kπ − π ∗ k2 ),
2 n
where (i) comes from the Cauchy-Schwarz inequality, (ii) follows from the choice d = cd np ≥ 2np and (iii)
results from the triangle inequality. By Theorem 6, with high probability we have
r
r
log n ∗
log n √
∗
nkπ ∗ k∞ ,
kπ − π k2 ≤ CN
kπ k2 ≤ CN
Ld
Ld
thus indicating that
E [I4m
1
CN
| G−m , ỹ] ≤ √ kπ (m) − πk2 +
2
2 n
r
log n ∗
kπ k∞ .
Ld
(58)
When it comes to the fluctuation term, one can write
I4m − E[I4m | G−m , ỹ] =
X
(m)
(πj
j:j6=m
|
!
L
X
1
(l)
− πj∗ )
ỹj,m 1(j,m)∈E −P {(j, m) ∈ E} .
Ld
l=1
{z
}
(m)
:=βj
Since kβ (m) k2 ≤ d1 kπ (m) − π ∗ k2 and kβ (m) k∞ ≤ d1 kπ (m) − π ∗ k∞ , one can apply Lemma 18 to derive
√
I4m − E I4m | G−m , ỹ .
np log n + log n (m)
kπ
− π ∗ k∞
d
with high probability.
The bounds (58) and (59) taken together complete the proof.
26
(59)
B
Proofs in Section 4
This section gathers the proofs of the lemmas in Section 4.
B.1
Proof of Lemma 7
Observe that
∗
X
∗
∇Lλ (θ ; y) = λθ +
(i,j)∈E,i>j
1
= λθ +
L
∗
∗
eθi
−yj,i + θ∗
∗
e i + eθ j
(ei − ej )
∗
L
X
eθ i
(l)
(ei − ej ).
−yj,i + θ∗
∗
e i + eθ j
(i,j)∈E,i>j l=1 |
{z
}
X
(l)
:=zi,j
(l)
(l)
It is seen that E[zi,j ] = 0, kzi,j k ≤
√
2,
h
i
h i
(l) (l)>
(l)
>
>
E zi,j zi,j = Var yi,j (ei − ej ) (ei − ej ) (ei − ej ) (ei − ej )
and
h
i
h
i
(l)> (l)
(l) (l)>
E zi,j zi,j = Tr E zi,j zi,j
≤ 2.
This implies that with high probability (note that the randomness comes from G),
X
L
h
i
X
(l) (l)>
≤L
E zi,j zi,j
(i,j)∈E,i>j l=1
X
>
(ei − ej ) (ei − ej )
= LkLG k . Lnp
(i,j)∈E,i>j
and
X
L
h
i
X
(l)> (l)
E zi,j zi,j ≤ 2L
(i,j)∈E,i>j l=1
n P
X
1 . Ln2 p.
(i,j)∈E,i>j
h
i
h
io
PL
(l) (l)>
(l)> (l)
(l)
Letting V := L12 max
E
z
z
,
E
z
z
and B := maxi,j,l kzi,j k,
i,j i,j
i,j
i,j
(i,j)∈E
l=1
(i,j)∈E
l=1
we can invoke the matrix Bernstein inequality (Tropp, 2012, Theorem 1.6) to reach
r
p
n2 p log n log n
∗
∗
k∇Lλ (θ ; y) − E [∇Lλ (θ ; y) | G]k2 . V log n + B log n .
+
L
L
r
2
n p log n
.
L
PL
P
with probability at least 1 − O(n−10 ). Combining this with the identity E [∇Lλ (θ ∗ ; y) | G] = λθ ∗ yields
k∇Lλ (θ ∗ ; y)k2 ≤ kE [∇Lλ (θ ∗ ; y) | G]k2 + k∇Lλ (θ ∗ ; y) − E [∇Lλ (θ ∗ ; y) | G]k2
r
r
n2 p log n
n2 p log n
∗
. λkθ k +
,
L
L
thus concluding the proof.
B.2
Proof of Lemma 12
First, Lemma 11 reveals that
kθ t − θk2 ≤ kθ 0 − θk2 = kθk2 ≤ kθ − θ ∗ k2 + kθ ∗ k2
27
for all t ≥ 0. Thus, by the triangle inequality we have
kθ t − θ ∗ k2 ≤ kθ t − θk2 + kθ − θ ∗ k2 ≤ 2kθ − θ ∗ k2 + kθ ∗ k2 .
Next, it follows from the optimality of θ as well as the mean value theorem that
Lλ (θ ∗ ; y) ≥ Lλ (θ; y) = Lλ (θ ∗ ; y) + h∇Lλ (θ ∗ ; y) , θ − θ ∗ i +
1
>
(θ − θ ∗ ) ∇2 Lλ θ̃; y (θ − θ ∗ ) ,
2
where θ̃ is between θ and θ ∗ . This together with the Cauchy-Schwarz inequality gives
1
>
(θ − θ ∗ ) ∇2 Lλ θ̃; y (θ − θ ∗ ) ≤ − h∇Lλ (θ ∗ ; y) , θ − θ ∗ i ≤ k∇Lλ (θ ∗ ; y)k2 kθ − θ ∗ k2 .
2
Recall from Fact 1 and the assumption (33) that 1> θ = 1> θ ∗ = 0. Then the above inequality gives
kθ − θ ∗ k2 ≤
2 k∇Lλ (θ ∗ ; y)k2
,
λmin,⊥ ∇2 Lλ θ̃; y
(60)
where λmin,⊥ (A) is the largest quantityof µ obeying
x> Ax ≥ µkxk22 for all x obeying 1> x = 0.
2
From the trivial lower bound λmin,⊥ ∇ Lλ θ̃; y
≥ λ, the preceding inequality gives
kθ − θ ∗ k2 ≤
2 k∇Lλ (θ ∗ ; y)k2
λ
(61)
and therefore
4 k∇Lλ (θ ∗ ; y)k2
+ kθ ∗ k2 .
λ
q
q
2
np log n
On the event A2 = k∇Lλ (θ ∗ ; y)k2 . n pLlog n and in the presence of the choice λ
, we
L
√
obtain kθ t − θ ∗ k2 ≤ C2 n for all t ≥ 0.
kθ t − θ ∗ k2 ≤
B.3
Proof of Lemma 13
The update rule of Algorithm 2 is given by
t+1
∗
t
∗
θm
− θm
= θm
− θm
− η ∇L θ t ; y + λθ t m .
(62)
From the gradient (34), we can rewrite (62) as
(
t+1
θm
−
∗
θm
=
t
θm
=
I1m
−
∗
θm
+
I2m
X
−η
i:(m,i)∈E
+
I3m
t
eθ m
−yi,m + θt
t
e i + eθ m
)
t
− ηλθm
+ I4m ,
where the four terms are defined as
X
∗
I1m = −η
−yi,m + yi,m
,
i:(m,i)∈E
)
(
)
t
t
θm
θm
X
e
e
∗
∗
,
= −η
−yi,m
+ θt
−p
−yi,m
+ θt
t
t
i + eθ m
i + eθm
e
e
i6=m
i:(m,i)∈E
(
)
t
θ
X
em
t
∗
∗
t
∗
= θm
− θm
− ηp
−yi,m
+ θt
− ηλ θm
− θm
,
t
θ
e i +e m
i6=m
I2m
I3m
(
X
∗
I4m = −ηλθm
.
28
(63)
Here, we set
∗
eθm
:= θ∗
∗ .
e i + eθm
It then comes down to analyzing the
preceding four terms.
3np
we have
First of all, on the event A0 := np
2 ≤ dmin ≤ dmax ≤ 2
∗
yi,m
I1m ≤ 2ηdmax ≤ 3ηnp
and
I2m ≤ η (2dmax + 2np) ≤ 5ηnp.
In addition, I4m is simple and bounded by
I4m ≤ ηλkθ ∗ k∞ .
Hence, we are only left with I3m .
To analyze I3m , we define the population loss
X
∗
L∗ (θ) :=
−yj,i
(θi − θj ) + log 1 + eθi −θj ,
i>j
whose gradient and Hessian can be computed as
X
∗
∇L∗ (θ) =
−yj,i
+
eθ i
θ
e i + eθ j
i>j
∇2 L∗ (θ) =
X
eθi eθj
i>j
(eθi + eθj )
(ei − ej ) ;
(64)
>
(ei − ej ) (ei − ej ) .
(65)
With these notation in place, one can simplify I3m as
t
∗
t
∗
I3m = θm
− θm
− ηp ∇L∗ (θ t ) m − ηλ θm
− θm
.
(66)
2
∗
∗
Furthermore, it is straightforward to see that ∇L (θ ) = 0, which together with Taylor’s theorem yields
[∇L∗ (θ t )]m = [∇L∗ (θ t )]m − [∇L∗ (θ ∗ )]m = [∇2 L∗ θ̃ t ]m,· θ t − θ ∗ ,
where θ˜t lies between θ t and θ ∗ . Combine the preceding two identities to get
h
h
i
X
i
m
2 ∗
t
t
∗
−ηp ∇2 L∗ θ̃ t
I3 = (1 − ηλ) − ηp ∇ L θ̃
θm − θm +
m,m
m,i
i6=m
{z
|
}
:=J1m
{z
|
:=J2m
θit − θi∗ .
}
Therefore, it remains to control J1m and J2m .
For J1m , we have
h
i
J1m ≤ (1 − ηλ) − ηp ∇2 L∗ θ̃ t
h
i
Hence we need to lower bound ∇2 L∗ θ̃ t
m,m
on the event A2 and, as a result,
m,m
t
∗
· θm
− θm
.
√
. Recall from Lemma 12 that we have kθ t − θ ∗ k2 ≤ C2 n
√
kθ˜t − θ ∗ k2 ≤ C2 n.
A straightforward consequence is that1 at least 60% of the entries in θ̃ t − θ ∗ must obey θ̃it − θi∗ ≤
With this observation in mind, one has
t
h
i
∇2 L∗ θ̃ t
m,m
=
t
i:i6=m
t
eθ̃i + eθ̃m
we would have kθ˜t − θ ∗ k22 ≥ 0.4n ×
√
kθ˜t − θ ∗ k2 ≤ C2 n.
1 Otherwise,
√
√
t
eθ̃i eθ̃m
X
3C2
2 ≥ 0.6 (n − 1)
2
29
e−2
1+
√
3C2 .
3C2
√
2 .
e−2 3C2
= 1.2C22 n > C22 n, which is contradictory to the fact that
This gives
√
J1m
≤
1 − 0.3ηnp
e−2
!
3C2
√
1 + e−2
t
∗
θm
.
− θm
3C2 2
When it comes to J2m , we see that
J2m
X h
i
∇2 L∗ θ̃ t
= ηp
m,i
i:i6=m
1/2
≤ ηp
θit − θi∗
X h
∇2 L∗ θ̃
i 2
t
m,i
i:i6=m
1/2
h
∇2 L∗ θ̃ t
i
i:i6=m
1/2
X h
∇2 L∗ θ̃
i 2
t
m,i
i:i6=m
Hence we need to upper bound
θit − θi
≤ ηpkθ t − θ ∗ k2
∗ 2
X
.
for i 6= m. By (65), we have
m,i
t
h
∇2 L∗ θ̃ t
i
=
m,i
t
eθ̃i eθ̃m
t
eθ̃i
+
2
t
eθ̃m
≤
1
,
4
which gives us
√
C2
1 √
ηp n − 1C2 n ≤
ηnp.
4
4
Combing all of the above bounds, we deduce that
!
√
e−2 3C2
C2
t+1
∗
t
∗
∗
θm − θm ≤ 1 − 0.3ηnp
ηnp
√
2 θm − θm + 8ηnp + ηλkθ k∞ +
−2
3C
4
2
1+e
|J2m | ≤
on the event A0 ∩ A2 . This finishes the proof.
B.4
Proof of Lemma 14
We first provide here a proof for the bound on kθ − θ ∗ k2 .
Recall from (60) that
2 k∇Lλ (θ ∗ ; y)k2
,
kθ − θ ∗ k2 ≤
λmin,⊥ ∇2 Lλ θ̃; y
(67)
where θ̃ is between θ and θ ∗ . We have also learned from Section 4.2.1 that kθ − θ ∗ k∞ . 1, and hence
θ̃ − θ ∗
∞
≤ kθ − θ ∗ k∞ . 1.
(68)
This together with Corollary 1 reveals that
λmin,⊥ ∇2 Lλ θ̃; y
& np,
and therefore
kθ − θ ∗ k2 .
k∇Lλ (θ ∗ ; y)k2
.
np
Therefore, combining Lemma 7 with (67) leads to
k∇Lλ (θ ∗ ; y)k2
kθ ∗ − θk2 .
.
np
30
s
log n
pL
(69)
as claimed.
The proof concerning θ (m) − θ ∗
that
2
follows an almost identical argument, except that we need to establish
kθ (m) − θ ∗ k∞ . 1.
And this can be easily checked using the same argument in Section 4.2.1.
B.5
Proof of Lemma 15
Repeating the argument in Appendix B.4 gives
kθ
(m)
∇Lλ θ (m) ; y
2 ∇Lλ θ (m) ; y 2
.
− θk2 ≤
np
λmin ∇2 Lλ θ̃ (m) ; y
It then suffices to upper bound ∇Lλ θ (m) ; y
2
.
(70)
. Recognizing that
(m)
∇Lλ θ (m) ; y = 0,
2
we can deduce that
(m)
∇Lλ θ (m) ; y = ∇Lλ θ (m) ; y − ∇Lλ θ (m) ; y
(
!
!)
(m)
(m)
∗
X
eθ i
eθi
eθ i
=
−ym,i + (m)
1{(i,m)∈E} −p − θ∗
(ei − em )
∗ +
(m)
(m)
(m)
e i + eθ m
eθi + eθm
eθi + eθm
i:i6=m
∗
L
1 X X
eθi
(l)
=
−ym,i + θ∗
(ei − em )
∗
L
e i + eθ m
i:(i,m)∈E l=1
|
{z
}
:=um
(
!
)
(m)
∗
X
eθ i
eθ i
(71)
+
− θ∗
1{(i,m)∈E} −p (ei − em ).
∗ +
(m)
(m)
e i + eθ m
eθi + eθm
i:i6=m
|
{z
}
:=v m
In the sequel, we control the two terms of (71) separately.
For the first term um in (71), we make the observation that
P
∗
(l)
L
1
eθi
,
if (i, m) ∈ E;
∗
L l=1 −ym,i + eθi∗ +eθm
∗
P
P
θ
m
(l)
e i
ui = 1 i:(i,m)∈E L
, if i = m;
∗
l=1 ym,i − eθi∗ +eθm
L
0,
else.
(l)
θ∗
Since ym,i − θi∗e i θm
≤ 1 and card ({i : (i, m) ∈ E}) np, we can apply Hoeffding’s inequality and union
∗
e +e
bounds to get for all 1 ≤ m ≤ n,
r
r
np log n
log n
m
m
|um | .
and
|ui | .
∀i obeying (i, m) ∈ E,
L
L
which further gives
m
ku k2 ≤
|um
m|
+
s X
r
2
(um
i ) .
i:(i,m)∈E
np log n
, ∀1 ≤ m ≤ n.
L
We then turn to the second term v m of (71). This is a zero-mean random vector that satisfies
− p) ,
if (i, m) ∈ E,
ξi (1
P
m
vi = − i:i6=m ξi 1{(i,m)∈E} −p , if i = m,
−ξi p,
else,
31
where
(m)
∗
1
eθ i
1
eθ i
=−
+
ξi := − θ∗
.
∗ −θ ∗ +
∗
(m)
(m)
(m)
(m)
θm
i
θ
θ
θ
e i + eθ m
1
+
e
e i +e m
1 + e m −θi
The first step is to bound the size of the coefficient ξi . Define g(x) = (1 + ex )−1 for x ∈ R. We have
|g 0 (x)| ≤ 1 and thus
(m)
(m)
|ξi | = |g(θm
− θi
(m)
∗
(m)
) − g(θm
− θi∗ )| ≤ |(θm
− θi
This indicates that
|ξi | ≤ 2kθ (m) − θ ∗ k∞
(m)
∗
) − (θm
− θi∗ )| ≤ θi∗ − θi
n
X
and
∗
(m)
.
− θm
+ θm
ξi2 ≤ 4nkθ (m) − θ ∗ k2∞ .
i=1
Applying the Bernstein inequality in Lemma 18 we obtain
v
!
u
n
p
u X
m
2
t
|vm | .
p
np log n + log n kθ (m) − θ ∗ k∞
ξi log n + max |ξi | log n .
1≤i≤n
i=1
with high probability. As a consequence,
s
s X
m
(vim )2 +
kv m k2 ≤ |vm
|+
X
(vim )2
i:(i,m)∈E
/ and i6=m
i:(i,m)∈E
p
√
√
.
np log n + log n kθ (m) − θ ∗ k∞ + npkθ (m) − θ ∗ k∞ + p nkθ (m) − θ ∗ k∞
p
.
np log n + log n kθ (m) − θ ∗ k∞ .
Putting the above results together, we see that
kθ (m) − θk2 .
.
kum k2 + kv m k2
np
q
√
np log n
+
np log n + log n kθ (m) − θ ∗ k∞
L
np
(72)
.
(73)
Upper bounding kθ (m) − θ ∗ k∞ by kθ − θ ∗ k∞ + kθ − θ (m) k, one has
q
√
(m)
√
np log n +
np log n + log n kθ − θ ∗ k∞
np
log
n
+
log
n
kθ
−
θk
L
2
kθ (m) − θk2 ≤ c1
+
np
np
c0 log n
for some
for some sufficiently large constant c0 > 0 so that np >
n
√ constant c1 > 0. When p >
4c1
np log n + log n , the above inequality guarantees that
c1
√ np log n
L
kθ (m) − θk2 ≤
1
B.6
√
+c1 ( np log n+log n)kθ−θ ∗ k∞
np
√
c1 ( np log n+log n)
−
np
≤
4c1
q
np log n
L
3np
1
+ kθ − θ ∗ k∞ .
3
Proof of Lemma 16
(m)
From the definition of θ (m) , we know that θm
entries of θ (m) , that is,
is also the coordinate-wise optimizer when given all other
(m)
θm
= arg min `(m) (τ ; y) ,
32
where
`(m) (τ ; y) = p
X
−
i:i6=m
∗
λ
(m)
eθ i
(m)
θi −τ
+ τ2
θ
−
τ
+
log
1
+
e
∗
∗
2
eθi + eθm i
is the coordinate-wise MLE when fixing all other entries of θ (m) . Similar to (60), we obtain from the
optimality condition that
d`(m) (τ ;y)
dτ
2
(m)
∗
≤
θm
− θm
dτ 2
where θ̃m is some quantity between
Observing that
d`(m) (τ ; y)
dτ
and that
∗
θm
and
(74)
τ =θ̃m
(m)
θm .
(
X
=p
∗
τ =θm
∗
τ =θm
d2 `(m) (τ ;y)
i:i6=m
(m)
∗
eθ i
eθ i
∗
∗ −
(m)
θ
θ
∗
e i +e m
eθ i + eθ m
)
∗
+ λθm
(m)
∗
eθ i
1
1
eθi
(m)
−
=
≤ θi∗ − θi ,
∗
∗ −θ ∗ −
∗
(m)
θm
∗ −θ (m)
∗
i
θ
θ
θ
eθi + eθm
1
+
e
e i +e m
1+e m i
we arrive at
d`(m) (τ ; y)
dτ
≤p
∗
τ =θm
X
(m)
θi∗ − θi
i:i6=m
r
(ii)
np log n
,
.
L
(i)
∗
+ λθm
≤ pn
kθ ∗ − θ (m) k2
∗
√
+ λθm
n
where (i) follows from Cauchy-Schwarz and (ii) arises from Lemma 14. Additionally, `(m) (τ ; y) is strongly
convex in τ since
(m)
X
d2 `(m) (τ ; y)
eτ eθ i
λ
=p
2 + & np.
2
(m)
dτ
2
τ =θ̃m
i:i6=m eτ + eθi
These taken collectively with (74) give
s
(m)
∗
θm
− θm
.
C
log n
.
npL
Hoeffding’s and Bernstein’s inequalities
This section collects two standard concentration inequalities used throughout the paper, which can be easily
found in textbooks such as Boucheron et al. (2013). The proofs are omitted.
Lemma 17 (Hoeffding’s inequality). Let P
{Xi }1≤i≤n be a sequence of independent random variables where
n
Xi ∈ [ai , bi ] for each 1 ≤ i ≤ n, and Sn = i=1 Xi . Then
P(|Sn − E [Sn ] | ≥ t) ≤ 2e−2t
2
/
Pn
2
i=1 (bi −ai )
.
The next lemma is about a user-friendly version of the Bernstein inequality.
Lemma 18 (Bernstein’s inequality). Consider n independent random variables zl (1 ≤ l ≤ n), each satisfying |zl | ≤ B. For any a ≥ 2, one has
v
" n #
u
n
n
X
X
X
u
2a
E [zl2 ] + B log n
zl − E
zl ≤ t2a log n
3
l=1
l=1
l=1
with probability at least 1 − 2n−a .
33
D
Proof of Theorem 6
By Theorem 5, we obtain
kπ ∗ − πkπ∗ ≤
π ∗> (P ∗ − P )
(i)
π∗
1 − max {λ2 (P ∗ ), −λn (P ∗ )} − P ∗ − P
(ii)
(iii)
. kπ ∗> (P ∗ − P )k2 . √
. π ∗> (P ∗ − P )
π∗
π∗
1
kπ ∗ k2 ,
npL
where (i) is a consequence of Lemma 5, (ii) follows from the relationship between k · k2 and k · kπ∗ , and (iii)
follows as long as one can justify that
π ∗> (P − P ∗ )
2
.√
1
kπ ∗ k2 .
npL
(75)
Therefore, the rest of the proof is devoted to establishing (75). To simplify the notations hereafter, we denote
∆ := P − P ∗ . In fact, it is easy to check that for any i 6= j,
!
L
1 1 X (l)
∗
∗
yi,j − yi,j
(76)
∆i,j = Pi,j − Pi,j = 1{(i,j)∈E}
d L
l=1
and for 1 ≤ i ≤ n, one has
X
∗
∆i,i = Pi,i − Pi,i
= 1 −
Pi,j − 1 −
j:j6=i
=−
X
j:j6=i
∗
Pi,j
j:j6=i
∗
Pi,j − Pi,j
=−
X
X
∆i,j
j:j6=i
= −∆lower
− ∆upper
.
i,i
i,i
(77)
where
∆lower
:=
i,i
X
∆i,j
and
∆upper
:=
i,i
j:j<i
X
∆i,j .
j:j>i
Towards proving (75), we decompose ∆ into four parts ∆ = ∆lower + ∆upper + ∆diag,l + ∆diag,u , where
∆lower is the lower triangular part (excluding the diagonal) of ∆, ∆upper is the upper triangular part, and
∆diag,l = −diag [∆lower
]1≤i≤n
and
∆diag,u = −diag [∆upper
]1≤i≤n .
i,i
i,i
The triangle inequality then gives
π ∗> ∆
2
≤ π ∗> ∆lower 2 + π ∗> ∆upper 2 + π ∗> ∆diag,l 2 + π ∗> ∆diag,u 2 .
{z
} |
{z
} |
{z
} |
{z
}
|
:=Ilower
:=Iupper
:=Idiag,l
:=Idiag,u
In what follows, we will focus on controlling the first term Ilower . The other three terms can be bounded
using nearly identical arguments.
Note that the j component of π ∗> ∆lower can be expressed as
X
∗>
π ∆lower j =
πi∗ ∆i,j .
i:i>j
Recall
∈ E, ∆i,j is a sum of L independent zero-mean random variables, and hence
∗> that for any pair (i, j)lower
π ∆lower j is a sum of Ldj
independent zero-mean random variables, where
dlower
:= {(i, j) | (i, j) ∈ E and i > j} .
j
34
In view of Hoeffding’s inequality (Lemma 17), one has, when conditional on G, that
!
2
2t
π ∗> ∆lower j ≥ t ≤ 2 exp − 1 lower
.
P
2
L (2 kπ ∗ k∞ )
(Ld)2 dj
Hence π ∗> ∆lower j can be treated as a sub-Gaussian random variable with variance proxy
σ2
dlower
1
j
2
2
kπ ∗ k∞ .
kπ ∗ k∞ .
2
d L
dL
Given that the entries of π ∗> ∆lower are independent, we see that
2
(Ilower ) =
n
X
∗>
2
π ∆lower j ,
j=1
is a quadratic form of a sub-Gaussian vector. On the one hand, E[Ilower ]2 . nσ 2 . On the other hand, we
invoke (?, Theorem 1.1) to reach
2
t
t
2
2
,
P{Ilower
− E[Ilower
] ≥ t} ≤ 2 exp −c min
nσ 4 σ 2
√
for some constant c > 0. By choosing t σ 2 n log n, we see that with probability at least 1 − O(n−10 ),
2
2
Ilower
. E[Ilower
] + σ2
p
n
1
2
2
n log n . nσ 2 .
kπ ∗ k∞ .
kπ ∗ k2 .
dL
dL
The same upper bounds can be derived for other terms using the same arguments. We have thus established
(75) by recognizing that d & np.
References
Abbe, E., Fan, J., Wang, K. and Zhong, Y. (2017). Eigenvectors are sharp: Entry-wise perturbation
analysis. working paper .
Agarwal, A., Agarwal, S., Assadi, S. and Khanna, S. (2017). Learning with limited rounds of adaptivity: Coin tossing, multi-armed bandits, and ranking from pairwise comparisons. In Conference on
Learning Theory.
Agresti, A. (2013). Categorical data analysis. 3rd ed. Wiley Series in Probability and Statistics, WileyInterscience [John Wiley & Sons], Hoboken, NJ.
Ailon, N. (2012). An active learning algorithm for ranking from pairwise preferences with an almost optimal
query complexity. J. Mach. Learn. Res. 13 137–164.
Ammar, A. and Shah, D. (2011). Ranking: Compare, don’t score. In 2011 49th Annual Allerton Conference
on Communication, Control, and Computing (Allerton).
Baltrunas, L., Makcinskas, T. and Ricci, F. (2010). Group recommendations with rank aggregation
and collaborative filtering. In Proceedings of the Fourth ACM Conference on Recommender Systems.
RecSys ’10, ACM, New York, NY, USA.
Borda, J. d. (1784). Mémoire sur les élections au scrutin. Histoire de l’Academie Royale des Sciences pour
1781 (Paris, 1784) .
Borkar, V. S., Karamchandani, N. and Mirani, S. (2016). Randomized kaczmarz for rank aggregation
from pairwise comparisons. In 2016 IEEE Information Theory Workshop (ITW).
35
Boucheron, S., Lugosi, G. and Massart, P. (2013). Concentration inequalities. Oxford University
Press, Oxford. A nonasymptotic theory of independence, With a foreword by Michel Ledoux.
Bradley, R. A. and Terry, M. E. (1952). Rank analysis of incomplete block designs. I. The method of
paired comparisons. Biometrika 39 324–345.
Braverman, M. and Mossel, E. (2008). Noisy sorting without resampling. In Proceedings of the Nineteenth
Annual ACM-SIAM Symposium on Discrete Algorithms. ACM, New York.
Brémaud, P. (1999). Markov chains, vol. 31 of Texts in Applied Mathematics. Springer-Verlag, New York.
Gibbs fields, Monte Carlo simulation, and queues.
Bubeck, S. (2015). Convex optimization: Algorithms and complexity. Found. Trends Mach. Learn. 8
231–357.
Busa-Fekete, R., Szörényi, B., Weng, P., Cheng, W. and Hüllermeier, E. (2013). Top-k selection
based on adaptive sampling of noisy preferences. In Proceedings of the 30th International Conference on
International Conference on Machine Learning - Volume 28. ICML’13, JMLR.org.
Candès, E. J., Li, X. and Soltanolkotabi, M. (2015). Phase retrieval via Wirtinger flow: theory and
algorithms. IEEE Trans. Inform. Theory 61 1985–2007.
Chaudhuri, K., Chung, F. and Tsiatas, A. (2012). Spectral clustering of graphs with general degrees in
the extended planted partition model. In Conference on Learning Theory.
Chen, X., Bennett, P. N., Collins-Thompson, K. and Horvitz, E. (2013). Pairwise ranking aggregation in a crowdsourced setting. In Proceedings of the Sixth ACM International Conference on Web Search
and Data Mining. WSDM ’13, ACM, New York, NY, USA.
Chen, X., Gopi, S., Mao, J. and Schneider, J. (2017). Competitive analysis of the top-k ranking
problem. In Proceedings of the Twenty-Eighth Annual ACM-SIAM Symposium on Discrete Algorithms.
SODA ’17, Society for Industrial and Applied Mathematics, Philadelphia, PA, USA.
Chen, Y. and Candes, E. (2016). The projected power method: An efficient algorithm for joint alignment
from pairwise differences. arXiv preprint arXiv:1609.05820 .
Chen, Y. and Candès, E. J. (2017). Solving random quadratic systems of equations is nearly as easy as
solving linear systems. Comm. Pure Appl. Math. 70 822–883.
Chen, Y. and Suh, C. (2015). Spectral mle: Top-k rank aggregation from pairwise comparisons. In
Proceedings of the 32Nd International Conference on International Conference on Machine Learning Volume 37. ICML’15, JMLR.org.
Chung, F. R. K. (1997). Spectral graph theory, vol. 92 of CBMS Regional Conference Series in Mathematics.
Published for the Conference Board of the Mathematical Sciences, Washington, DC; by the American
Mathematical Society, Providence, RI.
Davis, C. and Kahan, W. M. (1970). The rotation of eigenvectors by a perturbation. III. SIAM J. Numer.
Anal. 7 1–46.
Durrett, R. (2007). Random graph dynamics, vol. 20 of Cambridge Series in Statistical and Probabilistic
Mathematics. Cambridge University Press, Cambridge.
Dwork, C., Kumar, R., Naor, M. and Sivakumar, D. (2001). Rank aggregation methods for the web.
In Proceedings of the 10th International Conference on World Wide Web. WWW ’01, ACM, New York,
NY, USA.
El Karoui, N. (2017). On the impact of predictor geometry on the performance on high-dimensional
ridge-regularized generalized robust regression estimators. Probability Theory and Related Fields .
36
El Karoui, N., Bean, D., Bickel, P. J., Lim, C. and Yu, B. (2013). On robust regression with
high-dimensional predictors. Proceedings of the National Academy of Sciences 110 14557–14562.
Fan, J., Wang, W. and Zhong, Y. (2016). An `∞ eigenvector perturbation bound and its application to
robust covariance estimation. arXiv preprint arXiv:1603.03516 .
Ford, L. R., Jr. (1957). Solution of a ranking problem from binary comparisons. Amer. Math. Monthly
64 28–33.
Hajek, B., Oh, S. and Xu, J. (2014). Minimax-optimal inference from partial rankings. In Proceedings
of the 27th International Conference on Neural Information Processing Systems. NIPS’14, MIT Press,
Cambridge, MA, USA.
Heckel, R., Shah, N. B., Ramchandran, K. and Wainwright, M. J. (2016). Active ranking from
pairwise comparisons and when parametric assumptions don’t help. arXiv preprint arXiv:1606.08842 .
Hunter, D. R. (2004). MM algorithms for generalized Bradley-Terry models. Ann. Statist. 32 384–406.
Jain, P., Netrapalli, P. and Sanghavi, S. (2013). Low-rank matrix completion using alternating minimization (extended abstract). In STOC’13—Proceedings of the 2013 ACM Symposium on Theory of
Computing. ACM, New York, 665–674.
Jamieson, K. G. and Nowak, R. D. (2011). Active ranking using pairwise comparisons. In Proceedings of
the 24th International Conference on Neural Information Processing Systems. NIPS’11, Curran Associates
Inc., USA.
Jang, M., Kim, S., Suh, C. and Oh, S. (2016). Top-k ranking from pairwise comparisons: When spectral
ranking is optimal. arXiv preprint arXiv:1603.04153 .
Javanmard, A. and Montanari, A. (2015). De-biasing the lasso: Optimal sample size for gaussian designs.
arXiv preprint arXiv:1508.02757 .
Jiang, X., Lim, L.-H., Yao, Y. and Ye, Y. (2011). Statistical ranking and combinatorial Hodge theory.
Math. Program. 127 203–244.
Keshavan, R. H., Montanari, A. and Oh, S. (2010). Matrix completion from noisy entries. J. Mach.
Learn. Res. 11 2057–2078.
Lu, Y. and Negahban, S. N. (2014). Individualized rank aggregation using nuclear norm regularization.
arXiv preprint arXiv:1410.0860 .
Luce, R. D. (1959). Individual choice behavior: A theoretical analysis. John Wiley & Sons, Inc., New York;
Chapman & Hall, Ltd., London.
Masse, K. (1997). Statistical models applied to the rating of sports teams. Technical report Bluefield College
.
Mohajer, S. and Suh, C. (2016). Active top-k ranking from noisy comparisons. In 2016 54th Annual
Allerton Conference on Communication, Control, and Computing (Allerton).
Negahban, S., Oh, S. and Shah, D. (2017a). Rank centrality: ranking from pairwise comparisons. Oper.
Res. 65 266–287.
Negahban, S., Oh, S., Thekumparampil, K. K. and Xu, J. (2017b). Learning from comparisons and
choices. arXiv preprint arXiv:1704.07228 .
Oh, S. and Shah, D. (2014). Learning mixed multinomial logit model from ordinal data. In Proceedings
of the 27th International Conference on Neural Information Processing Systems. NIPS’14, MIT Press,
Cambridge, MA, USA.
37
Osting, B., Xiong, J., Xu, Q. and Yao, Y. (2016). Analysis of crowdsourced sampling strategies for
HodgeRank with sparse random graphs. Appl. Comput. Harmon. Anal. 41 540–560.
Page, L., Brin, S., Motwani, R. and Winograd, T. (1999). The pagerank citation ranking: Bringing
order to the web. Tech. rep., Stanford InfoLab.
Pananjady, A., Mao, C., Muthukumar, V., Wainwright, M. J. and Courtade, T. A. (2017).
Worst-case vs average-case design for estimation from fixed pairwise comparisons. arXiv preprint
arXiv:1707.06217 .
Park, D., Neeman, J., Zhang, J., Sanghavi, S. and Dhillon, I. S. (2015). Preference completion:
Large-scale collaborative ranking from pairwise comparisons. In Proceedings of the 32Nd International
Conference on International Conference on Machine Learning - Volume 37. ICML’15, JMLR.org.
Rajkumar, A. and Agarwal, S. (2014). A statistical convergence perspective of algorithms for rank
aggregation from pairwise data. In Proceedings of the 31st International Conference on International
Conference on Machine Learning - Volume 32. ICML’14, JMLR.org.
Rajkumar, A. and Agarwal, S. (2016). When can we rank well from comparisons of o(n log n) non-actively
chosen pairs? In 29th Annual Conference on Learning Theory.
Rohe, K., Chatterjee, S. and Yu, B. (2011). Spectral clustering and the high-dimensional stochastic
blockmodel. Ann. Statist. 39 1878–1915.
Shah, N. B., Balakrishnan, S., Bradley, J., Parekh, A., Ramchandran, K. and Wainwright, M.
(2014). When is it better to compare than to score? arXiv:1406.6618 .
Shah, N. B., Balakrishnan, S., Guntuboyina, A. and Wainwright, M. J. (2017). Stochastically
transitive models for pairwise comparisons: statistical and computational issues. IEEE Trans. Inform.
Theory 63 934–959.
Shah, N. B. and Wainwright, M. J. (2015). Simple, robust and optimal ranking from pairwise comparisons. arXiv preprint arXiv:1512.08949 .
Singer, A. (2011). Angular synchronization by eigenvectors and semidefinite programming. Appl. Comput.
Harmon. Anal. 30 20–36.
Soufiani, H. A., Chen, W. Z., Parkes, D. C. and Xia, L. (2013). Generalized method-of-moments for
rank aggregation. In Proceedings of the 26th International Conference on Neural Information Processing
Systems. NIPS’13, Curran Associates Inc., USA.
Soufiani, H. A., Parkes, D. C. and Xia, L. (2014). Computing parametric ranking models via rankbreaking. In Proceedings of the 31st International Conference on International Conference on Machine
Learning - Volume 32. ICML’14, JMLR.org.
Suh, C., Tan, V. Y. F. and Zhao, R. (2017). Adversarial top-K ranking. IEEE Trans. Inform. Theory
63 2201–2225.
Sur, P., Chen, Y. and Candès, E. J. (2017). The likelihood ratio test in high-dimensional logistic
regression is asymptotically a rescaled Chi-square. arXiv preprint arXiv:1706.01191 .
Tropp, J. A. (2012). User-friendly tail bounds for sums of random matrices. Found. Comput. Math. 12
389–434.
Tropp, J. A. (2015). An introduction to matrix concentration inequalities. Found. Trends Mach. Learn. 8
1–230.
Wu, R., Xu, J., Srikant, R., Massoulie, L., Lelarge, M. and Hajek, B. (2015). Clustering and
inference from pairwise comparisons. In Proceedings of the 2015 ACM SIGMETRICS International Conference on Measurement and Modeling of Computer Systems. SIGMETRICS ’15, ACM, New York, NY,
USA.
38
Xu, Q., Huang, Q., Jiang, T., Yan, B., Lin, W. and Yao, Y. (2012). Hodgerank on random graphs for
subjective video quality assessment. IEEE Transactions on Multimedia 14 844–857.
Zhong, Y. and Boumal, N. (2017). Near-optimal bounds for phase synchronization. arXiv preprint
arXiv:1703.06605 .
39
| 10math.ST |
Subsets and Splits
Random Sampled ArXiv Labels
Provides a random sample of 100 rows, 20 from each of 5 labels, which is useful for a basic overview but does not reveal deep insights.
Random 20 Rows Per Class
Provides a basic sample of 20 rows per label, useful for a quick overview but does not offer deep insights.